Concrete structure internal defect nondestructive testing method fusing big data feature extraction and deep learning

Abstract

The invention discloses a nondestructive testing method for internal defects of a concrete structure fusing big data feature extraction and deep learning. According to the method, through multi-source data collaboration and dynamic feature fusion, the accuracy and robustness of concrete structure defect detection are remarkably improved. In a data processing link, ultrasonic electromagnetic induction infrared thermal imaging data and the like acquired by a multi-source nondestructive testing technology are subjected to collaborative preprocessing, so that the influence of noise interference and environmental fluctuation is eliminated, and standardized input is provided for feature extraction. The dynamic weight distribution network further combines the relevance of each modal feature in a historical defect sample, adjusts fusion weights of different modals in real time, reinforces ultrasonic features with great contribution to cavity recognition or infrared features sensitive to cracks, effectively compresses redundant information, and improves the accuracy of cavity recognition. According to the method, features and data-driven deep features of the fused feature vectors are manually designed at the same time, so that the limitation of single-modal data is avoided, and a model can more accurately capture multi-dimensional features of defects.

Description

Technical Field

[0001] This invention belongs to the field of concrete testing technology, specifically a non-destructive testing method for internal defects in concrete structures that integrates big data feature extraction and deep learning. Background Technology

[0002] Internal testing of concrete structures is a key technical means to assess the safety and durability of existing or under-construction concrete buildings, aiming to non-destructively or minimally destructively identify internal defects and performance conditions. This technology comprehensively utilizes various physical methods such as ultrasonic waves, rebound hammer testing, radar detection, impact echo, and infrared thermography. By analyzing the propagation characteristics of sound waves, electromagnetic waves, or heat energy within concrete, it identifies potential hazards such as internal voids, cracks, honeycombing, reinforcement distribution, and corrosion. The test results provide a scientific basis for structural safety assessment, repair and reinforcement plan development, and lifespan prediction. It is a crucial link in ensuring project quality and preventing major accidents, playing an irreplaceable role in the operation and maintenance management of infrastructure such as bridges, dams, and high-rise buildings.

[0003] However, existing technologies mostly rely on single non-destructive testing techniques or simple data splicing, which makes it difficult to eliminate noise interference and environmental fluctuations from different modal data. Feature fusion often uses fixed weight allocation, which cannot dynamically adjust the contribution of each modal feature according to the defect type, resulting in redundant information interfering with the identification of key features. Deep learning models have fixed training samples and lack an incremental learning mechanism for new defect samples, resulting in insufficient adaptability to new defects. Furthermore, they lack closed-loop optimization of feature extraction parameters, making it difficult to continuously improve the accuracy and robustness of defect detection in complex structures. Summary of the Invention

[0004] The purpose of this invention is to provide a non-destructive testing method for internal defects in concrete structures that integrates big data feature extraction and deep learning in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: a non-destructive testing method for internal defects in concrete structures that integrates big data feature extraction and deep learning, the method comprising the following steps: S1: For the surface and internal areas of concrete structures, raw data are collected simultaneously using ultrasonic transmission, electromagnetic induction, infrared thermal imaging, and rebound hammer. The physical parameters corresponding to different detection technologies are recorded, such as sound wave propagation time, electromagnetic induction intensity, surface temperature distribution, and rebound value. This ensures that the data covers key stress-bearing parts of the structure and areas with a high incidence of historical defects, providing multimodal raw materials for subsequent data preprocessing.

[0006] S2: Based on the multimodal data acquired in S1, wavelet threshold denoising, adaptive baseline correction, and temperature field normalization methods are used to process the noise interference in ultrasonic signals, baseline drift of electromagnetic induction data, and ambient temperature fluctuations in infrared thermal imaging, respectively. Missing data are filled by interpolation based on samples with similar structures. The processed modal data are scaled to the same dimension and a spatial index is established according to the detection location to provide a standardized dataset for feature extraction in S3.

[0007] S3: Based on the standardized data output by S2, the system automatically extracts time-domain features such as signal peak value, mean, and variance, frequency-domain features such as dominant frequency and energy ratio, and texture features such as gray-level co-occurrence matrix parameters by traversing each detection area through a sliding window. Differentiated feature extraction window sizes are set for different modal data, for example, a 5mm window is used for ultrasonic data and a 10mm window is used for infrared thermal imaging. The extracted features are classified and organized into feature sets according to modality, providing the basic input for feature fusion in S4.

[0008] S4: Combining the multimodal feature set generated in S3, a dynamic weight allocation network is constructed. By analyzing the correlation between each modal feature and historical defect samples, such as the contribution rate of ultrasonic features to the identification of void defects and the sensitivity of infrared features to cracks, the fusion weights of different modal features are adjusted in real time. Key features with weights higher than the threshold are strengthened and retained, while redundant features with lower weights are compressed. The fused feature vector contains both artificially designed features with clear original physical meaning and data-driven deep features, providing highly discriminative input for the model training in S5.

[0009] S5: Using the feature vector fused from S4 as input, a hybrid deep learning network of "convolutional feature extraction - recurrent temporal modeling - fully connected classification" is constructed: convolutional layers extract local defect features, recurrent layers capture the spatial correlation of structural regions, and fully connected layers output the probability of defect types; historical defect samples in the feature set of S3 are used as the training set, and the network parameters are optimized through backpropagation; new samples initially identified by S6 are introduced in real time during the training process for incremental learning, and the model weights are dynamically updated to improve the adaptability to new defects.

[0010] S6: Input the preprocessed data of the structure to be detected from S2 into the model trained in S5, and output the spatial coordinates and type judgment of the defect area, such as crack, void, and looseness; set a confidence threshold for the identification results. Results higher than the threshold are directly entered into S8 for quantitative analysis, and suspicious areas lower than the threshold are marked as samples to be verified. The original data and identification results are fed back to S3 to optimize the feature extraction parameters.

[0011] S7: For the samples to be verified marked in S6, extract abnormal signal segments from their original data, such as abrupt changes in ultrasonic propagation time and abnormal temperature areas in infrared thermography. Compare these segments with the features extracted in S3 to analyze the reasons for missing or incorrectly extracted features. Adjust the window sliding step size and threshold parameters for feature extraction in S3, re-extract the features of the sample, input them into S4 for secondary fusion, and then re-identify them through the S5 model. Update the optimized feature extraction rules to the feature library of the full detection data to improve the effectiveness of the overall feature set.

[0012] S8: Defect Parameter Quantitative Analysis. Based on the defect area coordinates output by S6, physical parameters of the corresponding area are extracted from the raw data preprocessed by S2, such as ultrasonic wave velocity attenuation rate and rebound value standard deviation. Combined with the texture features and frequency domain features extracted by S3, the geometric parameters of the defect, namely length, width, depth, and severity level, are calculated according to the proportion of the defect size to the structural cross section. The quantitative results are compared with engineering safety standards to generate defect risk assessment indicators, providing a quantitative basis for the inspection report of S9.

[0013] S9: Integrate the defect quantification results and risk assessment indicators from S8, and generate an inspection report by structural part that includes a defect location distribution map, type parameters, severity, and maintenance suggestions; at the same time, store the original data of this inspection (S1), preprocessing results (S2), feature set (S3), fused features (S4), identification results (S6), and quantification parameters (S8) into the database as new samples to supplement the feature extraction sample library of S3 and the model training set of S5, forming a closed loop of "data acquisition - model optimization - inspection application" to continuously improve inspection accuracy.

[0014] In a preferred embodiment, in step S1, preprocessing is performed modally in sequence: the ultrasonic signal is decomposed into 5 levels using the db4 wavelet basis function, and soft threshold denoising is performed, with the threshold set to 1.2 times the signal standard deviation to remove environmental noise with a frequency below 5kHz; electromagnetic induction data is corrected by adaptive baseline correction to eliminate drift, with the correction window length set to 100 sampling points and the sliding step size to 20 points, so that the baseline fluctuation is controlled within ±0.5mV; infrared thermal imaging data is first normalized according to the ambient temperature, converting the temperature values ​​into relative temperatures in the [0,1] interval, and then smoothing the temperature field distribution through median filtering (3×3 window); the rebound data is averaged after removing the three maximum and three minimum values, and missing data is filled by interpolation of the rebound mean of adjacent grid units of the same type of component. The processed data are uniformly indexed according to the grid unit coordinates, and each modal data is converted into a 256×256 pixel matrix format, with the pixel values ​​corresponding to the normalized results of physical parameters, providing standardized input for feature extraction in S3.

[0015] In a preferred embodiment, in step S2, preprocessing is performed modally in sequence: the ultrasonic signal is decomposed into 5 levels using the db4 wavelet basis function, and soft threshold denoising is performed, with the threshold set to 1.2 times the signal standard deviation to remove environmental noise with a frequency below 5kHz; electromagnetic induction data is adjusted to eliminate drift through adaptive baseline correction, with a correction window length of 100 sampling points and a sliding step of 20 points to control baseline fluctuations within ±0.5mV; infrared thermal imaging data is first normalized to ambient temperature, converting temperature values ​​into relative temperatures in the [0,1] range, and then smoothed through median filtering (3×3 window); rebound data is averaged after removing the three maximum and three minimum values, and missing data is filled by interpolation of the rebound mean of adjacent grid cells of the same type of component. The processed data are uniformly indexed according to the grid cell coordinates, and each modal data is converted into a 256×256 pixel matrix format, with pixel values ​​corresponding to the normalized results of physical parameters, providing standardized input for feature extraction in S3.

[0016] In a preferred embodiment, in step S3, a sliding window is used to extract features: the ultrasonic data window size is set to 5mm×5mm with a step size of 2mm, and the signal peak value, mean, variance, and kurtosis in the time domain features are extracted; the frequency domain features are calculated by Fourier transform to determine the energy ratio of the main frequency and the three harmonic frequencies in the 0-10kHz frequency band; the infrared thermal imaging data window size is 10mm×10mm with a step size of 5mm, and the temperature gradient (horizontal and vertical directions), hot spot area ratio, and gray-level co-occurrence matrix parameters (contrast and entropy at a distance of 1 pixel and angles of 0°, 45°, 90°, and 135°) are extracted; the electromagnetic induction data window is 8mm×8mm with a step size of 3mm, and the maximum, minimum, and rate of change of the induction intensity are extracted; the rebound data is extracted as a whole by grid cell, including the mean, standard deviation, and coefficient of variation. All features are stored in a three-level structure of “modality-mesh cell-feature type”. A single mesh cell generates a feature vector containing 87 dimensions of features, including 32 dimensions of time-domain features, 28 dimensions of frequency-domain features, and 27 dimensions of texture features, providing a feature set for dynamic fusion of S4.

[0017] In a preferred embodiment, in step S4, firstly, based on a historical defect sample database, the basic contribution of each modal feature to different defect types is established through statistical analysis: for the ultrasonic feature set, its recognition accuracy and recall in void defect samples are calculated to obtain the basic contribution of the ultrasonic mode to void defects; for the infrared thermal imaging feature set, its recognition sensitivity in crack defect samples is statistically analyzed as the basic contribution of the infrared mode to cracks; similarly, the contribution of electromagnetic induction features to steel corrosion defects and the contribution of springback features to surface porosity can be obtained. Based on this, the network introduces a real-time feature anomaly score—for the input multimodal features, the deviation from the feature mean of normal structural samples is calculated to obtain the real-time anomaly degree of each modal feature. The basic contribution and anomaly score are coupled through a dynamic adjustment coefficient to generate dynamic weights for each modality: high weights are assigned to modal features with high basic contribution and significant anomaly scores, and low weights are assigned to modal features with low contribution or insignificant anomalies. After the weights are calculated, a global weight threshold is set. Feature vectors with weights higher than the threshold retain their original dimensions and enhance their values. Features with weights lower than the threshold are dimensionality-reduced and compressed. Finally, the feature vectors processed by each modality are spliced ​​together according to their spatial positions to form a fusion vector. This vector retains both physically meaningful artificially designed features such as ultrasonic wave propagation time and deep features such as texture energy extracted by the network, providing input with both physical interpretability and data discriminability for subsequent defect recognition model training.

[0018] In a preferred embodiment, in step S4, the dynamic weights need to combine the modal baseline contribution and the real-time anomaly score, and obtain the fusion weights of each modality through normalization processing, as shown in the following formula: ; In the formula: m is the modal index (e.g., 1 represents ultrasound, 2 represents infrared, 3 represents electromagnetic induction, etc.). M represents the total number of modes participating in the fusion; C m The base contribution of modality m is calculated by weighting the accuracy and recall of this modality in identifying the target defect type in historical samples: ; A m The real-time anomaly score for mode m is obtained by normalizing the Euclidean distance between the current mode feature vector and the mean of the normal sample features (the value ranges from 0 to 1, with a larger value indicating a more abnormal feature). W m The dynamic fusion weight of mode m reflects the priority of information contribution of this mode in the current detection area; To highlight key features and suppress redundant information, the feature vectors of each modality are differentiated according to their weights, as shown in the following formula: ; In the formula: Fm is the original feature vector of mode m; Fm′ is the processed feature vector of mode m; β is the key feature enhancement coefficient (taken as 1.2~1.5, determined based on the signal-to-noise ratio optimization of high-weight features in historical samples). τ is the weight threshold (taken from 0.3 to 0.5, set by statistically analyzing the weight distribution of effective features in historical defect samples). α is the redundancy feature compression coefficient (taken as 0.1~0.2, adjusted according to the structural noise level to ensure that the interference of the redundancy features on the fusion result after compression is less than 5%).

[0019] The above formula achieves dynamic weight adjustment by coupling the basic contribution and real-time anomalies, avoiding the problem of low modal information utilization in traditional fixed weight fusion. The feature enhancement and compression mechanism achieves feature selection through simple coefficient adjustment, which not only retains key features with clear physical meaning, but also reduces redundant information interference, providing highly discriminative input for subsequent deep learning models.

[0020] In a preferred embodiment, in step S5, a hybrid deep learning network is constructed using the 87-dimensional feature vector fused in S4 as input: the input layer is followed by three convolutional blocks. The first convolutional block contains 32 3×3 convolutional kernels (stride 1 padding 1) and a ReLU activation function, the second block contains 64 5×5 convolutional kernels (stride 2 padding 2), and the third block contains 128 3×3 convolutional kernels (stride 1 padding 1). Each convolutional block is followed by a 2×2 max pooling layer. The output of the convolutional layer is connected to a two-layer bidirectional LSTM network (64 hidden units per layer) to capture the spatial correlation between grid units. Finally, the probability of the defect type (cracks, cavities, looseness, corrosion, normal) is output through three fully connected layers (128→64→5). The training set was divided into 7:3 from the S3 feature set. The Adam optimizer was used (learning rate 0.001, β1=0.9 β2=0.999), and the loss function was cross-entropy loss. The training was conducted for 50 rounds (128 batches per round). Suspicious samples from the S6 feedback (accounting for 5% of the training set) were introduced every 10 rounds during training for incremental learning. The forget gate parameter of the LSTM layer was adjusted (initial value 0.8) to improve the model's adaptability to novel defects. The model convergence criterion was that the accuracy of the validation set fluctuated by less than 0.5% for 5 consecutive rounds.

[0021] In a preferred embodiment, in step S6, the preprocessed data of the structure to be detected from S2 is input into the model trained in S5 according to grid cells, and the probability and confidence level (0-1) of the defect type for each grid cell are output. A confidence threshold of 0.7 is set. For grid cells with the highest probability defect type and a confidence level ≥ 0.7, they are directly marked as defect areas, and their coordinates (accurate to 0.5 cm) and type are recorded. Grid cells with a confidence level < 0.7 are marked as suspicious areas, and their original detection data (such as the infrared temperature matrix of ultrasonic signal segments) and the feature importance ranking of the model output (the top 10 features with the highest contribution) are extracted and fed back to the feature extraction module in S3. Suspicious area data are summarized weekly and used to adjust the feature extraction window size of the corresponding mode in S3 (e.g., the infrared window is reduced to 8mm × 8mm in the suspected crack area), and the proportion of abnormal feature samples in the feature set is updated synchronously (controlled within 15% of the total samples).

[0022] In a preferred embodiment, in step S7, for the suspicious areas marked in S6, abnormal segments are extracted by detecting signal mutation points: Ultrasonic signals whose propagation time deviates from the mean of normal samples by ±3 standard deviations are considered mutation points, and segments of 200 sampling points before and after the mutation point are extracted; areas in infrared thermal images with a temperature gradient greater than 0.5℃ / mm are marked as abnormal areas, and a 10×10 pixel temperature matrix is ​​extracted for these areas. The abnormal segments are compared with the features extracted in S3. If the peak deviation rate in the temporal features is >15% or the entropy fluctuation in the texture features is >20%, it is determined that the features are missing or incorrectly extracted. At this time, the window sliding step size in S3 is adjusted (±1mm from the original step size) and the feature threshold (e.g., the main frequency energy ratio threshold is reduced from 50% to 45%). After re-extracting features from the suspicious area data, the data is input into S4, and a secondary fusion is performed using the original dynamic weight allocation network (the weight threshold remains unchanged at 0.4). Then, the model in S5 is used for re-identification. The optimized feature extraction rules are fully updated to the feature library every two weeks to ensure that the new rules cover the feature extraction needs of at least three types of typical defects.

[0023] In a preferred embodiment, in step S8, based on the defect area coordinates output in S6, the physical parameters of the corresponding grid cells are extracted from the preprocessed data in S2: the ultrasonic wave velocity attenuation rate is calculated as (normal wave velocity - measured wave velocity) / normal wave velocity (normal wave velocity is taken as 3800 m / s), and the rebound value standard deviation is calculated using data from 16 measurement points; combined with the features extracted in S3, the crack length in the geometric parameters is taken as the longest axis distance of the temperature anomaly zone in the infrared thermography, the width is calculated as the lateral span of the ultrasonic signal interruption area, and the depth is inferred from the inflection point of the electromagnetic induction intensity attenuation curve (the depth value when the induction intensity drops to 50%). The severity level is divided into 5 levels: Level 1 (defect size < 5% of structural cross-section), Level 2 (5%-10%), Level 3 (10%-20%), Level 4 (20%-30%), and Level 5 (> 30%); the risk assessment index adopts the ratio of defect size to structural safety limit, such as when the crack width > 0.2 mm or the cavity depth > 10 cm, it is marked as high risk, providing quantitative data support for the S9 report.

[0024] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, the accuracy and robustness of concrete structure defect detection are significantly improved through multi-source data collaboration and dynamic feature fusion. In the data processing stage, data collected by multi-source non-destructive testing technologies, such as ultrasonic, electromagnetic induction, and infrared thermal imaging, undergo collaborative preprocessing to eliminate noise interference and environmental fluctuations, providing standardized input for feature extraction. The dynamic weight allocation network further combines the correlation of modal features in historical defect samples, adjusting the fusion weights of different modalities in real time. Ultrasonic features that contribute significantly to void identification or infrared features sensitive to cracks are strengthened, and redundant information is effectively compressed. This ensures that the fused feature vector simultaneously retains physically meaningful artificially designed features and data-driven deep features, avoiding the limitations of single-modal data. This allows the model to more accurately capture the multi-dimensional features of defects, especially achieving higher identification specificity for multiple defect types such as cracks and voids coexisting in complex structures.

[0025] 2. In this invention, a hybrid deep learning network trained based on high-discrimination fusion features, through the synergistic effect of convolutional and recurrent layers, can extract local defect features and capture the spatial correlation of structural regions, improving the accuracy of defect localization and type judgment. Suspicious regions marked after initial identification are iteratively optimized through feedback, adjusting window size and thresholds to allow the feature set to continuously adapt to new detection scenarios, avoiding biases in feature extraction under fixed parameters. A closed-loop data process stores the original data feature set and identification results of each detection in a database, continuously supplementing training samples. This allows the model to gradually adapt to new defects or complex structural environments in long-term applications. Simultaneously, the quantitative analysis of the output defect geometric parameters and risk levels provides specific and reliable decision-making basis for engineering maintenance, balancing the technical rigor and engineering practicality of the detection. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating the process principle of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] Reference Figure 1 A non-destructive testing method for internal defects in concrete structures, integrating big data feature extraction and deep learning, includes the following steps: S1: For the surface and internal areas of concrete structures, raw data are collected simultaneously using ultrasonic transmission, electromagnetic induction, infrared thermal imaging, and rebound hammer. The physical parameters corresponding to different detection technologies are recorded, such as sound wave propagation time, electromagnetic induction intensity, surface temperature distribution, and rebound value. This ensures that the data covers key stress-bearing parts of the structure and areas with a high incidence of historical defects, providing multimodal raw materials for subsequent data preprocessing.

[0029] S2: Based on the multimodal data acquired in S1, wavelet threshold denoising, adaptive baseline correction, and temperature field normalization methods are used to process the noise interference in ultrasonic signals, baseline drift of electromagnetic induction data, and ambient temperature fluctuations in infrared thermal imaging, respectively. Missing data are filled by interpolation based on samples with similar structures. The processed modal data are scaled to the same dimension and a spatial index is established according to the detection location to provide a standardized dataset for feature extraction in S3.

[0030] S3: Based on the standardized data output by S2, the system automatically extracts time-domain features such as signal peak value, mean, and variance, frequency-domain features such as dominant frequency and energy ratio, and texture features such as gray-level co-occurrence matrix parameters by traversing each detection area through a sliding window. Differentiated feature extraction window sizes are set for different modal data, for example, a 5mm window is used for ultrasonic data and a 10mm window is used for infrared thermal imaging. The extracted features are classified and organized into feature sets according to modality, providing the basic input for feature fusion in S4.

[0031] S4: Combining the multimodal feature set generated in S3, a dynamic weight allocation network is constructed. By analyzing the correlation between each modal feature and historical defect samples, such as the contribution rate of ultrasonic features to the identification of void defects and the sensitivity of infrared features to cracks, the fusion weights of different modal features are adjusted in real time. Key features with weights higher than the threshold are strengthened and retained, while redundant features with lower weights are compressed. The fused feature vector contains both artificially designed features with clear original physical meaning and data-driven deep features, providing highly discriminative input for the model training in S5.

[0032] S5: Using the feature vector fused from S4 as input, a hybrid deep learning network of "convolutional feature extraction - recurrent temporal modeling - fully connected classification" is constructed: convolutional layers extract local defect features, recurrent layers capture the spatial correlation of structural regions, and fully connected layers output the probability of defect types; historical defect samples in the feature set of S3 are used as the training set, and the network parameters are optimized through backpropagation; new samples initially identified by S6 are introduced in real time during the training process for incremental learning, and the model weights are dynamically updated to improve the adaptability to new defects.

[0033] S6: Input the preprocessed data of the structure to be detected from S2 into the model trained in S5, and output the spatial coordinates and type judgment of the defect area, such as crack, void, and looseness; set a confidence threshold for the identification results. Results higher than the threshold are directly entered into S8 for quantitative analysis, and suspicious areas lower than the threshold are marked as samples to be verified. The original data and identification results are fed back to S3 to optimize the feature extraction parameters.

[0034] S7: For the samples to be verified marked in S6, extract abnormal signal segments from their original data, such as abrupt changes in ultrasonic propagation time and abnormal temperature areas in infrared thermography. Compare these segments with the features extracted in S3 to analyze the reasons for missing or incorrectly extracted features. Adjust the window sliding step size and threshold parameters for feature extraction in S3, re-extract the features of the sample, input them into S4 for secondary fusion, and then re-identify them through the S5 model. Update the optimized feature extraction rules to the feature library of the full detection data to improve the effectiveness of the overall feature set.

[0035] S8: Defect Parameter Quantitative Analysis. Based on the defect area coordinates output by S6, physical parameters of the corresponding area are extracted from the raw data preprocessed by S2, such as ultrasonic wave velocity attenuation rate and rebound value standard deviation. Combined with the texture features and frequency domain features extracted by S3, the geometric parameters of the defect, namely length, width, depth, and severity level, are calculated according to the proportion of the defect size to the structural cross section. The quantitative results are compared with engineering safety standards to generate defect risk assessment indicators, providing a quantitative basis for the inspection report of S9.

[0036] S9: Integrate the defect quantification results and risk assessment indicators from S8, and generate an inspection report by structural part that includes a defect location distribution map, type parameters, severity, and maintenance suggestions; at the same time, store the original data of this inspection (S1), preprocessing results (S2), feature set (S3), fused features (S4), identification results (S6), and quantification parameters (S8) into the database as new samples to supplement the feature extraction sample library of S3 and the model training set of S5, forming a closed loop of "data acquisition - model optimization - inspection application" to continuously improve inspection accuracy.

[0037] In step S1, preprocessing is performed modally in sequence: the ultrasonic signal is decomposed into 5 levels using the db4 wavelet basis function, and soft thresholding is applied for noise reduction. The threshold is set to 1.2 times the signal standard deviation to remove environmental noise with a frequency below 5kHz. Electromagnetic induction data is adjusted to eliminate drift through adaptive baseline correction. The correction window length is set to 100 sampling points, and the sliding step size is 20 points to keep the baseline fluctuation within ±0.5mV. Infrared thermal imaging data is first normalized to ambient temperature, converting the temperature values ​​into relative temperatures in the [0,1] range, and then smoothed by median filtering (3×3 window). The rebound data is averaged after removing the three maximum and three minimum values. Missing data is filled in by interpolating the mean rebound values ​​of adjacent grid cells of the same type of component. The processed data are uniformly indexed according to the grid cell coordinates, and each modal data is converted into a 256×256 pixel matrix format. The pixel values ​​correspond to the normalized results of physical parameters, providing standardized input for feature extraction in S3.

[0038] In step S2, preprocessing is performed modally in sequence: the ultrasonic signal is decomposed into 5 levels using the db4 wavelet basis function, and soft thresholding is applied for noise reduction. The threshold is set to 1.2 times the signal standard deviation to remove environmental noise with frequencies below 5kHz. Electromagnetic induction data is adjusted to eliminate drift through adaptive baseline correction. The correction window length is set to 100 sampling points, and the sliding step size is 20 points to keep the baseline fluctuation within ±0.5mV. Infrared thermal imaging data is first normalized to ambient temperature, converting the temperature values ​​into relative temperatures in the [0,1] range, and then smoothed by median filtering (3×3 window). The rebound data is averaged after removing the three maximum and three minimum values. Missing data is filled in by interpolating the mean rebound values ​​of adjacent grid cells of the same type of component. The processed data are uniformly indexed according to the grid cell coordinates, and each modal data is converted into a 256×256 pixel matrix format. The pixel values ​​correspond to the normalized results of physical parameters, providing standardized input for feature extraction in S3.

[0039] In step S3, a sliding window is used to extract features: the ultrasonic data window size is set to 5mm×5mm with a step size of 2mm, and the signal peak value, mean, variance, and kurtosis in the time domain are extracted. The frequency domain features are calculated by Fourier transform to calculate the energy ratio of the dominant frequency and the three harmonic frequencies in the 0-10kHz frequency band. The infrared thermal imaging data window size is 10mm×10mm with a step size of 5mm, and the temperature gradient (horizontal and vertical directions), hot spot area ratio, and gray-level co-occurrence matrix parameters (contrast and entropy at a distance of 1 pixel and angles of 0°, 45°, 90°, and 135°) are extracted. The electromagnetic induction data window is 8mm×8mm with a step size of 3mm, and the maximum, minimum, and rate of change of the induction intensity are extracted. The rebound data is extracted by the grid cell as a whole, including the mean, standard deviation, and coefficient of variation. All features are stored in a three-level structure of "modality-grid cell-feature type". A single grid cell generates a feature vector containing 87 dimensions of features, including 32 dimensions of time domain features, 28 dimensions of frequency domain features, and 27 dimensions of texture features, providing a feature set for dynamic fusion in S4.

[0040] In step S4, firstly, based on the historical defect sample database, the basic contribution of each modal feature to different defect types is established through statistical analysis: for the ultrasonic feature set, its recognition accuracy and recall in void defect samples are calculated to obtain the basic contribution of the ultrasonic mode to void defects; for the infrared thermal imaging feature set, its recognition sensitivity in crack defect samples is statistically analyzed as the basic contribution of the infrared mode to cracks; similarly, the contribution of electromagnetic induction features to steel corrosion defects and the contribution of springback features to surface porosity can be obtained. Based on this, the network introduces real-time feature anomaly scores—for the input multimodal features, the deviation from the feature mean of normal structural samples is calculated to obtain the real-time anomaly degree of each modal feature. The basic contribution and anomaly scores are coupled through a dynamic adjustment coefficient to generate dynamic weights for each modality: high weights are assigned to modal features with high basic contribution and significant anomaly scores, while low weights are assigned to modal features with low contribution or insignificant anomalies. After the weights are calculated, a global weight threshold is set. Feature vectors with weights higher than the threshold retain their original dimensions and enhance their values. Features with weights lower than the threshold are dimensionality-reduced and compressed. Finally, the feature vectors processed by each modality are spliced ​​together according to their spatial positions to form a fusion vector. This vector retains both physically meaningful artificially designed features such as ultrasonic wave propagation time and deep features such as texture energy extracted by the network, providing input with both physical interpretability and data discriminability for subsequent defect recognition model training.

[0041] In step S4, the dynamic weights need to combine the modality's basic contribution and the real-time anomaly score, and obtain the fusion weights of each modality through normalization, as shown in the following formula: ; In the formula: m is the modal index (e.g., 1 represents ultrasound, 2 represents infrared, 3 represents electromagnetic induction, etc.). M represents the total number of modes participating in the fusion; C m The base contribution of modality m is calculated by weighting the accuracy and recall of this modality in identifying the target defect type in historical samples: ; A m The real-time anomaly score for mode m is obtained by normalizing the Euclidean distance between the current mode feature vector and the mean of the normal sample features (the value ranges from 0 to 1, with a larger value indicating a more abnormal feature). W m The dynamic fusion weight of mode m reflects the priority of information contribution of this mode in the current detection area; To highlight key features and suppress redundant information, the feature vectors of each modality are differentiated according to their weights, as shown in the following formula: ; In the formula: Fm is the original feature vector of mode m; Fm′ is the processed feature vector of mode m; β is the key feature enhancement coefficient (taken as 1.2~1.5, determined based on the signal-to-noise ratio optimization of high-weight features in historical samples). τ is the weight threshold (taken from 0.3 to 0.5, set by statistically analyzing the weight distribution of effective features in historical defect samples). α is the redundancy feature compression coefficient (taken as 0.1~0.2, adjusted according to the structural noise level to ensure that the interference of the redundancy features on the fusion result after compression is less than 5%).

[0042] The above formula achieves dynamic weight adjustment by coupling the basic contribution and real-time anomalies, avoiding the problem of low modal information utilization in traditional fixed weight fusion. The feature enhancement and compression mechanism achieves feature selection through simple coefficient adjustment, which not only retains key features with clear physical meaning, but also reduces redundant information interference, providing highly discriminative input for subsequent deep learning models.

[0043] In step S5, the 87-dimensional feature vector fused from S4 is used as input to construct a hybrid deep learning network: the input layer is followed by three convolutional blocks. The first convolutional block contains 32 3×3 convolutional kernels (stance 1 padding 1) and a ReLU activation function, the second block contains 64 5×5 convolutional kernels (stance 2 padding 2), and the third block contains 128 3×3 convolutional kernels (stance 1 padding 1). Each convolutional block is followed by a 2×2 max pooling layer. The output of the convolutional layer is connected to a two-layer bidirectional LSTM network (64 hidden units per layer) to capture the spatial correlation between grid units. Finally, the probability of the defect type (cracks, cavities, looseness, corrosion, normal) is output through three fully connected layers (128→64→5). The training set was divided into 7:3 from the S3 feature set. The Adam optimizer was used (learning rate 0.001, β1=0.9 β2=0.999), and the loss function was cross-entropy loss. The training was conducted for 50 rounds (128 batches per round). Suspicious samples from the S6 feedback (accounting for 5% of the training set) were introduced every 10 rounds during training for incremental learning. The forget gate parameter of the LSTM layer was adjusted (initial value 0.8) to improve the model's adaptability to novel defects. The model convergence criterion was that the accuracy of the validation set fluctuated by less than 0.5% for 5 consecutive rounds.

[0044] In step S6, the preprocessed data of the structure to be detected from S2 is input into the model trained in S5 according to grid cells, and the probability and confidence level (0-1) of the defect type for each grid cell are output. A confidence threshold of 0.7 is set. For grid cells with the highest probability defect type and a confidence level ≥ 0.7, they are directly marked as defect areas, and their coordinates (accurate to 0.5 cm) and type are recorded. Grid cells with a confidence level < 0.7 are marked as suspicious areas, and their original detection data (such as the infrared temperature matrix of ultrasonic signal segments) and the feature importance ranking of the model output (the top 10 features with the highest contribution) are extracted and fed back to the feature extraction module in S3. Suspicious area data are summarized weekly and used to adjust the feature extraction window size of the corresponding mode in S3 (e.g., the infrared window is reduced to 8mm×8mm in the suspected crack area), and the proportion of abnormal feature samples in the feature set is updated synchronously (controlled within 15% of the total samples).

[0045] In step S7, for the suspicious areas marked in S6, abnormal segments are extracted through signal mutation point detection: Ultrasonic signals with propagation times deviating from the mean of normal samples by ±3 standard deviations are considered mutation points, and segments of 200 sampling points before and after the mutation point are extracted; areas with temperature gradients greater than 0.5℃ / mm in infrared thermal images are marked as abnormal areas, and a 10×10 pixel temperature matrix is ​​extracted for these areas. The abnormal segments are compared with the features extracted in S3. If the peak deviation rate in the temporal features is >15% or the entropy fluctuation in the texture features is >20%, it is determined to be a missing or incorrectly extracted feature. At this point, the window sliding step size in S3 is adjusted (±1mm from the original step size) and the feature threshold (e.g., the main frequency energy percentage threshold is reduced from 50% to 45%). Features are re-extracted from the suspicious area data and input into S4. The original dynamic weight allocation network is used for secondary fusion (the weight threshold remains unchanged at 0.4), and then the model in S5 is used for re-identification. The optimized feature extraction rules are fully updated to the feature library every two weeks to ensure that the new rules cover the feature extraction needs of at least three typical defects.

[0046] In step S8, based on the defect area coordinates output in S6, the physical parameters of the corresponding grid cells are extracted from the preprocessed data in S2: the ultrasonic wave velocity attenuation rate is calculated as (normal wave velocity - measured wave velocity) / normal wave velocity (normal wave velocity is taken as 3800m / s), and the rebound value standard deviation is calculated using data from 16 measurement points; combined with the features extracted in S3, the crack length in the geometric parameters is taken as the longest axis distance of the temperature anomaly zone in the infrared thermogram, the width is calculated as the lateral span of the ultrasonic signal interruption area, and the depth is inferred from the inflection point of the electromagnetic induction intensity attenuation curve (the depth value when the induction intensity drops to 50%). The severity level is divided into 5 levels: Level 1 (defect size < 5% of structural cross-section), Level 2 (5%-10%), Level 3 (10%-20%), Level 4 (20%-30%), and Level 5 (> 30%); the risk assessment index uses the ratio of defect size to structural safety limit, such as when the crack width > 0.2mm or the cavity depth > 10cm, it is marked as high risk, providing quantitative data support for the S9 report.

[0047] In step S9, the quantitative results and risk assessment indicators from S8 are integrated, and a three-level inspection report is generated: "component-floor-overall structure". The component-level report includes a grid unit defect distribution map (color-coded, red for level 5 and yellow for level 4), a geometric parameter table (length, width, and depth accurate to 0.1cm), and a risk level. The floor-level report summarizes the defect distribution density (number of defects per m²) for each component on the same floor. The overall structure report analyzes the correlation between defects and stress state (e.g., the proportion of defects at beam ends) and proposes repair suggestions (epoxy resin injection for cracks and pressure grouting for voids). Simultaneously, the inspection data is stored in an SQL database according to six categories: "raw data - preprocessing results - feature set - fused features - identification results - quantitative parameters". Each record contains 200MB of raw data and 50KB of feature data. New samples are added to the feature extraction sample library of S3 (total sample size maintained at 5000+) and the model training set of S5 (training set sample size 8000+) at a rate of 20% per month. A full-process closed-loop verification is conducted every quarter. By comparing the defect identification accuracy before and after optimization (target improvement ≥3% / quarter), the detection accuracy is continuously improved.

[0048] From the above, we can conclude that: In this invention, the accuracy and robustness of concrete structure defect detection are significantly improved through multi-source data collaboration and dynamic feature fusion. In the data processing stage, data from multi-source non-destructive testing technologies, including ultrasonic, electromagnetic induction, and infrared thermal imaging, undergo collaborative preprocessing to eliminate noise interference and environmental fluctuations, providing standardized input for feature extraction. The dynamic weight allocation network further combines the correlation of modal features in historical defect samples, adjusting the fusion weights of different modalities in real time. Ultrasonic features that contribute significantly to void identification or infrared features sensitive to cracks are strengthened, and redundant information is effectively compressed. This ensures that the fused feature vector simultaneously retains physically meaningful artificially designed features and data-driven deep features, avoiding the limitations of single-modal data. This allows the model to more accurately capture the multi-dimensional features of defects, especially achieving higher identification specificity for multiple defect types such as cracks and voids coexisting in complex structures.

[0049] In this invention, a hybrid deep learning network trained based on high-discrimination fusion features, through the synergistic effect of convolutional and recurrent layers, can extract local defect features and capture the spatial correlation of structural regions, improving the accuracy of defect localization and type judgment. Suspicious regions marked after initial identification are iteratively optimized through feedback, adjusting window size and thresholds to allow the feature set to continuously adapt to new detection scenarios, avoiding the bias of feature extraction under fixed parameters. A closed-loop data process stores the original data feature set and identification results of each detection in a database, continuously supplementing training samples. This allows the model to gradually adapt to new defects or complex structural environments in long-term applications. Simultaneously, the output defect geometric parameters and risk levels are quantitatively analyzed, providing specific and reliable decision-making basis for engineering maintenance, balancing the technical rigor and engineering practicality of the detection process.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-destructive testing method for internal defects in concrete structures that integrates big data feature extraction and deep learning, characterized by: The method includes the following steps: S1: The concrete structure data is collected simultaneously using ultrasonic transmission method, electromagnetic induction method, infrared thermal imaging technology and rebound hammer, and multimodal physical parameters are recorded. S2: Preprocessing multi-source data: wavelet threshold denoising of ultrasonic signals, adaptive baseline correction of electromagnetic induction data, normalization of infrared thermal imaging temperature field; interpolation completion of missing data, unified scale and establishment of spatial index, output of standardized dataset; S3: Sliding window feature extraction: For standardized data, set differential windows according to modality to extract time domain, frequency domain and texture features, and organize them into a feature set; S4: Dynamic weight fusion: Construct the correlation between network analysis modal features and historical defects, adjust the fusion weights in real time, strengthen key features, compress redundant features, and generate a fusion feature vector; S5: Model Training: Using the fused features as input, construct a hybrid deep learning network of "convolution-recurrent-fully connected" and train and optimize the parameters using historical samples; S6: Defect Identification and Feedback: Input preprocessed data, output defect coordinates and type; set a confidence threshold, high threshold results go to S8, low threshold suspicious samples are fed back to S3 to optimize feature extraction parameters; S7: Parameter optimization: Extract abnormal signals from suspicious samples, adjust the S3 window step size and threshold, perform secondary fusion recognition, and update the feature extraction rules to the feature library; S8: Parameter Quantification: Extract physical parameters of the defect area, combine features to calculate geometric parameters and severity level, and generate risk assessment indicators; S9: Reporting and Closure: Generate a report containing defect distribution, parameters, levels, and recommendations; store the data in the database, supplement samples to the feature library and training set, and form a closure.

2. The non-destructive testing method for internal defects in concrete structures that integrates big data feature extraction and deep learning as described in claim 1, characterized in that: In step S1, preprocessing is performed modally in sequence: the ultrasonic signal is decomposed into 5 layers using the db4 wavelet basis function, and soft threshold denoising is performed, with the threshold set to 1.2 times the signal standard deviation to remove environmental noise with a frequency lower than 5kHz; the electromagnetic induction data is eliminated by adaptive baseline correction to eliminate drift, with the correction window length set to 100 sampling points and the sliding step size set to 20 points to keep the baseline fluctuation within ±0.5mV; the infrared thermal imaging data is first normalized according to the ambient temperature, converting the temperature values ​​into relative temperatures in the [0,1] interval, and then smoothing the temperature field distribution through median filtering; the rebound data is averaged after removing the 3 maximum and 3 minimum values, and missing data is filled by interpolation of the rebound mean of adjacent grid units of the same type of component; the processed data are uniformly indexed according to the grid unit coordinates, and each modal data is converted into a 256×256 pixel matrix format, with the pixel values ​​corresponding to the normalized results of physical parameters, providing standardized input for feature extraction in S3.

3. The non-destructive testing method for internal defects in concrete structures integrating big data feature extraction and deep learning as described in claim 1, characterized in that: In step S2, preprocessing is performed modally in sequence: the ultrasonic signal is decomposed into 5 levels using the db4 wavelet basis function, and soft threshold denoising is performed, with the threshold set to 1.2 times the signal standard deviation to remove environmental noise with a frequency lower than 5kHz; the electromagnetic induction data is eliminated by adaptive baseline correction to eliminate drift, with the correction window length set to 100 sampling points and the sliding step size to 20 points, so that the baseline fluctuation is controlled within ±0.5mV; the infrared thermal imaging data is first normalized according to the ambient temperature, converting the temperature values ​​into relative temperatures in the [0,1] interval, and then smoothing the temperature field distribution through median filtering; the rebound data is averaged after removing the 3 maximum and 3 minimum values, and missing data is filled by interpolation of the rebound mean of adjacent grid units of the same type of component; the processed data are uniformly indexed according to the grid unit coordinates, and each modal data is converted into a 256×256 pixel matrix format, with the pixel values ​​corresponding to the normalized results of physical parameters, providing standardized input for feature extraction in S3.

4. The non-destructive testing method for internal defects in concrete structures that integrates big data feature extraction and deep learning as described in claim 1, characterized in that: In step S3, a sliding window is used to extract features: the ultrasonic data window size is set to 5mm×5mm with a step size of 2mm, and the signal peak value, mean, variance, and kurtosis in the time domain feature are extracted; the frequency domain feature is calculated by Fourier transform to calculate the energy ratio of the main frequency and the three harmonic frequencies in the 0-10kHz frequency band; the infrared thermal imaging data window size is 10mm×10mm with a step size of 5mm, and the temperature gradient, hot spot area ratio, and gray-level co-occurrence matrix parameters are extracted; the electromagnetic induction data window is 8mm×8mm with a step size of 3mm, and the maximum, minimum, and rate of change of the induction intensity are extracted; the rebound data is extracted by the grid cell as a whole, and the mean, standard deviation, and coefficient of variation are extracted; all features are stored in a three-level structure of "modality-grid cell-feature type", and a single grid cell generates a feature vector containing 87 dimensions of features, including 32 dimensions of time domain features, 28 dimensions of frequency domain features, and 27 dimensions of texture features, providing a feature set for dynamic fusion in S4.

5. The non-destructive testing method for internal defects in concrete structures integrating big data feature extraction and deep learning as described in claim 1, characterized in that: In step S4, firstly, based on the historical defect sample database, the basic contribution of each modal feature to different defect types is established through statistical analysis: for the ultrasonic feature set, its recognition accuracy and recall rate in void defect samples are calculated to obtain the basic contribution of the ultrasonic mode to void defects; for the infrared thermal imaging feature set, its recognition sensitivity in crack defect samples is statistically analyzed as the basic contribution of the infrared mode to cracks; similarly, the contribution of electromagnetic induction features to steel corrosion defects and the contribution of springback features to surface porosity can be obtained; on this basis, the network introduces real-time feature anomaly scores—for the input multimodal features, the deviation from the feature mean of normal structural samples is calculated to obtain the real-time anomaly degree of each modal feature; the basic contribution and anomaly scores are coupled through dynamic adjustment coefficients to generate dynamic weights for each modality: high weights are assigned to modal features with high basic contribution and significant anomaly scores, and low weights are assigned to modal features with low contribution or insignificant anomalies. After the weights are calculated, a global weight threshold is set. Feature vectors with weights higher than the threshold retain their original dimensions and enhance their values. Feature vectors with weights lower than the threshold are dimensionality-reduced and compressed. Finally, the feature vectors processed by each modality are spliced ​​together according to their spatial positions to form a fusion vector.

6. The non-destructive testing method for internal defects in concrete structures integrating big data feature extraction and deep learning as described in claim 1, characterized in that: In step S4, the dynamic weights need to combine the modality's basic contribution and the real-time anomaly score, and obtain the fusion weights of each modality through normalization processing, as shown in the following formula: ； In the formula: m is the modal index; M represents the total number of modes participating in the fusion; C m The base contribution of modality m is calculated by weighting the accuracy and recall of this modality in identifying the target defect type in historical samples: ; A m The real-time anomaly score for mode m is obtained by normalizing the Euclidean distance between the current mode feature vector and the mean of the normal sample features (the value ranges from 0 to 1, with a larger value indicating a more abnormal feature). W m The dynamic fusion weight of mode m reflects the priority of information contribution of this mode in the current detection area; To highlight key features and suppress redundant information, the feature vectors of each modality are differentiated according to their weights, as shown in the following formula: ； In the formula: Fm is the original feature vector of mode m; Fm′ is the processed feature vector of mode m; β is the key feature enhancement coefficient (taken as 1.2~1.5, determined based on the signal-to-noise ratio optimization of high-weight features in historical samples). τ is the weight threshold (taken from 0.3 to 0.5, set by statistically analyzing the weight distribution of effective features in historical defect samples). α is the redundancy feature compression coefficient (taken as 0.1~0.2, adjusted according to the structural noise level to ensure that the interference of the redundancy features on the fusion result after compression is less than 5%).

7. The non-destructive testing method for internal defects in concrete structures integrating big data feature extraction and deep learning as described in claim 1, characterized in that: In step S5, a hybrid deep learning network is constructed using the 87-dimensional feature vector fused from S4 as input: the input layer is followed by three convolutional blocks. The first convolutional block contains 32 3×3 convolutional kernels and a ReLU activation function, the second block contains 64 5×5 convolutional kernels, and the third block contains 128 3×3 convolutional kernels. Each convolutional block is followed by a 2×2 max pooling layer. The output of the convolutional layer is fed into a two-layer bidirectional LSTM network to capture the spatial correlation between grid cells. Finally, the defect type probability is output through three fully connected layers. The training set is divided into a 7:3 ratio from the S3 feature set. The Adam optimizer is used, and the loss function is cross-entropy loss. The model is iterated for 50 rounds. Suspicious samples from S6 are introduced every 10 rounds for incremental learning. The model's adaptability to new defects is improved by adjusting the forget gate parameters of the LSTM layer. The model convergence criterion is that the accuracy of the validation set fluctuates by less than 0.5% for five consecutive rounds.

8. The non-destructive testing method for internal defects in concrete structures integrating big data feature extraction and deep learning as described in claim 1, characterized in that: In step S6, the preprocessed data of the structure to be detected in S2 is input into the model trained in S5 according to the grid cells, and the probability and confidence of the defect type of each grid cell are output; the confidence threshold is set to 0.7, and the grid cells with the highest probability of defect type confidence ≥ 0.7 are directly marked as defect areas and their coordinates and types are recorded. Grid cells with a confidence level < 0.7 are marked as suspicious regions. Their original detection data and the feature importance ranking of the model output are extracted and fed back to the feature extraction module of S3. Suspicious area data is summarized weekly and used to adjust the feature extraction window size of the corresponding modality in S3, and to update the proportion of abnormal feature samples in the feature set synchronously.

9. The non-destructive testing method for internal defects in concrete structures integrating big data feature extraction and deep learning as described in claim 1, characterized in that: In step S7, for the suspicious areas marked in S6, abnormal segments are extracted by detecting signal mutation points: in ultrasonic signals, a deviation of the propagation time from the mean of normal samples by ±3 times the standard deviation is considered a mutation point, and segments of 200 sampling points before and after the mutation point are extracted; in infrared thermal images, areas with a temperature gradient greater than 0.5℃ / mm are marked as abnormal areas, and a 10×10 pixel temperature matrix of the area is extracted; the abnormal segments are compared with the features extracted in S3. If the peak deviation rate in the time domain features is >15% or the entropy fluctuation of the texture features is >20%, it is determined that the features are missing or misextracted; at this time, the window sliding step size and feature threshold of S3 are adjusted, and the features of the suspicious area data are re-extracted and input into S4. The original dynamic weight allocation network is used for secondary fusion, and then the model in S5 is used for re-identification; the optimized feature extraction rules are fully updated to the feature library every two weeks to ensure that the new rules cover the feature extraction needs of at least 3 types of typical defects.

10. The non-destructive testing method for internal defects in concrete structures integrating big data feature extraction and deep learning as described in claim 1, characterized in that: In step S8, based on the defect area coordinates output in S6, the physical parameters of the corresponding grid cells are extracted from the preprocessed data in S2: the ultrasonic wave velocity attenuation rate is calculated as / normal wave velocity, and the standard deviation of the rebound value is calculated using data from 16 measurement points; combined with the features extracted in S3, the crack length in the geometric parameters is taken as the longest axis distance of the temperature anomaly zone in the infrared thermogram, the width is calculated as the lateral span of the ultrasonic signal interruption area, and the depth is inferred from the inflection point of the electromagnetic induction intensity attenuation curve; the severity level is divided into 5 levels: Level 1, Level 2, Level 3, Level 4, and Level 5; the risk assessment index adopts the ratio of the defect size to the structural safety limit, such as when the crack width is >0.2mm or the cavity depth is >10cm, it is marked as high risk, providing quantitative data support for the S9 report.

Images