Ultrasonic array based imaging method for internal defects detection of concrete structures

Abstract

This invention discloses an imaging detection method for internal defects in concrete structures based on an ultrasonic array, comprising: arranging an ultrasonic array, receiving echoes in multiple rounds according to a preset sequence to acquire multi-channel data; time-aligning the echo data, extracting amplitude and phase consistency features, and constructing a scattering potential energy distribution; calculating the scattering gradient field, clustering along negative gradient streamlines, and determining candidate defect regions; inputting the amplitude and phase matrices into a shared NAFNet, frequency-domain gating and potential energy modulation, and outputting an enhanced map; extracting the center-to-time of sparse wavefront clustering, performing curvature canonical inversion, and constructing a sound velocity curvature field; generating propagation paths and folding and compressing them to construct reachable domain time-delay superimposed imaging. This invention achieves high-resolution imaging detection of internal defects in concrete structures by constructing a scattering potential energy distribution and combining it with an improved NAFNet network to enhance echo features and sound velocity-curvature coupled propagation modeling.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic nondestructive testing technology, and in particular to an imaging detection method for internal defects in concrete structures based on ultrasonic arrays. Background Technology

[0002] Concrete structures are widely used in bridges, tunnels, dams, and building construction. However, during long-term service, they are prone to internal defects such as cracks, voids, loose areas, and delamination. To ensure structural safety, it is necessary to detect and assess these internal defects. Since the interior of concrete structures cannot be directly observed, non-destructive testing (NDT) techniques are commonly used in current engineering inspections. Among these, ultrasonic wave propagation methods are widely used for detecting internal defects in concrete structures due to their advantages such as strong penetration, large detection depth, and applicability to large-volume structures. In recent years, with the development of array transducer technology, ultrasonic array-based imaging detection methods have gradually become an important technique for detecting internal defects in concrete. These methods utilize multiple transducers to collaboratively transmit and receive ultrasonic signals, and then reconstruct the internal reflection information of the structure using time-delay superposition or imaging algorithms, thereby achieving spatial localization and imaging of defects.

[0003] However, concrete is composed of a multiphase structure including aggregates, mortar, and pores, resulting in a highly non-uniform internal structure. Ultrasonic waves are prone to strong scattering during propagation, leading to significant speckle noise and random reflections in the echo signal. The sound velocity distribution within concrete is also uneven, causing the ultrasonic wave propagation path to often bend or deflect. Traditional methods typically assume a uniform sound velocity for imaging reconstruction, which increases the propagation time estimation error, resulting in blurred images or defect location misalignment. Furthermore, in complex scattering environments, numerous non-defect reflections can superimpose to form artifacts, making defect signals easily obscured by noise and reducing detection accuracy.

[0004] Existing methods for detecting concrete defects based on ultrasonic arrays mostly rely on simple time-delay superposition imaging or conventional signal enhancement processing, which have limited adaptability to the complex scattering environment inside concrete and make it difficult to effectively distinguish between aggregate scattering and reflection from actual defects. Current methods typically fail to adequately consider the effects of propagation path curvature and its changes, and lack screening mechanisms for spatial consistency of the propagation path, resulting in low imaging resolution, numerous artifacts, and insufficient localization accuracy for deep defects. Therefore, how to provide an imaging detection method for internal defects in concrete structures based on ultrasonic arrays is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an imaging detection method for internal defects in concrete structures based on ultrasonic arrays. This invention identifies candidate defect regions in strong scattering environments within concrete by introducing scattering potential energy distribution modeling, potential energy streamline tracking, and potential energy cluster localization methods. It also combines an improved NAFNet network to enhance the amplitude and phase of residual reflection signals. Furthermore, it constructs a sound velocity-curvature coupling field through sparse wavefront clustering and curvature canonical inversion, and generates a propagation reachability domain based on a propagation path folding and compression strategy, achieving precise delay compensation and coherent superposition imaging of multi-channel ultrasonic echo signals. This invention effectively suppresses scattering noise from concrete aggregates, improves the resolution of defect reflection signals, reduces imaging artifacts, and enhances the accuracy and clarity of deep defect localization. It possesses advantages such as strong anti-scattering capability, high imaging accuracy, and good detection stability.

[0006] An imaging detection method for internal defects in concrete structures based on an ultrasonic array, according to an embodiment of the present invention, includes: An array of ultrasonic transducers is arranged on the surface of the concrete structure to be tested. Multiple pulse excitations are performed according to a preset excitation sequence, and echo signals are received to obtain multi-channel ultrasonic echo data. Time alignment processing is performed on multi-channel ultrasonic echo data to extract phase change features, amplitude distribution features and channel consistency features, and to construct the scattering potential energy distribution data of the detection area. The scattering gradient vector field is calculated based on the scattering potential energy distribution data. Potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to form a set of potential energy streamlines. Density clustering is performed on the set of potential energy streamlines to obtain potential energy well clusters and determine the candidate core region of defects. The amplitude residual matrix and phase residual matrix are constructed and a residual reflection map is generated. The magnitude and phase matrices of the residual reflection map are input into the improved NAFNet network with dual-branch weight sharing. Frequency domain self-attention gating units are inserted in each convolutional layer. The scattering potential energy distribution data is used as the gating weight generation condition to output the defect reflection enhancement map. Based on the defect reflection enhancement map, sparse wavefront clustering is performed on the early arrival waves of each transmitting transducer and receiving transducer channel to determine the center arrival time of the early arrival wave cluster. Curvature canonical inversion is performed with the center arrival time as a constraint to obtain the sound velocity field and second-order propagation curvature field of the detection area, forming a sound velocity-curvature coupled field. Based on the sound velocity-curvature coupling field, the propagation paths of each transmit-receive channel are generated, and the propagation reachable domain is obtained by folding and compression. Delay compensation and coherent superposition are performed on the multi-channel ultrasonic echo data to output the defect imaging results.

[0007] Optionally, the step of performing multiple rounds of pulse excitation according to a preset excitation sequence and receiving echo signals to acquire multi-channel ultrasonic echo data includes: The ultrasonic transducer array is subjected to multiple rounds of pulse excitation according to a preset excitation sequence. The number of rounds of pulse excitation is 16. In each round of excitation, different transducers in the array are used as transmitters to emit ultrasonic pulses in sequence, and the remaining transducers are used as receivers to synchronously collect the echo signals after propagating inside the concrete. The corresponding transmitter transducer number, receiver transducer number and echo time series data are recorded to form a multi-channel ultrasonic echo dataset.

[0008] Optionally, the scattering potential energy distribution data of the constructed detection region includes: The multi-channel ultrasonic echo data is labeled and time-series organized. Each set of echo signals is marked according to the transmitting transducer number, receiving transducer number and excitation cycle, and the data is arranged according to a unified sampling time sequence to form a standardized multi-channel echo signal sequence. Time alignment processing is performed on the standardized multi-channel echo signal sequence. The initial arrival time of the echo signal is determined by calculating the position of the maximum value of the envelope energy of the echo signal of each channel, and time shift compensation is performed on the echo signal of each channel based on a unified reference time. Phase change characteristics are extracted from the time-aligned echo signal. The instantaneous phase change degree of the same transmitting and receiving transducer channels under different excitation cycles is calculated, and the average value of the phase change amplitude is statistically analyzed to obtain the phase change characteristics of the corresponding channel. Amplitude distribution features and channel consistency features are extracted from the time-aligned echo signals. The amplitude distribution features are obtained by statistically analyzing the average distribution of the echo signal amplitudes of each excitation round, and the channel consistency features are obtained by comparing the similarity of the echo signals of different excitation rounds at the same time position. The detection area is discretized into multiple spatial voxels. The phase change features, amplitude distribution features, and channel consistency features are mapped to the corresponding voxel positions. The scattering potential energy value of each voxel is calculated based on the weighted combination of the three types of features. The scattering potential energy values ​​of all voxels constitute the scattering potential energy distribution data of the detection area.

[0009] Optionally, the step of constructing the amplitude residual matrix and the phase residual matrix and generating the residual reflection map includes: Based on the scattering potential energy distribution data, the detection area is divided into multiple spatial voxels, and the scattering potential energy change direction of each voxel is calculated according to the change direction of the scattering potential energy value between adjacent voxels. The scattering potential energy change directions of all voxels together constitute the scattering gradient vector field of the detection area. Using each voxel in the detection area as a starting seed, potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to obtain a set of potential energy streamlines covering the entire detection area; The spatial distribution density of each streamline endpoint in the statistical potential energy streamline set is used to form multiple potential energy well clusters by density clustering of the endpoint locations, and the potential energy well clusters whose streamline endpoint density is higher than a preset density threshold are identified as candidate core areas of defects. Within the candidate core region of defects, an amplitude residual matrix is ​​constructed based on multi-channel ultrasonic echo data, specifically as follows: The echo amplitude values ​​obtained by each transmitting and receiving transducer channel under different excitation rounds are statistically analyzed, the reference amplitude of the channel echo amplitude is calculated, and the difference between the echo amplitude of each round and the corresponding reference amplitude is arranged according to the channel number and time series to form an amplitude residual matrix. Within the candidate core region of defects, a phase residual matrix is ​​constructed based on multi-channel ultrasonic echo data, specifically as follows: Extract the instantaneous phase information of the echo signals from each channel, calculate the difference between the phase of each round and the average phase of the channel, and arrange them according to the channel number and time sequence to form a phase residual matrix; Based on the streamline density of potential energy streamlines in the candidate core region of the defect, the amplitude residual matrix and the phase residual matrix are spatially weighted and fused to obtain the residual reflection map.

[0010] Optionally, the potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to obtain a set of potential energy streamlines covering the entire detection area, including: Read the scattering gradient vector for each seed voxel, and use the inverse of the scattering gradient vector as the current movement direction; The streamline movement step size is determined based on the gradient magnitude range of the current voxel. When the gradient magnitude is within the preset first gradient range, the first step size is used, and when the gradient magnitude is within the preset second gradient range, the second step size is used. When jumping to the next voxel along the movement direction, compare the angle between the two movement directions. If the angle exceeds the set bending threshold, insert an intermediate transition voxel to eliminate trajectory dispersion caused by excessive back-and-forth. The gradient magnitude is detected in real time during continuous movement. When the gradient magnitude is lower than the termination threshold or the detection area boundary is reached, the tracking stops and the movement path is recorded as a complete potential energy streamline. The streamline tracing process is repeated for all voxels within the detection area. The spatial overlap of the generated streamlines is judged. Streamlines with an overlap degree higher than the preset overlap threshold are merged or deleted. Streamlines with representative spatial distribution are retained to obtain a set of potential energy streamlines covering the entire detection area.

[0011] Optionally, the output defect reflection enhancement map includes: The residual reflection map is decomposed into amplitude and phase. The residual amplitudes of each transmitting and receiving transducer channel at different time sampling points are arranged according to the channel dimension and time dimension to form an amplitude matrix. The residual phases of the corresponding sampling points are arranged according to the same dimension to form a phase matrix. The amplitude matrix is ​​input into the amplitude branch of the improved NAFNet network, and the phase matrix is ​​input into the phase branch of the improved NAFNet network. The amplitude branch and the phase branch are composed of multiple levels of residual blocks connected in series. In each level of residual block, the two branches share the same set of convolution kernel weights. Independent channel scaling coefficients are set only at the output of the branches to extract amplitude features and phase features synchronously. In each residual block of the improved NAFNet network, a frequency domain self-attention gating unit is set up. The feature map output by the residual block is divided into frequency bands along the time dimension. The feature response intensity in each frequency band is calculated. Frequency band attention weights are generated based on the response intensity of each frequency band. The feature map of the corresponding frequency band is weighted using the frequency band attention weights to obtain the amplitude feature map and phase feature map after frequency domain self-attention gating. Potential energy guided gating units are set in each residual block of the improved NAFNet network. The scattered potential energy distribution data is mapped to the feature map spatial resolution to generate a potential energy weight map. The weight map is then multiplied element-wise with the amplitude feature map and phase feature map after frequency domain self-attention gating. Differential weighting is applied to the feature responses of different potential energy ranges to obtain the amplitude feature map and phase feature map after potential energy guided gating. An amplitude-phase joint reconstruction unit is set at the output of the improved NAFNet network. The amplitude feature map and phase feature map output by the final first-level residual block are concatenated in the channel dimension. An amplitude-phase joint feature map is generated by a set of one-dimensional convolution and channel mixing operations. The defect reflection intensity is calculated on the amplitude-phase joint feature map in spatial position to obtain the defect reflection enhancement value corresponding to each voxel. The defect reflection enhancement values ​​of all voxels constitute the defect reflection enhancement map.

[0012] Optionally, obtaining the sound velocity field and the second-order propagation curvature field of the detection area to form a sound velocity-curvature coupled field includes: Based on the defect reflection enhancement map, the enhanced echo signals corresponding to each transmitting and receiving transducer channel are organized according to channel number and time sequence to obtain the enhanced echo sequence of each channel. Early arrival candidate points are extracted for the enhanced echo sequence. The method of energy threshold and rising edge consistency is used to determine the set of time points in the time series of each channel that meet the condition of first exceeding the threshold and maintaining an upward trend for a number of consecutive sampling points. These are used as the early arrival candidate point set for the channel. Sparse wavefront clustering is performed on the early arrival wave candidate point set of all channels. The early arrival wave candidate points are associated according to the channel spatial adjacency relationship and the arrival time proximity relationship to form several early arrival wave clusters. For each early arrival wave cluster, the statistical center value of the arrival time of all candidate points in the cluster is calculated, and the center arrival time is assigned to the channel set corresponding to the cluster. The detection area is discretized into multiple spatial voxels, and a set of parameters to be solved is established, which consists of voxel sound velocity parameters and voxel second-order propagation curvature parameters. The propagation path length of each channel at each voxel is determined based on the position of the transmitting transducer, the position of the receiving transducer, and the position of the voxel. The predicted arrival time of each channel is calculated using the voxel sound velocity parameters and the propagation path length. The difference between the center arrival time and the predicted arrival time of each channel is used as the arrival time constraint term, and the smoothness of the change of the second-order propagation curvature parameter between adjacent voxels is used as the curvature regularization term. The set of parameters to be solved is iteratively updated until convergence, and the sound speed field and the second-order propagation curvature field of the detection area are output. The sound speed-curvature coupling field is constructed from the sound speed field and the second-order propagation curvature field.

[0013] Optionally, the output defect imaging results include: Based on the sound velocity-curvature coupling field, the propagation path from the transmitting transducer to the receiving transducer via voxels is generated. The propagation time at each voxel is calculated based on the sound velocity field and the path length. The propagation time is compared with the arrival time of the early arrival wave center. Voxel points with propagation time deviations within the preset time tolerance range are selected to form an initial propagation path point set. The initial propagation path point set is processed by path association, connecting path points that belong to the same transmitting and receiving transducer channels and are spatially continuous to form path segments, and recording the propagation curvature change information corresponding to each path segment. Based on the propagation curvature change of the path segment, folding and compression processing is performed. Path segments with continuous curvature and consistent direction are spatially folded and projected onto the same propagation trajectory centerline. Path segments that deviate from the center trajectory by more than a preset spatial deviation threshold are deleted. The folded path segments are spatially connected and filtered. Path segments that appear repeatedly in multiple transmit-receive channels and have the same position are counted. Path segments that appear more than a preset path consistency threshold are retained to form the propagation reachability domain. Delay compensation processing is performed on multi-channel ultrasonic echo data within the propagation reach domain. Coherent superposition calculation is performed on the multi-channel echo signals after delay compensation to obtain the defect reflection intensity corresponding to each spatial voxel, thereby generating imaging results of internal defects in concrete structures.

[0014] The beneficial effects of this invention are: Compared with existing methods for detecting concrete defects based on ultrasonic arrays, this invention constructs scattering potential energy distribution data and performs potential energy flow tracking based on the scattering gradient vector field to form a set of potential energy streamlines and identify potential energy well clusters, thereby determining the candidate core area of ​​defects and achieving effective localization of potential defect areas in the strong scattering environment inside concrete. This method can spatially distinguish aggregate scattering signals using the variation law of scattering potential energy, making the true defect reflection area stand out in a complex scattering background, reducing the interference of random scattering on defect identification, and improving the accuracy of defect area extraction.

[0015] This invention generates a residual reflection map by constructing amplitude and phase residual matrices. The amplitude and phase features are then input into an improved NAFNet network with shared weights across two branches for joint enhancement. Simultaneously, potential energy gating weights are generated using scattering potential energy distribution data to spatially modulate the network features, thereby enhancing the reflection characteristics of defect areas and suppressing noise in non-defect areas. This approach effectively improves the signal-to-noise ratio of defect reflection signals, reduces the impact of speckle noise on imaging results, and makes the response characteristics of minute defects within concrete more pronounced in the enhanced map.

[0016] This invention obtains the arrival time of the early-arriving wave center by performing sparse wavefront clustering on the enhanced echo signal, and uses this as a constraint to perform curvature regularization inversion to obtain the sound velocity field and the second-order propagation curvature field. A sound velocity-curvature coupled field is generated, and the propagation path is folded and compressed to form the propagation reachability domain. Delay compensation and coherent superposition imaging are then performed on the multi-channel ultrasonic echo data. By introducing propagation path consistency screening and propagation domain compression mechanisms, the interference of unreasonable propagation paths on the imaging results can be effectively reduced, imaging artifacts can be decreased, and the positioning accuracy and imaging resolution of deep defects can be improved, making the detection results of internal defects in concrete structures more accurate and reliable. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the imaging detection method for internal defects in concrete structures based on ultrasonic arrays proposed in this invention. Figure 2 This is a schematic diagram of the improved NAFNet network structure with dual-branch weight sharing for the imaging detection method of internal defects in concrete structures based on ultrasonic arrays proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 An imaging detection method for internal defects in concrete structures based on ultrasonic arrays includes: An array of ultrasonic transducers is arranged on the surface of the concrete structure to be tested. Multiple pulse excitations are performed according to a preset excitation sequence, and echo signals are received to obtain multi-channel ultrasonic echo data. Time alignment processing is performed on multi-channel ultrasonic echo data to extract phase change features, amplitude distribution features and channel consistency features, and to construct the scattering potential energy distribution data of the detection area. The scattering gradient vector field is calculated based on the scattering potential energy distribution data. Potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to form a set of potential energy streamlines. Density clustering is performed on the set of potential energy streamlines to obtain potential energy well clusters and determine the candidate core region of defects. The amplitude residual matrix and phase residual matrix are constructed and a residual reflection map is generated. The magnitude and phase matrices of the residual reflection map are input into the improved NAFNet network with dual-branch weight sharing. Frequency domain self-attention gating units are inserted in each convolutional layer. The scattering potential energy distribution data is used as the gating weight generation condition to output the defect reflection enhancement map. Based on the defect reflection enhancement map, sparse wavefront clustering is performed on the early arrival waves of each transmitting transducer and receiving transducer channel to determine the center arrival time of the early arrival wave cluster. Curvature canonical inversion is performed with the center arrival time as a constraint to obtain the sound velocity field and second-order propagation curvature field of the detection area, forming a sound velocity-curvature coupled field. Based on the sound velocity-curvature coupling field, the propagation paths of each transmit-receive channel are generated, and the propagation reachable domain is obtained by folding and compression. Delay compensation and coherent superposition are performed on the multi-channel ultrasonic echo data to output the defect imaging results.

[0020] In this embodiment, the step of performing multiple rounds of pulse excitation according to a preset excitation sequence and receiving echo signals to acquire multi-channel ultrasonic echo data includes: The ultrasonic transducer array is subjected to multiple rounds of pulse excitation according to a preset excitation sequence. The number of rounds of pulse excitation is 16. In each round of excitation, different transducers in the array are used as transmitters to emit ultrasonic pulses in sequence, and the remaining transducers are used as receivers to synchronously collect the echo signals after propagating inside the concrete. The corresponding transmitter transducer number, receiver transducer number and echo time series data are recorded to form a multi-channel ultrasonic echo dataset.

[0021] In this embodiment, the scattering potential energy distribution data of the constructed detection region includes: The multi-channel ultrasonic echo data is labeled and time-series organized. Each set of echo signals is marked according to the transmitting transducer number, receiving transducer number and excitation cycle, and the data is arranged according to a unified sampling time sequence to form a standardized multi-channel echo signal sequence. Time alignment processing is performed on the standardized multi-channel echo signal sequence. The initial arrival time of the echo signal is determined by calculating the location of the maximum value of the envelope energy of each channel's echo signal. Time shift compensation is then performed on the echo signals of each channel based on a unified reference time. Specifically, the location of the maximum value of the envelope energy of each channel's echo signal is calculated as follows: For each channel echo signal, the envelope is first extracted to obtain the corresponding envelope amplitude sequence; The envelope energy value within the window is calculated using a sliding time window on the envelope amplitude sequence. The envelope energy value is the sum of the squares of the envelope amplitudes at each sampling point within the time window. Traverse all sliding time windows and compare the envelope energy values ​​corresponding to each time window. Select the time window with the largest envelope energy value, determine the center sampling time of the maximum energy time window as the position of the maximum envelope energy value of the channel echo signal, and determine the time corresponding to the position of the maximum envelope energy value as the initial arrival time of the channel echo signal. Phase change features are extracted based on the time-aligned echo signal. The instantaneous phase change degree of the same transmitting and receiving transducer channels under different excitation cycles is calculated, and the average value of the phase change amplitude is statistically analyzed to obtain the phase change features of the corresponding channels. Specifically, the calculation of the instantaneous phase change degree of the same transmitting and receiving transducer channels under different excitation cycles is as follows: The time-aligned echo signals of the same transmitting and receiving transducer channels under each excitation cycle are analyzed to construct the corresponding instantaneous phase sequences. Under a unified time reference, the instantaneous phase sequences of each round are aligned point by point, and the average phase sequence of the instantaneous phase sequences of each round is used as the reference phase sequence. For each excitation cycle, the phase difference between the instantaneous phase and the reference phase is calculated at each sampling point, and the phase difference is processed by phase expansion to eliminate phase jump; The absolute value of the phase difference at each sampling point is accumulated within a preset time window and divided by the number of sampling points to obtain the instantaneous phase change. The instantaneous phase change of the channel under different excitation cycles is obtained by averaging the instantaneous phase change of all excitation cycles. Amplitude distribution features and channel consistency features are extracted from the time-aligned echo signals. The amplitude distribution features are obtained by statistically analyzing the average distribution of the echo signal amplitudes of each excitation round, and the channel consistency features are obtained by comparing the similarity of the echo signals of different excitation rounds at the same time position. The detection area is discretized into multiple spatial voxels. Phase change features, amplitude distribution features, and channel consistency features are mapped to the corresponding voxel locations. The scattering potential energy value of each voxel is calculated based on a weighted combination of the three types of features. The scattering potential energy values ​​of all voxels constitute the scattering potential energy distribution data of the detection area. The calculation of the scattering potential energy value is as follows: At each spatial voxel location, the multi-channel echo signal features passing through the voxel propagation path are collected, and the corresponding phase change feature values, amplitude distribution feature values, and channel consistency feature values ​​are extracted respectively. The three types of feature values ​​are then normalized to a uniform scale. The normalized phase change characteristic value, amplitude distribution characteristic value and channel consistency characteristic value are weighted and combined according to the preset weight coefficients to obtain the comprehensive scattering response value corresponding to the voxel. The phase change characteristic is used to characterize the phase stability, the amplitude distribution characteristic is used to characterize the echo energy distribution, and the channel consistency characteristic is used to characterize the spatial consistency of the multi-channel echo signal. The scattering potential energy of each spatial voxel is obtained by mapping the comprehensive scattering response value according to the inverse correspondence between the comprehensive scattering response value and the scattering potential energy.

[0022] In this embodiment, the step of constructing the amplitude residual matrix and the phase residual matrix and generating the residual reflection map includes: Based on the scattering potential energy distribution data, the detection area is divided into multiple spatial voxels. The direction of scattering potential energy change for each voxel is calculated according to the direction of change of scattering potential energy values ​​between adjacent voxels. The scattering potential energy change directions of all voxels together constitute the scattering gradient vector field of the detection area. Specifically, the calculation of the scattering potential energy change direction for each voxel is as follows: Taking the current voxel as the center, select its neighboring voxels in three spatial directions, calculate the difference in scattering potential energy between the current voxel and each neighboring voxel, and use the difference as the potential energy change in the corresponding spatial direction. The trend of voxel scattering potential energy change in space is determined by the potential energy change in three spatial directions. The spatial direction with the largest potential energy change is taken as the main change direction of the voxel, and the directional relationship of the change direction is recorded. The main potential energy change directions of each voxel are combined according to their spatial positions to form a set of direction vectors describing the trend of scattering potential energy change, which constitutes the scattering gradient vector field of the detection area. Using each voxel in the detection area as a starting seed, potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to obtain a set of potential energy streamlines covering the entire detection area; The spatial distribution density of each streamline endpoint in the statistical potential energy streamline set is used to form multiple potential energy well clusters by density clustering of the endpoint locations, and the potential energy well clusters whose streamline endpoint density is higher than a preset density threshold are identified as candidate core areas of defects. Within the candidate core region of defects, an amplitude residual matrix is ​​constructed based on multi-channel ultrasonic echo data, specifically as follows: The echo amplitudes obtained by each transmitting and receiving transducer channel under different excitation cycles are statistically analyzed, and a reference amplitude for the channel echo amplitude is calculated. The differences between the echo amplitudes of each cycle and the corresponding reference amplitudes are arranged according to the channel number and time series to form an amplitude residual matrix. Specifically, the reference amplitude for calculating the channel echo amplitude is as follows: The amplitude of the echo signals obtained from the same transmitting and receiving transducer channels under different excitation cycles is extracted to obtain the echo amplitude set on the corresponding time series, and then arranged uniformly according to the time sampling order. Under a unified time reference, the echo amplitude of each excitation round corresponding to the time sampling point is statistically calculated to obtain the average value of the echo amplitude of all excitation rounds at the time sampling point, and the average value is used as the reference amplitude of the time sampling point. The reference amplitudes corresponding to all time sampling points are combined in chronological order to form a reference amplitude sequence for the channel. This reference amplitude sequence is then used as the reference amplitude for calculating the amplitude residual. Within the candidate core region of defects, a phase residual matrix is ​​constructed based on multi-channel ultrasonic echo data, specifically as follows: Extract the instantaneous phase information of the echo signals from each channel, calculate the difference between the phase of each round and the average phase of the channel, and arrange them according to the channel number and time sequence to form a phase residual matrix; Based on the streamline density of potential energy streamlines in the candidate core region of the defect, the amplitude residual matrix and the phase residual matrix are spatially weighted and fused to obtain the residual reflection map.

[0023] In this embodiment, the step of performing potential energy flow tracking along the negative direction of the scattering gradient vector field to obtain a set of potential energy streamlines covering the entire detection area includes: Read the scattering gradient vector for each seed voxel, and use the inverse of the scattering gradient vector as the current movement direction; The streamline movement step size is determined based on the gradient magnitude range of the current voxel. When the gradient magnitude is within the preset first gradient range, the first step size is used, and when the gradient magnitude is within the preset second gradient range, the second step size is used. When jumping to the next voxel along the movement direction, compare the angle between the two movement directions. If the angle exceeds the set bending threshold, insert an intermediate transition voxel to eliminate trajectory dispersion caused by excessive back-and-forth. The gradient magnitude is detected in real time during continuous movement. When the gradient magnitude is lower than the termination threshold or the detection area boundary is reached, the tracking stops and the movement path is recorded as a complete potential energy streamline. The streamline tracing process is repeated for all voxels within the detection area. The spatial overlap of the generated streamlines is judged. Streamlines with an overlap degree higher than the preset overlap threshold are merged or deleted. Streamlines with representative spatial distribution are retained to obtain a set of potential energy streamlines covering the entire detection area.

[0024] In this embodiment, the output defect reflection enhancement map includes: The residual reflection map is decomposed into amplitude and phase. The residual amplitudes of each transmitting and receiving transducer channel at different time sampling points are arranged according to the channel dimension and time dimension to form an amplitude matrix. The residual phases of the corresponding sampling points are arranged according to the same dimension to form a phase matrix. The amplitude matrix is ​​input into the amplitude branch of the improved NAFNet network, and the phase matrix is ​​input into the phase branch of the improved NAFNet network. The amplitude branch and the phase branch are composed of multiple levels of residual blocks connected in series. In each level of residual block, the two branches share the same set of convolution kernel weights. Independent channel scaling coefficients are set only at the output of the branches to extract amplitude features and phase features synchronously. In each residual block of the improved NAFNet network, a frequency-domain self-attention gating unit is set. The feature map output by the residual block is divided into frequency bands along the time dimension. The feature response intensity in each frequency band is calculated, and frequency band attention weights are generated based on the response intensity of each frequency band. The feature maps of the corresponding frequency bands are then weighted using the frequency band attention weights to obtain the amplitude feature map and phase feature map after frequency-domain self-attention gating. Specifically, the calculation of the feature response intensity in each frequency band is as follows: The feature map output by the residual block is transformed in the frequency domain along the time dimension, converting the time domain feature representation into the frequency domain feature representation, and the frequency domain feature is divided into multiple frequency band regions according to the preset frequency range. In each frequency band region, the amplitude information of each characteristic response value within the corresponding frequency range is statistically analyzed, and the amplitudes of all characteristic responses within the frequency band are accumulated and statistically analyzed to obtain the overall response energy value corresponding to the frequency band. The overall response energy value corresponding to each frequency band is normalized so that the response intensity of different frequency bands is within a uniform scale range, thus obtaining the characteristic response intensity of each frequency band. Potential energy guided gating units are set in each residual block of the improved NAFNet network. The scattered potential energy distribution data is mapped to the feature map spatial resolution to generate a potential energy weight map. The weight map is then multiplied element-wise with the amplitude feature map and phase feature map after frequency domain self-attention gating. Differential weighting is applied to the feature responses of different potential energy ranges to obtain the amplitude feature map and phase feature map after potential energy guided gating. An amplitude-phase joint reconstruction unit is set at the output of the improved NAFNet network. The amplitude feature map and phase feature map output from the final first-level residual block are concatenated along the channel dimension. An amplitude-phase joint feature map is generated through a set of one-dimensional convolution and channel mixing operations. The defect reflection intensity is calculated on the amplitude-phase joint feature map in spatial location to obtain the defect reflection enhancement value corresponding to each voxel. The defect reflection enhancement values ​​of all voxels constitute the defect reflection enhancement map. The calculation of defect reflection intensity is specifically as follows: Extract the corresponding channel feature vector from the amplitude-phase joint feature map at each spatial voxel location, and perform amplitude statistics on the feature values ​​of each channel in the channel feature vector to obtain the set of feature response amplitudes at the voxel location. The feature response amplitude set is weighted and accumulated, where the feature values ​​of each channel are weighted according to the preset channel weights to obtain the comprehensive feature response value of the voxel position; The comprehensive characteristic response value is normalized and mapped to a numerical value that characterizes the intensity of defect reflection. The numerical value is the defect reflection enhancement value of the corresponding spatial voxel. The defect reflection enhancement values ​​of all voxels are combined to form a defect reflection enhancement map.

[0025] In this embodiment, obtaining the sound velocity field and the second-order propagation curvature field of the detection area to form a sound velocity-curvature coupled field includes: Based on the defect reflection enhancement map, the enhanced echo signals corresponding to each transmitting and receiving transducer channel are organized according to channel number and time sequence to obtain the enhanced echo sequence of each channel. Early arrival candidate points are extracted for the enhanced echo sequence. The method of energy threshold and rising edge consistency is used to determine the set of time points in the time series of each channel that meet the condition of first exceeding the threshold and maintaining an upward trend for a number of consecutive sampling points. These are used as the early arrival candidate point set for the channel. Sparse wavefront clustering is performed on the early arrival wave candidate point set of all channels. The early arrival wave candidate points are associated according to the channel spatial adjacency relationship and the arrival time proximity relationship to form several early arrival wave clusters. For each early arrival wave cluster, the statistical center value of the arrival time of all candidate points in the cluster is calculated, and the center arrival time is assigned to the channel set corresponding to the cluster. The detection area is discretized into multiple spatial voxels, and a set of parameters to be solved is established, which consists of voxel sound velocity parameters and voxel second-order propagation curvature parameters. The propagation path length of each channel at each voxel is determined based on the position of the transmitting transducer, the position of the receiving transducer, and the position of the voxel. The predicted arrival time of each channel is calculated using the voxel sound velocity parameters and the propagation path length. The difference between the center arrival time and the predicted arrival time of each channel is used as the arrival time constraint term, and the smoothness of the change of the second-order propagation curvature parameter between adjacent voxels is used as the curvature regularization term. The set of parameters to be solved is iteratively updated until convergence, and the sound speed field and the second-order propagation curvature field of the detection area are output. The sound speed-curvature coupling field is constructed from the sound speed field and the second-order propagation curvature field.

[0026] In this embodiment, the output defect imaging result includes: Based on the sound velocity-curvature coupling field, the propagation path from the transmitting transducer to the receiving transducer via voxels is generated. The propagation time at each voxel is calculated based on the sound velocity field and the path length. The propagation time is compared with the arrival time of the early arrival wave center. Voxel points with propagation time deviations within the preset time tolerance range are selected to form an initial propagation path point set. The initial propagation path point set is processed by path association, connecting path points that belong to the same transmitting and receiving transducer channels and are spatially continuous to form path segments, and recording the propagation curvature change information corresponding to each path segment. Based on the propagation curvature change of the path segment, folding and compression processing is performed. Path segments with continuous curvature and consistent direction are spatially folded and projected onto the same propagation trajectory centerline. Path segments that deviate from the center trajectory by more than a preset spatial deviation threshold are deleted. The folded path segments are spatially connected and filtered. Path segments that appear repeatedly in multiple transmit-receive channels and have the same position are counted. Path segments that appear more than a preset path consistency threshold are retained to form the propagation reachability domain. Within the propagation reach domain, delay compensation processing is performed on the multi-channel ultrasonic echo data. Coherent superposition calculation is then performed on the delay-compensated multi-channel echo signals to obtain the defect reflection intensity corresponding to each spatial voxel, thereby generating imaging results of internal defects in the concrete structure. Specifically, the coherent superposition calculation of the delay-compensated multi-channel echo signals involves: At each spatial voxel location, select all transmitting and receiving transducer channels that pass through the voxel propagation path, align the echo signals after delay compensation of each channel at the same sampling point, and extract the corresponding echo amplitude and phase information. Under a unified time reference, the phase consistency of the echo signals after alignment of each channel is judged, and the channel signals with phase deviation within the preset phase tolerance range are retained. The retained channel signals are then weighted and accumulated according to the channel weight to obtain the comprehensive echo response value of the voxel position. The amplitude of the composite echo response value is normalized, and the normalized value is used as the defect reflection intensity corresponding to the spatial voxel. The above calculation process is repeated for all voxels in the detection area to generate the imaging results of internal defects in the concrete structure.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to the inspection of precast beams in a viaduct project. The inspection object was a precast concrete box girder specimen at the bridge construction site. This box girder was cast using C50 concrete, with a length of 30m, a web thickness of approximately 0.32m, and aggregate particle size ranging from approximately 5mm to 25mm. Since precast beams may develop internal defects such as voids, honeycombing, and microcracks during production and transportation, non-destructive testing of its internal structure is necessary after the beam has been cast and reached its design strength to ensure that the structural quality meets engineering requirements. Traditional inspection methods typically employ single-probe ultrasonic testing or simple time-delay superposition imaging. However, due to strong scattering from the aggregate within the concrete, the echo signal contains a large amount of noise, making defect echoes easily masked, especially for deep defects that are difficult to locate accurately.

[0028] During the inspection process, an ultrasonic transducer array was first deployed on the outer surface of the concrete box girder web. The array consisted of a linear array structure of 32 piezoelectric transducers with a transducer spacing of 20 mm, a center operating frequency of 50 kHz, and a sampling frequency of 2 MHz. During inspection, each transducer was sequentially pulse-excited according to a preset excitation sequence, while the remaining transducers synchronously received the echo signals propagating through the concrete. Each excitation round lasted approximately 3 microseconds, and each acquisition round lasted 1.8 milliseconds, for a total of 48 rounds of excitation and acquisition, thus obtaining a large amount of multi-channel ultrasonic echo data. Subsequently, the acquired echo data underwent time alignment processing to extract the phase change characteristics, amplitude distribution characteristics, and channel consistency characteristics of the echo signals, and scattered potential energy distribution data was constructed within the inspection area. Based on the scattered potential energy distribution data, a scattered gradient vector field was calculated. Potential energy flow tracking was performed along the negative direction of the scattered gradient vector field within the inspection area to form a set of potential energy streamlines. Potential energy well clusters were identified through streamline density clustering, thereby determining candidate core regions where defects may exist.

[0029] After obtaining the candidate defect regions, amplitude and phase residual matrices are constructed from the multi-channel echo signals, and residual reflection maps are generated. The amplitude and phase information from the residual reflection maps are then input into an improved NAFNet network with shared weights across two branches for feature enhancement. Frequency-domain self-attention gating units and potential energy-guided gating units are introduced into the network structure, enabling the network to spatially modulate the features using scattering potential energy distribution data while extracting features, thereby enhancing the reflection characteristics of the defect region and suppressing aggregate scattering noise. The enhanced defect reflection map is obtained after this enhancement process. Subsequently, sparse wavefront clustering is performed on the enhanced echo signals to determine the early arrival clusters of each transmitting and receiving transducer channel, and the corresponding center arrival times are calculated. Using the center arrival time as a constraint, the sound velocity field and second-order propagation curvature field of the detection region are calculated through curvature canonical inversion, forming a sound velocity-curvature coupling field. This generates propagation paths from each transmitting transducer to each spatial voxel and then to the receiving transducer. The paths are then folded and compressed according to the propagation curvature characteristics to form the propagation reachability domain. Finally, delay compensation and coherent superposition processing are performed on the multi-channel ultrasonic echo data within the propagation reach domain to obtain imaging results of internal defects in the concrete structure.

[0030] To verify the detection effect of the method of the present invention, several artificial defects were pre-embedded inside the concrete specimen before testing, including spherical cavities with a diameter of 12 mm to 20 mm and simulated cracks with a width of about 2 mm to 3 mm, and their actual embedding depth was recorded. After the test, the imaging results were compared and analyzed with the actual defect locations.

[0031] Table 1 Comparison of Detection Results of Typical Internal Defects in Concrete

[0032] As shown in Table 1, the detection results demonstrate that the method of this invention exhibits good positioning accuracy in detecting spherical void defects inside concrete. For spherical void defects numbered A1 to A3, their actual depths are 620 mm, 780 mm, and 920 mm, respectively. The detection depths of this invention are 638 mm, 801 mm, and 948 mm, respectively, with positioning errors of 18 mm, 21 mm, and 28 mm, all within 30 mm. In contrast, the traditional method detects depths of 684 mm, 846 mm, and 996 mm, with positioning errors of 64 mm, 66 mm, and 76 mm, respectively, showing significantly larger positioning deviations.

[0033] For crack-type defects A4 and A5, the detection depths of this invention are 754 mm and 1079 mm, respectively, with corresponding positioning errors of 24 mm and 29 mm, while the detection errors of traditional methods are 82 mm and 92 mm, respectively. Crack-type defects have weak reflection signals and are easily affected by aggregate scattering, but by enhancing the echo signal using the method of this invention and combining it with propagation path constraints, the crack location can still be identified relatively stably.

[0034] For deep defects A6 and A7, with actual depths of 1280mm and 1450mm respectively, the positioning errors of this invention are 34mm and 33mm respectively, while the errors of traditional methods reach 108mm and 92mm respectively. This demonstrates that the method of this invention can maintain high positioning accuracy in deep defect detection, effectively reduce imaging errors, and improve the reliability of internal defect detection in concrete structures.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for imaging and detecting internal defects in concrete structures based on ultrasonic arrays, characterized in that, include: An array of ultrasonic transducers is arranged on the surface of the concrete structure to be tested. Multiple pulse excitations are performed according to a preset excitation sequence, and echo signals are received to obtain multi-channel ultrasonic echo data. Time alignment processing is performed on multi-channel ultrasonic echo data to extract phase change features, amplitude distribution features and channel consistency features, and to construct the scattering potential energy distribution data of the detection area. The scattering gradient vector field is calculated based on the scattering potential energy distribution data. Potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to form a set of potential energy streamlines. Density clustering is performed on the set of potential energy streamlines to obtain potential energy well clusters and determine the candidate core region of defects. The amplitude residual matrix and phase residual matrix are constructed and a residual reflection map is generated. The magnitude and phase matrices of the residual reflection map are input into the improved NAFNet network with dual-branch weight sharing. Frequency domain self-attention gating units are inserted in each convolutional layer. The scattering potential energy distribution data is used as the gating weight generation condition to output the defect reflection enhancement map. Based on the defect reflection enhancement map, sparse wavefront clustering is performed on the early arrival waves of each transmitting transducer and receiving transducer channel to determine the center arrival time of the early arrival wave cluster. Curvature canonical inversion is performed with the center arrival time as a constraint to obtain the sound velocity field and second-order propagation curvature field of the detection area, forming a sound velocity-curvature coupled field. Based on the sound velocity-curvature coupling field, the propagation paths of each transmit-receive channel are generated, and the propagation reachable domain is obtained by folding and compression. Delay compensation and coherent superposition are performed on the multi-channel ultrasonic echo data to output the defect imaging results.

2. The method for imaging and detecting internal defects in concrete structures based on ultrasonic arrays according to claim 1, characterized in that, The process of performing multiple rounds of pulse excitation according to a preset excitation sequence and receiving echo signals to acquire multi-channel ultrasonic echo data includes: The ultrasonic transducer array is subjected to multiple rounds of pulse excitation according to a preset excitation sequence. The number of rounds of pulse excitation is 16. In each round of excitation, different transducers in the array are used as transmitters to emit ultrasonic pulses in sequence, and the remaining transducers are used as receivers to synchronously collect the echo signals after propagating inside the concrete. The corresponding transmitter transducer number, receiver transducer number and echo time series data are recorded to form a multi-channel ultrasonic echo dataset.

3. The method for imaging and detecting internal defects in concrete structures based on ultrasonic arrays according to claim 1, characterized in that, The scattering potential energy distribution data of the constructed detection region includes: The multi-channel ultrasonic echo data is labeled and time-series organized. Each set of echo signals is marked according to the transmitting transducer number, receiving transducer number and excitation cycle, and the data is arranged according to a unified sampling time sequence to form a standardized multi-channel echo signal sequence. Time alignment processing is performed on the standardized multi-channel echo signal sequence. The initial arrival time of the echo signal is determined by calculating the position of the maximum value of the envelope energy of the echo signal of each channel, and time shift compensation is performed on the echo signal of each channel based on a unified reference time. Phase change characteristics are extracted from the time-aligned echo signal. The instantaneous phase change degree of the same transmitting and receiving transducer channels under different excitation cycles is calculated, and the average value of the phase change amplitude is statistically analyzed to obtain the phase change characteristics of the corresponding channel. Amplitude distribution features and channel consistency features are extracted from the time-aligned echo signals. The amplitude distribution features are obtained by statistically analyzing the average distribution of the echo signal amplitudes of each excitation round, and the channel consistency features are obtained by comparing the similarity of the echo signals of different excitation rounds at the same time position. The detection area is discretized into multiple spatial voxels. The phase change features, amplitude distribution features, and channel consistency features are mapped to the corresponding voxel positions. The scattering potential energy value of each voxel is calculated based on the weighted combination of the three types of features. The scattering potential energy values ​​of all voxels constitute the scattering potential energy distribution data of the detection area.

4. The method for imaging and detecting internal defects in concrete structures based on ultrasonic arrays according to claim 1, characterized in that, The process of constructing the amplitude residual matrix and the phase residual matrix and generating the residual reflection map includes: Based on the scattering potential energy distribution data, the detection area is divided into multiple spatial voxels, and the scattering potential energy change direction of each voxel is calculated according to the change direction of the scattering potential energy value between adjacent voxels. The scattering potential energy change directions of all voxels together constitute the scattering gradient vector field of the detection area. Using each voxel in the detection area as a starting seed, potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to obtain a set of potential energy streamlines covering the entire detection area; The spatial distribution density of each streamline endpoint in the statistical potential energy streamline set is used to form multiple potential energy well clusters by density clustering of the endpoint locations, and the potential energy well clusters whose streamline endpoint density is higher than a preset density threshold are identified as candidate core areas of defects. Within the candidate core region of defects, an amplitude residual matrix is ​​constructed based on multi-channel ultrasonic echo data, specifically as follows: The echo amplitude values ​​obtained by each transmitting and receiving transducer channel under different excitation rounds are statistically analyzed, the reference amplitude of the channel echo amplitude is calculated, and the difference between the echo amplitude of each round and the corresponding reference amplitude is arranged according to the channel number and time series to form an amplitude residual matrix. Within the candidate core region of defects, a phase residual matrix is ​​constructed based on multi-channel ultrasonic echo data, specifically as follows: Extract the instantaneous phase information of the echo signals from each channel, calculate the difference between the phase of each round and the average phase of the channel, and arrange them according to the channel number and time sequence to form a phase residual matrix; Based on the streamline density of potential energy streamlines in the candidate core region of the defect, the amplitude residual matrix and the phase residual matrix are spatially weighted and fused to obtain the residual reflection map.

5. The imaging detection method for internal defects in concrete structures based on ultrasonic arrays according to claim 4, characterized in that, The potential energy flow tracking is performed along the negative direction of the scattering gradient vector field to obtain a set of potential energy streamlines covering the entire detection area, including: Read the scattering gradient vector for each seed voxel, and use the inverse of the scattering gradient vector as the current movement direction; The streamline movement step size is determined based on the gradient magnitude range of the current voxel. When the gradient magnitude is within the preset first gradient range, the first step size is used, and when the gradient magnitude is within the preset second gradient range, the second step size is used. When jumping to the next voxel along the movement direction, compare the angle between the two movement directions. If the angle exceeds the set bending threshold, insert an intermediate transition voxel to eliminate trajectory dispersion caused by excessive back-and-forth. The gradient magnitude is detected in real time during continuous movement. When the gradient magnitude is lower than the termination threshold or the detection area boundary is reached, the tracking stops and the movement path is recorded as a complete potential energy streamline. The streamline tracing process is repeated for all voxels within the detection area. The spatial overlap of the generated streamlines is judged. Streamlines with an overlap degree higher than the preset overlap threshold are merged or deleted. Streamlines with representative spatial distribution are retained to obtain a set of potential energy streamlines covering the entire detection area.

6. The imaging detection method for internal defects in concrete structures based on ultrasonic arrays according to claim 1, characterized in that, The output defect reflection enhancement map includes: The residual reflection map is decomposed into amplitude and phase. The residual amplitudes of each transmitting and receiving transducer channel at different time sampling points are arranged according to the channel dimension and time dimension to form an amplitude matrix. The residual phases of the corresponding sampling points are arranged according to the same dimension to form a phase matrix. The amplitude matrix is ​​input into the amplitude branch of the improved NAFNet network, and the phase matrix is ​​input into the phase branch of the improved NAFNet network. The amplitude branch and the phase branch are composed of multiple levels of residual blocks connected in series. In each level of residual block, the two branches share the same set of convolution kernel weights. Independent channel scaling coefficients are set only at the output of the branches to extract amplitude features and phase features synchronously. In each residual block of the improved NAFNet network, a frequency domain self-attention gating unit is set up. The feature map output by the residual block is divided into frequency bands along the time dimension. The feature response intensity in each frequency band is calculated. Frequency band attention weights are generated based on the response intensity of each frequency band. The feature map of the corresponding frequency band is weighted using the frequency band attention weights to obtain the amplitude feature map and phase feature map after frequency domain self-attention gating. Potential energy guided gating units are set in each residual block of the improved NAFNet network. The scattered potential energy distribution data is mapped to the feature map spatial resolution to generate a potential energy weight map. The weight map is then multiplied element-wise with the amplitude feature map and phase feature map after frequency domain self-attention gating. Differential weighting is applied to the feature responses of different potential energy ranges to obtain the amplitude feature map and phase feature map after potential energy guided gating. An amplitude-phase joint reconstruction unit is set at the output of the improved NAFNet network. The amplitude feature map and phase feature map output by the final first-level residual block are concatenated in the channel dimension. An amplitude-phase joint feature map is generated by a set of one-dimensional convolution and channel mixing operations. The defect reflection intensity is calculated on the amplitude-phase joint feature map in spatial position to obtain the defect reflection enhancement value corresponding to each voxel. The defect reflection enhancement values ​​of all voxels constitute the defect reflection enhancement map.

7. The imaging detection method for internal defects in concrete structures based on ultrasonic arrays according to claim 1, characterized in that, The obtained sound velocity field and second-order propagation curvature field of the detection area form a sound velocity-curvature coupled field, including: Based on the defect reflection enhancement map, the enhanced echo signals corresponding to each transmitting and receiving transducer channel are organized according to channel number and time sequence to obtain the enhanced echo sequence of each channel. Early arrival candidate points are extracted for the enhanced echo sequence. The method of energy threshold and rising edge consistency is used to determine the set of time points in the time series of each channel that meet the condition of first exceeding the threshold and maintaining an upward trend for a number of consecutive sampling points. These are used as the early arrival candidate point set for the channel. Sparse wavefront clustering is performed on the early arrival wave candidate point set of all channels. The early arrival wave candidate points are associated according to the channel spatial adjacency relationship and the arrival time proximity relationship to form several early arrival wave clusters. For each early arrival wave cluster, the statistical center value of the arrival time of all candidate points in the cluster is calculated, and the center arrival time is assigned to the channel set corresponding to the cluster. The detection area is discretized into multiple spatial voxels, and a set of parameters to be solved is established, which consists of voxel sound velocity parameters and voxel second-order propagation curvature parameters. The propagation path length of each channel at each voxel is determined based on the position of the transmitting transducer, the position of the receiving transducer, and the position of the voxel. The predicted arrival time of each channel is calculated using the voxel sound velocity parameters and the propagation path length. The difference between the center arrival time and the predicted arrival time of each channel is used as the arrival time constraint term, and the smoothness of the change of the second-order propagation curvature parameter between adjacent voxels is used as the curvature regularization term. The set of parameters to be solved is iteratively updated until convergence, and the sound speed field and the second-order propagation curvature field of the detection area are output. The sound speed-curvature coupling field is constructed from the sound speed field and the second-order propagation curvature field.

8. The method for imaging and detecting internal defects in concrete structures based on ultrasonic arrays according to claim 1, characterized in that, The output defect imaging results include: Based on the sound velocity-curvature coupling field, the propagation path from the transmitting transducer to the receiving transducer via voxels is generated. The propagation time at each voxel is calculated based on the sound velocity field and the path length. The propagation time is compared with the arrival time of the early arrival wave center. Voxel points with propagation time deviations within the preset time tolerance range are selected to form an initial propagation path point set. The initial propagation path point set is processed by path association, connecting path points that belong to the same transmitting and receiving transducer channels and are spatially continuous to form path segments, and recording the propagation curvature change information corresponding to each path segment. Based on the propagation curvature change of the path segment, folding and compression processing is performed. Path segments with continuous curvature and consistent direction are spatially folded and projected onto the same propagation trajectory centerline. Path segments that deviate from the center trajectory by more than a preset spatial deviation threshold are deleted. The folded path segments are spatially connected and filtered. Path segments that appear repeatedly in multiple transmit-receive channels and have the same position are counted. Path segments that appear more than a preset path consistency threshold are retained to form the propagation reachability domain. Delay compensation processing is performed on multi-channel ultrasonic echo data within the propagation reach domain. Coherent superposition calculation is performed on the multi-channel echo signals after delay compensation to obtain the defect reflection intensity corresponding to each spatial voxel, thereby generating imaging results of internal defects in concrete structures.

Images