A method for detecting leakage anomalies based on the ternary correlation of surface, geological layer, and tunnel
By using the SGT-LADNet anomaly detection model, combined with short-time modeling and long-term path modules, the problems of lag in correlation identification and misjudgment in existing leakage anomaly detection methods are solved, achieving efficient, accurate identification and timely response to tunnel leakage anomalies.
Patent Information
- Application Number
- CN202611213674.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for detecting leakage anomalies cannot simultaneously reflect the correlation between surface water input, geological seepage conditions, and tunnel structural response, resulting in lag and misjudgment in leakage anomaly identification, which fails to effectively improve tunnel operation safety and maintenance efficiency.
The SGT-LADNet anomaly detection model is adopted, and an anomaly detection method for multivariate leakage monitoring data is constructed by combining short-time modeling module and long-term path module. This includes short-window mean calculation, deviation change magnitude matrix generation, bidirectional crystallization intensity construction, directed transmission matrix and initial short-time anomaly field, long-term drift skeleton value extraction and hysteresis transmission relationship graph. Finally, leakage monitoring anomaly results are generated through fully connected layer.
It improves the accuracy and timeliness of leakage anomaly identification, can identify short-term abnormal fluctuations of a single variable and the relationship between different variables, capture long-term trend shifts, reduce the false judgment rate, and enhance the reliability and interpretability of detection results.
Smart Images

Figure CN123116052A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of anomaly detection, specifically relating to a leakage anomaly detection method based on the ternary correlation of the surface, geological layer, and tunnel. Background Technology
[0002] During long-term operation, tunnel projects are susceptible to factors such as surface rainfall and water accumulation, changes in groundwater levels, and aging of the lining structure, leading to problems such as localized leakage, crack expansion, and structural damping damage. These issues affect tunnel operation safety and maintenance efficiency. Existing leakage anomaly detection methods mostly rely on manual inspections, single-sensor monitoring, or image recognition within the tunnel. These methods struggle to simultaneously reflect the correlation between surface water input, geological seepage conditions, and tunnel structural responses, resulting in delays and misjudgments in leakage anomaly identification. Therefore, a leakage anomaly detection method that integrates the spatial characteristics of the surface, geological layers, and tunnel structure is needed to improve the accuracy and timeliness of tunnel leakage anomaly identification.
[0003] In the existing technology, there are already some invention patents that apply data analysis and machine learning methods to detection. For example, patent publication number CN121917753A discloses a method and system for detecting the quality of highway pavement. It acquires multi-source data of the pavement through three-dimensional ground penetrating radar, falling weight deflectometer, coplanar capacitance sensor and environmental parameters, and outputs the road health index using deep learning model, temperature-dielectric hybrid model, fatigue damage model and AFG-KAN model. This scheme is suitable for identifying highway pavement defects, inverting asphalt condition and evaluating road health level. However, its detection object is the formed pavement structure, focusing on pavement service defects such as cracks, poor interlayer structure, looseness and fatigue damage. It does not address the parameter anomalies, variable coordination imbalances and long-term drift problems in the design and preparation process of leakage monitoring. Therefore, it is difficult to directly use it for leakage anomaly detection scenarios.
[0004] Another patent, CN121936996A, discloses a transmission line tower inspection and quality management system. It mainly realizes the informatization of inspection business and automatic report output through modules such as user authentication, permission management, inspection plan, sample information management, inspection data entry and report generation. The advantage of this solution is that it improves the efficiency of inspection records, collaborative editing and report generation. However, its core is still inspection process management. It does not establish an anomaly identification model for inspection data, nor does it characterize the short-term disturbance, collaborative conflict and long-term drift relationship between variables, making it difficult to solve the dynamic identification problem of leakage monitoring anomalies. Summary of the Invention
[0005] This invention provides a leakage anomaly detection method based on the ternary correlation of surface, geological layer and tunnel. For multivariate leakage monitoring data, an SGT-LADNet anomaly detection model is proposed, which consists of a short-time modeling module, a long-term path module and an anomaly judgment module.
[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps: Collect leakage monitoring data, construct datasets, and preprocess them; Construct a short-time modeling module: Input leakage monitoring data to construct a short-window mean, calculate the instantaneous deviation value of the short window to generate a deviation change magnitude matrix; A deviation correlation strength matrix is generated by measuring the changes in same-direction, opposite-direction, and synchronous deviations of leakage monitoring variable pairs; Based on the deviation variation magnitude matrix, a non-negative relational competitive allocation is performed, and bidirectional coupling generates bidirectional crystallization intensity and crystallization mean. An adaptive crystallization threshold is constructed based on bidirectional crystallization intensity to screen effective relation edges, forming a sparse graph of positive correlation between variables; A directed propagation matrix and an initial short-term anomaly field are constructed from the sparse relation graph, the deviation change magnitude matrix, and the short-window instantaneous deviation value matrix. The inter-order changes of the abnormal response values are calculated by propagating them step by step through the directed transmission matrix, and then weighted and fused to generate a short-term anomaly relationship characterization matrix. Based on the short-term anomaly relation representation, the time-series aggregated descriptor is calculated and the order mapping is performed to generate the short-term anomaly relation embedding output; The long-term path construction module extracts long-term drift skeleton values from long-term historical sequences and generates a hysteresis propagation relationship graph based on the differences between adjacent segments. Long-term path output is generated based on the strength of the relationship between long-term drift skeleton values and hysteresis propagation. An anomaly detection module is constructed: the short-term anomaly relationship embedding output and the long-term path output are compressed, spliced and fused, and then passed through a fully connected layer to obtain the anomaly detection value. The leakage monitoring anomaly result is obtained by threshold judgment.
[0007] Preferably, relevant data on leakage monitoring are collected, including data on surface subsidence, surface water depth, groundwater level, formation permeability coefficient, formation moisture content, lining crack width, and vapor lining humidity. The original dataset is constructed and preprocessed, and the dataset is divided into two parts in a ratio of 7:2:1.
[0008] Preferably, input leakage monitoring data. , set the first A short window is ,in, For the first The leakage monitoring variable is the first one in the window. The sampled value at each moment. The number of variables monitored for leakage. For the short-term window length, based on each leakage monitoring variable The short-window mean is constructed from the sampled values at all times within the current short window. The specific mathematical model is as follows: ; In the formula, The mean of the short window is taken at any time within the short window. As a benchmark, the first The sampled value of the leakage monitoring variable at this moment is compared with the sampled value of the same leakage monitoring variable at other moments within the short window. Sample values By comparing the sampling deviation values of the same variable at different times one by one, and combining them with the time... With time The sampling time intervals between samples are subjected to sampling time step interval attenuation processing. Combined with the overall deviation of the leakage monitoring variable relative to the short window mean throughout the entire short window, short window fluctuation amplitude normalization processing is performed to obtain the leakage monitoring variable at time [time value missing]. The instantaneous deviation value of the short window, the specific mathematical model is as follows: ; In the formula, It is the stability constant. The time interval smoothing constant, The short-window instantaneous deviation value is then used as the basis for calculating the short-window instantaneous deviation value matrix by iterating through all leakage monitoring variables and time points. According to the chronological order, the changes in the instantaneous deviation values of the same leakage monitoring variable between adjacent moments are compared moment by moment. The change range between the instantaneous deviation value of the current moment and the instantaneous deviation value of the previous moment is determined as the deviation range between adjacent moments. The deviation amplitude between adjacent times is set to a preset initial value at the start time of the short window, forming a matrix of deviation amplitude between adjacent times. .
[0009] Preferably, for any two leakage monitoring variables and Construct a deviation matrix among leakage monitoring variables, and categorize the leakage monitoring variables according to the order of data collection time within a short window. and Instantaneous deviation of the short window at each corresponding moment and Synchronous comparison is performed, and the product value of the two variables that show the same deviation in the same direction at the same acquisition time is retained. This product value is then accumulated sequentially according to the acquisition time order within a short window to obtain the cumulative value of the same deviation in the same direction for the leakage monitoring variable pair. The specific mathematical model is as follows: ; In the formula, This represents the cumulative deviation in the same direction. For short-term window length, For a moment, To maximize the value, the product of the two variables that show a reverse deviation at the same acquisition time is retained, and then accumulated sequentially according to the acquisition time order within the short window to obtain the cumulative reverse deviation value of the leakage monitoring variable pair. The specific mathematical model is as follows: ; In the formula, To determine the cumulative value in the reverse direction, leakage monitoring variables are analyzed according to the order of data collection within the short-term window. and The deviation changes at corresponding times within the short window are compared. When the deviation changes in the same direction at the same acquisition time are consistent, the corresponding product value is retained and accumulated item by item to obtain the cumulative value of synchronous deviation changes of the leakage monitoring variable pair. The specific mathematical model is as follows: ; In the formula, This represents the cumulative value of synchronous deviation changes. They are respectively Constant leakage monitoring variables and The deviation range of change To achieve the minimum value, the difference in deviation between the two variables at each corresponding time point is statistically analyzed time-by-time, and the absolute value of the difference is calculated for all time points to obtain the cumulative value of the deviation difference between the leakage monitoring variables. The specific mathematical model is as follows: ; In the formula, The cumulative value of deviation changes is used as the basis for determining the relationship enhancement component. Then, the cumulative value of same-direction deviation and the cumulative value of synchronous deviation change are used as the relationship suppression component, while the cumulative value of opposite-direction deviation and the cumulative value of synchronous deviation change are used as the relationship inhibition component, combined with the transition synergistic enhancement coefficient. Anisotropic conflict suppression coefficient and the transition mismatch suppression coefficient The enhancement and suppression components are jointly constrained to generate the deviation correlation strength of the leakage monitoring variable pairs. The specific mathematical model is as follows: ; In the formula, To deviate from the correlation strength, For the relationship enhancement part, For the relation suppression part, traverse all variable pairs to form the deviation correlation strength matrix. .
[0010] Preferably, leakage monitoring variables The relationship strength between the variables and other leakage monitoring variables is processed row by row to construct a sparse graph of positive correlations between the variables, and the leakage monitoring variables are fixed. Obtain the correlation strength of all deviations within the corresponding row for the leakage monitoring variable. Only the non-negative relation components are retained. These retained non-negative relation components are then normalized using a competition amplification index, and this normalization is applied to the current leakage monitoring variables. Within the corresponding entire row range, combined with the stability constant Normalization was performed to obtain leakage monitoring variables. Pointing to leakage monitoring variables The positive correlation weight value, specifically the mathematical model is as follows: ; In the formula, The positive correlation weight value. The number of variables monitored for leakage. To amplify the index for competition, For leakage monitoring variables With the The strength of the deviation association between the leakage monitoring variables As an index for leakage monitoring variables, after traversing all leakage monitoring variables, a positive competition allocation matrix is obtained. The positive correlation weight values of any pair of leakage monitoring variables with the leakage monitoring variables point to Positive correlation weight value A two-way coupling is performed, and the two-way crystallization intensity of leakage monitoring variable pairs is generated by combining the stability constant. The specific mathematical model is as follows: ; In the formula, The strength of bidirectional crystallization is used as a variable for leakage monitoring. All bidirectional crystallization in the row Based on this, the crystallization mean of this row is constructed, and the specific mathematical model is as follows: ; In the formula, This represents the average crystallization value.
[0011] Preferably, the average crystallization value is input. With all bidirectional crystallization strength By combining the threshold adjustment coefficient and the dispersion of the bidirectional crystallization intensity of that row relative to the crystallization mean, we obtain The adaptive crystallization threshold corresponding to the row is specifically modeled as follows: ; In the formula, To adapt the crystallization threshold, This is the threshold adjustment coefficient. To determine the number of leakage monitoring variables, an adaptive crystallization threshold set is obtained by traversing all variable rows. The bidirectional crystallization intensity is then correlated with the leakage monitoring variables. Adaptive crystallization threshold and leakage monitoring variables The adaptive crystallization threshold comparison is performed. When the bidirectional crystallization intensity of a leakage monitoring variable pair meets the corresponding crystallization threshold retention condition, the bidirectional crystallization intensity of that leakage monitoring variable pair is retained as the sparse relation edge intensity. After traversing all leakage monitoring variable pairs, a sparse graph of positive correlations between variables is formed. .
[0012] Preferably, a sparse graph of positive correlations between input variables. Deviation variation matrix and short window instantaneous deviation matrix A directed conservation anomaly propagation model is constructed, which, for any pair of leakage monitoring variables, is applied according to the time sequence within a short window to the leakage monitoring variables. and Cross-compare the instantaneous deviation values of short windows between adjacent time points to construct a leading direction quantity. The specific mathematical model is as follows: ; In the formula, For the leading direction quantity, Leakage monitoring variables The instantaneous deviation of the short window at the previous moment and the leakage monitoring variables The cross-correspondence between the instantaneous deviation values of the short window at the current moment. Leakage monitoring variables The instantaneous deviation of the short window at the previous moment and the leakage monitoring variables The cross-correspondence between the instantaneous deviation values of the short window at the current moment. Given the short-term window length, the directional constraint of the leader directional quantity is applied in conjunction with the stability constant to obtain the directional enhancement coefficient. The specific mathematical model is as follows: ; In the formula, For directional enhancement coefficient, It is the stability constant. For the maximum value operator, based on the sparse relation edge strength And directional enhancement coefficients, normalize all relation edges in the row corresponding to each leakage monitoring variable to construct a directed transmission matrix, the specific mathematical model is as follows: ; In the formula, It is a directed transmission matrix. For leakage monitoring variable index, For all-directional enhancement factors, The strength of all sparse relation edges is then combined with the instantaneous deviation value of the short window and the inflection strength. Furthermore, a mutation enhancement coefficient is introduced to construct an initial short-term anomaly field. The specific mathematical model is as follows: ; In the formula, This is the initial abnormal response value. This represents the mutation enhancement coefficient.
[0013] Preferably, according to the preset maximum conduction order Perform stepwise conservation conduction modeling, for the first... The first-order transmission process constructs the current-order abnormal response value based on the previous-order abnormal response value. The current-order abnormal response value is jointly constructed by the previous-order abnormal response portion of this leakage monitoring variable and the abnormal response portions transmitted from other leakage monitoring variables via the directed transmission matrix, and combined with the first-order transmission process... The first-order conduction coefficient is conserved and allocated to the two parts to obtain the second-order conduction coefficient. The specific mathematical model for the order-order abnormal response value is as follows: ; In the formula, For the first Each leakage monitoring variable is in The first moment Level abnormal response value, For the first The conductivity coefficient, its value is... , For the first Several leakage monitoring variables in The first moment Level abnormal response value, For the first Each leakage monitoring variable is in The first moment Level abnormal response value, The number of variables monitored for leakage. It is a directed transmission matrix. This is the first-order abnormal response part of the leakage monitoring variable. For the abnormal response portion of the remaining leakage monitoring variables transmitted through the directed transmission matrix, the overall change between the current abnormal response value and the previous abnormal response value is statistically analyzed for each order of transmission results to obtain the inter-order change. The specific mathematical model is as follows: ; In the formula, The inter-order variation is then normalized and attenuated using an attenuation adjustment coefficient to generate multi-order fusion weights for each order. The specific mathematical model is as follows: ; In the formula, For multi-level fusion weights, It is an exponential function. As the attenuation adjustment coefficient, the anomalous response values of all orders are weighted and fused according to the corresponding multi-order fusion weights to obtain the short-term anomaly relationship representation. After traversing all leakage monitoring variables and time points, a short-term anomaly relationship representation matrix is formed. .
[0014] Preferably, an anomaly condensation power-order expansion method is designed to map the short-term anomaly relationship representation matrix into a multi-order embedding for leakage monitoring variables. At any moment The short-term anomaly relationship is characterized by statistically analyzing the overall amplitude distribution of the short-term anomaly relationship representation of the leakage monitoring variable at all times within the current short window. The overall anomaly amplitude of the leakage monitoring variable within the short window is then used as a normalization reference to extract the target time. The short-term anomaly relationship is characterized by its relative proportion in the overall anomaly amplitude, forming the temporal cluster descriptive quantity at that moment, with a preset... To determine the embedding order, time... The short-term anomaly relationship representation is mapped to the temporal aggregation descriptor through an order-wise association mapping to obtain multi-order embedding values. The specific mathematical model is as follows: ; In the formula, For multi-level embedding values, For short-term window length, This is a temporal aggregation description quantity. It is the stability constant. To characterize short-term anomaly relationships, after traversing leakage monitoring variables, time periods, and embedding orders, a short-term anomaly relationship embedding output is generated. .
[0015] Preferably, input leakage monitoring data. Set time Take the length forward as The time period will be from Time's up Leakage monitoring data over time are combined in chronological order to obtain a long-term historical sequence. The sequence is divided into multiple long-term segments according to a preset time span, and a time index set is established for each segment. Leakage monitoring variables In the Within a long-term segment, data from all time-indexed data of the leakage monitoring variable are statistically analyzed. Then, the changes in the leakage monitoring variable between adjacent moments within the long-term segment are extracted moment-by-moment. A drift compensation coefficient is used to compensate for and adjust the adjacent changes. The segment baseline representation and the compensated adjacent changes are jointly constructed to generate the long-term drift skeleton value. The specific mathematical model is as follows: ; In the formula, This represents the long-term drift skeleton value. It is the stability constant. For time position index within a long segment, To determine the drift compensation coefficients, a long-term drift skeleton matrix is obtained by iterating through all leakage monitoring variables and long-term segments. , For the first In the first long historical sequence, the first Each leakage monitoring variable at time [time] The value, For the same leakage monitoring variable In the previous moment The value is then used to establish hysteresis propagation relationships between adjacent long segments, for the current long segment. Any leakage monitoring variable in And the previous long segment Any leakage monitoring variable Extract leakage monitoring variables Long-term drift skeleton value of the current long-term fragment Leakage monitoring variables Long-term drift skeleton value corresponding to the previous long-term segment The difference between the two is then normalized, and the hysteresis correlation strength between different leakage monitoring variables is constrained by the variable interval attenuation coefficient to generate the hysteresis transmission relationship strength. The specific mathematical model is as follows: ; In the formula, The strength of the hysteresis conduction relationship, The variable interval decay coefficient, As an exponential function, after traversing all combinations of leakage monitoring variables and long-term segments, a hysteresis transmission relationship diagram is formed between long-term segments. .
[0016] Preferably, a long-term abnormal trend embedding is constructed for any leakage monitoring variable. In the Long-term drift skeleton value under a long-term segment Combining the long-term drift skeleton value of this leakage monitoring variable in the current long-term segment with the hysteresis propagation relationship diagram of all leakage monitoring variables in the previous long-term segment, The skeleton response after conduction is used, and a hysteresis conduction fusion coefficient is introduced to fuse the two parts of information to obtain the long-term abnormal trend embedding. The specific mathematical model is as follows: ; In the formula, Embedded for long-term abnormal trends The hysteresis conduction fusion coefficient, The number of variables monitored for leakage. The strength of the hysteresis conduction relationship, This represents the long-term drift skeleton value corresponding to the previous long-term segment. The long-term path output is formed by traversing all leakage monitoring variables and long-term segments to obtain the skeleton response after propagation. .
[0017] Preferably, the short-term anomaly relationship is embedded in the output. Compression is performed based on the dimension of the leakage monitoring variable. Within the entire embedding order and time range corresponding to the leakage monitoring variable, the embedding values of each multi-order are... Overall statistics were performed, and short-time compression features were obtained after equalization. Then output the long-term path. Compression was performed based on the leakage monitoring variable dimension, and long-term abnormal trends were embedded across the entire long-term segment. Overall statistics were performed, and long-term compression characteristics were obtained after equalization. The fused input vector is constructed using the following mathematical model: ; In the formula, To fuse the input vectors, For the splicing operation, the fused input vector is input into the fully connected layer for anomaly detection to obtain the leakage monitoring anomaly detection value. The specific mathematical model is as follows: ; In the formula, For abnormal detection values in leakage monitoring, , These are the weight matrix and bias vector of the first fully connected layer, respectively. , These are the weight matrix and bias vector of the second fully connected layer, respectively. It is the ReLU activation function. For the Sigmoid function, a preset threshold is used. The abnormal detection values of leakage monitoring are determined to obtain the abnormal leakage monitoring results. .
[0018] The advantages of the technical effects provided by the present invention in the above technical solution are as follows: The short-term modeling module of this invention can extract key information from short-term window data of leakage monitoring. It can not only identify abnormal fluctuations of a single variable in a short period of time, but also further characterize the relationship between different variables. Through the sparse graph of positive correlation between variables and the directed conservation anomaly transmission mechanism, it can filter out weak correlations and invalid relationship edges, retain variable relationship paths that are more sensitive to anomaly propagation, and improve the accuracy and anti-interference ability of short-term anomaly feature expression. Finally, through multi-level fusion and embedding output, the model can take into account the influence of local mutations and cross-variable transmission, providing more stable and discriminative short-term anomaly relationship features for subsequent anomaly judgment.
[0019] The long-term path module of this invention can extract segment-level change features from long-term historical sequences of leakage monitoring and establish hysteresis transmission relationships between variables between adjacent long-term segments. Instead of judging anomalies solely based on the current short-term window, it further incorporates the overall response level within the long-term segment, adjacent change information, and drift compensation results, enabling the model to capture long-term trend shifts. By depicting the delayed impact of the previous long-term segment on the current segment through a hysteresis transmission relationship graph, it can identify hysteresis correlations and cross-segment transmission phenomena between different variables, avoiding misjudging short-term random fluctuations as anomalies. The resulting long-term path output provides a more stable trend basis for the anomaly judgment module, enhancing the model's ability to identify chronic drift, latent anomalies, and staged shifts, and improving the reliability of the overall detection results.
[0020] The anomaly detection module of this invention can uniformly compress and fuse short-term anomaly relationship embedding output and long-term path output, and generate leakage monitoring anomaly detection results through a fully connected layer. It incorporates short-term features within a short window and long-term features from long-term historical segments into the judgment process, avoiding misjudgment or missed judgment due to reliance on a single path. By compressing and fusing input vectors through variable dimensions, it can reduce redundant information brought by multi-order embedding and long-term segment output, making anomaly detection more focused and stable. By using a fully connected layer to perform nonlinear discrimination on fused features, it can improve the model's ability to distinguish complex anomaly patterns, and finally output clear anomaly detection values and anomaly results, improving the completeness, interpretability, and reliability of the detection process in engineering applications. Attached Figure Description
[0021] Figure 1 This is a step diagram of a leakage anomaly detection method based on the ternary correlation of the surface, geological layer, and tunnel.
[0022] Figure 2 This is a short-term modeling module.
[0023] Figure 3This is a long-term path module.
[0024] Figure 4 This is an error distribution diagram of the SGT-LADNet anomaly detection model.
[0025] Figure 5 The image shows the performance of the SGT-LADNet anomaly detection model.
[0026] Figure 6 Histogram of the distribution of comprehensive judgment indexes for the SGT-LADNet anomaly detection model. Detailed Implementation
[0027] This invention proposes a method for detecting leakage anomalies based on a ternary correlation of the surface, geological layers, and tunnels, the steps of which are as follows: Figure 1 As shown, an SGT-LADNet anomaly detection model is proposed for multivariate leakage monitoring data. It consists of a short-time modeling module, a long-term path module, and an anomaly judgment module. The short-time modeling module is used to extract the instantaneous fluctuations of variables and the abnormal correlations between variables within a short window, and to characterize the local mutation features. The long-term path module characterizes the long-term segment drift and hysteresis transmission relationship, and forms the long-term path output. The anomaly judgment module integrates the two types of outputs to generate anomaly detection values and anomaly results.
[0028] Collect leakage monitoring data, construct a dataset, and preprocess it.
[0029] Furthermore, the monitoring sections were selected from areas along the tunnel that were prone to leakage, areas with concentrated deformation, areas rich in groundwater, and normal areas. The spacing between monitoring sections was set at 20m, increasing to 10m in areas with suspected leakage or significant geological changes. Each section had surface monitoring points, geological layer monitoring points, and tunnel monitoring points. Surface settlement data was collected using an automated leveling instrument or a hydrostatic leveling instrument, in mm, with a range of -50mm to 50mm and a collection accuracy of no less than 0.1mm. The sampling frequency was set to once daily, adjusted to once every 6 hours after rain or during construction disturbances. Surface water depth data was collected using an ultrasonic level gauge or a pressure-type water depth sensor, in cm, with a range of 0cm to 50cm and a collection accuracy of [missing information]. The minimum depth is 0.1 cm, with a sampling frequency of once every 30 minutes, increasing to once every 10 minutes during rainfall. Groundwater level data is collected using pore water pressure gauges or automatic groundwater level monitoring instruments, in meters (m). The range is set from 0m to 30m depending on the burial depth, with a sampling accuracy of no less than 0.01m, and a sampling frequency of once every hour. Formation permeability coefficient data is obtained through field borehole water injection tests, pumping tests, or indoor soil sample permeability tests, in m / s, with a range of 10^-8 m / s to 10^-3 m / s. At least one formation test point is set up at each monitoring section, and the permeability coefficient is recorded separately for different strata at the same section. During the routine phase, retesting is conducted monthly, and during the rainy season or when groundwater levels are abnormal. During the change phase, measurements were repeated every 15 days. Formation moisture content data were collected using a TDR soil moisture sensor or an FDR moisture sensor, in percentage (%), with a range of 0% to 60%, an accuracy of at least ±2%, and a sampling frequency of once every 1 hour. Lining crack width data were collected using crack gauges, displacement gauges, or tunnel wall image recognition, in mm, with a range of 0mm to 10mm, an accuracy of at least 0.01mm, and a sampling frequency of once daily, with increased frequency every 6 hours for existing and newly added cracks. Lining humidity data were collected using a wall humidity sensor or an integrated temperature and humidity sensor, in %RH, with a range of 0%RH to 100%RH, and an accuracy of at least... ±3%RH, sampling frequency set to once every 30 minutes; all acquisition devices are uniformly connected to the data acquisition terminal, timestamp accuracy is uniform to the minute level, and data from different sampling frequencies are aligned according to a 1-hour time window. For continuous data within the window, the average value, maximum value and rate of change are used as the model input features. Anomaly labels are determined based on on-site inspection records, water leakage traces, lining wet area and manual verification results. Normal samples are marked as 0, and leakage anomaly samples are marked as 1. Finally, each sample is composed of surface settlement, surface water depth, groundwater level, formation permeability coefficient, formation water content, lining crack width, lining humidity and leakage anomaly label, and is divided into training set, validation set and test set in a ratio of 7:2:1.
[0030] Build short-time modeling modules, such as Figure 2 As shown, the module is as follows: inputting leakage monitoring data to construct a short window mean, and calculating the instantaneous deviation value of the short window to generate a deviation change magnitude matrix.
[0031] Furthermore, input relevant leakage monitoring data. , set the first A short window is ,in, For the first The leakage monitoring variable is the first one in the window. The sampled value at each moment. The number of variables monitored for leakage. For the short-term window length, based on each leakage monitoring variable The short-window mean is constructed from the sampled values at all times within the current short window. The specific mathematical model is as follows: ; In the formula, The mean of the short window is taken at any time within the short window. As a benchmark, the first The sampled value of the leakage monitoring variable at this moment is compared with the sampled value of the same leakage monitoring variable at other moments within the short window. Sample values By comparing the sampling deviation values of the same variable at different times one by one, and combining them with the time... With time The sampling time intervals between samples are subjected to sampling time step interval attenuation processing. Combined with the overall deviation of the leakage monitoring variable relative to the short window mean throughout the entire short window, short window fluctuation amplitude normalization processing is performed to obtain the leakage monitoring variable at time [time value missing]. The instantaneous deviation value of the short window, the specific mathematical model is as follows: ; In the formula, It is the stability constant. The time interval smoothing constant, The short-window instantaneous deviation value is then used as the basis for calculating the short-window instantaneous deviation value matrix by iterating through all leakage monitoring variables and time points. According to the chronological order, the changes in the instantaneous deviation values of the same leakage monitoring variable between adjacent moments are compared moment by moment. The change range between the instantaneous deviation value of the current moment and the instantaneous deviation value of the previous moment is determined as the deviation range between adjacent moments. The deviation amplitude between adjacent times is set to a preset initial value at the start time of the short window, forming a matrix of deviation amplitude between adjacent times. .
[0032] In this embodiment, the short-term window length is preferably set in the range of 12 to 24. In specific implementation, 12 is selected, that is, each short-term window is composed of 12 consecutive moments or 12 batches of leakage monitoring data, and the window length will not be too short; the stability constant is set to 1×10. -6 When calculating the degree of variable deviation, scale constraints, or normalized denominator, this method can avoid the instability caused by a denominator that is zero or close to zero. At the same time, this value is small enough that it will not significantly change the difference relationship between the original data, ensuring that the instantaneous deviation value of the short window is mainly determined by the actual data change. The time interval smoothing constant is preferably set in the range of 0.1 to 0.5. In the specific implementation, 0.25 is selected. This value is suitable for use with a short window of 12 time points, which can take into account the sensitivity of the nearest neighbor response and the suppression effect of far-distance fluctuations. The initial value of the turning point intensity at the beginning of the short window is preset to 0. Since there is no previous time point for comparison at the beginning of the short window, its turning point intensity is set to 0.
[0033] The deviation correlation strength matrix is generated by measuring the changes in same-direction deviation, opposite-direction deviation, and synchronous deviation of leakage monitoring variable pairs.
[0034] Furthermore, for any two leakage monitoring variables and Construct a deviation matrix among leakage monitoring variables, and categorize the leakage monitoring variables according to the order of data collection time within a short window. and Instantaneous deviation of the short window at each corresponding moment and Synchronous comparison is performed, and the product value of the two variables that show the same deviation in the same direction at the same acquisition time is retained. This product value is then accumulated sequentially according to the acquisition time order within a short window to obtain the cumulative value of the same deviation in the same direction for the leakage monitoring variable pair. The specific mathematical model is as follows: ; In the formula, This represents the cumulative deviation in the same direction. For short-term window length, For a moment, To maximize the value, the product of the two variables that show a reverse deviation at the same acquisition time is retained, and then accumulated sequentially according to the acquisition time order within the short window to obtain the cumulative reverse deviation value of the leakage monitoring variable pair. The specific mathematical model is as follows: ; In the formula, To determine the cumulative value in the reverse direction, leakage monitoring variables are analyzed according to the order of data collection within the short-term window. and The deviation changes at corresponding times within the short window are compared. When the deviation changes in the same direction at the same acquisition time are consistent, the corresponding product value is retained and accumulated item by item to obtain the cumulative value of synchronous deviation changes of the leakage monitoring variable pair. The specific mathematical model is as follows: ; In the formula, This represents the cumulative value of synchronous deviation changes. They are respectively Constant leakage monitoring variables and The deviation range of change To achieve the minimum value, the difference in deviation between the two variables at each corresponding time point is statistically analyzed time-by-time, and the absolute value of the difference is calculated for all time points to obtain the cumulative value of the deviation difference between the leakage monitoring variables. The specific mathematical model is as follows: ; In the formula, The cumulative value of deviation changes is used as the basis for determining the relationship enhancement component. Then, the cumulative value of same-direction deviation and the cumulative value of synchronous deviation change are used as the relationship suppression component, while the cumulative value of opposite-direction deviation and the cumulative value of synchronous deviation change are used as the relationship inhibition component, combined with the transition synergistic enhancement coefficient. Anisotropic conflict suppression coefficient and the transition mismatch suppression coefficient The enhancement and suppression components are jointly constrained to generate the deviation correlation strength of the leakage monitoring variable pairs. The specific mathematical model is as follows: ; In the formula, To deviate from the correlation strength, For the relationship enhancement part, For the relation suppression part, traverse all variable pairs to form the deviation correlation strength matrix. .
[0035] In this embodiment, the transition synergy enhancement coefficient is preferably set in the range of 1.0 to 3.0, and 1.20 is selected in the specific implementation. When two leakage monitoring variables show synchronous transitions within the same short-term window, the contribution of such common changes to the overall deviation correlation strength can be appropriately amplified, making it easier for the model to capture synergistic abnormal changes. The anisotropic conflict suppression coefficient is preferably set in the range of 0.5 to 1.0, and 0.85 is selected in the specific implementation. When two variables show opposite imbalance directions at the same time, their relationship strength can be significantly weakened, avoiding the incorrect classification of variable fluctuations with opposite directions into a stable synergistic relationship. The transition mismatch suppression coefficient is preferably set in the range of 0.2 to 1.2, and 0.65 is selected in the specific implementation. When the transition strength of two variables differs greatly, the overall deviation correlation strength can be more strongly suppressed, reducing the interference of asynchronous mutations or isolated fluctuations on variable relationship modeling.
[0036] Based on the deviation variation magnitude matrix, a non-negative relational competitive allocation is performed, and bidirectional coupling generates bidirectional crystallization intensity and crystallization mean.
[0037] Furthermore, regarding leakage monitoring variables The relationship strength between the variables and other leakage monitoring variables is processed row by row to construct a sparse graph of positive correlations between the variables, and the leakage monitoring variables are fixed. Obtain the correlation strength of all deviations within the corresponding row for the leakage monitoring variable. Only the non-negative relation components are retained. These retained non-negative relation components are then normalized using a competition amplification index, and this normalization is applied to the current leakage monitoring variables. Within the corresponding entire row range, combined with the stability constant Normalization was performed to obtain leakage monitoring variables. Pointing to leakage monitoring variables The positive correlation weight value, specifically the mathematical model is as follows: ; In the formula, The positive correlation weight value. The number of variables monitored for leakage. To amplify the index for competition, For leakage monitoring variables With the The strength of the deviation association between the leakage monitoring variables As an index for leakage monitoring variables, after traversing all leakage monitoring variables, a positive competition allocation matrix is obtained. The positive correlation weight values of any pair of leakage monitoring variables with the leakage monitoring variables point to Positive correlation weight value A two-way coupling is performed, and the two-way crystallization intensity of leakage monitoring variable pairs is generated by combining the stability constant. The specific mathematical model is as follows: ; In the formula, The strength of bidirectional crystallization is used as a variable for leakage monitoring. All bidirectional crystallization in the row Based on this, the crystallization mean of this row is constructed, and the specific mathematical model is as follows: ; In the formula, This represents the average crystallization value.
[0038] In this embodiment, the competition amplification index is preferably set in the range of 1.0 to 5.0. In specific implementation, 2.0 is selected. When the overall deviation correlation strength between a certain variable and other variables is high, an index of 2.0 can square amplify the strong relationship, making the variable relationship with true cooperative anomaly characteristics more likely to stand out in the competitive allocation.
[0039] An adaptive crystallization threshold is constructed based on bidirectional crystallization intensity to screen effective relation edges, forming a sparse graph of positive correlation between variables.
[0040] Furthermore, input the average crystallization value. With all bidirectional crystallization strength By combining the threshold adjustment coefficient and the dispersion of the bidirectional crystallization intensity of that row relative to the crystallization mean, we obtain The adaptive crystallization threshold corresponding to the row is specifically modeled as follows: ; In the formula, To adapt the crystallization threshold, This is the threshold adjustment coefficient. To determine the number of leakage monitoring variables, an adaptive crystallization threshold set is obtained by traversing all variable rows. The bidirectional crystallization intensity is then correlated with the leakage monitoring variables. Adaptive crystallization threshold and leakage monitoring variables The adaptive crystallization threshold comparison is performed. When the bidirectional crystallization intensity of a leakage monitoring variable pair meets the corresponding crystallization threshold retention condition, the bidirectional crystallization intensity of that leakage monitoring variable pair is retained as the sparse relation edge intensity. After traversing all leakage monitoring variable pairs, a sparse graph of positive correlations between variables is formed. .
[0041] In this embodiment, the threshold adjustment coefficient is preferably set in the range of 0.5 to 1.0. In a specific implementation, 0.70 is selected. When the threshold adjustment coefficient is 0.70, the adaptive crystallization threshold will take into account the dispersion of the relationship strength based on the average bidirectional crystallization strength of the variable row, so that the variable relationship that is significantly higher than the ordinary correlation level is retained, while weak correlation, accidental correlation or low-intensity relationship caused by noise is eliminated.
[0042] A directed transmission matrix and an initial short-term anomaly field are constructed from the sparse relation graph, the deviation change magnitude matrix, and the short-window instantaneous deviation value matrix.
[0043] Furthermore, a sparse graph showing positive correlations between input variables. Deviation variation matrix and short window instantaneous deviation matrix A directed conservation anomaly propagation model is constructed, which, for any pair of leakage monitoring variables, is applied according to the time sequence within a short window to the leakage monitoring variables. and Cross-compare the instantaneous deviation values of short windows between adjacent time points to construct a leading direction quantity. The specific mathematical model is as follows: ; In the formula, For the leading direction quantity, Leakage monitoring variables The instantaneous deviation of the short window at the previous moment and the leakage monitoring variables The cross-correspondence between the instantaneous deviation values of the short window at the current moment. Leakage monitoring variables The instantaneous deviation of the short window at the previous moment and the leakage monitoring variables The cross-correspondence between the instantaneous deviation values of the short window at the current moment. Given the short-term window length, the directional constraint of the leader directional quantity is applied in conjunction with the stability constant to obtain the directional enhancement coefficient. The specific mathematical model is as follows: ; In the formula, For directional enhancement coefficient, It is the stability constant. For the maximum value operator, based on the sparse relation edge strength And directional enhancement coefficients, normalize all relation edges in the row corresponding to each leakage monitoring variable to construct a directed transmission matrix, the specific mathematical model is as follows: ; In the formula, It is a directed transmission matrix. For leakage monitoring variable index, For all-directional enhancement factors, The strength of all sparse relation edges is then combined with the instantaneous deviation value of the short window and the inflection strength. Furthermore, a mutation enhancement coefficient is introduced to construct an initial short-term anomaly field. The specific mathematical model is as follows: ; In the formula, This is the initial abnormal response value. This represents the mutation enhancement coefficient.
[0044] In this embodiment, the mutation enhancement coefficient is preferably set in the range of 0.3 to 1.3. In a specific implementation, 0.80 is selected. When a certain leakage monitoring variable shows a significant turning point within a short window, a mutation enhancement coefficient of 0.80 can moderately increase the influence of the turning point intensity on the initial short-term anomaly field, making it easier to amplify and identify short-term anomalies such as abnormal disturbances.
[0045] The inter-order changes of the abnormal response values are calculated by propagating them step by step through the directed transmission matrix, and then weighted and fused to generate a short-term anomaly relationship characterization matrix.
[0046] Furthermore, according to the preset maximum conduction order Perform stepwise conservation conduction modeling, for the first... The first-order transmission process constructs the current-order abnormal response value based on the previous-order abnormal response value. The current-order abnormal response value is jointly constructed by the previous-order abnormal response portion of this leakage monitoring variable and the abnormal response portions transmitted from other leakage monitoring variables via the directed transmission matrix, and combined with the first-order transmission process... The first-order conduction coefficient is conserved and allocated to the two parts to obtain the second-order conduction coefficient. The specific mathematical model for the order-order abnormal response value is as follows: ; In the formula, For the first Each leakage monitoring variable is in The first moment Level abnormal response value, For the first The conductivity coefficient, its value is... , For the first Each leakage monitoring variable is in The first moment Level abnormal response value, For the first Each leakage monitoring variable is in The first moment Level abnormal response value, The number of variables monitored for leakage. It is a directed transmission matrix. This is the first-order abnormal response part of the leakage monitoring variable. For the abnormal response portion of the remaining leakage monitoring variables transmitted through the directed transmission matrix, the overall change between the current abnormal response value and the previous abnormal response value is statistically analyzed for each order of transmission results to obtain the inter-order change. The specific mathematical model is as follows: ; In the formula, The inter-order variation is then normalized and attenuated using an attenuation adjustment coefficient to generate multi-order fusion weights for each order. The specific mathematical model is as follows: ; In the formula, For multi-level fusion weights, It is an exponential function. As the attenuation adjustment coefficient, the anomalous response values of all orders are weighted and fused according to the corresponding multi-order fusion weights to obtain the short-term anomaly relationship representation. After traversing all leakage monitoring variables and time points, a short-term anomaly relationship representation matrix is formed. .
[0047] In this embodiment, the maximum transmission order is preset to 3. The first-order transmission can reflect the direct abnormal influence between variables, the second-order transmission can reflect the indirect diffusion of abnormalities through intermediate variables, and the third-order transmission can further capture deeper chain-like abnormal correlations. For leakage monitoring and abnormal detection scenarios, there are a large number of variables, and setting it to 3 orders can cover the main short-term abnormal propagation paths. The attenuation adjustment coefficient is preferably set in the range of 0.4 to 1.4. In the specific implementation, 0.75 is selected. 0.75 can moderately attenuate the results of higher-order transmission, so that the direct abnormal relationship in the lower-order transmission occupies the main weight, while retaining the indirect abnormal relationship with practical significance in the second-order and third-order transmission.
[0048] Based on the short-term anomaly relation representation, the time-series aggregated descriptor is calculated and the order mapping is performed to generate the short-term anomaly relation embedding output.
[0049] Furthermore, an anomaly condensation power-order expansion method is designed to map the short-term anomaly relationship representation matrix into a multi-order embedding for leakage monitoring variables. At any moment The short-term anomaly relationship is characterized by statistically analyzing the overall amplitude distribution of the short-term anomaly relationship representation of the leakage monitoring variable at all times within the current short window. The overall anomaly amplitude of the leakage monitoring variable within the short window is then used as a normalization reference to extract the target time. The short-term anomaly relationship is characterized by its relative proportion in the overall anomaly amplitude, forming the temporal cluster descriptive quantity at that moment, with a preset... To determine the embedding order, time... The short-term anomaly relationship representation is mapped to the temporal aggregation descriptor through an order-wise association mapping to obtain multi-order embedding values. The specific mathematical model is as follows: ; In the formula, For multi-level embedding values, For short-term window length, This is a temporal aggregation description quantity. It is the stability constant. To characterize short-term anomaly relationships, after traversing leakage monitoring variables, time periods, and embedding orders, a short-term anomaly relationship embedding output is generated. .
[0050] In this embodiment, the preset embedding order is set to 3. When the embedding order is set to 3, the basic response, enhanced response and high-order condensed response representing the short-term anomaly relationship can be retained respectively, so that the local aggregation degree of leakage monitoring anomalies within the short-term window can be more fully expressed.
[0051] Build long-term path modules, such as Figure 3 As shown, the module is: constructing a long-term historical sequence to extract long-term drift skeleton values, and generating a hysteresis transmission relationship diagram based on the differences between adjacent segments.
[0052] Furthermore, input relevant leakage monitoring data. Set time Take the length forward as The time period will be from Time's up Leakage monitoring data over time are combined in chronological order to obtain a long-term historical sequence. The sequence is divided into multiple long-term segments according to a preset time span, and a time index set is established for each segment. Leakage monitoring variables In the Within a long-term segment, data from all time-indexed data of the leakage monitoring variable are statistically analyzed. Then, the changes in the leakage monitoring variable between adjacent moments within the long-term segment are extracted moment-by-moment. A drift compensation coefficient is used to compensate for and adjust the adjacent changes. The segment baseline representation and the compensated adjacent changes are jointly constructed to generate the long-term drift skeleton value. The specific mathematical model is as follows: ; In the formula, This represents the long-term drift skeleton value. It is the stability constant. For time position index within a long segment, To determine the drift compensation coefficients, a long-term drift skeleton matrix is obtained by iterating through all leakage monitoring variables and long-term segments. , For the first In the first long historical sequence, the first Each leakage monitoring variable at time [time] The value, For the same leakage monitoring variable In the previous moment The value is then used to establish hysteresis propagation relationships between adjacent long segments, for the current long segment. Any leakage monitoring variable in And the previous long segment Any leakage monitoring variable Extract leakage monitoring variables Long-term drift skeleton value of the current long-term fragment Leakage monitoring variables Long-term drift skeleton value corresponding to the previous long-term segment The difference between the two is then normalized, and the hysteresis correlation strength between different leakage monitoring variables is constrained by the variable interval attenuation coefficient to generate the hysteresis transmission relationship strength. The specific mathematical model is as follows: ; In the formula, The strength of the hysteresis conduction relationship, The variable interval decay coefficient, As an exponential function, after traversing all combinations of leakage monitoring variables and long-term segments, a hysteresis transmission relationship diagram is formed between long-term segments. .
[0053] In this embodiment, the preset time span is set to 24, meaning that every 24 consecutive sampling moments or consecutive production batches are divided into a long-term segment. 24 moments can cover a relatively complete long-term trend of changes in leakage monitoring process parameters, unlike short-term windows which only focus on local mutations. The drift compensation coefficient is preferably set in the range of 0.2 to 0.5, and in specific implementation, 0.30 is selected. 0.30 can moderately compensate for adjacent changes within the long-term segment. The variable interval attenuation coefficient is preferably set in the range of 0.1 to 0.2, and in specific implementation, 0.15 is selected. When constructing the hysteresis transmission relationship between different leakage monitoring variables, 0.15 can moderately attenuate the relationship between variables with large index intervals, so that variables with closer physical correlation and more direct process connection have higher hysteresis correlation expression.
[0054] Long-term path outputs are generated based on the strength of the relationship between long-term drift skeleton values and hysteresis conduction.
[0055] Furthermore, a long-term abnormal trend embedding is constructed for any leakage monitoring variable. In the Long-term drift skeleton value under a long-term segment Combining the long-term drift skeleton value of this leakage monitoring variable in the current long-term segment with the hysteresis propagation relationship diagram of all leakage monitoring variables in the previous long-term segment, The skeleton response after conduction is used, and a hysteresis conduction fusion coefficient is introduced to fuse the two parts of information to obtain the long-term abnormal trend embedding. The specific mathematical model is as follows: ; In the formula, Embedded for long-term abnormal trends The hysteresis conduction fusion coefficient, The number of variables monitored for leakage. The strength of the hysteresis conduction relationship, This represents the long-term drift skeleton value corresponding to the previous long-term segment. The long-term path output is formed by traversing all leakage monitoring variables and long-term segments to obtain the skeleton response after propagation. .
[0056] In this embodiment, the hysteresis propagation fusion coefficient is preferably set in the range of 0.1 to 1.0. In a specific implementation, 0.40 is selected. 0.40 means that the long-term abnormal trend embedding is based primarily on the long-term drift skeleton value of the current long-term segment itself, while the skeleton response of the previous long-term segment after hysteresis propagation is introduced as a supplement.
[0057] An anomaly detection module is constructed: the short-term anomaly relationship embedding output and the long-term path output are compressed, spliced and fused, and then passed through a fully connected layer to obtain the anomaly detection value. The leakage monitoring anomaly result is obtained by threshold judgment.
[0058] Furthermore, the short-term anomaly relationship is embedded in the output. Compression is performed based on the dimension of the leakage monitoring variable. Within the entire embedding order and time range corresponding to the leakage monitoring variable, the embedding values of each multi-order are... Overall statistics were performed, and short-time compression features were obtained after equalization. Then output the long-term path. Compression was performed based on the leakage monitoring variable dimension, and long-term abnormal trends were embedded across the entire long-term segment. Overall statistics were performed, and long-term compression characteristics were obtained after equalization. The fused input vector is constructed using the following mathematical model: ; In the formula, To fuse the input vectors, For the splicing operation, the fused input vector is input into the fully connected layer for anomaly detection to obtain the leakage monitoring anomaly detection value. The specific mathematical model is as follows: ; In the formula, For abnormal detection values in leakage monitoring, , These are the weight matrix and bias vector of the first fully connected layer, respectively. , These are the weight matrix and bias vector of the second fully connected layer, respectively. It is the ReLU activation function. For the Sigmoid function, a preset threshold is used. The abnormal detection values of leakage monitoring are determined to obtain the abnormal leakage monitoring results. .
[0059] Furthermore, the SGT-LADNet anomaly detection model was written in Python, and the experiments were run on a Windows operating system. PyTorch was used as the framework in the CUDA 11.27 environment, and training was performed on a GeForce RTX 3090. The optimizer used was Adam, the initial learning rate was set to 0.001, the training batch size was set to 64, the training period was set to 100, and the dataset consisted of 60 days of leakage monitoring data, which was preprocessed and then input into the SGT-LADNet anomaly detection model.
[0060] Furthermore, the SGT-LADNet anomaly detection model demonstrates good detection stability and discriminative ability in leakage monitoring anomaly detection tasks, such as... Figure 4 As shown in the error distribution plot, the prediction error is mainly concentrated near zero, with an average error of approximately -0.010, a MAE of 0.044, and a STD of 0.076. This indicates that the overall deviation between the model's predicted values and the actual anomaly detection values is small, and the error fluctuation range is relatively concentrated. Only a small number of samples show large positive or negative deviations, indicating that the model fits well on most samples. The average error is slightly less than 0, suggesting that the model slightly underestimates the severity of some anomalies, but the underestimation is small and will not significantly affect the overall anomaly judgment. Figure 5In the anomaly detection results shown, the predicted anomaly detection values and the actual anomaly detection values generally follow the same trend. Within the normal sample range, both remain below the 0.5 anomaly threshold. However, in the anomaly ranges approximately 23-28, 48-56, and 77-83, the predicted values rise synchronously with the actual anomaly values and exceed the threshold, indicating that the model can effectively identify anomalies. Furthermore, while the predicted curve shows slight differences from the actual curve near the anomaly peak, the overall peak position and duration largely match, demonstrating that the model can not only determine whether anomalies exist but also reflect the changing trend of the anomaly degree well. Figure 6 The histogram showing the distribution of the comprehensive judgment index reveals a clear bimodal distribution. Most normal samples are concentrated around 0.15 to 0.42, while abnormal samples are mainly concentrated around 0.60 to 0.90. The anomaly threshold of 0.5 falls between the two sample distributions, demonstrating good separation. The mean value in the figure is 0.401, the standard deviation is 0.228, the minimum value is 0.070, and the maximum value is 0.982. This indicates that the comprehensive judgment index covers the entire range from low to high anomaly risk, and there is a clear numerical interval between normal and abnormal samples. Overall, the SGT-LADNet anomaly detection model maintains strong anomaly response capability with relatively small errors, demonstrating good identification effect, trend tracking ability, and comprehensive judgment reliability for leakage monitoring anomalies.
Claims
1. A method for detecting leakage anomalies based on a ternary correlation of surface, geological layer, and tunnel, characterized in that, Includes the following steps: Collect monitoring section data and construct relevant data for leakage monitoring; After setting a short-term window for the leakage monitoring data, the mean value and instantaneous deviation value of the short-term window are calculated, and the instantaneous deviation value matrix and the deviation change magnitude matrix are constructed. Based on the short-window instantaneous deviation value matrix and the deviation change magnitude matrix, the deviation correlation strength matrix is obtained through item-by-item accumulation and multiplication. By combining the deviation intensity matrix with the competition amplification index and through bidirectional coupling, the bidirectional crystallization intensity and crystallization mean are obtained; Based on bidirectional crystallization intensity and crystallization mean, an adaptive crystallization threshold is constructed and the sparse relation edge strength is screened to form a sparse graph of positive correlation between variables; After combining the sparse graph of positive correlation between variables with the deviation change magnitude matrix and the short-window instantaneous deviation value matrix, the directed transmission matrix and the initial short-time anomaly field are calculated. The abnormal response values are calculated step by step for the directed transmission matrix and the initial short-term anomaly field. After weighted fusion, a short-term anomaly relationship characterization matrix is generated. The temporal clustering descriptor is calculated by using the short-term anomaly relationship representation matrix, and the order association mapping is performed to generate the short-term anomaly relationship embedding output, thus constructing a short-term modeling module; After combining leakage monitoring data into a long-term historical sequence, the long-term drift skeleton value is calculated to form a hysteresis transmission relationship diagram; Based on the long-term drift skeleton value and hysteresis propagation relationship graph, the long-term path output is calculated and the long-term path module is constructed. The short-term anomaly relationship embedding output and long-term path output are compressed, spliced and fused, and the anomaly detection value is calculated. The anomaly detection result is obtained by threshold judgment, and an anomaly judgment module is constructed.
2. The method for detecting leakage anomalies based on the ternary correlation of surface, geological layer, and tunnel as described in claim 1, characterized in that, The process of constructing a short-window instantaneous deviation value matrix and a deviation change magnitude matrix is as follows: Input leakage monitoring data, divide the leakage monitoring data into short-windows according to a preset short-window length, and for each leakage monitoring variable within each short-window, construct a short-window mean based on the values at all times within the current short-window. Using any time within the current short-window as a benchmark, compare the leakage monitoring variable value at that time with the leakage monitoring variable values at other times within the window, and combine the time interval relationship between different times and the overall deviation of the leakage monitoring variable from the short-window mean to generate the short-window instantaneous deviation value of the leakage monitoring variable. Iterate through all leakage monitoring variables and times to form a short-window instantaneous deviation value matrix. According to the chronological order, compare the change magnitude of the short-window instantaneous deviation value of the same leakage monitoring variable between adjacent times, determine the change magnitude as the deviation change magnitude at the corresponding time, and assign a preset initial value to the deviation change magnitude at the beginning of the short-window to obtain the deviation change magnitude matrix.
3. The method for detecting leakage anomalies based on the ternary correlation of surface, geological layer, and tunnel as described in claim 1, characterized in that, The deviation correlation strength matrix is obtained as follows: Based on the short-window instantaneous deviation value matrix and the deviation change magnitude matrix, a deviation relationship matrix is constructed for any two leakage monitoring variables. Along the short-term time dimension, the short-window instantaneous deviation values of the two leakage monitoring variables at each corresponding time are synchronously compared. The portions showing the same-direction imbalance at the same time are extracted and accumulated over time to obtain the cumulative value of the same-direction deviation. Simultaneously, the portions showing the opposite-direction imbalance at the same time are extracted and accumulated over time to obtain the cumulative value of the opposite-direction deviation. Along the time dimension, the deviation change magnitudes of the two leakage monitoring variables at each corresponding time are synchronously compared, and the deviation change magnitudes at the same time are extracted and compared. When the deviations of the two variables are in the same direction at the same acquisition time, the corresponding product values are retained and accumulated item by item to obtain the cumulative value of synchronous deviation change. The difference in the magnitude of the deviation change at each corresponding time is statistically accumulated to obtain the cumulative value of the deviation change difference. The cumulative value of the same-direction deviation and the cumulative value of synchronous deviation change are used as the relationship enhancement part, and the cumulative value of the opposite deviation and the cumulative value of the deviation change difference are used as the relationship suppression part. Combined with the preset turning point synergistic enhancement coefficient, opposite-direction conflict suppression coefficient and turning point mismatch suppression coefficient, joint constraint processing is performed to obtain the deviation correlation strength between the two leakage monitoring variables. All leakage monitoring variable pairs are traversed to form the deviation correlation strength matrix.
4. The method for detecting leakage anomalies based on a ternary correlation of surface, geological layer, and tunnel as described in claim 1, characterized in that, The bidirectional crystallization intensity and crystallization mean are obtained as follows: A sparse graph of positive correlation between variables is constructed based on the deviation correlation intensity matrix. For any leakage monitoring variable, the comprehensive deviation correlation intensity between it and other leakage monitoring variables is processed row by row, retaining only the non-negative relation part. The non-negative relation part is then normalized by combining a preset competition amplification index. Within the entire row corresponding to the current leakage monitoring variable, the relationship intensity after competition enhancement is normalized and allocated by combining a stability constant to obtain the positive correlation weight value of the current leakage monitoring variable pointing to other leakage monitoring variables. All leakage monitoring variables are traversed to form a positive competition allocation matrix. The bidirectional positive correlation weight values between any two leakage monitoring variables are coupled to generate the bidirectional crystallization intensity of the corresponding variable pair. The crystallization mean is constructed based on all bidirectional crystallization intensities in the row where the same leakage monitoring variable is located.
5. The method for detecting leakage anomalies based on a ternary correlation of surface, geological layer, and tunnel as described in claim 1, characterized in that, To form a sparse graph of positive correlation between variables, specifically: an adaptive crystallization threshold is constructed based on the crystallization mean and all bidirectional crystallization intensities. For any leakage monitoring variable, an adaptive crystallization threshold is generated based on the dispersion of the bidirectional crystallization intensity of the row containing the variable relative to the crystallization mean, combined with a preset threshold adjustment coefficient. All leakage monitoring variables are traversed to obtain a set of adaptive crystallization thresholds. For any two leakage monitoring variables, the bidirectional crystallization intensity is compared with the adaptive crystallization thresholds corresponding to the two variables respectively. When the bidirectional crystallization intensity simultaneously meets the crystallization threshold retention condition, the bidirectional crystallization intensity is determined as the sparse relation edge strength. All leakage monitoring variable pairs are traversed to form a sparse graph of positive correlation between variables.
6. The method for detecting leakage anomalies based on the ternary correlation of surface, geological layer, and tunnel as described in claim 1, characterized in that, The directed transmission matrix and initial short-term anomaly field are calculated as follows: a directed conservation anomaly transmission model is constructed based on the sparse graph of positive correlation between variables, the deviation change magnitude matrix, and the short-window instantaneous deviation value matrix; for any two leakage monitoring variables, the short-window instantaneous deviation values of the two leakage monitoring variables at adjacent times are cross-compared according to the time sequence within the short-term window to obtain the leading direction quantity between variables; based on the leading direction quantity and combined with the stability constant, directional constraint processing is performed to generate the directional enhancement coefficient. Based on the strength and direction enhancement coefficient of sparse relation edges in the sparse graph of positive correlation between variables, all relation edges in the corresponding row of each leakage monitoring variable are normalized and direction-enhanced to obtain a directed transmission matrix; then, combined with the short-window instantaneous deviation value and deviation change amplitude of each leakage monitoring variable, and introducing a preset mutation enhancement coefficient, an initial short-term anomaly field is constructed.
7. The method for detecting leakage anomalies based on a ternary correlation of surface, geological layer, and tunnel as described in claim 1, characterized in that, The short-term anomaly relationship representation matrix is generated as follows: Multi-order conserved anomaly propagation is performed based on the directed transmission matrix and the initial short-term anomaly field; propagation modeling is performed step-by-step according to the preset maximum propagation order. In each propagation process, the current-order anomaly response value is constructed based on the previous-order anomaly response value. The current-order anomaly response value is composed of the previous-order anomaly response part of the current leakage monitoring variable and the anomaly response parts of the other leakage monitoring variables transmitted through the directed transmission matrix. The two parts are conserved and allocated using the corresponding order's propagation coefficient to obtain the current-order anomaly response value; the overall change degree between adjacent order anomaly response values is statistically analyzed to obtain the inter-order change amount corresponding to each order; based on the inter-order change amount and combined with the attenuation adjustment coefficient, normalization attenuation processing is performed on all orders to generate multi-order fusion weights corresponding to each order; the anomaly response values of all orders are weighted and fused according to the multi-order fusion weights to obtain the short-term anomaly relationship representation, and this process is iterated through all leakage monitoring variables and all time points to form the short-term anomaly relationship representation matrix.
8. The method for detecting leakage anomalies based on a ternary correlation of surface, geological layer, and tunnel as described in claim 17, characterized in that, The short-term anomaly relationship embedding output is generated as follows: a short-term anomaly relationship embedding output is constructed based on the short-term anomaly relationship representation matrix; for any leakage monitoring variable at any time, the short-term anomaly relationship representation is statistically analyzed, and the overall amplitude distribution of the short-term anomaly relationship representation of the leakage monitoring variable at all times within the current short-term window is statistically analyzed. The overall anomaly amplitude of the variable within the short-term window is used as a normalization reference to obtain the relative proportion of the short-term anomaly relationship representation at the target time in the overall anomaly amplitude, thus forming a time-series aggregated descriptive quantity. According to the preset embedding order, the short-term anomaly relationship representation at the target time is mapped with the temporal aggregation descriptor to obtain the corresponding multi-order embedding value. Iterate through all leakage monitoring variables, time points, and embedding orders to form short-term anomaly relationship embedding output.
9. The method for detecting leakage anomalies based on the ternary correlation of surface, geological layer, and tunnel as described in claim 1, characterized in that, To form a hysteresis transmission relationship diagram, the following steps are taken: input leakage monitoring data, select a time period of preset length from the current time, and combine the leakage monitoring data within the time period in chronological order to obtain a long-term historical sequence; divide the long-term historical sequence according to a preset time span to form multiple long-term segments, and establish a corresponding time index set for each long-term segment. For any leakage monitoring variable in any long-term segment, the overall response level of the variable in the corresponding time index set is statistically analyzed, and the change information of the variable between adjacent time points in the long-term segment is extracted. The change information is compensated and adjusted by combining the preset drift compensation coefficient, and the segment reference representation is jointly constructed with the compensated change information to generate the corresponding long-term drift skeleton value. Traverse all leakage monitoring variables and all long-term segments to form a long-term drift skeleton matrix; construct hysteresis transmission relationships between adjacent long-term segments, based on the degree of difference between the long-term drift skeleton values of leakage monitoring variables in the current long-term segment and the long-term drift skeleton values of leakage monitoring variables in the previous long-term segment, and combine the variable interval attenuation coefficient for distance constraints to generate the strength of hysteresis transmission relationships between variables; traverse all variable combinations and all long-term segments to form a hysteresis transmission relationship diagram between long-term segments.
10. The method for detecting leakage anomalies based on the ternary correlation of surface, geological layer, and tunnel as described in claim 9, characterized in that, The long-term path output is calculated as follows: a long-term abnormal trend embedding is constructed based on the long-term drift skeleton matrix and the hysteresis transmission relationship graph; for any leakage monitoring variable in any long-term segment, the long-term drift skeleton value of the variable in the current long-term segment is combined with the skeleton response of all leakage monitoring variables in the previous long-term segment after being transmitted through the hysteresis transmission relationship graph, and a preset hysteresis transmission fusion coefficient is introduced to fuse the skeleton response of the current segment itself and the transmitted skeleton response of the previous segment to obtain the long-term abnormal trend embedding of the variable in the corresponding long-term segment; Iterate through all leakage monitoring variables and all long-term segments to form a long-term path output.
Citation Information
Patent Citations
Highway pavement quality detection method and system
CN121917753A
Power transmission line iron tower detection and quality management system
CN121936996A
