Deep learning-based ultrahigh performance concrete defect vibration frequency identification method and system

Abstract

The present application relates to the technical field of concrete detection, and specifically to a method and system for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning, which comprises: collecting the vibration response signal of an ultra-high performance concrete test piece under wideband sweep excitation, decomposing the vibration response signal by variational modal decomposition to obtain intrinsic modal components and central frequencies, and constructing a time-frequency energy attenuation atlas based on the instantaneous energy attenuation curve of each intrinsic modal component; and extracting the frequency band boundary steepness of each intrinsic modal component based on the time-frequency energy attenuation atlas. The present application uses wideband sweep excitation and a non-contact laser vibration meter, and can quickly collect signals without the need for applying a coupling agent. By introducing coherence spectrum and a sensitive frequency band mask to guide variational modal decomposition, the present application solves the problems of fast signal attenuation caused by the dense matrix of ultra-high performance concrete and the difficulty of penetration by traditional ultrasonic methods, and can capture internal micro-vibration responses, thereby improving the accuracy of detection.

Description

Technical Field

[0001] This invention relates to the field of concrete testing technology, specifically to a method and system for identifying vibration frequencies of defects in ultra-high performance concrete based on deep learning. Background Technology

[0002] Ultra-high performance concrete (UHVPC), with its exceptionally high compressive strength, toughness, and durability, is widely used in long-span bridges, high-rise buildings, and nuclear facility protective structures. However, due to its dense matrix and extremely low water-cement ratio, UHVPC is highly susceptible to internal defects such as microcracks, pores, or inclusions during pouring and curing. These defects, often hidden within the structure, not only weaken its mechanical properties but may also induce sudden failure under long-term loads. Therefore, developing highly sensitive and precise internal defect detection technologies is crucial for ensuring the safety of UHVPC structures.

[0003] However, among existing defect detection methods, firstly, the traditional ultrasonic method attenuates extremely quickly in ultra-high performance concrete, and the signal is often too weak to be distinguished before it reaches the defect. Moreover, it requires the application of a coupling agent, making it difficult to achieve non-contact rapid detection. Secondly, the impact echo method relies on a single frequency stress wave, and for millimeter-level microcracks or pores inside ultra-high performance concrete, its reflection wavelength is too long to identify fine boundaries, often resulting in missed detections. Furthermore, in actual frequency sweep excitation tests, the overall stiffness vibration amplitude of the ultra-high performance concrete specimen itself is often very large, while the nonlinear vibration characteristic amplitude generated by internal defects is extremely small. Traditional Fourier transform or simple filtering methods will average the background noise and defect signal, making weak defect characteristics easy to be masked, leading to missed detection. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for identifying the vibration frequency of defects in ultra-high performance concrete, comprising: Vibration response signals of ultra-high performance concrete specimens under broadband sweep frequency excitation were collected. The intrinsic modal components and center frequencies were obtained by variational mode decomposition of the vibration response signals. A time-frequency energy decay spectrum was constructed based on the instantaneous energy decay curves of each intrinsic modal component. Based on the time-frequency energy attenuation spectrum, the frequency band boundary steepness of each intrinsic mode component is extracted. Using the center frequency as the anchor point and the steepness being greater than the threshold as the condition, the defect-sensitive sub-band and the structural background sub-band are divided. The frequency offset of the energy centroid within the defect-sensitive sub-band is calculated, and the average energy of the structural background sub-band is introduced as a normalization benchmark to obtain the first defect indication parameter. The first defect indication parameter and the defect-free reference parameter corresponding to the structural background subband are differentially divided at each frequency point to obtain the residual spectrum sequence; the residual spectrum sequence is demodulated to extract the modulation sideband features, the sideband features are arranged according to the frequency energy distribution to form a modulation feature vector and its atomic decomposition is performed to obtain sparse atoms and sparse coefficients. A deep analytical convolutional layer is constructed using sparse atoms as interpretable convolutional kernels. The input modulated feature vector is propagated forward, and the energy decay gradient is calculated to determine the vibrational frequency signal of the pure defect. The vibration frequency signal of the pure defect is weighted and corrected based on the sparse coefficient. Multi-scale entropy analysis is then performed on the weighted and corrected signal to construct the entropy value change curve and extract the inflection point frequency and the width of the stable platform, thereby determining the defect depth value and type parameters.

[0005] Preferably, the vibration response signal of ultra-high performance concrete specimens under broadband sweep excitation is acquired, and the intrinsic modal components and center frequencies are obtained by variational mode decomposition of the vibration response signal, including: Under the geometric boundary conditions of ultra-high performance concrete specimens, a broadband sweep excitation device is used to arrange non-contact laser vibrometers along a preset measuring line array. During the sweep excitation process, vibration response signals of each measuring point are collected synchronously to form the original vibration response sequence. Modal coherence indices for each frequency band are calculated based on the original vibration response sequence, generating a coherence spectrum sequence. High coherence frequency bands are identified based on the coherence spectrum sequence, and the frequency bands are dynamically calibrated using the boundary reflection characteristics of the specimen, generating a sensitive frequency band mask sequence. Using the sensitive frequency band mask sequence as a constraint, the initial iteration of variational mode decomposition is performed on the original vibration response sequence. The initial intrinsic mode component sequence and the initial center frequency estimation sequence are extracted by frequency clustering guided by the mask. The energy decay gradient of each mode is calculated based on the initial intrinsic mode component sequence to generate a decay gradient sequence. The initial center frequency estimation sequence is adjusted using the decay gradient sequence as a feedback signal to extract the intrinsic mode components and center frequency.

[0006] Preferably, the sensitive frequency band mask sequence is used as a constraint condition to perform initial iteration of variational mode decomposition on the original vibration response sequence. Initial intrinsic mode component sequences and initial center frequency estimation sequences are extracted through mask-guided frequency clustering, including: Extract the high-confidence frequency band boundaries from the sensitive frequency band mask sequence to generate a frequency domain constrained window sequence; The frequency domain constraint window sequence and the original vibration response sequence are weighted and superimposed in the frequency domain to generate the initial signal sequence driven by the mask. The peak power spectral density of each potential mode is calculated based on the mask-driven initial signal sequence to generate a candidate set of center frequencies; Using the candidate set of center frequencies as the cluster center, frequency stripping is performed on the initial signal sequence driven by the mask to extract the initial intrinsic mode component sequence and the initial center frequency estimation sequence.

[0007] Preferably, the time-frequency energy decay spectrum is constructed based on the instantaneous energy decay curves of each intrinsic mode component, including: For each intrinsic mode component signal, the instantaneous amplitude of the analytical signal is extracted and an instantaneous energy sequence is generated; Based on the timestamp information of the frequency sweep excitation, the instantaneous energy sequence is aligned with the frequency sweep time axis to generate an instantaneous energy decay curve sequence that characterizes the energy dissipation characteristics over time. Envelope extraction and slope analysis are performed on the instantaneous energy decay curve sequence to calculate the energy decay rate of the curve under different time windows and generate a decay characteristic parameter sequence. Using the center frequency of each intrinsic mode component as the abscissa and the attenuation characteristic parameter sequence as the ordinate, the two-dimensional data is mapped to the time-frequency plane to generate a continuous time-frequency energy attenuation spectrum.

[0008] Preferably, the frequency band boundary steepness of each intrinsic mode component is extracted based on the time-frequency energy attenuation spectrum. Using the center frequency as the anchor point and with the steepness exceeding a threshold as a condition, the defect-sensitive sub-band and the structural background sub-band are divided, including: Using the center frequency of each intrinsic mode component as the geometric anchor point, the gradient trend of energy decay with frequency in the neighborhood of the anchor point is extracted on the time-frequency energy decay spectrum to generate an initial steepness sequence characterizing the boundary undulation characteristics. The initial kurtosis sequence is smoothed by filtering to generate a smoothed kurtosis sequence; The maximum points in the smooth kurtosis sequence are extracted as potential frequency band boundary candidate points. The relative offset between the candidate points and the center frequency is calculated to generate the boundary offset feature sequence. Set a steepness threshold, and mark the regions in the boundary offset feature sequence with offsets less than the preset offset threshold and steepness greater than the steepness threshold as high gradient regions, and mark the remaining regions as low gradient regions. Based on the frequency ranges of the high gradient region and the low gradient region, the corresponding frequency bands in the time-frequency energy attenuation spectrum are extracted and defined as the defect-sensitive sub-band and the structural background sub-band.

[0009] Preferably, the energy centroid frequency shift within the defect-sensitive sub-band is calculated, and the average energy of the structural background sub-band is introduced as a normalization benchmark to obtain the first defect indication parameter, including: By traversing the time-frequency energy attenuation spectrum within the defect-sensitive sub-band, extracting the energy amplitude at each frequency point and performing first-order moment calculation, a centroid frequency sequence characterizing the degree of energy accumulation is obtained. The barycenter frequency sequence is compared with the center frequencies of each intrinsic mode component, and the absolute value of the frequency difference is calculated to generate an energy barycenter frequency offset sequence. Extract the time-frequency energy attenuation spectrum within the structural background sub-band, calculate its average energy level throughout the frequency sweep, and generate a background energy mean sequence. Using the background energy mean sequence as the denominator and the energy centroid frequency offset sequence as the numerator, a dimensionless ratio calculation is performed to generate the first defect indicator parameter sequence. The trend term is extracted from the first defect indication parameter sequence, random fluctuation interference is filtered out, and monotonic components are retained to obtain the final first defect indication parameters.

[0010] Preferably, the first defect indication parameter and the defect-free reference parameter corresponding to the structural background sub-band are differentially analyzed at each frequency point to obtain a residual spectrum sequence; the residual spectrum sequence is demodulated to extract modulation sideband features, and the sideband features are arranged according to frequency energy distribution to form a modulation feature vector and decomposed into atoms to obtain sparse atoms and sparse coefficients, including: The first defect indicator parameter sequence is aligned and subtracted point by point on the frequency axis with the pre-stored defect-free reference parameter sequence corresponding to the structural background sub-band to remove the structural background trend term and generate a residual spectrum sequence characterizing abnormal fluctuations. The residual spectrum sequence is demodulated in time and frequency to extract the sideband frequency components and their amplitudes generated by the defect nonlinear modulation, and to generate a set of sideband features containing frequency position and energy intensity. Based on the energy weights of each component in the sideband feature set, the sideband features are rearranged and vector quantized in ascending order of frequency to form a high-dimensional modulation feature vector. A complete time-frequency atom library was constructed. The high-dimensional modulation feature vector was projected and optimized using a sparse decomposition iterative algorithm. Under the premise of satisfying the preset sparsity constraint, the optimal combination of time-frequency basis functions was selected as sparse atoms, and the projection weights corresponding to each atom were calculated as sparse coefficients.

[0011] Preferably, a deep analytical convolutional layer is constructed using sparse atoms as interpretable convolutional kernels, and the input modulated feature vector is propagated forward to calculate the energy decay gradient to determine the pure defect vibrational frequency signal, including: The time-frequency structure parameters of sparse atoms are analyzed, and the time-frequency center, bandwidth and oscillation mode of sparse atoms are mapped to the spatial weight distribution of convolution kernels to generate an initial convolution kernel group with physical meaning. The initial convolutional kernel group is concatenated according to the depth level to construct a deep analytical convolutional layer, and the modulated feature vector is used as input to perform forward propagation operation to extract the feature activation map under the response of each level of convolutional kernel; The energy distribution evolution of the feature activation map is traced along the depth dimension of the convolutional layer, the rate of change of the feature response amplitude with increasing layer depth is calculated, and an energy decay gradient sequence characterizing the signal transmission loss characteristics is generated. Based on the energy decay gradient sequence, the connected path with the smallest decay rate is selected, and the sparse atomic structure corresponding to the path is marked as the dominant sparse atom transport path. By tracing back along the dominant sparse atom propagation path, the phase consistency index is extracted, and signal components with cross-correlation coefficients greater than a preset threshold are retained. Combined with adaptive threshold denoising processing, a pure defect vibration frequency signal is obtained.

[0012] Preferably, the pure defect vibration frequency signal is weighted and corrected according to the sparsity coefficients. Multi-scale entropy analysis is then performed on the weighted and corrected signal to construct an entropy change curve and extract the inflection point frequency and stable platform width. The defect depth and type parameters are then determined, including: The vibration frequency signal of the pure defect and the sparse coefficients are obtained. The sparse coefficients are used as weighting factors to perform weighted superposition correction on the pure signal to highlight the characteristics of high-confidence defects and generate a weighted correction signal sequence. Based on the number of sampling points of the weighted corrected signal sequence, a scale factor sequence is generated: with the signal length as the upper limit, a set of integers at preset intervals is taken as the scale parameter for multi-scale analysis. Based on the scale parameters in the scale factor sequence, the corresponding observation window length is set; the observation window is used to perform sliding truncation and mean calculation on the weighted modified signal sequence, mapping the long sequence to a short sequence, and generating coarse-grained time series at different scales. Calculate the probability distribution of each coarse-grained time series and quantify its disorder level to generate a multi-scale entropy sequence that characterizes the evolution of signal complexity with scale. Using the scale factor sequence as the x-axis and the multi-scale entropy sequence as the y-axis, an entropy change curve reflecting the nonlinear change in signal complexity is constructed. On the entropy change curve, identify the frequency points where the slope changes significantly, and use them as the inflection point frequency to characterize the geometric boundary of the defect; at the same time, identify the intervals where the curve enters the stable segment, and calculate the span of the interval as the width of the stable platform to characterize the morphological stability of the defect. Input the inflection point frequency and the width of the stable platform into a pre-built defect feature mapping library, and output the corresponding defect depth value and type parameters.

[0013] A deep learning-based vibration frequency identification system for ultra-high performance concrete defects, applicable to the aforementioned deep learning-based vibration frequency identification method for ultra-high performance concrete defects, includes: The signal acquisition and decomposition module is used to acquire the vibration response signal of ultra-high performance concrete specimens under broadband sweep frequency excitation, obtain the intrinsic modal components and center frequency by variational mode decomposition of the vibration response signal, and construct a time-frequency energy attenuation spectrum based on the instantaneous energy attenuation curve of each intrinsic modal component. The sensitive sub-band segmentation module is used to extract the frequency band boundary steepness of each intrinsic mode component based on the time-frequency energy attenuation spectrum. It uses the center frequency as the anchor point and the steepness as a condition to divide the defect sensitive sub-band and the structural background sub-band. The defect indication parameter acquisition module is used to calculate the frequency offset of the energy centroid within the defect-sensitive sub-band and introduce the average energy of the structural background sub-band as a normalization benchmark to obtain the first defect indication parameter. The residual demodulation sparse decomposition module is used to perform frequency-point differential analysis between the first defect indication parameter and the defect-free reference parameter corresponding to the structural background subband to obtain the residual spectrum sequence; the residual spectrum sequence is demodulated to extract the modulation sideband features, the sideband features are arranged according to the frequency energy distribution to form a modulation feature vector and its atoms are decomposed to obtain sparse atoms and sparse coefficients. The convolutional signal purification module is used to construct a deep analytical convolutional layer with sparse atoms as interpretable convolutional kernels, input the modulation feature vector for forward propagation, and calculate the energy decay gradient to determine the pure defect vibration frequency signal. The entropy analysis parameter determination module is used to perform weighted correction on the pure defect vibration frequency signal based on the sparse coefficient, perform multi-scale entropy analysis on the weighted corrected signal, construct the entropy value change curve, extract the inflection point frequency and the width of the stable platform, and determine the defect depth value and type parameters.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a wideband sweep frequency excitation and a non-contact laser vibrometer, which can quickly acquire signals without the need for coupling agent. By introducing coherence spectrum and sensitive frequency band mask to guide variational mode decomposition, it solves the problems of rapid signal attenuation caused by the dense matrix of ultra-high performance concrete and the difficulty of penetration by traditional ultrasonic methods. It can capture the tiny internal vibration response and improve the accuracy of detection. This invention constructs a time-frequency energy attenuation spectrum and accurately divides the defect-sensitive sub-band and the structural background sub-band according to the steepness of the frequency band boundary. By calculating the frequency offset of the energy centroid and introducing the mean of the background energy for normalization, the strong background noise interference caused by the overall structural stiffness is effectively eliminated, so that the weak nonlinear vibration characteristics generated by microcracks or pores are no longer masked, and the false detection rate is reduced. This invention directly maps time-frequency atoms obtained from sparse decomposition into physically meaningful convolutional kernels, constructs a deep analytical convolutional layer, and can separate pure defect vibration frequency signals from complex modulation sideband features by tracking energy decay gradients and phase consistency, avoiding the averaging of background noise. After weighting and correcting the pure signal using sparse coefficients, multi-scale entropy analysis is performed. By constructing entropy change curves, the inflection point frequency reflecting the defect boundary and the stable plateau width reflecting the stability of the defect morphology are extracted. Combined with the feature mapping library, the defect depth value and type parameters are directly output, realizing a closed-loop analysis from signal detection to defect quantization. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall system architecture in one embodiment of the present invention.

[0016] In the diagram: 1. Signal acquisition and decomposition module; 2. Sensitive subband division module; 3. Defect indication parameter acquisition module; 4. Residual demodulation sparse decomposition module; 5. Convolution signal purification module; 6. Entropy analysis parameter determination module. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1, please refer to Figure 1 This invention provides a technical solution: a method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning, comprising: S1. Collect the vibration response signal of ultra-high performance concrete specimens under broadband sweep frequency excitation, perform variational mode decomposition on the vibration response signal to obtain intrinsic mode components and center frequencies, and construct a time-frequency energy decay spectrum based on the instantaneous energy decay curves of each intrinsic mode component. S2. Based on the time-frequency energy attenuation spectrum, extract the frequency band boundary steepness of each intrinsic mode component, and divide the defect-sensitive sub-band and the structural background sub-band with the center frequency as the anchor point and the steepness being greater than the threshold as the condition. S3. Calculate the frequency offset of the energy centroid within the defect-sensitive sub-band, and introduce the average energy of the structural background sub-band as a normalization benchmark to obtain the first defect indication parameter. S4. Differentiate the first defect indicator parameter with the defect-free reference parameter corresponding to the structural background sub-band at each frequency point to obtain the residual spectrum sequence; demodulate the residual spectrum sequence to extract the modulation sideband features, arrange the sideband features according to the frequency energy distribution to form a modulation feature vector and decompose it into atoms to obtain sparse atoms and sparse coefficients. S5. Construct a deep analytical convolutional layer using sparse atoms as interpretable convolutional kernels, input the modulation feature vector and propagate it forward, calculate the energy decay gradient and determine the pure defect vibration frequency signal. S6. The pure defect vibration frequency signal is weighted and corrected according to the sparsity coefficient. Multi-scale entropy analysis is performed on the weighted and corrected signal to construct the entropy value change curve and extract the inflection point frequency and the width of the stable platform to determine the defect depth value and type parameters.

[0019] It should be noted that in the process of identifying the vibration frequency of defects in ultra-high performance concrete, the technical effects of each step need to be achieved through the following specific methods: First, the vibration response signal of the concrete specimen under broadband sweep frequency excitation is collected. For example, a vibrator is used to apply a sweep frequency excitation with a frequency continuously varying from 10Hz to 1000Hz to the specimen. At the same time, the vibration response signal of the specimen is recorded by an accelerometer. When performing variational mode decomposition on the signal, the complex vibration signal can be decomposed into multiple intrinsic mode components. Each component corresponds to a specific center frequency. For example, three intrinsic mode components with center frequencies of 50Hz, 200Hz, and 500Hz are obtained. The instantaneous energy decay curve of each component reflects the rate of energy decay of the frequency component over time. By superimposing the instantaneous energy decay curves of all intrinsic mode components in terms of time and frequency, a time-frequency energy decay spectrum can be constructed. The color intensity in the spectrum represents the energy strength, which can intuitively display the energy distribution at different time and frequency points. When extracting the band boundary steepness of each intrinsic modal component based on the time-frequency energy attenuation spectrum, it is necessary to observe the degree of energy change at the band edge. For example, if the center frequency of an intrinsic modal component is 150Hz, the energy in the 140-150Hz range on its left side rises rapidly from 20 units to 100 units, and the energy in the 150-160Hz range on its right side drops sharply from 100 units to 30 units. This abrupt change in edge energy is high steepness. On the other hand, for another component with a center frequency of 300Hz, the energy changes on both sides are gentle and the steepness is low. Using the center frequency as the anchor point, the band with a steepness greater than a preset threshold (such as a steepness exceeding 50 units / Hz) is divided into a defect-sensitive subband, and the band with a lower steepness is divided into a structural background subband. The defect-sensitive subband usually contains abnormal vibration frequencies caused by defects inside the concrete (such as microcracks and pores), while the structural background subband mainly reflects the inherent vibration characteristics of the concrete matrix. When calculating the frequency offset of the energy centroid within the defect-sensitive subband, it is necessary to first determine the centroid position of the energy in that subband (i.e., the energy-weighted average frequency). For example, if the energy centroid of the defect-sensitive subband of a defect-free specimen is 155Hz, and the energy centroid of the currently tested specimen shifts to 160Hz, the offset is 5Hz. The average energy of the structural background subband is introduced as a normalization benchmark. For example, if the average energy of the structural background subband is 80 units, dividing the offset of 5Hz by this benchmark value of 80 units yields the first defect indication parameter of 0.0625Hz / unit. This parameter eliminates the influence of factors such as specimen size and excitation intensity, and can more accurately reflect the defect-induced... The frequency change; when the first defect indication parameter is differentially compared with the defect-free reference parameter corresponding to the structural background sub-band at each frequency point, the parameter value at each frequency point needs to be subtracted. For example, the defect-free reference parameter is 0.05 Hz / unit at 160 Hz, and the current first defect indication parameter is 0.08 Hz / unit at 160 Hz. After differential, a residual spectrum sequence of 0.03 Hz / unit is obtained. The residual spectrum sequence is demodulated to extract the modulation sideband features. For example, the sideband peaks at 155 Hz and 165 Hz appear on both sides of the main frequency of 160 Hz in the residual sequence. This is the feature generated by the defect vibration on the main frequency modulation. The sideband features are arranged from high to low according to the frequency energy distribution to form a modulation feature vector. For example, the sideband with the highest energy is 155 Hz (energy 0.3) and the second highest is 165 Hz (energy 0.25). The vector after arrangement is [0.3, 0.25]. When performing atomic decomposition on the vector, the matching pursuit algorithm is used to select the sparse atoms that best match the vector from the pre-built atom library. For example, two atoms corresponding to 155Hz and 165Hz are selected, with sparsity coefficients of 0.8 and 0.7, respectively. The sparse atoms represent the key feature basis of the defect vibration, and the sparsity coefficient reflects the degree of contribution of each feature. When constructing a deep analytical convolutional layer using sparse atoms as interpretable convolutional kernels, the selected sparse atoms (such as atoms corresponding to 155Hz and 165Hz) are used as the convolutional kernels of the convolutional layer, and the input modulation feature vector is used for forward propagation calculation. The pure defect signal is separated from noise by calculating the energy decay gradient (i.e., the rate of change of energy over time). For example, a significant peak in the energy decay gradient near 160Hz indicates an energy abrupt change caused by a defect at that frequency, thus determining the pure defect vibration frequency signal as the 160Hz main frequency and the 155Hz and 165Hz sidebands. When weighting and correcting the pure defect vibration frequency signal according to the sparsity coefficients, the sparsity coefficients 0.8 and 0.7 are multiplied by... The signal amplitude of the corresponding sideband is used to enhance the weight of key features. Multi-scale entropy analysis is performed on the weighted and corrected signal. The signal entropy value at different time scales (such as 1ms, 5ms, and 10ms) needs to be calculated to construct the entropy value change curve. For example, the entropy value first decreases rapidly with the increase of scale, and an inflection point appears at the scale of 5ms (the entropy value drops from 1.2 to 0.8). After that, the entropy value remains stable in the range of 5ms-8ms (platform width of 3ms). By matching with the pre-trained defect feature library, the inflection point frequency corresponds to the defect type (such as the 5ms inflection point corresponding to microcracks), and the stable platform width corresponds to the defect depth (such as the 3ms platform width corresponding to a 2cm depth), thereby determining the defect depth value and type parameters.

[0020] In an optional embodiment, the vibration response signal of an ultra-high performance concrete specimen under broadband sweep excitation is acquired, and the intrinsic modal components and center frequencies are obtained by variational mode decomposition of the vibration response signal, including: Under the geometric boundary conditions of ultra-high performance concrete specimens, a broadband sweep excitation device is used to arrange non-contact laser vibrometers along a preset measuring line array. During the sweep excitation process, vibration response signals of each measuring point are collected synchronously to form the original vibration response sequence. It should be noted that signal acquisition under the geometric boundary conditions of ultra-high performance concrete specimens refers to arranging the testing equipment according to the actual shape and support method of the specimen. For example, for a long strip-shaped ultra-high performance concrete simply supported beam specimen, its two ends are placed on fixed supports to simulate the geometric boundary. A broadband sweep frequency excitation device (such as an electromagnetic vibrator) is placed close to the surface of the specimen, and a non-contact laser vibrometer is placed along a preset array of measuring lines (such as arranging a measuring point every 10 cm from left to right). When the vibrator emits a sweep frequency signal with continuously changing frequency (such as from 10 Hz to 1000 Hz), the laser vibrometer synchronously records the vibration velocity or displacement of each measuring point at different times. These continuous time series data are arranged according to the measuring point and time sequence to form the original vibration response sequence. Modal coherence indices for each frequency band are calculated based on the original vibration response sequence, generating a coherence spectrum sequence. High coherence frequency bands are identified based on the coherence spectrum sequence, and the frequency bands are dynamically calibrated using the boundary reflection characteristics of the specimen, generating a sensitive frequency band mask sequence. It should be noted that calculating the modal coherence index for each frequency band based on the original vibration response sequence refers to analyzing the similarity of vibration waveforms at different measuring points at the same frequency. For example, calculating the coherence of vibration signals from measuring points A and B at 150Hz: if the waveforms are almost synchronized, the coherence index is close to 1; if the waveforms are chaotic, the coherence index is close to 0. Generating a coherence spectrum sequence involves connecting the coherence values ​​of all frequency points into a curve. Identifying high-coherence frequency bands based on this curve means finding the frequency range where the coherence value is greater than a preset threshold (e.g., 0.8), for example, finding the 150Hz frequency band. The coherence between 150 Hz and 250 Hz is generally high. Dynamic calibration of the frequency band using the reflection characteristics of the specimen boundary refers to the consideration that the reflection of stress waves at the specimen boundary (such as cracks or end faces) will cause interference, and the frequency band range needs to be corrected according to the arrival time of the reflected waves. For example, the originally identified high coherence frequency band is 150 to 250 Hz, but according to the boundary reflection characteristics, it is found that the frequency band above 240 Hz includes reflected wave interference. Therefore, the frequency band is calibrated to 150 Hz to 235 Hz, and a sensitive frequency band mask sequence is generated. This sequence is marked as valid in the range of 150 Hz to 235 Hz, and the rest of the range is marked as invalid. Using the sensitive frequency band mask sequence as a constraint, the initial iteration of variational mode decomposition is performed on the original vibration response sequence. The initial intrinsic mode component sequence and the initial center frequency estimation sequence are extracted by frequency clustering guided by the mask. It should be noted that using the sensitive frequency band mask sequence as a constraint for the initial iteration of variational mode decomposition means that in the first step of signal decomposition, the algorithm is forced to focus only on the effective frequency band marked by the mask. For example, if the sensitive frequency band mask sequence specifies that only 190Hz to 240Hz is the range of interest, then during the iteration process, the algorithm will ignore all frequency components outside this range. Extracting the initial intrinsic mode component sequence through mask-guided frequency clustering means that, according to the range defined by the mask, components with similar frequencies in the signal are grouped into one category. For example, in the range of 190Hz to 240Hz, the algorithm automatically clusters fluctuations near 200Hz into one category and fluctuations near 220Hz into another category, with each category forming an initial intrinsic mode component. At the same time, the center frequency estimate of each category (such as 200Hz and 220Hz) is recorded to form an initial center frequency estimate sequence. The energy decay gradient of each mode is calculated based on the initial intrinsic mode component sequence to generate a decay gradient sequence; the initial center frequency estimation sequence is adjusted using the decay gradient sequence as a feedback signal to extract the intrinsic mode components and center frequency. It should be noted that calculating the energy decay gradient of each mode based on the initial intrinsic mode component sequence refers to calculating the rate at which the energy of each intrinsic mode component decreases over time. For example, for a mode component with a center frequency of 220 Hz, the energy rapidly decreases from 100 units to 40 units in the first 10 milliseconds after the excitation stops, but only decreases to 30 units in the next 10 to 20 milliseconds. This rate of change, which is fast at first and then slow, is the energy decay gradient. Generating the decay gradient sequence involves connecting the gradient values ​​at different time intervals into a curve. Adjusting the initial center frequency estimation sequence using this sequence as a feedback signal means that if an abnormal energy decay is found at a certain frequency (such as excessively slow decay, which may be due to noise), its center frequency is fine-tuned. For example, if the initial estimated center frequency is 220 Hz, but its energy decay gradient is significantly lower than that of adjacent frequencies, it indicates that the frequency may be impure. Therefore, the center frequency is fine-tuned to 221 Hz, making the energy decay of the newly extracted intrinsic mode components more consistent with physical laws, and finally determining the accurate intrinsic mode components and center frequencies.

[0021] In an optional embodiment, the sensitive frequency band mask sequence is used as a constraint condition to perform initial iteration of variational mode decomposition on the original vibration response sequence. Initial intrinsic mode component sequences and initial center frequency estimation sequences are extracted through mask-guided frequency clustering, including: Extract the high-confidence frequency band boundaries from the sensitive frequency band mask sequence to generate a frequency domain constrained window sequence; It should be noted that the high-confidence frequency band boundary refers to the frequency cutoff point that has been verified multiple times and is highly likely to contain defect information. For example, through testing multiple sets of specimens, it was found that the vibration caused by the defect is mainly concentrated between 195Hz and 245Hz, and the boundary of this range fluctuates very little in repeated tests. Therefore, 195Hz and 245Hz are used as high-confidence boundaries. Generating a frequency domain constraint window sequence involves constructing a window function that has a value of 1 only within the boundary and a value of 0 outside the boundary, which is used for subsequent signal filtering. For example, the generated window sequence on the frequency axis has 0 for 194Hz and below, 1 for 195 to 245Hz, and 0 for 246Hz and above. The frequency domain constraint window sequence and the original vibration response sequence are weighted and superimposed in the frequency domain to generate the initial signal sequence driven by the mask. It should be noted that weighting and superimposing the frequency-domain constraint window sequence with the original vibration response sequence in the frequency domain to generate the mask-driven initial signal sequence means transforming the original signal to the frequency domain and then multiplying it with the window sequence. For example, if the original signal contains components from 0 to 1000 Hz in the frequency domain, after multiplying it with the aforementioned window sequence, the amplitudes of frequency components below 195 Hz and above 245 Hz are forcibly reduced to near 0, while components in the range of 195 to 245 Hz retain their original values ​​or are amplified, thereby generating a mask-driven initial signal sequence that only contains information of the sensitive frequency band. For example, if the original amplitude at 200 Hz is 50 units, it will still be 50 units after superposition, while the amplitude at 300 Hz will be suppressed from 30 units to 0.1 units. The peak power spectral density of each potential mode is calculated based on the mask-driven initial signal sequence to generate a candidate set of center frequencies; It should be noted that calculating the peak power spectral density of each potential mode based on the initial signal sequence driven by the mask and generating a candidate set of center frequencies means analyzing the energy distribution of the signal in the frequency domain and finding the frequency points where the energy is most concentrated. For example, spectral analysis of the initial signal driven by the mask reveals three obvious energy peaks at 198Hz, 220Hz, and 242Hz. The frequencies corresponding to these peaks are the potential center frequencies. Generating a candidate set of center frequencies involves collecting these three peak frequencies (198Hz, 220Hz, and 242Hz) as candidate cluster centers. Using the candidate set of center frequencies as the cluster center, frequency stripping is performed on the initial signal sequence driven by the mask to extract the initial intrinsic mode component sequence and the initial center frequency estimation sequence. It should be noted that using the candidate set of center frequencies as cluster centers to perform frequency stripping on the initial signal sequence driven by the mask, and extracting the initial intrinsic mode component sequence and the initial center frequency estimation sequence, means separating the corresponding signal component with each frequency in the candidate set as the core. For example, using 220Hz as the cluster center, a small frequency range (such as 210Hz to 230Hz) is set, and the signal components within this range are extracted from the mixed signal as an independent initial intrinsic mode component. Similarly, two other components are extracted with 198Hz and 242Hz as centers. After extraction, the center frequencies of these three components (220Hz, 198Hz, and 242Hz) constitute the initial center frequency estimation sequence. For example, in this way, the originally mixed vibration signal is split into three relatively pure narrowband signals, each corresponding to a different physical mode.

[0022] In an optional embodiment, a time-frequency energy decay map is constructed based on the instantaneous energy decay curves of each intrinsic mode component, including: For each intrinsic mode component signal, the instantaneous amplitude of the analytical signal is extracted and an instantaneous energy sequence is generated; It should be noted that extracting the instantaneous amplitude of the analytical signal from each intrinsic modal component signal and generating an instantaneous energy sequence refers to processing each single-frequency vibration signal obtained from the decomposition, obtaining its vibration intensity at each moment, and converting it into an energy value. For example, for an intrinsic modal component with a center frequency of 200Hz, its instantaneous amplitude at moment 1 second is 0.5mm, and at moment 1.1 seconds it is 0.45mm. The amplitude at each moment is squared to represent the energy magnitude, and then arranged in chronological order to form a series of energy values ​​that change with time, such as 0.25, 0.2025, 0.16, etc. This sequence is the instantaneous energy sequence. Based on the timestamp information of the frequency sweep excitation, the instantaneous energy sequence is aligned with the frequency sweep time axis to generate an instantaneous energy decay curve sequence that characterizes the energy dissipation characteristics over time. It should be noted that aligning the instantaneous energy sequence with the frequency sweep time axis based on the timestamp information of the frequency sweep excitation to generate an instantaneous energy decay curve sequence that characterizes the energy dissipation characteristics over time means using precise time markers at the start and end of the frequency sweep excitation to map the energy data to the corresponding excitation time points, thereby observing the energy consumption over time. For example, if the frequency sweep excitation starts at 0 seconds, the excitation frequency is 10Hz at 0 seconds, and the collected energy is 100 units. At 1 second, the excitation frequency changes to 500Hz, and the energy is 80 units. At 2 seconds, the excitation frequency is 1000Hz, and the energy drops to 50 units. By aligning the timestamps in this way, the energy values ​​at different time points are connected to form a curve in which the energy gradually decreases as the frequency sweep time progresses. Each of the multiple intrinsic mode components forms such a curve, and together they constitute the instantaneous energy decay curve sequence. Envelope extraction and slope analysis are performed on the instantaneous energy decay curve sequence to calculate the energy decay rate of the curve under different time windows and generate a decay characteristic parameter sequence. It should be noted that performing envelope extraction and slope analysis on the instantaneous energy decay curve sequence, calculating the energy decay rate of the curve under different time windows, and generating a decay characteristic parameter sequence means first outlining the overall contour of the energy decay curve, and then calculating the steepness of its descent segment by segment. For example, selecting a time window of 100ms, within the window of 0 to 100ms, the energy rapidly decreases from 100 units to 60 units, and the calculated decay rate for this segment is 400 units per second; within the window of 100 to 200ms, the energy slowly decreases from 60 units to 50 units, and the decay rate is 100 units per second. Arranging the rate values ​​calculated for these different time periods in sequence forms a decay characteristic parameter sequence, which quantifies the characteristics of the rate of energy dissipation. Using the center frequency of each intrinsic mode component as the abscissa and the attenuation characteristic parameter sequence as the ordinate, the two-dimensional data is mapped to the time-frequency plane to generate a continuous time-frequency energy attenuation spectrum. It should be noted that mapping two-dimensional data onto the time-frequency plane using the center frequency of each intrinsic mode component as the x-axis and the attenuation characteristic parameter sequence as the y-axis to generate a continuous time-frequency energy attenuation spectrum means constructing a coordinate system where the horizontal axis represents frequency and the vertical axis represents the energy attenuation characteristics. The calculation results are then filled into this system to form an image. For example, center frequency points such as 50Hz, 200Hz, and 500Hz are marked on the horizontal axis, and the corresponding attenuation rate values ​​are marked on the vertical axis. For a 50Hz mode component, its attenuation rate is 50 units per second, so the corresponding color or grayscale value is filled at the intersection of the 50Hz horizontal and 50 vertical axes in the spectrum. For a 200Hz mode component, its attenuation rate is 200 units per second, so a different color is filled at 200Hz. After filling in the data for all frequency points, a color spectrum that can simultaneously display the frequency distribution and energy attenuation characteristics is formed. The intensity of the color in the spectrum represents the magnitude of the attenuation rate.

[0023] In an optional embodiment, the frequency band boundary steepness of each intrinsic mode component is extracted based on the time-frequency energy attenuation spectrum. Using the center frequency as the anchor point and the steepness being greater than a threshold as a condition, a defect-sensitive sub-band and a structural background sub-band are divided, including: Using the center frequency of each intrinsic mode component as the geometric anchor point, the gradient trend of energy decay with frequency in the neighborhood of the anchor point is extracted on the time-frequency energy decay spectrum to generate an initial steepness sequence characterizing the boundary undulation characteristics. It should be noted that extracting gradient trends and generating an initial kurtosis sequence using the center frequency of each intrinsic mode component as a geometric anchor point means taking the center frequency position of each intrinsic mode component as the center point and observing the degree of energy change with frequency on both sides of the center point in the time-frequency energy decay spectrum. For example, if the center frequency of an intrinsic mode component is 200Hz, and the energy distribution in the range of 190Hz to 210Hz is observed in the spectrum, if the energy rises rapidly from 195Hz to 200Hz and drops sharply from 200Hz to 205Hz, this drastic change represents the boundary undulation characteristics. Generating an initial kurtosis sequence involves quantifying the degree of change at these frequency points into numerical values, such as a kurtosis of 0.8 at 195Hz, 0.9 at 200Hz, and 0.85 at 205Hz, forming a series of numerical values ​​reflecting the boundary undulations. The initial kurtosis sequence is smoothed by filtering to generate a smoothed kurtosis sequence; It should be noted that smoothing the initial kurtosis sequence to generate a smooth kurtosis sequence refers to averaging the quantized values ​​to eliminate spurious fluctuations caused by random noise. For example, if the values ​​of three consecutive points in the initial kurtosis sequence are 0.8, 0.9, and 0.3, where 0.3 may be noise interference, smoothing filtering (such as moving average) can adjust these three points to 0.6, 0.65, and 0.6, making the curve smoother and more realistic, removing sharp spikes, and generating a smooth kurtosis sequence that better reflects the true frequency band boundary characteristics. The maximum points in the smooth kurtosis sequence are extracted as potential frequency band boundary candidate points. The relative offset between the candidate points and the center frequency is calculated to generate the boundary offset feature sequence. It should be noted that extracting the maximum points in the smoothed kurtosis sequence as potential frequency band boundary candidate points and calculating the boundary offset feature sequence means finding the peak positions on the smoothed curve. These peaks correspond to the frequency points with the most drastic energy changes, i.e., the potential frequency band boundaries. For example, in the smoothed kurtosis sequence, two obvious peaks appear at 185Hz to the left of the 200Hz center frequency and at 215Hz to the right. These two points are potential frequency band boundary candidate points. Calculating the relative offset of the candidate points from the center frequency means calculating the difference of 15Hz between 185Hz and 200Hz, and the difference of 15Hz between 215Hz and 200Hz. Arranging these offset distances in order forms the boundary offset feature sequence, which records the distance of the boundary from the center frequency. Set a steepness threshold, and mark the regions in the boundary offset feature sequence with offsets less than the preset offset threshold and steepness greater than the steepness threshold as high gradient regions, and mark the remaining regions as low gradient regions. It should be noted that setting a steepness threshold and marking regions with offsets less than the preset offset threshold and steepness greater than the steepness threshold as high-gradient regions means establishing a screening criterion. Only regions that are not too far from the center frequency (small offset) and change very drastically (large steepness) are considered valid boundaries. For example, setting the offset threshold to 20Hz and the steepness threshold to 0.7, at 215Hz, the offset is 15Hz, which is less than 20Hz, and the steepness is 0.8, which is greater than 0.7, so this region is marked as a high-gradient region. At 250Hz, although the steepness may be relatively large, the offset reaches 50Hz, exceeding the 20Hz limit, so it is not marked. The remaining regions that do not meet these two conditions are marked as low-gradient regions. Based on the frequency ranges of the high gradient region and the low gradient region, the corresponding frequency bands in the time-frequency energy attenuation spectrum are extracted and defined as the defect-sensitive sub-band and the structural background sub-band. It should be noted that defining the corresponding frequency bands as the defect-sensitive sub-band and structural background sub-band based on the frequency ranges of the high and low gradient regions means dividing the specific frequency ranges on the time-frequency energy attenuation spectrum according to the marked regions. For example, the high gradient region is concentrated between 205Hz and 225Hz, and vibrations in this frequency band are most sensitive to defects, so it is defined as the defect-sensitive sub-band. The low gradient region is distributed between 180Hz and 205Hz and between 225Hz and 250Hz, and these frequency bands mainly reflect the overall characteristics of the concrete matrix, so they are defined as the structural background sub-band. In this way, the complex vibration signal is physically segmented on the frequency axis.

[0024] In an optional embodiment, the energy centroid frequency offset within the defect-sensitive sub-band is calculated, and the average energy of the structural background sub-band is introduced as a normalization benchmark to obtain a first defect indication parameter, including: By traversing the time-frequency energy attenuation spectrum within the defect-sensitive sub-band, extracting the energy amplitude at each frequency point and performing first-order moment calculation, a centroid frequency sequence characterizing the degree of energy accumulation is obtained. It should be noted that extracting the energy amplitude at each frequency point and calculating the first moment by traversing the time-frequency energy attenuation spectrum within the defect-sensitive sub-band refers to finding the equilibrium point of energy distribution within the selected defect-sensitive frequency range. For example, if the defect-sensitive sub-band covers 205Hz to 225Hz, on the time-frequency spectrum, the energy amplitude is 10 units at 205Hz, 50 units at 210Hz, 100 units at 215Hz, 80 units at 220Hz, and 30 units at 225Hz. When calculating the first moment, each frequency value is multiplied by its corresponding energy amplitude, all products are added together, and finally divided by the sum of all energy amplitudes. The result, 217Hz, is the location where the energy is most concentrated at that moment, which is the barycenter frequency. Repeating this operation for each time point in the spectrum yields a barycenter frequency sequence that varies over time. The barycenter frequency sequence is compared with the center frequencies of each intrinsic mode component, and the absolute value of the frequency difference is calculated to generate an energy barycenter frequency offset sequence. It should be noted that the absolute value of the frequency difference calculated by comparing the barycenter frequency sequence with the center frequencies of each intrinsic mode component refers to the degree of deviation between the observed energy accumulation point and the theoretical center position. For example, the theoretical center frequency of a certain intrinsic mode component is fixed at 220Hz, while the calculated barycenter frequency at a certain moment is 217Hz, with a difference of 3Hz; at another moment, the barycenter frequency changes to 225Hz, with a difference of 5Hz; taking the absolute value of these differences and arranging them in chronological order generates the energy barycenter frequency offset sequence, where a larger value indicates a more severe frequency offset. Extract the time-frequency energy attenuation spectrum within the structural background sub-band, calculate its average energy level throughout the frequency sweep, and generate a background energy mean sequence. It should be noted that extracting the time-frequency energy attenuation spectrum within the structural background sub-band and statistically analyzing the average energy level throughout the frequency sweep refers to calculating the baseline energy value within the frequency range representing the characteristics of the concrete matrix. For example, the structural background sub-band is from 180Hz to 205Hz, and this region mainly reflects the vibration characteristics of the concrete itself. During the frequency sweep, the energy values ​​at all time points within this frequency band are statistically analyzed. For instance, the average energy is 70 units at 180Hz, 80 units at 190Hz, and 90 units at 200Hz. These values ​​are summed and averaged to obtain a background energy average of 80 units. Considering that the energy may fluctuate slightly over time, the calculated background energy average at different times will be slightly different (e.g., 78, 82, 80), thus forming a background energy average sequence. Using the background energy mean sequence as the denominator and the energy centroid frequency offset sequence as the numerator, a dimensionless ratio calculation is performed to generate the first defect indicator parameter sequence. It should be noted that the dimensionless ratio calculation, using the background energy mean sequence as the denominator and the energy centroid frequency offset sequence as the numerator, refers to dividing the degree of frequency offset by the magnitude of the background energy to eliminate the influence of different specimen sizes or excitation intensities. For example, at a certain moment, the energy centroid frequency offset is 3Hz, corresponding to a background energy mean of 80 units. Dividing 3 by 80 yields 0.0375. At another moment, the offset is 5Hz, and the background energy mean is 82 units. Dividing 5 by 82 yields approximately 0.061. Through this division operation, physical quantities with units are converted into pure numerical values, and the generated sequence is the first defect indicator parameter sequence. The larger the value, the more obvious the defect characteristics. The trend term is extracted from the first defect indication parameter sequence, random fluctuation interference is filtered out, and monotonic components are retained to obtain the final first defect indication parameters; It should be noted that extracting the trend term and filtering out random fluctuations in the first defect indicator parameter sequence means removing random jumps in the data and retaining only the overall direction of change. For example, the original first defect indicator parameter sequence is 0.0375, 0.061, 0.04, and 0.07, where 0.04 may be a drop caused by random noise. By using a trend extraction algorithm (such as moving average) to calculate the average of several adjacent points to smooth the curve, filtering out instantaneous random fluctuations, and finally retaining the monotonically rising component, a smooth sequence of 0.03, 0.05, and 0.07 is obtained. This final value is the first defect indicator parameter after filtering out interference, which can more accurately reflect the true development state of the defect.

[0025] In an optional embodiment, the first defect indication parameter and the defect-free reference parameter corresponding to the structural background subband are differentially analyzed at each frequency point to obtain a residual spectrum sequence; the residual spectrum sequence is demodulated to extract modulation sideband features, and the sideband features are arranged according to frequency energy distribution to form a modulation feature vector and decomposed into atoms to obtain sparse atoms and sparse coefficients, including: The first defect indicator parameter sequence is aligned and subtracted point by point on the frequency axis with the pre-stored defect-free reference parameter sequence corresponding to the structural background sub-band to remove the structural background trend term and generate a residual spectrum sequence characterizing abnormal fluctuations. It should be noted that subtracting the first defect indication parameter sequence from the pre-stored defect-free reference parameter sequence corresponding to the structural background sub-band point by point on the frequency axis means subtracting the currently tested parameter curve from the standard defect-free curve at the same frequency to eliminate the inherent response of the concrete matrix itself. For example, at a frequency of 190Hz, the currently tested first defect indication parameter is 0.05, while the pre-stored defect-free reference parameter is 0.02, and the difference is 0.03; at 200Hz, the current parameter is 0.06, and the reference parameter is 0.02, and the difference is 0.04; at 210Hz, the current parameter is 0.04, and the reference parameter is 0.01, and the difference is 0.03. Arranging these differences in frequency order forms a residual spectrum sequence that fluctuates around zero, where non-zero values ​​represent abnormal fluctuations deviating from the normal background. The residual spectrum sequence is demodulated in time and frequency to extract the sideband frequency components and their amplitudes generated by the defect nonlinear modulation, and to generate a set of sideband features containing frequency position and energy intensity. It should be noted that time-frequency demodulation of the residual spectral sequence to extract sideband frequency components and their amplitudes refers to further processing the residual signal to separate the sideband components generated by nonlinear modulation caused by defects. For example, if the residual sequence exhibits abnormal fluctuations near the 200Hz main frequency, demodulation analysis reveals two distinct amplitude peaks at 195Hz to the left and 205Hz to the right of the main frequency. These two peaks are the modulation sideband features. The specific frequency positions (195Hz, 205Hz) and corresponding energy intensities (e.g., amplitudes of 0.25, 0.3) of these sidebands are extracted, and these paired frequency and energy data are combined to generate a set of sideband features containing both frequency positions and energy intensities. Based on the energy weights of each component in the sideband feature set, the sideband features are rearranged and vector quantized in ascending order of frequency to form a high-dimensional modulation feature vector. It should be noted that rearranging and vectorizing the sideband features in ascending frequency order based on the energy weights of each component in the sideband feature set refers to sorting the sideband components according to their energy magnitude and converting them into vectors of a uniform format. For example, if the sideband feature set contains three components: 195Hz (energy 0.25), 200Hz (energy 0.3), and 205Hz (energy 0.25), the component at 200Hz has the highest weight based on its energy. After sorting in ascending frequency order, the energy values ​​of these three components are arranged sequentially or normalized to form a three-dimensional high-dimensional modulation feature vector, such as [0.25, 0.3, 0.25] or the normalized [0.29, 0.42, 0.29]. This vector describes the distribution pattern of the sidebands in digital form. A complete time-frequency atom library was constructed. The projection optimization of the high-dimensional modulation feature vector was performed using the sparse decomposition iterative algorithm. Under the premise of satisfying the preset sparsity constraint, the optimal combination of time-frequency basis functions was selected as sparse atoms, and the projection weights corresponding to each atom were calculated as sparse coefficients. It should be noted that constructing an overcomplete time-frequency atom library and using a sparse decomposition iterative algorithm for projection optimization refers to pre-establishing a database containing a large number of time-frequency basis functions of different shapes, and iteratively calculating to find a few basis functions that best match the current feature vector. For example, the constructed overcomplete atom library contains various basis functions such as Gaussian atoms, sine atoms, and chirp atoms. The sparse decomposition algorithm (such as matching pursuit) compares the high-dimensional modulation feature vector with the atoms in the atom library. Under the condition of satisfying the preset sparsity constraints (such as only allowing the selection of 3 atoms), it selects the combination of time-frequency basis functions that are most similar to the feature vector. For example, the algorithm selects Gaussian atoms that match the 195Hz sideband and sine atoms that match the 205Hz sideband as sparse atoms, and calculates the projection weights of these two atoms in the synthesized signal (such as 0.8 and 0.7) as sparse coefficients. The larger the sparse coefficient, the greater the contribution of the atom to the signal.

[0026] In an optional embodiment, a deep analytical convolutional layer is constructed using sparse atoms as interpretable convolutional kernels. The input modulated feature vector is forward-propagated, and the energy decay gradient is calculated to determine the pure defect vibrational frequency signal, including: The time-frequency structure parameters of sparse atoms are analyzed, and the time-frequency center, bandwidth and oscillation mode of sparse atoms are mapped to the spatial weight distribution of convolution kernels to generate an initial convolution kernel group with physical meaning. It should be noted that analyzing the time-frequency structural parameters of sparse atoms and mapping them to the spatial weight distribution of convolution kernels means transforming the physical characteristics of the specific time-frequency basis functions obtained in the previous step into the weight values ​​of the filters in the convolutional neural network. For example, a sparse atom is a Gaussian time-frequency atom with a center frequency of 200Hz and a bandwidth of 10Hz, characterized by high energy in the middle and low energy at both ends. When generating the initial convolution kernel set, the weight corresponding to the 200Hz position in the convolution kernel is set to the maximum value, and the weights corresponding to the 195Hz and 205Hz positions are successively reduced, forming a weight distribution array with a prominent middle and smooth transition on both sides. This array is the initial convolution kernel with physical meaning. It is no longer a randomly initialized value, but directly corresponds to the specific time-frequency morphology of defect vibration. The initial convolutional kernel group is concatenated according to the depth level to construct a deep analytical convolutional layer, and the modulated feature vector is used as input to perform forward propagation operation to extract the feature activation map under the response of each level of convolutional kernel; It should be noted that constructing a deep analytical convolutional layer by concatenating the initial convolutional kernel group according to depth levels and performing forward propagation operation means stacking multiple convolutional kernels with physical meaning into a multi-layer network structure, and inputting the modulation feature vector into it to extract features layer by layer. For example, constructing a three-layer deep analytical convolutional layer, the first layer convolutional kernel is responsible for extracting sideband features, the second layer convolutional kernel is responsible for extracting combined sideband features, and the third layer is responsible for extracting more abstract defect patterns. When the input modulation feature vector is such as 0.25, 0.3, 0.25, the first layer convolutional kernel performs a weighted sum operation on it, outputting a first-layer feature activation map, showing which frequency components are activated. This map is passed as input to the second layer, and the second-layer convolutional kernel performs another weighted operation to generate a more focused feature activation map. As the layers deepen, irrelevant background noise in the map is gradually suppressed, and defect features are amplified layer by layer. The energy distribution evolution of the feature activation map is traced along the depth dimension of the convolutional layer, the rate of change of the feature response amplitude with increasing layer depth is calculated, and an energy decay gradient sequence characterizing the signal transmission loss characteristics is generated. It's important to note that tracing the evolution of the energy distribution of the feature activation map along the depth dimension of the convolutional layer and calculating the energy decay gradient sequence refers to observing the rate at which the amplitude of the feature response changes as the number of network layers increases. For example, in the first layer's feature activation map, the energy amplitude of a defect feature is 1.0; it becomes 0.7 in the second layer and 0.4 in the third layer. Calculating the rate of change between each layer, the energy decays by 30% from the first to the second layer and by approximately 40% from the second to the third layer. Arranging these decay rates in hierarchical order generates the energy decay gradient sequence, which characterizes the signal transmission loss. If the decay gradient of a signal path is very small, it indicates that the signal remains strong in the deep network and is a genuine defect signal rather than noise. Based on the energy decay gradient sequence, the connected path with the smallest decay rate is selected, and the sparse atomic structure corresponding to the path is marked as the dominant sparse atom transport path. It should be noted that selecting the connected path with the smallest decay rate based on the energy decay gradient sequence and marking it as the dominant sparse atom transmission path means finding the path with the minimum energy loss among all possible signal transmission paths. For example, the decay gradient of path A is 10%, 15%, and 20% per layer, while the decay gradient of path B is 30%, 35%, and 40% per layer. Obviously, path A has a smaller decay rate and more stable signal transmission. The system determines path A as the dominant sparse atom transmission path, which means that the set of sparse atom structures corresponding to path A best represents the actual propagation law of defect vibrations in the network. By tracing back along the dominant sparse atom propagation path, the phase consistency index is extracted, and signal components with cross-correlation coefficients greater than a preset correlation coefficient threshold are retained. Combined with adaptive threshold denoising processing, a pure defect vibration frequency signal is obtained. It should be noted that backtracking along the dominant sparse atom propagation path and extracting the phase consistency index refers to searching backward along the selected optimal path to check whether the signals of each layer are aligned in time phase. For example, in the feature maps of each layer of the dominant path, 200Hz signal components are extracted, and the phase difference between them is calculated. If the phase difference of the 200Hz components of the first, second, and third layers is within 10 degrees, it indicates high phase consistency, and the cross-correlation coefficient is greater than a preset threshold of 0.8. These highly correlated signal components are retained, and adaptive threshold denoising is combined with this process. That is, a cutoff line is automatically set according to the noise level of the current signal, and weak and chaotic signals below the cutoff line are filtered out. The final signal obtained is a pure defect vibration frequency signal with noise interference removed and phase locked, such as a clear 200Hz main frequency and its accompanying stable sideband signal.

[0027] In an optional embodiment, the pure defect vibration frequency signal is weighted and corrected according to the sparsity coefficients. Multi-scale entropy analysis is then performed on the weighted and corrected signal to construct an entropy change curve and extract the inflection point frequency and stable platform width. The defect depth value and type parameters are then determined, including: The vibration frequency signal of the pure defect and the sparse coefficients are obtained. The sparse coefficients are used as weighting factors to perform weighted superposition correction on the pure signal to highlight the characteristics of high-confidence defects and generate a weighted correction signal sequence. It should be noted that obtaining the pure defect vibration frequency signal and sparse coefficients, and using the sparse coefficients as weighting factors for weighted superposition correction, refers to weighting the different frequency components in the pure signal according to the magnitude of the sparse coefficients calculated in the previous step. The larger the sparse coefficient, the more representative the frequency component is of the true defect characteristics. For example, the pure defect signal contains a 200Hz main frequency signal and sideband signals of 195Hz and 205Hz, with corresponding sparse coefficients of 0.8, 0.7, and 0.3, respectively. During weighted correction, the amplitude of the 200Hz signal is multiplied by 0.8, the 195Hz signal by 0.7, and the 205Hz signal by 0.3, and then these results are superimposed. The purpose of this is to highlight the high-confidence defect characteristics and suppress the noise components with lower weights. In the generated weighted corrected signal sequence, the 200Hz main frequency characteristic is significantly enhanced, while the sidebands with lower weights are weakened, thus obtaining a purer and more characteristic signal sequence. Based on the number of sampling points of the weighted corrected signal sequence, a scale factor sequence is generated: with the signal length as the upper limit, a set of integers at preset intervals is taken as the scale parameter for multi-scale analysis. It should be noted that generating a scale factor sequence based on the number of sampling points of a weighted corrected signal sequence refers to setting a series of different observation scales according to the total length of the signal; taking a set of integers with a preset interval as the upper limit of the signal length as the scale parameter for multi-scale analysis; for example, if the weighted corrected signal sequence has 1000 sampling points and the preset interval is 10 points, then the scale factor sequence will contain an integer set of 1, 11, 21, 31, etc., up to 991; these scale factors represent the size or step size of the observation window when analyzing the signal, with small scales used to observe details and large scales used to observe the overall trend; Based on the scale parameters in the scale factor sequence, the corresponding observation window length is set; the observation window is used to perform sliding truncation and mean calculation on the weighted modified signal sequence, mapping the long sequence to a short sequence, and generating coarse-grained time series at different scales. It should be noted that setting the corresponding observation window length based on the scale parameters in the scale factor sequence, and using this observation window to perform sliding truncation and mean calculation on the weighted corrected signal sequence, refers to coarsening the signal at different scales. For example, when the scale factor is 10, the observation window length is set to 10 sampling points. For a signal sequence of 1000 points, the first truncation is performed on the 1st to 10th points to calculate the average value, resulting in the first coarsened point. Then, the window is slid one unit to the right to truncate the 2nd to 11th points to calculate the average value, resulting in the second coarsened point, and so on. In this way, the original long sequence of 1000 points is mapped to a short sequence of 100 points. This process generates coarse-grained time series at different scales. The larger the scale, the shorter the sequence and the smoother the signal. Calculate the probability distribution of each coarse-grained time series and quantify its disorder level to generate a multi-scale entropy sequence that characterizes the evolution of signal complexity with scale. It should be noted that calculating the probability distribution of each coarse-grained time series and quantifying its disorder level to generate a multi-scale entropy sequence refers to analyzing the unpredictability of signals under different smoothing levels. For example, for a coarse-grained sequence generated at scale 10, the probability of values ​​falling into different intervals in the sequence is statistically analyzed. If the value distribution is concentrated, the disorder level is low, and the entropy value is small; if the value distribution is dispersed, the disorder level is high, and the entropy value is large. An entropy value is calculated for the coarse-grained sequence corresponding to each scale factor, and these entropy values ​​are arranged in ascending order of scale to generate a multi-scale entropy sequence that characterizes the evolution of signal complexity with scale; for example, the entropy value is 1.5 at scale 1, 1.2 at scale 10, and 0.8 at scale 50. Using the scale factor sequence as the x-axis and the multi-scale entropy sequence as the y-axis, an entropy change curve reflecting the nonlinear change in signal complexity is constructed. It should be noted that constructing an entropy change curve with the scale factor sequence as the x-axis and the multi-scale entropy value sequence as the y-axis refers to plotting the above calculation results into a graph; the x-axis represents the scale of observation, and the y-axis represents the degree of signal disorder; for example, in the graph, the x-axis ranges from 1 to 100, and the y-axis ranges from 0 to 2; connecting the points corresponding to the entropy value of 1.5 at scale 1, the entropy value of 1.2 at scale 10, etc., forms a curve reflecting the nonlinear change in signal complexity; this curve usually decreases as the scale increases, but the rate of decrease will change; On the entropy change curve, identify the frequency points where the slope changes significantly, and use them as the inflection point frequency to characterize the geometric boundary of the defect; at the same time, identify the intervals where the curve enters the stable segment, and calculate the span of the interval as the width of the stable platform to characterize the morphological stability of the defect. It should be noted that identifying the frequency points where the slope of the entropy change curve changes significantly as inflection point frequencies, and simultaneously identifying the intervals where the curve enters a stable phase and calculating the span as the width of the stable plateau, refers to finding key features of the curve's shape. For example, observing the entropy change curve, we find that before scale 20, the curve drops rapidly with a large slope, while after scale 20, the drop becomes very slow, almost horizontal. The frequency point corresponding to scale 20 is then the inflection point frequency where the slope changes significantly. Meanwhile, the interval where the curve enters a stable phase may start from scale 30 and continue to scale 50; the span of this interval, 20, is the width of the stable plateau. The inflection point frequency reflects the boundary characteristics of the defect, while the width of the stable plateau reflects the stability of the defect's shape. Input the inflection point frequency and the width of the stable platform into a pre-built defect feature mapping library, and output the corresponding defect depth value and type parameter; It should be noted that inputting the inflection point frequency and the stable platform width into the pre-built defect feature mapping library to output the corresponding defect depth value and type parameters means converting the extracted features into specific defect information by looking up a table. The pre-built defect feature mapping library stores a large number of correspondences summarized from experimental data. For example, the mapping library records that when the inflection point frequency corresponds to a scale of 20 and the stable platform width is 20, the corresponding defect depth is 1.5 cm and the defect type is microcrack; when the inflection point frequency corresponds to a scale of 50 and the stable platform width is 10, the corresponding defect depth is 3 cm and the defect type is hole. By inputting the currently calculated inflection point frequency and stable platform width into this library for matching, the final judgment result can be directly output, such as determining that the current specimen has a microcrack defect with a depth of 1.5 cm.

[0028] Example 2, please refer to Figure 2 This invention provides a technical solution: a deep learning-based ultra-high performance concrete defect vibration frequency identification system, applicable to the aforementioned deep learning-based ultra-high performance concrete defect vibration frequency identification method, comprising: Signal acquisition and decomposition module 1 is used to acquire the vibration response signal of ultra-high performance concrete specimens under broadband sweep frequency excitation, obtain the intrinsic modal components and center frequency by variational mode decomposition of the vibration response signal, and construct a time-frequency energy attenuation spectrum based on the instantaneous energy attenuation curve of each intrinsic modal component. Sensitive sub-band segmentation module 2 is used to extract the frequency band boundary steepness of each intrinsic mode component based on the time-frequency energy attenuation spectrum, and to segment the defect sensitive sub-band and the structural background sub-band with the center frequency as the anchor point and the steepness being greater than the threshold as the condition. The defect indication parameter acquisition module 3 is used to calculate the frequency offset of the energy centroid within the defect-sensitive sub-band and introduce the average energy of the structural background sub-band as a normalization benchmark to obtain the first defect indication parameter. The residual demodulation sparse decomposition module 4 is used to perform frequency-point differential analysis between the first defect indication parameter and the defect-free reference parameter corresponding to the structural background sub-band to obtain the residual spectrum sequence; the residual spectrum sequence is demodulated to extract the modulation sideband features, the sideband features are arranged according to the frequency energy distribution to form a modulation feature vector and its atoms are decomposed to obtain sparse atoms and sparse coefficients. Convolutional signal purification module 5 is used to construct a deep analytical convolutional layer with sparse atoms as interpretable convolutional kernels, input the modulation feature vector forward propagation, calculate the energy decay gradient to determine the pure defect vibration frequency signal; Entropy analysis parameter determination module 6 is used to perform weighted correction on the pure defect vibration frequency signal based on the sparse coefficient, perform multi-scale entropy analysis on the weighted corrected signal, construct the entropy value change curve and extract the inflection point frequency and stable platform width, and determine the defect depth value and type parameters.

[0029] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for identifying vibration frequencies of defects in ultra-high performance concrete based on deep learning, characterized in that, include: Vibration response signals of ultra-high performance concrete specimens under broadband sweep frequency excitation were collected. The intrinsic modal components and center frequencies were obtained by variational mode decomposition of the vibration response signals. A time-frequency energy decay spectrum was constructed based on the instantaneous energy decay curves of each intrinsic modal component. Based on the time-frequency energy attenuation spectrum, the frequency band boundary steepness of each intrinsic mode component is extracted. Using the center frequency as the anchor point and the steepness being greater than the threshold as the condition, the defect-sensitive sub-band and the structural background sub-band are divided. The frequency offset of the energy centroid within the defect-sensitive sub-band is calculated, and the average energy of the structural background sub-band is introduced as a normalization benchmark to obtain the first defect indication parameter. The first defect indication parameter and the defect-free reference parameter corresponding to the structural background subband are differentially divided at each frequency point to obtain the residual spectrum sequence; the residual spectrum sequence is demodulated to extract the modulation sideband features, the sideband features are arranged according to the frequency energy distribution to form a modulation feature vector and its atomic decomposition is performed to obtain sparse atoms and sparse coefficients. A deep analytical convolutional layer is constructed using sparse atoms as interpretable convolutional kernels. The input modulated feature vector is propagated forward, and the energy decay gradient is calculated to determine the vibrational frequency signal of the pure defect. The vibration frequency signal of the pure defect is weighted and corrected based on the sparse coefficient. Multi-scale entropy analysis is then performed on the weighted and corrected signal to construct the entropy value change curve and extract the inflection point frequency and the width of the stable platform, thereby determining the defect depth value and type parameters.

2. The method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning according to claim 1, characterized in that, Vibration response signals of ultra-high performance concrete specimens under broadband sweep excitation were collected. The intrinsic modal components and center frequencies were obtained by variational mode decomposition of the vibration response signals, including: Under the geometric boundary conditions of ultra-high performance concrete specimens, a broadband sweep excitation device is used to arrange non-contact laser vibrometers along a preset measuring line array. During the sweep excitation process, vibration response signals of each measuring point are collected synchronously to form the original vibration response sequence. Modal coherence indices for each frequency band are calculated based on the original vibration response sequence, generating a coherence spectrum sequence. High coherence frequency bands are identified based on the coherence spectrum sequence, and the frequency bands are dynamically calibrated using the boundary reflection characteristics of the specimen, generating a sensitive frequency band mask sequence. Using the sensitive frequency band mask sequence as a constraint, the initial iteration of variational mode decomposition is performed on the original vibration response sequence. The initial intrinsic mode component sequence and the initial center frequency estimation sequence are extracted by frequency clustering guided by the mask. The energy decay gradient of each mode is calculated based on the initial intrinsic mode component sequence to generate a decay gradient sequence. The initial center frequency estimation sequence is adjusted using the decay gradient sequence as a feedback signal to extract the intrinsic mode components and center frequency.

3. The method for identifying vibration frequencies of defects in ultra-high performance concrete based on deep learning according to claim 2, characterized in that, Using the sensitive frequency band mask sequence as a constraint, the original vibration response sequence is subjected to initial iteration of variational mode decomposition. Initial intrinsic mode component sequences and initial center frequency estimation sequences are extracted through mask-guided frequency clustering, including: Extract the high-confidence frequency band boundaries from the sensitive frequency band mask sequence to generate a frequency domain constrained window sequence; The frequency domain constraint window sequence and the original vibration response sequence are weighted and superimposed in the frequency domain to generate the initial signal sequence driven by the mask. The peak power spectral density of each potential mode is calculated based on the mask-driven initial signal sequence to generate a candidate set of center frequencies; Using the candidate set of center frequencies as the cluster center, frequency stripping is performed on the initial signal sequence driven by the mask to extract the initial intrinsic mode component sequence and the initial center frequency estimation sequence.

4. The method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning according to claim 3, characterized in that, A time-frequency energy decay spectrum is constructed based on the instantaneous energy decay curves of each intrinsic mode component, including: For each intrinsic mode component signal, the instantaneous amplitude of the analytical signal is extracted and an instantaneous energy sequence is generated; Based on the timestamp information of the frequency sweep excitation, the instantaneous energy sequence is aligned with the frequency sweep time axis to generate an instantaneous energy decay curve sequence that characterizes the energy dissipation characteristics over time. Envelope extraction and slope analysis are performed on the instantaneous energy decay curve sequence to calculate the energy decay rate of the curve under different time windows and generate a decay characteristic parameter sequence. Using the center frequency of each intrinsic mode component as the abscissa and the attenuation characteristic parameter sequence as the ordinate, the two-dimensional data is mapped to the time-frequency plane to generate a continuous time-frequency energy attenuation spectrum.

5. The method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning according to claim 4, characterized in that, Based on the time-frequency energy attenuation spectrum, the frequency band boundary steepness of each intrinsic mode component is extracted. Using the center frequency as the anchor point and the steepness being greater than a threshold, a defect-sensitive sub-band and a structural background sub-band are divided, including: Using the center frequency of each intrinsic mode component as the geometric anchor point, the gradient trend of energy decay with frequency in the neighborhood of the anchor point is extracted on the time-frequency energy decay spectrum to generate an initial steepness sequence characterizing the boundary undulation characteristics. The initial kurtosis sequence is smoothed by filtering to generate a smoothed kurtosis sequence; The maximum points in the smooth kurtosis sequence are extracted as potential frequency band boundary candidate points. The relative offset between the candidate points and the center frequency is calculated to generate the boundary offset feature sequence. Set a steepness threshold, and mark the regions in the boundary offset feature sequence with offsets less than the preset offset threshold and steepness greater than the steepness threshold as high gradient regions, and mark the remaining regions as low gradient regions. Based on the frequency ranges of the high gradient region and the low gradient region, the corresponding frequency bands in the time-frequency energy attenuation spectrum are extracted and defined as the defect-sensitive sub-band and the structural background sub-band.

6. The method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning according to claim 5, characterized in that, The energy centroid frequency shift within the defect-sensitive sub-band is calculated, and the average energy of the structural background sub-band is introduced as a normalization benchmark to obtain the first defect indication parameter, including: By traversing the time-frequency energy attenuation spectrum within the defect-sensitive sub-band, extracting the energy amplitude at each frequency point and performing first-order moment calculation, a centroid frequency sequence characterizing the degree of energy accumulation is obtained. The barycenter frequency sequence is compared with the center frequencies of each intrinsic mode component, and the absolute value of the frequency difference is calculated to generate an energy barycenter frequency offset sequence. Extract the time-frequency energy attenuation spectrum within the structural background sub-band, calculate its average energy level throughout the frequency sweep, and generate a background energy mean sequence. Using the background energy mean sequence as the denominator and the energy centroid frequency offset sequence as the numerator, a dimensionless ratio calculation is performed to generate the first defect indicator parameter sequence. The trend term is extracted from the first defect indication parameter sequence, random fluctuation interference is filtered out, and monotonic components are retained to obtain the final first defect indication parameters.

7. The method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning according to claim 6, characterized in that, The first defect indication parameter and the defect-free reference parameter corresponding to the structural background subband are differentially analyzed at each frequency point to obtain the residual spectrum sequence. The residual spectrum sequence is demodulated to extract modulation sideband features. These sideband features are arranged according to their frequency energy distribution to form a modulation feature vector, which is then decomposed into atoms to obtain sparse atoms and sparse coefficients, including: The first defect indicator parameter sequence is aligned and subtracted point by point on the frequency axis with the pre-stored defect-free reference parameter sequence corresponding to the structural background sub-band to remove the structural background trend term and generate a residual spectrum sequence characterizing abnormal fluctuations. The residual spectrum sequence is demodulated in time and frequency to extract the sideband frequency components and their amplitudes generated by the defect nonlinear modulation, and to generate a set of sideband features containing frequency position and energy intensity. Based on the energy weights of each component in the sideband feature set, the sideband features are rearranged and vector quantized in ascending order of frequency to form a high-dimensional modulation feature vector. A complete time-frequency atom library was constructed. The high-dimensional modulation feature vector was projected and optimized using a sparse decomposition iterative algorithm. Under the premise of satisfying the preset sparsity constraint, the optimal combination of time-frequency basis functions was selected as sparse atoms, and the projection weights corresponding to each atom were calculated as sparse coefficients.

8. The method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning according to claim 7, characterized in that, A deep analytical convolutional layer is constructed using sparse atoms as interpretable convolutional kernels. The input modulated feature vector is propagated forward, and the energy decay gradient is calculated to determine the pure defect vibrational frequency signal, including: The time-frequency structure parameters of sparse atoms are analyzed, and the time-frequency center, bandwidth and oscillation mode of sparse atoms are mapped to the spatial weight distribution of convolution kernels to generate an initial convolution kernel group with physical meaning. The initial convolutional kernel group is concatenated according to the depth level to construct a deep analytical convolutional layer, and the modulated feature vector is used as input to perform forward propagation operation to extract the feature activation map under the response of each level of convolutional kernel; The energy distribution evolution of the feature activation map is traced along the depth dimension of the convolutional layer, the rate of change of the feature response amplitude with increasing layer depth is calculated, and an energy decay gradient sequence characterizing the signal transmission loss characteristics is generated. Based on the energy decay gradient sequence, the connected path with the smallest decay rate is selected, and the sparse atomic structure corresponding to the path is marked as the dominant sparse atom transport path. By tracing back along the dominant sparse atom propagation path, the phase consistency index is extracted, and signal components with cross-correlation coefficients greater than a preset threshold are retained. Combined with adaptive threshold denoising processing, a pure defect vibration frequency signal is obtained.

9. The method for identifying the vibration frequency of defects in ultra-high performance concrete based on deep learning according to claim 8, characterized in that, The pure defect vibration frequency signal is weighted and corrected based on sparse coefficients. Multi-scale entropy analysis is then performed on the weighted and corrected signal to construct an entropy change curve and extract the inflection point frequency and stable plateau width. The defect depth and type parameters are then determined, including: The vibration frequency signal of the pure defect and the sparse coefficients are obtained. The sparse coefficients are used as weighting factors to perform weighted superposition correction on the pure signal to highlight the characteristics of high-confidence defects and generate a weighted correction signal sequence. Based on the number of sampling points of the weighted corrected signal sequence, a scale factor sequence is generated: with the signal length as the upper limit, a set of integers at preset intervals is taken as the scale parameter for multi-scale analysis. Based on the scale parameters in the scale factor sequence, the corresponding observation window length is set; the observation window is used to perform sliding truncation and mean calculation on the weighted modified signal sequence, mapping the long sequence to a short sequence, and generating coarse-grained time series at different scales. Calculate the probability distribution of each coarse-grained time series and quantify its disorder level to generate a multi-scale entropy sequence that characterizes the evolution of signal complexity with scale. Using the scale factor sequence as the x-axis and the multi-scale entropy sequence as the y-axis, an entropy change curve reflecting the nonlinear change in signal complexity is constructed. On the entropy change curve, identify the frequency points where the slope changes significantly, and use them as the inflection point frequency to characterize the geometric boundary of the defect; at the same time, identify the intervals where the curve enters the stable segment, and calculate the span of the interval as the width of the stable platform to characterize the morphological stability of the defect. Input the inflection point frequency and the width of the stable platform into a pre-built defect feature mapping library, and output the corresponding defect depth value and type parameters.

10. A deep learning-based ultra-high performance concrete defect vibration frequency identification system, applicable to the deep learning-based ultra-high performance concrete defect vibration frequency identification method according to any one of claims 1-9, characterized in that, include: The signal acquisition and decomposition module is used to acquire the vibration response signal of ultra-high performance concrete specimens under broadband sweep frequency excitation, obtain the intrinsic modal components and center frequency by variational mode decomposition of the vibration response signal, and construct a time-frequency energy attenuation spectrum based on the instantaneous energy attenuation curve of each intrinsic modal component. The sensitive sub-band segmentation module is used to extract the frequency band boundary steepness of each intrinsic mode component based on the time-frequency energy attenuation spectrum. It uses the center frequency as the anchor point and the steepness as a condition to divide the defect sensitive sub-band and the structural background sub-band. The defect indication parameter acquisition module is used to calculate the frequency offset of the energy centroid within the defect-sensitive sub-band and introduce the average energy of the structural background sub-band as a normalization benchmark to obtain the first defect indication parameter. The residual demodulation sparse decomposition module is used to perform frequency-point differential analysis between the first defect indication parameter and the defect-free reference parameter corresponding to the structural background subband to obtain the residual spectrum sequence; the residual spectrum sequence is demodulated to extract the modulation sideband features, the sideband features are arranged according to the frequency energy distribution to form a modulation feature vector and its atoms are decomposed to obtain sparse atoms and sparse coefficients. The convolutional signal purification module is used to construct a deep analytical convolutional layer with sparse atoms as interpretable convolutional kernels, input the modulation feature vector for forward propagation, and calculate the energy decay gradient to determine the pure defect vibration frequency signal. The entropy analysis parameter determination module is used to perform weighted correction on the pure defect vibration frequency signal based on the sparse coefficient, perform multi-scale entropy analysis on the weighted corrected signal, construct the entropy value change curve, extract the inflection point frequency and the width of the stable platform, and determine the defect depth value and type parameters.

Images