A highway tunnel disease occurrence prediction algorithm based on drift perception multi-scale section modeling

Abstract

The application discloses a tunnel disease prediction method based on drift perception and multi-scale section modeling, the method obtains tunnel per-meter detection data formed according to years, and carries out unified normalization processing of fields, years, mileage directions, missing states and repeated records; continuous mileage is divided into multiple sections according to a preset section length, and a section-year sample is constructed; a multi-scale space-time feature is constructed by fusing a section history state, adjacent section space influence, a short-term and long-term time window, trend change, state transfer and data quality marking, and a target risk score, a prediction label and a risk sorting result of each section in the next year are output. The application reduces per-meter positioning error and isolated noise interference, improves the robustness and interpretability of cross-year prediction, and provides a sortable high-risk section list for inspection review, key monitoring and maintenance resource allocation.

Description

Technical Field

[0001] This invention relates to the fields of highway tunnel structural health monitoring, defect detection data processing, spatiotemporal prediction, and maintenance decision support, and particularly to a tunnel defect prediction method based on drift perception and multi-scale segment modeling. Background Technology

[0002] During the long-term operation of highway tunnels, the lining, pavement, drainage facilities, and ancillary structures are often threatened by various defects such as cracks, water leakage, spalling, deformation, voids, and pavement damage. To understand the service status of the tunnel structure, annual inspections typically generate inspection tables, defect record tables, or statistical tables arranged meter by meter according to mileage. This type of data exhibits significant characteristics such as fine mileage granularity, strong spatial correlation between adjacent locations, potential fluctuations in recording caliber and inspection standards between years, and interference from local repair works on original defect traces, bringing complexity to the basic data level for subsequent modeling and analysis.

[0003] If meter-by-meter positions are used directly as independent prediction units, factors such as detection positioning errors, disease extension across meters, differences in manual verification, local maintenance, and missing measurement records will lead to high noise in sample labels, seriously affecting the stability of model training. At the same time, prediction results at the single-meter scale are prone to isolated false alarms or missed alarms, lack spatial continuity, and are difficult to effectively support the actual operational needs in maintenance projects, and are difficult to directly transform into decision-making basis for engineering.

[0004] Existing statistical forecasting and machine learning methods often employ single-year cross-sectional data, single model structures, or fixed discrimination thresholds, failing to fully consider multi-dimensional dynamic information such as the historical continuity of the target section, the spatial transmission of influence between adjacent sections, and short-term mutations and long-term evolutionary trends. Especially when adjustments to detection strategies, the implementation of repair projects, changes in environmental conditions, or updates to recording standards cause significant shifts in the number of defects in recent years, fixed prediction scores or discrimination thresholds weaken the adaptability and robustness of cross-year prediction strategies. Therefore, there is an urgent need to construct a section-level prediction scheme for annual tunnel inspection data. Through multi-scale spatiotemporal feature extraction, historical persistence risk quantification, and a drift-aware adaptive selection mechanism, this scheme can effectively assess the risk of defects in the following year and output a rankable list of high-risk sections, providing scientific support for the precise allocation of maintenance resources. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies, such as meter-by-meter detection noise causing prediction unit failure, the difficulty of adapting a single fixed strategy to the annual distribution drift of defects, and the lack of segment risk ranking and decision interpretation fields in the output results. It provides a tunnel defect prediction method based on drift perception and multi-scale segment modeling that integrates multi-scale features, continuous quantization, and adaptive selection.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] A tunnel defect prediction method based on drift sensing and multi-scale segment modeling is used to predict the risk of segment-level defects in the following year based on historical annual tunnel inspection data, including: S1, acquire meter-by-meter tunnel inspection data for at least two years, the meter-by-meter tunnel inspection data including tunnel identification, route direction, mileage location, inspection year, and defect attribute fields; S2, standardize the fields, align the years, unify the mileage and direction, mark missing measurements and merge duplicate records in the meter-by-meter tunnel detection data to obtain standardized meter-by-meter data; S3, divide the continuous mileage into multiple segments according to the preset segment length, and aggregate the standardized meter-by-meter data into segment-year samples; S4. Construct multi-scale spatiotemporal features based on the segment-year samples. The multi-scale spatiotemporal features include historical features of the current segment, spatial features of adjacent segments, short-term time window features, long-term time window features, state transition features, and data quality features. S5, use the multi-scale spatiotemporal features to train a category-weighted disease occurrence prediction model, and obtain the model risk score of the section to be predicted; S6. Calculate the persistence risk score based on the historical persistence status of the target section and the historical persistence status of the neighboring sections, and obtain the fused risk score based on the model risk score and the persistence risk score. S7. Based on the recent rate of change in the number of diseases and historical verification criteria, perform drift-aware adaptive selection among the model risk score, the persistent risk score, and the fused risk score to obtain the target risk score. S8. Generate the next year's segment-level disease occurrence prediction label and risk ranking results based on the target risk score.

[0008] This invention provides an end-to-end method for predicting road tunnel defects at the segment level. By segmenting and aggregating high-noise meter-by-meter data, it transforms the data into engineering-scale risk units, integrates multi-scale spatiotemporal features with three types of candidate risk scores, and introduces a drift-aware adaptive selection mechanism. This effectively solves the technical problems of meter-by-meter positioning error, annual data drift, and insufficient adaptability of a single model, and achieves stable, interpretable, and directly applicable segment-level risk prediction that can be used for maintenance decisions across years.

[0009] The optimized step S2 includes: unifying the mileage representation in different detection tables into a continuous meter index in the same direction; mapping different disease names or disease codes to a unified disease dictionary; merging duplicate records of the same meter position, the same year, and the same disease according to the principle of maximum risk, the highest level, or a preset source priority; and marking missing meter positions, invalid detection records, and low-confidence records as untrainable samples, low-confidence samples, or samples with data quality weights. This method, through unifying mileage representation, mapping disease dictionaries, merging duplicate records, and marking missing data, eliminates the differences in format, scope, and quality among multi-source detection data. It provides clean, standardized basic data that retains quality information for subsequent modeling, avoiding model bias caused by inconsistent data quality. Simultaneously, it allows the missing data information itself to participate in the analysis as an auxiliary feature, improving data utilization.

[0010] In the optimized version of step S3, the preset segment length is 5 meters, 10 meters, or 20 meters, or other lengths determined based on detection and positioning accuracy, spatial continuity of defects, and maintenance operation scale. Each segment is jointly identified by tunnel markers, route direction, segment number, segment start mileage, and segment end mileage, ensuring that the same physical segment has a stable number across different years. The defect occurrence label of the segment-year sample is obtained by aggregating meter-by-meter defect labels within the same segment: if a target defect record exists at any valid meter-by-meter location within the same segment, then the segment-year sample is marked as a defect occurrence sample; if no target defect record exists at any valid meter-by-meter location within the same segment, then the segment-year sample is marked as a non-occurrence sample; if the proportion of valid meter-by-meter locations is lower than a preset proportion, then the segment-year sample is removed, downweighted, or marked as a low-confidence sample. This invention, through a flexible and adjustable segment length division and a stable segment numbering mapping mechanism, enables the same physical location to have a traceable fixed number in different years, taking into account both the tolerance of detection and positioning accuracy and the actual needs of engineering maintenance work units, ensuring the spatial consistency of cross-year samples, and at the same time eliminating low-confidence samples through effective detection ratio constraints, thus ensuring the reliability of training data.

[0011] The optimized historical features of this segment include one or more of the following: disease status in the previous year, the number of disease occurrences in the most recent K historical years, the proportion of disease occurrences, the number of consecutive years of occurrence, the time interval between the most recent disease occurrence and the year to be predicted, the number of historical recovery times, and the number of historical new occurrences. The spatial features of adjacent segments include one or more of the following: disease status of the left and right adjacent segments of the target segment in one or more recent years, the number of disease segments within a neighborhood window centered on the target segment with a radius of R, the proportion of disease occurrence within the neighborhood window, the length of the longest consecutive disease segment within the neighborhood window, and the historical persistence intensity of the neighborhood segments. By constructing the historical features of this segment and the spatial features of its neighborhood, this scheme enables the model to simultaneously capture the disease evolution patterns of the target segment itself and the spatial propagation effects of diseases in adjacent segments, enhancing the ability to characterize the continuous development, spatial diffusion, and local recovery processes of diseases, and providing a comprehensive and complementary spatiotemporal information foundation for risk prediction.

[0012] The optimized short-term and long-term time window features are respectively composed of the number of disease occurrences, the proportion of disease occurrences, the window mean, the difference between windows, the disease trend slope, and the annual state transition type within different historical windows. The annual state transition type includes transition from no disease to disease, from disease to no disease, continuous disease, and continuous no disease. The data quality features include the proportion of effective detection meters within a segment, missing detection markers, and data source reliability scores. This scheme, by constructing short-term and long-term time window features, trend slopes, and state transition types, enables the model to perceive the dynamic changes of diseases at different time scales, accurately identify four state transition modes: new occurrence, recovery, continuation, and absence of disease, and improves the comprehensive discrimination ability against short-term mutations and long-term evolution trends of diseases. Simultaneously, the introduction of data quality features allows for the deweighting of low-confidence samples during training, further enhancing the model's robustness.

[0013] The optimized category-weighted disease occurrence prediction model sets category weights based on the ratio of disease-occurring to non-occurring samples in the training samples. Under the rolling annual training method, it uses only historical year samples prior to the year to be predicted for training, validation, model selection, and threshold determination. The category-weighted disease occurrence prediction model can be any one of the following: logistic regression, gradient boosting tree, random forest, support vector machine, temporal convolutional network, or graph neural network. The model risk score is the disease occurrence probability, calibration probability, or normalized risk score output by the model. This invention effectively alleviates the model bias problem caused by the imbalance between disease-occurring and non-occurring samples through a category-weighted strategy and a rolling annual training mechanism, while strictly avoiding future data leakage and ensuring the accuracy and reliability of the prediction. It supports multiple base model replacements, giving the method flexible adaptability under different data scales and application scenarios.

[0014] The optimized persistent risk score is obtained by weighting the recent annual disease status of the target segment, the disease occurrence ratio of the target segment in the last K years, the continuous occurrence intensity of the target segment, the recent annual disease status of neighboring segments, and the historical continuous occurrence intensity of neighboring segments, and is normalized to between 0 and 1. The fused risk score is a weighted combination of the model risk score and the persistent risk score, with the weights determined by historical verification criteria, preset business rules, data quality weights, or rolling verification results. Here, the persistent risk score characterizes the historical continuity of the disease in the time dimension and is weighted and fused with the model risk score, so that the prediction results remain stable in years with continuous disease and remain sensitive in years with fluctuating characteristics. This achieves the complementary advantages of historical patterns and current characteristics, effectively suppressing score jumps caused by isolated noise.

[0015] In the optimized version of step S7, the recent disease quantity change rate is calculated from the number of diseased sections in the two most recent known years, and the calculation formula is Δ = |C(t-1) - C(t-2)| / max(C(t-2), 1), where C(t-1) and C(t-2) are the number of diseased sections in the two most recent known years t-1 and t-2, respectively, and t is the year to be predicted. When Δ reaches the preset drift threshold τ, it is determined that there is a significant data distribution drift in the most recent year, and a risk score with a higher response to multi-scale spatiotemporal feature changes is selected as the target risk score. When Δ does not reach the preset drift threshold τ, the risk score that performs best in historical verification is selected as the target risk score from the model risk score, the persistent risk score, and the fused risk score according to the historical verification criteria. When there are insufficient historical verification years, the model risk score or the preset fused score is used as the target risk score by default. This invention achieves automatic perception of annual data distribution drift by calculating the recent rate of change in the number of diseased sections and comparing it with the drift threshold. In stable years, it prioritizes the use of historical persistence scores to maintain forecast robustness, and in drifting years, it automatically switches to scores that are more sensitive to feature changes, thereby improving the adaptability and robustness of the cross-year forecasting strategy.

[0016] In an optimized manner, before generating prediction labels, spatial smoothing is performed on the target risk score, ensuring that the risk score of the target segment is simultaneously constrained by its own risk score and the risk scores of adjacent segments. This spatial smoothing includes neighborhood mean smoothing, weighted neighborhood smoothing, or smoothing based on continuous segment constraints. The prediction label is generated from a fixed threshold, a dynamic threshold determined by the proportion of historical defects, a Top-N threshold determined by the quantity of maintenance resources, or a combination thereof. The risk ranking result includes tunnel identification, route direction, segment number, segment start mileage, segment end mileage, year to be predicted, target risk score, prediction label, risk ranking, candidate score source identifier, drift judgment marker, and a summary of key features used to interpret the prediction. This invention, by performing spatial smoothing on the target risk score, ensures that the risk scores of adjacent segments mutually constrain each other, further suppressing isolated false alarm segments. It supports multiple label generation strategies, including fixed thresholds, dynamic thresholds, and Top-N, and the output multidimensional risk ranking list can directly serve practical engineering decision-making processes such as high-risk segment review, key monitoring deployment, and maintenance resource priority allocation.

[0017] This invention also introduces a highway tunnel defect occurrence prediction system, which employs the tunnel defect prediction method described above, including: The data access module is used to read annual tunnel inspection data; The data normalization module is used to complete field mapping, mileage direction unification, disease dictionary unification, missing measurement marking, and duplicate record merging; The segment aggregation module is used to generate segment-year samples and corresponding aggregation tags according to a preset segment length; The feature construction module is used to generate a multi-scale spatiotemporal feature matrix that depends only on historical years for each year to be predicted. The model training module is used to train a category-weighted disease occurrence prediction model; the candidate score calculation module is used to generate model risk scores, persistent risk scores, and fusion risk scores. The drift perception selection module is used to determine the target risk score based on the recent rate of change in the number of diseases, the preset drift threshold, and historical verification criteria. The threshold determination module is used to generate predicted labels based on fixed thresholds, dynamic thresholds, or Top-N strategies. It also includes a prediction output module, which generates a risk ranking list, stores prediction records, and provides a query interface to the maintenance business system.

[0018] This solution uses a modular design to decouple data access, standardization, segment aggregation, feature construction, model training, candidate score calculation, drift perception selection, threshold determination, and prediction output into independent functional modules. The modules interact with each other through standardized interfaces, supporting flexible deployment on stand-alone machines, servers, or cloud platforms. It is easy to integrate with existing tunnel detection and management systems and maintenance business platforms, and has a high degree of engineering implementation capability and promotion and application value.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. Compared with the traditional method of directly using meter-by-meter location as the prediction unit, this invention transforms the high-noise meter-by-meter detection records into continuous risk units with practical engineering significance through segmented modeling. This processing method effectively reduces the disturbance of detection and positioning errors, disease extension across meters, and isolated abnormal records on model training, so that the prediction results have good continuity and consistency in space. It can directly match the work units of on-site inspection and verification and maintenance construction organization, thus improving the engineering usability of the prediction results.

[0020] 2. At the feature expression level, this invention breaks through the limitations of single temporal features or single spatial features, and simultaneously constructs a multi-scale spatiotemporal feature system covering the historical evolution of diseases in this section, the spatial influence transmission of adjacent sections, short-term changes and fluctuations, long-term trends, state transition patterns and data quality markers. This feature set enhances the ability to characterize the local continuous development, spatial spread, sudden deterioration and recovery process of diseases in multiple dimensions, laying a solid information foundation for subsequent risk quantification.

[0021] 3. Regarding prediction strategies and output applications, this invention constructs three categories of candidate indicators: model risk score, persistent risk score, and fusion risk score. It also introduces a drift-aware adaptive selection mechanism, dynamically determining the target risk score for each segment based on the annual disease quantity change rate and historical verification criteria. This mechanism ensures that historical persistent information is fully utilized in years with consistently stable disease incidence, while in drift years caused by adjustments to detection strategies, implementation of remediation projects, or sudden environmental changes, it promptly adapts to the data-driven model's response to new trends. The final output includes multi-dimensional explanatory fields such as risk score, prediction label, risk ranking, score source identifier, drift marker, and summary of key features. This can directly serve practical production decision-making processes such as high-risk segment review, key monitoring deployment, annual maintenance plan preparation, and priority allocation of maintenance resources, demonstrating strong engineering practical value and promotional significance. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the highway tunnel defect prediction method based on drift sensing and multi-scale segment modeling of the present invention. Figure 2This is a schematic diagram illustrating the data normalization and year / mileage alignment process in this invention; Figure 3 This is a schematic diagram illustrating the segment division, tag aggregation, and segment-year sample generation in this invention; Figure 4 This is a schematic diagram of the construction of multi-scale spatiotemporal features in this invention; Figure 5 This is a schematic diagram illustrating the training and risk score output of the category-weighted model in this invention; Figure 6 This is a schematic diagram illustrating the calculation of persistent risk, fusion risk, and spatial smoothing in this invention; Figure 7 This is a schematic diagram of the drift-sensing adaptive selector decision process in this invention; Figure 8 This is a schematic diagram illustrating threshold determination, Top-N screening, and risk ranking output in this invention; Figure 9 This is a schematic diagram of the system module architecture and data flow in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figures 1-9 As shown, this invention provides a method for predicting highway tunnel defects based on drift perception and multi-scale segment modeling. This method transforms high-noise meter-by-meter detection data into segment-level risk units, integrates multi-scale spatiotemporal features and continuous quantization, and introduces a drift perception adaptive selection mechanism to achieve effective prediction and sortable output of tunnel defect risks for the next year.

[0025] First, step S1 is executed to obtain tunnel meter-by-meter inspection data generated annually. This meter-by-meter inspection data originates from Excel files, CSV files, database tables, or the API interface of the inspection platform. The data content includes at least the following fields: tunnel identifier, route direction, mileage location, inspection year, defect type, defect level, and defect start and end mileage. After data access, data standardization processing is performed: field names are uniformly mapped, mapping defect codes and names from different sources to a standard defect dictionary; mileage representation is standardized, converting raw mileage into a continuous meter index m based on route direction, tunnel start and end mileage, or preset direction rules; and the inspection year is standardized to a four-digit year format. For cases where multiple defect records exist for the same meter location and year, the system uses the maximum risk principle for merging, meaning that if even one valid defect record exists, the defect flag AnyDefect for that meter location is recorded as 1; alternatively, the highest level principle or data source priority can be used for merging. For missing meter positions or missing years, the system does not simply treat them as disease-free, but adds a missing marker field and removes, downweights, or encodes such samples separately in subsequent training to avoid label bias caused by missing data.

[0026] Execute step S2, dividing the continuous mileage into multiple segments according to the preset segment length L, and aggregating the meter-by-meter records into segment-year samples. Let m represent the continuous meter index under a unified direction, t represent the detection year, and m0 be the starting meter position under the current tunnel or line direction, then the segment number g is determined according to the following formula: g = floor((m - m0) / L) + 1 This definition ensures that the same physical location can be mapped to a stable segment number across different years, guaranteeing spatial consistency of samples across years. For year t and segment g, if any valid meter-by-meter position within the segment has AnyDefect=1, then the segment label Y(g,t)=1; if all valid meter-by-meter positions within the segment have AnyDefect=0, then Y(g,t)=0. When the original disease record is represented by start and end mileage and spans multiple segments, the system splits or projects the disease record onto all segments it covers, preventing cross-segment diseases from falling into a single meter position and causing label bias. If the proportion of valid meter-by-meter positions within a segment is lower than a preset threshold (e.g., lower than 50%), then the segment-year sample is marked as a low-confidence sample or an untrainable sample, and will be removed or have its weight reduced in subsequent model training. The segment length L can be adjusted according to the detection and positioning accuracy, disease continuity and maintenance scale: a 5-meter segment can be used when more precise positioning is required; a 20-meter or longer segment can be used when mainly used for macro-maintenance planning; in a preferred embodiment of the present invention, the segment length L is 10 meters to take into account the tolerance of positioning error, the expression of disease spatial continuity and the sufficiency of sample quantity.

[0027] Execute step S3 to construct a feature vector X(g,t) for the target segment g and the year t to be predicted. The feature vector includes the historical features H(g,t) of this segment, the neighborhood spatial features N(g,t), the time window and trend features T(g,t), and the data quality features Q(g,t), i.e.: X(g,t) = [ H(g,t), N(g,t), T(g,t), Q(g,t) ] The historical characteristics H(g,t) of this section include, but are not limited to: the previous year's state Y(g,t-1), the number of disease occurrences in the last K years, the proportion of disease occurrences in the last K years, the number of consecutive years of disease occurrence, the time interval between the most recent disease occurrence and the year to be predicted, the number of historical new occurrences, and the number of historical recovery times. In one embodiment of the present invention, the historical review period K is set to 5 years to balance the long-term trend of disease evolution with data availability.

[0028] The neighborhood spatial features N(g,t) include, but are not limited to: the previous year's states Y(g-1,t-1) and Y(g+1,t-1) of the left and right adjacent segments, the number of diseased segments within a neighborhood window of radius R, the proportion of disease occurrence in the neighborhood, the length of the longest consecutive diseased segment in the neighborhood, and the historical persistence intensity of the neighboring segments. In one embodiment of the present invention, the spatial neighborhood radius R is set to 5 segments to cover the spatial propagation effect of adjacent areas.

[0029] The time window and trend feature T(g,t) include the number of disease occurrences, the proportion of occurrences, the window mean, the difference between windows, the trend slope, and the state transition type within the short-term window W1 and the long-term window W2. The state transition types include four categories: new occurrence from 0 to 1, recovery from 1 to 0, continued presence of disease from 1 to 1, and continued absence of disease from 0 to 0, to characterize the dynamic evolution pattern of disease status in a segment over time. In one embodiment of the invention, the short-term window W1 is 3 years, and the long-term window W2 is 5 years.

[0030] The data quality feature Q(g,t) includes the proportion of effective detection meters within the segment, missing detection markers, and data source reliability scores, which are used to characterize the credibility of the samples and assist in adjusting sample weights during subsequent training. To avoid target leakage, when constructing samples for the year t to be predicted, the normalization process, feature generation, threshold determination, and candidate score selection do not use the true label of year t, ensuring the authenticity and reliability of the prediction.

[0031] Step S4 is executed to calculate the model risk score P_model using a category-weighted prediction model. There is often a class imbalance between disease occurrence samples and non-occurrence samples at the segment level. This invention sets class weights based on the ratio of positive to negative samples during model training, ensuring that minority disease occurrence samples receive higher weights in the training objective, thus preventing the model from biased towards the majority class and causing missed reports of disease segments. In one embodiment of this invention, logistic regression is used as the base model, and the model risk score is expressed as: P_model(g,t) = σ(w X(g,t) + b) Where σ is the Sigmoid function, and w and b are model parameters, estimated by maximizing the weighted log-likelihood function. The prediction model can also be replaced with gradient boosting trees (such as XGBoost, LightGBM), random forests, support vector machines, temporal convolutional networks, or graph neural networks, as long as its input is the aforementioned multi-scale spatiotemporal features and its output is a sortable model risk score, all of which fall under the implementation methods of this invention. In rolling annual training, if the target prediction year is t, the training samples only come from historical years less than t. For each historically verifiable year, the system regenerates the training set, validation set, and feature matrix according to the same rules to ensure that model selection, hyperparameter determination, and weight determination do not use future year labels, thus ensuring the authenticity and reliability of the prediction results.

[0032] Step S5 is executed to calculate the persistence risk score P_persist(g,t), which characterizes the historical continuity of the disease within the same segment and its neighborhood. The persistence risk score is calculated by weighting the disease status Y(g,t-1) of the target segment in the most recent year, the occurrence rate FreqK(g,t) in the most recent K years, the continuous occurrence intensity ContK(g,t), and the disease status NeighR(g,t-1) of the neighboring area in the most recent year, and then normalizing it to between 0 and 1. P_persist(g,t) = Norm[λ1Y(g,t-1) + λ2FreqK(g,t) + λ3ContK(g,t) + λ4NeighR(g,t-1)] λ1, λ2, λ3, and λ4 are the weight coefficients for each item, which satisfy the normalization constraint and can be determined based on business experience or historical verification results. The introduction of continuous risk scores ensures high temporal consistency of prediction results in consecutive years of disease occurrence, avoiding unreasonable jumps in model risk scores due to annual characteristic fluctuations.

[0033] The fusion risk score P_hybrid(g,t) is a weighted combination of the model risk score and the persistent risk score: P_hybrid(g,t) = α·P_model(g,t) + (1-α)·P_persist(g,t) Here, α is the fusion weight between 0 and 1, which can be determined through historical verification criteria, business rules, or rolling verification results. In a preferred embodiment of the present invention, α is set to 0.6 to balance the model's sensitivity to new trends with the stability of historical information.

[0034] To reduce the impact of isolated false positives on prediction results, this invention can also perform spatial smoothing on the target score or candidate score: P_smooth(g,t) = β·P(g,t) + (1-β)·mean{P(n,t) | n ∈ Ω(g)} Where β is the smoothing weight between 0 and 1, and Ω(g) is the set of neighboring segments of the target segment g. Spatial smoothing can be applied to the model risk score, persistent risk score, fused risk score, or final target risk score. In one embodiment of the present invention, the spatial smoothing weight β is set to 0.7, so that the risk score of the target segment simultaneously reflects its own characteristics and the average state of its neighborhood, effectively suppressing abnormal scores of isolated noise segments.

[0035] Execute step S6, and determine the target risk score P_target(g,t) for each segment using the drift-aware adaptive selector. The system statistically analyzes the number of diseased segments C(t-1) and C(t-2) for the two most recent known years t-1 and t-2, and calculates the recent rate of change in the number of diseased segments: Δ = |C(t-1) - C(t-2)| / max(C(t-2), 1) When Δ reaches the preset drift threshold τ, it is considered that there is a significant data distribution drift in the most recent year. In this case, the risk score with a higher response to changes in multi-scale spatiotemporal characteristics is preferentially selected as P_target. In one embodiment of the present invention, the risk score is the model risk score P_model or a candidate score mainly weighted by P_model, to ensure that in years where the data distribution undergoes abrupt changes, the prediction results can reflect the latest trend of changes in the feature space in a timely manner, rather than relying excessively on historical persistence information.

[0036] When Δ does not reach the drift threshold τ, the system selects a target risk score from P_model, P_persist, and P_hybrid based on historical validation criteria. Historical validation criteria can be set according to maintenance objectives, including but not limited to: prioritizing high-risk section review coverage, prioritizing reducing isolated false alarm rates, prioritizing maintaining stable annual prediction quantities, prioritizing improving the recall rate of consecutive diseased sections, or comprehensive evaluation based on data quality weights. Specifically, the system divides historical annual data into multiple validation periods in chronological order, calculates evaluation indicators for each candidate score strategy on the same validation set, and selects the candidate score that performs best in historical validation as the target risk score strategy for the current year. When there are insufficient historical validation years (e.g., less than 3 historical years), the system defaults to using P_model or a preset fusion score as the target risk score.

[0037] In one embodiment of the present invention, the drift threshold τ is set to 0.25, that is, when the number of diseased sections changes by more than 25% between the two most recent known years, the drift perception selection mechanism is triggered, and the prediction strategy is switched to model risk-based.

[0038] In step S7, for the year t to be predicted, the system outputs a target risk score P_target(g,t) for each segment and generates prediction labels based on a fixed threshold, a dynamic threshold determined by the historical positive rate, a Top-N threshold determined by the number of maintenance resources, or a combination thereof. The fixed threshold is suitable for scenarios where the baseline rate of disease occurrence is stable; the dynamic threshold is determined based on the positive rate ranked by the target risk score in historical verification years, ensuring that the predicted positive number matches the actual historical distribution; the Top-N threshold determines the number of high-risk segments to be extracted based on the actual constraints of maintenance resources (such as the upper limit of the number of verifiable segments per year). In a preferred embodiment of the invention, the system simultaneously employs a combination strategy of dynamic threshold and Top-N threshold: first, a baseline threshold is determined based on the historical positive rate, and then the predicted positive segments are truncated based on the upper limit of maintenance resources, ensuring recall while controlling the output number to meet actual engineering constraints.

[0039] The predicted output is presented in the form of tables, database records, interface responses, or visual bars. Each output record includes at least the tunnel identifier, route direction, section number, section start mileage, section end mileage, predicted year, target risk score, prediction label, risk ranking, target score source identifier, drift judgment marker, and a summary of key features. Maintenance units can arrange verification inspections, key monitoring, or preventative maintenance operations for high-risk sections based on the risk ranking, achieving precise allocation of maintenance resources.

[0040] The present invention also provides a highway tunnel defect occurrence prediction system, which includes a data access module, a data normalization module, a section aggregation module, a feature construction module, a model training module, a candidate score calculation module, a drift perception selection module, a threshold determination module, and a prediction output module. The data access module reads annual inspection data; the data normalization module completes field mapping, mileage direction unification, disease dictionary unification, missing measurement marking, and duplicate record merging; the segment aggregation module generates segment-year samples and corresponding aggregation labels according to segment length L; the feature construction module generates a feature matrix that depends only on historical years for each year to be predicted; the model training module trains a category-weighted prediction model; the candidate score calculation module generates the model risk score P_model, persistent risk score P_persist, and fusion risk score P_hybrid; the drift perception selection module determines the target risk score P_target based on the rate of change Δ, drift threshold τ, and historical verification criteria; the threshold determination module generates prediction labels based on fixed thresholds, dynamic thresholds, or Top-N strategies; and the prediction output module generates a risk ranking list, stores prediction records, and provides a query interface to the maintenance business system.

[0041] The above system can be deployed on a standalone machine, server, or cloud platform. It can also run as a prediction submodule of a tunnel inspection data management system, structural health monitoring system, or maintenance management platform. Data can be transferred between modules through memory objects, files, database tables, or service interfaces. It is not limited to a specific programming language, specific database, or specific hardware platform.

[0042] The above system parameters can be flexibly adjusted according to the actual engineering situation: the section length L can be adjusted according to the detection and positioning accuracy, the continuity of defects, and the scale of maintenance operations; a 10-meter section is preferred in the preferred embodiment; the number of historical review years K can be set to 3 to 10 years, preferably 5 years; the spatial neighborhood radius R can be set to 1 to 10 sections, preferably 5 sections; the short-term window W1 can be set to 2 to 4 years, preferably 3 years; the long-term window W2 can be set to 4 to 8 years, preferably 5 years; the drift threshold τ can be set to 0.15 to 0.40, preferably 0.25; the fusion weight α can be set to 0.3 to 0.8, preferably 0.6; and the smoothing weight β can be set to 0.5 to 0.9, preferably 0.7. All of the above parameters can be adaptively adjusted according to the tunnel type, detection cycle, defect type, historical defect fluctuation range, and maintenance resource constraints to obtain optimal predictive performance and engineering applicability in different application scenarios.

[0043] In summary, this invention provides a method and system for predicting defects at the section level based on annual inspection data of highway tunnels. By aggregating high-noise meter-by-meter data into engineering-scale risk units, and integrating multi-scale spatiotemporal characteristics such as the section's history, neighborhood space, multiple time windows, and data quality, a dual evaluation system combining a category-weighted prediction model and continuous risk quantification is constructed. Furthermore, a drift-aware adaptive selection mechanism is introduced, enabling the prediction strategy to automatically and flexibly switch between model-driven and historical continuous-driven approaches based on annual changes in the number of defects. This effectively overcomes core technical challenges such as meter-by-meter positioning errors, insufficient adaptability of a single model, and annual data drift. The final output is a multi-dimensional interpretive list including risk ranking, score source identification, and drift judgment markers, directly serving practical decision-making processes such as high-risk section review, key monitoring deployment, and maintenance resource allocation. This invention has high engineering practical value and promotional significance, providing reliable technical support for the accurate prediction and proactive maintenance of highway tunnel defects.

[0044] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions made to the technical solutions of the present invention cannot depart from the spirit and scope of the present invention and should be covered within the scope of the claims of the present invention.

Claims

1. A tunnel defect prediction method based on drift sensing and multi-scale segment modeling, characterized in that, Used to predict the risk of section-level defects in the following year based on historical tunnel inspection data, including: S1, acquire meter-by-meter tunnel inspection data for at least two years, the meter-by-meter tunnel inspection data including tunnel identification, route direction, mileage location, inspection year, and defect attribute fields; S2, standardize the fields, align the years, unify the mileage and direction, mark missing measurements and merge duplicate records in the meter-by-meter tunnel detection data to obtain standardized meter-by-meter data; S3, divide the continuous mileage into multiple segments according to the preset segment length, and aggregate the standardized meter-by-meter data into segment-year samples; S4. Construct multi-scale spatiotemporal features based on the segment-year samples. The multi-scale spatiotemporal features include historical features of the current segment, spatial features of adjacent segments, short-term time window features, long-term time window features, state transition features, and data quality features. S5, use the multi-scale spatiotemporal features to train a category-weighted disease occurrence prediction model, and obtain the model risk score of the section to be predicted; S6. Calculate the persistence risk score based on the historical persistence status of the target section and the historical persistence status of the neighboring sections, and obtain the fused risk score based on the model risk score and the persistence risk score. S7. Based on the recent rate of change in the number of diseases and historical verification criteria, perform drift-aware adaptive selection among the model risk score, the persistent risk score, and the fused risk score to obtain the target risk score. S8. Generate the next year's segment-level disease occurrence prediction label and risk ranking results based on the target risk score.

2. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, Step S2 includes: unifying the mileage expression in different detection tables into a continuous meter index in the same direction; mapping different disease names or disease codes into a unified disease dictionary; merging duplicate records of the same meter, the same year, and the same disease according to the principle of maximum risk, the principle of highest level, or a preset source priority; and marking missing meter positions, invalid detection records, and low-confidence records as untrainable samples, low-confidence samples, or samples with data quality weights.

3. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, In step S3, the preset section length is 5 meters, 10 meters, 20 meters, or other lengths determined based on the detection and positioning accuracy, the spatial continuity of the disease, and the scale of maintenance operations. Each segment is identified by a tunnel marker, route direction, segment number, segment start mileage, and segment end mileage, ensuring that the same physical segment has a stable number across different years. The defect occurrence label of the segment-year sample is obtained by aggregating the meter-by-meter defect labels within the same segment: if a target defect record exists at any valid meter-by-meter location within the same segment, then the segment-year sample is marked as a defect occurrence sample; if no target defect record exists at any valid meter-by-meter location within the same segment, then the segment-year sample is marked as a non-occurrence sample; if the proportion of valid meter-by-meter locations is lower than a preset proportion, then the segment-year sample is removed, downweighted, or marked as a low-confidence sample.

4. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, The historical characteristics of this segment include one or more of the following: disease status in the previous year, number of disease occurrences in the most recent K historical years, disease occurrence ratio, number of consecutive occurrence years, time interval between the most recent disease occurrence and the year to be predicted, historical recovery count, and historical new occurrence count. The spatial characteristics of adjacent segments include one or more of the following: disease status of the left and right adjacent segments of the target segment in the most recent one or more years, number of disease segments within a neighborhood window centered on the target segment with a radius of R, disease occurrence ratio within the neighborhood window, length of the longest consecutive disease segment within the neighborhood window, and historical persistence intensity of the neighborhood segment.

5. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, The short-term and long-term time window features are respectively composed of the number of disease occurrences, the proportion of disease occurrences, the window mean, the difference between windows, the slope of disease trend, and the annual state transition type within different historical windows. The annual state transition type includes the transition from no disease to disease, the transition from disease to no disease, the continuous presence of disease, and the continuous absence of disease. The data quality features include the proportion of effective detection meters within the segment, missing measurement markers, and data source reliability scores.

6. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, The category-weighted disease occurrence prediction model sets category weights based on the ratio of disease occurrence samples to non-disease occurrence samples in the training samples, and uses only historical year samples before the year to be predicted for training, validation, model selection, and threshold determination under the rolling annual training method; the category-weighted disease occurrence prediction model is any one of the following: logistic regression model, gradient boosting tree model, random forest model, support vector machine model, temporal convolutional network model, or graph neural network model; the model risk score is the disease occurrence probability, calibration probability, or normalized risk score output by the model.

7. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, The persistent risk score is obtained by weighting the recent annual disease status of the target segment, the disease occurrence ratio of the target segment in the most recent K years, the continuous occurrence intensity of the target segment, the recent annual disease status of neighboring segments, and the historical continuous occurrence intensity of neighboring segments, and is normalized to between 0 and 1; the fused risk score is a weighted combination of the model risk score and the persistent risk score, and the weights are determined by historical verification criteria, preset business rules, data quality weights, or rolling verification results.

8. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, In step S7, the recent disease quantity change rate is calculated from the number of diseased sections in the two most recent known years, and the calculation formula is Δ = |C(t-1) - C(t-2)| / max(C(t-2), 1), where C(t-1) and C(t-2) are the number of diseased sections in the two most recent known years t-1 and t-2, respectively, and t is the year to be predicted; when Δ reaches the preset drift threshold τ, it is determined that there is a significant data distribution drift in the most recent year, and a risk score with a higher response to changes in multi-scale spatiotemporal characteristics is selected as the target risk score; When Δ does not reach the preset drift threshold τ, the risk score that performs best in historical verification is selected as the target risk score from the model risk score, the persistent risk score and the fusion risk score according to the historical verification criteria. When there are insufficient historical verification years, the model risk score or the preset fusion score is used as the target risk score by default.

9. The tunnel defect prediction method based on drift sensing and multi-scale segment modeling according to claim 1, characterized in that, Before generating the prediction label, the target risk score is spatially smoothed so that the risk score of the target segment is simultaneously constrained by its own risk score and the risk scores of adjacent segments. The spatial smoothing includes neighborhood mean smoothing, weighted neighborhood smoothing, or smoothing based on continuous segment constraints. The prediction label is generated by a fixed threshold, a dynamic threshold determined by the proportion of historical defects, a Top-N threshold determined by the amount of maintenance resources, or a combination thereof. The risk ranking result includes tunnel identification, route direction, segment number, segment start mileage, segment end mileage, year to be predicted, target risk score, prediction label, risk ranking, candidate score source identifier, drift judgment mark, and a summary of main features used to interpret the prediction.

10. A highway tunnel defect prediction system, employing the tunnel defect prediction method according to any one of claims 1-9, characterized in that, include: The data access module is used to read annual tunnel inspection data; The data normalization module is used to complete field mapping, mileage direction unification, disease dictionary unification, missing measurement marking, and duplicate record merging; The segment aggregation module is used to generate segment-year samples and corresponding aggregation tags according to a preset segment length; The feature construction module is used to generate a multi-scale spatiotemporal feature matrix that depends only on historical years for each year to be predicted. The model training module is used to train a category-weighted disease occurrence prediction model; The candidate score calculation module is used to generate model risk scores, persistent risk scores, and fusion risk scores. The drift perception selection module is used to determine the target risk score based on the recent rate of change in the number of diseases, the preset drift threshold, and historical verification criteria. The threshold determination module is used to generate predicted labels based on fixed thresholds, dynamic thresholds, or Top-N strategies. It also includes a prediction output module, which generates a risk ranking list, stores prediction records, and provides a query interface to the maintenance business system.

Images