A concrete surface crack detection method and system based on machine learning algorithm

Abstract

The present invention relates to a concrete surface crack detection method and system based on a machine learning algorithm, which obtains concrete wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient, analyzes characteristic parameters, establishes a mathematical relationship model, fits data using a machine learning algorithm, determines model parameters, inputs relevant parameters of the concrete to be detected into a multi-parameter digital quantity plate model, and the model calculates the potential risk and predicted value of the crack. The multi-parameter digital quantity plate model is applied to real-time detection of concrete structures to realize quality assessment and crack prediction, and the concrete structure is regularly monitored to update model data. The system includes data acquisition, data preprocessing, feature extraction, crack measurement, multi-parameter digital quantity plate model establishment, multi-parameter comprehensive analysis, real-time monitoring and early warning, database management, and user interface and interaction modules.

Description

Technical Field

[0001] The present invention relates to the technical field of concrete surface crack detection, and in particular to a concrete surface crack detection method and system based on a machine learning algorithm. Background Art

[0002] With the economic growth and the vigorous development of transportation infrastructure construction, the health of various bridges and dams is closely related to the traffic environment and the safety of pedestrians and vehicles. The surfaces of bridges, dams, etc. are all concrete structures. Common diseases developed in concrete include cracks, honeycombs, pits and pits. The treatment method should be to strengthen until they are eradicated. Among the cracks, honeycombs, pits and pits of concrete structure diseases, cracks are the most serious and threatening. If the disease is not treated in a timely and thorough manner, it will not only affect the normal operation of bridges and dams, but once the disease is further developed, it will increase the difficulty of maintenance.

[0003] In response to the above problems, Chinese patent CN118015412A discloses a method and device for detecting cracks on the concrete surface based on a machine learning algorithm, which relates to the field of concrete structure health detection, in order to solve the problem that the current traditional manual naked eye detection and digital image processing technology are mainly used. Manual detection is simple to operate, but it is time-consuming and labor-intensive, requires high experience from technicians, and is easily affected by subjective factors of technicians, resulting in the problem of low accuracy of measurement results. Compared with the prior art, this invention patent first uses a computer vision algorithm to preliminarily determine whether there are cracks on the concrete surface, and then further pre-processes the image with cracks, so as to calculate the maximum width of the cracks on the concrete surface using the pixel matrix of the pre-processed image, thereby realizing rapid identification of concrete cracks and having stronger operability and accuracy. However, during the use of this patent, the morphology of concrete cracks may be very complex and diverse, including straight lines, curves, bifurcated types, etc., and may be accompanied by interference factors such as stains and water stains. The computer vision algorithm is insufficient in its ability to recognize cracks of various shapes.

[0004] Chinese patent CN111242851B discloses a method and device for detecting cracks on the surface of concrete structures, which relates to the technical field of concrete crack detection, in order to solve the problem that the current image-based crack detection method is mainly based on the processing process at the whole pixel level, and cannot identify fine cracks at the sub-pixel level, and has low detection accuracy. This invention patent adopts the regional image of the surface of the concrete structure to be tested by an industrial camera to realize the remote and long-term detection of the width, length and distribution of cracks within a certain area of ​​the surface of the structure, and uses the regional image collected by the low-resolution industrial camera and the crack detection algorithm based on the seepage model to detect cracks on the panoramic image, which can detect fine cracks larger than 1 / 5 of the pixel size, greatly improving the crack detection resolution. Uneven illumination and image quality, limitations of crack detection algorithms, complexity of gray value adjustment, dependence on verification and correction, real-time and automation, and adaptability to complex environments. However, during the use of this patent, although the entire process is automated, manual intervention or adjustment is still required in gray value allocation, broken connections, etc., and the result verification and correction are highly dependent on manual operation, which not only increases the workload and cost, but also may introduce human errors.

[0005] Therefore, the invention team proposed a method and system for detecting concrete surface cracks based on machine learning algorithm. Summary of the invention

[0006] The purpose of the present invention is to provide a method and system for detecting concrete surface cracks based on a machine learning algorithm.

[0007] On the one hand, the present invention provides a method for detecting concrete surface cracks based on a machine learning algorithm, comprising the following steps:

[0008] S1 data acquisition: obtain the concrete wave velocity, frequency, amplitude characteristic parameters, and obtain the mixed soil strength, elastic modulus, temperature, and linear expansion coefficient;

[0009] S2 Data preprocessing: perform data cleaning, remove invalid data, standardize wave velocity, frequency, amplitude, mixed soil strength, elastic modulus, temperature, and linear expansion coefficient, and eliminate the impact of different testing equipment and testing conditions on the data;

[0010] S3 characteristic parameter analysis: analyze the relationship between wave velocity and concrete strength and elastic modulus, evaluate the density and integrity of concrete, analyze the changing rules of frequency and amplitude, identify defects and cracks inside concrete, record temperature data during detection, and analyze the impact of temperature on concrete performance in combination with the linear expansion coefficient of concrete;

[0011] S4 multi-parameter digital panel model establishment: Based on wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient, a mathematical relationship model is established between them. Statistical analysis and machine learning algorithms are used to fit the collected data and determine the model parameters;

[0012] S5 Crack prediction and assessment: The collected relevant data parameters of the concrete to be tested are input into the multi-parameter digital plate model, and the model calculates the potential risk and prediction value of the cracks based on the input parameters;

[0013] S6 model application and feedback: Apply the multi-parameter digital plate model to the real-time detection of hydraulic concrete structures to achieve quality assessment and crack prediction, regularly monitor the concrete structure, update the model data, and improve the accuracy and adaptability of the model.

[0014] Furthermore, in the S1 data collection, an ultrasonic detector is used to measure the ultrasonic wave propagation speed in concrete, a tester capable of generating and detecting impact elastic waves is used to analyze the dynamic response characteristics of concrete, including wave velocity, frequency, amplitude, strength and elastic modulus, a radar wave detector is selected to identify tiny cracks and voids in concrete, and a temperature sensor is used to measure the temperature of the concrete surface. The linear expansion coefficient is an inherent property of concrete material, which is obtained through experimental measurement or reference to literature. An ultrasonic detector, impact elastic wave tester, radar wave detector, high-definition camera and temperature sensor are arranged on the concrete surface to be tested to collect raw data.

[0015] When collecting data, the specific inspection locations are determined according to the structural characteristics and inspection requirements of the dam. The key structural points of the dam, such as the dam body, dam foundation, dam shoulders and spillway, are usually the focus of inspection. These locations are subject to greater water pressure and load, and are prone to defects such as cracks and voids. If the dam has had historical problems such as cracks and leakage in the past, these areas should be the focus of inspection. By re-inspecting these areas, the effectiveness of repair measures can be evaluated and the recurrence of potential problems can be monitored. A variety of materials may be used in the dam, such as concrete, rock, and soil-rock mixtures. Due to the different physical and chemical properties of these materials, cracks and seepage channels are prone to occur at the junctions of these materials. These locations are also the focus of inspection. The appropriate inspection frequency is determined based on the purpose of the inspection and the characteristics of the concrete. If the purpose of the inspection is to evaluate the overall safety of the dam, regular and comprehensive inspections are required; if the purpose of the inspection is to monitor performance changes in a specific area, more frequent inspections can be carried out for that area. Different types of concrete have different physical and chemical properties, and the requirements for inspection frequency are also different. High-performance concrete may require a lower testing frequency due to its excellent durability and crack resistance; ordinary concrete may require more frequent testing to monitor its performance changes. If the dam is operating at a high water level, more frequent testing is required to ensure the safety of the dam; if the dam is at a low water level or is out of service, the testing frequency can be appropriately reduced. Under extreme climatic conditions, the performance of concrete may change, so the testing frequency needs to be increased to monitor its changing trends. A testing schedule is formulated based on concrete of different ages. During the dam construction phase, the testing schedule is coordinated with the construction progress. During the dam operation phase, the testing schedule is determined based on the dam's operating status and testing needs. Testing before the rainy season can promptly detect and deal with potential water seepage problems; testing under low temperature conditions in winter can evaluate the frost resistance of concrete, etc.

[0016] Furthermore, the S2 data preprocessing is to preprocess the raw data collected by S1, specifically, to remove invalid data, including processing missing values, deleting duplicate values ​​and processing abnormal values, the processing of missing values ​​includes deletion, prediction and filling, the deletion is to delete the rows or columns containing missing values, specifically, if the missing proportion of missing values ​​in the data set is less than 10%, the mean square error and the mean absolute error are 0-5%, the correlation coefficient is between 0.85-0.99 or -0.85-0.99, the recall rate is 0.80-1.00, the F1 score is 0.87-0.97, and the statistical test method finds that the significance level p value is greater than 0.05, then the rows or columns containing missing values ​​are deleted.

[0017] Assume that in a dataset of concrete surface crack detection, the number of missing values ​​of a variable accounts for 3% of the total data, and the variable is not a key variable for crack detection. After deleting these missing values, the accuracy of the model is greater than 95%, the recall rate is between 0.80 and 1.00, and the F1 score is between 0.87 and 0.97, which has not changed significantly. At the same time, the statistical test method found that the significance level p value is greater than 0.05, indicating that there is no significant difference in the data distribution before and after processing the missing values. In this case, it can be considered that these missing values ​​have little effect on the analysis results, so the rows or columns containing missing values ​​can be directly deleted.

[0018] The prediction is to use machine learning algorithms such as linear regression and decision trees to predict missing values, and the filling is to use mean filling:

[0019] ;

[0020] Among them, X old is the original data, X new is the filled data, i and j represent the row and column indexes respectively;

[0021] The deduplication is to identify duplicate items in the data set and delete them to retain unique records, which can be achieved by deduplicating the data set;

[0022] The processing of outliers includes removal, pruning, replacement and conversion. Removal refers to directly deleting outliers, pruning refers to retaining only a specified percentage of data and discarding extreme values, replacement refers to replacing outliers with values ​​closer to other data points, and conversion refers to converting outliers to reduce their impact on the results.

[0023] Data standardization uses Z-score standardization to convert the original data into the Z value of its standard normal distribution, that is, all data values ​​are converted to standard deviations relative to the mean. The formula is:

[0024] ;

[0025] Where Xnew represents the standardized data, x represents the original data, μ represents the mean of the data set, and σ represents the standard deviation of the data set.

[0026] In the processing of outliers, the specified percentage is 90%, the value closer to other data points is the median or mean, and the outliers are converted to logarithmic transformation.

[0027] Furthermore, the S3 characteristic parameter analysis specifically shows that there is a positive correlation between wave velocity and elastic modulus:

[0028] ;

[0029] Among them, V is the wave velocity, S is the strength, E is the elastic modulus, a, b, c are unknown coefficients, which are determined by experimental data;

[0030] The speed of the wave can also reflect the density of the concrete. Concrete with high density has small internal porosity and tight structure, and the ultrasonic wave propagates quickly. Conversely, the propagation speed is slow:

[0031] ;

[0032] Among them, V is the wave velocity, k is a constant, and σ is the density index, which needs to be determined or calculated through experiments;

[0033] The decay of the amplitude is proportional to the cross-sectional area of ​​the defect or crack:

[0034] ;

[0035] Among them, A1 is the amplitude measured after passing through the defect or crack, A0 is the initial amplitude, β is the attenuation coefficient related to the characteristics of concrete and defects or cracks, S is the cross-sectional area of ​​the defect or crack, and V is the propagation volume of the ultrasonic wave in the concrete, which is regarded as a characteristic volume and used to normalize the cross-sectional area;

[0036] When concrete is subjected to temperature changes, thermal stress will be generated inside it. The thermal stress is proportional to the elastic modulus, linear expansion coefficient and temperature change of the concrete:

[0037] ;

[0038] Among them, σ represents thermal stress, E represents elastic modulus, α represents linear expansion coefficient of concrete, and ΔT represents temperature change.

[0039] Furthermore, the S4 multi-parameter digital quantity board model is established, and the specific steps are as follows:

[0040] S41 Data processing: normalize the wave velocity, frequency, amplitude, strength, elastic modulus, temperature and linear expansion coefficient data to the same scale, convert the data to values ​​between 0 and 1, and divide the data set into training set, validation set and test set. The training set is used to train the neural network, the validation set is used to adjust the network parameters, and the test set is used to evaluate the final performance of the model.

[0041] S42 Neural Network Design:

[0042] S421 input layer design: contains a corresponding number of nodes and uses different activation functions, such as softmax for classification and linear activation for multi-output regression;

[0043] S422 hidden layer design:

[0044] Convolutional layer: Use one-dimensional convolutional layer or special two-dimensional convolutional layer configuration, where the width of the convolution kernel is the same as the width of the input feature map, that is, each parameter is processed as a row, and the convolution layer is followed by an activation function such as ReLU to increase nonlinearity. The special two-dimensional convolutional layer includes deformable convolution, hole convolution or dilated convolution, where deformable convolution helps the model better capture crack features with irregular shapes or directions, and hole convolution or dilated convolution allows the convolution kernel to capture features in a larger spatial range, which is particularly useful for capturing long-distance dependencies or subtle features of cracks;

[0045] Multiple convolutional layer stacking: Multiple convolutional layers are stacked, and each convolutional layer is followed by an activation function;

[0046] S423 Flattening layer: At the end of the neural network, a flattening layer is used to convert the multi-dimensional output feature map into a one-dimensional array so that it can be connected to the fully connected layer;

[0047] S424 fully connected layer: After the flattening layer, one or more fully connected layers are added to learn the nonlinear combination of features and finally output the predicted value. A batch normalization layer is added after the fully connected layer to improve the training speed and stability.

[0048] S43 trains the neural network: selects an optimization algorithm such as gradient descent or Adam optimizer to adjust the network weights to minimize the loss function, sets the learning rate, which determines the step size of the weight adjustment, inputs the training data into the neural network, calculates the output value through forward propagation, and then calculates the loss function such as mean square error and cross entropy, and updates the network weights through the back propagation algorithm;

[0049] S44 Model Verification: Use the data sets that did not participate in the model training, namely the validation set and the test set, to verify the model and calculate the evaluation indicators: the proportion of correctly classified samples to the total number of samples, the accuracy rate,

[0050] ;

[0051] The ratio of the instances predicted as positive samples to the ones that are actually positive samples, and the precision is,

[0052] ;

[0053] The ratio of true positive examples to all positive examples, and the recall rate is,

[0054] ;

[0055] The harmonic mean of precision and recall, the F1 score, is,

[0056] ;

[0057] Among them, TP, TN, FP, and FN represent the number of samples of true positive examples, true negative examples, false positive examples, and false negative examples, respectively;

[0058] Furthermore, in S43, the neural network is trained, and the selected optimization algorithm is the gradient descent method or the Adam optimizer, and the loss function is the mean square error or the cross entropy.

[0059] Mean square error: The average of the squares of the differences between the predicted value and the true value, used to measure the size of the prediction error;

[0060] S45 model optimization: adjust and optimize the model according to the verification results. The specific steps are as follows;

[0061] S451 adjusts the network architecture: increases or decreases the number of hidden layers and nodes;

[0062] S452 adjusts hyperparameters: adjusts learning rate, batch size, and regularization term;

[0063] S453 uses regularization techniques: L1 regularization, L2 regularization, and Dropout between fully connected layers to prevent overfitting;

[0064] S454 Ensemble learning: Integrate multiple neural network models to improve overall prediction performance.

[0065] Furthermore, the S6 model is applied and fed back to monitor the concrete structure regularly. Specifically, if the dam uses high-performance concrete and is in normal operation and the environmental conditions are relatively stable, a comprehensive inspection is conducted every two years to evaluate the overall safety of the dam. If cracks or water seepage problems occur in a certain area of ​​the dam, the area needs to be inspected quarterly or monthly to monitor its performance changes and take corresponding repair measures. If the dam has been built and put into operation, a schedule for comprehensive inspections every two years can be established, avoiding the rainy season and extreme weather during each inspection cycle. Arrange to carry out inspections once in spring and once in autumn to ensure the comprehensiveness and accuracy of the inspections. At the same time, if abnormal situations occur during the operation of the dam or specific inspection tasks are required, such as monitoring the development of cracks and evaluating the repair effect, the inspection schedule can be adjusted according to the actual situation and corresponding inspection tasks can be added. If cracks have appeared in the middle of the dam body, the middle of the dam body and its surrounding areas should be taken as key inspection locations when formulating the inspection plan. At the same time, if the spillway of the dam uses special impact-resistant and wear-resistant concrete, detailed inspections should also be carried out in the spillway area to evaluate the durability and performance of the concrete.

[0066] Another aspect of the present invention provides a system used in a method for detecting concrete surface cracks based on a machine learning algorithm, comprising the following modules:

[0067] Data acquisition module: responsible for receiving raw data from ultrasonic detectors, impact elastic wave testers, radar wave detectors, high-definition cameras, and temperature sensor detection equipment;

[0068] Data preprocessing module: responsible for cleaning, denoising and standardizing the collected raw data to improve the quality and reliability of the data;

[0069] Feature extraction module: responsible for using feature extraction algorithms to convert raw data into useful information for crack detection;

[0070] Crack measurement module: responsible for identifying cracks on the concrete surface and measuring the width and length parameters of the identified cracks;

[0071] Multi-parameter digital plate model building module: responsible for building a digital plate model based on wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient;

[0072] Multi-parameter comprehensive analysis module: Use multi-parameter digital plate model to comprehensively evaluate the quality status and crack risk of concrete;

[0073] Real-time monitoring and early warning module: input relevant parameters of the concrete to be tested into the multi-parameter digital plate model, and calculate the potential risk and prediction value of cracks;

[0074] Database management module: responsible for managing all data of the device, including original data, processing results, and crack information;

[0075] User interface and interaction module: Provides a user interface to facilitate user operations and data viewing.

[0076] Furthermore, the data acquisition module serves as the input end of the device, the data preprocessing module receives the original data transmitted by the data acquisition module, the data processed by the data preprocessing module is transmitted to the feature extraction module, the feature extraction module transmits the extracted feature parameters and image features to the crack measurement module, the multi-parameter digital quantity board model establishment module and the real-time monitoring and early warning module, the crack measurement module transmits the measurement data of the crack and the location and shape information of the crack to the multi-parameter comprehensive analysis module, the multi-parameter digital quantity board model establishment module transmits the established digital quantity board model to the multi-parameter comprehensive analysis module, and the multi-parameter comprehensive analysis module transmits the comprehensive evaluation results to the user interface and interaction module.

[0077] Furthermore, the multi-parameter comprehensive analysis module includes a multi-source data fusion unit, a model evaluation unit and a result transmission unit. The multi-source data fusion unit is responsible for fusing the data obtained by different sensors and detection means to improve the comprehensiveness of the analysis. The model evaluation unit is responsible for using the multi-parameter digital quantity plate model to comprehensively evaluate the quality status and crack risk of concrete. The result display unit transmits the evaluation results to the user interface and interaction module. The real-time monitoring and early warning module includes a real-time data collection unit, an abnormality detection unit and an early warning notification unit. The real-time data collection unit continuously collects real-time feature parameters and image features extracted by the feature extraction module. The abnormality detection unit detects abnormal changes in the data according to preset thresholds or rules. Once abnormal changes are found, the early warning notification unit immediately notifies the staff by sound, light signals, text messages, or emails.

[0078] The beneficial effects of the present invention are:

[0079] 1. The present invention establishes a multi-parameter digital quantity plate model based on the physical parameters of concrete strength, elastic modulus, temperature, linear expansion coefficient, and the correlation between the wave velocity, frequency, and amplitude characteristic parameters of its impact elastic wave, ultrasonic wave, and radar wave. While quickly and accurately measuring concrete defects, the present invention predicts cracks and defects inside the concrete, discovers potential problems in advance, and provides a basis for maintenance and reinforcement.

[0080] 2. The present invention is based on the advantages of ultrasonic detection method, which has strong penetration, no damage to the structure, and good directionality. It combines the impact elastic wave and radar wave technology. The impact elastic wave has the advantages of high energy and large test range, and the radar wave has higher resolution and deeper detection capabilities, which expands the detection range. At the same time, it obtains characteristic parameters of more dimensions, thereby more comprehensively reflecting the quality status of concrete.

[0081] 3. The detection process of the present invention is highly automated, reducing the possibility of manual intervention and human error. The model can process large amounts of data, quickly generate detection reports and analysis results, and improve work efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 is a flow chart of a method for detecting cracks on a concrete surface based on a machine learning algorithm according to Embodiment 1 of the present invention;

[0083] Figure 2 This is a block diagram of a concrete surface crack detection system based on a machine learning algorithm according to Embodiment 2 of the present invention;

[0084] Figure 3 This is a device block diagram of a multi-parameter comprehensive analysis module and a real-time monitoring and early warning module of a concrete surface crack detection system based on a machine learning algorithm in Example 2 of the present invention. DETAILED DESCRIPTION

[0085] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The exemplary embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0086] Example 1

[0087] See also Figure 1 As shown, the present invention provides a method for detecting concrete surface cracks based on a machine learning algorithm, comprising the following steps:

[0088] S1 Data Collection:

[0089] Taking the Aoshuihe Reservoir in Qianxi County as an example, the wave velocity, frequency, and amplitude characteristic parameters of concrete of different ages and different magnesium oxide content were obtained through the comprehensive detection method of shock wave, ultrasonic wave, radar wave and indoor mechanical test, and the mixed soil strength, elastic modulus, temperature, and linear expansion coefficient were obtained; the RSM-SY6 (C) non-metallic ultrasonic detector was used to measure the ultrasonic propagation velocity in concrete, and the SCE-MATS-S impact echo meter was used to analyze the dynamic response characteristics of concrete, including wave velocity, frequency, amplitude, strength, and elastic modulus. The GSSI SIR-3000 radar wave detector was selected to identify tiny cracks and voids in concrete. The RSM-SY6 (C) ultrasonic detector, SCE-MATS-S impact elastic wave tester, GSSI SIR-3000 radar wave detector, temperature sensor, and D850 high-definition camera were arranged on the concrete surface to be tested to ensure that the equipment can cover the area to be tested and collect original data;

[0090] The collection scheme of this embodiment is as follows:

[0091] Scheme 1: The concrete age is 3 days, the magnesium oxide content is 3%, the concrete design strength is C25, and 2 groups are measured.

[0092] Scheme 2: The concrete age is 7 days, the magnesium oxide content is 3%, the concrete design strength is C25, and 2 groups are measured.

[0093] Scheme 3: The concrete age is 28 days, the magnesium oxide content is 3%, the concrete design strength is C25, and 2 groups are measured.

[0094] Scheme 4: The concrete age is 90 days, the magnesium oxide content is 3%, the concrete design strength is C25, and 2 groups are measured.

[0095] See Table 1 for data.

[0096] Table 1 Data collection of Example 1

[0097] ;

[0098] S2 Data preprocessing: perform data cleaning, remove invalid data, standardize wave velocity, frequency, amplitude, concrete strength, elastic modulus, temperature, and linear expansion coefficient, and eliminate the influence of different testing equipment and testing conditions on the data; preprocess the original data collected by S1, specifically, remove invalid data, including processing missing values, deleting duplicate values, and processing outliers. The processing of missing values ​​includes deletion, prediction, and filling. The deletion is that if the missing ratio of missing values ​​in the data set is less than 10%, the mean square error and mean absolute error are 0-5%, the correlation coefficient is between 0.85-0.99 or -0.85-0.99, the recall rate is 0.80-1.00, and the F1 score is 0.87-0.97. The statistical test method finds that the significance level p value is greater than 0.05, then delete the rows or columns containing missing values. The prediction is to use machine learning algorithms such as linear regression and decision trees to predict missing values, and the filling is to use mean filling:

[0099] ;

[0100] Among them, X old is the original data, X new is the filled data, i and j represent the row and column indexes respectively;

[0101] The deduplication is to identify duplicates in the data set and delete them to retain unique records, which can be achieved by deduplicating the data set, such as using the drop_duplicates() function in the Pandas library;

[0102] The processing of outliers includes removal, pruning, replacement and conversion. Removal refers to directly deleting outliers, pruning refers to retaining only a specified percentage of data and discarding extreme values, replacement refers to replacing outliers with values ​​closer to other data points, and conversion refers to converting outliers to reduce their impact on the results.

[0103] Data standardization uses Z-score standardization to convert the original data into the Z value of its standard normal distribution, that is, all data values ​​are converted to standard deviations relative to the mean. The formula is:

[0104] ;

[0105] Where X new represents the standardized data, x represents the original data, μ represents the mean of the data set, and σ represents the standard deviation of the data set.

[0106] In the processing of outliers, the specified percentage is 90%, the value closer to other data points is the median or mean, and the outliers are converted to logarithmic transformation;

[0107] S3 characteristic parameter analysis: Analyze the relationship between wave velocity and concrete strength and elastic modulus, evaluate the density and integrity of concrete, analyze the changing rules of frequency and amplitude, identify defects and cracks inside concrete, record temperature data during detection, and analyze the impact of temperature on concrete performance in combination with the linear expansion coefficient of concrete; specifically, there is a positive correlation between wave velocity and elastic modulus:

[0108] ;

[0109] Among them, V is the wave velocity, S is the strength, E is the elastic modulus, a, b, c are unknown coefficients, which are determined by experimental data;

[0110] The speed of the wave can also reflect the density of the concrete. Concrete with high density has small internal porosity and tight structure, and the ultrasonic wave propagates quickly. Conversely, the propagation speed is slow:

[0111] ;

[0112] Among them, V is the wave velocity, k is a constant, and σ is the density index, which needs to be determined or calculated through experiments;

[0113] The decay of the amplitude is proportional to the cross-sectional area of ​​the defect or crack:

[0114] ;

[0115] Among them, A1 is the amplitude measured after passing through the defect or crack, A0 is the initial amplitude, β is the attenuation coefficient related to the characteristics of concrete and defects or cracks, S is the cross-sectional area of ​​the defect or crack, and V is the propagation volume of the ultrasonic wave in the concrete, which is regarded as a characteristic volume and used to normalize the cross-sectional area. When the concrete is subjected to temperature changes, thermal stress will be generated inside it, and the thermal stress is proportional to the elastic modulus, linear expansion coefficient and temperature change of the concrete:

[0116] ;

[0117] Among them, σ represents thermal stress, E represents elastic modulus, α represents linear expansion coefficient of concrete, and ΔT represents temperature change;

[0118] S4 multi-parameter digital plate model establishment: Based on the wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient collected from the Aoshuihe Reservoir in Qianxi County, their mathematical relationship models were established respectively, and statistical analysis and machine learning algorithms were used to fit the collected data to determine the model parameters;

[0119] The specific steps are:

[0120] S41 Data processing: normalize the wave velocity, frequency, amplitude, strength, elastic modulus, temperature and linear expansion coefficient data to the same scale, convert the data to values ​​between 0 and 1, and divide the data set into training set, validation set and test set. The training set is used to train the neural network, the validation set is used to adjust the network parameters, and the test set is used to evaluate the final performance of the model.

[0121] The normalized processing results of this embodiment are shown in Table 2.

[0122] Table 2 Data normalization results of Example 1

[0123] ;

[0124] S42 Neural Network Design:

[0125] S421 Input layer design: contains a corresponding number of nodes and uses different activation functions;

[0126] S422 hidden layer design:

[0127] Convolutional layers: Use one-dimensional convolutional layers or special two-dimensional convolutional layer configurations where the width of the convolution kernel is the same as the width of the input feature map and the convolutional layer is followed by an activation function to add nonlinearity. The special two-dimensional convolutional layers include deformable convolution, dilated convolution, or expanded convolution.

[0128] Multiple convolutional layer stacking: Multiple convolutional layers are stacked, and each convolutional layer is followed by an activation function;

[0129] S423 Flattening layer: At the end of the neural network, a flattening layer is used to convert the multi-dimensional output feature map into a one-dimensional array so that it can be connected to the fully connected layer;

[0130] S424 fully connected layer: After the flattening layer, one or more fully connected layers are added to learn the nonlinear combination of features and finally output the predicted value. A batch normalization layer is added after the fully connected layer to improve the training speed and stability.

[0131] S43 trains the neural network: selects an optimization algorithm to adjust the network weights to minimize the loss function, sets the learning rate, which determines the step size of the weight adjustment, inputs the training data into the neural network, calculates the output value through forward propagation, then calculates the loss function, and updates the network weights through the back propagation algorithm;

[0132] S44 Model Validation: Use the data sets that did not participate in the model training, namely the validation set and the test set, to validate the model.

[0133] Calculate the evaluation indicators: the proportion of correctly classified samples to the total number of samples, the accuracy is,

[0134] ;

[0135] The ratio of the instances predicted as positive samples to the ones that are actually positive samples, and the precision is,

[0136] ;

[0137] The ratio of true positive examples to all positive examples, and the recall rate is,

[0138] ;

[0139] The harmonic mean of precision and recall, the F1 score is

[0140] ;

[0141] Among them, TP, TN, FP, and FN represent the number of samples of true positive examples, true negative examples, false positive examples, and false negative examples, respectively;

[0142] Mean square error: The average of the squares of the differences between the predicted value and the true value, used to measure the size of the prediction error;

[0143] S45 model optimization: adjust and optimize the model according to the verification results. The specific steps are as follows:

[0144] S451 adjusts the network architecture: increases or decreases the number of hidden layers and nodes;

[0145] S452 adjusts hyperparameters: adjusts learning rate, batch size, and regularization term;

[0146] S453 uses regularization techniques: L1 regularization, L2 regularization, and Dropout between fully connected layers to prevent overfitting;

[0147] S454 Ensemble learning: Integrate multiple neural network models to improve overall prediction performance;

[0148] S5 Crack prediction and assessment: The relevant data parameters of the concrete to be tested collected from Aoshuihe Reservoir in Qianxi County and Yousha Reservoir in Wangmo County are input into the multi-parameter digital plate model. The model calculates the potential risk and prediction value of the cracks based on the input parameters;

[0149] S6 Model Application and Feedback: The multi-parameter digital plate model is applied to the real-time detection of hydraulic concrete structures in the Aoshuihe Reservoir in Qianxi County to achieve quality assessment and crack prediction, regularly monitor the concrete structure, update the model data, and improve the accuracy and adaptability of the model. Practice has proved that the model has a good application effect in the Aoshuihe Reservoir in Qianxi County.

[0150] The crack prediction results of this embodiment are compared in Table 3.

[0151] Table 3 Comparison of crack prediction calculation results in Example 1

[0152] ;

[0153] Regular monitoring of concrete structures means that if the dam uses high-performance concrete and is in normal operation and the environmental conditions are relatively stable, a comprehensive inspection should be carried out every two years to evaluate the overall safety of the dam. If cracks or water seepage problems occur in a certain area of ​​the dam, the area needs to be inspected once a quarter or once a month to monitor its performance changes and take corresponding repair measures. If the dam has been built and put into operation, a schedule for comprehensive inspections every two years can be formulated. During each inspection cycle, an inspection should be arranged in spring and autumn to avoid the rainy season and extreme weather to ensure the comprehensiveness and accuracy of the inspection. At the same time, if abnormal conditions occur during the operation of the dam or specific inspection tasks are required, such as monitoring the development of cracks and evaluating the repair effect, the inspection schedule can be adjusted according to the actual situation and the corresponding inspection tasks can be added. If cracks have appeared in the middle of the dam body, the middle of the dam body and its surrounding areas should be taken as key inspection locations when formulating the inspection plan. At the same time, if the spillway of the dam uses special impact-resistant and wear-resistant concrete, detailed inspections should be carried out in the spillway area to evaluate the durability and performance of the concrete.

[0154] Example 2

[0155] A method for detecting concrete surface cracks based on a machine learning algorithm comprises the following steps:

[0156] S1 Data Collection:

[0157] Taking Yousha Reservoir in Wangmo County as an example, the wave velocity, frequency, amplitude, concrete strength, elastic modulus, temperature, and linear expansion coefficient characteristic parameters of concrete of different ages and different mix ratios were obtained through the comprehensive detection method of shock wave, ultrasonic wave, radar wave, and indoor mechanical test; RSM-SY6 (C) non-metallic ultrasonic detector was used to measure the ultrasonic wave propagation velocity in concrete, and SCE-MATS-S impact echo meter was used to analyze the dynamic response characteristics of concrete, including wave velocity, frequency, amplitude, strength, and elastic modulus. GSSI SIR-3000 radar wave detector was selected to identify tiny cracks and voids in concrete. RSM-SY6 (C) ultrasonic detector, SCE-MATS-S impact elastic wave tester, GSSI SIR-3000 radar wave detector, temperature sensor, and D850 high-definition camera were arranged on the concrete surface to be tested to ensure that the equipment can cover the area to be tested and collect original data;

[0158] The collection scheme of this embodiment is as follows:

[0159] Scheme 1: Concrete age is 3d, concrete design strength is C25, and 2 groups are measured.

[0160] Scheme 2: Concrete age is 7d, concrete design strength is C25, and 2 groups are measured.

[0161] Scheme 3: The concrete age is 28 days, the concrete design strength is C25, and 2 groups are measured.

[0162] Scheme 4: Concrete age is 90d, concrete design strength is C25, and 2 groups are measured.

[0163] The data of Example 2 are shown in Table 4.

[0164] Table 4 Data collection of Example 2

[0165] ;

[0166] S2 Data preprocessing: perform data cleaning, remove invalid data, standardize wave velocity, frequency, amplitude, concrete strength, elastic modulus, temperature, and linear expansion coefficient, and eliminate the influence of different testing equipment and testing conditions on the data; preprocess the original data collected by S1, specifically, remove invalid data, including processing missing values, deleting duplicate values, and processing outliers. The processing of missing values ​​includes deletion, prediction, and filling. The deletion is that if the missing ratio of missing values ​​in the data set is less than 10%, the mean square error and mean absolute error are 0-5%, the correlation coefficient is between 0.85-0.99 or -0.85-0.99, the recall rate is 0.80-1.00, the F1 score is 0.87-0.97, and the statistical test method finds that the significance level p value is greater than 0.05, then delete the rows or columns containing missing values. The prediction is to use machine learning algorithms such as linear regression and decision tree to predict missing values, and the filling is to use mean filling:

[0167] ;

[0168] Among them, X old is the original data, X new is the filled data, i and j represent the row and column indexes respectively;

[0169] The deduplication is to identify duplicates in the data set and delete them to retain unique records, which can be achieved by deduplicating the data set, such as using the drop_duplicates() function in the Pandas library;

[0170] The processing of outliers includes removal, pruning, replacement and conversion. Removal refers to directly deleting outliers, pruning refers to retaining only a specified percentage of data and discarding extreme values, replacement refers to replacing outliers with values ​​closer to other data points, and conversion refers to converting outliers to reduce their impact on the results.

[0171] Data standardization uses Z-score standardization to convert the original data into the Z value of its standard normal distribution, that is, all data values ​​are converted to standard deviations relative to the mean. The formula is:

[0172] ;

[0173] Where X new represents the standardized data, x represents the original data, μ represents the mean of the data set, and σ represents the standard deviation of the data set.

[0174] In the processing of outliers, the specified percentage is 90%, the value closer to other data points is the median or mean, and the outliers are converted to logarithmic transformation;

[0175] S3 characteristic parameter analysis: analyze the relationship between wave velocity and concrete strength and elastic modulus, evaluate the density and integrity of concrete, analyze the changing rules of frequency and amplitude, identify defects and cracks inside concrete, record temperature data during detection, and analyze the impact of temperature on concrete performance in combination with the linear expansion coefficient of concrete; specifically,

[0176] There is a positive correlation between wave velocity and elastic modulus:

[0177] ;

[0178] Among them, V is the wave velocity, S is the strength, E is the elastic modulus, a, b, c are unknown coefficients, which are determined by experimental data;

[0179] The speed of the wave can also reflect the density of the concrete. Concrete with high density has small internal porosity and tight structure, and the ultrasonic wave propagates quickly. Conversely, the propagation speed is slow:

[0180] ;

[0181] Among them, V is the wave velocity, k is a constant, and σ is the density index, which needs to be determined or calculated through experiments;

[0182] The decay of the amplitude is proportional to the cross-sectional area of ​​the defect or crack:

[0183] ;

[0184] Among them, A1 is the amplitude measured after passing through the defect or crack, A0 is the initial amplitude, β is the attenuation coefficient related to the characteristics of concrete and defects or cracks, S is the cross-sectional area of ​​the defect or crack, and V is the propagation volume of the ultrasonic wave in the concrete, which is regarded as a characteristic volume and used to normalize the cross-sectional area. When the concrete is subjected to temperature changes, thermal stress will be generated inside it, and the thermal stress is proportional to the elastic modulus, linear expansion coefficient and temperature change of the concrete:

[0185] ;

[0186] Among them, σ represents thermal stress, E represents elastic modulus, α represents linear expansion coefficient of concrete, and ΔT represents temperature change;

[0187] S4 multi-parameter digital plate model establishment: Based on the wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient collected from the Aoshuihe Reservoir in Qianxi County, their mathematical relationship models were established respectively, and statistical analysis and machine learning algorithms were used to fit the collected data to determine the model parameters;

[0188] The specific steps are:

[0189] S41 Data processing: normalize the wave velocity, frequency, amplitude, strength, elastic modulus, temperature and linear expansion coefficient data to the same scale, convert the data to values ​​between 0 and 1, and divide the data set into training set, validation set and test set. The training set is used to train the neural network, the validation set is used to adjust the network parameters, and the test set is used to evaluate the final performance of the model.

[0190] The normalized processing results of this embodiment are shown in Table 5.

[0191] Table 5 Data normalization results of Example 2

[0192] ;

[0193] S42 Neural Network Design:

[0194] S421 Input layer design: contains a corresponding number of nodes and uses different activation functions;

[0195] S422 hidden layer design:

[0196] Convolutional layers: Use one-dimensional convolutional layers or special two-dimensional convolutional layer configurations where the width of the convolution kernel is the same as the width of the input feature map and the convolutional layer is followed by an activation function to add nonlinearity. The special two-dimensional convolutional layers include deformable convolution, dilated convolution, or expanded convolution.

[0197] Multiple convolutional layer stacking: Multiple convolutional layers are stacked, and each convolutional layer is followed by an activation function;

[0198] S423 Flattening layer: At the end of the neural network, a flattening layer is used to convert the multi-dimensional output feature map into a one-dimensional array so that it can be connected to the fully connected layer;

[0199] S424 fully connected layer: After the flattening layer, one or more fully connected layers are added to learn the nonlinear combination of features and finally output the predicted value. A batch normalization layer is added after the fully connected layer to improve the training speed and stability.

[0200] S43 trains the neural network: selects an optimization algorithm to adjust the network weights to minimize the loss function, sets the learning rate, which determines the step size of the weight adjustment, inputs the training data into the neural network, calculates the output value through forward propagation, then calculates the loss function, and updates the network weights through the back propagation algorithm;

[0201] S44 Model Validation: Use the data sets that did not participate in the model training, namely the validation set and the test set, to validate the model.

[0202] Calculate the evaluation indicators: the proportion of correctly classified samples to the total number of samples, the accuracy is,

[0203] ;

[0204] The ratio of the instances predicted as positive samples to the ones that are actually positive samples, and the precision is,

[0205] ;

[0206] The ratio of true positive examples to all positive examples, and the recall rate is,

[0207] ;

[0208] The harmonic mean of precision and recall, the F1 score is

[0209] ;

[0210] Among them, TP, TN, FP, and FN represent the number of samples of true positive examples, true negative examples, false positive examples, and false negative examples, respectively;

[0211] Mean square error: The average of the squares of the differences between the predicted value and the true value, used to measure the size of the prediction error;

[0212] S45 model optimization: adjust and optimize the model according to the verification results. The specific steps are as follows:

[0213] S451 adjusts the network architecture: increases or decreases the number of hidden layers and nodes;

[0214] S452 adjusts hyperparameters: adjusts learning rate, batch size, and regularization term;

[0215] S453 uses regularization techniques: L1 regularization, L2 regularization, and Dropout between fully connected layers to prevent overfitting;

[0216] S454 Ensemble learning: Integrate multiple neural network models to improve overall prediction performance;

[0217] S5 Crack prediction and assessment: The relevant data parameters of the concrete to be tested collected from Aoshuihe Reservoir in Qianxi County and Yousha Reservoir in Wangmo County are input into the multi-parameter digital plate model. The model calculates the potential risk and prediction value of the cracks based on the input parameters;

[0218] S6 Model Application and Feedback: The multi-parameter digital plate model is applied to the real-time detection of hydraulic concrete structures in the Aoshuihe Reservoir in Qianxi County to achieve quality assessment and crack prediction, regularly monitor the concrete structure, update the model data, and improve the accuracy and adaptability of the model. Practice has proved that the model has a good application effect in the Aoshuihe Reservoir in Qianxi County.

[0219] The crack prediction results of this embodiment are compared in Table 6.

[0220] Table 6 Comparison of crack prediction calculation results in Example 2

[0221] ;

[0222] Regular monitoring of concrete structures means that if the dam uses high-performance concrete and is in normal operation and the environmental conditions are relatively stable, a comprehensive inspection should be carried out every two years to evaluate the overall safety of the dam. If cracks or water seepage problems occur in a certain area of ​​the dam, the area needs to be inspected once a quarter or once a month to monitor its performance changes and take corresponding repair measures. If the dam has been built and put into operation, a schedule for comprehensive inspections every two years can be formulated. During each inspection cycle, avoid the rainy season and extreme weather and arrange for inspections in spring and autumn to ensure the comprehensiveness and accuracy of the inspections. At the same time, if abnormal conditions occur during the operation of the dam or specific inspection tasks are required, such as monitoring the development of cracks and evaluating the repair effect, the inspection schedule can be adjusted according to the actual situation and the corresponding inspection tasks can be added. If cracks have appeared in the middle of the dam body, the middle of the dam body and its surrounding areas should be taken as key inspection locations when formulating the inspection plan. At the same time, if the spillway of the dam uses special impact-resistant and wear-resistant concrete, detailed inspections should also be carried out in the spillway area to evaluate the durability and performance of the concrete.

[0223] Example 3

[0224] See also Figure 2-Figure 3As shown, a concrete surface crack detection system based on machine learning algorithm includes the following modules:

[0225] Data acquisition module: responsible for receiving raw data from ultrasonic detectors, impact elastic wave testers, radar wave detectors, high-definition cameras, and temperature sensor detection equipment;

[0226] Data preprocessing module: responsible for cleaning, denoising and standardizing the collected raw data to improve the quality and reliability of the data;

[0227] Feature extraction module: responsible for using feature extraction algorithms to convert raw data into useful information for crack detection;

[0228] Crack measurement module: responsible for identifying cracks on the concrete surface and measuring the width and length parameters of the identified cracks;

[0229] Multi-parameter digital plate model building module: responsible for building a digital plate model based on wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient;

[0230] Multi-parameter comprehensive analysis module: Use multi-parameter digital plate model to comprehensively evaluate the quality status and crack risk of concrete;

[0231] Real-time monitoring and early warning module: input relevant parameters of the concrete to be tested into the multi-parameter digital plate model, and calculate the potential risk and prediction value of cracks;

[0232] Database management module: responsible for managing all data of the device, including original data, processing results, and crack information;

[0233] User interface and interaction module: Provides a user interface to facilitate user operations and data viewing;

[0234] The data acquisition module serves as the input end of the device, the data preprocessing module receives the original data transmitted by the data acquisition module, the data processed by the data preprocessing module is transmitted to the feature extraction module, the feature extraction module transmits the extracted feature parameters and image features to the crack measurement module, the multi-parameter digital quantity board model establishment module and the real-time monitoring and early warning module, the crack measurement module transmits the measurement data of the crack and the position and shape information of the crack to the multi-parameter comprehensive analysis module, the multi-parameter digital quantity board model establishment module transmits the established digital quantity board model to the multi-parameter comprehensive analysis module, and the multi-parameter comprehensive analysis module transmits the comprehensive evaluation result to the user interface and interaction module;

[0235] The multi-parameter comprehensive analysis module includes a multi-source data fusion unit, a model evaluation unit and a result transmission unit. The multi-source data fusion unit is responsible for fusing the data obtained by different sensors and detection means to improve the comprehensiveness of the analysis. The model evaluation unit is responsible for using the multi-parameter digital quantity plate model to comprehensively evaluate the quality status and crack risk of concrete. The result display unit transmits the evaluation results to the user interface and interaction module. The real-time monitoring and early warning module includes a real-time data collection unit, an abnormality detection unit and an early warning notification unit. The real-time data collection unit continuously collects real-time feature parameters and image features extracted by the feature extraction module. The abnormality detection unit detects abnormal changes in the data according to preset thresholds or rules. Once an abnormal change is found, the early warning notification unit immediately notifies the staff through sound, light signals, text messages, emails, etc.

[0236] The working principle is: prepare ultrasonic detectors, impact elastic wave testers, radar wave detectors, high-definition cameras, and temperature sensor detection equipment, and arrange the detection equipment on the concrete surface to be detected to ensure that the equipment can cover the area that needs to be detected. According to the actual situation, set multiple detection points or scanning paths; ultrasonic detectors and radar wave detectors will emit sound waves or electromagnetic waves into the concrete and receive reflected signals. These signals contain information about the internal structure and defects of the concrete. The data preprocessing module cleans, denoises, and standardizes the collected raw data. The feature extraction module extracts physical characteristic parameters such as wave velocity, frequency, and amplitude from the preprocessed data, and extracts the image features of the cracks, such as edges, shapes, size, etc.; the crack measurement module uses image processing technology and machine learning algorithms to analyze and identify the extracted image features, identify cracks on the concrete surface, and determine their position, shape and approximate size, and accurately measure the identified cracks, including measurement of parameters such as width and length; according to the distribution law of the relationship between concrete wave velocity, frequency, amplitude and mixed soil strength, elastic modulus, temperature, linear expansion coefficient, etc., a multi-parameter digital plate model for concrete defect detection crack prediction is established, combining multiple parameters such as wave velocity, frequency, amplitude, strength, elastic modulus, temperature, linear expansion coefficient, etc., as well as the measurement results of cracks, and using the multi-parameter digital plate model or other evaluation methods to conduct a comprehensive assessment of the quality status and crack risk of concrete.

[0237] Although the present invention has been described in detail above by general explanation and specific implementation methods, it is obvious to those skilled in the art that some modifications or improvements can be made on the basis of the present invention. Therefore, these modifications or improvements made on the basis of not departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for detecting concrete surface cracks based on a machine learning algorithm, characterized in that: The following steps are involved: S1 data acquisition: obtain the concrete wave velocity, frequency, amplitude characteristic parameters, and obtain the mixed soil strength, elastic modulus, temperature, and linear expansion coefficient; S2 Data preprocessing: perform data cleaning, remove invalid data, and perform Z-score standardization on wave velocity, frequency, amplitude, mixed soil strength, elastic modulus, temperature, and linear expansion coefficient; S3 characteristic parameter analysis: Analyze the relationship between wave velocity and concrete strength and elastic modulus, evaluate the density and integrity of concrete, analyze the changing rules of frequency and amplitude, identify defects and cracks inside concrete, use temperature sensors to record temperature data during detection, and analyze the impact of temperature on concrete performance in combination with the linear expansion coefficient of concrete. Specifically, There is a positive correlation between wave velocity and elastic modulus: ； Among them, V is the wave velocity, S is the strength, E is the elastic modulus, a, b, c are unknown coefficients, which are determined by experimental data; ; Among them, V is the wave velocity, k is a constant, and σ is the density index, which needs to be determined or calculated through experiments; The decay of the amplitude is proportional to the cross-sectional area of ​​the defect or crack: ； Among them, A1 is the amplitude measured after passing through the defect or crack, A0 is the initial amplitude, β is the attenuation coefficient related to the characteristics of concrete and defects or cracks, S is the cross-sectional area of ​​the defect or crack, and V is the propagation volume of the ultrasonic wave in the concrete, which is regarded as a characteristic volume and used to normalize the cross-sectional area; When concrete is subjected to temperature changes, thermal stress will be generated inside it. The thermal stress is proportional to the elastic modulus, linear expansion coefficient and temperature change of the concrete: ； Among them, σ represents thermal stress, E represents elastic modulus, α represents linear expansion coefficient of concrete, and ΔT represents temperature change; S4 multi-parameter digital panel model establishment: Based on wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient, a mathematical relationship model is established between them. Statistical analysis and machine learning algorithms are used to fit the collected data and determine the model parameters. The specific steps are as follows: S41 data processing: normalizing input data to the same scale, the input data including wave velocity, frequency, amplitude, strength, elastic modulus, temperature and linear expansion coefficient data, converting the data into values ​​between 0 and 1, dividing the data set into a training set, a validation set and a test set, the training set is used to train the neural network, the validation set is used to adjust the network parameters, and the test set is used to evaluate the final performance of the model; S42 Neural Network Design: S421 Input layer design: contains a corresponding number of nodes and uses different activation functions; S422 hidden layer design: Convolutional layer: Use one-dimensional convolutional layer or special two-dimensional convolutional layer configuration, where the width of the convolution kernel is the same as the width of the input feature map, and the convolution layer is followed by an activation function to increase nonlinearity. The special two-dimensional convolutional layer includes deformable convolution, hole convolution or dilated convolution, where deformable convolution helps the model better capture crack features with irregular shapes or directions, and hole convolution or dilated convolution allows the convolution kernel to capture features in a larger spatial range, which is particularly useful for capturing long-distance dependencies or subtle features of cracks; Multiple convolutional layer stacking: Multiple convolutional layers are stacked, and each convolutional layer is followed by an activation function; S423 Flattening layer: At the end of the neural network, a flattening layer is used to convert the multi-dimensional output feature map into a one-dimensional array and connect it to the fully connected layer; S424 fully connected layer: After the flattening layer, one or more fully connected layers are added to learn the nonlinear combination of features and finally output the predicted value. A batch normalization layer is added after the fully connected layer to improve the training speed and stability. S43 trains the neural network: selects an optimization algorithm to adjust the network weights to minimize the loss function, sets the learning rate, which determines the step size of the weight adjustment, inputs the training data into the neural network, calculates the output value through forward propagation, then calculates the loss function, and updates the network weights through the back propagation algorithm; S44 Model Validation: Use the data sets that did not participate in the model training, namely the validation set and the test set, to validate the model. Calculate the evaluation indicators: the proportion of correctly classified samples to the total number of samples, the accuracy is, ； The ratio of the instances predicted as positive samples to the ones that are actually positive samples, and the precision is, ； The ratio of true positive examples to all positive examples, and the recall rate is, ； The harmonic mean of precision and recall, the F1 score, is, ； Among them, TP, TN, FP, and FN represent the number of samples of true positive examples, true negative examples, false positive examples, and false negative examples, respectively; S45 model optimization: adjust and optimize the model according to the verification results. The specific steps are as follows: S451 adjusts the network architecture: increases or decreases the number of hidden layers and nodes, S452 adjusts hyperparameters: adjusts learning rate, batch size, regularization term, S453 uses regularization techniques: L1 regularization, L2 regularization, and Dropout between fully connected layers to prevent overfitting. S454 Ensemble learning: Integrate multiple neural network models to improve overall prediction performance; S5 Crack prediction and evaluation: The collected relevant data parameters of the concrete to be tested are input into the multi-parameter digital plate model. The model calculates the potential risk and prediction value of the crack according to the input parameters. The prediction value is the physical size of the corresponding crack output by the model, including the width, length and depth of the crack; S6 model application and feedback: Apply the multi-parameter digital plate model to the real-time detection of hydraulic concrete structures to achieve quality assessment and crack prediction, regularly monitor the concrete structure, and update the model data.

2. The method for detecting concrete surface cracks based on a machine learning algorithm according to claim 1, characterized in that: The S1 data acquisition uses an ultrasonic detector to measure the ultrasonic wave propagation speed in concrete, and uses a tester that can generate and detect impact elastic waves to analyze the dynamic response characteristics of concrete, including wave velocity, frequency, amplitude, strength and elastic modulus. A radar wave detector is selected to identify tiny cracks and voids in concrete, and a temperature sensor is used to measure the temperature of the concrete surface. The linear expansion coefficient is determined experimentally or obtained by consulting literature. An ultrasonic detector, an impact elastic wave tester, a radar wave detector, a high-definition camera, and a temperature sensor are arranged on the concrete surface to be tested to collect raw data.

3. The method for detecting concrete surface cracks based on a machine learning algorithm according to claim 1, characterized in that: The S2 data preprocessing is to preprocess the raw data collected by S1, specifically, to remove invalid data, including processing missing values, deleting duplicate values ​​and processing abnormal values; The processing of missing values ​​includes deletion, prediction and filling. The deletion is to delete the rows or columns containing missing values, the prediction is to use a machine learning algorithm to predict the missing values, and the filling is to use the mean value to fill: ； Among them, X old is the original data, X new is the filled data, i and j represent the row and column indexes respectively; The deduplication is to identify duplicate items in the data set and delete them to retain unique records, which is achieved by deduplicating the data set; The processing of outliers includes removal, pruning, replacement and conversion. Removal refers to directly deleting outliers, pruning refers to retaining only a specified percentage of data and discarding extreme values, replacement refers to replacing outliers with values ​​closer to other data points, and conversion refers to converting outliers to reduce their impact on the results. Data standardization uses Z-score standardization to convert the original data into the Z value of its standard normal distribution, that is, all data values ​​are converted to standard deviations relative to the mean. The formula is: ； Where X new represents the standardized data, x represents the original data, μ represents the mean of the data set, and σ represents the standard deviation of the data set.

4. The method for detecting concrete surface cracks based on a machine learning algorithm according to claim 3, characterized in that: The deletion of rows or columns containing missing values ​​is specifically, if the missing proportion of missing values ​​in the data set is less than 10%, the mean square error and the mean absolute error are 0-5%, the correlation coefficient is between 0.85-0.99 or -0.85-0.99, the recall rate is 0.80-1.00, the F1 score is 0.87-0.97, and the significance level p value is greater than 0.05 through the statistical test method, then the rows or columns containing missing values ​​are deleted.

5. The method for detecting concrete surface cracks based on a machine learning algorithm according to claim 3, characterized in that: The specified percentage is 90%, the value closer to other data points is the median or mean, and the outliers are converted to logarithmic transformation.

6. A concrete surface crack detection system based on machine learning algorithm, characterized in that: The system is used to implement the concrete surface crack detection method based on machine learning algorithm according to any one of claims 1 to 5, and includes the following modules: Data acquisition module: responsible for receiving raw data from ultrasonic detectors, impact elastic wave testers, radar wave detectors, high-definition cameras and temperature sensors; Data preprocessing module: responsible for cleaning, denoising and standardizing the collected raw data; Feature extraction module: responsible for using feature extraction algorithms to convert raw data into useful information for crack detection; Crack measurement module: Utilizes image processing technology and machine learning algorithms to analyze and identify the image features extracted by the feature extraction module, identify cracks on the concrete surface, determine their location, shape and size, measure the identified cracks, and transmit the crack measurement results to the multi-parameter comprehensive analysis module and the real-time monitoring and early warning module; Multi-parameter digital plate model building module: responsible for building a digital plate model based on wave velocity, frequency, amplitude, strength, elastic modulus, temperature, and linear expansion coefficient; Multi-parameter comprehensive analysis module: Use multi-parameter digital plate model to comprehensively evaluate the quality status and crack risk of concrete; Real-time monitoring and early warning module: input relevant parameters of the concrete to be tested into the multi-parameter digital plate model, and calculate the potential risk and prediction value of cracks; Database management module: responsible for managing all data of the device, including original data, processing results, and crack information; User interface and interaction module: Provides a user interface to facilitate user operations and data viewing.

7. The concrete surface crack detection system based on machine learning algorithm according to claim 6 is characterized in that: The multi-parameter comprehensive analysis module includes a multi-source data fusion unit, a model evaluation unit and a result transmission unit. The multi-source data fusion unit is responsible for fusing the data obtained by different sensors and detection means to improve the comprehensiveness of the analysis. The model evaluation unit is responsible for using a multi-parameter digital plate model to comprehensively evaluate the quality status and crack risk of concrete. The result transmission unit transmits the evaluation results to the user interface and interaction module. The real-time monitoring and early warning module includes a real-time data collection unit, an abnormality detection unit and an early warning notification unit. The real-time data collection unit continuously collects real-time feature parameters and image features extracted by the feature extraction module. The abnormality detection unit detects abnormal changes in the data according to preset thresholds or rules. Once abnormal changes are found, the early warning notification unit immediately notifies the staff.

8. The concrete surface crack detection system based on machine learning algorithm according to claim 6 is characterized in that: The data acquisition module serves as the input end of the device, the data preprocessing module receives the original data transmitted by the data acquisition module, the data processed by the data preprocessing module is transmitted to the feature extraction module, the feature extraction module transmits the extracted feature parameters and image features to the crack measurement module, the multi-parameter digital quantity board model establishment module and the real-time monitoring and early warning module, the crack measurement module transmits the measurement data of the crack and the location and shape information of the crack to the multi-parameter comprehensive analysis module, the multi-parameter digital quantity board model establishment module transmits the established digital quantity board model to the multi-parameter comprehensive analysis module, and the multi-parameter comprehensive analysis module transmits the comprehensive evaluation result to the user interface and interaction module.

Images