Tunnel structure long-term stability evaluation method based on intelligent detection and monitoring data

Abstract

The application discloses a tunnel structure long-term stability evaluation method based on intelligent detection and monitoring data, relates to the technical field of tunnel engineering monitoring and evaluation, and comprises the following steps: S1, tunnel monitoring data is extracted from a preset database, the data set after classification is processed by adopting a branch condition setting and feature selection mechanism, and a mapped user demand set is obtained; S2, the demand difference is analyzed by adopting a node splitting rule according to the mapped user demand set, and if the demand difference points to a management personnel, a safety level quantitative index is extracted, and preliminary deconstruction information is judged by information gain calculation; the tunnel structure long-term stability evaluation method based on intelligent detection and monitoring data not only improves the accuracy of information expression and the efficiency of report generation, but also establishes a closed-loop information chain from original data to a final report, and effectively guarantees the reliability and scientificity of tunnel structure long-term stability evaluation.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering monitoring and evaluation technology, specifically to a method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data. Background Technology

[0002] Tunnel engineering, as an important component of infrastructure construction, is widely used in transportation, water conservancy, energy, and other fields. Its long-term stability is directly related to public safety and economic benefits. During long-term operation, tunnels are affected by geological conditions, environmental changes, and structural aging, making stability assessment a crucial step in ensuring project safety.

[0003] However, current evaluation systems are significantly inadequate in meeting the needs of different user groups, making it difficult to effectively support diverse application scenarios at the management, technical, and decision-making levels. This makes the efficient and accurate delivery of targeted information a pressing issue in the tunnel engineering field. Existing methods typically generate generic evaluation reports based on unified monitoring data, but neglect the differences in information needs among different user groups. Managers require intuitive safety levels and early warning prompts to quickly respond to potential risks; technicians rely on detailed monitoring data and analysis results for in-depth research and problem identification; and decision-makers prioritize concise comprehensive reports and maintenance recommendations for strategic planning. Current solutions often struggle to flexibly adapt to these differentiated needs based on the same set of data, leading to inefficient information delivery and even affecting the timeliness of critical decisions. The core technical challenge lies in how to achieve customized information expression based on the same set of monitoring data. First, the flexibility of data deconstruction and reconstruction is insufficient. Monitoring data typically contains multi-dimensional information such as time series and mechanical parameters, but existing systems struggle to dynamically extract and reconstruct relevant data according to user roles. For example, managers might only need security levels derived from data, while technicians require complete mechanical calculations; current technology struggles to efficiently achieve this layered output. Secondly, due to limitations in data deconstruction, report generation systems face bottlenecks in output format diversity. Different user groups have significantly different needs for report formats; for instance, managers prefer intuitive graphical presentations, while decision-makers require concise text summaries, and existing systems struggle to support seamless switching between multiple formats within a single framework. Summary of the Invention

[0004] The purpose of this invention is to provide a long-term stability evaluation method for tunnel structures based on intelligent detection and monitoring data, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a long-term stability evaluation method for tunnel structures based on intelligent detection and monitoring data, comprising: S1, extracting tunnel monitoring data from a preset database, performing preliminary classification based on user roles, and processing the classified data set using branch condition setting and feature selection mechanisms to obtain a mapped user demand set; S2, analyzing demand differences based on the mapped user demand set using node splitting rules, and extracting safety level quantitative indicators if the demand differences point to management personnel, and determining preliminary structural information through information gain calculation; S3, obtaining preliminary structural information, grouping mechanical parameters through leaf node output, and if the grouping results show that there are parameter groups with risks exceeding a preset threshold, reorganizing them into a graphical expression of the impact of water pressure fluctuations, and determining the reorganization. S4. From the reorganized visualized data, obtain complete time series details according to the needs of technical personnel. If the details exceed the preset threshold, add analysis annotations and obtain enhanced technical output through tree depth control. S5. Based on the enhanced technical output, branch to the decision-making level path using pruning methods. Quantify the security level by maintaining the suggestion list and comparing risk thresholds to obtain a concise text summary. S6. From the concise text summary, obtain the output set of all user paths. If the output format needs to be switched, dynamically convert it to the specified format. Determine the final customized report by prioritizing suggestions. S7. Validate the compatibility of multiple formats through the final customized report. If the validation is successful, distribute it to the corresponding user group. Obtain the complete information transmission chain through text aggregation logic and output simplification processing.

[0006] Preferably, step S1 includes: acquiring tunnel monitoring data from a preset database, including time series and mechanical parameters; extracting the complete dataset using a standardized protocol to obtain the original monitoring dataset; classifying the original monitoring dataset according to user roles: if the user role is an administrator, all time series and mechanical parameters are extracted; if the user role is a technician, a subset under specific geological conditions is extracted to obtain the classified dataset; for the classified dataset, a branching condition mechanism is used: if the completeness of the time series is higher than a preset threshold, the complete data is retained; if it is lower than the threshold, missing data is removed to obtain the filtered dataset; processing the filtered dataset through a feature selection mechanism, using principal component analysis to extract key features from the time series and mechanical parameters to obtain the feature-extracted dataset; based on the feature-extracted dataset, using K-means clustering to group the data and determine the user demand mapping relationship for each group to obtain the initial mapping set; optimizing the initial mapping set using a decision tree algorithm: if the accuracy of the mapping set is higher than a preset threshold, the mapping relationship is retained; if it is lower than the threshold, the feature weights are adjusted to obtain the optimized mapping set; and generating the final user demand set using the optimized mapping set to determine the tunnel monitoring data mapping result that satisfies the user role.

[0007] Preferably, step S2 includes: obtaining demand difference data from the user demand set; classifying the difference data using node splitting rules; if the difference data points to management personnel, extracting security level quantitative indicators; determining preliminary deconstruction information through information gain calculation to obtain a deconstruction information set; grouping the security level quantitative indicators using a random forest algorithm for the deconstruction information set; retaining the grouped data if the accuracy of the grouping result is higher than a preset threshold, and removing abnormal data if it is lower than the preset threshold, to obtain a grouped data set; obtaining security level quantitative indicators from the grouped data set; formatting the indicator data using a standardization protocol to obtain a formatted indicator set; extracting key features of the security level quantitative indicators from the formatted indicator set using a principal component analysis algorithm to obtain a feature-extracted data set; obtaining key feature data from the feature-extracted data set; grouping the feature data using cluster analysis; determining the priority of each group of features to obtain a priority grouping set; generating security level assessment results from the priority grouping set if the group priority is higher than a preset threshold, using weighted calculation to obtain an assessment result set; obtaining the security level assessment results from the assessment result set; generating a final security level recommendation set using mapping rules to determine the security level mapping result that meets the needs of management personnel.

[0008] Preferably, step S3 includes obtaining preliminary deconstruction information from the user demand set, removing noisy data through data cleaning methods to obtain a cleaned deconstruction information set; classifying the cleaned deconstruction information set using a decision tree algorithm, generating mechanical parameter groups based on leaf node outputs to obtain a mechanical parameter group set; if there are parameter groups in the mechanical parameter group set whose risk exceeds a preset threshold, then further subdividing the parameter groups whose risk exceeds the preset threshold through cluster analysis to obtain a subdivided risk data set; for the subdivided risk data set, extracting water pressure fluctuation data using a data recombination method to obtain a water pressure fluctuation data set; obtaining key fluctuation features from the water pressure fluctuation data set, generating formatted fluctuation data through standardization processing to obtain a formatted fluctuation data set; for the formatted fluctuation data set, generating graphical data on water pressure fluctuations using a visualization transformation method to obtain a visualized data set; if the visualized data set meets preset risk assessment standards, then determining the final safety level mapping result through weighted calculation to obtain a safety level mapping set.

[0009] Preferably, step S4 includes: extracting local time segments from the time series data using a sliding window method to obtain a time segment set; decomposing the frequency components of the time segment set using a Fourier transform method to obtain a frequency component set; if the proportion of frequency components in the frequency component set whose frequencies exceed a set frequency threshold exceeds a preset threshold, removing noise from the frequency components whose frequencies exceed the set frequency threshold using a filtering method to obtain a filtered frequency set; generating a smoothed time series based on the filtered frequency set using a time series reconstruction method to obtain a smoothed time series set; extracting trend change features from the smoothed time series set to obtain a trend feature set; determining the trend stability of the trend feature set using a classification method to obtain a stability classification set; and generating output data with technical annotations based on the stability classification set to obtain the final technical output set.

[0010] Preferably, step S5 includes: processing redundant branches from the enhanced technology output using a decision tree pruning method to obtain a simplified path set; grouping similar paths into a path grouping set using a clustering analysis method; if the complexity of a path group exceeds a preset threshold, simplifying the path structure using a dimensionality reduction method to obtain a simplified path set; constructing a suggestion list based on the simplified path set, quantifying the security level of each path by comparing it with a preset risk threshold to obtain a security level set; organizing the level descriptions using a text generation method to obtain a structured text set; extracting key level information from the structured text set, generating a concise text summary, and obtaining the final text output; and associating the final text output with business scenarios using a mapping method to obtain a scenario adaptation description.

[0011] Preferably, step S6 includes obtaining all path data from the output set of user paths, grouping the path data using a clustering analysis method to obtain a path group set; for the path group set, if the format of a certain group of paths does not match the target format, adjusting the path data according to a preset format conversion rule to obtain a path set with consistent format; based on the path set with consistent format, extracting the suggested priority attribute of each path, and arranging them from high to low priority using a sorting algorithm to obtain a sorted path set.

[0012] Preferably, step S6 further includes obtaining the highest priority path data from the sorted path set, associating it with a preset business template through a mapping method to generate customized output data; for the customized output data, if the data volume exceeds a preset threshold, simplifying the data structure through a dimensionality reduction method to obtain a simplified output set; organizing the output content according to the simplified output set using a text generation method to obtain the final output; extracting key information fields from the final output, saving them through a structured storage method to obtain a persistent output result.

[0013] Preferably, step S7 includes obtaining format data from the final customized report, extracting the structural features of each format using a format parsing method to obtain a structured format dataset; if the structured format dataset does not match the preset format standard, adjusting the data through format conversion rules to obtain a standardized format dataset; and for the standardized format dataset, using a consistency verification algorithm to check the data integrity and standardization to obtain a set of verified formats.

[0014] Preferably, step S7 further includes matching group attributes according to the verified format set using a user group analysis method to generate a distribution mapping table corresponding to the group; obtaining distribution data from the distribution mapping table, integrating similar information using a content clustering method to obtain an aggregated information set; filtering information with a priority higher than a preset threshold using a keyword extraction method for the aggregated information set to obtain a simplified information set; extracting information chains from the simplified information set, organizing and outputting them using a structured coding method to obtain the final information transmission chain.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This method for evaluating the long-term stability of tunnel structures based on intelligent detection and monitoring data achieves the mapping of demand sets by extracting monitoring data from a pre-set database and combining it with user role classification, branch condition setting, and feature selection. It employs node splitting rules and information gain calculation to output quantitative indicators of safety levels and generate preliminary deconstruction information for management personnel. By grouping mechanical parameters through leaf nodes, when the parameter grouping results exceed a pre-set threshold, they are reorganized into graphical data representing the impact of water pressure fluctuations, meeting the visualization needs of technical personnel. Annotations are added to the time series analysis to form enhanced technical output. Pruning and risk threshold comparison are introduced into the decision-making path to obtain maintenance suggestions and concise text summaries. Finally, through dynamic format conversion, priority sorting, and verification mechanisms, customized reports adaptable to multiple formats are generated, and a complete and efficient information transmission chain is formed through text aggregation and result summarization. Compared with existing technologies, this invention can achieve differentiated information customization output based on different user roles, enabling efficient utilization of tunnel monitoring data at the management, technical and decision-making levels. It not only improves the accuracy of information expression and the efficiency of report generation, but also establishes a closed-loop information chain from raw data to the final report, effectively ensuring the reliability and scientific nature of the long-term stability evaluation of tunnel structures. Attached Figure Description

[0016] Figure 1 This is a flowchart of the long-term stability evaluation method for tunnel structures based on intelligent detection and monitoring data, as presented in this invention. Detailed Implementation

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

[0018] like Figure 1 As shown, this invention provides a technical solution: a long-term stability evaluation method for tunnel structures based on intelligent detection and monitoring data, characterized by comprising: S1, extracting tunnel monitoring data from a preset database, performing preliminary classification based on user roles, and processing the classified data set using branch condition setting and feature selection mechanisms to obtain a mapped user demand set; S2, analyzing demand differences based on the mapped user demand set using node splitting rules, and extracting safety level quantitative indicators if the demand differences point to management personnel, and determining preliminary deconstruction information through information gain calculation; S3, obtaining preliminary deconstruction information, grouping mechanical parameters through leaf node output, and if the grouping results show that there are parameter groups with risks exceeding a preset threshold, reorganizing them into a graphical representation of the impact of water pressure fluctuations, and determining the reorganization... S4. From the reorganized visualized data, obtain complete time series details according to the needs of technical personnel. If the details exceed the preset threshold, add analysis annotations and obtain enhanced technical output through tree depth control. S5. Based on the enhanced technical output, branch to the decision-making level path using pruning methods. Quantify the safety level by maintaining the suggestion list and comparing risk thresholds to obtain a concise text summary. S6. From the concise text summary, obtain the output set of all user paths. If the output format needs to be switched, dynamically convert it to the specified format. Determine the final customized report by prioritizing suggestions. S7. Validate the compatibility of multiple formats through the final customized report. If the validation is successful, distribute it to the corresponding user group. Obtain the complete information transmission chain through text aggregation logic and output simplification processing.

[0019] This implementation method constructs a logical framework based on intelligent detection and monitoring data processing to conduct multi-level, multi-role assessments of the long-term stability of tunnel structures. First, in step S1, tunnel monitoring data, including structural stress, displacement, and water pressure, is extracted from a preset database and initially categorized according to user roles (e.g., managers, technicians). By setting branching conditions and selecting execution features, a more targeted data input set is formed. Then, in step S2, a node splitting algorithm is used to analyze and judge demand differences. When relevant demands from managers are identified, the system automatically extracts the corresponding safety level quantification indicators and constructs preliminary deconstruction information using the information gain criterion. In step S3, this deconstruction information is used to cluster key mechanical parameters. If the parameters of a certain group exceed the safety threshold, the system converts them into a chart representing the impact of water pressure fluctuations, thereby improving data interpretability. In step S4, to meet the needs of technicians, the system further extracts the complete time series and controls the data depth, generating an annotated enhanced output. In step S5, the decision path is optimized through pruning and compared based on maintenance recommendations and risk thresholds, with the output being a structured, concise text summary. Subsequently, in step S6, the path outputs of different users are managed collectively, dynamically converted according to actual format requirements, and a final customized analysis report is generated through a suggestion priority algorithm. Finally, in step S7, the report format is verified for compatibility with multiple terminals. After verification, it is distributed to users of different roles, and a complete data transmission chain is constructed using text aggregation and output optimization algorithms.

[0020] This invention achieves a systematic and intelligent evaluation of the long-term stability of tunnel structures by combining multi-step intelligent data processing and user-oriented output strategies. Its advantages are mainly reflected in the following aspects: First, feature selection and classification mechanisms improve the targeting and efficiency of data processing; second, node splitting and information gain methods ensure the scientific validity and adaptability of the deconstructed information; third, parameter risk identification and chart visualization enhance the responsiveness to structural safety risks; furthermore, techniques such as pruning and priority ranking optimize the structure and content of the generated report, making the output more readable and providing decision support value; finally, multi-format verification and automatic distribution mechanisms ensure the integrity and timeliness of information transmission, greatly improving information sharing and management efficiency.

[0021] S1 includes: acquiring tunnel monitoring data from a pre-set database, including time series and mechanical parameters; extracting the complete dataset using a standardized protocol to obtain the original monitoring dataset; classifying the original monitoring dataset according to user roles: if the user role is administrator, all time series and mechanical parameters are extracted; if the user role is technician, a subset under specific geological conditions is extracted to obtain the classified dataset; for the classified dataset, a branching condition mechanism is used: if the time series completeness is higher than a pre-set threshold, complete data is retained; if it is lower than the threshold, missing data is removed to obtain the filtered dataset; processing the filtered dataset through a feature selection mechanism, using principal component analysis to extract key features from the time series and mechanical parameters to obtain the feature-extracted dataset; based on the feature-extracted dataset, using K-means clustering to group the data and determine the user demand mapping relationship for each group to obtain the initial mapping set; optimizing the initial mapping set using a decision tree algorithm: if the accuracy of the mapping set is higher than a pre-set threshold, the mapping relationship is retained; if it is lower than the threshold, the feature weights are adjusted to obtain the optimized mapping set; and generating the final user demand set using the optimized mapping set to determine the tunnel monitoring data mapping results that satisfy the user roles.

[0022] First, tunnel monitoring data is read from a pre-set database using a standardized protocol. This protocol includes strict field name comparison, unified time format (year, month, day, hour, minute, second), unified units (internationally accepted units), fixed anomaly marking rules (missing data marked with a negative one), and direct removal of non-compliant records. This yields the original monitoring dataset. The time series refers to a continuous set of time points obtained at a fixed sampling period. In this implementation, the sampling period is fixed at once every hour, and any missing data is directly marked by the standardized protocol without interpolation. Mechanical parameters refer to quantitative data that characterizes the tunnel's structural state. In this implementation, these include circumferential stress and axial stress. Stress, shear stress, radial displacement, axial displacement, convergence, pore water pressure, and surrounding rock pressure are all stored as numerical values ​​with units in the database. After reading, they are uniformly converted to a single unit system. After reading, the data is categorized according to the user's role. If the user is an administrator, the system directly selects all time series data and all mechanical parameters to form the administrator-categorized dataset. If the user is a technician, the system filters the data of monitoring points corresponding to specified geological conditions according to the geological condition field already labeled in the database to form the technician-categorized dataset. The geological condition field consists of deterministic enumerated values ​​and was written during the data acquisition phase. The system does not make inferences; after classification, it enters the branching condition mechanism to determine the integrity of the time series. The calculation method is to calculate the ratio of the number of samples to be sampled to the number of samples actually sampled within the selected statistical period and multiply it by one hundred to obtain the integrity percentage. In this implementation, the integrity threshold is fixed at 90. This threshold is given by the monitoring system design specification to ensure the continuity of the series and is a fixed value that does not change with the data. When the integrity of a certain monitoring point is not less than 90% within the statistical period, all time series and corresponding mechanical parameters of the monitoring point are retained. When the integrity is less than 90%, all records of the monitoring point in that period are removed, resulting in the filtered dataset. Subsequently, feature selection was performed on the filtered dataset, and principal component analysis was used only. First, each mechanical parameter and time series derived index were linearly stretched to 0 to 1. Specifically, the minimum value of a parameter in the sample was mapped to 0, the maximum value was mapped to 1, and the remaining values ​​were proportionally mapped to the range of 0 to 1. The time series derived indexes include the single-point change calculated by the sampling period, the moving average calculated by the 24-hour window, the moving average calculated by the 7-day window, and the intraday fluctuation range calculated by the 24-hour window. All of these derived indexes are calculated with a fixed window length and are not adaptively adjusted.Subsequently, the collaborative change relationship among multiple parameters is calculated to form a variance collaborative matrix. Numerical decomposition is then performed to obtain a set of unrelated new variable directions, which are sorted from high to low according to variance contribution. In this implementation, the cumulative variance contribution rate threshold is fixed at 90. When the cumulative contribution rate of several new variables reaches 90 or above, these new variables are retained as key features, and the remaining new variables are discarded. If the cumulative contribution rate is insufficient, it continues to increase sequentially until it reaches or exceeds 90, thus obtaining the feature-extracted dataset. Simultaneously, the contribution ratio of each original parameter in the key features is recorded as the initial feature weight. In the grouping stage, the mean clustering algorithm is used, and the number of clusters is fixed at 3 to correspond to the low-risk, medium-risk, and high-risk categories. The initial center is determined as follows: After sorting all samples from smallest to largest based on the value of the first key feature, the key feature vectors of the first, median, and last samples are selected as the three initial centers. The distance metric is a weighted Euclidean distance calculation method using each key feature within the interval of 0 to 1. The weights are taken from the aforementioned initial feature weights. The calculation process is as follows: for each key feature, the square of the difference between the sample value and the center value is multiplied by the feature weight, then the sum is taken and the square root is taken to obtain the distance value. After completing one full distance calculation, each sample is assigned to the cluster containing the center with the smallest distance. After classification, the arithmetic mean of all samples in each cluster is calculated dimension by dimension to obtain the new center. Then, the distance calculation and reclassification process is repeated until the maximum change in the displacement of all centers is less than 0.01 or 1 / 2. When the iteration reaches 100, the iteration stops and the final three clusters are output if either condition is met. In the user demand mapping stage, the system uses the data range obtained by the user role in the aforementioned classification stage as a constraint, and establishes a mapping relationship based on the sample source region and geological condition statistical results within each cluster. Specifically, the system calculates the proportion of samples from the administrator-classified dataset and the proportion of samples with specified geological conditions from the technician-classified dataset for each cluster. When the proportion of samples with specified geological conditions in a cluster is not less than 70%, the cluster is marked as a technician demand mapping; when the proportion is less than 70%, the cluster is marked as an administrator demand mapping. This yields the initial mapping set. To optimize the consistency and decidability of the mapping set, the system... A decision tree is constructed with key features as input and initial mapping labels as output. During the tree construction process, the difference between the uncertainty before splitting and the weighted uncertainty after splitting according to a certain feature and threshold is calculated as the information gain at each split node. The feature and threshold with the largest information gain are selected as the splitting condition for the current node. After splitting, the process is repeated for each child node. The stopping condition is that the consistency of sample labels in any leaf node is not less than 95, or the tree depth reaches 6, or the number of samples in the leaf node is not less than 30 and the information gain increases by less than 1 after further splitting. Growth stops when any of the conditions are met. After the tree is constructed, all samples are input into the tree to obtain the predicted mapping and are compared one by one with the initial mapping. The matching ratio is calculated as the mapping accuracy.This implementation fixes the mapping accuracy threshold at 90. When the matching ratio is not less than 90%, the current mapping relationship is retained as the optimized mapping set. When the matching ratio is less than 90%, feature weight adjustment is performed. The adjustment method is to count the splitting feature that appears most frequently in the whole tree and is located near the root node and increase its weight by 10. At the same time, the weight of the splitting feature that appears less frequently and is located deep in the tree is reduced by 10. Then, all weights are scaled proportionally so that the sum of the weights equals 1. The process of mean clustering grouping and decision tree optimization is returned until the mapping accuracy reaches not less than 90 or the maximum number of repetitions is reached (5 times), whichever comes first. Finally, the optimized mapping set generates the final user demand set. The generation method is to output the user role to which each cluster belongs, the list of monitoring points covered by the cluster, the numerical range of the key features of the cluster, and the value range of the corresponding time series derived indicators for each cluster. The administrator demand set is set as a three-cluster summary view covering all monitoring points, and the technical personnel demand set is set as a view covering specified geological conditions. The target cluster view is defined, thus determining the data mapping results that satisfy the user roles; the time step is 1 hour, determined by the monitoring system sampling specifications; the integrity threshold is 90, determined by data continuity control requirements; the interval linear stretching is 0 to 1, determined by feature same-scale comparison requirements; the cumulative contribution rate threshold is 90, determined by engineering requirements to maintain the main information after dimensionality reduction; the number of clusters is 3, determined by the three-level classification requirements of security assessment; the initial center selection rule is determined to avoid uncertainty caused by randomness; the center displacement convergence threshold is 0.01 and the maximum number of iterations is 100, determined by balancing operational stability and computational cost; the cluster-to-role mapping ratio threshold is 70, determined by the requirement of clear role attention classification; the consistency 95, maximum depth 6, minimum leaf node sample number 30, and information gain improvement of less than 1 in the decision tree stopping conditions are determined by avoiding overfitting and ensuring the interpretability of the tree structure; the mapping accuracy threshold is 90, the weight adjustment step size is 10, and the maximum number of repetitions is 5, determined jointly by model stability and convergence efficiency.

[0023] S2 includes: obtaining demand difference data from the user demand set; classifying the difference data using node splitting rules; if the difference data points to management personnel, extracting security level quantitative indicators; determining preliminary deconstruction information through information gain calculation to obtain a deconstruction information set; grouping the security level quantitative indicators using a random forest algorithm for the deconstruction information set; retaining the grouped data if the accuracy of the grouping result is higher than a preset threshold, and removing outlier data if it is lower than the preset threshold, resulting in a grouped data set; obtaining security level quantitative indicators from the grouped data set; formatting the indicator data using a standardization protocol to obtain a formatted indicator set; extracting key features of the security level quantitative indicators from the formatted indicator set using principal component analysis to obtain a feature-extracted data set; obtaining key feature data from the feature-extracted data set; grouping the feature data using cluster analysis to determine the priority of each group of features, resulting in a priority grouping set; generating security level assessment results from the priority grouping set if the group priority is higher than a preset threshold, using weighted calculation to obtain an assessment result set; obtaining the security level assessment results from the assessment result set; and generating a final security level recommendation set using mapping rules to determine the security level mapping result that meets the needs of management personnel.

[0024] First, each user requirement is compared against the role field, geological condition field, monitoring point identifier field, and time period field. Any difference in any field value is considered a requirement discrepancy and added to the discrepancy data set. Then, node splitting rules are applied to classify the discrepancy data set. Specifically, the overall uncertainty value for the role field is calculated first; this value is generated by the sample proportion of each role and represents the degree of classification confounding. Next, the weighted uncertainty value after splitting by role field is calculated; the difference between the two is the information gain value. The field with the largest information gain value is used as the current splitting field to generate two or more branches, including a branch pointing to management personnel. When the branch pointing to management personnel is reached... When personnel are branched, the system immediately extracts the safety level quantification indicators from the branch sample and generates an indicator list. In this implementation, the safety level quantification indicators are fixed as eight items: number of times exceeding limits, duration of exceeding limits, maximum exceedance range, average fluctuation range, maximum fluctuation range, recovery time, proportion of abnormal points in the monitoring segment, and cumulative risk duration. All of these are derived from the data output in step S1, requiring no additional calculation module. The system continues to iteratively apply information gain-driven node splitting to the data pointing to management personnel until any of the following conditions are triggered: information gain increase is less than 1, leaf node sample number is less than 30, or tree depth reaches 6. The system then combines the field values ​​on each leaf node with the corresponding indicators. The inventory collection serves as preliminary deconstruction information, and the preliminary deconstruction information from all leaf nodes is summarized to form a deconstruction information set. The system uses this deconstruction information set as input to build a random forest model to group the security level quantification indicators. The random forest consists of 100 decision trees, with a fixed maximum depth of 6 and a fixed minimum leaf node sample size of 30. The splitting criterion is fixed as information gain. Samples are drawn with replacement from the deconstruction information set using a bootstrap sampling method with the same sample size as the original. Half of the indicator subset is randomly selected at each splitting node to participate in the split. Training and validation are divided chronologically: 80% of the data from the early stage is used for training, and 20% of the data from the later stage is used for validation. After training is completed... The grouping accuracy is calculated on the validation set by dividing the number of samples correctly assigned to the target group by the total number of samples in the validation set and multiplying by 100 to obtain the percentage. The system fixes the accuracy threshold at 90. When the accuracy is not lower than 90, the grouping result is directly retained and output as the grouped data set. When the accuracy is lower than 90, the system performs anomaly removal and retraining. The anomaly removal criteria are that the highest vote percentage of a sample in the voting results of all trees is less than 50 or the sample is continuously assigned to different groups in three adjacent time periods. If either condition is met, it is judged as an anomaly and removed, and retraining is performed. If the threshold is not reached after 5 consecutive retrainings, the highest accuracy result in the 5th retraining is used as the output to ensure that the process terminates.The system then performs a standardization protocol on the grouped dataset and generates a formatted indicator set. The standardization protocol is fixed as follows: field names are aligned to uniform Chinese names; time is standardized to year, month, day, hour, minute, and second; units are standardized to internationally accepted units; missing values ​​are marked with a negative one and do not participate in subsequent calculations; all indicators are linearly scaled to the 0-100 range based on their minimum and maximum values ​​in the grouped dataset to achieve comparisons at the same scale. The system uses the formatted indicator set as input to perform principal component analysis to extract key features. The processing order is as follows: first, the variance and co-variance relationships of each indicator are calculated based on a 0-100 scale, resulting in a set of independent new feature directions, which are then analyzed according to their variance contribution. The features are sorted from highest to lowest contribution. When the cumulative contribution reaches at least 90%, a corresponding number of new features are extracted as key features to form a feature-extracted dataset. Typically, 3 to 5 key features are retained. The contribution ratio of each key feature to the original 8 indicators is recorded for subsequent weighting. The system reads the key feature data from the feature-extracted dataset and performs cluster analysis to determine grouping and priority. The cluster analysis uses mean clustering with a fixed number of 3 groups to correspond to low-risk, medium-risk, and high-risk groups, respectively. The initial center selection rule is fixed: the minimum, middle, and maximum values ​​of the first key feature are taken as the three initial centers. The distance metric uses a weighted distance of key features on a scale of 0 to 100, with the aforementioned contribution ratio as the weights. The iterative process involves calculating the weighted distance to the three centers for each sample and assigning it to the group with the smallest distance. Then, the centers are updated based on the arithmetic mean of the samples within each group on each key feature. This calculation is repeated until either the maximum center displacement is less than 0.01 or the number of iterations reaches 100, at which point the process stops and outputs three stable groups. The system calculates a priority score for each group. The priority score is calculated using a weighted comprehensive method, with the weights fixed at 40, 35, and 25 for the first three key features. If a fourth or fifth key feature exists, its weight is fixed at 0. The system performs a weighted calculation on the values ​​of the center of the group on each key feature, with the weights applied and kept within the range of 0 to 100. The priority threshold is fixed at 70. When a priority score of a group is not lower than 70, it is judged as high priority and included in the subsequent evaluation. When a priority score of a group is lower than 70, it is marked as low priority and is only used as a reference and not included in the evaluation result set. The system performs a weighted calculation on all high-priority groups to generate a security level evaluation result. Specifically, it performs a linear weighted calculation on the values ​​of the center of the group on the first three key features with the same weights as the priority and outputs a score range of 0 to 100. The higher the score, the higher the risk. The scores of all high-priority groups constitute the evaluation result set.Finally, the system converts the evaluation result set into a safety level recommendation set according to the mapping rules. The mapping rules are fixed as follows: a score of 80 or above is determined as safety level 3, a score between 50 and 79 is determined as safety level 2, and a score between 0 and 49 is determined as safety level 1. Each level result is output along with its corresponding deconstruction path, group number, key feature value range, and coverage time period to form a safety level mapping result that meets the needs of management personnel. The information gain threshold of 1, the minimum number of leaf node samples of 30, and the maximum depth of 6 in node splitting are used to limit the size of the tree to prevent overfitting and ensure interpretability. The number of trees in the random forest is 100 to balance stability and computational cost. The indicator subset ratio of 50 is used to enhance diversity. The training and validation split ratios of 80 and 20 are used to ensure that the model is trained on history and validated on recent periods. The grouping accuracy threshold is 9. 0 and the maximum number of retraining iterations (5) are used to ensure output quality and ensure the process terminates within a finite number of steps. The highest vote ratio (50) and the number of consecutive inconsistencies (3) in the anomaly detection are used to exclude unstable samples. Missing values ​​in the standardization protocol are marked as negative one to unify missing value handling and prevent them from participating in subsequent calculations. Linear scaling to 0-100 is used to eliminate dimensional differences. The cumulative contribution threshold (90) for principal component analysis is used to retain key information while reducing dimensionality. The number of clusters (3) is used to align with the three-tier management strategy. The central convergence threshold (0.01) and the maximum number of iterations (100) are used to ensure cluster convergence. The priority threshold (70) is used to screen key groups that need to be evaluated. Weighting coefficients (40, 35, and 25) are used to highlight the dominant role of the first three key features and fix the caliber. The tier mapping boundary (80 and 50) is used to align with the management plan threshold and establish clear tier boundaries.

[0025] S3 includes obtaining preliminary deconstruction information from the user demand set, removing noisy data through data cleaning methods to obtain a cleaned deconstruction information set; classifying the cleaned deconstruction information set using a decision tree algorithm, generating mechanical parameter groups based on leaf node outputs to obtain a mechanical parameter group set; if there are parameter groups in the mechanical parameter group set whose risk exceeds a preset threshold, then further subdividing these parameter groups through cluster analysis to obtain a subdivided risk data set; for the subdivided risk data set, extracting water pressure fluctuation data using a data recombination method to obtain a water pressure fluctuation data set; obtaining key fluctuation characteristics from the water pressure fluctuation data set, generating formatted fluctuation data through standardization processing to obtain a formatted fluctuation data set; for the formatted fluctuation data set, generating chart-based data on water pressure fluctuations using a visualization transformation method to obtain a visualization data set; if the visualization data set meets preset risk assessment criteria, then determining the final safety level mapping result through weighted calculation to obtain a safety level mapping set.

[0026] In this implementation, preliminary deconstruction information is first read from the user demand set in record order. Each piece of information includes the monitoring point number, start and end time, mechanical parameter name, corresponding numerical sequence, and risk label. Then, a data cleaning method is performed to remove noisy data. The cleaning order is fixed as four steps: time alignment, missing data handling, outlier removal, and mutation removal. Time alignment rearranges the sequences of the same monitoring point at equal intervals based on second-level timestamps without interpolation. Missing data handling replaces null and illegal values ​​with negative one and excludes them from subsequent calculations. Outlier removal calculates the mean and standard deviation of each sequence over the complete time period covered by the deconstruction information. When the difference between the value of a sampling point and the mean exceeds three times the standard deviation, it is judged as noise. Points are removed; mutation removal calculates the relative rate of change between two adjacent sampling points. When the relative rate of change exceeds 200%, the latter is marked as noise and removed. The three-standard-deviation threshold and the 200% threshold are fixed values ​​derived from the robustness requirements of tunnel monitoring to ensure the removal of occasional spikes and non-physical jumps. After completion, a cleaned deconstruction information set is formed. The system then uses the cleaned deconstruction information set as input to construct a decision tree and perform classification. The input fields are the time-period statistics and risk labels corresponding to the mechanical parameters. The statistics are fixed as the average, maximum, minimum, and over-limit counts within the time period. The tree splitting rule adopts the maximum information gain criterion, and the splitting stopping condition is fixed as the information gain increase being less than 1 or the number of leaf node samples being less than 30. If the tree depth reaches any of the conditions in section 6, after the tree is built, the sample set of each leaf node is used as a group of mechanical parameters and the groups are aggregated to form a set of mechanical parameter groups. The system checks each group to see if there are any parameter groups whose risk exceeds a preset threshold. The risk judgment threshold is a fixed ratio of the parameter value to the engineering benchmark value. For stress, the threshold is 80% of the design strength; for displacement and convergence, the threshold is 70% of the allowable deformation; and for pore water pressure, the threshold is 85% of the warning value. If any record in any group exceeds the corresponding threshold at any time point, the group is judged as a risk group and enters the subdivision process. The system performs cluster analysis on each risk group for subdivision. The clustering method is fixed as mean clustering, and the number of groups is fixed at 3 to correspond to low-risk subgroups. The system divides the risk data into medium-risk and high-risk subgroups. The initial center is taken from the minimum, median, and maximum value samples of the target indicator for each risk group. The distance metric is the Euclidean distance of the target indicator after interval scaling. The iteration stops when the maximum difference between the center positions of the current and previous iterations is less than 0.01 or the number of iterations reaches 100. After subdivision, the output is a subdivided risk data set. The system then performs data recombination on the subdivided risk data set to extract water pressure fluctuation data. The recombination step involves selecting only indicators whose names contain "water pressure" or "pore water pressure," concatenating the values ​​of these indicators at the same monitoring point on the same time axis in ascending time order to form a continuous sequence, retaining the original timestamps, and performing no interpolation or smoothing to form the water pressure fluctuation data set.The system extracts key fluctuation features from the water pressure fluctuation data set. These key features are fixed as four items: maximum fluctuation amplitude, average fluctuation amplitude, fluctuation period, and abrupt change frequency. The maximum fluctuation amplitude is calculated as the difference between the maximum and minimum values ​​in the sequence within the analysis period. The average fluctuation amplitude is calculated as the arithmetic mean of the absolute differences between adjacent sampling points within the analysis period. The fluctuation period is calculated by sequentially identifying continuous upward and downward segments within the analysis period, calculating their durations, and then taking the arithmetic mean. The abrupt change frequency is calculated by counting the number of times the difference between adjacent sampling points exceeds the average difference of the entire sequence plus twice the standard deviation of the difference, dividing by the period length, and converting this to the number of times per week. The analysis period is fixed at 7 days. A weekly sliding window with no overlap is used, and the aforementioned threshold of twice the standard deviation of the difference is used to identify only significant abrupt changes and align them with the weekly scale. The system then performs standardization on four key fluctuation features to generate a formatted fluctuation dataset. The standardization rule is to linearly scale each feature to the range of 0 to 100 using the minimum and maximum values ​​in the current evaluation samples as endpoints. Missing values ​​and the aforementioned negative one marker are not included in this scaling step. The system uses the formatted fluctuation dataset as input and performs a visualization transformation method to generate graphical data for water pressure fluctuations. The visualization is transformed into two deterministic graphical data structures: the first is a line graph data structure, with the horizontal axis representing the time stamp sequence and the vertical axis representing the water pressure value, containing one sampling point per hour. The second type uses a bar chart data structure, with the horizontal axis representing the fixed order of four key fluctuation characteristics and the vertical axis representing the corresponding standardized values. Both data structures output simultaneously to form a visualized data set. The system performs a risk assessment on this visualized data set and determines whether to proceed with weighted calculations accordingly. The risk assessment criteria are fixed as follows: when the standardized value of the maximum fluctuation amplitude is not less than 60, or the standardized value of the average fluctuation amplitude is not less than 50, or the standardized value of the fluctuation period is less than 20, or the frequency of mutation converted to weekly counts is not less than 30, the risk assessment criteria are considered met. These thresholds are derived from the alarm and intervention boundaries of the management plan and are fixed as read-only configurations before system deployment. When the risk assessment criteria are met, the system applies the following standards to the four key fluctuation characteristics: The standardized values ​​are weighted to determine the final safety level mapping result. The weighting coefficients are fixed as follows: maximum fluctuation amplitude weight 40, average fluctuation amplitude weight 30, fluctuation period weight 20, and mutation frequency weight 10. The four standardized values ​​are linearly combined with the weights in this order to obtain a risk score from 0 to 100. The risk score classification mapping rule is fixed as follows: when the score is not lower than 80, it is determined as safety level 3; when the score is between 50 and 79, it is determined as safety level 2; and when the score is between 0 and 49, it is determined as safety level 1. The safety level, the corresponding monitoring point number, the start and end time, the risk group number, the value range of the four key fluctuation characteristics, and the reference identifier of the visualization data structure are written into the safety level mapping set.Three standard deviations and two standard deviations thresholds are used to stably identify outliers and mutations; a 200% rate of change threshold is used to eliminate non-physical jumps; information gain enhancement of 1, minimum leaf node sample size of 30, and maximum depth of 6 are used to limit the size of the decision tree and ensure interpretability; 80%, 70%, and 85% of the engineering baseline proportions correspond to the risk red lines of stress, displacement, and water pressure, respectively, and are directly derived from the engineering contingency plan; the number of cluster groups of 3, the central convergence threshold of 0.01, and the maximum number of iterations of 100 are used to ensure that the subdivision results are consistent with the three-level management strategy and convergent; the analysis period of 7 days and the mutation frequency count weekly are used. To align with the weekly maintenance inspection schedule, a standardized range of 0 to 100 is used to eliminate dimensional differences and provide a unified scale for subsequent visualization and weighting. Risk assessment thresholds of 60, 50, 20, and 30 are used to quickly screen high-risk sequences after graphical representation. Weighting coefficients of 40, 30, 20, and 10 are used to emphasize the dominant role of magnitude classes in structural safety. Grading boundaries of 80 and 50 are used to align with the activation and attention boundaries of contingency plans. This ensures that the processes of cleaning, classification, subdivision, reorganization, extraction, standardization, visualization, and weighted mapping form a closed loop, outputting a set of safety level mappings that are verifiable, executable, and meet the decision-making needs of management personnel.

[0027] S4 includes: extracting local time segments from time series data using a sliding window method to obtain a set of time segments; decomposing the time segment set into frequency components using a Fourier transform method to obtain a set of frequency components; if the proportion of frequency components in the set whose frequencies exceed a set frequency threshold exceeds a preset threshold, removing noise from the frequency components whose frequencies exceed the set frequency threshold using a filtering method to obtain a filtered set of frequencies; generating a smoothed time series based on the filtered set of frequencies using a time series reconstruction method to obtain a smoothed time series set; extracting trend change features from the smoothed time series set to obtain a trend feature set; determining the trend stability based on the trend feature set using a classification method to obtain a stability classification set; and generating output data with technical annotations based on the stability classification set to obtain the final technical output set.

[0028] In this embodiment, the original time series from step S3 is first segmented using a fixed sliding window. The window length is set to 24 hours and the step size to 1 hour. Any time series is continuously truncated from its first full hour according to the above parameters, forming a set of time segments containing several segments of length 24. The system then processes each time segment and performs frequency decomposition, treating each segment as an equally spaced sequence with an interval of 1 hour. The amplitude and corresponding frequency of each frequency component contained in the segment are calculated to obtain a set of frequency components. To avoid interference from weak amplitude noise on component statistics, the system first uses the median of the amplitude of all frequency components in the segment as a benchmark to determine the effective component judgment threshold as the median. Components with amplitudes below the threshold are considered invalid and not included in the percentage statistics, while components with amplitudes above the threshold are considered valid and participate in subsequent judgments. The system sets a fixed frequency threshold of 12 times per day, determined by the sampling theorem under hourly sampling conditions. This value is used to distinguish between high-frequency disturbances and structure-dominant frequencies. The number of valid components with frequencies above the threshold in each time segment is counted and divided by the total number of valid components in that segment, multiplied by 100 to obtain the high-frequency percentage. When the high-frequency percentage exceeds 20%, the segment is considered to contain high-frequency noise and enters the filtering process; otherwise, no filtering is required, and the system directly enters the reconstruction process. The filtering process adopts a deterministic low-pass strategy, specifically retaining low-frequency components. The system identifies effective components with a set frequency threshold and sets the amplitude of components with frequencies above the threshold to zero. All processed components are collectively denoted as the filtered frequency set. Time series reconstruction uses this filtered frequency set as input. The system calculates the numerical contribution of each component's amplitude and phase in the time domain moment by moment and sums the contributions of all components at the same moment to obtain the smoothed value for that moment. This generates a smoothed time series of length 24 for each time segment. All smoothed series are aggregated into a smoothed time series set. The system extracts trend change features from the smoothed time series set. The fixed features are monotonic rate of change, fluctuation amplitude, number of inflection points, and trend duration (4 items in total). The monotonic rate of change is determined by... The percentage is calculated by subtracting the starting time value from the ending time value of the segment, dividing by 24 hours, and multiplying by 100. The fluctuation amplitude is obtained by taking the difference between the maximum and minimum values ​​within the segment. The number of inflection points is obtained by counting the number of times the sign changes from positive to negative or from negative to positive after taking the sign of the difference between adjacent times. The duration of the trend is obtained by the length of the longest continuous segment in the segment where the direction remains unchanged. The system then performs stability classification on the four features of each time segment. The classification rules are fixed and the boundary processing order is clear. First, the stability conditions are judged. If the monotonic rate of change is not higher than 10, the fluctuation amplitude is not higher than 5, the number of inflection points is not more than 3, and the duration of the trend is not less than 18 hours, then it is marked as stable.If the stability condition is not met, the fluctuation stability condition is determined. If any one of the following conditions is met: a monotonic rate of change greater than 10 and not higher than 30, a fluctuation amplitude greater than 5 and not higher than 15, a number of inflection points greater than 3 and not more than 8, or a trend duration not less than 6 and less than 18 hours, then the fluctuation is marked as stable. If neither of the first two conditions is met, the system is marked as unstable. Specifically, if any one of the following conditions is met—a monotonic rate of change greater than 30, a fluctuation amplitude greater than 15, a number of inflection points greater than 8, or a trend duration less than 6 hours—the system classifies the system as unstable. To avoid cross-conflict, the above conditions are applied sequentially, and each segment is classified into only one category. The system then associates the classification results with the corresponding monitoring point number. The start and end times, along with four feature values, are written into the stability classification set. After classification, the system generates output data with technical annotations based on the stability classification set. The annotation generation uses a fixed template and strictly corresponds to the classification labels. When the label is "stable," the output is "Structural response is stable, monitoring trend is stable." When the label is "fluctuating and stable," the output is "Slight structural fluctuations exist; it is recommended to pay closer attention and maintain the current monitoring frequency." When the label is "unstable," the output is "Abnormal structural response; it is recommended to initiate key monitoring and review." The annotation includes specific values ​​for the monotonic rate of change, fluctuation amplitude, number of inflection points, and duration of the trend for that segment. After verification, all annotation records are matched one-to-one with the original fragment identifiers to form the final technical output set; the window length of 24 hours and the step size of 1 hour are determined by the sampling frequency and the needs of diurnal periodic analysis; the frequency threshold of 12 times per day is determined by the sampling theorem under hourly sampling conditions; the high-frequency proportion threshold of 20 is determined by the statistical results of the proportion of high-frequency interference in historical data and is fixed during system deployment; the effective component judgment threshold is taken as 10 of the median amplitude to stably shield weak amplitude noise and ensure comparability between different fragments; smooth reconstruction is strictly based on the summation of retained components at each time step without interpolation or extrapolation to ensure traceability; the calculation caliber of the four trend features is divided into... The percentage, original unit amplitude, dimensionless count, and number of hours correspond to different values. The stability classification thresholds—10 and 30 for monotonic rate of change, 5 and 15 for fluctuation amplitude, 3 and 8 for the number of inflection points, and 6 and 18 for trend duration—are all determined by long-term monitoring experience and contingency plans, and are fixed in the system as read-only configurations. Annotation text uses fixed semantic templates to avoid subjective description differences. This achieves end-to-end deterministic calculation and consistency across the entire process from slicing, frequency domain analysis, filtering, reconstruction, feature extraction, stability classification to technical annotation output, ensuring that any input data under the same parameter configuration produces a completely consistent final technical output set.

[0029] S5 includes: processing redundant branches from the enhanced technology output using decision tree pruning to obtain a simplified path set; grouping similar paths into a path grouping set using clustering analysis; simplifying the path structure by using dimensionality reduction methods to obtain a simplified path set if the complexity of a path group exceeds a preset threshold; constructing a suggestion list based on the simplified path set and quantifying the security level of each path by comparing it with preset risk thresholds to obtain a security level set; organizing the level descriptions using text generation methods to obtain a structured text set; extracting key level information from the structured text set to generate a concise text summary to obtain the final text output; and associating the final text output with business scenarios using mapping methods to obtain a scenario adaptation description.

[0030] In this implementation, path information is first read sequentially from the enhanced technology output. Each path consists of a condition sequence from the root node to the leaf node, a coverage time range, the number of monitoring points involved, security level records appearing on the path, and risk warnings. The system performs decision tree pruning on each path to handle redundant branches. Pruning starts from the last branch and proceeds from bottom to top. Three deterministic indicators are calculated as deletion criteria: the number of leaf node samples, the percentage of incorrectly labeled leaf nodes, and the overall accuracy reduction. The number of leaf node samples is the number of records contained in that leaf node; the percentage of incorrectly labeled leaf nodes is the percentage of records in that leaf node that differ from the main label; and the overall accuracy reduction is the percentage of records in the entire leaf node that are different from the main label after deleting that branch. The system calculates the percentage decrease in the correct classification ratio of the path in the historical records compared to before deletion. When the number of leaf node samples is less than 30, the mislabeling ratio of leaf nodes is not higher than 10%, and the overall accuracy decrease is not higher than 1%, the terminal branch is determined to be a redundant branch and deleted. At the same time, its parent node and adjacent direct edges are merged to maintain structural connectivity. If two consecutive terminal branches that meet the conditions appear on the same path, they are deleted simultaneously. The pruning cycle continues until there are no branches that meet the deletion conditions, resulting in a simplified path set. The system then performs cluster analysis on the simplified path set to group similar paths. The clustering method is fixed as mean clustering, and the number of groups is fixed at 3 to correspond to the three categories of short, medium, and complex paths. The similarity measurement is based on five parameters. Qualitative features are linearly scaled to the range of 0 to 100 before entering clustering. The five features are, in order: total number of nodes, maximum depth, number of conditional field types, number of key feature entries, and number of risk exceedances triggered on the path. The scaling endpoints are the minimum and maximum values ​​of the feature across all paths. Distance calculation uses a weighted approach, with the five weights being: total number of nodes (30), maximum depth (25), number of conditional field types (20), number of key feature entries (15), and number of risk exceedances (10). Initial center selection involves the first, middle, and last paths sorted by total number of nodes from smallest to largest. In each iteration, the weighted distance from all paths to the three centers is calculated and the path with the lowest distance is selected. The system first divides the path into smaller groups. Then, it updates the three centers based on the arithmetic mean of the five features within each group. The iteration stops when the maximum change in the center after any round of updates is no higher than 0.5 or the number of iterations reaches 100, taking the first one as the standard. This results in a path group set. The system calculates the complexity score for each path group set and identifies whether it exceeds a preset threshold. The complexity score is the value of the linear combination of the aforementioned five standardized features with equal weights and is kept in the range of 0 to 100. The complexity threshold is fixed at 70 and is derived from the balance requirement between interpretability and display capacity. When the complexity score of any path in a group is higher than 70, the path is marked as a complex path and enters the dimensionality reduction process. Paths that do not exceed the threshold directly proceed to the next step.Dimensionality reduction methods are used to simplify path structure without altering path semantics. The processing object is five features of complex paths. The system first calculates the correlation between features and generates a set of uncorrelated combined features. These features are then sorted from highest to lowest according to their contribution to the overall variance. When the cumulative contribution reaches at least 90%, a corresponding number of combined features are truncated and mapped to two comprehensive indices, named the structural complexity comprehensive value and the risk density comprehensive value, respectively. Subsequently, the contribution ratio of each condition node to these two comprehensive indices is calculated from top to bottom along the path. Condition nodes with a contribution ratio less than 2 are considered redundant and deleted. Simultaneously, their adjacent direct edges are merged to maintain path connectivity. When three or more consecutive... For low-contribution nodes, only the one with the highest contribution percentage is retained to avoid oversimplification. After dimensionality reduction, these nodes are merged with unreduced paths to form a simplified path set. The system constructs a suggestion list based on the simplified path set and quantifies the security level. The suggestion list outputs specific fields for each path, including path number, number of covered monitoring points, coverage time length, proportion of security level 3 nodes in the path, proportion of security level 2 nodes in the path, highest risk warning on the path, comprehensive risk density value, and suggested action. The risk score consists of three parts and is maintained in the range of 0 to 100: a 50% weighting for the proportion of security level 3, a 30% weighting for the proportion of security level 2, and a 20% weighting for the comprehensive risk density value. The sum of these three parts is the final score. The path risk score is preset with two risk thresholds: 80 and 50. A risk score of 80 or higher is designated as safety level 3, a risk score between 50 and 79 is designated as safety level 2, and a risk score between 0 and 49 is designated as safety level 1. The recommended actions are given in the suggestion list: Safety level 3 corresponds to immediately arranging on-site verification and adjusting the monitoring frequency to once every 4 hours; safety level 2 corresponds to strengthening monitoring and adjusting the monitoring frequency to once every 12 hours; and safety level 1 corresponds to maintaining routine monitoring once every 24 hours. This generates a set of safety levels. The system uses this set of safety levels as input to generate a text description of the level, employing a fixed template and filling in the fields in order. The template fields are, in order: path number, start time, end time, number of monitoring points, highest level, risk score, suggested action, and review time. The time format is uniformly set to year, month, day, hour, minute, and second, and the numerical values ​​are uniformly set to Arabic numerals. The generated record set is a structured text set. The system extracts key level information from the structured text set to generate a concise text summary. The extraction rule is to retain the top three records of the highest level for each cluster group and sort them from high to low risk score. At the same time, adjacent records with the same level, the same path number, and the same time range are merged into one record to reduce redundancy. The output fields are limited to path number, level, risk score, and suggested action to form the final text output.The system finally executes a mapping method to associate business scenarios based on the final text output. The mapping rules are fixed as follows: when the security level is 3, it is associated with an emergency response scenario, and the scenario adaptation description includes three fixed phrases: on-site verification, temporary traffic control, and encrypted monitoring; when the security level is 2, it is associated with a maintenance scenario, and the fixed phrases: enhanced inspection, sensor verification, and phased pressure testing are added; when the security level is 1, it is associated with an operation management scenario, and the fixed phrases: routine inspection, weekly reporting, and trend tracking are added. The final output is a scenario adaptation description. The thresholds for the number of leaf node samples (30), the error labeling ratio (10), and the accuracy reduction (1) are used to control the pruning intensity and ensure that the deletion does not damage the overall discrimination ability. The number of cluster groups (3) is used to align with the three-stage complexity management strategy. The feature standardization range (0 to 100) is used to unify the units of measurement. The five clustering weights are related to the complexity. The numerical configuration of weights is used to highlight the impact of node scale and depth and is fixed as a read-only configuration during system deployment. A complexity threshold of 70 is used to balance readability and fidelity. A cumulative contribution threshold of 90 and a node contribution percentage threshold of 2 are used to achieve stable dimensionality reduction and controllable deletion while retaining key information. Risk scoring weights of 50, 30, and 20 are used to emphasize the dominant role of the level ratio in the final level. Risk thresholds of 80 and 50 are used to align with the activation and attention boundaries of the contingency plan. Monitoring frequencies of 4 hours, 12 hours, and 24 hours are used to match the intensity of emergency, enhanced, and routine operations. This achieves deterministic calculation and consistent standards across the entire chain, from pruning, clustering, dimensionality reduction, quantification, textualization, and refinement to business mapping, ensuring that the same input under the same parameters produces a completely consistent set of simplified paths, a suggestion list, final text output, and scenario-adaptive descriptions.

[0031] S6 includes: obtaining all path data from the user path output set; grouping the path data using clustering analysis to obtain a path group set; if the format of a path group does not match the target format, adjusting the path data according to preset format conversion rules to obtain a path set with consistent format; extracting the suggested priority attribute of each path from the path set with consistent format, and sorting them from high to low priority using a sorting algorithm to obtain a sorted path set; obtaining the highest priority path data from the sorted path set, associating it with a preset business template using a mapping method to generate customized output data; simplifying the data structure using dimensionality reduction methods to obtain a streamlined output set if the data volume exceeds a preset threshold; organizing the output content using text generation methods based on the streamlined output set to obtain the final output; and extracting key information fields from the final output, saving them using structured storage methods to obtain a persistent output result.

[0032] In this implementation, all path data is first read from the user path output set at once, and the completeness of each field is verified. Each path is fixed to include 14 items: path number, start time, end time, total number of nodes, maximum depth, number of condition field types, number of key feature entries, number of risk exceedances, proportion of occurrences of security level 3, proportion of occurrences of security level 2, number of covered monitoring points, coverage duration, existing text description, and priority suggestion placeholder fields. The time is uniformly set to year, month, day, hour, minute, and second; the proportion is uniformly set to integers from 0 to 100 or one decimal place; and the count is uniformly set to non-negative integers. The system then performs cluster analysis on all path data to group them. The clustering method is fixed to mean clustering, and the number of groups is fixed to 3. The nodes are labeled as short path, medium path, and long path groups. Similarity is measured based on five numerical features, which are linearly scaled to the range of 0 to 100 before clustering. The five features are, in order: total number of nodes, maximum depth, number of condition field types, number of key feature entries, and number of risk exceedances. The scaling endpoints are the minimum and maximum values ​​of the feature across all paths. Distance calculation uses a weighted Euclidean distance numerical process, but the formula is not shown. The weights are fixed as follows: total number of nodes (30), maximum depth (25), number of condition field types (20), number of key feature entries (15), and number of risk exceedances (10). Initial centers are selected from the first, middle, and last samples based on the total number of nodes, from smallest to largest. In each iteration, the weighted distance from all paths to the three centers is calculated sequentially and grouped into the group with the smallest distance. Then, the three centers are updated according to the arithmetic mean of the samples in each group across five features. The stopping condition is that the maximum change in the center after any round of updates is no greater than 0.5 or the number of iterations reaches 100, whichever comes first, and the path group set is output. The system checks the path group set group by group to see if the path format matches the target format. The target format strictly specifies the field order as follows: path number, start time, end time, total number of nodes, maximum depth, number of condition field types, number of key feature entries, number of risk exceedances, proportion of occurrence of security level 3, proportion of occurrence of security level 2, number of covered monitoring points, coverage duration, existing text description, and suggested priority. The placeholder rule is to set missing fields to 0 or empty strings. The conditions for determining mismatch are: the time field does not conform to the year, month, day, hour, minute, and second; the ratio field is not between 0 and 100; the count field is not an integer or is less than 0; and the field order is inconsistent with the target format. If a group has mismatched paths, it will be adjusted one by one according to the preset format conversion rules. The rules are: rename the fields to fixed Chinese names, rearrange the fields to the target order, parse all the time and unify them to year, month, day, hour, minute, and second; multiply the ratio field by 100 or round it to enter the 0 to 100 range; set the count to 0 when it is less than 0; fill the missing fields with 0 or empty strings and mark the source in the remarks. After the conversion is completed, a full verification is performed. After all passes, a set of paths with the same format is output.The system calculates a suggested priority attribute for each path based on a set of paths with a consistent format. The suggested priority score is obtained by a linear combination of three parts, ranging from 0 to 100. The weights of these three parts are: the proportion of occurrences of security level 3 (50%), the proportion of occurrences of security level 2 (30%), and the risk exceedance count normalized score (20%). The risk exceedance count normalized score is obtained by linearly scaling the minimum and maximum values ​​of this indicator within the group to 0 to 100. The sum of the three parts is the suggested priority score. Simultaneously, priority level labels are generated and divided using fixed thresholds: 80 or higher is considered high, 50 to 79 is medium, and 0 to 49 is low. The aforementioned weights and thresholds are fixed as read-only configurations during system deployment to ensure consistency. The system performs a sorting algorithm on all paths, arranging them from highest to lowest priority according to the suggested priority. The sorting method is fixed as a stable merge sort with deterministic tie-breaking rules. If priority scores are the same, they are sorted from highest to lowest based on the proportion of occurrences of security level 3. If they are still the same, they are sorted from highest to lowest based on the number of times risk limits are exceeded. If they are still the same, they are sorted from smallest to largest based on path number. After sorting, the sorted path set is output. The system retrieves the highest priority path data from the sorted path set. If there are ties, the path that is ranked first after ties is retrieved, and then the mapping is performed. The mapping method is associated with a preset business template. The mapping rule is fixed as follows: when the proportion of security level 3 occurrences within the path is higher than or equal to 50%, the emergency response template is selected; when the proportion of security level 3 occurrences is lower than 50% and the proportion of security level 2 occurrences is higher than or equal to 50%, the maintenance template is selected; when both are lower than 50%, the operation management template is selected. The template fields are fixed as title, path number, time range, number of monitoring points, main risk description, priority score, suggested action and review time. The system fills in the fields in order and generates customized output data. The system checks whether the amount of customized output data exceeds a preset threshold, which is fixed as the number of records. If the value exceeds 1000, a dimensionality reduction method is used to simplify the data structure. The dimensionality reduction is based on the field of variance contribution. The processing order is to calculate the cross-record variance for all numerical fields except for the title and path number and sort them from high to low. When the cumulative variance contribution reaches or exceeds 90, the corresponding field is truncated and the rest of the fields are discarded. At the same time, the existing text description is truncated to keep the first 200 words and an ellipsis mark is added at the end. The fields retained after dimensionality reduction are fixed as path number, start time, end time, proportion of occurrence of security level 3, proportion of occurrence of security level 2, number of times risk exceeds the limit, suggested action and priority score. The output is a simplified output set.The system organizes the output content using a text generation method based on the simplified output set. The text generation employs a fixed template, filling in each line sequentially. The template format is: a record with a path number covering a certain number of monitoring points between a certain time and a certain time, exhibiting a certain percentage of security level 3 occurrences, a certain percentage of security level 2 occurrences, a certain number of risk exceedances, a certain priority score, and a recommended action to be performed with a review completed at a certain time. All values ​​use Arabic numerals, all times use year, month, day, hour, minute, and second, and time zones are kept consistent. The generated result is the final output. The system then extracts key information fields from the final output and executes... Structured storage is used, with key information fields fixed as path number, start time, end time, priority score, priority level label, proportion of occurrences at security level 3, proportion of occurrences at security level 2, number of risk exceedances, suggested actions, and business template type. The structured storage method is fixed: a composite primary key is created using the path number, start time, and end time; a secondary index is created using the priority score and priority level label; and a time index is created using the time field. After successful storage, persistent output is returned. The number of clusters (3) and weights (30, 25, 20, 15, 10) are used to associate the three path segments. The complexity management strategy and interpretability trade-offs remain consistent. A center change threshold of 0.5 and a maximum iteration count of 100 are used to ensure cluster convergence and computational controllability. The target format and validation rules ensure consistency and verifiability across stages. Recommended priority weights of 50, 30, and 20 are used to emphasize the dominant role of higher-level proportions in the ranking results while also considering the frequency of risk events. Priority level thresholds of 80 and 50 are used to align with the initiation and attention boundaries of operational and maintenance procedures. In the business template mapping, a 50 threshold is used to provide clear boundaries between different handling scenarios. A data volume threshold of 1000 and the cumulative contribution of dimensionality reduction are also considered. A threshold of 90 and a text truncation length of 200 are used to control the interface and transmission load while retaining key information. The composite primary key and index of structured storage ensure deduplication and efficient retrieval, thereby achieving end-to-end deterministic computation from path grouping, format unification, priority calculation and stable sorting, template mapping, data volume control and dimensionality reduction, written output to structured persistence. This ensures that under the same input and parameter configuration, completely consistent path grouping sets, formatted path sets, sorted path sets, customized output data, simplified output sets, final output, and persistent output results are generated.

[0033] S7 includes obtaining format data from the final customized report, extracting the structural features of each format using a format parsing method to obtain a structured format dataset; if the structured format dataset does not match the preset format standard, adjusting the data through format conversion rules to obtain a standardized format dataset; for the standardized format dataset, using a consistency verification algorithm to check the data integrity and standardization to obtain a set of verified formats; based on the set of verified formats, matching group attributes using a user group analysis method to generate a distribution mapping table corresponding to the group; obtaining distribution data from the distribution mapping table, integrating similar information using a content clustering method to obtain an aggregated information set; for the aggregated information set, using a keyword extraction method to filter information with a priority higher than a preset threshold to obtain a simplified information set; extracting information chains from the simplified information set, organizing and outputting them using a structured coding method to obtain the final information transmission chain.

[0034] In this implementation, format data is first read line by line from the final customized report to generate format records. Each format record contains 13 fields: format type code, title, start time, end time, path number, number of monitoring points, security level, risk score, suggested action, text body, table cell data, graphic metadata, and attachment references. Time is uniformly set to year, month, day, hour, minute, and second; values ​​are uniformly retained to one decimal place; and counts are uniformly set to non-negative integers. The system then executes a format parsing method to extract structural features and form a structured format dataset. Specifically, for text records, the number of paragraphs, sentences, and words is counted, and the line number of the title, the position of the time field, and the position of the value field are located. For table records, the number of rows, columns, field names, units, and the proportion of empty cells are counted, and the data type of each column is recorded. For graphic records, the graphic type code, the number of time axis points, the number of value points, and the number of legend items are read, and the coordinate axis units are recorded. For structured field records, the field order and the proportion of field completeness are read, and the existence of required fields is recorded. The system then compares the structured format dataset with the preset format standards. The item matching check uses a fixed preset format standard: field order and field name must be exactly the same; time format must be year, month, day, hour, minute, and second; numerical precision must be one decimal place; units must use a unified unit system; table row count must not exceed 1000; table column count must not exceed 100; number of fields must not be less than 8; field integrity ratio must not be less than 95%; and security level can only be 1, 2, or 3. If any of these conditions are not met, the table is considered a mismatch and enters format conversion. The format conversion rules are executed in a fixed order: first, the fields are renamed to their standard Chinese names, then they are re-ordered according to the target order. The process involves sorting the fields, parsing all times and unifying them to year, month, day, hour, minute, and second; unifying all units into a common unit system using a unit conversion table; rounding all values ​​to one decimal place; setting missing values ​​to 0 or an empty string and writing the correction type and correction time in the remarks; truncating tables from the end to the top if the number of rows or columns exceeds the limit and recording the truncation position; and adding uniform placeholders to fill in any missing legend items for graphs. After the conversion is complete, a full validation is performed again, and only all compliant data is retained as a standardized format dataset.The system performs a consistency verification algorithm on the standardized format dataset and obtains a set of verified formats. The consistency verification includes four types of checks: The first is integrity checking, requiring that the title, start and end times, security level, and recommended actions are not empty, and the field integrity ratio is not less than 95%. The second is range checking, requiring the risk score to be between 0 and 100, the number of monitoring points to be a non-negative integer, and the proportion of empty cells in the table not exceeding 20%. The third is correlation consistency checking, requiring the start time to be earlier than the end time, path numbers to be unique within the same report, and the security level of the same path number to be consistent across different format types. The fourth is format consistency checking, requiring that the time and structure in the text content be consistent. The time in the fields must be completely consistent, and the table field names must use a unified naming convention within the same report. If any check fails, the record is removed and added to the verification failure list. The remaining records are summarized into a set of verified formats. The system generates a distribution mapping table based on the user group analysis method using the verified format set. The group types are fixed at three categories: management personnel, technical personnel, and operation and maintenance personnel. The matching rule is fixed as follows: when the security level is 3 or the risk score is not lower than 80, the record is distributed to the management personnel group; when the security level is 2 or the risk score is between 50 and 79, the record is distributed to the technical personnel group; and when the security level is 1 or the risk score is between 0 and 49, the record is distributed to the operation and maintenance group. Simultaneously, suggestions are... The action field is matched for keywords. If it contains emergency response phrases, it is distributed to the management personnel group; if it contains maintenance phrases, it is distributed to the technical personnel group; and if it contains routine inspection phrases, it is distributed to the operation and maintenance group. The system generates a distribution priority for each record and writes it to the distribution mapping table. The priority values ​​are fixed as follows: safety level 3 is labeled 3, safety level 2 is labeled 2, and safety level 1 is labeled 1. Within the same level, higher priority is assigned based on risk score from highest to lowest; if the scores are the same, priority is given to the earlier time point. The system extracts distribution data from each group from the distribution mapping table and performs content clustering to integrate similar information. Content representation uses phrase set representation, and phrase extraction is fixed from the label... Phrases of length 2 to 6 are extracted from titles, subheadings, paragraph opening sentences, and table field names and deduplicated to form a phrase set. Similarity is calculated using the percentage value of phrase overlap. The clustering process establishes initial clusters with a similarity threshold of 70. In each iteration, records with a maximum intra-cluster difference exceeding 30 are reassigned until the intra-cluster difference is no higher than 30 or the iteration reaches 50, stopping at the first to be determined. The aggregated output is the representative record with the most occurrences in each cluster, and the records in the same cluster are merged at the field level. The earliest value is taken for the start time, the latest value for the end time, the maximum value for the security level, the maximum value for the risk score, and the item with the widest coverage for the suggested action, resulting in the aggregated information set.The system performs keyword extraction on the aggregated information set and filters it according to priority thresholds to generate a concise information set. Keyword extraction involves three steps: First, frequency is counted, and the top 20 phrases are selected as candidates based on their frequency of occurrence. Second, position weighting is applied: 20 points are added for titles, 10 points for paragraph beginnings, and 10 points for table headers. Third, level weighting is applied: 30 points are added for phrases appearing in records with security level 3, 20 points for phrases appearing in records with security level 2, and 10 points for phrases appearing in records with security level 1. The scores from the three steps are summed to obtain a keyword priority score, which is limited to between 0 and 100. The filtering threshold is fixed at 60, and phrases with scores of 60 or higher are retained and extracted. The corresponding records form a concise information set, and other phrases are discarded and not retained. The system constructs information chains from the concise information set and performs structured coding to organize and output the information. The order of information chain construction is fixed as follows: first, sort by group type, with managers in the middle, technicians in the middle, and operation and maintenance last; then sort by security level from high to low; and finally sort by time starting point from early to late. The coding rule is fixed as consisting of 5 segments of numbers connected by hyphens. The first segment is a 2-digit group type code: managers 11, technicians 22, operation and maintenance 33; the second segment is a 1-digit security level code: level 1 1, level 2 2, level 3 3; the third segment is a 1-digit format type code: text 1, table 2, graphic 3, structured field. Class 4, the fourth segment is a 14-digit time code corresponding to year, month, day, hour, minute, and second, the fifth segment is a 4-digit sequence code starting from 0001 and increasing sequentially within the chain. A one-to-one mapping is established between the encoding and the path number, maintaining uniqueness within the same information chain. The system merges the encoding with the record content and sorting relationship to output the final information transmission chain and saves it. A field integrity ratio threshold of 95 is used to ensure record readability and verifiability. A table row limit of 1000 and a column limit of 100 are used to control rendering performance and processing overhead. A minimum field quantity of 8 is used to ensure sufficient information. Time format uniformity, numerical precision uniformity, and unit uniformity are determined by cross-stage consistency requirements. An empty cell ratio limit of 20% is used to restrict incomplete tables from entering distribution. The risk score range of 0 to 100 is determined by the preceding steps; the group matching thresholds of 80 and 50 are determined by the activation and attention boundaries in the management plan; the distribution priority level mapping and tie-breaking rules are used to ensure stable distribution order; the content clustering similarity threshold of 70, the intra-cluster difference threshold of 30, and the maximum number of iterations of 50 are used to achieve a balance between merging rate and semantic consistency; the keyword candidate quantity of 20, the position weighting values ​​of 20 and 10, and the level weighting values ​​of 30, 20, and 10 are used to highlight the importance of title position and high-level records; the keyword screening threshold of 60 is used to ensure the keyness and conciseness of the retained information; and the length of each segment and the value code table in the encoding are used to ensure that the encoding is parsable and traceable.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the long-term stability of tunnel structures based on intelligent detection and monitoring data, characterized in that, include: S1. Extract tunnel monitoring data from the preset database, perform preliminary classification based on user roles, and process the classified data set using branch condition setting and feature selection mechanisms to obtain the mapped user demand set. S2. Based on the mapped set of user needs, the node splitting rule is used to analyze the differences in needs. If the differences in needs point to management personnel, the security level quantitative indicators are extracted, and the preliminary deconstruction information is determined by information gain calculation. S3. Obtain preliminary deconstruction information, group the mechanical parameters through leaf node output, and if the grouping results show that there are parameter groups with risks exceeding the preset threshold, then reorganize them into a graphical representation of the impact of water pressure fluctuations, and determine the reorganized visualization data. S4. From the reorganized visualized data, obtain complete time series details according to the needs of technical personnel. If the details exceed the preset threshold, add analysis annotations and obtain enhanced technical output through tree depth control. S5. Based on the enhanced technology output, a pruning method is used to branch the path to the decision-making level. By maintaining the suggestion list and comparing risk thresholds, the safety level is quantified and a concise text summary is obtained. S6. Obtain the output set of all user paths from the concise text summary. If the output format needs to be switched, dynamically convert it to the specified format. Determine the final customized report by prioritizing the suggested outputs. S7. Verify compatibility with multiple formats through the final customized report. If the verification is successful, distribute it to the corresponding user group and obtain a complete information transmission chain through text aggregation logic and output simplification. S5 includes: From the enhanced technology output, redundant branches are processed using a decision tree pruning method to obtain a simplified path set; For the simplified path set, cluster analysis is used to group similar paths to obtain path group sets; If the complexity of a path group in the path group set exceeds a preset threshold, the path structure is simplified by a dimensionality reduction method to obtain a simplified path set. Based on the simplified path set, a suggestion list is constructed, and the security level of each path is quantified by comparing preset risk thresholds to obtain a set of security levels; For the set of security levels, a text generation method is used to organize the level descriptions, resulting in a set of structured text. Extract key hierarchical information from a structured text collection, generate a concise text summary, and obtain the final text output; Based on the final text output, a mapping method is used to associate business scenarios to obtain scenario adaptation descriptions.

2. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 1, characterized in that: S1 includes: Tunnel monitoring data, including time series and mechanical parameters, is obtained from a pre-set database. The complete dataset is extracted using a standardized protocol to obtain the original monitoring dataset. The original monitoring dataset is classified according to the user role. If the user role is an administrator, all time series and mechanical parameters are extracted. If the user role is a technician, a subset under specific geological conditions is extracted to obtain the classified dataset. For the classified dataset, a branching condition mechanism is adopted: if the time series integrity is higher than a preset threshold, the complete data is retained; if it is lower than the threshold, the missing data is removed to obtain the filtered dataset. The filtered dataset is processed by a feature selection mechanism, and the key features in the time series and mechanical parameters are extracted by the principal component analysis algorithm to obtain the feature-extracted dataset. Based on the dataset after feature extraction, the K-means clustering algorithm is used to group the data, determine the user demand mapping relationship of each group of data, and obtain the initial mapping set; For the initial mapping set, a decision tree algorithm is used for optimization. If the accuracy of the mapping set is higher than a preset threshold, the mapping relationship is retained; if it is lower than the threshold, the feature weights are adjusted to obtain the optimized mapping set. The optimized mapping set is used to generate the final user requirement set, and the tunnel monitoring data mapping results that meet the user roles are determined.

3. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 1, characterized in that: S2 includes: The system obtains demand difference data from the user demand set, classifies the difference data using node splitting rules, and extracts security level quantitative indicators if the difference data points to management personnel. The system also determines the preliminary deconstruction information through information gain calculation, thus obtaining the deconstruction information set. For the deconstructed information set, the random forest algorithm is used to group the security level quantification indicators. If the accuracy of the grouping result is higher than the preset threshold, the grouped data is retained; if it is lower than the preset threshold, abnormal data is removed to obtain the grouped data set. The security level quantitative indicators are obtained from the grouped dataset, and the indicator data is formatted using a standardized protocol to obtain a formatted indicator set. For the formatted indicator set, the key features of the security level quantitative indicators are extracted by the principal component analysis algorithm to obtain the feature-extracted data set. Key feature data are obtained from the feature extraction dataset, and cluster analysis is used to group the feature data to determine the priority of each group of features, thus obtaining a priority group set. For a priority group set, if the group priority is higher than a preset threshold, a security level assessment result is generated through weighted calculation, resulting in an assessment result set. The security level assessment results are obtained from the assessment result set, and the final security level recommendation set is generated by using mapping rules to determine the security level mapping results that meet the needs of management personnel.

4. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 1, characterized in that: S3 includes: Preliminary deconstruction information is obtained from the user demand set, and noisy data is removed through data cleaning methods to obtain a cleaned deconstruction information set. The cleaned deconstruction information set is classified using a decision tree algorithm, and mechanical parameter groups are generated based on the leaf node outputs to obtain a set of mechanical parameter groups. If there are parameter groups in the mechanical parameter grouping set whose risk exceeds a preset threshold, then cluster analysis is used to further subdivide the parameter groups whose risk exceeds the preset threshold to obtain a subdivided risk data set. For the segmented risk data set, a data recombination method is used to extract water pressure fluctuation data to obtain a water pressure fluctuation data set. Key fluctuation characteristics are obtained from the water pressure fluctuation data set, and formatted fluctuation data is generated through standardization processing to obtain the formatted fluctuation data set; For the formatted fluctuation data set, a visualization transformation method is used to generate chart-like data on water pressure fluctuations, resulting in a visualized data set; If the visualized data set meets the preset risk assessment criteria, the final security level mapping result is determined through weighted calculation, resulting in a security level mapping set.

5. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 1, characterized in that: S4 includes: From time series data, a sliding window method is used to extract local time segments to obtain a set of time segments; For a set of time segments, the Fourier transform method is used to decompose the frequency components to obtain a set of frequency components; If the proportion of frequency components in the frequency component set whose frequency exceeds the set frequency threshold exceeds the preset threshold, then the noise of the frequency components whose frequency exceeds the set frequency threshold is removed by filtering method to obtain the filtered frequency set. Based on the filtered frequency set, a smooth time series is generated using a time series reconstruction method, resulting in a smooth time series set. From the smoothed time series set, trend change features are extracted to obtain a trend feature set; For a set of trend features, a classification method is used to determine the stability of the trend, resulting in a set of stability classifications. Based on the stability classification set, output data with technical annotations is generated to obtain the final technical output set.

6. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 1, characterized in that: S6 includes: From the output set of user paths, obtain all path data, and use cluster analysis to group the path data to obtain a path group set; For a set of path groups, if the format of a certain path group does not match the target format, the path data is adjusted according to the preset format conversion rules to obtain a set of paths with the same format. Based on a set of paths with a consistent format, extract the suggested priority attribute for each path, and then sort them from highest to lowest priority using a sorting algorithm to obtain a sorted set of paths.

7. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 6, characterized in that: S6 further includes: From the sorted set of paths, retrieve the highest priority path data, associate it with a preset business template through a mapping method, and generate customized output data; For customized output data, if the data volume exceeds a preset threshold, the data structure is simplified by dimensionality reduction to obtain a streamlined output set. Based on the simplified output set, the output content is organized using a text generation method to obtain the final output. Extract key information fields from the final output and save them using a structured storage method to obtain persistent output results.

8. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 1, characterized in that: S7 includes: Format data is obtained from the final customized report, and the structural features of each format are extracted using format parsing methods to obtain a structured format dataset; If the structured dataset does not match the preset format standard, the data is adjusted through format conversion rules to obtain a standardized dataset. For standardized format datasets, a consistency verification algorithm is used to check the data integrity and standardization, resulting in a set of verified formats.

9. The method for long-term stability evaluation of tunnel structures based on intelligent detection and monitoring data according to claim 8, characterized in that: The S7 also includes: Based on the validated set of formats, user group analysis methods are used to match group attributes and generate a distribution mapping table corresponding to the group. The distribution data is obtained from the distribution mapping table, and similar information is integrated using content clustering methods to obtain an aggregated information set. For the aggregated information set, a keyword extraction method is used to filter information with a priority higher than a preset threshold, resulting in a simplified information set; Information chains are extracted from a simplified information set, and then organized and output using structured coding methods to obtain the final information transmission chain.

Images