Tunnel structure deterioration prediction method and system based on multi-source data fusion perception

Abstract

The invention provides a tunnel structure deterioration prediction method and system based on multi-source data fusion perception, and relates to the technical field of computers, and the method comprises the steps: obtaining original test data generated by a cold region heavy haul railway tunnel structure deterioration test system; feature extraction is carried out according to the original test data, and feature vector sets representing the dynamic state, the static force bearing state and the apparent damage state of the structure are obtained respectively; performing feature fusion according to the feature vector set and the environmental parameter time sequence data to obtain advanced feature representation; performing feature screening and dimension reduction according to the advanced feature representation to obtain an optimal feature subset after dimension reduction; performing model mapping according to the optimal feature subset to construct a deterioration prediction model; and inputting on-site original monitoring data into the deterioration prediction model to obtain a prediction result. According to the method, the depth of state perception, the reliability of evaluation and early warning and the scientificity of maintenance decision are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method and system for predicting tunnel structure deterioration based on multi-source data fusion sensing. Background Technology

[0002] As my country's railway network continues to extend into complex environments such as high-altitude and frigid regions, the long-term service safety of heavy-haul railway tunnels in cold climates faces severe challenges. Under the coupled effects of continuous freeze-thaw cycles and the dynamic loads of heavy-haul trains, tunnel lining structures are prone to various latent and overt damages. The evolution mechanisms are complex, and traditional monitoring methods, primarily relying on manual inspections and single sensors such as crack gauges and strain gauges, suffer from low efficiency, high risk, and incomplete coverage in harsh environments, making it difficult to comprehensively capture complete information on structural damage. While existing automated monitoring technologies can acquire long-term time-series data for specific cross-sections, these data often present a point-like, discrete spatial distribution, resulting in spatiotemporal discontinuity and an inability to construct a data field reflecting the continuous evolution of the tunnel's structural state. This restricts the overall assessment of the spatial distribution and evolution patterns of damage. At the data analysis level, existing methods largely rely on empirical formulas and shallow statistical analysis, making it difficult to effectively process massive, heterogeneous, multi-source monitoring data. They cannot automatically uncover deep damage evolution patterns and sensitive characteristics driven by the coupling of multiple factors such as freeze-thaw cycles and loads, resulting in insufficient accuracy and reliability in condition assessment. More importantly, current technological systems primarily focus on diagnosing historical and current conditions, lacking effective models capable of scientifically predicting future structural deterioration trends and safety status. This results in severely delayed maintenance decisions, leading to a passive, reactive approach. Therefore, overcoming technological bottlenecks such as limited sensing methods, data silos, superficial analysis, and a lack of predictive capabilities, and achieving a shift from discrete point inspection and passive response to continuous sensing, intelligent assessment, and proactive early warning, has become a core technological issue urgently needing to be addressed in the field of safety assurance for high-altitude, heavy-haul railway tunnels in cold regions.

[0003] Based on the shortcomings of the existing technologies, there is an urgent need for a method and system for predicting tunnel structure damage based on multi-source data fusion sensing. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting tunnel structure deterioration based on multi-source data fusion sensing, so as to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a method for predicting tunnel structural defects based on multi-source data fusion sensing, including: The raw test data generated by the cold-region heavy-haul railway tunnel structure deterioration test system is obtained. The raw test data includes vibration response time series data, internal force time series data, appearance image data and environmental parameter time series data of the lining specimen under the coupled action of freeze-thaw cycle and heavy load. Based on the original experimental data, feature extraction was performed to obtain feature vector sets representing the dynamic state, static bearing state, and apparent damage state of the structure, respectively. Feature fusion is performed based on the feature vector set and the environmental parameter time series data. By combining the environmental state, weights are adaptively assigned to features from different sources, and the weighted features are combined and then subjected to nonlinear interaction to obtain a high-level feature representation. Feature filtering and dimensionality reduction are performed based on the high-level feature representation. The optimal feature subset after dimensionality reduction is obtained by calculating the feature importance and arranging them in descending order of importance. Model mapping is performed based on the optimal feature subset, and a nonlinear regression function is trained based on the optimal feature subset and the known deterioration state labels to construct a deterioration prediction model; The original on-site monitoring data is input into the defect prediction model to obtain the prediction results of the future safety status of the tunnel structure.

[0005] Secondly, this application also provides a tunnel structure deterioration prediction system based on multi-source data fusion sensing, comprising: The acquisition module is used to acquire the raw test data generated by the cold-region heavy-load railway tunnel structure deterioration test system. The raw test data includes the vibration response time series data, internal force time series data, appearance image data and environmental parameter time series data of the lining specimen under the coupled action of freeze-thaw cycle and heavy load. The extraction module is used to extract features based on the original test data to obtain feature vector sets representing the dynamic state, static bearing state and apparent damage state of the structure, respectively. The fusion module is used to perform feature fusion based on the feature vector set and the environmental parameter time series data. By combining the environmental state, it adaptively assigns weights to features from different sources and performs nonlinear interaction on the weighted features to obtain a high-level feature representation. The filtering module is used to perform feature filtering and dimensionality reduction based on the high-level feature representation. By calculating the feature importance and sorting them in descending order of importance, the optimal feature subset after dimensionality reduction is obtained. The module is used to perform model mapping based on the optimal feature subset, and to construct a damage prediction model by training a nonlinear regression function based on the optimal feature subset and the known damage state labels. The prediction module is used to input the original on-site monitoring data into the deterioration prediction model to obtain the prediction results of the future safety status of the tunnel structure.

[0006] The beneficial effects of this invention are as follows: This invention integrates multi-source heterogeneous data and performs intelligent fusion and feature mining to construct a predictive model that can output a continuous deterioration index. This enables an objective quantitative assessment of the health status of tunnel structures and proactive early warning of future deterioration trends, thereby significantly improving the depth of condition perception, the reliability of assessment and early warning, and the scientific nature of maintenance decisions. Attached Figure Description

[0007] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart illustrating a method for predicting tunnel structure deterioration based on multi-source data fusion sensing, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of a tunnel structure deterioration prediction system based on multi-source data fusion sensing, as described in an embodiment of the present invention.

[0009] The diagram is labeled as follows: 901, Acquisition Module; 902, Extraction Module; 903, Fusion Module; 904, Filtering Module; 905, Construction Module; 906, Prediction Module. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0011] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0012] With the continuous expansion of railway networks in high-altitude and cold regions, the lining structures of heavy-haul railway tunnels are subjected to multiple coupled effects of extreme freeze-thaw cycles, repeated impacts from heavy-load trains, and complex ground stresses. This has led to increasingly prominent problems such as crack development, stiffness degradation, backfill voids, and concrete spalling, seriously threatening long-term operational safety. To address this challenge, the field has long relied on manual periodic inspections and localized fixed-point monitoring using single-type sensors such as crack gauges and strain gauges at typical cross-sections. However, in harsh, cold environments, manual inspections have inherent limitations such as low efficiency and high risk, and are difficult to detect intangible damage such as internal voids and material softening. Existing automated monitoring methods, on the other hand, perceive isolated information with discrete point coverage, failing to simultaneously acquire multi-dimensional information reflecting the overall structural condition, such as full-section vibration response and surface images. This makes it difficult to achieve a paradigm shift from local, discrete, single-point monitoring to full-line, continuous, and integrated intelligent assessment. Although current practices have attempted to incorporate technologies such as vibration detection and image recognition, existing methods are generally limited by insufficient multi-source data perception and fusion capabilities. They often rely on a single data source for analysis, lacking the ability to simultaneously and correlatedly perceive and fuse apparent damage and internal latent damage, resulting in biased and unreliable assessment results. Furthermore, these methods primarily depend on long-term time-series data from fixed cross-sections, making it difficult to construct a complete state field covering the entire tunnel's spatial continuity and correlated with time evolution. This limits the in-depth characterization of damage spatial distribution and the gradual change mechanism under the coupling effects of multiple factors. In addition, traditional assessments are mostly based on empirical formulas or shallow statistical analysis, making it difficult to automatically extract deterioration-sensitive features from massive amounts of heterogeneous multi-source data. Moreover, they cannot establish intelligent prediction models that can accurately characterize damage development patterns and output quantitative early warning indicators, resulting in a lack of foresight in maintenance decisions and a reactive approach. To address these problems, this invention provides a method and system for predicting tunnel structural deterioration based on multi-source data fusion perception. This invention aims to overcome the limitations of traditional monitoring models in terms of data dimensions, fusion methods, and intelligence. It constructs a full-cycle data acquisition and verification system combining a laboratory multi-field coupled experimental system with a field-based fixed multi-source monitoring system, forming a multi-dimensional heterogeneous database that integrates spatiotemporal evolution and surface-level characteristics. Furthermore, it utilizes deep learning and machine learning algorithms to achieve adaptive fusion of multi-source features and mining of key damage information, establishing an intelligent mapping model and time-series prediction model from multi-source data to continuous deterioration indices and safety levels. Ultimately, it achieves a shift from relying on localized inspections and post-incident handling to intelligent assessment, prediction, and early warning based on multi-source fixed monitoring, providing a complete and reliable technical solution for the accurate assessment, scientific maintenance, and risk management of the structural health status of high-altitude, heavy-load railway tunnels.

[0013] Example 1: This embodiment provides a method for predicting tunnel structural defects based on multi-source data fusion sensing.

[0014] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0015] Step S100: Obtain the original test data generated by the cold region heavy-load railway tunnel structure deterioration test system. The original test data includes the vibration response time series data, internal force time series data, appearance image data and environmental parameter time series data of the lining specimen under the coupled action of freeze-thaw cycle and heavy load. Understandably, this step is implemented using a specially constructed test system for the deterioration of heavy-haul railway tunnel structures in cold regions. In this controlled physical model test environment, the lining specimen is placed in a freeze-thaw cycle device capable of accurately simulating the periodic changes in extreme temperatures, while simultaneously subjected to repetitive mechanical and vibration loads simulating the passage of heavy-haul trains. Accelerated deterioration tests are conducted on the pre-determined damaged lining under controlled conditions. Under this coupled effect, multiple types of raw test data are simultaneously acquired: a high-sensitivity accelerometer captures the dynamic vibration response sequence of the lining during simulated train passage; strain gauges and other sensors densely distributed on the specimen surface record the evolution of internal stress and strain; a high-resolution imaging unit acquires high-definition images of the specimen surface under controlled illumination to capture visual evidence of crack formation and propagation; and simultaneously, temperature and humidity sensors continuously record the time sequence of environmental parameters within the test chamber to quantify the number of freeze-thaw cycles, the duration of negative temperatures, and the temperature gradient.

[0016] Step S200: Based on the original experimental data, feature extraction is performed to obtain feature vector sets representing the dynamic state, static bearing state, and apparent damage state of the structure, respectively. It should be noted that this step employs differentiated processing approaches for data sources with different physical properties: for time-series signals such as vibration responses, frequency domain and statistical features reflecting the structural dynamic characteristics and impact load effects are extracted; for internal force data, deep patterns reflecting changes in the structure's static bearing capacity are mined from its waveforms; and for apparent images, visual information such as cracks is transformed into precise geometric and morphological parameters. This process refines the raw monitoring data into a set of feature vectors that can characterize specific state attributes of the structure from different perspectives.

[0017] Step S300: Perform feature fusion based on the feature vector set and environmental parameter time series data. Adaptively assign weights to features from different sources by combining environmental states, and perform nonlinear interaction after combining the weighted features to obtain high-level feature representation. Understandably, this step is not simply a matter of concatenating various feature vectors, but rather introduces an intelligent fusion mechanism. This mechanism specifically considers the characteristics of cold-region scenarios, incorporating quantified environmental parameters (such as freeze-thaw stages) as important contextual information into the fusion decision. Through adaptive weight allocation, the model can dynamically adjust its focus on features from different data sources; for example, during freeze-thaw stages, it may place greater emphasis on features related to material stiffness softening or moisture content. Subsequently, the weighted features interact and combine in a nonlinear network, aiming to uncover deep and complex correlations between different modal data such as vibration, internal forces, and images, and their relationship to the overall structural degradation state, thereby generating a more informative high-level feature representation.

[0018] Step S400: Perform feature filtering and dimensionality reduction based on the high-level feature representation. Calculate the feature importance and sort them in descending order of importance to obtain the optimal feature subset after dimensionality reduction. It should be noted that this step calculates the contribution of each feature to the final determination of structural damage status and then filters and ranks the features based on their contribution. The aim is to automatically identify and retain the core features most relevant to the evolution of structural damage, forming a smaller but more information-pure optimal feature subset. This not only improves the computational efficiency of subsequent models but also effectively prevents overfitting and enhances the model's generalization ability.

[0019] Step S500: Perform model mapping based on the optimal feature subset. Train a nonlinear regression function based on the optimal feature subset and the known defective state labels to construct a defective prediction model. Understandably, this step uses the optimal feature subset obtained in step S400 as input and the known deterioration states from the experiment as supervision labels to train a nonlinear regression function. The learning objective of this function is to establish a complex mapping relationship from a multidimensional feature space to a continuous real value. Through training, the final deterioration prediction model possesses a core capability: when a new set of feature data is input, it can output a continuous deterioration index between 0 and 1. This index is a comprehensive quantitative indicator, and its value directly reflects the overall severity of deterioration of the lining structure.

[0020] Step S600: Input the original on-site monitoring data into the defect prediction model to obtain the prediction results of the future safety status of the tunnel structure.

[0021] It should be noted that in this step, a fixed monitoring system deployed inside the actual tunnel continuously collects raw monitoring data from the site. This data undergoes feature extraction and preprocessing consistent with the training phase, forming real-time feature vectors that conform to the model input format. These vectors are then input into the pre-trained deterioration prediction model. The model first outputs a deterioration index characterizing the current structural health status. Further, based on this index and its historical sequences, a time-series prediction model is used to analyze its changing trends, thereby inferring the evolution direction and potential risk level of the structural safety status over a future period. Finally, it outputs forward-looking prediction results, providing direct evidence for preventative maintenance decisions.

[0022] Further, step S200 includes steps S210 to S230.

[0023] Step S210: Perform dynamic feature extraction processing based on vibration response time series data. Obtain signal components of different frequency bands by performing wavelet packet decomposition on vibration response time series data, and calculate the energy of each frequency band as well as the amplitude factor and peak factor of the signal to obtain the feature vector characterizing the impact characteristics of the structure. Step S220: Based on the internal force time series data, the bearing feature extraction process is performed. By constructing a one-dimensional convolutional neural network, local and deep patterns are extracted from the original stress or strain waveforms to obtain feature vectors characterizing the bearing state of the structure. Step S230: Perform visual feature extraction processing based on the apparent image data. By performing pixel-level crack segmentation, skeleton extraction, and width measurement along the skeleton normal, the total length, average width, maximum width, and fractal dimension of the crack are quantified to obtain the feature vector characterizing the apparent damage state.

[0024] Specifically, to ensure the comparability of data with different physical dimensions, the acquired numerical time-series signals (such as stress, strain, and acceleration) are Z-score standardized: ; In the formula, These are the original data points; This is the mean of the dataset for this type of signal; Standard deviation; These are the standardized data points, with a mean of 0 and a standard deviation of 1.

[0025] After standardizing the original time-series signals, differentiated feature extraction was performed on the three types of heterogeneous data sources. First, for vibration response time-series data (such as acceleration signals), discrete wavelet packet transform was used to decompose it to a specific number of levels, obtaining a series of signal components in different frequency bands. Then, the energy characteristics of each frequency band component were calculated to characterize the energy distribution in different frequency bands. The calculation formula is as follows: ; In the formula, The wavelet packet frequency band energy of the acceleration response signal represents the first... In the layer decomposition, the first The energy value corresponding to each node (frequency band); This is the standardized timing signal; For time; To represent the first node in the sequence of node coefficients A discrete point; This represents the total number of wavelet packet coefficients; These are the wavelet packet decomposition coefficients; This is the differential symbol.

[0026] Simultaneously, the time-domain statistical characteristics of the original signal, such as the amplitude factor and crease factor, are directly calculated to characterize the signal's impulse characteristics. These frequency-domain and time-domain characteristics together constitute the feature vector representing the dynamic state of the structure. The calculation formula is as follows: ; ; In the formula, Amplitude factor; Peak factor; It is the root mean square value; The sampling point number; This represents the number of signal sampling points. These are discrete acceleration signal data points; This is the peak value.

[0027] Secondly, for time-series data of internal forces (such as stress and strain), a one-dimensional convolutional neural network is constructed for deep feature extraction. This network automatically learns and extracts abstract features reflecting the local patterns and deep dynamics of material stress and deformation by defining trainable convolutional kernel weights and biases and performing convolution, activation, and pooling operations on the original waveform. This forms a characterization of the static forces of the structure. The feature vector of the bearing state. Layer convolution operation is: ; In the formula, For the first The intensity of the internal force characteristics of the higher layers learned by the layer at time m; The activation function (such as ReLU); and For trainable convolutional kernel weights and biases; This is the index inside the convolution kernel; For time span; For time; For the first The internal force sequence of a layer (input layer or the layer above).

[0028] Finally, for the apparent image data, image preprocessing is performed first, including illumination normalization using Limiting Contrast Adaptive Histogram Equalization (CLAHE), linear contrast stretching, and noise suppression using nonlocal mean or Gaussian filtering. The preprocessed image is then input into an improved U-Net++ segmentation network for pixel-level crack identification. This network is optimized by combining binary cross-entropy and Dice coefficient loss, outputting a crack probability map. The Dice Loss formula is: ; In the formula, The loss function; The pixel value of the actual label; The probability value predicted by the model; It is a smoothing factor; These are the spatial coordinates of a pixel.

[0029] probability Figure 2 After obtaining the crack mask through quantization, skeletonization and connection are performed to obtain a crack skeleton with a single pixel width. The skeleton is scanned along the normal direction to the mask boundary, and the local width of each point is accurately measured by calculating the Euclidean distance between the two points. Then, statistical measures such as average width, maximum width, and width variance are obtained. The calculation formula is: ; In the formula, This refers to the width of the local crack. This is the left boundary point of the mask; This is the right boundary point of the mask.

[0030] Furthermore, using box counting, the crack skeleton is covered with meshes of different side lengths. The number of meshes containing skeleton pixels is counted, and the fractal dimension of the crack skeleton is calculated to quantify morphological complexity. Histograms or rose diagrams are plotted based on the skeleton tangent directions to analyze the dominant crack direction. The extracted geometric and morphological parameters, such as total length, width statistics, and fractal dimension, collectively constitute a feature vector characterizing the apparent damage state of the structure. The calculation formula is as follows: ; In the formula, For fractal dimension, The larger the value, the more complex the crack morphology and the more branches it has; This represents the number of grid cells containing the skeleton pixels; is the side length of the grid.

[0031] Further, step S300 includes steps S310 to S340.

[0032] Step S310: Standardize the feature vector set by calculating the mean and standard deviation of each feature vector to make all features comparable on the same scale, and obtain the standardized feature vector. Step S320: Adaptive weighting is performed on the standardized feature vector and environmental parameter time series data. The environmental feature vector, which quantifies the number of freeze-thaw cycles, the duration of negative temperature and the temperature gradient, is concatenated with the feature vectors of each data source. The feature channel is scaled by generating a weight vector related to the current freeze-thaw environment state using an attention mechanism, and the reweighted feature vector is obtained. Step S330: Perform concatenation processing based on all reweighted feature vectors. By connecting feature vectors from different data sources along the feature dimension, a concatenated fused feature vector is obtained. Step S340: Perform deep nonlinear interactive processing based on the fused feature vector. By inputting the fused feature vector into a fully connected neural network for nonlinear transformation and feature combination, the cross-modal damage correlation pattern under freeze-thaw and load coupling is mined to obtain a high-level feature representation.

[0033] Specifically, Z-score standardization is performed on each feature vector, calculating its mean and standard deviation on the training dataset and transforming them to eliminate the incomparability caused by differences in physical units and numerical ranges. This normalizes all features to a standard normal distribution with a mean of 0 and a standard deviation of 1, resulting in a standardized feature vector. The calculation formula is: ; In the formula, For the standardized vector; This is the original feature vector; and The first The mean and standard deviation of each feature vector are calculated on the training set.

[0034] Subsequently, adaptive weighted processing is performed. The core of this process is to construct an environmental feature vector from quantitative parameters reflecting the severity of the cold-region environment, including the number of freeze-thaw cycles, duration of sub-zero temperatures, and current temperature stage label. Specifically, the number of freeze-thaw cycles is calculated based on historical temperature data, counting the number of days within the assessment time window (e.g., the past 7 days, 30 days, total operational age, etc.) where the daily maximum temperature is >0℃ and the daily minimum temperature is <0℃; the duration of sub-zero temperatures is the number of hours or days with consecutive temperatures below 0℃; the temperature gradient is calculated by determining the temperature difference between different parts of the structure and the maximum temperature change rate within 24 hours; and the current temperature stage label represents the current ambient temperature and the freeze-thaw stage. This environmental feature vector is then fused with feature vectors from various standardized data sources in the early stages, i.e., concatenated along the feature dimensions to form... ,in, For each standardized data source feature vector, The environmental feature vector is used as input. The fused representation is fed into a lightweight attention subnetwork: first, the information from each feature channel is compressed into a scalar using global average pooling; then, it is activated by a two-layer fully connected network (bottleneck structure) containing ReLU and Sigmoid activation functions, generating an attention weight vector strongly correlated with the current freeze-thaw environment state. The calculation formula is: ; ; In the formula, This is the compressed vector; For location index; For feature dimensions; For feature vectors; This is an attention weight vector that is strongly correlated with the current environmental state; For the Sigmoid function; It is the ReLU activation function; and This is a trainable weight matrix.

[0035] The weight vector is then multiplied channel-by-channel with the corresponding original feature vector, enhancing important features and weakening less important features, thus obtaining the reweighted feature vector. The calculation formula is as follows: ; In the formula, The eigenvectors are the reweighted eigenvectors; This is the original feature vector; This is an attention weight vector that is strongly correlated with the current environmental state; This represents element-wise multiplication (Hadamard product).

[0036] Next, step S330 concatenates all the reweighted feature vectors along the feature dimension, integrating information from different modalities such as vibration, internal force, and images into a unified vector representation, resulting in the concatenated fused feature vector. The calculation formula is: ; In the formula, This is the concatenated fused feature vector; For splicing operations; , and These represent feature vectors containing multi-dimensional information such as vibration, internal force, and images.

[0037] Finally, step S340 performs deep nonlinear feature interaction. The concatenated fused feature vector is input into a feature interaction network consisting of one or more fully connected layers. This network, through trainable weights and bias parameters, performs complex nonlinear transformations and combinations on weighted and concatenated features from different data sources in a higher-dimensional feature space. This process aims to deeply explore the deep correlations between features characterizing different damage modes such as material stiffness softening, brittle response, and crack propagation under the coupled effects of freeze-thaw cycles and heavy-load train loads, ultimately outputting a highly fused and abstract high-level feature representation. The calculation formula is: ; ; In the formula, These are the hidden features output from the first layer; Represents the ReLU activation function; This is the concatenated fused feature vector; , , , These are trainable parameters. The final output, a deeply fused high-level feature representation, can be directly input into a subsequent classifier for severity level prediction.

[0038] Further, step S400 includes steps S410 to S430.

[0039] Step S410: Based on the high-level feature representation and the corresponding known damage state labels, perform model training. Train a random forest model and evaluate the contribution of each feature in the high-level feature representation to the judgment of the tunnel structure damage state based on the reduction of Gini impurity, and obtain the importance score of each feature. Step S420: Sort the features according to their importance scores. By arranging all features in descending order of importance scores, a feature list is obtained. Step S430: Perform feature subset selection processing based on the feature list. By selecting the top K features with the highest importance ranking from the feature list, the optimal feature subset is obtained.

[0040] Specifically, a random forest model is first trained quickly using all high-level feature representations and their corresponding known poor state labels. Then, the contribution of each feature is evaluated based on this model, with the core metric being the reduction in Gini impurity resulting from feature splitting at decision tree nodes. For any given feature, its importance score is calculated based on the cumulative contribution of all decision trees in the forest, using the following formula: ; In the formula, Representation of features Importance rating; The total number of decision trees; Indicates a single tree; The node number in the decision tree; For nodes The sample weight ratio; For nodes The reduction in Gini impurity before and after splitting; Indicates at node The feature used for splitting is This calculation quantifies the discriminative power of features in distinguishing different states of degradation.

[0041] Subsequently, based on the calculation results, feature sorting is performed, arranging all features in descending order according to their importance scores to generate an ordered feature list. Finally, based on a preset number of features K or an importance threshold, the top K features with the highest importance ranking are selected from the descending feature list to form the optimal feature subset used to train the prediction model, thus completing the process of selecting a concise and efficient feature subset from high-dimensional fused features.

[0042] Further, step S500 includes steps S510 to S530.

[0043] Step S510: Iterative model training is performed based on the optimal feature subset and the corresponding known poor state labels. By using the gradient boosting decision tree algorithm, the prediction residual of the previous model is used as the target to fit a new regression tree, and the prediction results of the gradient boosting decision tree are accumulated round by round to minimize the overall loss function, thus obtaining a preliminary regression model. Step S520: Perform numerical range constraint processing on the original output of the preliminary regression model. By applying the Sigmoid function, the continuous predicted values ​​are mapped to the interval between 0 and 1 to obtain the continuous defect index. Step S530: Perform engineering state mapping processing based on the continuous deterioration index. By dividing the numerical interval into four sub-intervals proportionally and assigning a safety level and engineering state description to each sub-interval, a deterioration prediction model is constructed.

[0044] Specifically, this process uses the optimal feature subset as input features and trains it using the gradient boosting decision tree algorithm. Due to its high accuracy and ability to handle complex nonlinear relationships, this algorithm is an ideal choice for learning the mapping relationship between lining defect levels from the optimal feature subset. Its training is an additive process, calculated using the following formula: ; In the formula, For the first The predicted output value for each sample; For prediction functions; For index variables; This represents the total number of base learners in the entire addition model; For the first decision trees for the first Output values ​​for each sample; Let the function space consist of all regression trees; In the first The regression decision tree is trained and added to the model during rounds of iteration.

[0045] Next, the training process is an iterative optimization process: First, a loss function is defined, such as multi-class cross-entropy loss, to measure the difference between the model's predictions and the true labels, expressed as: ; In the formula, For multi-class cross-entropy loss; The sample number; The total number of samples; This is the index of the sample category; The total number of sample categories; For the sample The true label (unique hot encoding); Predict samples for the model Category The probability of.

[0046] In the In each iteration, the negative gradient (i.e., pseudo-residual) of the current model for each sample is calculated using the following formula: ; In the formula, For the first In each iteration, the current model performs a certain number of iterations for each sample. The negative gradient; The symbol is for partial differentials; Let be the loss function, measuring the th The true value of each sample Compared with current forecast value The differences between them; Indicates the first The model prediction value corresponding to each sample; subscript This indicates that the current model The predicted value is substituted into the derivative formula for calculation.

[0047] Next, a new regression tree is fitted using this pseudo-residual as the objective; the optimal weights of this tree are determined through linear search, and the model is updated using the following formula: ; In the formula, For the first The model after rounds of iterations; For premenstrual The model after rounds of iterations; For the first The error value that should be corrected during each iteration; For indicator functions; The learning rate is used to prevent overfitting.

[0048] This process is performed round by round, minimizing the overall loss function by continuously fitting the residuals from the previous round, ultimately obtaining a preliminary regression model capable of capturing the complex nonlinear relationship between features and the poor state. Step S520 performs numerical range constraints on the original output of the preliminary regression model. Since the original output of the gradient boosting decision tree is a continuous real value, to ensure that its output strictly falls within the (0,1) interval and has probabilistic interpretation, the Sigmoid function is applied to the final output layer for transformation, as shown in the formula: ; In the formula, The defect index predicted by the model; The output value of the Sigmoid function; It is a natural constant; This function generates the raw output of the gradient boosting regression tree model. It ensures that regardless of... Why is it worth it? All values ​​are mapped to the range (0, 1) and monotonically increase, which aligns with the intuitive understanding that the larger the value, the more severe the damage.

[0049] Finally, step S530 performs engineering state mapping processing to complete the final construction of the prediction model. This step divides the numerical range of the continuous deterioration index [0, 1] into four sub-intervals proportionally, and assigns a clear safety level (e.g., level one to level four) and a corresponding engineering state description (e.g., safe to unsafe) to each sub-interval, thus establishing the tunnel structure safety level correspondence table shown in Table 1 below. By establishing a mapping rule from the continuous deterioration index to discrete safety levels, the numerical results output by the algorithm are transformed into early warning information that can be directly understood and applied in engineering practice, thereby constructing a complete deterioration prediction model with state assessment and graded early warning capabilities.

[0050] Table 1. Correspondence of Tunnel Structure Safety Levels Based on Defect Index

[0051] Further, step S600 includes steps S610 to S640.

[0052] Step S610: Data acquisition and synchronous processing are performed based on the field monitoring system for structural defects in heavy-haul railway tunnels in cold regions to obtain the original field monitoring data; Step S620: Perform real-time feature extraction and preprocessing based on the original on-site monitoring data. By performing feature extraction processing on the vibration, internal force, image and environmental data on-site, a real-time feature vector that meets the input requirements of the damage prediction model is obtained. Step S630: Input the real-time feature vector into the damage prediction model for current state evaluation. Calculate the current damage index, which characterizes the immediate health of the tunnel structure, using the nonlinear regression function trained in the model. Step S640: Perform degradation trend prediction processing based on the current degradation index. By inputting the current degradation index into the preset long short-term memory network time series model, learn the evolution law of the degradation index over time, and calculate the predicted value of the degradation index for the future preset time step. Step S650: Perform safety status prediction processing based on the predicted value of the damage index. By querying the preset damage index-safety level mapping table, the continuous index values ​​are converted into the corresponding engineering safety levels, thereby obtaining the qualitative-quantitative integrated prediction result of the future safety status of the tunnel structure.

[0053] Specifically, step S610 is the stage of synchronous acquisition of multi-source sensing data on site, which relies on a specially deployed on-site monitoring system for structural defects in heavy-haul railway tunnels in cold regions. This system is a comprehensive hardware platform integrating multi-dimensional physical quantity sensing and reliable data transmission, consisting of the following four functional modules working together: Structural Deformation, Stress, and Temperature Monitoring Module: This module directly and continuously senses the mechanical state, deformation information, and real-time temperature of the tunnel lining. It mainly includes surface-mounted strain gauges, crack gauges, and thermometers, used to monitor the strain distribution on the lining surface, the width variation of existing cracks, and the temperature variation of the lining, respectively.

[0054] Vehicle-induced vibration monitoring module: Used to capture dynamic vibrations exerted on the tunnel structure in real time as operating trains pass by. The core sensing element of this module is a modular accelerometer, which is deployed at key locations in the tunnel lining or track structure to measure the vibration acceleration response caused by train operation.

[0055] Image scanning module: Used to acquire visual information about the appearance of the tunnel lining. This module consists of a modular high-definition camera and a dedicated active lighting module, ensuring stable acquisition of high-definition images of the lining appearance even in low-light conditions inside the tunnel, for identifying surface damage such as cracks and spalling.

[0056] Information transmission module: As the system's data pathway, it is responsible for efficiently and reliably aggregating and remotely transmitting the raw data collected by the aforementioned sensing modules to the central data processing server. This module adopts a hybrid networking architecture combining wired transmission (such as industrial Ethernet) and wireless transmission (such as 4G / 5G, LoRa) to ensure the stability and real-time performance of the data link in complex tunnel environments.

[0057] Step S620 performs real-time feature engineering on the raw on-site monitoring data. To ensure consistency with the model training phase, this step performs the same feature extraction process as step S200 on various data collected on-site, including vibration response, internal forces, apparent images, and environmental parameters. This involves converting vibration data into time-frequency domain statistical features, extracting depth features from internal force data using a one-dimensional convolutional neural network, and converting image data into crack geometric features, thereby generating a real-time feature vector that meets the input format and dimensional requirements of the constructed deterioration prediction model. Step S630 uses the constructed model to quantitatively assess the tunnel's immediate health status. The real-time feature vector generated in step S620 is input into the deterioration prediction model, whose core is a trained gradient boosting decision tree nonlinear regression function. The model calculates based on the input features and directly outputs a continuous value between 0 and 1, i.e., the current deterioration index. This index is a comprehensive and quantitative representation of the overall deterioration level of the tunnel structure at this moment. Step S640 performs time-series prediction of the deterioration trend based on historical data. To achieve the leap from current state assessment to future trend prediction, the system pre-trains a Long Short-Term Memory (LSTM) time series model. This network can effectively learn and memorize complex temporal dependencies and long-term evolution patterns in multi-source monitoring data. During the training phase, historical multi-source monitoring data sequences from the database (such as time-series acceleration signals, image damage quantification features, stress data, and crack propagation data) are used as input, and corresponding structural degradation indices or state labels with temporal continuity are used as supervision targets to train the LSTM network, enabling it to master the mapping law from historical data sequences to future state changes. Current and historical degradation indices are input into this LSTM model, and the model, through its internal state propagation and calculation, can deduce predicted degradation index values ​​for one or more preset time steps in the future. Step S650 transforms the quantitative prediction results into early warning information that can directly guide engineering practice. This step automatically queries and calls a preset degradation index-safety level mapping table to convert the predicted future degradation index value into a corresponding qualitative engineering safety level. For example, when the model predicts that the safety level will degrade from Level 2 (basic safety) to Level 3 (potentially unsafe), maintenance intervention measures can be planned in advance. This enables tunnel maintenance management to shift from post-event handling to pre-event prediction and early warning in an intelligent maintenance model.

[0058] In summary, the core of the method described in this embodiment lies in constructing a closed-loop technical system that integrates holographic perception, intelligent fusion, precise assessment, and forward-looking prediction. The innovation of this system is reflected in several key, interconnected steps: First, it employs a mechanism-site dual-space coupling data production and verification approach. This involves calibrating the damage mechanism and generating standard samples in a controlled environment through indoor multi-field coupled model experiments, while simultaneously acquiring real operational data through long-term on-site monitoring. This lays the foundation for solving the problems of single data and difficulty in correlating superficial and intrinsic damage in traditional methods, forming a multi-source heterogeneous database with clear physical meaning. Second, it designs a feature intelligent fusion and cascaded decision-making model based on an attention mechanism. Through a process of adaptive weighted fusion of channel attention → feature importance optimization → gradient boosting regression tree precise mapping, it achieves deep feature mining and high-precision, interpretable quantitative assessment of complex damage processes, overcoming the drawbacks of isolated data utilization, simple models, and coarse results. Furthermore, a deterioration index with clear physical meaning was innovatively defined as the core continuous quantitative indicator, and a strictly corresponding four-level engineering safety system was established. This directly transforms the algorithm output into qualitative-quantitative integrated early warning information that engineers can understand, bridging the gap between intelligent models and maintenance decisions. Finally, a time-series prediction model based on long short-term memory networks was integrated on the basis of real-time assessment, enabling forward-looking prediction of future structural deterioration trends. Moreover, the system architecture supports continuous optimization of the model using new data, giving the entire system the ability to predict risks, support preventative decision-making, and self-evolve, thus forming a proactive closed-loop intelligent agent.

[0059] The core of this invention lies in the complete technical solution comprised of the aforementioned steps, specifically encompassing: First, a method and system for constructing a full-cycle data foundation from laboratory mechanism calibration to on-site data verification. This protects the specific technical path and implementation method for constructing and verifying a specific multi-source heterogeneous database by combining a multi-field coupled physical model test device integrating freeze-thaw cycles, mechanical loading, and image recognition with a fixed on-site monitoring system including structural deformation, vehicle-induced vibration, and image scanning. Second, an intelligent assessment and decision-making process and method based on intelligent feature fusion, continuous exponential regression, and engineering grade mapping. This protects the complete algorithm process and steps for generating high-order representations through adaptive weighted fusion and selection of multi-source heterogeneous features based on attention mechanisms, then using gradient boosting regression trees and other models to output a continuous degradation index, and mapping it to multi-level engineering safety levels. Third, an intelligent early warning system architecture and process based on real-time state assessment, degradation trend prediction, and system closed-loop evolution. This protects the complete system workflow and self-evolution method that, based on real-time state assessment, further integrates a degradation trend prediction model based on time series analysis to achieve forward-looking early warning, and can utilize new data to iteratively update the model. The scope of protection aims to cover the overall method for realizing the above complete technical solution, the corresponding intelligent monitoring, evaluation and early warning system, and the core special devices involved therein (such as the multi-field coupling test system and crack quantitative identification device).

[0060] Example 2: like Figure 2 As shown, this embodiment provides a tunnel structure deterioration prediction system based on multi-source data fusion sensing. The system includes: The acquisition module 901 is used to acquire the raw test data generated by the cold region heavy-load railway tunnel structure deterioration test system. The raw test data includes the vibration response time series data, internal force time series data, appearance image data and environmental parameter time series data of the lining specimen under the coupled action of freeze-thaw cycle and heavy load. The extraction module 902 is used to extract features from the original experimental data to obtain feature vector sets that characterize the dynamic state, static bearing state and apparent damage state of the structure, respectively. The fusion module 903 is used to perform feature fusion based on the feature vector set and environmental parameter time series data. By combining the environmental state, it adaptively assigns weights to features from different sources and performs non-linear interaction after combining the weighted features to obtain a high-level feature representation. The filtering module 904 is used to perform feature filtering and dimensionality reduction based on the high-level feature representation. By calculating the importance of features and sorting them in descending order of importance, the optimal feature subset after dimensionality reduction is obtained. Module 905 is used to perform model mapping based on the optimal feature subset. By training a nonlinear regression function based on the optimal feature subset and the known defect state labels, a defect prediction model is constructed. The prediction module 906 is used to input the original on-site monitoring data into the deterioration prediction model to obtain the prediction results of the future safety status of the tunnel structure.

[0061] In one specific embodiment of this application, the extraction module 902 includes: The first extraction unit is used to perform dynamic feature extraction processing based on vibration response time series data. By performing wavelet packet decomposition on the vibration response time series data, signal components of different frequency bands are obtained, and the energy of each frequency band and the amplitude factor and peak factor of the signal are calculated to obtain the feature vector characterizing the impact characteristics of the structure. The second extraction unit is used to perform load-bearing feature extraction processing based on internal force time series data. By constructing a one-dimensional convolutional neural network, it extracts local and deep patterns from the original stress or strain waveforms to obtain feature vectors that characterize the load-bearing state of the structure. The third extraction unit is used to perform visual feature extraction processing based on the apparent image data. By performing pixel-level crack segmentation, skeleton extraction and width measurement along the skeleton normal of the image, the total length, average width, maximum width and fractal dimension of the crack are quantified to obtain the feature vector characterizing the apparent damage state.

[0062] In one specific embodiment of this application, the fusion module 903 includes: The first fusion unit is used to perform standardization processing based on the feature vector set. By calculating the mean and standard deviation of each feature vector, all features are made to be on the same comparable scale, resulting in standardized feature vectors. The second fusion unit is used to perform adaptive weighting processing based on the standardized feature vector and environmental parameter time series data. It concatenates the environmental feature vector, which quantifies the number of freeze-thaw cycles, the duration of negative temperature and the temperature gradient, with the feature vectors of each data source, and uses an attention mechanism to generate a weight vector related to the current freeze-thaw environment state to scale the feature channel, thus obtaining the reweighted feature vector. The third fusion unit is used to concatenate all the reweighted feature vectors. By connecting the feature vectors from different data sources along the feature dimension, the concatenated fusion feature vector is obtained. The fourth fusion unit is used to perform deep nonlinear interactive processing based on the fused feature vector. By inputting the fused feature vector into a fully connected neural network for nonlinear transformation and feature combination, it mines the cross-modal damage correlation pattern under freeze-thaw and load coupling to obtain a high-level feature representation.

[0063] In one specific embodiment of this application, the screening module 904 includes: The first screening unit is used to train the model based on the high-level feature representation and the corresponding known defect state labels. By training a random forest model and evaluating the contribution of each feature in the high-level feature representation to the defect state of the tunnel structure based on the reduction of Gini impurity, the importance score of each feature is obtained. The second filtering unit is used to sort features according to their importance scores. By arranging all features in descending order of importance scores, a feature list is obtained. The third filtering unit is used to perform feature subset selection based on the feature list. The optimal feature subset is obtained by selecting the top K features with the highest importance ranking from the feature list.

[0064] In one specific embodiment of this application, the construction module 905 includes: The first building unit is used to perform iterative model training based on the optimal feature subset and the corresponding known poor state labels. By using the gradient boosting decision tree algorithm, the prediction residual of the previous model is used as the target to fit a new regression tree, and the prediction results of the gradient boosting decision tree are accumulated in each round to minimize the overall loss function and obtain the preliminary regression model. The second building unit is used to perform numerical range constraint processing on the original output of the preliminary regression model. By applying the Sigmoid function, the continuous predicted values ​​are mapped to the interval between 0 and 1 to obtain the continuous defect index. The third building unit is used to perform engineering state mapping processing based on the continuous deterioration index. By dividing the numerical interval into four sub-intervals proportionally and assigning a safety level and engineering state description to each sub-interval, a deterioration prediction model is constructed.

[0065] In one specific embodiment of this application, the prediction module 906 includes: The first prediction unit collects and synchronously processes data based on the field monitoring system for structural defects in heavy-haul railway tunnels in cold regions, and obtains the original field monitoring data. The second prediction unit is used to perform real-time feature extraction and preprocessing based on the original on-site monitoring data. By performing feature extraction and processing on-site vibration, internal force, image and environmental data, a real-time feature vector that meets the input requirements of the damage prediction model is obtained. The third prediction unit is used to input the real-time feature vector into the deterioration prediction model for current state evaluation. The current deterioration index, which characterizes the instantaneous health of the tunnel structure, is calculated through the nonlinear regression function trained in the model. The fourth prediction unit is used to predict the degradation trend based on the current degradation index. By inputting the current degradation index into a preset long short-term memory network time series model, it learns the evolution law of the degradation index over time and calculates the predicted value of the degradation index for a preset time step in the future. The fifth prediction unit is used to perform safety status forecasting based on the predicted value of the damage index. By querying the preset damage index-safety level mapping table, it converts continuous index values ​​into corresponding engineering safety levels, thereby obtaining a qualitative-quantitative integrated prediction result of the future safety status of the tunnel structure.

[0066] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting tunnel structural deterioration based on multi-source data fusion sensing, characterized in that, include: The raw test data generated by the cold-region heavy-haul railway tunnel structure deterioration test system is obtained. The raw test data includes vibration response time series data, internal force time series data, appearance image data and environmental parameter time series data of the lining specimen under the coupled action of freeze-thaw cycle and heavy load. Based on the original experimental data, feature extraction was performed to obtain feature vector sets representing the dynamic state, static bearing state, and apparent damage state of the structure, respectively. Feature fusion is performed based on the feature vector set and the environmental parameter time series data. By combining the environmental state, weights are adaptively assigned to features from different sources, and the weighted features are combined and then subjected to nonlinear interaction to obtain a high-level feature representation. Feature filtering and dimensionality reduction are performed based on the high-level feature representation. The optimal feature subset after dimensionality reduction is obtained by calculating the feature importance and arranging them in descending order of importance. Model mapping is performed based on the optimal feature subset, and a nonlinear regression function is trained based on the optimal feature subset and the known deterioration state labels to construct a deterioration prediction model; The original on-site monitoring data is input into the defect prediction model to obtain the prediction results of the future safety status of the tunnel structure.

2. The method for predicting tunnel structural deterioration based on multi-source data fusion sensing according to claim 1, characterized in that, Based on the original experimental data, feature extraction is performed to obtain feature vector sets representing the structural dynamic state, load-bearing state, and apparent damage state, respectively, including: Dynamic feature extraction processing is performed on the vibration response time series data. By performing wavelet packet decomposition on the vibration response time series data, signal components of different frequency bands are obtained, and the energy of each frequency band and the amplitude factor and peak factor of the signal are calculated to obtain the feature vector characterizing the impact characteristics of the structure. Based on the internal force time series data, load-bearing feature extraction processing is performed. By constructing a one-dimensional convolutional neural network, local and deep patterns are extracted from the original stress or strain waveforms to obtain feature vectors characterizing the load-bearing state of the structure. Visual feature extraction processing is performed on the apparent image data. By performing pixel-level crack segmentation, skeleton extraction, and width measurement along the skeleton normal, the total length, average width, maximum width, and fractal dimension of the crack are quantified to obtain a feature vector characterizing the apparent damage state.

3. The method for predicting tunnel structural defects based on multi-source data fusion sensing according to claim 1, characterized in that, Feature fusion is performed based on the feature vector set and the time-series environmental parameter data. By adaptively assigning weights to features from different sources in conjunction with the environmental state, and then performing nonlinear interaction on the weighted features, a high-level feature representation is obtained, including: The feature vector set is standardized by calculating the mean and standard deviation of each feature vector to make all features comparable, thus obtaining the standardized feature vector. Adaptive weighting is performed on the standardized feature vector and the environmental parameter time series data. The environmental feature vector, which quantifies the number of freeze-thaw cycles, the duration of negative temperature and the temperature gradient, is concatenated with the feature vectors of each data source. An attention mechanism is used to generate a weight vector related to the current freeze-thaw environment state to scale the feature channel, resulting in a reweighted feature vector. The concatenated feature vectors are obtained by concatenating the feature vectors from different data sources along the feature dimension. Based on the fused feature vector, deep nonlinear interactive processing is performed. By inputting the fused feature vector into a fully connected neural network for nonlinear transformation and feature combination, cross-modal damage correlation patterns under freeze-thaw and load coupling are mined to obtain the high-level feature representation.

4. The method for predicting tunnel structural deterioration based on multi-source data fusion sensing according to claim 1, characterized in that, Feature filtering and dimensionality reduction are performed based on the high-level feature representation. By calculating the feature importance and sorting them in descending order of importance, the optimal feature subset after dimensionality reduction is obtained, including: The model is trained based on the high-level feature representation and the corresponding known damage state labels. The contribution of each feature in the high-level feature representation to the judgment of the tunnel structure damage state is evaluated by training a random forest model and evaluating the amount of reduction in Gini impurity, and the importance score of each feature is obtained. The features are sorted according to their importance scores. By arranging all features in descending order of importance, a feature list is obtained. The feature subset selection process is performed based on the feature list. The optimal feature subset is obtained by selecting the top K features with the highest importance ranking from the feature list.

5. The method for predicting tunnel structural defects based on multi-source data fusion sensing according to claim 1, characterized in that, Model mapping is performed based on the optimal feature subset, and a nonlinear regression function is trained based on the optimal feature subset and known deterioration state labels to construct a deterioration prediction model, including: Iterative model training is performed based on the optimal feature subset and the corresponding known poor state labels. By using the gradient boosting decision tree algorithm, the prediction residual of the previous model is used as the target to fit a new regression tree, and the prediction results of the gradient boosting decision tree are accumulated round by round to minimize the overall loss function, thus obtaining a preliminary regression model. The numerical range constraint is performed on the original output of the preliminary regression model. By applying the Sigmoid function, the continuous predicted values ​​are mapped to the interval between 0 and 1 to obtain the continuous defect index. Based on the continuous deterioration index, an engineering state mapping process is performed. By dividing the numerical interval into four sub-intervals proportionally and assigning a safety level and engineering state description to each sub-interval, a deterioration prediction model is constructed.

6. A tunnel structure deterioration prediction system based on multi-source data fusion sensing, characterized in that, include: The acquisition module is used to acquire the raw test data generated by the cold-region heavy-load railway tunnel structure deterioration test system. The raw test data includes the vibration response time series data, internal force time series data, appearance image data and environmental parameter time series data of the lining specimen under the coupled action of freeze-thaw cycle and heavy load. The extraction module is used to extract features based on the original test data to obtain feature vector sets representing the dynamic state, static bearing state and apparent damage state of the structure, respectively. The fusion module is used to perform feature fusion based on the feature vector set and the environmental parameter time series data. By combining the environmental state, it adaptively assigns weights to features from different sources and performs nonlinear interaction on the weighted features to obtain a high-level feature representation. The filtering module is used to perform feature filtering and dimensionality reduction based on the high-level feature representation. By calculating the feature importance and sorting them in descending order of importance, the optimal feature subset after dimensionality reduction is obtained. The module is used to perform model mapping based on the optimal feature subset, and to construct a damage prediction model by training a nonlinear regression function based on the optimal feature subset and the known damage state labels. The prediction module is used to input the original on-site monitoring data into the deterioration prediction model to obtain the prediction results of the future safety status of the tunnel structure.

7. The tunnel structure deterioration prediction system based on multi-source data fusion sensing according to claim 6, characterized in that, The extraction module includes: The first extraction unit is used to perform dynamic feature extraction processing based on the vibration response time series data. By performing wavelet packet decomposition on the vibration response time series data, signal components of different frequency bands are obtained, and the energy of each frequency band and the amplitude factor and peak factor of the signal are calculated to obtain the feature vector characterizing the impact characteristics of the structure. The second extraction unit is used to perform load-bearing feature extraction processing based on the internal force time series data. By constructing a one-dimensional convolutional neural network, it extracts local and deep patterns from the original stress or strain waveform to obtain a feature vector characterizing the load-bearing state of the structure. The third extraction unit is used to perform visual feature extraction processing based on the apparent image data. By performing pixel-level crack segmentation, skeleton extraction, and width measurement along the skeleton normal, the total length, average width, maximum width, and fractal dimension of the crack are quantified to obtain a feature vector characterizing the apparent damage state.

8. The tunnel structure deterioration prediction system based on multi-source data fusion sensing according to claim 6, characterized in that, The fusion module includes: The first fusion unit is used to perform standardization processing on the feature vector set, and to make all features on the same comparable scale by calculating the mean and standard deviation of each feature vector to obtain the standardized feature vector. The second fusion unit is used to perform adaptive weighting processing based on the standardized feature vector and the environmental parameter time series data. It concatenates the environmental feature vector, which quantifies the number of freeze-thaw cycles, the duration of negative temperature and the temperature gradient, with the feature vectors of each data source, and uses an attention mechanism to generate a weight vector related to the current freeze-thaw environment state to scale the feature channel, thereby obtaining the reweighted feature vector. The third fusion unit is used to perform concatenation processing based on all the reweighted feature vectors. By connecting the feature vectors from different data sources along the feature dimension, a concatenated fusion feature vector is obtained. The fourth fusion unit is used to perform deep nonlinear interactive processing based on the fused feature vector. By inputting the fused feature vector into a fully connected neural network for nonlinear transformation and feature combination, it mines the cross-modal damage correlation pattern under freeze-thaw and load coupling to obtain the high-level feature representation.

9. The tunnel structure deterioration prediction system based on multi-source data fusion sensing according to claim 6, characterized in that, The filtering module includes: The first screening unit is used to perform model training based on the high-level feature representation and the corresponding known defect state labels. By training a random forest model and evaluating the contribution of each feature in the high-level feature representation to the defect state of the tunnel structure based on the reduction of Gini impurity, the importance score of each feature is obtained. The second filtering unit is used to sort features according to the importance score. By arranging all features in descending order of importance score, a feature list is obtained. The third filtering unit is used to perform feature subset selection processing based on the feature list, and to obtain the optimal feature subset by selecting the top K features with the highest importance ranking from the feature list.

10. The tunnel structure deterioration prediction system based on multi-source data fusion sensing according to claim 6, characterized in that, The building module includes: The first construction unit is used to perform iterative model training based on the optimal feature subset and the corresponding known poor state labels. By using the gradient boosting decision tree algorithm, the prediction residual of the previous model is used as the target to fit a new regression tree, and the prediction results of the gradient boosting decision tree are accumulated round by round to minimize the overall loss function, so as to obtain a preliminary regression model. The second construction unit is used to perform numerical range constraint processing based on the original output of the preliminary regression model. By applying the Sigmoid function, the continuous predicted values ​​are mapped to the interval between 0 and 1 to obtain the continuous defect index. The third construction unit is used to perform engineering state mapping processing based on the continuous deterioration index. By dividing the numerical interval into four sub-intervals proportionally and assigning a safety level and engineering state description to each sub-interval, a deterioration prediction model is constructed.

Images