Pile foundation defect three-dimensional imaging method based on deep learning
By using multi-angle, dense acoustic data acquisition and deep learning model inversion in pile foundation detection, a three-dimensional visualization model of pile foundation defects is generated, which solves the problem of inaccurate pile foundation defect location in traditional methods and achieves high-precision three-dimensional imaging.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies cannot achieve high-resolution three-dimensional imaging of pile foundation defect areas, leading to uncertainty and risk in detection results. Traditional methods are difficult to accurately locate and quantify internal defects in pile foundations.
A transducer array containing multiple ultrasonic transmitting and receiving crystals arranged at preset intervals is used. Combined with a deep learning model, dense acoustic data is acquired through multi-angle, sequential scanning. A three-dimensional probability matrix is generated by inversion and quantization is performed by applying a probability threshold to generate a three-dimensional visualization model of the defect.
It achieves ultra-high resolution three-dimensional imaging of the defect area of pile foundation, improves the accuracy of detection and the reliability of engineering decisions, and can intuitively and quantitatively show the true shape, size and distribution of defects.
Smart Images
Figure CN121962489A_ABST
Abstract
Description
A Deep Learning-Based Three-Dimensional Imaging Method for Pile Foundation Defects Technical Field
[0001] This application belongs to the field of pile foundation defect technology, and more specifically, relates to a three-dimensional imaging method for pile foundation defects based on deep learning. Background Technology
[0002] Non-destructive testing (NDT) is a safety guarantee for modern industrial systems. It involves using anomalies in the internal structure of materials to determine, evaluate, and locate macroscopic defects, mechanical properties, and microstructure within or on the surface of the object being tested, without damaging or affecting its performance. In the field of civil engineering, particularly in the pile foundation engineering of infrastructure such as bridges, buildings, and dams, NDT is a crucial link in ensuring project quality and assessing structural safety, yielding significant economic and social benefits.
[0003] Currently, the concrete pile foundation integrity testing industry generally uses the low-strain reflection method for rapid general surveys and relies on cross-aperture ultrasonic transmission and finite element method for precise diagnosis by layering concrete piles. However, these technologies are no longer sufficient to meet today's construction needs: the low-strain method struggles to accurately locate and quantify internal defects in pile foundations; while ultrasonic transmission can effectively locate abnormal areas, it relies too heavily on the original detection path, resulting in significant blind spots and ultimately relying heavily on the experience of the testing personnel for judgment. Even the current method of layering and displaying sound velocity changes to form a three-dimensional pile structure can only deduce simple sound velocity anomalies, not the specific location of defects. This leads to current detection cloud maps providing only approximate two-dimensional anomaly maps or incomplete three-dimensional cloud maps showing defects, failing to intuitively display the true shape, size, and distribution of defects in three dimensions, introducing uncertainty and risk into safety assessments and decisions.
[0004] Therefore, how to achieve high-resolution three-dimensional imaging of the defect area of pile foundation is a pressing problem that urgently needs to be studied. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a deep learning-based three-dimensional imaging method for pile foundation defects, which can achieve ultra-high resolution three-dimensional imaging of key areas of pile foundation defects, greatly improving the accuracy of detection and the reliability of decision-making.
[0006] To achieve the above objectives, in a first aspect, this application provides a deep learning-based three-dimensional imaging method for pile foundation defects, comprising the following steps: S10, acquiring acoustic data collected by multi-angle, sequential scanning of a suspected defect area of the pile foundation using a transducer array deployed in the pile foundation acoustic logging tube; the transducer array includes multiple ultrasonic transmitting crystals and ultrasonic receiving crystals arranged opposite each other, with identical crystals arranged at a preset spacing, and the acoustic data includes a sound velocity matrix and amplitude matrix densely distributed in three-dimensional space; S20, preprocessing the acoustic data; S30, inputting the preprocessed acoustic data into a pre-trained deep learning model, the deep learning model being trained using a dataset consisting of concrete models containing defects of different shapes, sizes, and locations, generated by finite element numerical simulation, and their corresponding acoustic data, to invert a three-dimensional probability matrix characterizing the probability of defects existing at each location within the suspected defect area through the deep learning model; S40, processing the three-dimensional probability matrix based on a preset probability threshold to generate and output a three-dimensional visualization model of the defects within the suspected defect area.
[0007] The deep learning-based three-dimensional imaging method for pile foundation defects provided in this application has the following advantages: First, the method employs a transducer array containing multiple transmitting and receiving crystals arranged at preset intervals, and executes a multi-angle, sequential scanning strategy on the suspected defect area. This strategy can acquire dense acoustic ray network data, namely the sound velocity matrix and amplitude matrix, which are highly intersecting and comprehensively covered in three-dimensional space. This data acquisition method fundamentally overcomes the limitations of traditional acoustic transmission methods, such as limited detection paths and blind spots, providing sufficient data conditions for generating high-resolution images. Next, instead of relying on simplified physical models or empirical formulas to interpret the dense data, this method inputs it into a deep learning model pre-trained using a large amount of numerical simulation of defects and their acoustic response data. This model can directly learn from high-dimensional, complex acoustic parameters and establish a complex mapping relationship between them and the three-dimensional spatial structure inside the concrete, thereby achieving a three-dimensional inversion of the defect probability distribution. Finally, by applying a preset threshold to quantize the obtained three-dimensional probability matrix, a three-dimensional visual entity model of the defect can be directly generated. This complete process, from intensive data acquisition to AI-driven data inversion and then to 3D visualization, works synergistically to enable ultra-high resolution 3D imaging of key areas of pile foundation defects. It can intuitively and quantitatively show the true shape, size and distribution of defects, thereby greatly improving the accuracy of detection results and the reliability of engineering decisions.
[0008] As a further preferred embodiment, in step S10, the sequential scanning includes a horizontal scanning mode and an oblique scanning mode corresponding to the receiving point within a 30° vertical range from the emission point, as shown in the specification; wherein, the horizontal scanning mode is: an ultrasonic transmitting chip emits a signal, and an ultrasonic receiving chip at a directly opposite position receives the signal; the oblique scanning mode is: an ultrasonic transmitting chip emits a signal, and an ultrasonic receiving chip at a non-directly opposite position within a 30° vertical range receives the signal.
[0009] As a further preferred embodiment, in step S10, the sequential scanning executes the horizontal measurement mode and the oblique measurement mode in turn, and the oblique measurement range is within a vertical range of 30° above and below; the oblique measurement mode is as follows: an ultrasonic transmitting chip transmits a signal, and the receiving chips above and below the receiving chip corresponding to the horizontal measurement of the transmitting chip receive the signal, and the receiving chips receive the signal in sequence at a vertical angle of 30° above and below the receiving chip.
[0010] As a further preferred embodiment, the sequential scanning is performed cyclically in the order of horizontal and oblique measurement modes until all preset transmission and reception combinations are completed.
[0011] As a further preferred embodiment, step S10 specifically comprises: S11, scanning the pile foundation using ultrasonic transmission method, and preliminarily determining the suspected defect area within the pile foundation based on the scanning data; S12, deploying the transducer array in the acoustic logging tube corresponding to the suspected defect area; S13, controlling the transducer array to scan the suspected defect area according to a preset transmission and reception sequence, and recording the sound velocity matrix and amplitude matrix.
[0012] As a further preferred embodiment, in step S13, the transmission and reception sequence is such that the final acoustic measurement line covers the central region of the transducer array, and the acoustic measurement line data corresponding to the upper and lower ends of the transducer array are discarded; the generated three-dimensional visualization model corresponds to the central region.
[0013] As a further preferred embodiment, in step S30, the deep learning model is a convolutional neural network, which extracts, combines, and spatially reconstructs features from the input acoustic data through multiple convolutional layers and upsampling layers.
[0014] As a further preferred embodiment, in step S40, the preset probability threshold is a value determined after the deep learning model has been trained; blocks with probability values greater than or equal to the threshold in the three-dimensional probability matrix are rendered as defects to generate the three-dimensional visualization model.
[0015] As a further preferred option, in step S40, the output results also include an inspection report containing a three-dimensional defect cloud map, a defect parameter table, and a quality assessment report.
[0016] Secondly, this application provides a deep learning-based three-dimensional imaging system for pile foundation defects, used to implement the method described in any one of the above descriptions, comprising: a data acquisition module for acquiring acoustic data collected by scanning the suspected defect area of the pile foundation through a transducer array; a preprocessing module for preprocessing the acoustic data; an inversion module including a pre-trained deep learning model for receiving the preprocessed acoustic data and inverting it to obtain a three-dimensional probability matrix; and a model generation module for processing the three-dimensional probability matrix based on a preset probability threshold to generate and output a three-dimensional visualization model of the defect.
[0017] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description
[0018] Figure 1 is a flowchart of the three-dimensional imaging method for pile foundation defects based on deep learning provided in this application; Figure 2 is a flowchart of the three-dimensional imaging method for pile foundation defects based on deep learning provided in an embodiment of this application; Figure 3 is a single-section acoustic logging line diagram provided in an embodiment of this application, wherein (a) is a schematic diagram of the planar measurement mode, (b) is a schematic diagram of the first oblique measurement mode, (c) is a schematic diagram of the second oblique measurement mode, (d) is a schematic diagram of the third oblique measurement mode, (e) is a schematic diagram of the fourth oblique measurement mode, and (f) is a schematic diagram of all acoustic logging lines; Figure 4 is a cross-sectional view of the pile foundation provided in an embodiment of this application, wherein (a) is a schematic diagram of the cross-sectional view of the arrangement of three acoustic logging pipes, and (b) is a schematic diagram of the cross-sectional view of the arrangement of four acoustic logging pipes. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0020] It should be understood that, in the description of this application, the term "multiple" means two or more, unless otherwise expressly and specifically defined.
[0021] As shown in Figure 1, this application provides a three-dimensional imaging method for pile foundation defects based on deep learning, including steps S10 to S40, which are detailed below: Step S10, acquiring acoustic data collected by multi-angle and sequential scanning of the suspected defect area of the pile foundation through a transducer array deployed in the pile foundation acoustic tube. The acoustic data includes a sound velocity matrix and amplitude matrix densely distributed in three-dimensional space.
[0022] The transducer array provided in this application includes multiple corresponding ultrasonic transmitting crystals and ultrasonic receiving crystals, with identical crystals arranged at a preset spacing. This step, by deploying a transducer array with a specific layout and detecting the initially identified suspected defect areas according to a preset sequential scanning method, can obtain a raw data matrix of sound velocity and amplitude that is much denser in three-dimensional space than conventional methods, providing the necessary data foundation for subsequent high-precision inversion.
[0023] Step S20: Preprocess the acoustic data.
[0024] This step involves preprocessing the raw acoustic data obtained in step S10, such as normalization, to standardize the data format, which is beneficial for the stable and efficient processing of subsequent deep learning models.
[0025] Step S30: Input the preprocessed acoustic data into the pre-trained deep learning model. The deep learning model is trained using a dataset consisting of concrete models with defects of different shapes, sizes and locations generated by finite element numerical simulation and their corresponding acoustic data. The deep learning model is used to invert and obtain a three-dimensional probability matrix to characterize the probability of defects at each location in the suspected defect area.
[0026] This step utilizes a pre-trained deep learning model to process the pre-processed dense acoustic data. This model is trained based on the correspondence between "defect model and acoustic parameters" from a large number of numerical simulations. It has the ability to learn the nonlinear mapping from complex acoustic response to internal structure, and can automatically and quickly invert the probability distribution of defects at various locations within the target area, forming a three-dimensional probability matrix.
[0027] Step S40: Process the three-dimensional probability matrix based on a preset probability threshold to generate and output a three-dimensional visualization model of the defect in the suspected defect area.
[0028] This step involves binarizing the three-dimensional probability matrix by setting a clear probability threshold, identifying blocks with probabilities higher than the threshold as defects, thereby directly generating a three-dimensional solid model that intuitively displays the shape, size, and spatial distribution of defects, and outputting a relevant inspection report.
[0029] The deep learning-based three-dimensional imaging method for pile foundation defects provided in this application has the following advantages: First, the method employs a transducer array containing multiple transmitting and receiving crystals arranged at preset intervals, and executes a multi-angle, sequential scanning strategy on the suspected defect area. This strategy can acquire dense acoustic ray network data, namely the sound velocity matrix and amplitude matrix, which are highly intersecting and comprehensively covered in three-dimensional space. This data acquisition method fundamentally overcomes the limitations of traditional acoustic transmission methods, such as limited detection paths and blind spots, providing sufficient data conditions for generating high-resolution images. Next, instead of relying on simplified physical models or empirical formulas to interpret the dense data, this method inputs it into a deep learning model pre-trained using a large amount of numerical simulation of defects and their acoustic response data. This model can directly learn from high-dimensional, complex acoustic parameters and establish a complex mapping relationship between them and the three-dimensional spatial structure inside the concrete, thereby achieving a three-dimensional inversion of the defect probability distribution. Finally, by applying a preset threshold to quantize the obtained three-dimensional probability matrix, a three-dimensional visual entity model of the defect can be directly generated. This complete process, from intensive data acquisition to AI-driven data inversion and then to 3D visualization, works synergistically to enable ultra-high resolution 3D imaging of key areas of pile foundation defects. It can intuitively and quantitatively show the true shape, size and distribution of defects, thereby greatly improving the accuracy of detection results and the reliability of engineering decisions.
[0030] In one embodiment, the technical solution to achieve the above objective can be as follows: As shown in Figure 2, this embodiment proposes a method for three-dimensional fine imaging of internal defects in concrete pile foundations that integrates dense acoustic scanning and artificial intelligence inversion.
[0031] The core of this embodiment lies in the fact that when conventional inspections detect a suspected abnormal area within the pile foundation (e.g., within a 1-meter pile length), the process does not stop at empirical qualitative judgment, but rather initiates a refined secondary diagnostic process. As shown in Figure 3 (how sequential emission is performed), this method utilizes a multi-angle, high-density, sequentially emitted ultrasonic scanning network to perform refined scanning and data extraction on the target area, obtaining a significantly larger amount of information on sound wave velocity and amplitude than conventional methods. This information is then used as input data to a deep neural network system pre-trained with extensive numerical simulations and experimental data. This network can learn the complex nonlinear mapping relationship from acoustic response to a three-dimensional model of the internal concrete structure, and ultimately quickly and automatically invert and generate a high-resolution three-dimensional solid model of the defects within the target area (approximately the middle 50 cm of the test section), greatly improving the accuracy of detection and the reliability of decision-making.
[0032] The following is a specific implementation example of this application: This embodiment takes the detection of a bored pile with a diameter of 1 meter and three pre-embedded sonic logging pipes (a total of 3 cross sections as shown in Figure 4(a), with effective data of 3×[10+2×(9+8+7+6)]=210) as an example (the same applies to four sonic logging pipes, and the arrangement of four sonic logging pipes should have 6 cross sections as shown in Figure 4(b), with effective data of 6×[2×(9+8+7+6)]=420), and the steps are as shown in Figure 1.
[0033] Step 1: Scan the entire pile using conventional ultrasonic transmission method to preliminarily determine the approximate range of defects within the pile.
[0034] First, a standard ultrasonic transduction method was used to conduct a general survey of the entire pile. Ordinary ultrasonic transmitters and receivers were placed in three acoustic logging tubes, and measurements were taken point-by-point. The testing instrument recorded the numerical changes at each measuring point. Data analysis revealed an anomaly within a one-meter range (multiple profiles showed sound velocity values below the normal 4000 m / s), preliminarily identifying this one-meter range as a suspected defect area.
[0035] Step 2: After determining the approximate range of the defect, deploy a new transducer array in the suspected area.
[0036] After identifying the suspected area, the transducer array involved in this embodiment is vertically placed into the acoustic logging tube. This array integrates 10 ultrasonic transmitting chips (A1~A10) and 10 receiving chips (B1~B10), with a designed distance of 10 millimeters between the chips. The array is controlled by a lifting mechanism, allowing it to precisely reach the suspected abnormal area.
[0037] Step 3: Arrange the transducer array, perform a refined and sequential scan of the suspected area, and record the relevant acoustic data.
[0038] After the transducers are set up, the transmission and reception of each profile are shown in Figure 3. Each transmission point and receiving point are arranged in a corresponding manner, and the transmission and reception sequence is as follows: First, the ultrasonic wave is tested in a horizontal mode (as shown in Figure 3(a)). As shown in Figure 3(a), if A1 is the transmission point, the corresponding receiving point is B1, and the same applies to the other transmission and receiving points. The next step is to start the oblique mode (as shown in Figure 3(b)-Figure 3(e)). In the first oblique mode, the transmission point transmits and the receiving point is the two adjacent horizontal points opposite it. As shown in Figure 3(b), if A2 transmits, its corresponding receiving points are B1 and B2, and the same applies to the other transmission and receiving points. Then, the second oblique mode is tested. One transmission point transmits and the receiving point is the receiving point that extends vertically from the opposite side of the first mode. As shown in Figure 3(c), if A3 is the transmission point, its corresponding receiving points are B1 and B5, and the same applies to the other transmission and receiving points. The latter two oblique measurement methods are similar to the oblique measurement described above. The acoustic data is transmitted and received sequentially according to the methods in Figure 3(a) to Figure 3(e), and the acoustic data is recorded. Finally, Figure 3(f) represents all the transmit and receive lines. Each acoustic measurement line represents the acoustic time and amplitude data. However, since the amount of acoustic data on both sides is small and may be inaccurate, this method discards the data of the upper and lower transducer areas (the blue acoustic measurement line data in Figure 2f) and only retains the data of the central area (the red acoustic measurement line data in Figure 3(f)). The detection range is changed from the original 1-meter detection area to the central area (the area represented by yellow in Figure 3(f)).
[0039] The scans are performed sequentially according to the order specified in Figure 2, acquiring the parameters of sound velocity and amplitude, and recording the corresponding positions as set by the program. Ultimately, a densely distributed sound velocity and amplitude matrix in three-dimensional space is obtained, far exceeding the quantity obtained by conventional methods.
[0040] Step 4: Preprocess the obtained data and then input it into the pre-trained neural network system for inversion calculation.
[0041] After preprocessing the collected acoustic data (such as normalization), it is input into a pre-trained deep learning model.
[0042] Training Model Overview: This model was trained before implementation of detection. The training data was generated using finite element numerical simulation, producing tens of thousands of concrete models with defects of different shapes, sizes, and locations. Acoustic data was then calculated to construct a dataset corresponding to the "defect model - acoustic parameters." This dataset was fed into a convolutional neural network, enabling efficient learning of the mapping relationship between acoustic data and the three-dimensional spatial structure.
[0043] After the trained neural network receives the actual test data from the field, it will extract, combine and spatially reconstruct the acoustic features through multiple convolutional layers and upsampling layers during forward propagation. Finally, a three-dimensional matrix will be directly generated at the output, where the value of each voxel represents the probability of its existence at that location.
[0044] Step 5: Based on the inversion results and specific threshold processing, generate a visualized 3D defect model and output an inspection report.
[0045] The computing device receives the network output and renders blocks above a certain threshold (such as 0.7) as defects, thereby generating a 3D visualization model and detection report (including a 3D defect cloud map, defect parameter table, quality assessment report, etc.) of the suspected abnormal area (the abnormal area is within a range of one meter, and the generated 3D model is only about 50 centimeters in the middle of the detection array).
[0046] The key point of this embodiment is that it combines routine general surveys with detailed local diagnosis, thereby improving detection efficiency and achieving ultra-high resolution three-dimensional imaging of key areas.
[0047] A novel, specific multi-transmitter, multi-receiver transducer array layout was adopted, and a highly intersecting, fully covered acoustic ray network was formed by multiple acoustic survey lines in the key area, providing a large amount of data support for the subsequent inversion of three-dimensional images.
[0048] Abandoning the traditional inversion method that relies on simplified physical models and empirical formulas, this method utilizes the powerful nonlinear mapping capabilities of deep learning neural networks to directly invert the 3D structure of defects in high-latitude data, solving the problem that traditional methods cannot directly generate complex 3D imaging of key areas.
[0049] The final output cloud map data is not a simple data list or two-dimensional data sample. It abandons the method of judging abnormal areas based on experience and directly generates a three-dimensional solid model that can be used for engineering analysis, so as to realize the visualization and quantification of the detected structure.
[0050] The effects of this embodiment are as follows: It eliminates detection blind spots and significantly enhances the detection capabilities for both small and complex-shaped defects. It directly generates three-dimensional defect imaging models, allowing for complete observation of defect size and location, providing reliable data support for safety assessments and acceptance. Relying on neural networks for inversion reduces the rigid requirements and reliance on observation personnel, improving detection accuracy and objectivity. This embodiment is an upgrade and improvement based on the existing ultrasonic transmission method, requiring no further modifications to the pile foundation (such as adding acoustic logging tubes or altering its original structure), making it convenient to implement, cost-effective, and easy to promote.
[0051] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A three-dimensional imaging method for pile foundation defects based on deep learning, characterized in that, The process includes the following steps: S10, acquiring acoustic data collected by multi-angle, sequential scanning of the suspected defect area of the pile foundation using a transducer array deployed in the pile foundation acoustic logging pipe; the transducer array includes multiple ultrasonic transmitting crystals and ultrasonic receiving crystals arranged opposite each other, with identical crystals arranged at a preset spacing; the acoustic data includes a sound velocity matrix and amplitude matrix densely distributed in three-dimensional space; S20, preprocessing the acoustic data; S30, inputting the preprocessed acoustic data into a pre-trained deep learning model; the deep learning model is trained using a dataset consisting of concrete models containing defects of different shapes, sizes, and locations, generated by finite element numerical simulation, and their corresponding acoustic data, to invert and obtain a three-dimensional probability matrix characterizing the probability of defects at each location within the suspected defect area; S40, processing the three-dimensional probability matrix based on a preset probability threshold to generate and output a three-dimensional visualization model of the defects within the suspected defect area.
2. The deep learning-based three-dimensional imaging method for pile foundation defects as described in claim 1, characterized in that, In step S10, the sequential scanning includes a horizontal scanning mode and an oblique scanning mode corresponding to the receiving point within a 30° vertical range from the emission point, as shown in the specification. The horizontal scanning mode is characterized by an ultrasonic transmitting chip emitting a signal, which is received by an ultrasonic receiving chip positioned directly opposite it. The oblique scanning mode is characterized by an ultrasonic transmitting chip emitting a signal, which is received by an ultrasonic receiving chip positioned within a 30° vertical range not directly opposite it.
3. The deep learning-based three-dimensional imaging method for pile foundation defects as described in claim 2, characterized in that, In step S10, the sequential scanning executes the horizontal measurement mode and the oblique measurement mode in turn. The oblique measurement range is within a vertical range of 30° above and below. The oblique measurement mode is as follows: an ultrasonic transmitting chip transmits a signal, and the two receiving chips adjacent to the horizontal receiving chip of the transmitting chip receive the signal in the opposite receiving chip. The receiving chips receive the signal in sequence at a vertical angle of 30° above and below the receiving chip.
4. The three-dimensional imaging method for pile foundation defects based on deep learning as described in claim 3, characterized in that, The sequential scanning is performed in a loop according to the order of flat measurement mode, first oblique measurement mode, and second oblique measurement mode until all preset transmission and reception combinations are completed.
5. The three-dimensional imaging method for pile foundation defects based on deep learning as described in claim 1, characterized in that, Step S10 specifically includes: S11, scanning the pile foundation using ultrasonic transmission method, and preliminarily determining the suspected defect area within the pile foundation based on the scanning data; S12, deploying the transducer array in the acoustic logging tube corresponding to the suspected defect area; S13, controlling the transducer array to scan the suspected defect area according to a preset transmission and reception sequence, and recording the sound velocity matrix and amplitude matrix.
6. The deep learning-based three-dimensional imaging method for pile foundation defects as described in claim 5, characterized in that, In step S13, the transmission and reception sequence ensures that the final acoustic measurement line covers the central region of the transducer array, and the acoustic measurement line data corresponding to the upper and lower ends of the transducer array are discarded; the generated three-dimensional visualization model corresponds to the central region.
7. The deep learning-based three-dimensional imaging method for pile foundation defects as described in claim 1, characterized in that, In step S30, the deep learning model is a convolutional neural network, which extracts, combines, and spatially reconstructs features from the input acoustic data through multiple convolutional layers and upsampling layers.
8. The three-dimensional imaging method for pile foundation defects based on deep learning as described in claim 1, characterized in that, In step S40, the preset probability threshold is a value determined after the deep learning model has been trained; blocks with probability values greater than or equal to the threshold in the three-dimensional probability matrix are rendered as defects to generate the three-dimensional visualization model.
9. The three-dimensional imaging method for pile foundation defects based on deep learning as described in claim 1, characterized in that, In step S40, the output also includes an inspection report containing a three-dimensional defect cloud map, a defect parameter table, and a quality assessment report.
10. A three-dimensional imaging system for pile foundation defects based on deep learning, characterized in that, The method for implementing any one of claims 1 to 9 comprises: a data acquisition module for acquiring acoustic data collected by scanning a suspected defect area of a pile foundation through a transducer array; a preprocessing module for preprocessing the acoustic data; an inversion module comprising a pre-trained deep learning model for receiving the preprocessed acoustic data and inverting it to obtain a three-dimensional probability matrix; and a model generation module for processing the three-dimensional probability matrix based on a preset probability threshold to generate and output a three-dimensional visualization model of the defect.