A stone masonry ancient building restoration process knowledge base and adaptive application system

Abstract

This invention belongs to the technical field of artificial intelligence and relates to a knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings. It includes: a data acquisition module for applying a combined acoustic and mechanical wave to the stone masonry, extracting energy attenuation spectra and phase features to generate broadband scattering data; a defect modeling module for parsing the data to obtain phase hysteresis boundary definitions, and inversely mapping to generate a three-dimensional wavefield penetration failure mask; a network construction module for loading a dynamic impedance compensation neural network, reshaping physical properties and operational primitives into repair factor nodes; a network pruning module for overwriting the mask as a network weight matrix, and performing forced circuit breaking pruning on physically mismatched nodes according to size thresholds; and a sequence generation module for assigning the mask as a feature matrix, calculating smooth paths, and decoding to generate an adaptive repair parameter sequence. This invention solves the problem of insufficient adaptability of repair schemes due to the lack of deep coupling of physical information between defect diagnosis, scheme formulation, and process selection.

Description

Technical Field

[0001] This invention belongs to the technical field of artificial intelligence and relates to a knowledge base and adaptive application system for the restoration process of ancient stone masonry buildings. Background Technology

[0002] Stone masonry ancient buildings are exposed to the external natural environment for a long time, making them susceptible to the combined effects of temperature variations, humidity erosion, and structural stress. This can easily lead to physical defects such as microscopic cracks and material weathering within the stone masonry, thereby reducing the overall structural stability and durability. Accurately detecting internal defects in stone masonry and applying targeted repair interventions is a crucial aspect of ancient building conservation projects.

[0003] Existing repair methods typically combine manual inspection with traditional non-destructive testing (NDT) techniques. In practice, technicians first conduct an external inspection and assessment, then use NDT equipment based on a single physical principle to scan a localized area, obtaining shallow data reflecting internal structural anomalies. Based on this data, the repair team combines manual interpretation with standard procedures to develop a repair plan, including selecting specific types of grouting materials and setting grouting parameters, and finally carrying out the corresponding construction and routine testing and evaluation.

[0004] However, the aforementioned combined detection and repair techniques have certain limitations when dealing with complex, heterogeneous, and anisotropic masonry structures. On the one hand, detection methods based on a single physical quantity are prone to multiple interpretations when analyzing the detection signals, making it difficult to accurately characterize the true geometric shape and physical properties of microscopic defects. On the other hand, since the formulation of solutions largely relies on manual qualitative interpretation, there is a lack of data-driven, deep technical correlation between the selection of repair processes and the microscopic compatibility of physical defects. This results in obstacles in the matching degree of the selected repair materials in terms of physical properties such as fluid dynamics, making it difficult to effectively penetrate deep into micro-cracks, and thus affecting the bonding state of the interface between new and old materials and the long-term repair quality to some extent. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a knowledge base and adaptive application system for the restoration techniques of ancient stone masonry buildings.

[0006] A knowledge base and adaptive application system for the restoration techniques of ancient stone masonry buildings, comprising: The data acquisition module is used to apply a sweeping sonic boom composite mechanical wave to the target masonry, capture the back-reflected wave train and the far-end penetrating wave train, extract the nonlinear interface energy attenuation spectrum and absolute phase delay characteristics, and merge them to generate broadband back-scattered wave train data. The defect modeling module is used to analyze broadband backscattered wave train data to obtain the phase hysteresis time span, thereby defining the external geometric boundary conditions, and inversely mapping and encapsulating to generate a three-dimensional wave field penetration failure mask. The network construction module is used to load the dynamic impedance compensation artificial neural network, obtain the disassembled independent operation perturbation primitives and the associated physical properties of the repair materials, reshape the independent operation perturbation primitives into coordinate tensors to define atomized repair factor nodes, and format them into an acousto-elastic composite wave impedance compensation tensor. The network pruning module is used to overwrite the three-dimensional wavefield penetration failure mask as a self-attention gating weight matrix of a dynamic impedance compensation artificial neural network. Based on the penetration size threshold in the physical properties of the repair material, it performs a forced circuit breaking pruning operation on the atomized repair factor nodes that do not match the physical boundary of the mask. The sequence generation module is used in the dynamic impedance compensation artificial neural network to characterize the three-dimensional wave field penetration failure mask as a disordered medium scattering feature matrix lattice, calculate the cumulative directed connected path of the smoothed disordered medium scattering feature matrix lattice, and inversely decode the path to generate an adaptive repair parameter sequence.

[0007] In a further embodiment of the present invention, the data acquisition module is configured to perform the following steps: Short-time Fourier transforms are applied to the captured back-reflected wave train and the far-end penetrating wave train to generate time-spectrum diagrams; The power spectral density of the far-end penetrating wave train under discrete frequency bands is calculated to the ratio of the preset power spectral density of the non-destructive healthy medium reference, and the nonlinear interface energy attenuation spectrum is obtained. The initial phase difference between the back-reflected wave train and the preset reference signal is extracted, and phase dewinding operation is performed along the discrete frequency band axis to obtain monotonically continuous absolute phase delay characteristics. The nonlinear interface energy decay spectrum and absolute phase delay features are stacked along the feature dimension and merged to generate broadband backscattered wave train data with a three-dimensional structure tensor.

[0008] In a further embodiment of the present invention, the defect modeling module is configured to perform the following steps: The broadband backscatter wave train data is sliced ​​along the detection time axis, and the absolute phase delay characteristic value under the discrete frequency band is divided by the corresponding angular frequency to calculate the true phase hysteresis time span. The acoustic wave phase velocity under the preset target frequency band is retrieved. The time-of-flight method is used to combine the acoustic wave phase velocity with the actual phase hysteresis time span to calculate and generate the defect depth coordinates, and define the external geometric boundary conditions formed by the spatial coordinate point group. Spatial voxels whose absolute value of phase hysteresis time span is greater than the failure judgment time threshold are selected and, together with the recorded nonlinear interface energy decay spectrum values, are inversely mapped to a preset three-dimensional mesh algorithm framework to encapsulate and generate a three-dimensional wave field penetration failure mask.

[0009] In a further embodiment of the present invention, the network construction module is configured to perform the following steps: The domain-specific language model obtained through unsupervised training is used to process the independent perturbation primitives decomposed from natural language, and the projection generates a preliminary mathematical representation in the latent vector multidimensional space. The preliminary mathematical representation is reshaped by fusing the retrieved fluid dynamics limit attributes using a graph embedding algorithm, generating a coordinate tensor with internalized physical constraints, and defining atomized repair factor nodes independently. A diagonal matrix consisting of diagonal elements is generated for the atomized repair factor nodes with coordinate tensors, and formatted as an acoustic-elastic composite wave impedance compensation tensor simulating energy attenuation and energy coupling effects.

[0010] In a further embodiment of the present invention, the network pruning module is configured to perform the following steps: Analyze the overwritten self-attention gating weight matrix and measure the topological specifications of the microscopic spatial gaps it defines; Compare the topological specifications of the microscopic spatial gaps with the penetration size threshold representing the repair material stored in the atomized repair factor nodes; If the microscopic spatial gap topology exhibits severe low-frequency resonance distortion and high-frequency signal cutoff characteristics, and is physically mismatched with the penetration size threshold, a zero-value attenuation mask slot is derived using a self-attention gating weight matrix. By using a zero-value decay mask slot, the self-attention connection weight between the corresponding atomized repair factor node and its predecessor node is locked to an absolute zero value, cutting off the forward information transmission of the node and completing the forced disconnection pruning operation.

[0011] In a further embodiment of the present invention, the sequence generation module is configured to perform the following steps: Before calculating the optimal path, clear the redundant link branches in the dynamic impedance compensation artificial neural network that have become completely dormant due to forced circuit closure pruning. The disordered medium scattering characteristic matrix lattice characterized by the three-dimensional wave field penetration failure mask is specified as the starting boundary of the latent space solution action. The non-convex optimization solver is used to iterate the unpruned adjacent atomized repair factor node clusters to calculate the cumulative directed connected path. Inversely decode the surviving graph nodes contained in the multiplication of the directed connected path, extract the engineering semantic information and reorganize it into an independent operation perturbation primitive flow; Following the spatiotemporal topological order of neurons, the independent operational perturbation primitive flow is loaded with a fluid perfusion miniaturized control pressure scalar and a millisecond-level waiting time scalar of the heterogeneous material layering solidification gap, and cascaded and encapsulated to generate an adaptive repair parameter sequence.

[0012] A further embodiment of the present invention includes an effect evaluation module, which is configured to perform the following steps: Relying on the hardware communication bus to analyze the adaptive repair parameter sequence, the system issues control commands to drive the arrayed numerical control micro-irrigation equipment to complete the physical intervention action; The waveform of the first detection frequency response bandwidth and frequency modulation slope was replicated, and the secondary reverse reflection wave train after the physical intervention was received synchronously, and the boundary feedback delay parameter of the reflection wave array was calculated from it. The boundary feedback delay parameter of the reflected wave array is subjected to frequency-band interferometric subtraction with the first extracted absolute phase delay feature to output a differential phase signal. After digital bandpass filtering, the low-frequency standing wave energy accumulation distortion is extracted from the differential phase signal and transformed into a shared quantity, which is then used to synthesize residual wave train characteristic data.

[0013] A further embodiment of the present invention also includes a model update module, which is configured to perform the following steps: Extract the residual energy spectrum from the residual wave train characteristic data, and calculate and obtain the second derivative eigenvalues ​​of the residual energy spectrum; Based on the preset inverse dynamics function, the second derivative eigenvalues ​​are transcribed and amplified into the initial impulse of the backward propagation momentum of the final layer of the decision network, forming the physical loss gradient flow; Load the physical loss gradient flow back to the specific atomized repair factor node position that caused the misleading decision, and generate a momentum repulsion vector from the defect anchor point mapped from the current spatial gap topology specification to the problem node based on the magnitude of the physical loss gradient flow; Based on the momentum repulsion vector, the repulsion penalty calculation is performed to push away and update the latent space geometric cluster coordinates of the node group associated with the poor repair effect in the graph manifold space cluster relative to the Euclidean alienation distance of the defect anchor point.

[0014] In a further embodiment of the present invention, the data acquisition module is configured to perform the following steps: Obtain the background vibration noise flow of the environment in which the target masonry is located, excluding test sources; Spectral decomposition of the background vibration noise stream is performed to extract the set of inherent resonant frequencies of the current detection environment; By comparing the environmental resonant frequency set with the preset mechanical wave sweep frequency band, the starting and ending sweep frequencies of the sweep frequency sonic projectile composite mechanical wave are dynamically adjusted to generate an adaptive sweep command that avoids the environmental resonant frequency set and send it to the excitation unit.

[0015] In a further embodiment of the present invention, the network construction module further includes an insulation initialization step before receiving external signals: Instantiate a dynamic impedance compensation artificial neural network based on a high-order undirected graph neural network logic structure in the memory of the computing cluster; In the initial state where no arbitrary mask signal generated by calculation is received, all off-diagonal elements in the adjacency matrix of the network internal graph are forcibly assigned the value of zero; The initial bias of the control unit of the hidden layer graph node is set to the limit of negative infinity, and the active connection channels between all hidden layer graph nodes are forcibly set to a non-connected and open state that does not contain any preset conduction path.

[0016] In summary, the present invention has the following beneficial technical effects: 1. This invention injects a three-dimensional wavefield penetration failure mask into a dynamic impedance compensation artificial neural network and overwrites its topological distribution matrix as the self-attention gating weight matrix of the network's hidden layers, performing forced circuit breaking and pruning operations on physically mismatched repair factor nodes. This setting directly transforms the detected physical defect topological features into numerical constraints of the network computation structure, enabling the adjacency relationships and impedance characteristics in the physical space to be isomorphically mapped to information transmission preferences in the network graph. When a specific repair process is not applicable under physical conditions such as pore permeation, the system can block the computational path at the network level before entering the optimization solution stage, thereby reducing the algorithm's search iterations in the invalid solution space and helping to improve the computational efficiency and convergence of subsequent optimal path solutions.

[0017] 2. This invention, based on a physically pruned dynamic impedance compensation artificial neural network, uses the disordered medium scattering feature matrix lattice represented by the failure mask as the initial boundary. A non-convex optimization solver iteratively optimizes the path to obtain a cumulative directed connected path that smooths the impedance, and then performs inverse decoding. This method transforms the multivariate repair scheme formulation problem into a clear graph path optimization process. By deriving network paths aimed at eliminating wavefield distortion, the system establishes a mapping relationship from the physical representation of the defect to the repair strategy, and then automatically outputs a repair parameter sequence containing quantitative indicators such as specific operation steps, miniaturized control pressure scalars, and waiting time sequences, improving the targeting and data-driven nature of the repair scheme generation.

[0018] 3. This invention collects secondary detection data after physical intervention, synthesizes residual wave train feature data, and transforms it into a physical loss gradient flow in a high-dimensional tensor space. Then, based on repulsive penalty calculation, it updates the latent space geometric aggregation coordinates of the atomized repair factor nodes in the neural network. This mechanism constructs a closed-loop iterative process based on real feedback from physical detection, reducing the subjective dependence on human experience evaluation. By tracing the quantization error calculated from the secondary detection back to the repair factor node that generated the decision, and adjusting the Euclidean distance between the inferior node and the problem anchor point in the graph manifold space, the network model can subsequently change the topological projection weights to optimize the decision allocation of repair options when facing similar defect features, promoting continuous adjustment and evolution of the model. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0020] Figure 1 This is a schematic diagram of the framework in the embodiments of this application.

[0021] Figure 2 This is a flowchart illustrating an embodiment of this application. Detailed Implementation

[0022] The following is in conjunction with the appendix Figures 1-2 A preferred description of the present invention is provided below.

[0023] See attached document Figures 1-2 This invention proposes a knowledge base and adaptive application system for the restoration techniques of ancient stone masonry buildings, including the following modules: The data acquisition module is used to apply a sweeping sonic boom composite mechanical wave to the target masonry, capture the back-reflected wave train and the far-end penetrating wave train, extract the nonlinear interface energy attenuation spectrum and absolute phase delay characteristics, and merge them to generate broadband back-scattered wave train data. The defect modeling module is used to analyze broadband backscattered wave train data to obtain the phase hysteresis time span, thereby defining the external geometric boundary conditions, and inversely mapping and encapsulating to generate a three-dimensional wave field penetration failure mask. The network construction module is used to load the dynamic impedance compensation artificial neural network, obtain the disassembled independent operation perturbation primitives and the associated physical properties of the repair materials, reshape the independent operation perturbation primitives into coordinate tensors to define atomized repair factor nodes, and format them into an acousto-elastic composite wave impedance compensation tensor. The network pruning module is used to overwrite the three-dimensional wavefield penetration failure mask as a self-attention gating weight matrix of a dynamic impedance compensation artificial neural network. Based on the penetration size threshold in the physical properties of the repair material, it performs a forced circuit breaking pruning operation on the atomized repair factor nodes that do not match the physical boundary of the mask. The sequence generation module is used in the dynamic impedance compensation artificial neural network to characterize the three-dimensional wave field penetration failure mask as a disordered medium scattering feature matrix lattice, calculate the cumulative directed connected path of the smoothed disordered medium scattering feature matrix lattice, and inversely decode the path to generate an adaptive repair parameter sequence.

[0024] In one embodiment of the present invention, the data acquisition module is configured to perform the following steps: Short-time Fourier transforms are applied to the captured backscattered wave train and the far-end penetrating wave train to generate time-frequency spectra. The ratio of the power spectral density of the far-end penetrating wave train in the discrete frequency band to the preset power spectral density of the non-destructive healthy medium reference is calculated to obtain the nonlinear interface energy attenuation spectrum. The initial phase difference between the backscattered wave train and the preset reference signal is extracted, and phase dewinding operation is performed along the discrete frequency band axis to obtain monotonically continuous absolute phase delay features. The nonlinear interface energy attenuation spectrum and the absolute phase delay features are stacked along the feature dimension and merged to generate broadband backscattered wave train data with a three-dimensional structure tensor.

[0025] In one embodiment of the present invention, before performing formal active frequency sweep detection, in order to eliminate the interference of the complex environment at the ancient building site, such as surrounding traffic vibrations and wind loads, on the weak acoustic signals, the system first performs an adaptive adjustment step to eliminate environmental background interference.

[0026] The silent monitoring mode is activated to acquire the background vibration noise flow of the environment in which the target masonry is located, which is not a test source. During this stage, the system does not transmit any active mechanical waves to the target masonry, but only turns on the receiving unit of the phased array acoustic missile flaw detection sensor antenna group to record the background vibration time-domain signal of the masonry caused by the slight excitation of the environment within a certain time window.

[0027] The background vibration noise stream is spectrally decomposed to extract the set of inherent resonant frequencies of the current detection environment. Specifically, the data acquisition and processing terminal performs a Fourier transform on the recorded background vibration time-domain signal to generate an environmental noise spectrum. The system sets a dynamic energy threshold and marks frequency bands in the spectrum with amplitudes exceeding this threshold as inherent resonant frequencies caused by environmental factors. For example, the system may identify significant mechanical resonance peaks in the 12kHz to 18kHz frequency band caused by nearby heavy vehicle traffic.

[0028] By comparing the inherent resonant frequency set of the environment with the preset mechanical wave sweep frequency band, the starting sweep frequency of the sweeping acoustic-bullet composite mechanical wave is dynamically adjusted. With the termination frequency of the sweep System analysis revealed a significant overlap between the originally planned 10kHz initial sweep frequency and the extracted 12kHz-18kHz intrinsic resonant frequency set. To prevent low-frequency environmental vibrations from masking the actual defect scattering signal, the terminal used mathematical optimization logic to adjust the initial sweep frequency. The frequency is dynamically adjusted to 20kHz. This generates an adaptive frequency sweep command that avoids the inherent resonant frequency set of the environment, and the updated sweep parameters are sent to the excitation unit. After completing the environmental adaptive calibration, the system enters the formal detection phase.

[0029] In one embodiment of the present invention, a sonic composite wave frequency sweep detection is performed on the target masonry, generating structured broadband backscattered wave train data. First, an embedded data acquisition and processing terminal controls a phased array sonic composite flaw detection sensor antenna group coupled to preset detection grid points on the surface of the target masonry, driving its excitation unit to emit a frequency sweep sonic composite mechanical wave of a predetermined waveform into the interior of the target masonry. This mechanical wave employs a linear frequency modulation mode, with its frequency fluctuating within the sweep period. Internal from the starting frequency Continuous change to the termination frequency After the mechanical wave propagates and interacts with the material interface inside the target masonry, the receiving unit of the phased array acoustic missile flaw detection sensing antenna group and the remotely deployed penetrating wave receiving array synchronously capture the response signal. Specifically, the phased array acoustic missile flaw detection sensing antenna group captures the back-reflected wave train formed by energy scattering and reflection at the heterogeneous weathering interface, while the remotely deployed penetrating wave receiving array captures the remotely penetrating wave train that penetrates the entire masonry medium and records its spatial polar coordinate information.

[0030] The embedded data acquisition and processing terminal digitizes the acquired back-reflected wave train and the far-end transmitted wave train, and applies a short-time Fourier transform to convert them from time-domain signals into a time-spectrum graph. Based on this time-spectrum graph, the terminal further calculates two core features for each discrete frequency band: The first characteristic is the nonlinear interface energy attenuation spectrum, calculated at specific discrete frequency bands. The energy loss is quantified by the ratio of the power spectral density of the far-end penetrating wave train to the power spectral density of the back-reflected wave train.

[0031] The second feature is the absolute phase delay feature, which is obtained by extracting the back-reflected wave train flow in the same discrete frequency band. The phase angle is obtained by subtracting it from the phase angle of a pre-calibrated reference signal, and is used to characterize the acoustic path time lag experienced by the sound wave as it travels to the interface of the internal heterogeneous defect and is reflected.

[0032] The terminal stacks the nonlinear interface energy attenuation spectrum values ​​and absolute phase delay characteristic values ​​calculated under all discrete frequency bands along the characteristic dimension, integrates the data from all detection grid points, and merges the output to have a three-dimensional structure (number of detection points). Number of feature types A broadband backscattered wave train data tensor (number of discrete frequency bands).

[0033] In this embodiment, the nonlinear interface energy decay spectrum The calculation formula is: Meanwhile, absolute phase delay characteristics The calculation formula is: In the formula, The unit is decibel (dB), which is essentially a logarithmic representation of the power ratio. The unit is radians (rad), which represents the difference in phase angles. A phase unwinding mathematical operator. Represents the far-end penetrating wave train flow in discrete frequency bands The power spectral density at that location. The transmission reference power spectral density of a pre-calibrated, intact test crystal or masonry in this frequency band. For the reverse reflected wave train in the discrete frequency band The complex spectrum value at that location. The reference signal is the complex spectral value of the reference signal in this frequency band. The reference signal is the reference waveform data obtained by calibration in a standard homogeneous medium. The function represents the phase angle operation of taking a complex number.

[0034] It should be noted that the frequency range of the swept-frequency sonic projectile composite mechanical wave is set to 10 kHz to 500 kHz. This frequency range is based on the fact that lower frequency components, i.e., 10 kHz to 100 kHz, have better penetration and sensitivity to macroscopic cracks and voids at the centimeter scale, while higher frequency components, i.e., 100 kHz to 500 kHz, can provide high-resolution detection capabilities for microscopic cracks and material weathering interfaces at the millimeter or even sub-millimeter scale.

[0035] For example, suppose at a detection point in granite masonry, the data acquisition terminal drives the probe to transmit data from... Linear frequency sweep to The acoustic projectile wave. When selecting the target frequency band as... Discrete frequency bands at During analysis, the processor performs a short-time Fourier transform on the acquired time-domain sequence to calculate the power spectral density of the far-end penetrating wave train that penetrates to the far-end receiver at that frequency band. for The transmission reference power spectral density of the intact test medium in the pre-calibrated frequency band. for Based on the aforementioned formula, the nonlinear interface energy attenuation spectrum at this frequency is... The calculation result is approximately At the same time, the processor extracts the phase of the back-reflected wave train in that frequency band from the spectral data. And retrieve the reference signal phase under the standard medium from the system's internal calibration database. Therefore, the absolute phase delay characteristic at this frequency is calculated as follows: After performing this calculation process across all discrete frequency bands, a set of attenuation spectral vectors and phase delay vectors for that detection point are obtained. These are then integrated with data from other detection points to ultimately form broadband backscattered wave train data.

[0036] In one embodiment of the present invention, the defect modeling module is configured to perform the following steps: Broadband backscattered wave train data is sliced ​​along the detection time axis. The absolute phase delay characteristic value under discrete frequency bands is divided by the corresponding angular frequency to calculate the true phase hysteresis time span. The sonic projectile wave phase velocity under the preset target frequency band is retrieved. Using the time-of-flight method, the sonic projectile wave phase velocity and the true phase hysteresis time span are combined to calculate and generate defect depth coordinates, defining the outward geometric boundary conditions formed by the spatial coordinate point group. Spatial voxels with an absolute value of phase hysteresis time span greater than the failure judgment time threshold are selected and, together with the recorded nonlinear interface energy attenuation spectrum value, are inversely mapped to the preset three-dimensional mesh algorithm framework to encapsulate and generate a three-dimensional wavefield penetration failure mask.

[0037] Broadband backscattered wave train data is analyzed, and a three-dimensional wavefield penetration failure mask that can intuitively reflect the damage distribution inside the target masonry is constructed. Specifically, the server-side data analysis and modeling module first receives and reads the generated broadband backscattered wave train data tensor. For the data at each detection point in the tensor, this module traverses along the discrete frequency band dimension. For each discrete frequency band... The module utilizes the absolute phase delay characteristics of this frequency band. Calculate the corresponding phase hysteresis time span This calculation is performed by dividing the phase delay at a specific frequency band by its corresponding center angular frequency, i.e.: In the formula, For this discrete frequency band The corresponding center frequency.

[0038] After obtaining the phase hysteresis time span, the module performs a three-dimensional spatial positioning operation. This involves probing the two-dimensional coordinates of the grid points. It is known that the defect is along the depth direction. The position is calculated using the phase hysteresis time span. The module first retrieves the sonic wave phase velocity from a pre-set material property library, based on the material type of the target masonry. Subsequently, the depth coordinates of the defect were calculated using the time-of-flight method. Applying this calculation to all detection points and all frequency bands exhibiting significant phase delay, the system obtains a series of points labeled as defects. Three-dimensional spatial coordinates.

[0039] To define the precise geometric boundaries of defects, the module sets a failure determination time threshold. Only when the calculated phase hysteresis time span Only when the absolute value of the coordinates is greater than the threshold is the corresponding spatial coordinate point confirmed as constituting the spatial coordinate point group of the external geometric boundary condition.

[0040] The module inversely maps the aggregated spatial coordinate points of the extended geometric boundary conditions to a pre-defined 3D mesh algorithm framework covering the entire detection area. This framework is essentially a regular array of voxels. The module iterates through each voxel, determining whether its spatial extent contains any point from the spatial coordinate point group of the extended geometric boundary conditions. If it does, the voxel is marked as a failed voxel. Furthermore, to encode energy attenuation information in this mask, for each failed voxel, the module also stores its nonlinear interface energy attenuation spectrum in the corresponding frequency band. The numerical value is recorded in the data structure of the voxel.

[0041] In this way, a three-dimensional physical impedance distribution tensor is generated, which is zero only in homogeneous stone regions and filled with specific frequency domain energy resonance gap characteristics and spatial topological hindrance shape values ​​in defect regions. This tensor is ultimately encapsulated as a whole and output as a three-dimensional wavefield penetration failure mask for use in subsequent steps.

[0042] In this embodiment, the phase hysteresis time span essentially represents the flight time of the acoustic wave emitted from the object's surface, reflected at the interface of the internal heterogeneous defect, and finally returning to the receiver. Failure determination time threshold. The setting is equivalent to a spatial filtering threshold for defect depth, based on experimental statistics. For example, by probing dozens of standard test blocks of the same material as the target masonry, including intact test blocks and prefabricated micro-crack test blocks, the distribution differences of phase hysteresis time spans between intact and defective test blocks in various frequency bands are statistically analyzed. A boundary value that can effectively distinguish between the two is selected as the threshold, with a typical threshold range of [missing information]. to .

[0043] The voxel resolution of the 3D mesh algorithm framework, such as 1mm×1mm×1mm, is set to match the detection accuracy of the highest frequency component of the swept signal used in the data acquisition module. The 3D physical impedance distribution tensor is a four-dimensional array. ,in It is a spatial index of voxels. It is a frequency index, and the value it stores is the energy attenuation value at a specific frequency at that location.

[0044] For example, the system is currently processing data from a detection point (X=15, Y=25) in the 120kHz frequency band. The absolute phase delay characteristic Δφ(120kHz) at this point is -0.6rad. Based on this, the system calculates the phase hysteresis time span as Δt(120kHz) = (-0.6rad) / (2π×120000Hz) ≈ -0.796μs.

[0045] In this embodiment, the system sets a failure determination time threshold based on the calibration of the granite. The calculated phase hysteresis time span is 0.5 μs. Since the absolute value of the calculated phase hysteresis time span, |-0.796 μs|, is greater than 0.5 μs, the system determines that there is a physical defect at this point. Next, the system retrieves the average phase velocity of the sonic elastic wave of granite from the material library. The velocity is 3500 m / s, and the defect depth is calculated. That is, 1.393mm.

[0046] Therefore, the spatial coordinates of a defect point are determined to be (15mm, 25mm, 1.393mm). In the reverse mapping stage, it is assumed that the voxel size of the 3D mesh algorithm framework is... Then the coordinate point will fall into the index of Within the voxel. Ultimately, in the generated three-dimensional wavefield penetrating the failure mask, the corresponding The voxel value is no longer zero, and the three-dimensional physical impedance distribution tensor stored within it will be assigned a value, and the voxel is associated with... Record the calculated nonlinear interface energy attenuation spectrum values ​​at the frequency band data points. .

[0047] In one embodiment of the present invention, the network construction module is configured to perform the following steps: The domain-specific language model obtained through unsupervised training is used to process the independent operational perturbation primitives decomposed from natural language, and a preliminary mathematical representation is generated in the latent vector multidimensional space. The preliminary mathematical representation is reshaped by fusing the retrieved fluid dynamic limit attributes using a graph embedding algorithm, generating a coordinate tensor with internalized physical constraints to independently define atomized repair factor nodes. A diagonal matrix consisting of diagonal elements is generated for the atomized repair factor nodes with coordinate tensors, and it is formatted as an acoustic-elastic composite wave impedance compensation tensor simulating energy attenuation and energy coupling effects.

[0048] Specifically, a fully initialized dynamic impedance compensation artificial neural network is loaded and structured within the computer computing architecture. Based on external repair technology knowledge, the basic computational units of this network, namely atomic repair factor nodes, are constructed. Specifically, the network construction and initialization module in the backend server instantiates a dynamic impedance compensation artificial neural network based on a high-order undirected graph neural network logic structure in the memory of the AI ​​computing cluster. The special feature of this network lies in its initialization state: when no external input signal is received, the activation connection channels between all hidden layer graph nodes within the network are rigidly set to a disconnected state. This is achieved, for example, by setting all off-diagonal elements in the graph adjacency matrix to zero, or by setting the initial bias of the gate unit to negative infinity, thereby logically realizing a physically infinite impedance disconnection state and ensuring that the network initially has no preset conduction paths.

[0049] Next, the readable restoration process knowledge is transformed into network-computable atomized units. This submodule accesses a structured database storing a large number of cross-regional records of ancient building restoration processes. It reads the steps of a complex processing procedure, such as "repairing cracks with tung oil, lime, and hemp rope," and uses dependency parsing techniques from natural language processing to break down the procedure into multiple sub-actions. Subsequently, the system removes subjective semantic labels such as "traditional" and "refined" from these sub-actions, and reduces them to a series of linearly executed and decoupled independent operational perturbation primitives based on the direct impact of each action on the physical state of the material, such as density, porosity, and elastic modulus, for example, "high-pressure airflow dust removal" or "specific pressure injection of grout."

[0050] After obtaining the independent perturbation primitives, the system executes the node construction process. Each independent perturbation primitive is first fed into an unsupervised trained domain-specific language model for processing, projecting it to generate its preliminary mathematical representation in the latent vector multidimensional space. Subsequently, this preliminary mathematical representation is reshaped by a graph embedding algorithm module, such as an encoder based on the Physical Information Neural Network (PINN). This encoder simultaneously fuses fluid dynamics or solid mechanical limit properties related to the primitive retrieved from a materials database, such as the maximum allowable viscosity of the slurry and the ultimate compressive strength after curing, thereby generating a coordinate tensor that internalizes the physical constraints. This coordinate tensor is used to independently define the atomized repair factor nodes.

[0051] At the underlying physical level of the algorithm, the core payload of this node is formatted as an acoustic-bullet composite wave impedance compensation tensor. This tensor is a mathematical operator whose structure and values ​​are carefully designed to mathematically simulate the effect of the physical operation represented by this node on the impedance of the acoustic-bullet composite wave propagation medium. Essentially, it is a carrier wave used to numerically smooth and cancel wave field distortion in subsequent calculations.

[0052] In this embodiment, the dynamic impedance compensation artificial neural network adopts the graph attention network (GAT) as its basic architecture. Its "high-order" nature is reflected in the fact that the message passing mechanism will consider the first-order and second-order neighborhood information of the node at the same time, thereby capturing the synergistic or antagonistic effects between non-adjacent steps in the repair process.

[0053] Independent operation perturbation primitives are the smallest functional units for repairing knowledge. They are indivisible, and their execution results depend only on their operation parameters and are independent of the context, achieving effective logical decoupling. Atomized repair factor nodes are digital twins of independent operation perturbation primitives and are basic information processing units in neural networks.

[0054] The acoustic-bullet composite wave impedance compensation tensor is a square matrix with dimensions matching the number of frequency bands in the broadband backscattered wave train data, for example, N×N, where N is the number of discrete frequency bands. The diagonal elements of this tensor simulate the effect of the repair action on the energy attenuation of each frequency band, while the off-diagonal elements simulate the effect on the energy coupling effect between different frequency bands.

[0055] For example, suppose the system reads a process from the repair process record diagram as "Repairing stone cracks with a width of 1-2mm using epoxy resin grouting". The knowledge extraction submodule decomposes this into independent operational perturbation primitives: "with..." "Pressure injection of EP-5 type epoxy resin grout." This primitive text was first converted by the BERT model in the field of materials science. The initial vector of dimension. Subsequently, the graph embedding algorithm module receives this vector and retrieves the hydrodynamic limit properties of EP-5 resin from the database, such as "viscosity". "and applicable pore size" The module fuses these physical parameters with the initial vectors, ultimately generating a high-dimensional coordinate tensor, for example... A vector of dimension is used to define the new atomized repair factor node. Next, the system generates its acoustoelastic composite wave impedance compensation tensor for this node. Considering that the main physical effects of resin grouting are filling voids and improving the homogeneity and density of the medium, this operation reduces the energy attenuation of acoustic waves. Therefore, the compensation tensor generated by the system... Its structure may be a diagonal matrix, with its diagonal elements... The values ​​are all less than 1, for example, between 0.8 and 0.95, and are used to multiply the measured energy decay spectrum in subsequent calculations to simulate the effect of reduced decay after it has been repaired. This is the row and column index of the diagonal matrix. This newly created node is loaded into the graph structure of the dynamic impedance compensation artificial neural network, but its connection weights with all other nodes are set to zero during initialization, placing it in an open-circuit state awaiting activation.

[0056] In one embodiment of the present invention, the network pruning module is configured to perform the following steps: The overwritten self-attention gating weight matrix is ​​analyzed, and the topological specifications of the microscopic spatial gap defined by it are measured. The topological specifications of the microscopic spatial gap are compared with the penetration size threshold representing the repair material stored in the atomized repair factor node. If the topological specifications of the microscopic spatial gap exhibit severe low-frequency resonance distortion and high-frequency signal cutoff characteristics, and there is a physical mismatch with the penetration size threshold, a zero-value attenuation mask slot is derived using the self-attention gating weight matrix. The self-attention connection weight between the corresponding atomized repair factor node and its predecessor node is locked to an absolute zero value through the zero-value attenuation mask slot, cutting off the forward information transmission of the node and completing the forced circuit pruning operation.

[0057] Specifically, after constructing and initializing the dynamic impedance compensation artificial neural network, a physical information-guided network structure pruning process is implemented. The physical spatial defect information contained in the three-dimensional wave field penetrating the failure mask is directly transformed into a mandatory constraint on the neural network computation topology.

[0058] The generated three-dimensional wavefield penetrates the three-dimensional topological distribution matrix of the failure mask, bypassing any conventional feature extraction or encoding layers. Its structure and values ​​are directly overwritten as the self-attention gating weight matrices of the hidden convergence layers at each level of a dynamic impedance-compensated artificial neural network. Through this operation, the spatial adjacency relationships and impedance characteristics of physical defects within the masonry are directly and isomorphically mapped to the initial attention scores for information transmission between nodes in the network graph, determining the original information flow direction when the network processes the defect.

[0059] Based on the overwritten self-attention gating weight matrix, the fit of each atomized repair factor node is examined. It measures and compares the microscopic spatial gap topology defined by the self-attention gating weight matrix with the penetration size threshold of the repair material stored in the metadata of each atomized repair factor node. Specifically, this measurement process involves the system analyzing the energy attenuation spectrum characteristics of voxel clusters marked as defects in the failure mask at different frequency bands to infer the equivalent size of their physical pores. Simultaneously, the system reads the preset material properties of each repair factor node, such as the molecular diameter of the slurry or the minimum permeable crack width.

[0060] When a specific mismatch is detected during fit review, a forced open-circuit pruning operation is triggered. This specific mismatch refers to a sub-millimeter-scale capillary pore physical characteristic, where the microscopic spatial gap topology exhibits severe low-frequency resonance distortion and complete cutoff of high-frequency signals through spectral analysis. However, the repair material represented by a certain atomized repair factor node with high attentional connectivity to this defect, such as macromolecular gel injection, has a penetration size threshold far exceeding that of the capillary pore. In this case, the repair factor is determined to be physically infeasible for this defect.

[0061] By utilizing the high weight values ​​of the self-attention gating weight matrix on this connection path, a zero-value decay mask slot is derived. This mask slot rigidly locks the self-attention connection weights or output activation mask of the incompatible atomic repair factor node to absolute zero. This achieves a forced circuit breaker pruning operation, from physical space penetration obstruction to isomorphic blocking of signal transmission in the computational network, thus permanently excluding this invalid repair option in subsequent solution paths.

[0062] It should be noted that the self-attention gating weight matrix is ​​a matrix used in the GAT network to calculate the influence weights between nodes. This invention innovatively sets its initial value directly using the topological information of the physical mask, rather than learning it through training. The microscopic spatial gap topology specification is a spectral description of the geometry and properties of the physical defect. For example, the cutoff frequency of high-frequency signals is inversely proportional to the pore size, while the position of the low-frequency resonant peak is related to the overall cavity volume of the defect. The penetration size threshold of the repair material is defined in the defect modeling module and is a set of physical constraint parameters stored in each atomized repair factor node, such as the minimum applicable crack width. .

[0063] Self-attention connection weights are adjustable parameters in a neural network neuron model. They control the information flow between nodes before the activation function takes effect. Locking them to zero is equivalent to completely disconnecting the computation graph branch during the forward propagation of the network, which is equivalent to being blocked in this forward propagation.

[0064] For example, continuing with the aforementioned repair scenario, the system injects a three-dimensional wave field that penetrates the failure mask. This mask is located at the voxel level. A defect marker was found at that location. Further analysis of the voxel's energy decay spectrum across the entire frequency sweep revealed that it... The above high-frequency signals are completely cut off, but... Significant resonance distortion occurs in the following low-frequency range. Based on these two characteristics, the system determines the topological specification of the microscopic spatial gap at this location to be "equivalent diameter approximately..." "Capillary pores". At this point, the three-dimensional wavefield penetration failure mask is overwritten as a self-attention gated weight matrix, causing the graph node corresponding to the capillary pore to generate high initial attention to all potential repair factor nodes. During the fit review, an atomized repair factor node representing "macromolecule gel injection" was detected, whose metadata recorded a penetration size threshold for the repair material as " ".because Much larger A physical mismatch was identified. The system immediately triggered a forced circuit breaker operation, deriving a zero-value decay mask slot. This mask slot, targeting the "Macromolecule Gel Injection" node, forcibly modified its self-attention connection weights from the default initial values ​​to absolute zero. Therefore, in subsequent path derivation, despite a higher initial attention level, this computational path from "Capillary Pores" to "Macromolecule Gel Injection" was blocked due to the node's dormancy.

[0065] In one embodiment of the present invention, the sequence generation module is configured to perform the following steps: Before calculating the optimal path, redundant network links in the dynamic impedance compensation artificial neural network that have become completely dormant due to forced circuit breaking and pruning are cleared. The disordered medium scattering feature matrix lattice characterized by the three-dimensional wave field penetration failure mask is specified as the starting boundary for the latent space solution. The non-convex optimization solver is used to iterate the adjacent atomized repair factor node clusters that have not been pruned to calculate the cumulative directed connected path. The surviving graph nodes contained in the cumulative directed connected path are decoded in reverse, and the engineering semantic information is extracted and reorganized into independent operation perturbation primitive flow. According to the spatiotemporal topological order of neurons, the independent operation perturbation primitive flow is loaded with the fluid infusion miniaturized control pressure scalar and the millisecond-level waiting time sequence scalar of the heterogeneous material layered solidification gap, and cascaded and encapsulated to generate an adaptive repair parameter sequence.

[0066] Specifically, after the topology of the dynamic impedance compensation artificial neural network has been pruned and optimized based on physical defect information, the final repair scheme is deduced and generated within this pruned network. A structural scan is performed on the dynamic impedance compensation artificial neural network after the forced circuit closure pruning operation. The purpose of this scan is to clear all redundant network links that have been completely disconnected from upstream activation sources due to the forced circuit closure operation. These branches, becoming disconnected redundant nodes, no longer participate in any valid path calculations.

[0067] The disordered medium scattering feature matrix lattice, composed of multiple non-zero voxels, revealed by the three-dimensional wave field penetrating the failed mask body, is designated as the starting boundary condition for this latent space optimization action or the initial state of the "problem to be solved".

[0068] The solution process officially begins: Starting from the initial state, it iteratively depth-first traverses all adjacent atomic repair factor node clusters that have not been deprived of channel activation weights in the network building blocks, along all branches that remain connected in the graph. At each traversal step, a non-convex optimization solver based on the Sequential Quadratic Programming (SQP) algorithm is invoked. The objective function of this solver is to compute the cumulative product sequence of the acoustic-elastic composite wave impedance compensation tensor that maximizes the smoothness of the disordered medium scattering characteristic matrix lattice. That is, the solver attempts to find a directed connected path composed of multiple surviving atomic repair factor nodes, such that the cumulative product of the acoustic-elastic composite wave impedance compensation tensors represented by all nodes on the path, when applied to the disordered medium scattering characteristic matrix lattice, makes it as close as possible, after mathematical transformation, to the ideal acoustic characteristic matrix state representing a homogeneous medium. This process ultimately produces a directed connected path with optimal cumulative product of the required acoustic-elastic composite wave impedance compensation tensor.

[0069] After obtaining the optimal path, the system performs the final decoding and encapsulation operations. Along the optimal path, all surviving graph nodes contained in the decoding path are decoded in reverse, their engineering semantic information is extracted again, and projected back into the engineering application physical space to reorganize them into a series of ordered independent perturbation primitive flows.

[0070] The decoding process not only reconstructs the operational actions themselves but also strictly adheres to the spatiotemporal topological order of neuronal signals propagating from one layer to the next in a neural network. This order is directly converted into a time series of the repair procedures and cascaded and encapsulated into a final output. This output is the adaptive repair parameter sequence, whose content carrier is a structured data object. It not only contains the step sequence of the repair operations but also includes the fluid infusion miniaturized control pressure scalar and the millisecond-level waiting time scalar of the heterogeneous material layering and curing gap, both parsed from the parameters of each node.

[0071] It should be noted that the scattering characteristic matrix lattice of the disordered medium is the mathematical essence of the three-dimensional wave field penetrating the failed mask. The distribution and value of its non-zero elements directly reflect the scattering and absorption effects of physical defects on the acoustic wave. The acoustic characteristic matrix of the homogeneous medium is an idealized target matrix, usually represented by an identity matrix or a matrix with constant diagonal elements and zero off-diagonal elements, representing the propagation characteristics of sound waves in a lossless medium without attenuation or scattering.

[0072] The optimal cumulative directed connected path is a concept in graph theory, specifically referring to a path from the starting point to the ending point in a graph that maximizes the product of weights or the sum of weights after logarithmic transformation. The fluid grouting micro-control pressure scalar refers to a precise pressure value, measured in Pa or MPa, for a specific grouting operation. The millisecond-level waiting time scalar for the layered curing gap of heterogeneous materials refers to the precise millisecond-level time interval during which the first material must reach a specific degree of curing before two different repair materials can be applied sequentially.

[0073] For example, the system clears isolated connected branches resulting from the pruning of the "macromolecule gel injection" nodes. The current disordered medium scattering characteristic matrix lattice is composed of... voxel The matrix consists of non-zero elements such as attenuation values. The non-convex optimization solver starts from this matrix and traverses the pruned network. The solver finds a path: first connecting to the matrix representing "low viscosity epoxy resin (...)". The atomic repair factor node is connected to the "penetration" node, which is then connected to the "ultrasonic-assisted degassing" node, and finally connected to the "48-hour natural curing" node.

[0074] The solver calculates the acoustic-elastic composite wave impedance compensation tensor corresponding to these three nodes. cumulative product It was found that after it acts on the initial scattering matrix, it can make The energy decay value at the voxel is from The closest to the Therefore, this path was determined to be the optimal cumulative directed connected path. Finally, the decoding module reverse-engineers this path to generate an adaptive repair parameter sequence. The sequence content is: [Step 1: {Action: 'Inject low-viscosity epoxy resin', Pressure: '0.1MPa'}, Step 2: {Action: 'Apply 20kHz ultrasound', Duration: '30s'}, Step 3: {Action: 'Wait', Waiting time: '172800000ms'}]. This sequence, containing specific pressure and millisecond-level timing, is the final output.

[0075] In one embodiment of the present invention, the effect evaluation module is used to perform the following steps: Relying on the hardware communication bus to analyze the adaptive repair parameter sequence, control commands are issued to drive the arrayed numerically controlled micro-irrigation equipment to complete the physical intervention action; the waveform of the initial detection frequency response bandwidth and frequency modulation slope is replicated, and the secondary back-reflected wave train flow after the physical intervention action is received synchronously, and the feedback delay parameter of the reflection wave array boundary is calculated from it; the feedback delay parameter of the reflection wave array boundary is subjected to frequency-band interferometric subtraction operation with the absolute phase delay feature extracted in the first step, and the differential phase signal is output; after digital bandpass filtering, the low-frequency standing wave energy accumulation distortion is extracted from the differential phase signal to generate the residual wave train feature data.

[0076] Specifically, after generating the parameter sequence to guide physical repair, physical intervention is performed, and the repair effect is then subjected to secondary detection and quantitative evaluation. The integrated repair execution and online monitoring system first obtains the adaptive repair parameter sequence from the previous stage and parses the control scalars it carries. Subsequently, the system sends a serialized instruction set to the lower-level execution unit via a hardware communication bus, such as an industrial Ethernet based on the Modbus / TCP protocol. These instruction sets precisely drive external arrayed CNC micro-irrigation equipment deployed in the physical repair area, such as high-precision piezoelectric metering pumps mounted on multi-axis robotic arms, and radio frequency polymerization curing transmitter clusters, to work collaboratively, strictly following the pressure, dosage, timing, and energy parameters specified in the sequence, to complete the actual destructive and blocking intervention action on the target masonry.

[0077] After the intervention was completed and the necessary material curing and sealing period had passed, the monitoring system seamlessly reused the phased array acoustic missile flaw detection sensor antenna group configured in the data acquisition module. This antenna group was precisely repositioned and aligned with the core coordinate region of the repaired physical area, replicating the emission of a verification-grade controlled frequency-sweeping acoustic missile composite mechanical wave identical to the initial detection. Its frequency response bandwidth remained between 10 kHz and 500 kHz, and waveform parameters such as the modulation slope and transmission power remained unchanged to ensure the comparability of the two detection results. The system simultaneously acquired the secondary back-reflected wave train current excited by this detection in the post-repair environment.

[0078] The system employs a phase difference comparison mathematical logic gate to process the secondary back-reflected wave train in real time, capturing and calculating the boundary feedback delay parameters of the reflected wave array at each discrete frequency band after repair. These parameters are then subjected to frequency-band interferometric subtraction with the original phase-sensitive data (absolute phase delay characteristics) stored in the first-round detection baseline archive, which is caused by the internal heterogeneous defect reflection interface. The subtraction result is then filtered by a digital bandpass filter to separate and remove randomly distributed conventional thermal noise, while retaining and extracting the low-frequency standing wave energy accumulation distortion that cannot be completely smoothed out by theoretical repair schemes. This remaining distortion energy is ultimately synthesized into residual wave train characteristic data capable of finely mapping residual defects such as interface micro-voids, and serves as the final output of this step.

[0079] The purpose of performing subtraction interferometry is to precisely quantify the changes in the physical medium introduced by the repair process. The core calculation formula is as follows: In the formula, Represents discrete frequency bands The differential phase signal at that location. This calculated differential phase signal needs to undergo filtering and energy conversion to obtain the final energy distortion component. This is used to construct residual wave train characteristic data. Both phase delay quantities in the formula are dimensionless radians, therefore the subtraction operation is dimensionally consistent. The physical repair region specifically refers to the set of three-dimensional spatial elements marked by the three-dimensional wavefield penetrating the failed mask, which requires physical repair operations. The feedback delay parameter of the reflected wave array boundary from the secondary detection is denoted as... Its calculation method is exactly the same as that used for processing the initial detection data, representing the acoustic response characteristics of the repaired medium. The phase-sensitive data caused by the internal heterogeneous defect reflection interface in the benchmark archive is the calculated and stored absolute phase delay characteristic, denoted as... It represents the acoustic response of the medium before repair.

[0080] Low-frequency standing wave energy accumulation distortion refers to the micron-level poor curing interface or microbubbles that may exist between the repair material and the original stone. These micro-void defects, like a Helmholtz resonator, can produce notch and standing wave effects on specific low-frequency sound waves, leading to abnormal energy accumulation in that frequency band. Its characteristic frequency range is typically within... to kilohertz. The residual wave train characteristic data is a tensor with a similar structure to the original acquired data, but its internal padding consists of energy distortion values ​​that quantify the degree of remaining defects after repair.

[0081] For example, the repair execution and online monitoring system receives an adaptive repair parameter sequence containing three steps and drives the robotic arm to perform grouting and ultrasonic-assisted operations at coordinates (15mm, 25mm, 1.393mm). While waiting... After a curing time of one hour, the system initiates a secondary detection. In the frequency band, new detection results were used to calculate the boundary feedback delay parameter of the penetration wave array. for Compared to the raw values ​​recorded in the data acquisition module Substantial improvements have been made. However, in In the low-frequency band, the system detects the original phase delay. Close to zero, but the new measurement after repair But it turned into The system performs an interferometric subtraction operation to obtain the phase residual for this frequency band. This non-zero residual indicates that the repair process may have introduced new microscopic defects. The system further... Extracting the power spectral density from the relevant signal spectrum yields a The low-frequency standing wave energy accumulation distortion becomes a sharing amount. This energy value is then written into a new data structure, and together with the residual energy values ​​calculated from other frequency bands, it constitutes the residual wave train characteristic data for this restoration assessment.

[0082] In one embodiment of the present invention, the model update module is configured to perform the following steps: The residual energy spectrum is extracted from the residual wave train feature data, and the second derivative eigenvalue of the residual energy spectrum is calculated and obtained. According to the preset inverse dynamics function, the second derivative eigenvalue is transcribed and amplified into the initial impulse of the backpropagation momentum of the final layer of the decision network, forming a physical loss gradient flow. The physical loss gradient flow is loaded and traced back to the specific atomized repair factor node position that caused the misleading decision. Based on the magnitude of the physical loss gradient flow, a momentum repulsion vector is generated from the defect anchor point mapped by the current spatial gap topology specification to the problem node. The repulsion penalty is calculated based on the momentum repulsion vector, and the latent space geometric clustering coordinates of the poor repair effect related node group relative to the Euclidean alienation distance of the defect anchor point in the graph manifold space cluster are pushed away and updated.

[0083] Specifically, after objectively and quantitatively evaluating the repair effect, the process enters a closed-loop self-correction and model evolution stage. The repair residuals in the physical world are transformed into the internal driving force for the evolution of the neural network structure. First, residual wave train feature data is received, and any evaluation annotation interfaces dependent on human experience are immediately removed. Then, directly from this data, a retained distortion energy spectral envelope is fitted by analyzing the distortion energy values ​​across all discrete frequency bands.

[0084] The second derivative of the envelope slope is calculated to obtain the second derivative eigenvalue of the residual energy spectrum. This operator quantifies the severity and trend of energy distortion. Next, based on a pre-defined inverse dynamics function, this operator is transcribed and nonlinearly amplified to generate a physical loss gradient flow. The magnitude and direction of this gradient flow directly correspond to the failure level and correction direction of the repair task and are designated as the initial impulse of the backpropagation momentum in the final layer of the decision network.

[0085] After obtaining the physical loss gradient flow, it is loaded into the loss function interface of the dynamic impedance compensation artificial neural network, and a targeted backpropagation calculation is initiated. The gradient flow enters from the end of the network, backtracks along the activated path, and reaches the positions of the effective activation layer operators in the concatenation and arrangement decision of this suboptimal repair sequence, that is, the specific atomic repair factor nodes that have been proven to have produced misleading decisions.

[0086] Within the latent space multidimensional coordinate system of the graph topology, for the anchor points formed by the microscopic spatial gap topological specifications of the input, and the identified problem nodes, the system will trigger a repulsion penalty calculation based on feature space distance. The core of this calculation is to generate a repulsion vector based on the magnitude of the physical loss gradient flow. This vector's direction points from the defect anchor point to the problem node. This momentum repulsion vector is used to directly increase the absolute Euclidean alienation distance of the cluster of nodes associated with the poor repair effect within the graph manifold space cluster relative to the defect anchor point.

[0087] After the vector offset calculation is completed, the latent space geometric cluster coordinates of the corresponding atomic repair factor nodes in the network are updated. This ensures that when the neural system receives a three-dimensional wave field that maps similar physical damage textures through the failure mask in the next time, it will naturally reduce the probability of assigning attention weights to the nodes with a history of failed applications due to the change in topological projection capability. This completes a completely objective closed-loop self-correction that directly constrains the evolution of network connection synapses by deep physical objective mapping.

[0088] The formula for calculating the latent space geometric aggregation coordinate update is: In the formula, The number of steps for the iterative update operation; The operator for finding the Euclidean norm (L2 norm) of a vector. This indicates a specific atomized repair factor node that was identified as causing misleading decisions. In the The updated latent space geometric cluster coordinates. Is this node at the ? Coordinates before the update. This is a hyperparameter representing the learning rate, and its value range is usually set empirically. to The interval is used to control the update step size. It is the scalar magnitude of the physical loss gradient flow, which is directly derived from the second derivative eigenvalue of the residual energy spectrum. These are the coordinates of the current input microscopic spatial gap topological specification anchor point within the same latent space. The fractional part calculates the coordinates from the anchor point... Point to the problem node The unit direction vector is the direction of the momentum repulsion force.

[0089] It should be noted that the second derivative eigenvalue of the residual energy spectrum is a scalar value used to describe the sharpness of the most significant distortion peak in the residual energy spectrum; the higher the sharpness, the larger the operator value. The inverse dynamics function is a predefined nonlinear mapping function, such as the sigmoid function or a piecewise linear function. Its role is to map the above operator values ​​to a more suitable numerical range as gradient signals and amplify them, ensuring that even small physical residuals can generate sufficient network update momentum. The latent space geometric cluster coordinates are the mathematical representation of each atomized repair factor node in the high-dimensional vector space of the neural network, and the Euclidean distance between nodes reflects their functional similarity.

[0090] For example, receiving a message containing frequency band After obtaining the residual wave train characteristic data of energy accumulation, the data is analyzed to lock in... The peak point of the energy distortion is identified, and the second derivative eigenvalue of the residual energy spectrum of this peak is calculated. Through a magnification factor of The inverse dynamics function generates a physical loss gradient flow, whose scalar magnitude is... identified as The backpropagation algorithm is initiated, tracing the loss back to ultimately pinpoint the cause of the micro-voids as the first node of the optimal path, namely "low viscosity epoxy resin ( "Penetration" node Assuming a simplified two-dimensional latent space, the current defect anchor point... The coordinates are [1.2, 3.4], and the "epoxy resin" node... Coordinates before update for Learning rate Set as The system begins calculating the repulsion penalty based on the feature space distance. First, the unit vector of the repulsion direction is calculated as ([1.5,3.2]-[1.2,3.4]) / ∥[0.3,-0.2]∥≈[0.83,-0.55]. Then, the updated coordinates are calculated. After this update, the "epoxy resin" node... With defect anchor point The absolute Euclidean distance between them is approximately [missing information] before the update. Added to the updated version The coordinates of the node were successfully pushed away, completing this closed-loop self-correcting evolution.

[0091] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0092] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A knowledge base and adaptive application system for the restoration techniques of ancient stone masonry buildings, characterized in that, include: The data acquisition module is used to apply a sweeping sonic boom composite mechanical wave to the target masonry, capture the back-reflected wave train and the far-end penetrating wave train, extract the nonlinear interface energy attenuation spectrum and absolute phase delay characteristics, and merge them to generate broadband back-scattered wave train data. The defect modeling module is used to analyze broadband backscattered wave train data to obtain the phase hysteresis time span, thereby defining the external geometric boundary conditions, and inversely mapping and encapsulating to generate a three-dimensional wave field penetration failure mask. The network construction module is used to load the dynamic impedance compensation artificial neural network, obtain the disassembled independent operation perturbation primitives and the associated physical properties of the repair materials, reshape the independent operation perturbation primitives into coordinate tensors to define atomized repair factor nodes, and format them into an acousto-elastic composite wave impedance compensation tensor. The network pruning module is used to overwrite the three-dimensional wavefield penetration failure mask as a self-attention gating weight matrix of a dynamic impedance compensation artificial neural network. Based on the penetration size threshold in the physical properties of the repair material, it performs a forced circuit breaking pruning operation on the atomized repair factor nodes that do not match the physical boundary of the mask. The sequence generation module is used in the dynamic impedance compensation artificial neural network to characterize the three-dimensional wave field penetration failure mask as a disordered medium scattering feature matrix lattice, calculate the cumulative directed connected path of the smoothed disordered medium scattering feature matrix lattice, and inversely decode the path to generate an adaptive repair parameter sequence.

2. The knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings according to claim 1, characterized in that, The data acquisition module is configured to perform the following steps: Short-time Fourier transforms are applied to the captured back-reflected wave train and the far-end penetrating wave train to generate time-spectrum diagrams; The power spectral density of the far-end penetrating wave train under discrete frequency bands is calculated to the ratio of the preset power spectral density of the non-destructive healthy medium reference, and the nonlinear interface energy attenuation spectrum is obtained. The initial phase difference between the back-reflected wave train and the preset reference signal is extracted, and phase dewinding operation is performed along the discrete frequency band axis to obtain monotonically continuous absolute phase delay characteristics. The nonlinear interface energy decay spectrum and absolute phase delay features are stacked along the feature dimension and merged to generate broadband backscattered wave train data with a three-dimensional structure tensor.

3. The knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings according to claim 1, characterized in that, The defect modeling module is configured to perform the following steps: The broadband backscatter wave train data is sliced ​​along the detection time axis, and the absolute phase delay characteristic value under the discrete frequency band is divided by the corresponding angular frequency to calculate the true phase hysteresis time span. The acoustic wave phase velocity under the preset target frequency band is retrieved. The time-of-flight method is used to combine the acoustic wave phase velocity with the actual phase hysteresis time span to calculate and generate the defect depth coordinates, and define the external geometric boundary conditions formed by the spatial coordinate point group. Spatial voxels whose absolute value of phase hysteresis time span is greater than the failure judgment time threshold are selected and, together with the recorded nonlinear interface energy decay spectrum values, are inversely mapped to a preset three-dimensional mesh algorithm framework to encapsulate and generate a three-dimensional wave field penetration failure mask.

4. The knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings according to claim 1, characterized in that, The network building module is configured to perform the following steps: The domain-specific language model obtained through unsupervised training is used to process the independent perturbation primitives decomposed from natural language, and the projection generates a preliminary mathematical representation in the latent vector multidimensional space. The preliminary mathematical representation is reshaped by fusing the retrieved fluid dynamics limit attributes using a graph embedding algorithm, generating a coordinate tensor with internalized physical constraints, and defining atomized repair factor nodes independently. A diagonal matrix consisting of diagonal elements is generated for the atomized repair factor nodes with coordinate tensors, and formatted as an acoustic-elastic composite wave impedance compensation tensor simulating energy attenuation and energy coupling effects.

5. The knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings according to claim 1, characterized in that, The network pruning module is configured to perform the following steps: Analyze the overwritten self-attention gating weight matrix and measure the topological specifications of the microscopic spatial gaps it defines; Compare the topological specifications of the microscopic spatial gaps with the penetration size threshold representing the repair material stored in the atomized repair factor nodes; If the microscopic spatial gap topology exhibits severe low-frequency resonance distortion and high-frequency signal cutoff characteristics, and is physically mismatched with the penetration size threshold, a zero-value attenuation mask slot is derived using a self-attention gating weight matrix. By using a zero-value decay mask slot, the self-attention connection weight between the corresponding atomized repair factor node and its predecessor node is locked to an absolute zero value, cutting off the forward information transmission of the node and completing the forced disconnection pruning operation.

6. The knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings according to claim 1, characterized in that, The sequence generation module is configured to perform the following steps: Before calculating the optimal path, clear the redundant link branches in the dynamic impedance compensation artificial neural network that have become completely dormant due to forced circuit closure pruning. The disordered medium scattering characteristic matrix lattice characterized by the three-dimensional wave field penetration failure mask is specified as the starting boundary of the latent space solution action. The non-convex optimization solver is used to iterate the unpruned adjacent atomized repair factor node clusters to calculate the cumulative directed connected path. Inversely decode the surviving graph nodes contained in the multiplication of the directed connected path, extract the engineering semantic information and reorganize it into an independent operation perturbation primitive flow; Following the spatiotemporal topological order of neurons, the independent operational perturbation primitive flow is loaded with a fluid perfusion miniaturized control pressure scalar and a millisecond-level waiting time scalar of the heterogeneous material layering solidification gap, and cascaded and encapsulated to generate an adaptive repair parameter sequence.

7. The knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings according to claim 1, characterized in that, It also includes an effectiveness evaluation module, which is configured to perform the following steps: Relying on the hardware communication bus to analyze the adaptive repair parameter sequence, the system issues control commands to drive the arrayed numerical control micro-irrigation equipment to complete the physical intervention action; The waveform of the first detection frequency response bandwidth and frequency modulation slope was replicated, and the secondary reverse reflection wave train after the physical intervention was received synchronously, and the boundary feedback delay parameter of the reflection wave array was calculated from it. The boundary feedback delay parameter of the reflected wave array is subjected to frequency-band interferometric subtraction with the first extracted absolute phase delay feature to output a differential phase signal. After digital bandpass filtering, the low-frequency standing wave energy accumulation distortion is extracted from the differential phase signal and transformed into a shared quantity, which is then used to synthesize residual wave train characteristic data.

8. The knowledge base and adaptive application system for the restoration technology of ancient stone masonry buildings according to claim 7, characterized in that, It also includes a model update module, which is configured to perform the following steps: Extract the residual energy spectrum from the residual wave train characteristic data, and calculate and obtain the second derivative eigenvalues ​​of the residual energy spectrum; Based on the preset inverse dynamics function, the second derivative eigenvalues ​​are transcribed and amplified into the initial impulse of the backward propagation momentum of the final layer of the decision network, forming the physical loss gradient flow; Load the physical loss gradient flow back to the specific atomized repair factor node position that caused the misleading decision, and generate a momentum repulsion vector from the defect anchor point mapped from the current spatial gap topology specification to the problem node based on the magnitude of the physical loss gradient flow; Based on the momentum repulsion vector, the repulsion penalty calculation is performed to push away and update the latent space geometric cluster coordinates of the node group associated with the poor repair effect in the graph manifold space cluster relative to the Euclidean alienation distance of the defect anchor point.

9. A knowledge base and adaptive application system for the restoration techniques of ancient stone masonry buildings according to claim 1, characterized in that, The data acquisition module is also configured to perform the following steps: Obtain the background vibration noise flow of the environment in which the target masonry is located, excluding test sources; Spectral decomposition of the background vibration noise stream is performed to extract the set of inherent resonant frequencies of the current detection environment; By comparing the environmental resonant frequency set with the preset mechanical wave sweep frequency band, the starting and ending sweep frequencies of the sweep frequency sonic projectile composite mechanical wave are dynamically adjusted to generate an adaptive sweep command that avoids the environmental resonant frequency set and send it to the excitation unit.

10. A knowledge base and adaptive application system for the restoration techniques of ancient stone masonry buildings according to claim 1, characterized in that, The network building block also includes an insulation initialization step before receiving external signals: Instantiate a dynamic impedance compensation artificial neural network based on a high-order undirected graph neural network logic structure in the memory of the computing cluster; In the initial state where no arbitrary mask signal generated by calculation is received, all off-diagonal elements in the adjacency matrix of the network internal graph are forcibly assigned the value of zero; The initial bias of the control unit of the hidden layer graph node is set to the limit of negative infinity, and the active connection channels between all hidden layer graph nodes are forced to be in a non-connected and open-circuit state that does not contain any preset conduction path.

Images