An artificial intelligence-based building main body structure defect identification method

Abstract

This invention provides an artificial intelligence-based method for identifying defects in building structures, relating to the field of deep learning. The method includes: acquiring structural observation image sequences, structural vibration response sequences, and structural acoustic wave propagation sequences to construct a set of structural observation units; extracting multimodal features based on spatial alignment to form a set of structural feature expressions; further constructing a lining structure interface state discrimination model and predicting the interface state; identifying potential voiding areas by combining vibration propagation anomaly analysis; and fusing spatial features of the prediction results and anomaly analysis results to construct a spatial voiding probability distribution. Finally, generating lining structure defect identification results through spatial threshold discrimination and region mapping. This invention can improve the accuracy of identifying latent defects such as early micro-voiding at the surrounding rock lining interface of deep underground building structures.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and specifically to an artificial intelligence-based method for identifying defects in the main structure of buildings. Background Technology

[0002] Existing technologies for identifying defects in building structures mainly rely on manual inspection, single-image detection, or localized sensor detection. Manual inspection typically involves visual inspection or simple measuring tools to identify visible defects such as cracks and spalling on the structural surface. Image-based detection methods usually use cameras to acquire images of the structural surface and then identify cracks and surface anomalies using traditional image processing algorithms or shallow machine learning models. Sensor-based detection methods acquire structural response data by deploying vibration sensors, acoustic sensors, or strain sensors and analyze the structural state based on threshold judgments or empirical models. With the development of artificial intelligence technology, some solutions have begun to use deep learning models to automatically identify structural images or combine single-modal vibration data to detect structural anomalies. However, overall, defect identification still mainly relies on single data sources or single feature types.

[0003] However, in complex environments such as deep underground laboratories, the rock lining interface may develop micro-void structures due to construction disturbance, stress release, or long-term creep. These micro-void defects typically do not initially manifest as obvious cracks or spalling on the structural surface, but rather as latent features such as altered vibration propagation paths and abnormal sound wave propagation characteristics. Existing technologies, relying on single image information or single sensor signals for discrimination, struggle to simultaneously characterize the correlation between visible surface features and internal propagation characteristics, resulting in insufficient ability to identify early micro-void regions. Furthermore, they lack the ability to spatially fuse multi-source features and determine continuous regions, easily leading to missed detections or misjudgments, thus hindering reliable identification of latent defects at the rock lining interface. Summary of the Invention

[0004] This invention provides an artificial intelligence-based method for identifying defects in the main structure of buildings, which can improve the accuracy of identifying latent defects such as early micro-voids at the interface of the surrounding rock lining in deep underground building structures.

[0005] In a first aspect of the present invention, an artificial intelligence-based method for identifying defects in the main structure of a building is provided, the method comprising: Acquire the structural observation image sequence, structural vibration response sequence, and structural acoustic wave propagation sequence of the monitored building lining surface, and perform preprocessing to form a set of structural observation units; Feature extraction processing is performed on the set of structural observation units, and a multimodal structural feature representation is constructed by spatial coordinate alignment, thereby forming a set of structural feature representations; Based on the set of structural feature expressions, a lining structure interface state discrimination model is constructed, and historical samples are used to perform model training processing. At the same time, interface state inference processing is performed on the set of structural feature expressions to obtain a set of predicted results for the structural interface state. Based on the predicted result set, vibration propagation anomaly analysis processing is performed on the structural vibration response signal to identify potential void areas at the surrounding rock lining interface and obtain the vibration propagation anomaly analysis results. Based on the anomaly analysis results and the prediction results set, spatial feature fusion processing is performed to construct the spatial void probability distribution of the lining structure; Based on the spatial void probability distribution, spatial threshold discrimination and spatial region mapping processes are performed to generate lining structure defect identification results.

[0006] In a second aspect of the invention, an artificial intelligence-based building structure defect identification device is provided. The device is used to execute an artificial intelligence-based building structure defect identification method as described in any of the above embodiments. The device includes an acquisition module, a processing module, and an output module, wherein: The acquisition module is used to acquire the structural observation image sequence, structural vibration response sequence, and structural acoustic wave propagation sequence of the monitored building lining structure surface, and perform preprocessing to form a set of structural observation units; The processing module is used to perform feature extraction processing on the set of structural observation units and construct a multimodal structural feature representation by aligning spatial coordinates, thereby forming a set of structural feature representations. The processing module is used to construct a lining structure interface state discrimination model based on the structural feature expression set, and to perform model training processing using historical samples. At the same time, it performs interface state inference processing on the structural feature expression set to obtain a set of predicted results of the structural interface state. The processing module is used to perform vibration propagation anomaly analysis processing on the structural vibration response signal based on the prediction result set, so as to identify potential void areas at the surrounding rock lining interface and obtain the vibration propagation anomaly analysis results. The processing module is used to perform spatial feature fusion processing based on the anomaly analysis results and the prediction result set, thereby constructing the spatial void probability distribution of the lining structure. The output module is used to perform spatial threshold discrimination and spatial region mapping processing based on the spatial void probability distribution to generate lining structure defect identification results.

[0007] In a third aspect of the invention, an electronic device is provided, including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the preceding embodiments.

[0008] In a fourth aspect of the invention, a non-transitory computer-readable storage medium is provided, the computer-readable storage medium storing instructions that, when executed, perform the method as described in any of the preceding claims.

[0009] In summary, one or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: This invention achieves collaborative modeling of surface visible features and internal propagation features within the same structural observation unit framework by unifying the spatial coordinates of structural observation image sequences, structural vibration response sequences, and structural acoustic wave propagation sequences, and constructing a multimodal feature representation of the structure. Simultaneously, it utilizes a lining structure interface state discrimination model to impose prior constraints on the interface state, and extracts propagation deviation information sensitive to interface voiding through vibration propagation anomaly analysis. Furthermore, it constructs a continuous spatial voiding probability distribution through spatial feature fusion and performs regional-level judgment, enabling latent defects to be amplified and stably expressed under the combined effects of multi-source information coupling, spatial continuity constraints, and temporal evolution consistency. This significantly improves the accuracy of identifying latent defects such as early micro-voiding at the surrounding rock lining interface. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a method for identifying defects in the main structure of a building based on artificial intelligence, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of a module of a building structure defect identification device based on artificial intelligence disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention.

[0011] Explanation of reference numerals in the attached drawings: 201, acquisition module; 202, processing module; 203, output module; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. Detailed Implementation

[0012] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification 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.

[0013] In the description of the embodiments of the present invention, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0014] In the description of the embodiments of the present invention, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0015] Existing methods for identifying defects in building structures mostly rely on manual inspections or single-modal data analysis, which makes it difficult to simultaneously characterize the relationship between visible surface features and internal propagation features. In complex environments such as deep underground laboratories, there is a lack of effective means to identify latent defects such as early micro-voids at the interface of the surrounding rock lining, which can easily lead to missed detections or misjudgments, making it difficult to achieve accurate defect determination and continuous spatial identification.

[0016] This invention discloses an artificial intelligence-based method for identifying defects in building structures, which is applied to a server. The server includes, but is not limited to, electronic devices such as mobile phones, tablets, wearable devices, and PCs (Personal Computers), and can also be a backend server running the artificial intelligence-based method for identifying defects in building structures. The server can be a standalone server or a server cluster composed of multiple servers.

[0017] This embodiment discloses an artificial intelligence-based method for identifying defects in building main structures, referring to... Figure 1 It includes the following steps: S110: Acquire the structural observation image sequence, structural vibration response sequence, and structural acoustic wave propagation sequence of the monitored building lining surface, and perform preprocessing to form a set of structural observation units.

[0018] S120 performs feature extraction processing on the set of structural observation units and constructs a multimodal feature representation of the structure through spatial coordinate alignment, thereby forming a set of structural feature representations.

[0019] S130: Construct a lining structure interface state discrimination model based on the set of structural feature expressions, and use historical samples to perform model training processing. At the same time, perform interface state inference processing on the set of structural feature expressions to obtain a set of predicted results for the structural interface state.

[0020] S140, based on the prediction result set, perform vibration propagation anomaly analysis processing on the structural vibration response signal to identify potential void areas at the surrounding rock lining interface and obtain the anomaly analysis results of vibration propagation.

[0021] S150 performs spatial feature fusion processing based on the set of anomaly analysis results and prediction results to construct the spatial void probability distribution of the lining structure.

[0022] S160, perform spatial threshold discrimination and spatial region mapping processing based on the spatial void probability distribution to generate lining structure defect identification results.

[0023] In one possible implementation, feature extraction processing is performed on the set of structural observation units, and a multimodal feature representation of the structure is constructed through spatial coordinate alignment, thereby forming a set of structural feature representations. Specifically, this includes: performing spatial coordinate identification binding and timestamp identification registration processing on the structural observation images, structural vibration response signals, and structural acoustic wave propagation signals in the set of structural observation units to establish a structural observation reference system; performing image enhancement processing and region slicing processing on the structural observation images, and extracting image texture feature vectors representing the surface state of the lining structure; and performing signal purification processing and multi-domain feature extraction processing on the structural vibration response signals to obtain a representation of the structure's surface state. The vibration response feature vector of the lining structure's dynamic response is obtained; the propagation waveform correction and propagation feature analysis of the structural acoustic wave propagation signal are performed to obtain the acoustic wave propagation feature vector characterizing the propagation state of the surrounding rock lining interface; based on the structural observation reference system, the image texture feature vector, vibration response feature vector, and acoustic wave propagation feature vector are spatially aligned and consistency checked to form candidate multimodal feature vectors; the candidate multimodal feature vectors are subjected to correlation compression and discriminative enhancement to generate structural multimodal feature expressions; the structural multimodal feature expressions are aggregated according to the correspondence of structural observation units to form a structural feature expression set.

[0024] Specifically, a unified structural observation reference system is first established around each structural observation unit. This reference system constrains structural observation images, structural vibration response signals, and structural acoustic wave propagation signals to the same spatial location and observation time for interpretation, preventing regional or temporal misalignment of features from different sources. Spatial coordinate identification binding can be achieved through the grid position in the lining structure geometric model, sensor installation position, image capture pose, and acoustic wave transmission and reception position, mapping each structural observation unit to a unique spatial location on the lining structure surface or surrounding rock lining interface. Timestamp identification registration is used to integrate multi-source observation data with different sampling frequencies and periods to a unified time reference. Typically, the clocks of each acquisition device are first synchronized and corrected, and then interpolation compensation, sampling alignment, and delay correction are used to form a unified observation time. Here, spatial coordinate identification refers to the coordinate code that uniquely represents the spatial location of the structural observation unit, timestamp identification refers to the time code that represents the time of observation data acquisition, and registration refers to the process of adjusting data from multiple sources to a unified reference system. For cases where device triggering delay exists, the original timestamp can be corrected based on the known propagation speed and system delay, as expressed by:

[0025] in, This represents the corrected unified timestamp, used for subsequent time alignment. Indicates the original data collection timestamp; This indicates the fixed delay introduced by the data acquisition device itself, which can be obtained through device calibration; This represents the propagation delay of the signal along its path, which can be calculated based on the sensor spacing and the propagation speed of the medium. This processing ensures that the three types of observation data have a unified spatial and temporal assignment within the same structural observation unit.

[0026] The focus of image processing for structural observations is to highlight abnormal textures on the surface of the lining structure and to segment large-format images into regions suitable for local analysis. Image enhancement typically includes brightness normalization, local contrast enhancement, noise suppression, edge preservation, and shadow reduction. Brightness normalization reduces grayscale deviations caused by uneven illumination; local contrast enhancement highlights crack edges, seepage marks, and surface erosion boundaries; noise suppression removes dust reflections, sensor noise, and low-light particle interference; and edge preservation preserves crack tips and abrupt boundary changes while smoothing noise. Region slicing involves dividing the enhanced image into image sub-blocks corresponding to specific structural observation units based on spatial coordinates, ensuring that each sub-block represents only the surface state of a particular lining structure region. The image texture feature vector is a set of numerical features extracted from the image sub-blocks, used to characterize local grayscale distribution, directional continuity, edge density, roughness variations, and the degree of anomalous patch aggregation. A common implementation method is to first extract deep texture features using a convolutional feature extraction network, and then superimpose local statistical features to form a composite feature representation. If it is necessary to extract local texture complexity, a grayscale dispersion index can be introduced, the expression of which is:

[0027] in, The grayscale dispersion of an image sub-block reflects the degree of surface texture undulation; N represents the number of pixels in the image sub-block. This represents the grayscale value of the i-th pixel; This represents the average grayscale value of all pixels within an image sub-block. A larger value generally indicates more drastic changes in the surface texture of the observed structural unit, and is more likely to contain cracks, peeling, or seepage marks.

[0028] The focus of processing structural vibration response signals is to restore the true dynamic response of the lining structure and extract features reflecting interface relaxation and local stiffness anomalies from three levels: time domain, frequency domain, and time-frequency domain. Signal purification generally begins with baseline drift elimination to remove low-frequency bias caused by sensor zero-point drift; then bandpass filtering is performed to retain effective frequency bands related to structural vibration; subsequently, abnormal pulse removal is performed to remove isolated spikes caused by mechanical shock or transient electrical interference; and, if necessary, multi-channel synchronous correction is performed to ensure that vibration waveforms at different measuring points have a unified time starting point. Multi-domain feature extraction refers to extracting features such as peak value, root mean square (RMS), and decay rate from the time domain; extracting features such as dominant frequency, frequency band energy distribution, and spectral centroid from the frequency domain; and extracting energy migration trajectory and local resonance enhancement features from the time-frequency domain. Here, time-domain features refer to features directly extracted from the amplitude variation of the original waveform over time, frequency-domain features refer to features extracted from the frequency distribution after spectral transformation, and time-frequency domain features refer to the combined features retaining both time and frequency positions. If it is necessary to characterize the degree of vibration energy concentration, the RMS response can be calculated, and its expression is:

[0029] in, The root mean square value of the vibration response is used to characterize the intensity of vibration energy in the current structural observation unit; N represents the number of sampling points involved in the calculation. This represents the vibration amplitude at the i-th sampling point. The larger the root mean square value, the higher the overall energy of the vibration response; if there is an abnormal weakening or amplification compared to the reference area, it may indicate a change in interface stiffness or local voiding.

[0030] The processing of structural acoustic wave propagation signals focuses on correcting the propagation path and waveform distortion, and extracting propagation characteristics reflecting the continuity of the surrounding rock lining interface. Propagation waveform correction typically begins with uniform calibration of the transmission and reception times, followed by eliminating gain differences and sampling delays between different channels. Subsequently, interference from multiple reflected waves and boundary echoes on first-wave identification is suppressed, and a path separation algorithm distinguishes between direct propagation components and diffracted propagation components. Propagation characteristic analysis revolves around the first-wave arrival time, amplitude attenuation, reflection peak distribution, transmission continuity, and cavity echo characteristics. Here, the first-wave arrival time refers to the earliest effective waveform moment of the acoustic signal reaching the receiving point, reflecting the propagation path and medium continuity; amplitude attenuation reflects energy loss during propagation; reflection peak distribution characterizes the reflection effect of interface discontinuities on the waveform; and transmission continuity characterizes the stable propagation capability of the acoustic wave after passing through the interface. To characterize the propagation attenuation, an attenuation coefficient can be defined, with the following expression:

[0031] in, This represents the sound wave propagation attenuation coefficient, used to characterize the strength of attenuation during sound wave propagation; This indicates the reference amplitude corresponding to the transmitter or reference path; This represents the effective amplitude measured at the receiving end. The larger this coefficient is, the more significant the energy attenuation during propagation, which usually indicates that there is medium mismatch, micro-voids, or local relaxation at the interface of the surrounding rock lining.

[0032] After obtaining the image texture feature vector, vibration response feature vector, and sound wave propagation feature vector, spatial alignment and consistency verification are performed under a unified structural observation reference system. The core of spatial alignment is to ensure that all three types of features originate from the same structural observation unit corresponding to the lining structure region. Typically, the intersection of the image sub-block center position, vibration sensor measurement point position, and sound wave propagation path is mapped onto a unified structural grid based on spatial coordinates. Then, nearest neighbor matching, weighted projection, or local interpolation is used to achieve position alignment. Consistency verification is used to determine whether these three types of features meet the fusion requirements in terms of spatial attribution, temporal attribution, and observation quality. If the image texture feature of a structural observation unit originates from the current moment, while the vibration response and sound wave propagation features lag behind earlier moments, it indicates insufficient temporal consistency, requiring re-registration or removal. Candidate multimodal feature vectors refer to the features to be fused, formed by splicing or associative encoding of the three types of features after spatial alignment and consistency verification. Here, multimodality refers to data sources from different physical observation mechanisms, and associative encoding refers to an encoding method that further introduces the correlation between modes based on splicing. To quantify the degree of consistency among the three types of features, a consistency score can be constructed, the expression of which is:

[0033] Where C represents the overall consistency score of the current structural observation unit; M represents the number of consistency items participating in the score, which may include spatial consistency, temporal consistency, and channel reliability; This represents the weight of the k-th consistency item, used to reflect the importance of different verification items; This represents the score result of the k-th consistency item. Only when the overall consistency score meets the preset conditions will the three types of features corresponding to the current structural observation unit enter the subsequent fusion stage.

[0034] After the candidate multimodal feature vectors are formed, further correlation compression and discriminative enhancement processing are required to generate structural multimodal feature representations. Correlation compression is used to eliminate redundant information in the three types of features. For example, image texture features and sound wave propagation features may simultaneously respond to local void boundaries. Directly superimposing all features would increase feature dimensional redundancy and weaken model stability. Therefore, methods such as principal component projection, low-dimensional embedding, or autoencoder compression are needed to compress repeatedly expressed feature components into a more compact feature space. Discriminative enhancement is used to highlight the feature components most relevant to interface state changes. This can be achieved through class separation constraints, contrastive learning constraints, or attention weighting to increase the weight of key features such as crack continuity, frequency band anomaly migration, and echo enhancement in the final representation. Here, correlation compression refers to reducing intermodal repetition or linear correlation, discriminative enhancement refers to improving features directly related to the ability to distinguish the target state, and structural multimodal feature representation refers to a unified feature carrier that comprehensively characterizes the surface state, dynamic response state, and interface propagation state of the lining structure. If a weighted enhancement mechanism is used, the enhanced features can be represented as:

[0035] in, This represents the generated structural multimodal feature representation; P represents the number of feature components retained after compression. The weight represents the enhancement weight of the j-th feature component, which reflects its importance in interface state discrimination; This represents the value of the j-th feature component. This processing allows the final feature representation to retain sensitivity to interface voids, structural cracks, and propagation anomalies in the surrounding rock lining while maintaining dimensional control.

[0036] Finally, all structural multimodal feature expressions are aggregated according to the correspondence of structural observation units to form a structural feature expression set. This aggregation is not a simple stacking; rather, it involves writing the structural multimodal feature expression corresponding to each structural observation unit, along with its spatial coordinates, timestamp, and channel reliability results, into a unified data organization structure. This forms a feature sample set that can be directly used for subsequent interface state discrimination model training and inference. The structural feature expression set is essentially a multi-temporal feature library built for the entire lining structure region, where each record corresponds to a specific structural observation unit. Therefore, the interface state prediction results output by the subsequent model can be unambiguously mapped back to the original lining structure location. The aggregation here refers to integrating discrete unit-level features into a globally searchable feature set, and the structural observation unit correspondence means that each feature expression maintains a unique mapping relationship with the original acquisition area. To ensure consistency in subsequent training and inference, sample index construction, missing item registration, and version identifier binding are typically performed on the structural feature expression set, so that the set can be continuously updated without disrupting the original mapping relationship after new observation data is added.

[0037] In one possible implementation, a lining structure interface state discrimination model is constructed based on a set of structural feature representations, and historical samples are used to perform model training. Simultaneously, interface state inference processing is performed on the set of structural feature representations to obtain a set of predicted structural interface states. Specifically, this includes: constructing a set of historical samples labeled with interface states based on the set of structural feature representations; performing sample balancing and subset partitioning on the historical sample set to obtain training sample subsets, validation sample subsets, and test sample subsets; constructing a lining structure interface state discrimination model based on the training sample subset, comprising an input layer, a feature association layer, a state discrimination layer, and a result output layer, and establishing a multimodal model in the feature association layer. The model establishes the correlation and mapping relationships between features; it trains parameters for the interface state discrimination model of the lining structure using a subset of training samples, and iteratively corrects the model parameters based on the state discrimination bias results; it performs validation and evaluation on the interface state discrimination model of the lining structure using a subset of validation samples, and performs collaborative adjustment on the feature association layer and state discrimination layer based on the validation and evaluation results to determine the target model parameters; it performs generalization testing on the interface state discrimination model of the lining structure using a subset of test samples, and performs model deployment and solidification when preset conditions are met; it inputs the set of structural feature expressions into the interface state discrimination model of the lining structure to perform interface state inference processing, and outputs the set of prediction results corresponding to the structural observation units.

[0038] Specifically, a historical sample set for supervised training is first established around the set of structural feature expressions. This historical sample set is not a simple storage of existing features, but rather a sample set formed by binding the multimodal structural feature expression corresponding to each structural observation unit with the interface state label corresponding to that structural observation unit during actual detection or annotation. The interface state label is used to characterize the true state of the surrounding rock lining interface and can be assigned values ​​based on borehole verification results, manual review results, rebound detection results, acoustic transmission diagnostic results, long-term monitoring results, or cross-confirmation results from multiple sources. To ensure a strict correspondence between the interface state label and the multimodal structural feature expression, sample consistency verification needs to be performed before the samples are entered into the database. The verification content includes at least whether the structural observation unit identifiers are consistent, whether the spatial coordinate identifiers are consistent, whether the timestamp identifiers are within a valid corresponding window, and whether the source of the interface state label is reliable. Here, the interface state label refers to the target state identifier used for supervised model learning, and the historical sample set refers to the training basis composed of structural observation unit samples whose states have been confirmed in historical periods. If it is necessary to quantify the reliability of the labels, a label credibility index can be established, the expression of which is:

[0039] in, This represents the credibility of the interface state label corresponding to the current sample, used to measure whether the label can participate in subsequent model training; N represents the number of evidence sources participating in the credibility assessment. This represents the weight of the i-th type of evidence source, used to reflect the differences in the reliability of different verification methods; The confidence score represents the source of evidence of type i, which can be obtained based on the degree of consistency in verification, the completeness of data collection, or the results of cross-validation. This process avoids low-reliability labels from directly entering the training process and affecting the model's discrimination boundary.

[0040] After forming the historical sample set, it is necessary to perform sample balancing and subset partitioning. The purpose of sample balancing is to alleviate training bias caused by excessive differences in the number of samples between different interface state labels. For example, samples of intact interface states are usually far more numerous than those of slightly decoupled interface states and damaged interface coupling states. Without balancing, the model is prone to over-biasing towards the majority class. This can be achieved through methods such as minority class sample augmentation, nearest neighbor sample amplification, spatiotemporal perturbation generation, and majority class sample constraint screening to maintain a relatively balanced distribution of samples corresponding to different interface state labels. Subset partitioning, after balancing, divides the historical sample set into training, validation, and testing subsets. The training subset is used for model parameter learning, the validation subset for model structure adjustment and threshold correction, and the testing subset for verifying the final model's adaptability to samples not used in training. To prevent spatially adjacent and temporally similar samples from entering different subsets simultaneously, leading to information leakage, the partitioning should be performed hierarchically based on the structural region and time interval to which the structural observation unit belongs. Here, sample equalization refers to balancing the quantity and features of samples with uneven class distribution. Sample subset partitioning refers to assigning samples to independent sets according to different uses. To measure the degree of sample equalization across classes, a sample equalization coefficient can be defined, whose expression is:

[0041] in, The equilibrium coefficient represents the historical sample set, used to reflect whether the distribution of the number of samples of different interface state labels is close to equilibrium; M represents the number of categories of interface state labels. This represents the number of samples corresponding to the k-th type of interface status label; This represents the average number of samples in each category. The closer this coefficient is to a higher level, the more balanced the number of samples in different categories, which is more conducive to the model learning the differences between various interface states.

[0042] After obtaining a subset of training samples, a state discrimination model for the lining structure interface is constructed around this subset. This model employs a hierarchical neural network structure, where the input layer receives multimodal feature representations of the structure, the feature association layer models the relationships between image texture features, vibration response features, and sound wave propagation features, the state discrimination layer extracts discrimination patterns directly related to interface state changes, and the output layer provides the interface state category and corresponding state confidence. The core function of the input layer is to preserve the original semantic meaning of the structural feature representation set and to standardize and organize the feature vectors. The core function of the feature association layer is to bring the multimodal features, which originally described surface texture, dynamic response, and propagation behavior separately, into the same discrimination space, enabling the model to learn the coupling rules of these features in the interface integrity state, interface micro-void state, interface expanded void state, and interface damaged coupling state. The state discrimination layer further transforms the features with established association mappings into state distinction boundaries. The output layer outputs the classification results for each interface state. Here, the feature association layer refers to a network layer structure specifically designed to learn the dependencies between features of different modalities. The association mapping relationship refers to the correspondence between multimodal features in the same latent space. To enhance the mutual constraints between multimodal features, an association strength matrix can be constructed in the feature association layer, the expression of which is:

[0043] in, This represents the correlation strength between the i-th and j-th feature classes, reflecting the consistency of different modal features in the current sample. and These represent the two types of feature vectors to be compared; This represents the inner product of two types of eigenvectors, used to characterize directional consistency. and These represent the lengths of the corresponding feature vectors, used to normalize the inner product result. Through this processing, the model can more stably learn the synergistic relationship between image texture anomalies, vibration propagation anomalies, and sound wave propagation anomalies.

[0044] After the initial model structure is built, parameter training is performed on the interface state discrimination model of the lining structure using a subset of training samples. The core of parameter training is to input the multimodal feature representations of the structure from the training sample subset into the model. After being transformed step-by-step through the input layer, feature association layer, and state discrimination layer, the model outputs a prediction result of the interface state category. This prediction result is then compared with the interface state labels of the corresponding historical samples, and the internal parameters of the model are updated based on the deviation. Here, the parameters refer to the weights and biases involved in feature mapping and state discrimination calculation within the model. Iterative correction refers to gradually optimizing these parameters through multiple rounds of repeated training. The state discrimination bias result reflects the degree of difference between the model's current output and the true label; essentially, it is a feedback signal indicating whether the model's learning is adequate. To enable the model to gradually converge to a better discrimination boundary, a state discrimination loss can be constructed, the expression of which is:

[0045] Where L represents the state discrimination loss of the current sample, which is used to measure the deviation between the model output and the real interface state label; M represents the number of interface state categories; This represents the actual label value corresponding to the k-th type of interface state. When the current sample belongs to this type, a preset valid value is used; otherwise, a zero value is used. This represents the predicted probability of the k-th type of interface state output by the model. The smaller this loss is, the closer the model output is to the actual interface state. Applying this loss to the model parameters essentially strengthens the correct feature association patterns and weakens invalid mapping relationships that are detrimental to distinguishing interface states.

[0046] After several rounds of parameter training, a validation sample subset is used to perform validation evaluation on the interface state discrimination model of the lining structure. Validation evaluation does not involve retraining the model; instead, it uses samples independent of the training sample subset to test the current model's recognition performance across different interface state categories. Based on the evaluation results, the feature association layer and the state discrimination layer are collaboratively adjusted. The adjustment of the feature association layer focuses on changing the distribution of association weights between multimodal features to prevent the model from over-relying on a single modality. The adjustment of the state discrimination layer focuses on correcting the boundary sensitivity between different interface states, enabling the model to maintain low false alarms for intact interface states while enhancing its early recognition ability for slightly detached interface states. Here, validation evaluation refers to using samples not involved in parameter learning to check the performance of the sub-model, and collaborative adjustment refers to simultaneously modifying the configuration of the feature association layer and the state discrimination layer to ensure consistency in their discrimination objectives. To measure the overall discrimination performance of the model on the validation sample subset, a comprehensive validation index can be calculated, expressed as:

[0047] Wherein, V represents the comprehensive validation index, used to rank the merits of the current model parameter combination; A represents the overall recognition accuracy, used to reflect the proportion of correctly classified samples in all samples; R represents the abnormal state recall, used to reflect the model's recognition coverage of interface micro-void state, interface expansion void state, and interface damage coupling state; P represents the abnormal state discrimination accuracy, used to reflect the proportion of truly abnormal samples among the samples judged as abnormal. , and These represent the weights of the three metrics, used to balance overall accuracy and anomaly detection sensitivity based on the task objective. Selecting the optimal parameter combination based on the comprehensive validation metrics allows the model to achieve a more reasonable performance distribution across multi-class discrimination tasks.

[0048] Once the validation evaluation results meet the preset requirements, a generalization test is performed on the interface state discrimination model of the lining structure using a subset of test samples. The purpose of the generalization test is to confirm that the model is not only applicable to the structural observation units covered by the training and validation sample subsets, but also capable of making stable discriminations for unseen structural regions, unseen time intervals, and samples with different background conditions. If the model maintains a high interface state discrimination ability in the test sample subset, it indicates that it has learned a universal structural feature and interface state mapping law, rather than a mechanical memorization of historical samples. When the preset conditions are met, model deployment and solidification can be performed. Model deployment and solidification includes freezing the target model parameters, binding the model version identifier, binding the training data source identifier, binding the applicable structural region range, and binding the output interface standard, so that the model can directly enter the actual inference stage. Here, the generalization test refers to testing the model's ability to maintain discrimination ability when facing new samples, and deployment and solidification refers to putting the tested model into actual use as a stable version. If it is necessary to quantify the generalization stability of the model, a generalization offset index can be defined, the expression of which is:

[0049] Where G represents the generalization bias of the model, which is used to reflect the performance difference of the model between the validation sample subset and the test sample subset; This represents the overall performance metric of the model on the validation sample subset; This represents the model's overall performance on a subset of test samples. A smaller value indicates that the model's performance on unseen samples is closer to its performance during the validation phase, and its generalization ability is more stable. Only when the generalization offset is within acceptable limits is it suitable to use the model for subsequent interface state inference tasks.

[0050] After the model deployment and solidification are completed, the current set of structural feature expressions to be judged is input into the lining structure interface state discrimination model to perform interface state inference processing. This inference processing refers to using the solidified model to perform forward computation on new input samples without updating model parameters, thereby outputting the interface state category result and state confidence score for each structural observation unit. In actual execution, the structural multimodal feature expressions in the set of structural feature expressions are read sequentially according to the correspondence of structural observation units. These are then input into the model's input layer for feature standardization and dimensional organization. Cross-modal correlation information is extracted through the feature association layer, and the interface state discrimination result is formed through the state discrimination layer. Finally, the result output layer provides the interface state category and state confidence score. The state confidence score refers to the model's degree of certainty regarding the current interface state category output result, which can be used for weight control in subsequent anomaly analysis and spatial feature fusion. If it is necessary to determine the final interface state category from the model output, the maximum state attribution principle can be used, the expression of which is:

[0051] in, M represents the final interface state category of the current structural observation unit; M represents the number of interface state categories. represents the predicted probability of the k-th interface state output by the model; arg max represents selecting the class with the highest predicted probability from all interface state classes as the final result. Based on this process, each structural observation unit can obtain a unique interface state class and its corresponding state confidence.

[0052] After all structural observation units have completed interface state inference, the interface state category results, state confidence scores, spatial coordinate identifiers, timestamp identifiers, and structural observation unit identifiers corresponding to each structural observation unit are aggregated according to a unified mapping relationship to form a prediction result set. This prediction result set is not simply a stack of classification outputs, but rather a carrier of interface state distribution results established for the entire lining structure region. Each result maintains a unique mapping relationship with the original structural observation unit. Therefore, when performing vibration propagation anomaly analysis on the structural vibration response signal, the corresponding interface state category results and state confidence scores can be directly retrieved based on the structural observation unit identifier and spatial coordinate identifier. The prediction result set can also be further subjected to spatial continuity verification and temporal stability verification to identify local anomaly jumps or conflicts between adjacent unit states. If necessary, conflict results can be sent back to the local verification process. Here, the prediction result set refers to the total set of interface state results output by the model for all structural observation units to be identified, and the unified mapping relationship refers to the unique correspondence between the result output and the original structural observation unit. Through this processing, the model inference stage and the subsequent anomaly analysis stage can maintain consistency in spatial, temporal, and unit references.

[0053] Furthermore, the structure of the lining structure interface state discrimination model can be constructed around an input layer, a feature standardization layer, a feature association layer, a state discrimination layer, a result output layer, and auxiliary constraint units. Its core objective is not simply to complete category classification, but rather to organize the image texture features corresponding to the structural observation image, the vibration response features corresponding to the structural vibration response signal, and the acoustic propagation features corresponding to the structural acoustic propagation signal into an interpretable interface state discrimination process within the same discrimination chain. Since anomalies in the surrounding rock lining interface typically manifest simultaneously as changes in surface texture, abnormal dynamic response, and propagation state mismatch, the model structure needs to possess multimodal correlation modeling capabilities, hierarchical state separation capabilities, and spatial unit consistency maintenance capabilities to avoid directly misjudging occasional anomalies in a single mode as interface voids or interface damage.

[0054] The input layer receives structural multimodal feature representations from the set of structural feature representations. Each structural multimodal feature representation is bound to a unique structural observation unit and retains spatial coordinate identifiers, timestamp identifiers, and channel quality identifiers. The input layer's role is twofold: first, to load features, ensuring that image texture features, vibration response features, and sound wave propagation features maintain a fixed order and fixed dimensional definition when entering the model; second, to unify the basic scale, preventing dominant shifts in subsequent association learning caused by excessive differences in the value ranges of features from different sources. This input layer does not undertake the final discrimination task but rather transforms the original structural multimodal feature representations into a unified input tensor suitable for propagation within the network. To suppress amplitude differences between different feature dimensions, a standardization process can be added after the input layer, with the expression:

[0055] in, This represents the result of standardizing the i-th input feature; This represents the i-th original input feature; This represents the statistical center value of the i-th feature in the training samples; This indicates the degree of dispersion of the i-th feature in the training samples; This represents a stable term to prevent the denominator from being too small. The principle behind this process is to remove the center offset of different features and compress discrete differences, so that multimodal features are in a comparable state in subsequent association layers.

[0056] Following the feature normalization layer, a modal branch coding unit is typically set up. The role of the modal branch coding unit is to first enhance the local representations of image texture features, vibration response features, and sound wave propagation features separately, before proceeding to subsequent cross-modal fusion. Because the statistical structures and physical semantics of the three types of features are not consistent—image texture features emphasize spatial local texture continuity, vibration response features emphasize frequency band shift, damping changes, and local resonance anomalies, and sound wave propagation features emphasize first-wave arrival changes, propagation attenuation, and echo enhancement—an independent coding branch can be set for each type of feature. The image texture feature branch can use a combination of fully connected mapping and local attention to enhance surface texture anomalies; the vibration response feature branch can use temporal convolutional mapping or frequency domain compression mapping to enhance dynamic features; and the sound wave propagation feature branch can use propagation mode embedding mapping to enhance interface propagation features. Essentially, the modal branch coding unit first aggregates features of different modalities into similar categories, then creates clearer semantic boundaries for subsequent cross-modal association.

[0057] Following the modal branching coding unit, the feature association layer is entered. This layer is a crucial structure in the entire lining structure interface state discrimination model. Its task is not simply to concatenate different modal features, but to learn the coupling relationships between image texture anomalies, vibration propagation anomalies, and sound wave propagation anomalies in a unified latent space. In the micro-void state of the surrounding rock lining interface, these three types of features often exhibit weak correlation changes simultaneously; in the extended void state, the common anomalies of the three types of features are significantly enhanced; and in the interface damage coupling state, multimodal synchronous mismatch also occurs. Therefore, the feature association layer needs to both preserve the features of each modality and explicitly model cross-modal dependencies. This can be achieved through interactive attention units, i.e., calculating the correlation weights between different modalities separately, so that the degree of anomaly of a certain modality in the current structural observation unit can influence the feature reweighting of other modalities. Let the image texture feature encoding result be... The vibration response feature encoding result is The sound wave propagation feature encoding result is The association weights of image texture features with vibration response features can be expressed as:

[0058] in, This represents the association weight between image texture features and vibration response features; The parameters represent the correlation mapping between image texture features and vibration response features; This represents the encoding results of other modal features involved in the association comparison. The principle behind this process is to determine the strength of the influence of the current modal feature on other modalities during fusion by calculating the degree of matching of different modal features in a unified mapping space, thereby allowing the model to automatically learn which modal combinations are better able to represent interface state changes.

[0059] The feature association layer can also contain feature gating units. Feature gating units are used to suppress feature components that are irrelevant to the current interface state or have low reliability. For example, in environments with dust, water stains, or strong local lighting changes, image texture features may exhibit false anomalies; during periods of strong mechanical disturbance, vibration response features may also show unstructured fluctuations; in interface areas with significant multipath reflections, sound wave propagation features may be interfered with by boundary echoes. Feature gating units can apply dynamic gating coefficients to each modal feature based on the feature consistency and channel quality information of the input samples, weakening the weights of low-reliability modes and strengthening the weights of high-reliability modes. The gating output can be expressed as:

[0060] in, This represents the result of gating modulation of the i-th modal feature; The gating coefficient represents the i-th mode, and its value is determined by the channel quality, feature consistency, and local anomaly features of the current structural observation unit. This represents the original encoded features of the i-th modality. The principle behind this process is to control the degree of participation of different modalities in the discrimination process through a gating mechanism, thereby reducing the interference of false anomaly propagation on the state discrimination boundary.

[0061] Following the feature association layer, the state discrimination layer begins. This layer undertakes the actual task of separating the interface states and can employ a hierarchical discrimination structure rather than a single-layer direct classification structure. The hierarchical discrimination structure is more suitable for the evolutionary nature of the lining structure's interface states, because the intact interface state, the slightly voided interface state, the expanded voided interface state, and the coupled interface damage state are not completely parallel categories, but rather have a continuous evolutionary relationship from mild to severe. The state discrimination layer can first set up a first-level discrimination unit to distinguish between the intact interface state and the abnormal interface state; then set up a second-level discrimination unit to distinguish between the slightly voided interface state and the significantly abnormal interface state within the abnormal interface state; finally, set up a third-level discrimination unit to distinguish between the expanded voided interface state and the coupled interface damage state within the significantly abnormal interface state. The advantage of this construction is that the model first learns coarse-grained boundaries and then fine-grained boundaries, which better reflects the actual damage evolution law of the surrounding rock lining interface and is also more conducive to enhancing the sensitivity of early slightly voided states.

[0062] A residual mapping structure can be introduced within the state discrimination layer. The purpose of the residual mapping structure is to prevent the discrimination features from degrading during propagation as the network depth increases, and to retain information from previous layers that is still useful for interface state discrimination. For the task of discriminating the interface state of lining structures, some low-order features, such as local gray-level abrupt changes, slight drift upon first wave arrival, and local vibration energy changes, although simple in form, are very sensitive to early micro-void states. Relying entirely on high-order, deep features may weaken these early anomalous signals. Therefore, the state discrimination layer can directly superimpose effective features from previous layers into the output of subsequent layers using residual units, as expressed by:

[0063] in, This represents the input features of the l-th layer; This represents the feature transformation result of the l-th layer after nonlinear mapping; This represents the mapping parameters of this layer; This represents the output features after adding residual connections. The principle behind this process is to retain the original valid information and superimpose the newly extracted discriminative information from the current layer, thereby improving the model's ability to preserve subtle interface anomalies.

[0064] The output layer, located after the state discrimination layer, outputs the interface state category and state confidence score for each structural observation unit. The output layer typically employs a multi-class classification mapping to compress the high-dimensional discriminative features output by the state discrimination layer to an output dimension equal to the number of interface state categories. Each dimension of the output represents the classification strength of the current structural observation unit to a certain interface state category, which is then normalized to obtain the state confidence score distribution. Let the original output of the output layer be... Then the predicted probability of the k-th interface state can be expressed as:

[0065] in, This represents the predicted probability that the current structural observation unit belongs to the k-th type of interface state; This represents the original output value of the output layer for the k-th interface state; M represents the total number of interface state categories. The principle of this processing is to map the original output of all categories to a sum-constrained attribution distribution, enabling the model to simultaneously output the final category and the corresponding state confidence.

[0066] To enhance the model's discriminative continuity for spatially adjacent structural observation units, spatial consistency constraint units can be inserted before the output layer. These units utilize the spatial adjacency relationships between structural observation units to perform local neighborhood correction on the discriminative features of the current unit. Since real-world rock lining interface voids typically exhibit regional continuity and do not appear isolated on only a single structural observation unit, spatial consistency constraint units can suppress isolated spurious anomalies and enhance the discriminative stability of continuous anomaly regions. Neighborhood aggregation results can be constructed based on the feature similarity and spatial distance of each structural observation unit in the neighborhood, expressed as:

[0067] in, This represents the spatial consistency characteristics of the i-th structural observation unit after neighborhood aggregation; Represents the spatial neighborhood set of the i-th structural observation unit; This represents the influence weight of the j-th structural observation unit in the neighborhood on the current unit. This weight can be determined comprehensively based on spatial distance, feature similarity, and temporal proximity. This represents the discrimination feature of the j-th structural observation unit within its neighborhood. The principle behind this processing is to utilize the continuous anomalous support of local regions to correct the unit-level discrimination results, making the model output closer to the actual interface state space distribution.

[0068] In addition to the main discriminant structure, the interface state discrimination model for the lining structure can also include auxiliary loss units. These auxiliary loss units constrain the model's learning process from multiple perspectives during training, rather than relying solely on the final classification loss. For example, class separation constraints can be set to maintain a larger gap between intact and abnormal interface states in the feature space; center clustering constraints can be set to make samples of the same interface state category more compact in the latent space; and modal consistency constraints can be set to ensure stronger collaborative responses of multimodal features under the same structural observation unit in abnormal regions. The advantage of this construction is that the model not only learns the final label results but also learns whether the intermediate layer structure conforms to the physical evolution law of the surrounding rock lining interface state.

[0069] From an overall structural perspective, the interface state discrimination model for lining structures can be understood as a hierarchical discrimination network consisting of front-end modal branch encoding, mid-section cross-modal correlation modeling, rear-section hierarchical state separation, and terminal probability output. The front end organizes heterogeneous features into comparable coded representations; the mid-section identifies whether there are common correlation patterns among multiple modalities pointing to interface state changes; the rear end separates different states layer by layer according to the interface state evolution relationship; and the terminal outputs the interface state category and state confidence at the structural observation unit level. This structural design is suitable for deep underground laboratory lining structure scenarios because interface anomalies in such scenarios are often not single manifestations but rather the result of changes in multiple aspects, including surface texture, dynamic propagation, and sound wave propagation. Through this structure, the model can uniformly map multi-source heterogeneous observation information into interface state prediction results with clear semantics, providing stable input for subsequent vibration propagation anomaly analysis and the construction of spatial void probability distributions.

[0070] In one possible implementation, vibration propagation anomaly analysis is performed on the structural vibration response signal based on the prediction result set to identify potential void areas at the surrounding rock lining interface and obtain anomaly analysis results for vibration propagation. Specifically, this includes: establishing a mapping relationship between structural observation units, structural vibration response signals, and structural interface states based on the prediction result set; performing benchmark construction processing on the structural vibration response signal corresponding to the interface stability reference unit in the structural observation unit to form a reference vibration benchmark; performing vibration propagation pre-analysis processing on the structural vibration response signal based on the mapping relationship to obtain vibration response segments with unified propagation starting points and propagation time windows; and performing vibration propagation feature reconstruction processing on the vibration response segments to... A vibration propagation characterization sequence corresponding to the structural observation unit is formed; vibration propagation deviation analysis is performed based on the reference vibration benchmark and the vibration propagation characterization sequence to obtain vibration propagation deviation results; joint constraint processing is performed on the vibration propagation deviation results and the set of prediction results to form vibration propagation anomaly judgment results corresponding to the interface state category; anomaly seat allocation processing is performed on the vibration propagation deviation results to determine the candidate units of propagation anomalies and the propagation anomaly clustering areas; propagation path backtracking processing and interface voiding direction analysis processing are performed around the propagation anomaly clustering areas to identify potential voiding areas of the surrounding rock lining interface; anomaly analysis result generation processing is performed on the potential voiding areas of the surrounding rock lining interface to form anomaly analysis results.

[0071] Specifically, a one-to-one mapping relationship is first established between structural observation units, structural vibration response signals, and structural interface states based on the prediction result set. This mapping relationship means that, under the constraints of the same structural observation unit identifier, the same spatial coordinate identifier, and the same timestamp identifier, the interface state category and state confidence information in the structural interface state prediction results are accurately bound to the corresponding structural vibration response signal. This ensures that subsequent vibration propagation anomaly analysis always revolves around the same lining structure area, avoiding the problem of misalignment between the state source and the signal source. In specific processing, each prediction record in the prediction result set is first read, and the structural observation unit identifier, spatial coordinate identifier, timestamp identifier, interface state category, and state confidence information are extracted. Then, the corresponding structural vibration response signal segment is retrieved from the vibration acquisition records. If the vibration acquisition frequency is higher than the prediction result generation frequency, the closest vibration sampling segment is selected within the allowed time window based on the timestamp identifier, and this segment is registered under the corresponding structural observation unit. If the same structural observation unit has continuous prediction results at multiple times, a temporal mapping chain is further established, enabling the vibration propagation anomaly analysis at the current time to continuously compare with previous interface state results. Here, the structural interface state refers to the interface integrity state, interface micro-void state, interface expanded void state, or interface damage coupling state output by the lining structure interface state discrimination model. State confidence information refers to the reliability of the model's judgment on the current interface state category. To quantify the mapping quality, a mapping matching degree can be constructed, whose expression is:

[0072] in, This indicates the mapping matching degree of the current structural observation unit, which is used to measure the reliability of the correspondence between the prediction results and the structural vibration response signal; It indicates the degree of spatial coordinate matching, which can be calculated based on whether the spatial positions are consistent or the spatial distance deviation. This indicates the degree of timestamp matching, which can be calculated based on whether the time difference is within the allowed window. This indicates the degree of matching between structural observation unit identifiers, reflecting whether the unit indices are completely consistent; , and These represent the weights of the three matching factors, used to adjust the influence of space, time, and element identification in the mapping. This process ensures that the vibration response signals entering subsequent anomaly analysis maintain a strict correspondence with the predicted result set.

[0073] After establishing the mapping relationship, benchmark construction processing is performed on the structural vibration response signals corresponding to the interface stable reference units to form a reference vibration benchmark. Here, an interface stable reference unit refers to a structural observation unit that is determined to be in an intact interface state in the prediction result set, has a high level of state confidence information, and maintains stable state across multiple consecutive time stamps. The role of the reference vibration benchmark is to provide a standard vibration propagation pattern under normal surrounding rock lining interface conditions, enabling comparison of the vibration propagation behavior of subsequent abnormal units with normal propagation behavior. Specifically, reference units that meet the interface stability conditions are first selected from all structural observation units. Then, noise suppression, excitation amplitude normalization, propagation distance normalization, and channel consistency correction are performed on the structural vibration response signals corresponding to these reference units. Subsequently, the benchmark amplitude envelope, benchmark frequency distribution, benchmark propagation delay, and benchmark attenuation trend are statistically formed. This benchmark construction processing is not a simple average waveform calculation, but rather a set of standard propagation templates that can be used for comparison, based on the relationship between vibration propagation path, propagation intensity, and propagation time. To obtain the average amplitude envelope in the reference vibration benchmark, the following formula can be used:

[0074] in, The average amplitude envelope of the reference vibration reference at time t is used to characterize the standard energy profile of vibration propagation under normal interface conditions; N represents the number of interface stable reference elements involved in the reference construction. This represents the vibration envelope value of the i-th stable interface reference element at time t. The principle behind this process is to use the stable propagation results of multiple reference elements to cancel out occasional disturbances, thereby obtaining a reference template that better characterizes the propagation law of the normal interface.

[0075] Based on the aforementioned mapping relationship, vibration propagation pre-analysis processing is performed on the structural vibration response signal to obtain vibration response segments with a unified propagation starting point and propagation time window. This vibration propagation pre-analysis processing refers to performing event identification, starting point location, time window segmentation, and channel synchronization correction on the original structural vibration response signal before entering anomaly analysis, ensuring that the vibration responses in different structural observation units are under a unified analysis benchmark. The vibration propagation starting point refers to the initial moment when vibration energy begins to enter the current propagation path, and the propagation time window refers to an effective analysis time interval intercepted around this starting point, used to ensure that the comparison between each structural observation unit is established at the same propagation stage. Specifically, effective excitation events are first detected in the continuous vibration signal, and then the propagation starting point is determined based on energy mutation points, waveform first jump points, or gradient rise points. Then, a background noise interval is reserved forward from the propagation starting point, and a complete propagation response interval is intercepted backward, thus obtaining a vibration response segment of uniform length. If multi-channel vibration acquisition exists, further time delay compensation is needed to align the records of the same vibration event across channels on the time axis. The propagation starting point can be detected based on instantaneous energy changes, and its expression is:

[0076] in, The value represents the local energy accumulation up to time t, used to determine whether an effective vibration propagation starting point has appeared near the current time; L represents the length of the local energy calculation window, used to limit the range of the statistical interval. This represents the vibration amplitude at the k-th sampling point. When the local energy accumulation value continuously increases relative to the background interval and exceeds a preset judgment condition, the corresponding moment can be determined as the vibration propagation starting point. Through this processing, vibration response segments corresponding to different structural observation units can have a unified starting benchmark and a unified analysis window.

[0077] After obtaining the vibration response segment, vibration propagation feature reconstruction processing is performed on the vibration response segment to form a vibration propagation characterization sequence corresponding to the structural observation unit. This vibration propagation feature reconstruction processing refers to not directly using the original waveform for anomaly analysis, but further extracting the most discriminative propagation features from the original waveform and organizing them into a continuous characterization sequence according to time sequence or propagation path sequence. The vibration propagation characterization sequence can include at least the propagation amplitude envelope sequence, frequency band energy transfer sequence, propagation time delay change sequence, phase continuous change sequence, and local resonance response sequence, thereby describing the true state of vibration propagation in the current structural observation unit from multiple perspectives. Essentially, this feature reconstruction transforms the scattered propagation information in the original waveform into a structured propagation representation that is easier to compare and discriminate. For the propagation amplitude envelope, Hilbert transform or local peak tracking can be used; for frequency band energy transfer, short-time spectrum analysis can be used to statistically analyze the energy changes over time in multiple frequency bands; for propagation time delay changes, it can be calculated through alignment offset with the reference template. If the propagation time delay offset needs to be characterized, the following formula can be used:

[0078] in, It indicates the propagation time delay offset of the current structural observation unit relative to the reference vibration reference, and is used to reflect whether the vibration propagation path is blocked, detoured, or has a local decrease in continuity; This represents the measured time delay of effective propagation arrival in the current vibration response segment; This represents the standard propagation delay in the reference vibration standard. When the propagation path is affected by interface delamination, the propagation delay usually deviates from the normal level, thus forming anomalies in the vibration propagation characterization sequence.

[0079] After forming the vibration propagation characterization sequence, vibration propagation deviation analysis is performed based on the reference vibration benchmark and the vibration propagation characterization sequence to obtain the vibration propagation deviation results. This vibration propagation deviation analysis refers to comparing the vibration propagation characterization sequence corresponding to the current structural observation unit with the corresponding propagation template in the reference vibration benchmark item by item, thereby quantifying the degree of deviation of the current propagation behavior from normal propagation behavior. Deviation analysis typically revolves around propagation amplitude deviation, frequency distribution deviation, propagation delay deviation, phase continuity deviation, and attenuation trend deviation. Specifically, propagation amplitude deviation describes whether vibration energy is abnormally weakened or abnormally concentrated; frequency distribution deviation describes whether the dominant frequency band has shifted; propagation delay deviation describes whether the propagation path is obstructed; and attenuation trend deviation describes the impact of interface stiffness changes on propagation dissipation. To obtain unified vibration propagation deviation results, multiple deviation indices can be combined and calculated, with the following expression:

[0080] in, This indicates the vibration propagation deviation result of the current structural observation unit, which is used to comprehensively characterize the overall degree to which the current vibration propagation behavior deviates from the reference vibration benchmark; This indicates the deviation in propagation amplitude, which can be calculated from the difference between the current envelope and the reference envelope; This indicates the deviation in frequency distribution, which can be calculated from the difference in energy distribution across frequency bands. Indicates the deviation in propagation delay; Indicates the continuous deviation of the phase; Indicates the deviation of propagation attenuation; to These represent the weights of the corresponding deviations, used to adjust the contribution of different deviation features in the overall deviation analysis. The principle behind this process is to compress multidimensional propagation anomalies into a unified deviation representation, enabling subsequent judgments to take into account multiple propagation mismatches rather than relying on a single anomaly indicator.

[0081] After obtaining the vibration propagation deviation results, joint constraint processing is performed on the vibration propagation deviation results and the prediction results to form a vibration propagation anomaly judgment result corresponding to the interface state category. This joint constraint processing means that anomalies are not judged solely based on the vibration propagation deviation results, but rather the interface state category and state confidence information from the preceding interface state prediction results are introduced into the current vibration propagation deviation judgment, applying different sensitivities and constraint boundaries to different interface state categories. Specifically, for interface stable reference units, stricter anomaly judgment conditions should be adopted to avoid normal fluctuations being misjudged as anomalies; for interface micro-void state units, the sensitivity to slight deviations should be increased to enhance early anomaly detection capabilities; for interface extended void state and interface damage coupled state units, the deviation triggering conditions can be relaxed, allowing persistent anomalies to enter high-risk judgment more quickly. The vibration propagation anomaly judgment result refers to the anomaly level determined for the vibration propagation behavior corresponding to the current structural observation unit after considering the interface state category and state confidence information. The anomaly judgment strength after joint constraints can be constructed using the following formula:

[0082] in, This indicates the intensity of vibration propagation anomaly determination after joint constraints, used to measure the degree of anomaly of the current structural observation unit under interface state constraints; This indicates a deviation in vibration propagation results; This represents the state modulation coefficient determined by the interface state category; different interface state categories can correspond to different modulation intensities. This represents the confidence modulation factor determined by state confidence information, used to enhance the impact of high-confidence prediction results on anomaly identification. The principle behind this process is to introduce the semantics of structural interface states into vibration propagation anomaly analysis, ensuring semantic consistency between anomaly identification and prior state reasoning results.

[0083] After determining the vibration propagation anomaly, anomaly seat allocation is performed on the vibration propagation deviation results to identify candidate units and clusters of propagation anomalies. This anomaly seat allocation process assigns different anomaly seats to anomaly units across all structural observation units based on their vibration propagation deviation results and anomaly determination intensities, thus distinguishing between general anomalies, key anomalies, and continuous anomalies. Candidate units of propagation anomalies are structural observation units that meet the anomaly determination conditions at the unit level. A cluster of propagation anomalies is a local anomaly region formed by multiple candidate units that are spatially adjacent, temporally continuous, and similar in anomaly intensity. Specifically, all structural observation units can be sorted based on the anomaly determination intensities after joint constraints. Then, similar units are merged into the same cluster based on spatial adjacency, while simultaneously checking the continuous occurrence of their cross-timestamp identifiers. Only when multiple candidate units of propagation anomalies simultaneously meet both spatial and temporal continuity conditions are they identified as a cluster of propagation anomalies. To quantify the priority of units in the anomaly seat allocation, the following formula can be used:

[0084] in, This represents the anomalous seat allocation value of the i-th structural observation unit, used to determine whether it enters the candidate unit for propagation anomalies and its priority in cluster analysis; This represents the intensity of the vibration propagation anomaly determination for the i-th structural observation unit; It indicates the number of spatial connections between it and adjacent anomalous units, used to reflect the degree of spatial clustering; This indicates the number of times the anomaly persists within a continuous time window, reflecting the duration of the anomaly. , and These represent the weights of the three factors, respectively. The principle behind this process is to incorporate the intensity of anomalies, spatial clustering, and temporal duration into the anomaly determination, thereby reducing misjudgments caused by sporadic disturbances at single points.

[0085] After identifying the cluster of propagation anomalies, propagation path tracing and interface voiding direction analysis are performed around these clusters to identify potential voiding areas at the surrounding rock lining interface. Propagation path tracing involves tracing the vibration response in the cluster along the direction of vibration propagation to pinpoint the initial location of the anomaly, its propagation direction, and its diffusion trajectory in different structural observation units. Interface voiding direction analysis, based on the known propagation anomaly propagation path, combines interface state type, vibration propagation deviation mode, spatial clustering morphology, and anomaly persistence pattern to determine whether the current anomaly cluster is more likely caused by voiding at the surrounding rock lining interface, rather than by local external disturbances, sensor malfunctions, or temporary structural vibrations. Specifically, the anomaly propagation chain can be reconstructed first along the propagation time sequence and spatial adjacency sequence. Then, it is analyzed whether the beginning and direction of the propagation chain are consistent with the micro-voiding or extended voiding state areas of the interface. If the anomaly propagation path shows a continuous trajectory expanding from a local initial anomaly to the surrounding area, and the propagation deviation is mainly characterized by increased time delay, weakened amplitude, and enhanced attenuation, then it can be determined that it has a clear interface voiding direction. The potential lining interface void area here refers to the lining structure area that is highly suspected of having interface voids, although it has not been verified by direct excavation, but is supported by multiple propagation anomalies and interface condition results.

[0086] After identifying potential void areas at the surrounding rock lining interface, anomaly analysis results are generated for these void areas. These anomaly analysis results are not simply isolated anomaly indicators, but rather a complete anomaly result carrier encompassing the potential void areas. Specifically, they include the spatial boundary of the area, anomaly intensity level, propagation path characteristics, temporal evolution state, the corresponding set of structural observation units, and consistency with the predicted result set. The anomaly intensity level can be determined comprehensively based on vibration propagation deviation results, anomaly determination intensity after joint constraints, and anomaly seat allocation values. The temporal evolution state can be determined as initial development, continuous expansion, or high-risk coupling based on changes in anomaly area, anomaly intensity, and expansion direction under multiple consecutive timestamps. The spatial boundary can be determined comprehensively based on the outer edge of the propagation anomaly cluster area, the endpoint of the propagation path expansion, and the location of anomaly interruption in adjacent units. To characterize the comprehensive anomaly level of the potential void area at the surrounding rock lining interface, the following formula can be used:

[0087] in, The comprehensive anomaly level indicates the potential void area at the surrounding rock lining interface, which is used to uniformly characterize the risk level of this area in vibration propagation anomaly analysis; This represents the average intensity of vibration propagation anomaly determination for each structural observation unit within the region; This represents the average level of the abnormal seat allocation values ​​for each structural observation unit within the region; This indicates the intensity of the anomalous evolution in the region under consecutive timestamps, reflecting whether the anomalous phenomenon continues to intensify or expand. , and These represent the weights of the three indicators. The principle behind this process is to elevate unit-level anomaly results to regional-level risk representation, enabling the generated anomaly analysis results to directly serve subsequent spatial feature fusion processing.

[0088] In one possible implementation, spatial feature fusion processing is performed based on the set of anomaly analysis results and prediction results to construct the spatial void probability distribution of the lining structure. Specifically, this includes: performing same-unit merging processing based on the set of anomaly analysis results and prediction results to form a set of structural fusion basic units; performing spatial neighborhood organization processing on the set of structural fusion basic units to construct spatial neighborhood association information corresponding to the structural observation units; performing interface state quantification mapping processing based on the set of structural fusion basic units and spatial neighborhood association information to obtain interface state quantification characterization values ​​and interface state confidence characterization values; and performing spatial neighborhood association processing based on the set of structural fusion basic units and spatial neighborhood association information. Vibration propagation anomaly quantization mapping is performed to obtain anomaly intensity characterization values, anomaly boundary constraint values, and anomaly trend direction values. Fusion processing is then performed on the interface state quantization characterization values, interface state confidence characterization values, anomaly intensity characterization values, anomaly boundary constraint values, and anomaly trend direction values ​​to obtain element voiding direction values. Based on the element voiding direction values, regional-level extended fusion processing is performed to form candidate voiding regions and select effective voiding regions. Spatial interpolation fusion processing is then performed around the effective voiding regions and structural observation units to construct a continuous spatial voiding characterization field. Finally, probabilistic mapping processing is performed on the spatial voiding characterization field to generate the spatial voiding probability distribution of the lining structure.

[0089] Specifically, the anomaly analysis results and prediction results are first merged into a single unit, forming a unified data carrier under the same structural observation unit. The process involves first reading the structural observation unit identifier, spatial coordinate identifier, timestamp identifier, anomaly region boundary information, anomaly intensity level, and anomaly evolution state from the anomaly analysis results; then reading the structural observation unit identifier, spatial coordinate identifier, timestamp identifier, interface state category, and state confidence information from the prediction results. Subsequently, based on the conditions of complete matching of structural observation unit identifiers, consistent spatial coordinate identifiers, and timestamp identifiers within the same observation period, the corresponding records in the anomaly analysis results and prediction results are merged into the same structural observation unit, thus generating a structural fusion basic unit set. This same-unit merging process refers to the unified binding of data results from different sources but pointing to the same lining structure region; the structural fusion basic unit set refers to a unit-level result set that simultaneously contains interface state characterization information and vibration propagation anomaly characterization information. To avoid mismatches during the merging process, the merging matching degree can be calculated first, with the expression:

[0090] in, This indicates the matching degree of the current records to be merged, used to determine whether the anomaly analysis results and the prediction results can be merged into the same structural observation unit; This indicates the matching result of the structural observation unit identifier. When the identifiers match, a higher value is used; otherwise, a lower value is used. This indicates the spatial coordinate matching result, used to reflect whether the spatial positions are consistent or whether the deviation is within the allowable range; The timestamp indicates the matching result and is used to reflect whether the two results belong to the same time base. , and These represent the weights of the three matching results. This process ensures that every record in the set of basic structural fusion units has a strictly unified spatial and temporal attribution.

[0091] After forming the set of basic structural units, spatial neighborhood organization processing is performed on the set to construct spatial neighborhood association information corresponding to the structural observation units. Specifically, based on the spatial coordinates of each structural observation unit in the lining structure geometric model, the position of each structural observation unit on the lining structure surface or the surrounding rock lining interface is determined. Then, taking the current structural observation unit as the center, and according to a preset spatial radius, adjacent grid topology, or interface extension direction, surrounding adjacent structural observation units are selected to form the spatial neighborhood set corresponding to that structural observation unit. Subsequently, the continuity of interface state categories, the trend of state confidence information changes, the connection relationship of anomaly intensity, the contact relationship of anomaly boundaries, and the consistency of anomaly evolution direction are extracted from each adjacent structural observation unit in the spatial neighborhood set, thereby constructing spatial neighborhood association information. Here, spatial neighborhood organization processing refers to the process of establishing spatial connections between a single structural observation unit and its surrounding units; spatial neighborhood association information refers to the set of information used to describe the state continuity relationship and anomaly extension relationship between the current structural observation unit and its adjacent structural observation units. Since actual voids at the interface of surrounding rock lining typically manifest as continuous anomalies within a certain range, rather than isolated single-point anomalies, it is necessary to transform discrete units into continuously interpretable regional analysis objects through spatial neighborhood organization. To quantify the degree of neighborhood correlation, a spatial neighborhood correlation coefficient can be defined, with the following expression:

[0092] in, represents the spatial neighborhood correlation coefficient of the i-th structural observation unit, used to characterize the degree of continuous support between it and its adjacent structural observation units; Represents the spatial neighborhood set of the i-th structural observation unit; The spatial influence weight of the adjacent structural observation unit j on the current structural observation unit i can be determined based on spatial distance, interface orientation consistency and boundary contact degree. This indicates the degree of state compatibility or anomalous compatibility between the current structural observation unit and its adjacent structural observation units. This process provides a clear neighborhood support relationship for subsequent spatial feature fusion.

[0093] Based on the structural fusion base unit set and spatial neighborhood association information, interface state quantization mapping is performed to obtain interface state quantization representation values ​​and interface state confidence representation values. Specifically, the interface state categories in the prediction result set are first converted into computable continuous quantization representations. For example, the complete interface state is mapped to a lower anomaly pointing value, the slightly detached interface state to a medium-low anomaly pointing value, the expanded detached interface state to a medium-high anomaly pointing value, and the interface damage coupling state to a higher anomaly pointing value. Then, the confidence of the interface state quantization results is modulated using state confidence information, so that high-confidence interface state categories have a stronger effect in subsequent fusion, while the effect of low-confidence interface state categories is suppressed. Furthermore, spatial neighborhood association information is combined to perform neighborhood smoothing correction on the interface state quantization results of the current structural observation unit, making it easier for units in continuous anomaly regions to maintain the continuity of anomaly pointing and appropriately constraining isolated anomaly units. The interface state quantization mapping process here refers to the process of converting discrete interface state categories into continuous, computable representations. The interface state quantization representation value is a continuous numerical value representing the anomaly level of the current structural observation unit's interface. The interface state confidence representation value is a continuous numerical value representing the reliability of the current interface state quantization result. To achieve the mapping from state categories to quantized values, the following expression can be used:

[0094] in, This represents the quantitative characterization value of the interface state, used to characterize the anomaly level of the interface state of the current structural observation unit. Indicates the interface state category; This represents a pre-defined category mapping function used to map different interface state categories to different quantization levels. Based on this, state confidence information can be combined to form a modulation result, the expression of which is:

[0095] in, This represents the quantization result of the interface state after confidence modulation; This represents the quantized representation value of the interface state; This represents the state confidence score, used to characterize the reliability of the current interface state judgment. Through this process, the discrete interface state output is converted into continuous features that can directly participate in spatial fusion calculations.

[0096] Based on the structural fusion of basic unit sets and spatial neighborhood association information, vibration propagation anomaly quantification mapping is performed to obtain anomaly intensity characterization values, anomaly boundary constraint values, and anomaly trend indication values. Specifically, the anomaly intensity level, anomaly region boundary information, and anomaly evolution state corresponding to the current structural observation unit are first extracted from the anomaly analysis results. These discrete or semi-continuous results are then uniformly mapped into continuous quantization features. The anomaly intensity characterization value represents the deviation intensity of vibration propagation anomalies from normal propagation behavior in the current structural observation unit; the anomaly boundary constraint value represents whether the anomalies between the current structural observation unit and adjacent structural observation units form a continuous boundary connection relationship; and the anomaly trend indication value represents whether the anomaly tends to intensify, maintain, or decay over time, and further reflects the persistence of the anomaly expansion direction. During the processing, spatial neighborhood association information can also be combined to perform neighborhood diffusion correction on the anomaly intensity characterization value, enabling units located in continuous anomaly regions to obtain more stable anomaly support, while suppressing isolated anomaly units. The vibration propagation anomaly quantification mapping process here refers to the process of converting the vibration propagation anomaly analysis results into continuous features that can participate in fusion calculations; the anomaly boundary constraint value refers to the numerical value characterizing the connection relationship between the anomaly boundary and the neighborhood boundary; the anomaly trend direction value refers to the numerical value characterizing the direction of anomaly development over time. To construct the anomaly boundary constraint value, the following expression can be used:

[0097] in, This represents the abnormal boundary constraint value, used to characterize the degree of connection between the abnormal boundaries of the current structural observation unit and the adjacent structural observation units; This represents the effective boundary contact length between the current structural observation unit and the neighboring anomaly unit; This represents the total length of the boundary of the current structural observation unit involved in the comparison. A higher value indicates that the current structural observation unit is more likely to be within a continuous anomalous boundary. To characterize the direction of the anomalous trend, the following expression can be used:

[0098] in, This indicates the value pointing to the abnormal trend, used to characterize the degree of enhancement or decay of the anomaly corresponding to the current structural observation unit in adjacent time windows; Indicates the intensity of the anomaly at the current moment; Indicates the intensity of the anomaly at the previous moment; This represents a stable term to prevent the denominator from becoming too small. Through this processing, the anomaly analysis results are transformed into continuous features that possess semantics of intensity, boundary, and time trend.

[0099] After obtaining the interface state quantification value, interface state confidence value, anomaly intensity value, anomaly boundary constraint value, and anomaly trend indication value, respectively, the above features are fused to obtain the unit voiding indication value. Specifically, the five types of features are first normalized to ensure their values ​​are within a comparable range, preventing any single feature from dominating the fusion result due to an excessively large value range. Then, based on the contribution of each feature to the determination of voiding at the surrounding rock lining interface, corresponding fusion weights are set. Subsequently, the interface state quantification value and the interface state confidence value are combined to form the interface anomaly indication component, and the anomaly intensity value, anomaly boundary constraint value, and anomaly trend indication value are combined to form the propagation anomaly indication component. Finally, the interface anomaly indication component and the propagation anomaly indication component are synergistically fused to obtain the unit voiding indication value for the current structural observation unit. Here, the unit voiding indication value refers to the comprehensive quantification result of the probability of voiding at the surrounding rock lining interface in the current structural observation unit. The fusion process is not a simple addition, but rather a unified determination formed while taking into account the semantics of interface state, propagation anomaly, and neighborhood continuity. To achieve the above fusion, the following expression can be used:

[0100] in, The value indicating the voiding direction of the unit is used to characterize the comprehensive probability of voiding at the interface between the surrounding rock and the lining of the current structural observation unit. This represents the quantization result of the interface state after state confidence modulation; Indicates the anomaly intensity characterization value; Indicates abnormal boundary constraint values; Indicates the value to which the abnormal trend points; to These represent the weights of various features in the fusion process. If we consider the interface state confidence representation value participating in modulation independently, a multiplicative constraint can be introduced on top of the above, allowing high-confidence state results to enhance the fusion output. Through this processing, the multi-source anomaly information of the current structural observation unit is uniformly mapped into a single vacancy-free pointer, providing direct input for subsequent regional-level extended fusion.

[0101] Based on the unit void-direction value, a regional-level expansion and fusion process is performed to form candidate void-direction regions and then filter out effective void-direction regions. Specifically, all structural observation units with high unit void-direction values ​​are first located in the geometric model of the lining structure, and these units are connected into several initial clustered regions based on spatial adjacency. Then, within each initial clustered region, the continuity of the unit void-direction value distribution, the compatibility of the interface state hierarchy, the continuity of the anomaly boundary, and the consistency of the anomaly trend are checked. Initial clustered regions that satisfy spatial connectivity, anomaly continuity, and development consistency are identified as candidate void-direction regions. Regions consisting only of scattered high-value units, lacking boundary connections, or lacking trend support are split or eliminated. Subsequently, regional integrity assessment, internal consistency assessment, and regional credibility assessment are further performed on the candidate void-direction regions. Only candidate regions that simultaneously meet these conditions are retained as effective void-direction regions. The region-level extended fusion processing here refers to the process of expanding from unit-level void removal determination to region-level void removal identification; candidate void removal regions refer to spatial regions that initially meet the void removal continuity characteristics; effective void removal regions refer to void removal regions that are considered to have high authenticity after further screening. To calculate the consistency within a region, the following expression can be used:

[0102] in, The consistency coefficient represents the candidate voided region and is used to characterize whether the voided direction value distribution of each structural observation unit within the region is concentrated. This indicates the number of structural observation units contained in the current candidate void region; This represents the cell void pointing value of the i-th structural observation unit; This represents the average value of the void pointing values ​​of all structural observation units within the current candidate void region. The higher this value, the more stable the void pointing distribution within the region, and the more suitable it is for preservation as an effective void region.

[0103] Spatial interpolation fusion is performed around the effective voiding region and structural observation units to construct a continuous spatial voiding characterization field. Specifically, high-confidence structural observation units within the effective voiding region are first used as core control points, and their corresponding voiding direction values ​​are used as interpolation source values. Then, combined with transition units near the boundary of the effective voiding region, interpolation estimation is performed on continuous spatial locations in the lining structure geometric model. This expands the voiding direction results, originally defined only on discrete structural observation units, to a continuous distribution across the entire lining structure surface or surrounding rock lining interface. For spatial locations near the center of the effective voiding region, stronger high-value support should be assigned; for locations near the boundary, boundary correction should be considered to ensure smooth attenuation of the interpolation results along the effective voiding region boundary; for locations far from the effective voiding region, a spatial attenuation mechanism is used to suppress the unsupported anomaly diffusion. Here, spatial interpolation fusion refers to the process of estimating continuous spatial voiding characterization values ​​using discrete unit results; the continuous spatial voiding characterization field refers to the continuously defined voiding anomaly distribution field on the lining structure geometric model. If distance-weighted interpolation is used, the following expression can be used:

[0104] in, The value represents the continuous spatial voiding characterization value at spatial location x, used to characterize the relative strength of the voiding at the surrounding rock lining interface at that location; N represents the number of structural observation units involved in the interpolation. This represents the cell void pointing value of the i-th structural observation unit; denoted by x, the spatial distance between spatial location x and the i-th structural observation unit; p represents the distance decay index, used to control the difference in influence between nearby and distant neighboring units in interpolation. The principle behind this processing is that structural observation units that are closer in distance and have higher void-free pointing values ​​contribute more to the void-free representation of the current spatial location, thus forming a representation field that conforms to the laws of spatial continuity.

[0105] After constructing a continuous spatial voiding characterization field, a probabilistic mapping process is performed on the spatial voiding characterization field to generate a spatial voiding probability distribution for the lining structure. Specifically, the relative strength values ​​in the continuous spatial voiding characterization field are first converted into probability values ​​in a unified probability domain, ensuring that different spatial locations correspond to a voiding probability within a uniform range. Then, the initial probability results are corrected by considering the credibility of the effective voiding region, the regional evolution trend, and the strength of the neighboring support, further enhancing the probability values ​​of highly credible continuous anomaly regions and moderately reducing the probability values ​​of blurred edge regions. Further spatial smoothing and boundary preservation corrections are then performed to ensure the high-probability areas remain spatially continuous while maintaining consistency between the probability abrupt change boundary and the effective voiding region boundary. Here, the probabilistic mapping process refers to converting continuous spatial voiding characterization values ​​into the probability of voiding occurrence; the spatial voiding probability distribution for the lining structure refers to the distribution of the voiding probability at the surrounding rock-lining interface continuously given on the geometric model of the lining structure. To map continuous spatial voiding characterization values ​​to probability values, the following expression can be used:

[0106] in, This represents the probability of spatial voiding of the lining structure at spatial location x, used to characterize the likelihood of voiding at the interface between the surrounding rock and the lining at that location. This represents the continuous spatial void characterization value at spatial location x; The steepness parameter represents the probability mapping and is used to control how sensitive changes in the representation value are to changes in the probability. This represents the probability mapping center position parameter, used to determine the center position of the transition from low probability to high probability. The processing principle lies in transforming continuous representation values ​​into probability distribution results more suitable for subsequent threshold discrimination and region mapping through smoothing nonlinear mapping. Through this processing, the final spatial void probability distribution of the lining structure retains both the local anomaly information of the unit-level fusion results and the continuous spatial relationships reflected by the region-level extended fusion and spatial interpolation fusion.

[0107] In one possible implementation, a fusion process is performed on the interface state quantification value, interface state confidence value, anomaly intensity value, anomaly boundary constraint value, and anomaly trend direction value to obtain the element detachment direction value. Specifically, this includes: binding the interface state quantification value, interface state confidence value, anomaly intensity value, anomaly boundary constraint value, and anomaly trend direction value to the same element to form a fusion feature carrier corresponding to the structural observation unit; performing interface constraint fusion processing on the interface state quantification value and interface state confidence value based on the fusion feature carrier to form an interface anomaly direction component; and based on the fusion... The feature carrier performs propagation constraint fusion processing on the anomaly intensity characterization value, anomaly boundary constraint value, and anomaly trend pointing value to form a propagation anomaly pointing component. Based on the interface anomaly pointing component and the propagation anomaly pointing component, a co-unit discrimination processing is performed to determine the dual-source unidirectional support condition, the single-source unverified condition, the conflict unverified condition, and the low-risk stability condition, and obtain the discrimination processing result. According to the discrimination processing result, a differential fusion processing is performed on the interface anomaly pointing component and the propagation anomaly pointing component to obtain the fusion result. Based on the fusion result, a neighborhood coupling correction processing and a trend consistency verification processing are performed sequentially to generate a unit vacancy pointing value.

[0108] Specifically, the interface state quantification value, interface state confidence value, anomaly intensity value, anomaly boundary constraint value, and anomaly trend indication value are first bound to the same unit to form a fusion feature carrier corresponding to the structural observation unit. In the specific processing, the structural observation unit identifier is used as a unique index to sequentially read the interface state quantification value, interface state confidence value, anomaly intensity value, anomaly boundary constraint value, and anomaly trend indication value corresponding to the current structural observation unit from the set of structural fusion basic units. Then, the spatial coordinate identifier, timestamp identifier, and spatial neighborhood association information of the structural observation unit are read simultaneously and written into the same data record to form the fusion feature carrier. Here, "binding to the same unit" refers to the unified association of multiple quantification results from different sources but all belonging to the same structural observation unit at the data level; the fusion feature carrier refers to the unified result object carrying all the fusion input features of the current structural observation unit. Since subsequent interface constraint fusion processing, propagation constraint fusion processing, and same-unit collaborative discrimination processing must all revolve around the same structural observation unit, this unified binding needs to be completed first to avoid fusion deviations caused by different features coming from adjacent units or different times. To quantify the completeness of binding within the same unit, a binding completeness can be constructed, expressed as:

[0109] in, It indicates the binding completeness of the current fusion feature carrier, and is used to characterize whether the fusion input features corresponding to the observation unit of this structure are complete; This indicates the number of valid feature items that have been successfully bound to the current fused feature carrier; This indicates the total number of feature items required for the current fusion step. The principle behind this process is to determine whether the current fusion feature carrier meets the conditions for entering subsequent fusion calculations by statistically analyzing the ratio between valid binding items and required binding items.

[0110] After forming the fusion feature carrier, interface constraint fusion processing is performed on the interface state quantification value and interface state confidence value based on the fusion feature carrier to form the interface anomaly pointing component. Specifically, the interface state quantification value is first used as the basic interface anomaly level of the current structural observation unit. Then, the interface state confidence value is introduced to modulate the confidence of this basic interface anomaly level, enhancing high-confidence interface anomaly representations and suppressing low-confidence interface anomaly representations. Subsequently, the interface state continuity in the spatial neighborhood association information is combined to perform neighborhood consistency correction on the interface anomaly level of the current structural observation unit, ensuring that structural observation units located in continuous interface anomaly regions maintain a strong anomaly pointing direction, while isolated interface anomaly units are appropriately suppressed. Here, interface constraint fusion processing refers to the process of constraining and shaping the anomaly degree of the current structural observation unit based on the interface state prediction results; the interface anomaly pointing component refers to the continuous component specifically used to characterize the current structural observation unit's pointing direction to the detachment of the surrounding rock lining interface in the interface state dimension. The interface state quantification value reflects the anomaly level corresponding to the interface state category itself, while the interface state confidence value reflects the reliability of the conclusion at that anomaly level. To form the interface anomaly pointing component, the following expression can be used:

[0111] in, This represents the interface anomaly pointing component, used to characterize the anomaly pointing intensity of the current structural observation unit in the interface state dimension; This represents the quantitative representation value of the interface state, used to characterize the level of interface anomalies. This represents the confidence level of the interface state, used to characterize the reliability of the conclusions drawn from the interface state. This represents the interface continuity support value extracted from spatial neighborhood association information, used to reflect whether there is anomalous support at compatible interfaces between adjacent structural observation units; This represents the neighborhood continuity support modulation coefficient, used to adjust the degree to which neighborhood support enhances the interface anomaly pointing component. The principle behind this processing is to incorporate the interface anomaly level, the reliability of the interface conclusion, and the continuity of the neighborhood interface into the same interface anomaly expression, making the interface anomaly pointing component more consistent with the actual distribution pattern of interface anomalies.

[0112] Based on the fusion feature carrier, propagation constraint fusion processing is performed on the anomaly intensity characterization value, anomaly boundary constraint value, and anomaly trend direction value to form a propagation anomaly direction component. Specifically, the anomaly intensity characterization value first reflects the dominant strength of the vibration propagation anomaly in the current structural observation unit; then, the anomaly boundary constraint value characterizes whether a continuous anomaly boundary is formed between the current structural observation unit and adjacent structural observation units; finally, the anomaly trend direction value characterizes whether the current anomaly is in a state of continuous enhancement, stable maintenance, or gradual attenuation. Subsequently, the three are combined at a unified scale, so that structural observation units possessing high anomaly intensity, strong boundary connection, and a continuous enhancement trend simultaneously obtain a higher propagation anomaly direction, while structural observation units with only a single anomaly intensity but lacking boundary support or trend support obtain a lower propagation anomaly direction. Here, propagation constraint fusion processing refers to the process of constructing a continuous propagation anomaly characterization based on the vibration propagation anomaly analysis results; the propagation anomaly direction component refers to the continuous component specifically used to characterize the current structural observation unit's direction towards the detachment of the surrounding rock lining interface in the vibration propagation dimension. The anomaly boundary constraint value reflects the continuity of the propagation anomaly at the spatial boundary, while the anomaly trend direction value reflects the development direction of the propagation anomaly over time. To generate a propagation anomaly pointing component, the following expression can be used:

[0113] in, This represents the propagation anomaly pointing component, used to characterize the anomaly pointing intensity of the current structural observation unit in the vibration propagation dimension; This represents the abnormal intensity characterization value, used to characterize the degree of deviation in vibration propagation; This represents the anomaly boundary constraint value, used to characterize the degree of connection with the neighboring anomaly boundaries; This value indicates an abnormal trend and is used to characterize an abnormal development trend. and These represent the modulation coefficients for boundary constraints and trend orientation, respectively. The principle behind this processing is that when the propagation anomaly is not only strong but also spatially continuous and temporally persistent, its orientation towards the voiding of the surrounding rock lining interface is stronger, and therefore it needs to be given higher weight in the fusion process.

[0114] After obtaining the interface anomaly pointing component and the propagation anomaly pointing component respectively, collaborative discrimination processing within the same unit is performed based on these components to determine the dual-source unidirectional support condition, single-source verification condition, conflict verification condition, and low-risk stability condition, thus obtaining the discrimination processing results. Specifically, the strength levels and directions of change of the interface anomaly pointing component and the propagation anomaly pointing component in the current structural observation unit are first compared. Then, it is determined whether both point to a high anomaly simultaneously, whether the main support is formed only from one side, whether there is a significant conflict in the anomaly direction, or whether both are at a low level. If both the interface anomaly pointing component and the propagation anomaly pointing component exceed the preset anomaly judgment condition, and their directions of change with the neighborhood and time remain consistent, the structural observation unit is judged as having a dual-source unidirectional support condition. If one reaches an anomaly level while the other does not, but no significant reverse conflict occurs, it is judged as having a single-source verification condition. If one shows a strong anomaly while the other shows stability or reverse change, it is judged as having a conflict verification condition. If both remain at a low level, it is judged as having a low-risk stability condition. The same-unit collaborative discrimination processing here refers to the process of judging the consistency and support relationship of two types of anomaly-oriented components within the same structural observation unit. The dual-source co-directional support condition indicates that interface anomalies and propagation anomalies jointly support the conclusion of vacancy removal; the single-source verification condition indicates that only one type of anomaly supports the conclusion of vacancy removal; the conflict verification condition indicates that there is a significant inconsistency between the two types of anomaly conclusions; the low-risk stability condition indicates that the overall anomaly degree of the current structural observation unit is low. To quantify the consistency degree of the two types of components, the following expression can be used:

[0115] in, It represents the degree of coordination and consistency between the interface anomaly pointing component and the propagation anomaly pointing component, and is used to characterize whether the two types of anomaly components exhibit a consistent support relationship in the same structural observation unit; This indicates an interface error pointing to a component; This indicates that the propagation anomaly points to a component; This indicates a stable term to prevent the denominator from being too small. The principle behind this process is that if the difference between the two types of anomalous components is small and both are relatively high, the degree of coordination is high, and it is more likely to enter a dual-source, same-direction support condition; if the difference between the two types of anomalous components is large, the degree of coordination is low, and it is more likely to enter a single-source verification condition or a conflict verification condition.

[0116] Based on the discrimination processing results, differentiated fusion processing is performed on the interface anomaly pointing component and the propagation anomaly pointing component to obtain the fusion result. Specifically, different discrimination processing results correspond to different fusion strategies. For dual-source co-directional support conditions, an enhanced fusion strategy should be adopted to allow the interface anomaly pointing component and the propagation anomaly pointing component to jointly improve the comprehensive anomaly result of the current structural observation unit. For single-source verification conditions, a conservative fusion strategy should be adopted to retain the main influence of high-value anomaly components, but the overall result is moderately suppressed to avoid the single-source anomaly being directly amplified into a highly reliable empty result. For conflict verification conditions, a conflict reduction fusion strategy should be adopted to perform mutual cancellation and weight reduction processing on the two components and generate conflict markers to provide a basis for verification in subsequent regional-level extended fusion processing. For low-risk stable conditions, a low-sensitivity preservation fusion strategy should be adopted to maintain a low fusion result for the current structural observation unit. Here, differentiated fusion processing refers to the process of using different fusion rules based on the collaborative discrimination conclusion; the fusion result refers to the intermediate comprehensive result of the current structural observation unit after completing the joint expression of interface anomalies and propagation anomalies. To represent enhanced fusion under dual-source co-directional support conditions, the following expression can be used:

[0117] in, This indicates the differentiated fusion results of the current structural observation unit; This indicates an interface error pointing to a component; This indicates that the propagation anomaly points to a component; and These represent the basic fusion weights for the two types of outlier components, respectively. This represents the collaborative enhancement weight, used to enhance the fusion result when both types of anomalous components are simultaneously high. The principle behind this is that when interface anomalies and propagation anomalies coexist, the collaborative term between them can significantly improve the null pointer orientation of the current unit; while under single-source verification conditions or conflict verification conditions, it can be improved by reducing... Alternatively, a suppressor term can be introduced to achieve conservative fusion or conflict reduction.

[0118] After obtaining the fusion results, neighborhood coupling correction and trend consistency verification are performed sequentially based on the fusion results to generate cell decoupling indices. Specifically, the neighborhood coupling correction is performed first, which involves reading the fusion results of adjacent structural observation units in the spatial neighborhood of the current structural observation unit and analyzing whether the current structural observation unit forms a continuous anomaly band with its neighbors, whether it is located at the edge of a high-value anomaly region, or whether it exhibits an isolated high-value cell. If the current structural observation unit and multiple adjacent units simultaneously show high fusion results with good boundary connections, the current fusion result is adjusted upwards; if the current structural observation unit is an isolated anomaly while the neighborhood is generally stable, the current fusion result is adjusted downwards. Subsequently, trend consistency verification is performed, which combines the historical fusion results and anomaly trend indices of the current structural observation unit under continuous timestamps to determine whether the current fusion result is consistent with the anomaly development direction. If the current fusion result is high and the anomaly trend continues to strengthen, the credibility of the result is further enhanced; if the current fusion result is high but the anomaly trend rapidly declines, a conservative correction is performed; if the current fusion result is moderate but the anomaly trend continues to strengthen, its potential for subsequent regional expansion analysis is preserved. The neighborhood coupling correction process here refers to the continuous correction of the current unit result using the results of spatially adjacent units; the trend consistency verification process refers to the dynamic rationality verification of the current unit result using the temporal evolution direction; the unit voiding direction value refers to the final comprehensive quantitative result of the possibility of the surrounding rock lining interface voiding in the current structural observation unit after simultaneously considering interface anomalies, propagation anomalies, spatial continuity, and temporal trends. To represent the result after neighborhood coupling correction, the following expression can be used:

[0119] in, This represents the fusion result after neighborhood coupling correction; This indicates the differentiated fusion results of the current structural observation unit; This represents the spatial neighborhood set of the current structural observation unit; This represents the differential fusion result of the j-th structural observation unit in the spatial neighborhood; N represents the number of structural observation units in the spatial neighborhood. This represents the neighborhood coupling correction coefficient, used to control the influence of the neighborhood average level on the current result. When combining it with time trends for verification, the following expression can be used:

[0120] in, This indicates the unit's delamination direction value, which is used to ultimately characterize the comprehensive probability of delamination at the surrounding rock lining interface in the current structural observation unit; This represents the fusion result after neighborhood coupling correction; Indicates the value to which the abnormal trend points; This represents the trend modulation coefficient, used to adjust the effect of the time trend on the cell vacancy pointing value. The principle behind this processing is that true vacancy anomalies usually exhibit both spatial continuity and temporal consistency. Therefore, only when the fusion result is supported in both the neighborhood and time dimensions will the final generated cell vacancy pointing value have higher reliability.

[0121] In one possible implementation, spatial threshold discrimination and spatial region mapping are performed based on the spatial void probability distribution to generate lining structure defect identification results. Specifically, this includes: performing probability hierarchy division based on the spatial void probability distribution to form probability levels; performing spatial threshold discrimination based on the probability levels to form spatial locations that distinguish between non-defect background locations, defect edge candidate locations, defect body candidate locations, and defect core locations; performing spatial connectivity analysis on the spatial locations to form defect candidate connected regions; performing region boundary extraction and region closure correction around the defect candidate connected regions to form candidate defect regions; performing region attribute discrimination on the candidate defect regions to filter out valid defect regions and regions to be reviewed; performing spatial region mapping on the valid defect regions to obtain lining structure defect regions; performing defect type merging and defect level determination on the lining structure defect regions; performing review identifier generation on the regions to be reviewed and result solidification on the valid defect regions; and converging the lining structure defect regions and the regions to be reviewed to form the lining structure defect identification results.

[0122] Specifically, a probability hierarchy is first formed based on the spatial voiding probability distribution. During this process, the spatial voiding probability value corresponding to each spatial location in the geometric model of the lining structure is read and segmented within a unified probability interval. This transforms the continuously changing spatial voiding probability distribution into a probability hierarchy with hierarchical semantics. Here, the probability hierarchy refers to the stratified result formed based on the magnitude of the spatial voiding probability value, used to characterize the strength of the likelihood of voiding at the surrounding rock lining interface at different spatial locations. In implementation, multiple levels, such as low-probability, low-to-medium-probability, medium-to-high-probability, and high-probability layers, can be determined by combining historical sample statistics, field calibration results, and the overall distribution characteristics of the current spatial voiding probability distribution. These layers correspond to different voiding risk levels. The purpose of this process is to transform subtle differences in the continuous probability field into a stratification basis that can be used for subsequent spatial threshold discrimination, thus avoiding subsequent processing that relies solely on a single probability value and lacks regional hierarchy. If a certain lining structure area has a high overall background probability, a local adaptive adjustment method can be adopted so that the probability hierarchy division reflects both the global risk level and retains local abrupt change characteristics.

[0123] After forming the probability hierarchy, spatial threshold discrimination processing is performed based on the probability hierarchy to form spatial locations that distinguish between non-defect background locations, defect edge candidate locations, defect body candidate locations, and defect core locations. This spatial threshold discrimination processing refers to classifying each spatial location based on the spatial void probability distribution and its probability hierarchy. Non-defect background locations are those with low void probability values ​​and lacking neighborhood anomaly support; defect edge candidate locations are transitional locations between high and low probability areas; defect body candidate locations are those that have formed relatively stable high-probability anomaly support; and defect core locations are those located at the local highest probability cluster center and have strong neighborhood continuity. In practice, it is necessary not only to determine the probability hierarchy of the current spatial location itself but also to combine the probability hierarchy distribution of adjacent spatial locations, the neighboring support of the effective void area, and the consistency of anomaly trends for joint judgment. The reason for this processing is that real lining structure defects are usually not a single high value but have a spatial distribution structure with a gradual transition from center to body to edge. Therefore, spatial threshold discrimination is needed to transform the spatial probability distribution into a hierarchical location type more suitable for region identification.

[0124] After spatial threshold discrimination, spatial connectivity analysis is performed on the resulting spatial locations to form candidate connected regions for defects. This spatial connectivity analysis involves analyzing the spatial locations identified as defect edge candidates, defect body candidates, and defect core locations based on their spatial adjacency, boundary contact, and extension direction relationships within the lining structure's geometric model. This process identifies continuous anomalous regions composed of multiple adjacent spatial locations. A candidate connected region for defects refers to a set of spatial locations that are spatially connected, probabilistically progressive, and geometrically constitute a local anomalous region. In practice, the defect core location is typically used as the center, and the search extends outwards to find directly or indirectly adjacent candidate defect body locations and defect edge candidates. The spatial distance, geometric connection method, and hierarchical transition relationship between these locations must meet preset connectivity conditions. Essentially, this process further organizes the discrete spatial location discrimination results into region-oriented candidate anomalous objects, providing a foundation for subsequent region boundary extraction and region attribute analysis. If certain high-probability spatial locations are too far apart, lack intermediate level transitions, or do not have reasonable geometric connections, they should not be classified into the same defective candidate connected region, but should be regarded as multiple independent candidate regions.

[0125] After forming candidate connected regions of defects, region boundary extraction and region closure correction are performed around these regions to form candidate defect regions. Region boundary extraction refers to extracting the outer boundary of the abnormal region along the probability gradient change interface between the candidate connected region of defects and the adjacent non-defect background region. Region closure correction refers to repairing and correcting any broken boundaries, jagged boundaries, local gaps, and internal voids that may appear after boundary extraction. A candidate defect region is a candidate region object that already has a clear boundary outline and can independently describe the spatial anomaly range. During implementation, an initial region boundary can be constructed first based on the boundary adjacency relationship between spatial locations. Then, the continuity of the boundary is checked. If gaps are found in the boundary due to local sampling sparsity, interpolation attenuation, or probability fluctuations, boundary closure compensation is performed. If there are local low-value voids inside the candidate region, but they are all supported by spatial high values, internal void filling is performed. If the boundary has obvious outward expansion but lacks neighboring high-value support, boundary contraction correction is performed. The candidate defect regions formed through this process have more complete boundaries, more continuous spatial morphology, and are closer to the distribution of defects in the actual lining structure in geometric space.

[0126] After obtaining candidate defect regions, regional attribute discrimination processing is performed on these regions to filter out valid defect regions and regions awaiting verification. This regional attribute discrimination processing refers to the analysis of the authenticity and stability of candidate defect regions at the overall regional level. Regional attributes can include regional probability intensity, regional area size, regional extension morphology, regional boundary smoothness, internal consistency, and regional temporal stability. Valid defect regions are those deemed to have high authenticity and strong continuity support after discrimination; regions awaiting verification are those that, while showing some abnormal signs, do not yet meet the complete defect determination criteria and require further observation or verification. In practice, the process first checks whether a high proportion of candidate defect locations and defect core locations exist within the candidate defect region. Then, it checks whether the spatial area of ​​the region reaches the minimum coverage range sufficient to characterize an actual defect. Next, it analyzes whether the region's boundaries are too fragmented, too sharp, or lack a reasonable extension direction. Finally, it checks whether the region can maintain a basically consistent spatial existence under multiple consecutive timestamps. If a candidate defect region simultaneously meets multiple of the above region attribute conditions, it is determined to be a valid defect region; if it only meets some conditions, or if there are obvious conflicting attributes, it is determined to be a region to be reviewed. This process can effectively reduce the probability of noisy regions, interpolation artifact regions, or transient anomaly regions being misjudged as real defects.

[0127] For the selected valid defect areas, spatial region mapping is performed to obtain the lining structure defect areas. This spatial region mapping refers to mapping the valid defect areas from the abstract probability space and analysis space back to the specific engineering space of the lining structure, so that the area corresponds to a specific location, specific structural segment, and specific interface range in the actual lining structure. The lining structure defect area refers to the defect spatial area object that can be directly identified and located by engineers. During implementation, based on the predefined structural partitions, axial mileage, circumferential position, surface development position, and interface hierarchy relationships in the lining structure geometric model, the valid defect area can be mapped to a specific lining structure segment, specific lining structure surface, or specific surrounding rock lining interface zone. If a valid defect area spans multiple structural partitions, its primary and secondary affected areas must also be determined based on the area's high-value center, main extension direction, and area distribution. Through this processing, the valid defect areas, which originally only had analytical significance, are transformed into lining structure defect areas with engineering guidance significance, thus facilitating subsequent result output, graphical annotation, and maintenance decision-making.

[0128] After identifying the defective areas in the lining structure, defect type merging and defect level determination are performed. Defect type merging involves classifying the defective areas based on the distribution of interface state categories, vibration propagation anomaly levels, spatial voiding probability levels, and regional extension characteristics. Defect level determination further identifies the risk level based on the area's risk intensity and evolution. Defect types can be categorized as early interface voiding defects, extended interface voiding defects, and high-risk interface coupling defects. Defect levels can be classified as low-level, medium-level, high-level, and emergency-level defects. In practice, if a lining structure defective area is primarily supported by a micro-void state at the interface, and the vibration propagation anomaly intensity is moderate with a medium-to-high spatial voiding probability, it can be classified as an early interface voiding defect. If the area simultaneously exhibits continuous high probability, large-area extension, and a clear abnormal development trend, it can be classified as an extended interface voiding defect or a higher-risk type. When determining the risk level, factors such as the area size, spatial core proportion, temporal duration, and abnormal growth rate can be considered to further categorize the risk level. Through this dual approach of type and level, defective areas in the lining structure are not only identified but also assigned clear semantic and risk attributes.

[0129] For areas awaiting review, a review identifier generation process is performed; for valid defect areas, a result solidification process is performed. The review identifier generation process involves generating a specific review status marker, a review reason explanation, and subsequent processing suggestions for the areas awaiting review, ensuring that these areas are retained in the final result as pending confirmation, rather than being ignored. Review reasons may include insufficient area, unstable boundaries, insufficient temporal continuity, or conflicts between interface states and propagation anomalies. Result solidification involves writing the final boundary information, spatial mapping information, defect type information, defect level information, center location, extension direction, and corresponding temporal status of valid defect areas into a unified result carrier, ensuring consistency during subsequent retrieval, display, and archiving. This process guarantees that the system output includes not only confirmed high-confidence defect results but also results for abnormal areas requiring focused attention and further verification, thus avoiding information omissions. The simultaneous existence of review identifiers and result solidification ensures that the entire defect identification process possesses both deterministic output and the ability to manage uncertain anomalies.

[0130] Finally, the defective areas and the areas to be reviewed in the lining structure are aggregated to form the lining structure defect identification results. This aggregation refers to integrating the confirmation results corresponding to all valid defective areas and the review results corresponding to all areas to be reviewed into a single defect identification result set, under a unified structural observation unit mapping relationship, a unified spatial coordinate benchmark, and a unified timestamp benchmark. The lining structure defect identification results should at least include the defect area boundary, defect center location, defect type, defect level, time status, area mapping relationship, and review identifier. In implementation, the results can be organized according to the spatial location order of the lining structure, risk level order, or time order, making it convenient for subsequent graphical display and subsequent maintenance operations. The resulting lining structure defect identification results have completed the full transformation from spatial void probability distribution to engineering defect objects, and can directly serve structural condition assessment, defect location management, and maintenance decision support.

[0131] This embodiment also discloses an artificial intelligence-based building structure defect identification device, referring to... Figure 2 The device includes an acquisition module 201, a processing module 202, and an output module 203. It is used to execute any of the above-described artificial intelligence-based methods for identifying defects in the main structure of a building, wherein: The acquisition module 201 is used to acquire the structural observation image sequence, structural vibration response sequence and structural acoustic wave propagation sequence of the surface of the building lining structure, and perform preprocessing to form a set of structural observation units; Processing module 202 is used to perform feature extraction processing on the set of structural observation units and construct a multimodal feature representation of the structure through spatial coordinate alignment, thereby forming a set of structural feature representations; The processing module 202 is used to construct a lining structure interface state discrimination model based on the set of structural feature expressions, and to perform model training processing using historical samples. At the same time, it performs interface state inference processing on the set of structural feature expressions to obtain a set of predicted results of the structural interface state. The processing module 202 is used to perform vibration propagation anomaly analysis processing on the structural vibration response signal based on the prediction result set, so as to identify potential void areas at the interface of the surrounding rock lining and obtain the anomaly analysis results of vibration propagation. Processing module 202 is used to perform spatial feature fusion processing based on the set of anomaly analysis results and prediction results, thereby constructing the spatial void probability distribution of the lining structure; The output module 203 is used to perform spatial threshold discrimination and spatial region mapping processing based on the spatial void probability distribution to generate lining structure defect identification results.

[0132] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0133] This embodiment also discloses an electronic device, as shown in the reference. Figure 3 The electronic device may include: at least one processor 301, at least one communication bus 302, user interface 303, network interface 304, and at least one memory 305.

[0134] The communication bus 302 is used to enable communication between these components.

[0135] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0136] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0137] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 305, and by calling data stored in memory 305. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0138] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory may include a non-transitory computer-readable storage medium. The memory 305 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 305 may also be at least one storage device located remotely from the aforementioned processor 301. As a computer storage medium, the memory 305 may include an operating system, a network communication module, a user interface 303 module, and an application program for an artificial intelligence-based method for identifying defects in building structures.

[0139] exist Figure 3In the electronic device shown, the user interface 303 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 301 can be used to call an application program stored in the memory 305 that is an artificial intelligence-based method for identifying defects in the main structure of a building. When executed by one or more processors 301, the electronic device executes one or more methods as described in the above embodiments.

[0140] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps can be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0141] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0142] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0143] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0144] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0145] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 305 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned memory 305 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0146] The present invention also discloses a non-transitory computer-readable storage medium storing instructions. When executed by one or more processors 301, these instructions cause an electronic device to perform one or more methods as described in the above embodiments.

[0147] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and the disclosure of practical truths. This invention is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for identifying defects in the main structure of a building based on artificial intelligence, characterized in that, The method includes: Acquire the structural observation image sequence, structural vibration response sequence, and structural acoustic wave propagation sequence of the monitored building lining surface, and perform preprocessing to form a set of structural observation units; Feature extraction processing is performed on the set of structural observation units, and a multimodal structural feature representation is constructed by spatial coordinate alignment, thereby forming a set of structural feature representations; Based on the set of structural feature expressions, a lining structure interface state discrimination model is constructed, and historical samples are used to perform model training processing. At the same time, interface state inference processing is performed on the set of structural feature expressions to obtain a set of predicted results for the structural interface state. Based on the predicted result set, vibration propagation anomaly analysis processing is performed on the structural vibration response signal to identify potential void areas at the surrounding rock lining interface and obtain the vibration propagation anomaly analysis results. Based on the anomaly analysis results and the prediction results set, spatial feature fusion processing is performed to construct the spatial void probability distribution of the lining structure; Based on the spatial void probability distribution, spatial threshold discrimination and spatial region mapping processes are performed to generate lining structure defect identification results.

2. The method for identifying defects in the main structure of a building based on artificial intelligence according to claim 1, characterized in that, The process of performing feature extraction on the set of structural observation units and constructing a multimodal structural feature representation through spatial coordinate alignment, thereby forming a set of structural feature representations, specifically includes: Spatial coordinate identification binding and timestamp identification registration are performed on the structural observation images, structural vibration response signals and structural acoustic wave propagation signals in the set of structural observation units to establish a structural observation reference system; Image enhancement and region slicing processing are performed on the observed images of the structure, and image texture feature vectors representing the surface state of the lining structure are extracted; The vibration response signal of the structure is subjected to signal purification processing and multi-domain feature extraction processing to obtain a vibration response feature vector characterizing the dynamic response of the lining structure. The propagation waveform correction process and propagation feature analysis process are performed on the acoustic wave propagation signal of the structure to obtain the acoustic wave propagation feature vector characterizing the propagation state of the surrounding rock lining interface; Based on the structural observation reference system, spatial alignment and consistency verification are performed on the image texture feature vector, the vibration response feature vector, and the sound wave propagation feature vector to form candidate multimodal feature vectors; The candidate multimodal feature vectors are subjected to correlation compression and discriminative enhancement processing to generate the structural multimodal feature representation. The structural multimodal feature representations are aggregated according to the correspondence of structural observation units to form the structural feature representation set.

3. The method for identifying defects in the main structure of a building based on artificial intelligence according to claim 1, characterized in that, The process involves constructing a lining structure interface state discrimination model based on the structural feature expression set, training the model using historical samples, and simultaneously performing interface state inference processing on the structural feature expression set to obtain a set of predicted structural interface states. Specifically, this includes: Construct a historical sample set with interface state labels based on the aforementioned structural feature expression set; The historical sample set is subjected to sample balancing and sample subset partitioning to obtain training sample subsets, validation sample subsets and test sample subsets. Based on the training sample subset, a lining structure interface state discrimination model is constructed, which includes an input layer, a feature association layer, a state discrimination layer, and a result output layer. In the feature association layer, the association mapping relationship between multimodal features is established. The training sample subset is used to perform parameter training on the interface state discrimination model of the lining structure, and the model parameters are iteratively corrected based on the state discrimination bias results. The validation sample subset is used to perform validation evaluation processing on the interface state discrimination model of the lining structure, and the feature association layer and the state discrimination layer are coordinated and adjusted according to the validation evaluation results to determine the target model parameters. The generalization test process of the interface state discrimination model of the lining structure is performed using the subset of test samples, and the model deployment solidification process is performed when the preset conditions are met. The set of structural feature representations is input into the interface state discrimination model of the lining structure to perform interface state inference processing, and the set of prediction results corresponding to the structural observation unit is output.

4. The method for identifying defects in the main structure of a building based on artificial intelligence according to claim 1, characterized in that, The vibration propagation anomaly analysis processing of the structural vibration response signal based on the predicted result set is performed to identify potential void areas at the surrounding rock lining interface, and the anomaly analysis results of vibration propagation are obtained, specifically including: Based on the predicted result set, a mapping relationship is established between the structural observation unit, the structural vibration response signal, and the structural interface state. A reference construction process is performed on the structural vibration response signal corresponding to the interface stability reference unit in the structural observation unit to form a reference vibration reference. Based on the mapping relationship, vibration propagation pre-analysis processing is performed on the structural vibration response signal to obtain vibration response segments with a unified propagation starting point and propagation time window; The vibration response segment is subjected to vibration propagation feature reconstruction processing to form a vibration propagation characterization sequence corresponding to the structural observation unit; Based on the reference vibration benchmark and the vibration propagation characterization sequence, a vibration propagation deviation analysis is performed to obtain the vibration propagation deviation result. The vibration propagation deviation result and the prediction result set are subjected to joint constraint processing to form a vibration propagation anomaly determination result corresponding to the interface state category; Anomaly seat allocation processing is performed on the vibration propagation deviation results to determine the candidate units of propagation anomalies and the propagation anomaly clustering areas; Propagation path backtracking and interface void pointing analysis are performed around the aforementioned propagation anomaly cluster area to identify potential void areas at the surrounding rock lining interface. Anomaly analysis results are generated by performing anomaly analysis on the potential void area at the surrounding rock lining interface.

5. The method for identifying defects in the main structure of a building based on artificial intelligence according to claim 1, characterized in that, The step of performing spatial feature fusion processing based on the anomaly analysis results and the prediction results set to construct the spatial void probability distribution of the lining structure specifically includes: Based on the anomaly analysis results and the prediction result set, perform same-unit merging processing to form a structural fusion basic unit set; Spatial neighborhood organization processing is performed on the set of basic structural fusion units to construct spatial neighborhood association information corresponding to the structural observation units; Based on the set of structural fusion basic units and the spatial neighborhood association information, perform interface state quantization mapping processing to obtain interface state quantization representation value and interface state confidence representation value. Based on the structural fusion basic unit set and the spatial neighborhood association information, vibration propagation anomaly quantization mapping processing is performed to obtain anomaly intensity characterization value, anomaly boundary constraint value and anomaly trend direction value. The interface state quantization characterization value, the interface state confidence characterization value, the anomaly intensity characterization value, the anomaly boundary constraint value, and the anomaly trend pointing value are fused to obtain the cell vacancy pointing value. Based on the unit's empty pointing value, perform a region-level extended fusion process to form candidate empty regions and filter out effective empty regions; Spatial interpolation and fusion processing is performed around the effective void region and the structural observation unit to construct a continuous spatial void characterization field; A probabilistic mapping process is performed on the spatial void characterization field to generate the spatial void probability distribution of the lining structure.

6. The method for identifying defects in the main structure of a building based on artificial intelligence according to claim 5, characterized in that, The process of fusing the interface state quantization value, the interface state confidence value, the anomaly intensity value, the anomaly boundary constraint value, and the anomaly trend indication value to obtain the cell vacancy indication value specifically includes: The interface state quantification value, the interface state confidence value, the anomaly intensity value, the anomaly boundary constraint value, and the anomaly trend direction value are bound to the same unit to form a fusion feature carrier corresponding to the structural observation unit. Based on the fusion feature carrier, the interface state quantization representation value and the interface state confidence representation value are subjected to interface constraint fusion processing to form an interface anomaly pointing component. Based on the fusion feature carrier, the anomaly intensity characterization value, the anomaly boundary constraint value, and the anomaly trend direction value are subjected to propagation constraint fusion processing to form a propagation anomaly direction component; Based on the interface anomaly pointing component and the propagation anomaly pointing component, perform co-unit collaborative discrimination processing to determine the dual-source unidirectional support condition, the single-source unverified condition, the conflict unverified condition, and the low-risk stable condition, and obtain the discrimination processing result. Based on the discrimination processing result, a differential fusion processing is performed on the interface anomaly pointing component and the propagation anomaly pointing component to obtain a fusion result; Based on the fusion result, neighborhood coupling correction processing and trend consistency verification processing are performed sequentially to generate the cell null pointer value.

7. The method for identifying defects in the main structure of a building based on artificial intelligence according to claim 1, characterized in that, The step of performing spatial threshold discrimination and spatial region mapping processing based on the spatial void probability distribution to generate lining structure defect identification results specifically includes: Based on the spatial voiding probability distribution, a probability hierarchy division process is performed to form a probability hierarchy; Based on the probability hierarchy, spatial threshold discrimination processing is performed to form spatial locations that distinguish between non-defect background locations, defect edge candidate locations, defect body candidate locations, and defect core locations. Spatial connectivity analysis is performed on the spatial location to form candidate connected regions for defects; The candidate defect connectivity region is subjected to region boundary extraction processing and region closure correction processing to form a candidate defect region; Perform region attribute discrimination processing on the candidate defect regions to filter out valid defect regions and regions to be reviewed; Perform spatial region mapping processing on the effective defect region to obtain the lining structure defect region; The defective areas of the lining structure are subjected to defect type merging and defect level determination processing. The area to be reviewed is processed to generate a review identifier, and the effective defect area is processed to solidify the result. The defective area of ​​the lining structure is combined with the area to be reviewed to form the defect identification result of the lining structure.

8. A building structure defect identification device based on artificial intelligence, characterized in that, The device is used to execute the artificial intelligence-based method for identifying defects in the main structure of a building as described in any one of claims 1-7. The device includes an acquisition module, a processing module, and an output module, wherein: The acquisition module is used to acquire the structural observation image sequence, structural vibration response sequence, and structural acoustic wave propagation sequence of the monitored building lining structure surface, and perform preprocessing to form a set of structural observation units; The processing module is used to perform feature extraction processing on the set of structural observation units and construct a multimodal structural feature representation by aligning spatial coordinates, thereby forming a set of structural feature representations. The processing module is used to construct a lining structure interface state discrimination model based on the structural feature expression set, and to perform model training processing using historical samples. At the same time, it performs interface state inference processing on the structural feature expression set to obtain a set of predicted results of the structural interface state. The processing module is used to perform vibration propagation anomaly analysis processing on the structural vibration response signal based on the prediction result set, so as to identify potential void areas at the surrounding rock lining interface and obtain the vibration propagation anomaly analysis results. The processing module is used to perform spatial feature fusion processing based on the anomaly analysis results and the prediction result set, thereby constructing the spatial void probability distribution of the lining structure. The output module is used to perform spatial threshold discrimination and spatial region mapping processing based on the spatial void probability distribution to generate lining structure defect identification results.

9. An electronic device, characterized in that, The device includes a processor, a communication bus, a user interface, a network interface, and a memory. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The communication bus is used to enable communication between the components within the electronic device. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images