Intelligent analysis method for ultrasonic detection data of cast-in-place pile based on deep learning

Abstract

The invention discloses a cast-in-place pile ultrasonic detection data intelligent analysis method based on deep learning, and belongs to the technical field of cast-in-place pile ultrasonic detection. The cast-in-place pile ultrasonic detection data intelligent analysis method comprises the steps that cast-in-place pile ultrasonic detection data is obtained based on an ultrasonic detector, and preprocessing and feature extraction are conducted on the cast-in-place pile ultrasonic detection data based on natural language processing; ultrasonic detection characteristic data of the cast-in-place pile are determined; and a cast-in-place pile intelligent analysis defect recognition model is constructed based on artificial intelligence, the cast-in-place pile ultrasonic detection feature data is analyzed and recognized according to the cast-in-place pile intelligent analysis defect recognition model, the cast-in-place pile intelligent analysis defect recognition result is determined, and the cast-in-place pile is evaluated and managed according to the cast-in-place pile intelligent analysis defect recognition result. The problem that the ultrasonic detection efficiency and accuracy of the cast-in-place pile are low due to the fact that the defects of the pile body cannot be automatically recognized and the integrity level of the pile body cannot be automatically evaluated in the prior art is solved. The method can automatically identify the defects of the pile body and evaluate the integrity level of the pile body, and can improve the ultrasonic detection efficiency and accuracy of the cast-in-place pile.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic testing technology for cast-in-place piles, specifically to an intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning. Background Technology

[0002] A cast-in-place pile is a type of pile made by drilling a hole in place and pouring in concrete or reinforced concrete. The traditional ultrasonic testing process for cast-in-place piles is as follows: a sonic logging tube is pre-embedded in the pile body, and acoustic parameters such as sound time, amplitude, and frequency are obtained by emitting and receiving ultrasonic waves, thereby judging the integrity of the pile body, such as defects such as segregation, mud inclusion, necking, and broken piles.

[0003] Traditional ultrasonic testing of cast-in-place piles mainly relies on inspectors manually analyzing acoustic parameter profiles and judging the type of defects based on experience. This method is highly dependent on personal experience, inefficient, subjective, and difficult to detect small and complex defect patterns.

[0004] Existing technologies cannot automatically identify pile defects and assess pile integrity levels, resulting in low efficiency and accuracy of ultrasonic testing for cast-in-place piles. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning, which can automatically identify pile defects and assess the pile integrity level, thereby improving the efficiency and accuracy of ultrasonic testing of cast-in-place piles and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A deep learning-based intelligent analysis method for ultrasonic testing data of cast-in-place piles includes:

[0008] Ultrasonic testing data of cast-in-place piles is obtained using an ultrasonic testing instrument, and the ultrasonic testing data of cast-in-place piles is preprocessed and feature extracted based on natural language processing to determine the ultrasonic testing feature data of cast-in-place piles.

[0009] An intelligent analysis and defect identification model for cast-in-place piles is constructed based on artificial intelligence. The ultrasonic testing feature data of the cast-in-place piles are analyzed and identified based on the intelligent analysis and defect identification model to determine the intelligent analysis and defect identification results of the cast-in-place piles. The cast-in-place piles are then evaluated and managed based on the intelligent analysis and defect identification results.

[0010] Preferably, the ultrasonic testing data of the cast-in-place pile is acquired based on an ultrasonic testing instrument, including:

[0011] The ultrasonic testing instrument is used to measure and test the cast-in-place piles and record the sound velocity, amplitude, frequency and waveform parameters of the sound waves to obtain ultrasonic testing data of the cast-in-place piles.

[0012] Among them, the speed of sound is the speed at which sound waves propagate in concrete, and is used to reflect the elastic modulus and density of concrete.

[0013] Amplitude is the peak amplitude of the first cycle of the received ultrasonic signal, used to reflect the energy attenuation of ultrasonic waves during propagation;

[0014] The frequency is the main frequency or frequency component of the received ultrasonic signal; the waveform is the voltage and time signal curve received over the entire time domain.

[0015] Preferably, the ultrasonic testing data of the cast-in-place piles is preprocessed based on natural language processing, and the following operations are performed:

[0016] Denoising is performed on ultrasonic testing data of cast-in-place piles based on wavelet denoising to improve the signal-to-noise ratio of the ultrasonic testing data of cast-in-place piles. Specifically, wavelets with similar shapes to ultrasonic waves are selected, and the original waveform at each depth point is decomposed into N layers of wavelets to obtain approximation coefficients and detail coefficients. The detail coefficients are processed by applying a threshold function, and the signal is reconstructed using the processed coefficients to obtain the denoised waveform.

[0017] The ultrasonic testing data of cast-in-place piles is examined based on linear traversal to identify missing and outlier values ​​in the data. The missing and outlier values ​​in the data are then evaluated to determine whether they are valuable for intelligent analysis and defect identification of cast-in-place piles.

[0018] When missing and outlier values ​​in the ultrasonic testing data of cast-in-place piles are valuable for intelligent analysis and defect identification of cast-in-place piles, interpolation is used to fill in the missing values ​​and the mean value is used to replace the outlier values.

[0019] When missing or outlier values ​​in the ultrasonic testing data of cast-in-place piles are of no value for intelligent analysis and defect identification of cast-in-place piles, then the missing or outlier values ​​in the ultrasonic testing data of cast-in-place piles are removed.

[0020] Preferably, feature extraction is performed on the ultrasonic testing data of cast-in-place piles based on natural language processing, and the following operations are performed:

[0021] Based on Z-Score standardization, the ultrasonic testing data of cast-in-place piles is normalized, which transforms the dimensional ultrasonic testing data of cast-in-place piles into dimensionless data expression, removes the dimensional differences between the ultrasonic testing data of cast-in-place piles, and forms standardized ultrasonic testing data of cast-in-place piles.

[0022] Enhancement processing is performed on the ultrasonic testing data of cast-in-place piles. Specifically, for the images in the ultrasonic testing data of cast-in-place piles, rotation, flipping, cropping, adding Gaussian noise, and adjusting brightness and contrast are performed. For the sequences in the ultrasonic testing data of cast-in-place piles, noise is added and small-amplitude scaling transformations are performed.

[0023] Feature extraction is performed on the ultrasonic testing data of cast-in-place piles. Features related to intelligent analysis and defect identification of cast-in-place piles are extracted from the ultrasonic testing data of cast-in-place piles to determine the ultrasonic testing feature data of cast-in-place piles.

[0024] Preferably, an intelligent analytical defect identification model for cast-in-place piles is constructed based on artificial intelligence, and the following operations are performed:

[0025] Based on the intelligent analysis requirements of ultrasonic testing data of cast-in-place piles, historical data on defect identification of cast-in-place piles are collected and divided into training set and test set.

[0026] The deep learning model was trained based on the training set to determine the intelligent analytical defect identification model for cast-in-place piles.

[0027] The intelligent analytical defect identification model for cast-in-place piles was tested using a test set to evaluate its performance, determine the model test evaluation results, and adjust and optimize the model based on these results to determine the optimal intelligent analytical defect identification model for cast-in-place piles.

[0028] Preferably, the intelligent analytical defect identification model for cast-in-place piles is tested using a test set to evaluate its performance. The following operations are performed:

[0029] The test set is input into the intelligent analytical defect identification model for cast-in-place piles. The performance of the intelligent analytical defect identification model for cast-in-place piles is tested based on the test set. The model output results are used to determine whether the intelligent analytical defect identification model for cast-in-place piles can achieve the expected effect of automatically identifying defects in cast-in-place piles.

[0030] When the intelligent analysis and defect identification model for cast-in-place piles fails to achieve the expected effect of automatically identifying defects in cast-in-place piles, the parameters of the intelligent analysis and defect identification model for cast-in-place piles are adjusted and optimized. The performance of the intelligent analysis and defect identification model for cast-in-place piles after parameter adjustment and optimization is tested again, forming a closed-loop test and optimization of the intelligent analysis and defect identification model for cast-in-place piles, until the intelligent analysis and defect identification model for cast-in-place piles can achieve the expected effect of automatically identifying defects in cast-in-place piles, and the optimal intelligent analysis and defect identification model for cast-in-place piles is determined.

[0031] Preferably, the ultrasonic testing feature data of the cast-in-place pile is analyzed and identified based on the intelligent analytical defect identification model of the cast-in-place pile, and the following operations are performed:

[0032] The ultrasonic testing feature data of the cast-in-place pile is input into the intelligent analysis and defect identification model of the cast-in-place pile. The model analyzes and identifies the ultrasonic testing feature data of the cast-in-place pile, automatically identifies defects in the cast-in-place pile, determines the intelligent analysis and defect identification results of the cast-in-place pile, and displays the intelligent analysis and defect identification results of the cast-in-place pile in a visual form.

[0033] Preferably, the cast-in-place piles are evaluated and managed based on the intelligent analysis and defect identification results, and the following operations are performed:

[0034] The integrity level of the cast-in-place pile is assessed based on the intelligent analysis and defect identification results, including Class I, Class II, Class III and Class IV piles. The corresponding suggestions are automatically selected from a predefined document library according to the integrity level of the pile, and a structured cast-in-place pile inspection report is generated.

[0035] For Class I piles, the pile body structure is intact and there are no obvious defects;

[0036] For Class II piles, the pile body has minor defects that do not affect normal use, but these need to be addressed during subsequent construction.

[0037] For Class III piles, there are obvious defects in the pile body. It is necessary to analyze the impact on the pile bearing capacity together with the design and supervision units and determine the treatment plan.

[0038] For Class IV piles, which have serious defects, core drilling is required for verification, and reinforcement or scrapping is necessary.

[0039] Preferably, the ultrasonic testing feature data of the cast-in-place pile is analyzed and identified based on the intelligent analytical defect identification model of the cast-in-place pile, including:

[0040] Obtain an ultrasonic training dataset for cast-in-place piles; the ultrasonic training dataset includes sound velocity, amplitude, frequency, and waveform parameters; add depth labels and defect pre-labels to the ultrasonic training dataset to form a labeled feature set;

[0041] The labeled feature set is generated based on the Physically Constrained Variational Autoencoder (PC-VAE). The PC-VAE includes an input layer, an encoder, a physical constraint layer, and a decoder. The physical constraint layer adds hard constraints on sound speed and density and soft constraints on feature distribution corresponding to different depths to generate enhanced samples that conform to the physical laws of ultrasonic propagation. The enhanced samples in the training dataset are spliced ​​together with the enhanced samples according to the depth segmentation rules to form an expanded dataset.

[0042] Multimodal cross-scale feature extraction is performed on the expanded dataset to obtain statistical feature vectors, frequency feature vectors, and waveform feature vectors;

[0043] A cross-attention fusion network (CAFN) is constructed, which maps the statistical feature vector, frequency feature vector, and waveform feature vector to the same dimension and constructs a feature matrix. The attention weights of each modal feature are calculated through cross-attention. Based on the attention weights, the modal features are weighted and fused to obtain a cross-modal fusion feature vector.

[0044] A TCN-Transformer hybrid core network is constructed, comprising an input layer, a local feature extraction layer, a global association capture layer, a feature interaction layer, and an output layer. The cross-modal fused feature vector and the depth label are concatenated and input into the hybrid core network. The local feature extraction layer captures depth local defect features through a 4-layer 1D-TCN, the global association capture layer captures cross-depth defect associations through a 2-layer Transformer encoder, the feature interaction layer fuses local and global features through residual connections and layer normalization, and the output layer outputs the defect type probability distribution and severity regression value through two parallel fully connected branches.

[0045] The total loss function is constructed based on classification loss, regression loss and physical consistency loss. The TCN-Transformer hybrid core network is trained using a two-stage training strategy. In the first stage, the Transformer layer is frozen using the original training set to train the TCN layer and the output layer. In the second stage, all layers are unfrozen on the expanded dataset for fine-tuning. When the training results meet the preset convergence conditions, the initial defect recognition model is obtained.

[0046] An adaptive dynamic decision-making module is constructed, which generates probability confidence thresholds, physical constraint distance thresholds, and time series consistency thresholds based on a normal sample metric library through kernel density estimation.

[0047] The initial defect identification model is combined with the adaptive dynamic decision-making module to obtain a trained intelligent analytical defect identification model for cast-in-place piles.

[0048] Preferably, the method further includes supplementing the intelligent analytical defect identification model of the cast-in-place pile with additional training to obtain a supplemented intelligent analytical defect identification model of the cast-in-place pile.

[0049] The intelligent analytical defect identification model for cast-in-place piles is further trained to obtain a supplemented intelligent analytical defect identification model for cast-in-place piles, including:

[0050] A supplementary data system is constructed, which includes core supplementary data and auxiliary supplementary data. The core supplementary data includes manual defect assessment records, non-destructive testing raw data, full data of the construction process, and historical defect verification cases. The auxiliary supplementary data includes industry standards and specifications data.

[0051] The supplementary data is subjected to a four-dimensional progressive filtering process to obtain a high-quality supplementary dataset;

[0052] Based on the four-dimensional index, the diagnostic fit score is calculated, and the high-quality supplementary dataset is classified into high-fit dataset, medium-fit dataset and low-fit dataset.

[0053] The high-fit, medium-fit, and low-fit datasets are divided into different training batches according to a preset ratio. The intelligent analytical defect identification model for cast-in-place piles is dynamically batch-supplemented to obtain the supplemented intelligent analytical defect identification model for cast-in-place piles.

[0054] Compared with the prior art, the beneficial effects of the present invention are:

[0055] This invention uses an ultrasonic testing instrument to measure and record the sound velocity, amplitude, frequency, and waveform parameters of cast-in-place piles, acquiring ultrasonic testing data. Based on natural language processing, the ultrasonic testing data is preprocessed and features are extracted. Features related to intelligent defect identification of cast-in-place piles are extracted from the ultrasonic testing data, determining the ultrasonic testing feature data. An intelligent defect identification model for cast-in-place piles is constructed based on artificial intelligence. The ultrasonic testing feature data is analyzed and identified according to the model, automatically identifying defects in the cast-in-place piles, determining the intelligent defect identification results, and displaying the results in a visual format. Furthermore, the cast-in-place piles are evaluated and managed based on the intelligent defect identification results. This invention can automatically identify pile defects and assess the pile integrity level, improving the efficiency and accuracy of ultrasonic testing of cast-in-place piles. Attached Figure Description

[0056] Figure 1 This is a flowchart of the intelligent analysis method for ultrasonic testing data of cast-in-place piles according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] To address the current limitations of automatic pile defect identification and pile integrity assessment, which lead to low efficiency and accuracy in ultrasonic testing of cast-in-place piles, please refer to [link to relevant documentation]. Figure 1 This embodiment provides the following technical solution:

[0059] Example 1

[0060] A deep learning-based intelligent analysis method for ultrasonic testing data of cast-in-place piles includes:

[0061] Ultrasonic testing data of cast-in-place piles is acquired using an ultrasonic testing instrument, and preprocessed and feature extracted based on natural language processing to determine the characteristic data of ultrasonic testing of cast-in-place piles.

[0062] In this embodiment, acquiring ultrasonic testing data of the cast-in-place pile based on an ultrasonic testing instrument includes:

[0063] The ultrasonic testing instrument is used to measure and test the cast-in-place piles and record the sound velocity, amplitude, frequency and waveform parameters of the sound waves to obtain ultrasonic testing data of the cast-in-place piles.

[0064] Among them, the speed of sound is the speed at which sound waves propagate in concrete, which is used to reflect the elastic modulus and density of concrete. According to the elastic wave theory, the speed of sound wave propagation in a solid medium depends on the elastic modulus and density of the medium. For composite materials like concrete, the speed of sound is highly related to the strength of the cement matrix, the properties and distribution of aggregates, and the internal density. The denser and harder the material, the faster the speed of sound wave propagation.

[0065] Amplitude is the peak amplitude of the first cycle of the received ultrasonic signal. It is used to reflect the energy attenuation of ultrasonic waves during propagation. When ultrasonic waves encounter defects (such as bubbles or mud), they will scatter, reflect, and diffract, resulting in a sharp decrease in energy. A significant drop in amplitude usually means that there is a medium inside the concrete that strongly absorbs sound wave energy, such as mud interlayers or honeycomb structures. It is very effective in judging the severity of defects.

[0066] The frequency is the main frequency or frequency component of the received ultrasonic signal, which can be obtained through spectrum analysis. Among them, the high-frequency component of ultrasonic waves is more easily attenuated during propagation. When the quality of concrete is poor, the high-frequency component will be lost in large quantities, causing the main frequency of the received signal to shift to a lower frequency. The frequency reduction is an indicator of uneven concrete quality or the presence of minor defects. It can be used to judge the amplitude anomaly.

[0067] The waveform is the voltage and time signal curve received over the entire time domain. The waveform contains all the information about the propagation of the sound wave. The presence of defects not only affects the arrival time and amplitude of the wave, but also changes the shape of the wave.

[0068] An intelligent analysis and defect identification model for cast-in-place piles is constructed based on artificial intelligence. The ultrasonic testing feature data of the cast-in-place piles are analyzed and identified based on the intelligent analysis and defect identification model to determine the intelligent analysis and defect identification results of the cast-in-place piles. The cast-in-place piles are then evaluated and managed based on the intelligent analysis and defect identification results.

[0069] In this embodiment, the ultrasonic testing feature data of the cast-in-place pile is analyzed and identified according to the intelligent analytical defect identification model of the cast-in-place pile, and the following operations are performed:

[0070] The ultrasonic testing feature data of the cast-in-place pile is input into the intelligent analysis and defect identification model of the cast-in-place pile. The model analyzes and identifies the ultrasonic testing feature data of the cast-in-place pile, automatically identifies defects in the cast-in-place pile, determines the intelligent analysis and defect identification results of the cast-in-place pile, and displays the intelligent analysis and defect identification results of the cast-in-place pile in a visual form.

[0071] In this embodiment, the cast-in-place piles are evaluated and managed based on the intelligent analysis and defect identification results, and the following operations are performed:

[0072] The integrity level of the cast-in-place pile is assessed based on the intelligent analysis and defect identification results, including Class I, Class II, Class III and Class IV piles. The corresponding suggestions are automatically selected from a predefined document library according to the integrity level of the pile, and a structured cast-in-place pile inspection report is generated.

[0073] For Class I piles, the pile body structure is intact and there are no obvious defects;

[0074] For Class II piles, the pile body has minor defects that do not affect normal use, but these need to be addressed during subsequent construction.

[0075] For Class III piles, there are obvious defects in the pile body. It is necessary to analyze the impact on the pile bearing capacity together with the design and supervision units and determine the treatment plan.

[0076] For Class IV piles, which have serious defects, core drilling is required for verification, and reinforcement or scrapping is necessary.

[0077] Example 2

[0078] In this embodiment, the ultrasonic testing data of the cast-in-place pile is preprocessed based on natural language processing, and the following operations are performed:

[0079] Denoising is performed on ultrasonic testing data of cast-in-place piles based on wavelet denoising to improve the signal-to-noise ratio of the ultrasonic testing data of cast-in-place piles. Specifically, wavelets with similar shapes to ultrasonic waves are selected, and the original waveform at each depth point is decomposed into N layers of wavelets to obtain approximation coefficients and detail coefficients. The detail coefficients are processed by applying a threshold function, and the signal is reconstructed using the processed coefficients to obtain the denoised waveform.

[0080] The ultrasonic testing data of cast-in-place piles is examined based on linear traversal to identify missing and outlier values ​​in the data. The missing and outlier values ​​in the data are then evaluated to determine whether they are valuable for intelligent analysis and defect identification of cast-in-place piles.

[0081] When missing and outlier values ​​in the ultrasonic testing data of cast-in-place piles are valuable for intelligent analysis and defect identification of cast-in-place piles, interpolation is used to fill in the missing values ​​and the mean value is used to replace the outlier values.

[0082] When missing or outlier values ​​in the ultrasonic testing data of cast-in-place piles are of no value for intelligent analysis and defect identification of cast-in-place piles, then the missing or outlier values ​​in the ultrasonic testing data of cast-in-place piles are removed.

[0083] In this embodiment, feature extraction is performed on the ultrasonic testing data of the cast-in-place piles based on natural language processing, and the following operations are performed:

[0084] Based on Z-Score standardization, the ultrasonic testing data of cast-in-place piles is normalized, which transforms the dimensional ultrasonic testing data of cast-in-place piles into dimensionless data expression, removes the dimensional differences between the ultrasonic testing data of cast-in-place piles, and forms standardized ultrasonic testing data of cast-in-place piles.

[0085] Enhancement processing is performed on the ultrasonic testing data of cast-in-place piles. Specifically, for the images in the ultrasonic testing data of cast-in-place piles, rotation, flipping, cropping, adding Gaussian noise, and adjusting brightness and contrast are performed. For the sequences in the ultrasonic testing data of cast-in-place piles, noise is added and small-amplitude scaling transformations are performed.

[0086] Feature extraction is performed on the ultrasonic testing data of cast-in-place piles. Features related to intelligent analysis and defect identification of cast-in-place piles are extracted from the ultrasonic testing data of cast-in-place piles to determine the ultrasonic testing feature data of cast-in-place piles.

[0087] It should be noted that by performing noise reduction, inspection, normalization, enhancement processing, and feature extraction on the ultrasonic testing data of cast-in-place piles, features related to intelligent analysis and defect identification of cast-in-place piles can be extracted from the ultrasonic testing data. This allows for the determination of ultrasonic testing feature data of cast-in-place piles, which facilitates subsequent analysis and identification of ultrasonic testing feature data of cast-in-place piles based on the intelligent analysis and defect identification model of cast-in-place piles, automatically identifying defects in cast-in-place piles, and determining the intelligent analysis and defect identification results of cast-in-place piles.

[0088] Example 3

[0089] In this embodiment, an intelligent analytical defect identification model for cast-in-place piles is constructed based on artificial intelligence, and the following operations are performed:

[0090] Based on the intelligent analysis requirements of ultrasonic testing data of cast-in-place piles, historical data on defect identification of cast-in-place piles are collected and divided into training set and test set.

[0091] The deep learning model was trained based on the training set to determine the intelligent analytical defect identification model for cast-in-place piles.

[0092] The intelligent analytical defect identification model for cast-in-place piles was tested using a test set to evaluate its performance, determine the model test evaluation results, and adjust and optimize the model based on these results to determine the optimal intelligent analytical defect identification model for cast-in-place piles.

[0093] In this embodiment, a test set is used to test the intelligent analytical defect identification model for cast-in-place piles and evaluate its performance. The following operations are performed:

[0094] The test set is input into the intelligent analytical defect identification model for cast-in-place piles. The performance of the intelligent analytical defect identification model for cast-in-place piles is tested based on the test set. The model output results are used to determine whether the intelligent analytical defect identification model for cast-in-place piles can achieve the expected effect of automatically identifying defects in cast-in-place piles.

[0095] When the intelligent analysis and defect identification model for cast-in-place piles fails to achieve the expected effect of automatically identifying defects in cast-in-place piles, the parameters of the intelligent analysis and defect identification model for cast-in-place piles are adjusted and optimized. The performance of the intelligent analysis and defect identification model for cast-in-place piles after parameter adjustment and optimization is tested again, forming a closed-loop test and optimization of the intelligent analysis and defect identification model for cast-in-place piles, until the intelligent analysis and defect identification model for cast-in-place piles can achieve the expected effect of automatically identifying defects in cast-in-place piles, and the optimal intelligent analysis and defect identification model for cast-in-place piles is determined.

[0096] In summary, by using an ultrasonic testing instrument to measure and record the sound velocity, amplitude, frequency, and waveform parameters of the cast-in-place piles, ultrasonic testing data of the cast-in-place piles is obtained. Natural language processing is used to preprocess and extract features from this data, extracting features related to intelligent defect identification of the cast-in-place piles. Ultrasonic testing feature data of the cast-in-place piles is determined, and an intelligent defect identification model for cast-in-place piles is constructed based on artificial intelligence. This model is then used to analyze and identify the ultrasonic testing feature data of the cast-in-place piles, automatically identifying defects and determining the intelligent defect identification results. These results are then visualized and used for evaluation and management of the cast-in-place piles. This method can automatically identify pile defects and assess the pile integrity level, thus improving the efficiency and accuracy of ultrasonic testing of cast-in-place piles.

[0097] The ultrasonic testing feature data of cast-in-place piles are analyzed and identified based on the intelligent analytical defect identification model for cast-in-place piles, including:

[0098] Obtain an ultrasonic training dataset for cast-in-place piles; the ultrasonic training dataset includes sound velocity, amplitude, frequency, and waveform parameters; add depth labels and defect pre-labels to the ultrasonic training dataset to form a labeled feature set;

[0099] The labeled feature set is generated based on the Physically Constrained Variational Autoencoder (PC-VAE). The PC-VAE includes an input layer, an encoder, a physical constraint layer, and a decoder. The physical constraint layer adds hard constraints on sound speed and density and soft constraints on feature distribution corresponding to different depths to generate enhanced samples that conform to the physical laws of ultrasonic propagation. The enhanced samples in the training dataset are spliced ​​together with the enhanced samples according to the depth segmentation rules to form an expanded dataset.

[0100] Multimodal cross-scale feature extraction is performed on the expanded dataset to obtain statistical feature vectors, frequency feature vectors, and waveform feature vectors;

[0101] A cross-attention fusion network (CAFN) is constructed, which maps the statistical feature vector, frequency feature vector, and waveform feature vector to the same dimension and constructs a feature matrix. The attention weights of each modal feature are calculated through cross-attention. Based on the attention weights, the modal features are weighted and fused to obtain a cross-modal fusion feature vector.

[0102] A TCN-Transformer hybrid core network is constructed, comprising an input layer, a local feature extraction layer, a global association capture layer, a feature interaction layer, and an output layer. The cross-modal fused feature vector and the depth label are concatenated and input into the hybrid core network. The local feature extraction layer captures depth local defect features through a 4-layer 1D-TCN, the global association capture layer captures cross-depth defect associations through a 2-layer Transformer encoder, the feature interaction layer fuses local and global features through residual connections and layer normalization, and the output layer outputs the defect type probability distribution and severity regression value through two parallel fully connected branches.

[0103] The total loss function is constructed based on classification loss, regression loss and physical consistency loss. The TCN-Transformer hybrid core network is trained using a two-stage training strategy. In the first stage, the Transformer layer is frozen using the original training set to train the TCN layer and the output layer. In the second stage, all layers are unfrozen on the expanded dataset for fine-tuning. When the training results meet the preset convergence conditions, the initial defect recognition model is obtained.

[0104] An adaptive dynamic decision-making module is constructed, which generates probability confidence thresholds, physical constraint distance thresholds, and time series consistency thresholds based on a normal sample metric library through kernel density estimation.

[0105] The initial defect identification model is combined with the adaptive dynamic decision-making module to obtain a trained intelligent analytical defect identification model for cast-in-place piles.

[0106] In this embodiment, sound speed and amplitude features are processed based on depth gradient calculation and statistical analysis to obtain statistical feature vectors; frequency features are processed based on wavelet packet transform and CNN network to obtain frequency feature vectors; and waveform features are processed based on segmented time-series coding and 1D-TCN to obtain waveform feature vectors.

[0107] In this embodiment, the normal sample measurement library includes the probability confidence level of defect-free cast-in-place piles, physical constraint distance, and temporal consistency data. The probability confidence level is the maximum value of the probability of the defect type. The physical constraint distance is the weighted distance of sound speed, amplitude, and frequency, and the weights are calculated by the inverse of the variance of historical normal samples. The temporal consistency is the consistency ratio of multiple detections of defect types.

[0108] In this embodiment, depth labels and defect pre-labels are added to the ultrasonic training dataset to form a labeled feature set, including: depth label z: the depth of the sample is marked by dividing the depth data of the cast-in-place pile into 1m intervals, and the theoretical physical properties of the concrete at that depth are associated; defect pre-label D: the ultrasonic data is labeled with the defect type and severity based on the core sample detection results.

[0109] In this embodiment, the PC-VAE network structure design input layer includes 4-dimensional ultrasonic features (sound velocity v, amplitude A, frequency f, waveform subsequence). The waveform passes through a length ( The waveform is divided into subsequences by a sliding window of 200. The encoder uses a 3-layer fully connected network and a 1-layer LSTM. The fully connected layers extract statistical features (mean / gradient of v, decay rate of A), and the LSTM captures the temporal dependence of the waveform subsequences and outputs latent variables. (Dimension 32); Physical constraint layer: Add hard constraints to the latent variable space (speed of sound v and density) physical relationship k is the elastic coefficient of concrete, obtained by fitting historical normal samples) and soft constraints (feature distribution constraints corresponding to different depths z, such as the mean of the latent variable v being 15% lower when z≥15m than when z<5m); Decoder: symmetric to the encoder, outputting generated samples with the same dimension as the input, the loss function is reconstruction loss + physical constraint loss: ; For the i-th sample in the training dataset; for The i-th generated sample; is the theoretical density corresponding to the i-th sample; , is the weight; is the total number of samples; is the sound velocity value of the i-th sample; is the elastic coefficient of concrete; is the depth label; is the conditional latent variable distribution; is the standard normal distribution; KL(⋅||⋅): is the KL divergence; is the loss function; For the problem of scarce defective samples, cross-generation is performed according to the defect type - depth training dataset of the training dataset, and 5 enhanced samples are generated for each defective sample; Considering the physical property differences of different depths of cast-in-place piles (shallow part: large pouring pressure, high density; middle part: uniform stress; deep part: easy accumulation of sediment), it is divided into 3 segments according to depth: shallow segment ( ): (0 < z ≤ 5m); middle segment : (5 < z ≤ 15m); deep segment : (z > 15m); splicing rule: within the same depth segment, splice the samples in the training set and the samples of the same defect type generated by PC-VAE according to the 7:3 training dataset ratio of the training dataset, and the generated samples maintain the overall consistency of sound velocity, wave amplitude, frequency, and waveform; Finally, form an expanded dataset of the training dataset training set + enhanced samples , and the sample size is increased by 3 - 5 times compared with the samples in the training set.

[0110] In this embodiment, a cross-attention fusion network CAFN is constructed, mapping the statistical feature vector, frequency feature vector, and waveform feature vector to the same dimension and constructing a feature matrix, calculating the attention weights of each modal feature through cross-attention, and performing weighted fusion on each modal feature based on the attention weights to obtain a cross-modal fusion feature vector, including: sound velocity (v) and wave amplitude (A) feature processing: calculating the depth gradient feature of the training dataset training dataset (sound velocity difference between adjacent depths), ; (relative attenuation rate of wave amplitude, is the wave amplitude at the pile top), reflecting the change trend of sound velocity / wave amplitude with depth; constructing a statistical feature vector of the training dataset training dataset: within each depth segment , calculate the mean value of v, variance , proportion of outliers , peak value of A, decay half-life , and form a 12-dimensional statistical feature vector Frequency (f) feature processing: Wavelet packet transform is used to perform three-level wavelet packet decomposition on the frequency signal of each depth, resulting in time-frequency maps of eight frequency bands; energy distribution features of the training dataset of the time-frequency maps are extracted; the energy proportion of each frequency band is calculated. ; ( (energy of the k-th frequency band) and dominant frequency offset ( This is the measured main frequency. Using the reference frequency as a reference, spatial features of the time-frequency map are extracted through a CNN (3 convolutional layers + 2 pooling layers) to form a 64-dimensional frequency feature vector. Waveform W(t) Feature Processing: The time-domain waveform W(t) is segmented into training datasets for temporal encoding. The training dataset is divided into subsequences with a time step (t=0.1μs), each subsequence being 200 bytes long. Local temporal dependencies of the waveform (such as peak intervals and valley depths) are captured using 1D-TCN (3 layers of dilated convolution with dilation coefficients of 1, 2, and 4). Waveform morphology features are extracted from the training dataset. The number of peaks in each subsequence is calculated. Waveform distortion ( (The average value of the normal waveform), combined with the 64-dimensional time-series features output by the TCN, forms a 128-dimensional waveform feature vector. ;

[0111] Construct a Cross-Attention Fusion Network (CAFN) to capture the physical correlations between different modal features (e.g., low sound speed → large amplitude attenuation → frequency shift): Modal feature alignment: ... (12-dimensional) is mapped to 64-dimensional through a fully connected layer; (64 dimensions) remain unchanged; The 128-dimensional feature matrix is ​​projected to 64-dimensional space through a fully connected layer to obtain the aligned feature matrix. (3×64 dimensions); The aligned 3×64-dimensional feature matrix; This is the transpose of the frequency eigenvector; The waveform feature vector after projection; cross-attention calculation: query: (Sound speed-amplitude characteristics, as the core reference mode); key: (( Training dataset ); value: ( Training dataset Attention weights: ( =64 is the feature dimension; Frequency characteristics Attention weights; This is the mapped sound speed-amplitude eigenvector; This is a transpose operation; It is a frequency eigenvector; Modal features; This is the scaling factor; similarly, it is calculated. ; Fusion feature output: This results in 64-dimensional cross-modal fusion features; This is a cross-modal fusion feature vector.

[0112] In this embodiment, the hybrid core network includes an input layer, a local feature extraction layer, a global association capture layer, a feature interaction layer, and an output layer. The cross-modal fused feature vector and depth labels are concatenated and input into the hybrid core network. The local feature extraction layer captures depth-specific local defect features through a 4-layer 1D-TCN. The global association capture layer captures cross-depth defect associations through a 2-layer Transformer encoder. The feature interaction layer fuses local and global features through residual connections and layer normalization. The output layer outputs the defect type probability distribution and severity regression value through two parallel fully connected branches, including:

[0113] Input layer: Input (64-dimensional) + depth label z (1-dimensional), where z is normalized to the [0,1] interval, and concatenated to form a 65-dimensional input vector;

[0114] Local feature extraction layer: 1D-TCN (4 layers), convolution kernel size 3, dilation coefficient 1→2→4→8, activation function ReLU, output 64-dimensional local features. Global correlation capture layer: Transformer encoder (2 layers), multi-head attention (8 heads), attention dropout=0.1, outputting 64-dimensional global features. Feature interaction layer: residual connections + layer normalization Output 64-dimensional interactive features; For layer normalization function; This is the 64-dimensional interactive feature vector output by the feature interaction layer; It is a 64-dimensional cross-modal fusion feature vector;

[0115] Output layer: 2 parallel fully connected branches: Defect type branch: Outputs the probability distribution of 5 types of defects. (Softmax activation); Severity branch: Outputs regression values ​​S from level 1 to 5, calculated as follows: ( For the sigmoid function, (for network output), ensuring S∈[1,5]; This represents the severity value in the final output.

[0116] In this embodiment, a total loss function is constructed based on classification loss, regression loss, and physical consistency loss. A two-stage training strategy is used to train the TCN-Transformer hybrid core network. In the first stage, the Transformer layers are frozen using the original training set to train the TCN layer and the output layer. In the second stage, the dataset is expanded. After thawing all layers and fine-tuning, an initial defect recognition model is obtained when the training results meet the preset convergence conditions, including: classification loss. : ;in, One-hot encoding for defect types; The probability value of defect type i predicted by the model; regression loss. : ;in, S represents the severity of the labeling, and S represents the model's predicted value; Physical consistency loss: Training dataset , , , , The defects are classified into five categories: sediment, uneven concrete, necking, mud inclusion, and voids. This represents the percentage of sound velocity anomalies at this depth. This represents the energy percentage of frequency band 3. For the sound speed gradient, The amplitude attenuation rate, Waveform distortion degree; for The probability of deep sediment defects; for The probability of unevenness defects in deep concrete; for The probability of a deep necking defect; for The probability of deep mud inclusion defects; for The probability of deep void defects;

[0117] ; , , These are the weighting coefficients.

[0118] In this embodiment, an adaptive dynamic decision-making module is constructed, which generates probability confidence thresholds, physical constraint distance thresholds, and time series consistency thresholds based on a normal sample metric library through kernel density estimation, including:

[0119] The normal sample measurement library includes the probability confidence level, physical constraint distance, and temporal consistency data of defect-free cast-in-place piles. The probability confidence level is the maximum value of the probability of the defect type. The physical constraint distance is the weighted distance of sound speed, amplitude, and frequency. The temporal consistency is the consistency ratio of the defect type in three repeated detections.

[0120] Dynamic thresholds are generated using kernel density estimation (KDE): A metric library of normal samples from the training dataset is constructed. Training dataset: 1000 defect-free cast-in-place piles were collected. , , Data; Threshold calculation: for Zhong Zhong The 95th percentile; for middle The 95th percentile; =0.8; Threshold update: Update using KDE every 100 newly added cast-in-place pile test data. , Defect judgment: For each depth z in the longitudinal direction, if the following conditions are met simultaneously... > , > , > Then z is marked as the defect depth; the defect type is taken as... corresponding Severity quantification: Outputs a level by combining S-value with the degree of abnormality of ultrasonic parameters; Result visualization: Generates a three-dimensional heat map of depth, defect type, and severity.

[0121] The working principle and beneficial effects of the above technical solution are as follows: By adding depth labels and defect pre-labels to the ultrasonic training dataset, a labeled feature set is formed, giving the data more explicit attribute information. This helps the subsequent model learn the correlation between different depth locations and defect features, improving the model's ability to identify defects at different depths in cast-in-place piles. The training dataset samples are concatenated with augmented samples according to depth segmentation rules to form an expanded dataset. This method fully integrates the advantages of the original and augmented data, allowing the model to access more comprehensive data during training, further improving the model's performance and stability. Multimodal cross-scale feature extraction is performed on the expanded dataset to obtain statistical feature vectors and frequencies. Feature vectors and waveform feature vectors; different types of feature vectors describe the characteristics of ultrasonic testing data of cast-in-place piles from multiple perspectives, providing richer information for subsequent defect identification and helping to improve the accuracy and reliability of identification; a cross-attention fusion network is constructed to map feature vectors of different modalities to the same dimension and perform weighted fusion to obtain cross-modal fusion feature vectors; an adaptive dynamic decision module is constructed to generate probability confidence thresholds, physical constraint distance thresholds, and time series consistency thresholds based on a normal sample metric library; these thresholds can be dynamically adjusted according to the distribution and characteristics of actual data, making the model more flexible and accurate in defect identification and able to adapt to changes in different testing environments and cast-in-place pile types.

[0122] It also includes supplementing the training of the intelligent analytical defect identification model for cast-in-place piles to obtain a supplemented training intelligent analytical defect identification model for cast-in-place piles;

[0123] The intelligent analytical defect identification model for cast-in-place piles is further trained to obtain a supplemented intelligent analytical defect identification model for cast-in-place piles, including:

[0124] A supplementary data system is constructed, which includes core supplementary data and auxiliary supplementary data. The core supplementary data includes manual defect assessment records, non-destructive testing raw data, full data of the construction process, and historical defect verification cases. The auxiliary supplementary data includes industry standards and specifications data.

[0125] The supplementary data is subjected to a four-dimensional progressive filtering process to obtain a high-quality supplementary dataset;

[0126] Based on the four-dimensional index, the diagnostic fit score is calculated, and the high-quality supplementary dataset is classified into high-fit dataset, medium-fit dataset and low-fit dataset.

[0127] The high-fit, medium-fit, and low-fit datasets are divided into different training batches according to a preset ratio. The intelligent analytical defect identification model for cast-in-place piles is dynamically batch-supplemented to obtain the supplemented intelligent analytical defect identification model for cast-in-place piles.

[0128] In this embodiment, the four-dimensional progressive filtering of supplementary data includes:

[0129] First-dimensional timeliness-completeness screening: Remove overdue data and data with a core field missing rate >20%; Standardization processing: Normalize non-destructive testing signals to the [-1,1] interval, and convert them into a 256×256 two-dimensional time-frequency graph through short-time Fourier transform; Construction parameters are normalized to the [0,10] scoring space according to industry standard thresholds; Defect annotations are uniformly mapped to the training dataset defect type + severity training dataset label system;

[0130] The second dimension of working condition adaptability screening: construct the benchmark working condition feature vector (including pile diameter, geological layer, and concrete grade), calculate the cosine similarity between the working condition feature vector of the supplementary data and the benchmark vector, and the similarity threshold = 0.8-0.001 × total amount of supplementary data (dynamically adjusted). Data with similarity below the threshold are removed.

[0131] Third-dimensional reliability verification: Actively submitted data is used to verify the quality of data content (e.g., signal-to-noise ratio > 20dB), and passively collected data is used to verify the compliance of sensor calibration records and random inspections, eliminating unqualified data.

[0132] Fourth dimension feature effectiveness screening: extract core features of defects in cast-in-place piles (sound velocity gradient, amplitude attenuation rate, frequency band energy distribution), evaluate feature completeness (missing rate <10%) and discriminative power (between-class variance / within-class variance >2), and remove substandard data; the remaining data are divided into training set, validation set and test set in a 7:2:1 ratio to form a high-quality supplementary dataset.

[0133] In this embodiment, the calculation of the diagnostic fit score based on four-dimensional indicators includes:

[0134] For manual defect assessment records: if it is actively submitted, S1 = institutional credit score × 0.4 + working condition similarity × 0.3 + feature discrimination × 0.3; if it is passively collected, S1 = normalized value of personal experience × 0.4 + working condition similarity × 0.3 + feature discrimination × 0.3; all variables are normalized to [0,1].

[0135] For non-destructive testing data: S2 = signal quality score × 0.5 + operating condition similarity × 0.2 + feature completeness × 0.3 (signal quality score is based on signal-to-noise ratio and waveform integrity).

[0136] For construction process data: S3 = Parameter deviation compliance rate × 0.4 + Working condition similarity × 0.3 + Data integrity × 0.3;

[0137] For historical defect cases: S4 = Case verification credibility × 0.5 + Operating condition similarity × 0.3 + Feature detail × 0.2;

[0138] The classification criteria are as follows: S≥0.85 is a highly adapted dataset (accounting for 55%); 0.6<S<0.85 is a medium-adapted dataset (accounting for 35%); S≤0.6 is a poorly adapted dataset, which retains only scarce samples under special working conditions (such as karst geology) and has a training contribution value >0.2 (training contribution value = gradient magnitude of the sample / average gradient magnitude of the batch), accounting for 10%.

[0139] The dynamic batch supplementary training includes:

[0140] Training order: first strengthen batches with scarce samples → generalization training batches → core training batches, to avoid high-fit data dominating the model;

[0141] Rare sample reinforcement batch: using a low-fit dataset, batch size 16, 20 iterations, and learning rate 0.0001;

[0142] Generalization training batches: using the medium-adapted dataset, batch size 32, 30 iterations, learning rate 0.0005;

[0143] Core training batches: using a highly adaptable dataset, batch size 32, 50 iterations, and a learning rate of 0.001;

[0144] Training mechanism: Real-time feedback adjustment: Calculate the F1 score on the validation set in each round. If the score decreases by more than 0.01 for three consecutive rounds, reduce the learning rate by 50%; Overfit suppression: Early stopping mechanism.

[0145] The working principle and beneficial effects of the above technical solution are as follows: The constructed supplementary data system includes core supplementary data and auxiliary supplementary data; the manual defect assessment records in the core supplementary data contain the experience judgment of professionals, the original non-destructive testing data is the most direct data reflecting the condition of the cast-in-place piles, the full data of the construction process can reflect the information of the entire process of the cast-in-place piles from construction to completion, and the historical defect verification cases provide confirmed defect samples; the industry standards and specifications data in the auxiliary supplementary data provide authoritative reference for the model; these rich data types provide the model with more comprehensive information from multiple dimensions, which helps the model learn more complex and accurate defect characteristics; comprehensively considering the test data, construction data, historical cases, and... Information from industry standards and other sources enables the model to not only identify defects from the detection data itself, but also to make more accurate judgments by combining the construction process and industry specifications. A four-dimensional progressive screening process for supplementary data removes noisy, erroneous, and redundant data, resulting in a high-quality supplementary dataset. A diagnostic fit score is calculated based on four-dimensional indicators, and the high-quality supplementary dataset is categorized into high-fit, medium-fit, and low-fit datasets. This categorization method classifies data according to its fit with the model's diagnostic task, allowing the model to utilize data of different fits more effectively during training, improving training efficiency and effectiveness. Through dynamic batch supplementary training, the model can continuously learn new data and knowledge, continuously optimizing its performance. With the addition of new supplementary data, the model can adapt to new situations and problems in the field of cast-in-place pile inspection, improve its ability and accuracy in identifying various defects, and thus better meet the needs of actual engineering projects. The intelligent analytical defect identification model for cast-in-place piles, after supplementary training, can more accurately identify the types and severity of defects in cast-in-place piles due to the use of richer and higher-quality data for training. This is of great significance for the quality assessment of cast-in-place piles and engineering safety, as it can promptly detect potential safety hazards and avoid engineering accidents caused by defects not being detected in time.

[0146] It should be noted that in this document, relational terms such as "first" and "second" are used only 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 terms "training dataset includes training dataset," "training dataset contains training dataset," or any other variations thereof are 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.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent analysis of ultrasonic testing data of cast-in-place piles based on deep learning, characterized in that, include: Ultrasonic testing data of cast-in-place piles is obtained using an ultrasonic testing instrument, and the ultrasonic testing data of cast-in-place piles is preprocessed and feature extracted based on natural language processing to determine the ultrasonic testing feature data of cast-in-place piles. An intelligent analysis and defect identification model for cast-in-place piles is constructed based on artificial intelligence. The ultrasonic testing feature data of the cast-in-place piles are analyzed and identified based on the intelligent analysis and defect identification model to determine the intelligent analysis and defect identification results of the cast-in-place piles. The cast-in-place piles are then evaluated and managed based on the intelligent analysis and defect identification results.

2. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 1, characterized in that, Ultrasonic testing data of cast-in-place piles is obtained using an ultrasonic testing instrument, including: The ultrasonic testing instrument is used to measure and test the cast-in-place piles and record the sound velocity, amplitude, frequency and waveform parameters of the sound waves to obtain ultrasonic testing data of the cast-in-place piles. Among them, the speed of sound is the speed at which sound waves propagate in concrete, and is used to reflect the elastic modulus and density of concrete. Amplitude is the peak amplitude of the first cycle of the received ultrasonic signal, used to reflect the energy attenuation of ultrasonic waves during propagation; The frequency is the main frequency or frequency component of the received ultrasonic signal; the waveform is the voltage and time signal curve received over the entire time domain.

3. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 2, characterized in that, Preprocessing of ultrasonic testing data for cast-in-place piles based on natural language processing involves the following steps: Denoising is performed on ultrasonic testing data of cast-in-place piles based on wavelet denoising to improve the signal-to-noise ratio of the ultrasonic testing data of cast-in-place piles. Specifically, wavelets with similar shapes to ultrasonic waves are selected, and the original waveform at each depth point is decomposed into N layers of wavelets to obtain approximation coefficients and detail coefficients. The detail coefficients are processed by applying a threshold function, and the signal is reconstructed using the processed coefficients to obtain the denoised waveform. The ultrasonic testing data of cast-in-place piles is examined based on linear traversal to identify missing and outlier values ​​in the data. The missing and outlier values ​​in the data are then evaluated to determine whether they are valuable for intelligent analysis and defect identification of cast-in-place piles. When missing and outlier values ​​in the ultrasonic testing data of cast-in-place piles are valuable for intelligent analysis and defect identification of cast-in-place piles, interpolation is used to fill in the missing values ​​and the mean value is used to replace the outlier values. When missing or outlier values ​​in the ultrasonic testing data of cast-in-place piles are of no value for intelligent analysis and defect identification of cast-in-place piles, then the missing or outlier values ​​in the ultrasonic testing data of cast-in-place piles are removed.

4. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 3, characterized in that, Based on natural language processing, feature extraction is performed on the ultrasonic testing data of cast-in-place piles, and the following operations are performed: Based on Z-Score standardization, the ultrasonic testing data of cast-in-place piles is normalized, which transforms the dimensional ultrasonic testing data of cast-in-place piles into dimensionless data expression, removes the dimensional differences between the ultrasonic testing data of cast-in-place piles, and forms standardized ultrasonic testing data of cast-in-place piles. Enhancement processing is performed on the ultrasonic testing data of cast-in-place piles. Specifically, for the images in the ultrasonic testing data of cast-in-place piles, rotation, flipping, cropping, adding Gaussian noise, and adjusting brightness and contrast are performed. For the sequences in the ultrasonic testing data of cast-in-place piles, noise is added and small-amplitude scaling transformations are performed. Feature extraction is performed on the ultrasonic testing data of cast-in-place piles. Features related to intelligent analysis and defect identification of cast-in-place piles are extracted from the ultrasonic testing data of cast-in-place piles to determine the ultrasonic testing feature data of cast-in-place piles.

5. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 4, characterized in that, Based on artificial intelligence, an intelligent analytical defect identification model for cast-in-place piles is constructed, and the following operations are performed: Based on the intelligent analysis requirements of ultrasonic testing data of cast-in-place piles, historical data on defect identification of cast-in-place piles are collected and divided into training set and test set. The deep learning model was trained based on the training set to determine the intelligent analytical defect identification model for cast-in-place piles. The intelligent analytical defect identification model for cast-in-place piles was tested using a test set to evaluate its performance, determine the model test evaluation results, and adjust and optimize the model based on these results to determine the optimal intelligent analytical defect identification model for cast-in-place piles.

6. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 5, characterized in that, The intelligent analytical defect identification model for cast-in-place piles was tested using a test set to evaluate its performance. The following steps were performed: The test set is input into the intelligent analytical defect identification model for cast-in-place piles. The performance of the intelligent analytical defect identification model for cast-in-place piles is tested based on the test set. The model output results are used to determine whether the intelligent analytical defect identification model for cast-in-place piles can achieve the expected effect of automatically identifying defects in cast-in-place piles. When the intelligent analysis and defect identification model for cast-in-place piles fails to achieve the expected effect of automatically identifying defects in cast-in-place piles, the parameters of the intelligent analysis and defect identification model for cast-in-place piles are adjusted and optimized. The performance of the intelligent analysis and defect identification model for cast-in-place piles after parameter adjustment and optimization is tested again, forming a closed-loop test and optimization of the intelligent analysis and defect identification model for cast-in-place piles, until the intelligent analysis and defect identification model for cast-in-place piles can achieve the expected effect of automatically identifying defects in cast-in-place piles, and the optimal intelligent analysis and defect identification model for cast-in-place piles is determined.

7. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 6, characterized in that, Based on the intelligent analytical defect identification model for cast-in-place piles, the ultrasonic testing feature data of the cast-in-place piles are analyzed and identified, and the following operations are performed: The ultrasonic testing feature data of the cast-in-place pile is input into the intelligent analysis and defect identification model of the cast-in-place pile. The model analyzes and identifies the ultrasonic testing feature data of the cast-in-place pile, automatically identifies defects in the cast-in-place pile, determines the intelligent analysis and defect identification results of the cast-in-place pile, and displays the intelligent analysis and defect identification results of the cast-in-place pile in a visual form.

8. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 7, characterized in that, Based on the intelligent analysis and defect identification results of the cast-in-place piles, the cast-in-place piles are evaluated and managed, and the following operations are performed: The integrity level of the cast-in-place pile is assessed based on the intelligent analysis and defect identification results, including Class I, Class II, Class III and Class IV piles. The corresponding suggestions are automatically selected from a predefined document library according to the integrity level of the pile, and a structured cast-in-place pile inspection report is generated. For Class I piles, the pile body structure is intact and there are no obvious defects; For Class II piles, the pile body has minor defects that do not affect normal use, but these need to be addressed during subsequent construction. For Class III piles, there are obvious defects in the pile body. It is necessary to analyze the impact on the pile bearing capacity together with the design and supervision units and determine the treatment plan. For Class IV piles, which have serious defects, core drilling is required for verification, and reinforcement or scrapping is necessary.

9. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 5, characterized in that, The ultrasonic testing feature data of cast-in-place piles are analyzed and identified based on the intelligent analytical defect identification model for cast-in-place piles, including: Obtain an ultrasonic training dataset for cast-in-place piles; the ultrasonic training dataset includes sound velocity, amplitude, frequency, and waveform parameters; add depth labels and defect pre-labels to the ultrasonic training dataset to form a labeled feature set; The labeled feature set is generated based on the Physically Constrained Variational Autoencoder (PC-VAE). The PC-VAE includes an input layer, an encoder, a physical constraint layer, and a decoder. The physical constraint layer adds hard constraints on sound speed and density and soft constraints on feature distribution corresponding to different depths to generate enhanced samples that conform to the physical laws of ultrasonic propagation. The enhanced samples in the training dataset are spliced ​​together with the enhanced samples according to the depth segmentation rules to form an expanded dataset. Multimodal cross-scale feature extraction is performed on the expanded dataset to obtain statistical feature vectors, frequency feature vectors, and waveform feature vectors; A cross-attention fusion network (CAFN) is constructed, which maps the statistical feature vector, frequency feature vector, and waveform feature vector to the same dimension and constructs a feature matrix. The attention weights of each modal feature are calculated through cross-attention. Based on the attention weights, the modal features are weighted and fused to obtain a cross-modal fusion feature vector. A TCN-Transformer hybrid core network is constructed, comprising an input layer, a local feature extraction layer, a global association capture layer, a feature interaction layer, and an output layer. The cross-modal fused feature vector and the depth label are concatenated and input into the hybrid core network. The local feature extraction layer captures depth local defect features through a 4-layer 1D-TCN, the global association capture layer captures cross-depth defect associations through a 2-layer Transformer encoder, the feature interaction layer fuses local and global features through residual connections and layer normalization, and the output layer outputs the defect type probability distribution and severity regression value through two parallel fully connected branches. The total loss function is constructed based on classification loss, regression loss and physical consistency loss. The TCN-Transformer hybrid core network is trained using a two-stage training strategy. In the first stage, the Transformer layer is frozen using the original training set to train the TCN layer and the output layer. In the second stage, all layers are unfrozen on the expanded dataset for fine-tuning. When the training results meet the preset convergence conditions, the initial defect recognition model is obtained. An adaptive dynamic decision-making module is constructed, which generates probability confidence thresholds, physical constraint distance thresholds, and time series consistency thresholds based on a normal sample metric library through kernel density estimation. The initial defect identification model is combined with the adaptive dynamic decision-making module to obtain a trained intelligent analytical defect identification model for cast-in-place piles.

10. The intelligent analysis method for ultrasonic testing data of cast-in-place piles based on deep learning according to claim 5, characterized in that, It also includes supplementing the training of the intelligent analytical defect identification model for cast-in-place piles to obtain a supplemented training intelligent analytical defect identification model for cast-in-place piles; The intelligent analytical defect identification model for cast-in-place piles is further trained to obtain a supplemented intelligent analytical defect identification model for cast-in-place piles, including: A supplementary data system is constructed, which includes core supplementary data and auxiliary supplementary data. The core supplementary data includes manual defect assessment records, non-destructive testing raw data, full data of the construction process, and historical defect verification cases. The auxiliary supplementary data includes industry standards and specifications data. The supplementary data is subjected to a four-dimensional progressive filtering process to obtain a high-quality supplementary dataset; Based on the four-dimensional index, the diagnostic fit score is calculated, and the high-quality supplementary dataset is classified into high-fit dataset, medium-fit dataset and low-fit dataset. The high-fit, medium-fit, and low-fit datasets are divided into different training batches according to a preset ratio. The intelligent analytical defect identification model for cast-in-place piles is dynamically batch-supplemented to obtain the supplemented intelligent analytical defect identification model for cast-in-place piles.

Images