A reservoir dam dynamic early warning method and system based on working condition migration and double baseline competition

Abstract

The present application belongs to the field of safety monitoring and intelligent early warning of hydraulic structures, and discloses a reservoir dam dynamic early warning method and system based on working condition migration and double baseline competition. Environmental load, monitoring time and measuring point response data are obtained to identify current working condition migration; matched historical migration segments are screened and aligned according to stage change nodes to form a historical baseline of the same working condition, and a component model baseline is formed according to the environmental load influence component and the time effect influence component; the difference between the two baselines is calculated, and when the difference does not exceed the threshold, a dynamic early warning band is formed by fusion, and when the difference exceeds the threshold, the baseline is selected or the abnormality is determined according to the position of the current response value relative to the range of the two baselines; early warning is generated by combining data quality and associated measuring point responses, and a to-be-verified updated baseline is formed for the mismatched baseline, which is replaced after independent verification. The present application can reduce the false alarm and baseline error update risk during working condition migration.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and intelligent early warning technology for hydraulic structures, specifically to a dynamic early warning method and system for reservoir dams based on working condition migration and dual baseline competition. Background Technology

[0002] During long-term operation, the monitoring responses of reservoir dams, such as displacement, seepage flow, seepage pressure, and stress-strain, are simultaneously affected by environmental loads such as water level, temperature, and rainfall, as well as time-related factors such as material aging and foundation deformation. Existing dam safety monitoring systems typically employ fixed thresholds, statistical regression models, or machine learning models to compare measured values ​​with thresholds or model predictions, thereby identifying anomalies and generating early warnings.

[0003] For example, Chinese patent CN111508216A discloses an intelligent early warning method for dam safety monitoring data. Through gross error identification and processing, it selects a stepwise regression model, a grey system model, or a correlation vector machine model based on the monitoring project, the correlation of independent variables, the quantity and distribution of historical data, and combines this with thresholds, the number of consecutive early warnings, and the situation of associated monitoring points to determine the early warning type. This scheme can improve the quality of monitoring data and reduce manual judgment work. However, its early warning benchmark still mainly comes from the selected model and preset indicators, and it does not establish a competitive processing mechanism between the historical migration process and the component models for situations where the operating conditions migrate from one stable state to another during water storage, water discharge, or seasonal temperature changes.

[0004] For example, Chinese patent CN112381309A discloses a method for monitoring and early warning of reservoir dam safety. This method receives monitoring data such as dam deformation, seepage pressure, seepage flow, water level, and rainfall, compares the measured values ​​with the predicted values ​​of a prediction model, assesses the dam's safety risk based on the deviation, and distinguishes between collected anomalies and potential safety anomalies through subsequent continuous monitoring. This scheme achieves multi-source monitoring and risk assessment, but still primarily relies on a single early warning benchmark output by the prediction model.

[0005] In actual operation, rapid rises and falls in water levels, seasonal temperature changes, and other operational conditions can cause dam responses to exhibit significant lag, phased nature, and path correlation. If models established under steady-state conditions or fixed thresholds are directly used, normal migration responses may be misjudged as structural anomalies. Historical migration processes of similar nature can reflect the actual response patterns of a specific dam, but different migration processes have different durations, making direct comparison based on absolute time difficult. Models based on environmental loads and time-dependent components have good continuous prediction capabilities, but mismatches may occur when operational conditions change, parameters drift, or unmodeled factors appear.

[0006] Therefore, existing technologies lack a method to identify operational condition shifts, evaluate the consistency between historical shift baselines and component model baselines, and determine the supported baseline based on the current response when significant discrepancies occur between the two baselines. Furthermore, directly updating the early warning baseline online using current operational data may inadvertently write actual structural anomalies or data acquisition anomalies into the baseline, causing the anomalies to be absorbed by the baseline and reducing the sensitivity of subsequent early warnings. Therefore, it is also necessary to address the issues of securely updating mismatched baselines and independently verifying the update results. Summary of the Invention

[0007] This invention aims to provide a dynamic early warning method for reservoir dams based on working condition migration identification, dual baseline competition, and controlled updates, in order to solve the technical problems of easy mismatch of a single early warning baseline during working condition migration, high false alarm rate of anomalies, and the possibility of unverified baseline updates absorbing abnormal data.

[0008] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution.

[0009] In a first aspect, the present invention provides a dynamic early warning method for reservoir dams based on operating condition migration and dual baseline competition.

[0010] This includes: S1, acquiring environmental load data, monitoring time information, and response data from monitoring points awaiting early warning and related monitoring points of the reservoir dam to form a working condition sequence and determine the current response value; S2, identifying the current working condition migration based on changes in the relatively stable sample domain of the working condition sequence, and determining the working condition migration window, source stable working condition, target stable working condition, and migration direction; S3, aligning historical migration segments matching the current working condition migration according to the migration process to form a historical baseline for the same working condition containing historical benchmark values ​​and historical fluctuation ranges, and forming a component model baseline containing model predicted values ​​and model uncertainty ranges through component models based on environmental load data and time-dependent data; S4, determining the baseline based on the difference in benchmark values ​​and the degree of overlap in range between the two baselines. Line divergence degree; if it does not exceed the divergence threshold, the two baselines are merged to form a dynamic early warning zone; if it exceeds the divergence threshold, if the current response value is only in one of the two ranges, the corresponding baseline is determined as the selected baseline and the other baseline is determined as the mismatched baseline. If it is in both ranges at the same time, the two baselines are kept running in parallel. If it is outside both ranges at the same time, it is transferred to anomaly judgment; S5, after the data quality check is qualified, the structural anomaly or the anomaly to be verified is determined according to the baseline divergence degree, the current response value range exceeding the status and the response of the associated measuring point; when the working condition migration window ends and the selected baseline and the mismatched baseline have been determined, the updated baseline to be verified is formed according to the sample without anomalies and conforming to the selected baseline. After the verification is passed, the mismatched baseline is replaced.

[0011] Specifically, the stable sample domain is formed by the historical environmental load characteristics during the normal operation of the reservoir dam; the moment when the current operating condition feature leaves the stable sample domain corresponding to the source stable operating condition is determined as the starting point of the operating condition migration window, and the moment when the current operating condition feature enters the stable sample domain corresponding to the target stable operating condition and reaches the preset duration is determined as the ending point of the operating condition migration window; the migration direction is determined according to the change direction between the feature centers of the source stable operating condition and the target stable operating condition.

[0012] Specifically, each historical migration segment is associated with a source stable condition, a target stable condition, a migration direction, a measurement point identifier, and a quality status. From the historical migration segments with qualified quality status, historical migration segments whose source stable condition, target stable condition, and migration direction match the current condition migration are selected. Using the migration start point, the extreme point of the response change rate, and the migration end point as stage change nodes, the stage change nodes of the selected historical migration segments are aligned sequentially, and resampling is performed between adjacent stage change nodes according to the relative migration progress. The historical reference value is determined based on the median of the historical measurement point response at each aligned position, and the historical fluctuation range is determined based on the preset upper and lower quantiles.

[0013] Specifically, the current response value is the response value of the monitoring point to be warned at the current judgment time, or the statistical characteristic value of the response data of the monitoring point to be warned within the current judgment time window; the time-effect data is determined according to the time interval between the current monitoring time and the preset benchmark time; the component model includes an environmental load component sub-model and a time-effect component sub-model, the environmental load component sub-model and the time-effect component sub-model respectively calculate the environmental load influence component and the time-effect influence component, the composite value of each influence component is determined as the model prediction value, and the model uncertainty range is determined according to the model residual distribution of historical verification samples that match the current working conditions.

[0014] Specifically, determining the baseline divergence level includes: dividing the absolute difference between the historical benchmark value and the model prediction value by the sum of the historical fluctuation range width and the model uncertainty range width to obtain the central difference term; determining the ratio of the intersection length and union length of the historical fluctuation range and the model uncertainty range as the range overlap coefficient; weighting the central difference term with the result obtained by subtracting the range overlap coefficient to obtain the baseline divergence level; when forming a dynamic early warning zone, determining the fusion weight based on the recent error and range coverage of the two baselines, and using the fusion weight to fuse the historical benchmark value and the model prediction value, as well as the corresponding boundaries of the historical fluctuation range and the model uncertainty range.

[0015] More specifically, when the baseline divergence does not exceed the divergence threshold, and the current response value continuously exceeds the dynamic warning zone, historical fluctuation range, and model uncertainty range in the same direction for multiple consecutive judgment times, structural anomaly or anomaly to be verified is determined based on the response of associated measurement points; when the baseline divergence exceeds the divergence threshold, and the current response value is simultaneously outside the historical fluctuation range and the model uncertainty range, structural anomaly or anomaly to be verified is determined based on the response of associated measurement points; when the baseline divergence exceeds the divergence threshold, and the current response value is within the historical fluctuation range and outside the model uncertainty range, the historical baseline under the same operating condition is determined as the selected baseline, and the component model baseline is determined as the mismatched baseline; when the current response value is within the model uncertainty range and outside the historical fluctuation range, the component model baseline is determined as the selected baseline, and the historical baseline under the same operating condition is determined as the mismatched baseline; when the current response value is simultaneously within the historical fluctuation range and the model uncertainty range, both baselines are kept running in parallel.

[0016] Specifically, the quality inspection includes integrity checks, time synchronization checks, and outlier checks on environmental load data, response data of monitoring points to be warned, and response data of associated monitoring points. If any check fails, an abnormal acquisition marker is generated and the structural anomaly determination is suspended. Within a preset associated determination time window, when a preset number of associated monitoring points and monitoring points to be warned have the same response change direction, the response time difference does not exceed a preset time difference threshold, and the response change amount reaches the corresponding change amount threshold, it is determined that the associated monitoring points have a consistent response and a structural anomaly warning is generated; otherwise, an anomaly marker to be verified is generated.

[0017] Specifically, the formation and verification of the updated baseline to be verified includes: after the end of the operating condition transition window and the selected baseline and mismatched baseline have been determined, selecting samples that have passed quality inspection, have not been judged as structural anomalies or anomalies to be verified, and whose current response values ​​are continuously within the range corresponding to the selected baseline, to form an updated sample set; when the mismatched baseline is a historical baseline of the same operating condition, forming a historical updated baseline of the same operating condition to be verified based on the updated sample set; when the mismatched baseline is a component model baseline, updating the component model and its model residual distribution based on the updated sample set, forming a component model updated baseline to be verified; determining the historical updated baseline of the same operating condition to be verified or the component model updated baseline to be verified as the updated baseline to be verified, and running it in parallel with the mismatched baseline on an independent verification sample set; when the error of the updated baseline to be verified is lower than that of the mismatched baseline, the range coverage is not lower than that of the mismatched baseline, and the false alarm rate is not higher than that of the mismatched baseline, replacing the mismatched baseline with the updated baseline to be verified, otherwise retaining the mismatched baseline; the updated baseline to be verified does not participate in the structural anomaly determination before replacing the mismatched baseline.

[0018] In a second aspect, the present invention also provides a dynamic early warning system for reservoir dams. The system is used to implement the dynamic early warning method for reservoir dams described in the first aspect, including: a data processing module, used to determine the monitoring points to be warned and their associated monitoring points, acquire environmental load data, monitoring time information, response data of the monitoring points to be warned and response data of associated monitoring points, form a working condition sequence, and determine the current response value based on the response data of the monitoring points to be warned;

[0019] The working condition migration identification module is used to identify the current working condition migration based on the change of the working condition sequence relative to the stable sample domain, and to determine the working condition migration window, source stable working condition, target stable working condition and migration direction corresponding to the current working condition migration.

[0020] The dual baseline construction module is used to align historical migration segments that match the current working condition migration according to the migration process to form a historical baseline of the same working condition, and to form a component model baseline through component model based on environmental load data and time data determined by monitoring time information.

[0021] The baseline competition module is used to determine the degree of baseline divergence. When the degree of baseline divergence does not exceed the divergence threshold, a dynamic early warning zone is formed. When the degree of baseline divergence exceeds the divergence threshold and the current response value is only within one of the historical fluctuation range and the model uncertainty range, the selected baseline and the mismatched baseline are determined.

[0022] The early warning update module is used to determine structural anomalies or anomalies to be verified based on the degree of baseline divergence, the current response value's range exceeding the limit, and the responses of related measurement points after the environmental load data and the measurement point response data have passed the quality check. When the working condition migration window ends and the selected baseline and mismatched baseline have been determined, a verification update baseline is formed. The verification update baseline and the mismatched baseline are verified in parallel. When the update conditions are met, the verification update baseline replaces the mismatched baseline. The verification update baseline does not participate in the structural anomaly determination before replacing the mismatched baseline.

[0023] Thirdly, the present invention further provides a computer program product, including a computer program, which, when executed by a processor, implements the dynamic early warning method for reservoir dams described in the first aspect.

[0024] This invention identifies source stable conditions, target stable conditions, and migration directions, filters similar historical migration segments, and aligns them according to stage change nodes, reducing the impact of migration duration differences on historical baselines. By forming historical baselines and component model baselines under the same conditions in parallel, and determining the degree of baseline divergence based on benchmark value differences and range overlap, it comprehensively utilizes actual historical response patterns and environmental load and time-dependent mechanisms. When the two baselines are consistent, fusion is performed; when there is a significant divergence, the supported baseline is determined based on the positional relationship of the current response values, or an anomaly assessment is initiated. Data quality checks and responses from associated measurement points are combined to distinguish between acquisition anomalies and structural anomalies, thereby reducing the risk of false alarms and missed alarms. For mismatched baselines, only samples that meet quality standards and have not been identified as anomalies are used to form updated baselines to be verified. These updated baselines are then replaced after independent verification, preventing unverified update results from directly participating in early warning and preventing the erroneous absorption of anomalous data, thus improving the long-term stability and traceability of the early warning baseline. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the dynamic early warning method for reservoir dams according to the present invention;

[0026] Figure 2 This is a schematic diagram of the dual baseline formation and competition process of the present invention;

[0027] Figure 3 This is a schematic diagram of the anomaly determination process of the present invention;

[0028] Figure 4 This is a schematic diagram of the baseline-controlled update process of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are used to illustrate the implementation process of the present invention, and not to limit the scope of protection of the present invention. Those skilled in the art can adaptively set the sampling period, threshold, model form and data storage method without changing the data processing relationship, baseline competition relationship and update control relationship of the technical solution of the present invention.

[0030] I. Definition of Technical Objects

[0031] In this embodiment, the reservoir dam can be a concrete gravity dam, arch dam, earth-rock dam, or other water-retaining structure equipped with a safety monitoring system. The monitoring objects can include the dam's horizontal displacement, vertical displacement, crack opening and closing degree, uplift pressure, seepage flow, stress, strain, or temperature, etc.

[0032] 1.1 Environmental Load Data

[0033] Environmental load data refers to external forces that can have a measurable impact on the monitoring response of reservoir dams, including one or more of the following: upstream water level, downstream water level, reservoir water temperature, air temperature, rainfall, and water storage / discharge status. For different types of monitoring points to be monitored, different environmental load data can be selected to form a working condition sequence.

[0034] For example, when issuing an early warning for horizontal displacement of the dam body, environmental load data may include upstream water level, downstream water level, and air temperature; when issuing an early warning for seepage flow, it may also include rainfall and previous cumulative rainfall.

[0035] 1.2 Monitoring Time Information and Timeliness Data

[0036] Monitoring time information includes the monitoring time, sampling period, timestamp of the monitoring record, and runtime relative to the preset reference time. Time-dependent data is used to characterize the slow changes in the monitoring response of the reservoir dam over time.

[0037] The preset reference time can be the first time the reservoir impounds water, the time when monitoring equipment is replaced, the time when the dam reaches the designed operating water level, or the start time of a historical analysis period. Time-sensitive data can be obtained by directly using the time interval between the current monitoring time and the preset reference time, or by using the logarithm of the time interval, the piecewise linear transformation value, or other monotonic transformation values.

[0038] 1.3. Monitoring points to be warned and related monitoring points

[0039] The monitoring points to be warned are those that currently require anomaly assessment. Associated monitoring points are other monitoring points that have a pre-defined correlation with the monitoring points to be warned in terms of structural location, force path, monitoring object, or historical response relationship.

[0040] The associated measuring points can be determined from the following measuring points:

[0041] (1) Similar measuring points located in the same or adjacent dam sections as the measuring points to be warned;

[0042] (2) Similar measuring points located at the same elevation or adjacent elevations;

[0043] (3) Displacement, strain or stress measurement points located on the same structural force path;

[0044] (4) During the normal operation phase in history, the response change after direction normalization is a measurement point that has reached the preset correlation with the measurement point to be warned.

[0045] For measurement points with opposite response directions but corresponding structural mechanisms, a direction transformation coefficient can be preset to unify their response change direction to the same physical judgment direction before performing the associated response judgment.

[0046] 1.4 Operating Condition Sequence

[0047] The load condition sequence is a sequence of environmental load characteristics arranged according to monitoring time. For the first... At each decision point, the operating condition characteristics can be expressed as:

[0048]

[0049] In the formula, For the first The operating condition characteristics at each determination point; This refers to the upstream water level. This refers to the downstream water level. Temperature; The temperature of the reservoir water; Characteristics of rainfall; Characteristics of water storage and release; superscript This represents the transpose of a vector.

[0050] In practical applications, only environmental load characteristics related to the response of the monitoring point to be warned are selected, and it is not required that each working condition feature contain all the above data.

[0051] 1.5 Stable Sample Domain

[0052] A stable sample domain is a characteristic range formed by the historical environmental load characteristics of a reservoir dam under relatively stable operating conditions during its normal operation. Each stable sample domain corresponds to a single stable operating condition.

[0053] A relatively stable operating condition refers to an operating state in which the environmental load characteristics remain within the corresponding range for a preset duration, and there are no significant migration processes such as rapid water storage, rapid water discharge, extreme temperature changes, or concentrated heavy rainfall.

[0054] Stable sample domains can be represented by multidimensional boundaries, clustered regions, or statistical distance ranges. In one implementation, historical environmental load characteristics are standardized and clustered, and stable sample domains are formed based on the feature centers and covariance matrices of each cluster. Current operating conditions and the... The statistical distance between the feature centers is:

[0055]

[0056] In the formula, For the current operating condition characteristics and the first Statistical distance between stable sample domains; This describes the current operating condition characteristics; For the first Feature centers of a stable sample domain; For the first The covariance matrix of a stable sample domain; is the inverse of the covariance matrix.

[0057] when If the value is not greater than the corresponding domain boundary threshold, the current operating condition feature is determined to be located in the first... Within a stable sample domain. When the covariance matrix cannot be stably obtained, a stable sample domain can be formed by combining the upper and lower limits of various environmental load characteristics.

[0058] 1.6 Current operating condition migration

[0059] Current condition migration refers to the process by which current condition characteristics leave one stable sample domain and change to another stable sample domain. Current condition migration includes at least a migration window, a source stable condition, a target stable condition, and a migration direction.

[0060] The source stable condition is the condition corresponding to the stable sample domain before the migration begins; the target stable condition is the condition corresponding to the stable sample domain after the migration ends; the migration direction is used to distinguish migration processes that have the same source stable condition and target stable condition but different directions of environmental load change.

[0061] For example, a change in water level from a low-water-level stable condition to a high-water-level stable condition can be labeled as an upward direction; a change in water level from a high-water-level stable condition to a low-water-level stable condition can be labeled as a downward direction. For multi-dimensional operating condition characteristics, the migration direction can be from the feature center of the source stable condition to the feature center of the target stable condition.

[0062] 1.7 Historical Migration Fragments

[0063] Historical migration segments are complete operating condition migration data segments captured during normal historical operation. Each historical migration segment is associated with at least the segment identifier, measurement point identifier, start time, end time, source stable operating condition, target stable operating condition, migration direction, environmental load sequence, measurement point response sequence, and quality status.

[0064] Historical migration segments are used to form a historical baseline under the same operating conditions. Data segments that are determined to have structural anomalies, acquisition anomalies, equipment maintenance impacts, or human intervention impacts are not considered as historical migration segments with acceptable quality status.

[0065] 1.8 Historical baseline under the same operating conditions

[0066] The historical baseline for the same operating condition is a statistical baseline formed based on multiple historical migration segments that match the current operating condition migration, including historical benchmark values ​​and historical fluctuation ranges that change with the migration process.

[0067] Historical baseline values ​​are used to represent the typical change process of the measurement point response under the same migration conditions; historical fluctuation range is used to represent the allowable fluctuation range of this typical change process under normal operating conditions.

[0068] 1.9 Component Model Baseline

[0069] The component model baseline is a prediction baseline jointly formed by the environmental load component sub-model and the long-term aging component sub-model, including the model prediction values ​​and the model uncertainty range. The component model baseline calculates the normal response of the monitoring points to be warned from both the environmental load effect and the long-term aging effect.

[0070] 1.10. Selecting the baseline, mismatched baseline, and updated baseline to be verified.

[0071] The selected baseline is a provisional, reliable baseline supported by the current response value when the baseline divergence exceeds the divergence threshold. The mismatched baseline is a formal baseline not supported by the current response value and requiring further verification to determine if a decrease in applicability has occurred. The current response value refers to the median of the three most recent valid 10-minute samples from the monitoring point to be warned at the current warning determination time, used for comparison with historical baselines under the same operating conditions, component model baselines, or fused dynamic warning bands.

[0072] The updated baseline to be verified is a candidate updated version formed based on the mismatched baseline. The updated baseline to be verified does not participate in the structural anomaly determination until it is verified and replaces the mismatched baseline, and therefore will not directly change the current warning result.

[0073] II. Data Structure and Module Relationships

[0074] The reservoir dam dynamic early warning system includes a data processing module, an operating condition migration identification module, a dual baseline construction module, a baseline competition module, and an early warning update module.

[0075] The data processing module connects to the monitoring database, sensor acquisition system, or dam safety monitoring platform to read environmental load data, monitoring time information, and measurement point response data, perform time alignment and quality checks, and generate operating condition sequences and current response values.

[0076] The operating condition migration identification module receives the operating condition sequence, accesses the stable sample domain library, and outputs the current operating condition migration record. The current operating condition migration record includes at least the source stable operating condition identifier, the target stable operating condition identifier, the migration direction, the migration start point, the migration end point or migration not finished marker, and the current migration progress.

[0077] The dual baseline construction module accesses the historical migration fragment library and the component model library respectively. The dual baseline construction module outputs historical baseline records and component model baseline records under the same operating conditions.

[0078] Historical baseline records under the same working conditions may include: current migration identifier; identifier of measurement points to be warned; current migration progress; historical benchmark value; lower boundary of historical fluctuation range; upper boundary of historical fluctuation range; number of historical migration segments involved in the calculation; and baseline version number.

[0079] The baseline record of the component model may include: the identifier of the monitoring point to be warned; the current monitoring time; the environmental load influence component; the time-dependent influence component; the model prediction value; the lower boundary of the model uncertainty range; the upper boundary of the model uncertainty range; and the component model version number.

[0080] The baseline competition module receives two baselines and the current response value, calculates the degree of baseline divergence, and outputs one of the following states: fusion operation, selection of historical baseline under the same working condition, selection of component model baseline, parallel operation of dual baselines, or transition to anomaly determination.

[0081] The early warning update module receives the quality inspection results, baseline competition status, and related measurement point responses, and outputs one of the following processing results: normal, continuous observation, abnormal acquisition, structural abnormality, abnormality pending verification, or baseline update.

[0082] Each module can be deployed on the same server, or separately on a monitoring data server, analysis server, and early warning server. Different servers can transmit data records with measurement point identifiers, timestamps, and version numbers via the internal network. To avoid out-of-order data affecting current judgments, the data processing module determines the recording order based on the monitoring timestamp, not the time the data arrives at the server.

[0083] III. Construction Phase

[0084] 3.1 Historical Data Compilation

[0085] Environmental load data and monitoring point response data during normal operation are obtained from the historical monitoring database of the reservoir dam. Data is aligned with a unified sampling period. The unified sampling period can be 10 minutes, 30 minutes, 1 hour, or 1 day.

[0086] When the sampling periods of different data sources are inconsistent, linear interpolation, hold-attribute interpolation, or window statistics are used to unify continuous environmental load data and response data; for discrete data such as water storage and discharge status and equipment operation status, the most recent valid status is used for alignment.

[0087] Interpolated data should retain interpolation markers. When the duration of consecutive missing data exceeds the preset allowable duration, the corresponding data segment will not be used for the construction of stable sample domains, historical migration fragments, or component models.

[0088] 3.2 Construction of Stable Sample Domain

[0089] First, data from periods of rapid water level changes, concentrated rainfall, equipment maintenance, and known structural anomalies were excluded from the historical data of normal operation. Then, relatively stable samples were selected based on the rate of change of environmental load.

[0090] When water level is the main operating condition feature, if the rate of change of water level within a continuous preset time does not exceed the stable rate threshold, and other environmental load features such as temperature and rainfall do not exceed their respective change thresholds, then the corresponding data will be determined as a stable sample.

[0091] Clustering or interval partitioning is performed on stable samples to obtain multiple stable sample domains. For example, stable operating conditions at low water levels, medium water levels, and high water levels can be formed. Each stable sample domain stores feature centers, boundary parameters, number of samples, and effective time range.

[0092] 3.3 Construction of Historical Migration Fragment Library

[0093] Identify historical operating condition migrations between adjacent stable sample domains. The moment when a historical operating condition feature leaves the source stable sample domain is determined as the starting point of the historical migration segment, and the moment when a historical operating condition feature enters the target stable sample domain and reaches a preset duration is determined as the ending point of the historical migration segment.

[0094] For cases where the source stable sample domain is briefly re-entered during migration, a hysteresis time can be set. If the duration of re-entry into the source stable sample domain does not exceed the hysteresis time, the current historical migration segment is not terminated; if the hysteresis time is exceeded, the current segment is terminated or canceled to avoid repeated segmentation of the migration segment due to fluctuations in environmental loads near the boundary.

[0095] After each historical migration segment is completed, its environmental load data and measurement point response data are checked for completeness, time synchronization, outliers, and event records. Segments that pass the checks and have no known structural anomalies are set to qualified; otherwise, they are set to unqualified or pending verification.

[0096] 3.4 Component Model Construction

[0097] The component model includes an environmental load component model and a time-dependent component model. In one implementation method suitable for dam displacement monitoring, the model predicts the following values:

[0098]

[0099] In the formula, For the first Model prediction values ​​at each monitoring time point; For constant terms; For the first Water level characteristics at each monitoring time; This is a function of the water level load components; For the present and the future Temperature characteristics over a period of time; This is a function of the temperature load component; For timely data; For time-dependent components; This represents the temperature hysteresis order.

[0100] The environmental load component quantum model includes and The time-sensitive quantum model includes For seepage flow monitoring, rainfall or upstream / downstream water level difference components can be further added to the environmental load component sub-model.

[0101] The water level load component function can be a polynomial function or a piecewise spline function. The time-dependent component function can be a logarithmic function.

[0102]

[0103] In the formula, The timeliness affects the quantity; The time efficiency coefficient to be fitted; This is the time interval between the current monitoring time and the preset baseline time. It is a positive scaling parameter used to unify the units of time; It is the natural logarithm function.

[0104] Model parameters can be obtained by fitting historical samples from normal operation using least squares, weighted least squares, or regression methods with regularization constraints. During construction, historical samples are divided into training and validation sets according to time sequence to avoid the same continuous data segment from being included in both sets simultaneously.

[0105] The model residuals for the historical validation samples are:

[0106]

[0107] In the formula, For the first The model residuals of a historical validation sample; This corresponds to the measured response value; This is the corresponding model prediction value.

[0108] The model uncertainty range is determined based on the residual distribution of historical validation samples that match the current operating conditions:

[0109]

[0110] In the formula, For the first The model uncertainty range at each monitoring time; The preset lower quantile of the model residuals; The preset upper quantile of the model residuals; These are the model's predicted values.

[0111] For example, the 5th percentile and 95th percentile of the residuals can be used as... and When there are significant differences in the residual distribution under different stable operating conditions, the residual distribution corresponding to each stable operating condition or each migration direction is saved separately.

[0112] 3.5 Threshold and Validation Data Settings

[0113] The divergence threshold can be determined based on the distribution of baseline divergence in historical verification data from normal operation. For example, the 95th percentile of the baseline divergence in the normal operation verification samples can be used as the divergence threshold. Alternatively, the divergence threshold can be determined through historical playback based on the allowable false alarm rate of the early warning system.

[0114] The independent validation sample set is separate from the updated sample set used to form the updated baseline to be validated. The independent validation sample set can be normal operating data obtained after the updated sample, or it can be historical data that has been reserved in advance and has not been involved in model fitting and updating.

[0115] IV. Operation Phase

[0116] like Figure 1 As shown, the overall process of this invention includes: monitoring data acquisition, operating condition migration identification, dual baseline formation, baseline competition judgment, anomaly judgment, and baseline update.

[0117] 4.1 S1: Data acquisition, operating condition sequence formation, and current response value determination

[0118] The data processing module acquires environmental load data, monitoring time information, and response data of monitoring points awaiting early warning and related monitoring points of the reservoir dam.

[0119] The data processing module first reads the sensor type, unit, direction definition, and sampling period of the monitoring point to be warned based on the monitoring point identifier, and then reads the list of associated monitoring points. Subsequently, it aligns various types of data according to the monitoring timestamp.

[0120] The current response value can be the response value of the monitoring point to be warned at the current judgment time, or it can be the statistical characteristic value within the current judgment time window. For displacement data with small fluctuations and stable sampling, the current value or the median of the window can be used; for seepage flow data affected by instantaneous disturbances, the window mean, median, or the estimated value of the stable segment can be used.

[0121] When using a windowed value, the current response value is:

[0122]

[0123] In the formula, For the first The current response value at each decision point; to This refers to the response data of the monitoring points awaiting early warning within the current judgment time window; The number of samples within the determination time window; This indicates median operations.

[0124] If the number of valid samples in the current judgment time window is lower than the preset number of valid samples, no new structural anomaly judgment result will be generated. Instead, a collection anomaly marker will be generated or the previous valid judgment state will be maintained.

[0125] In one implementation, environmental load data and measurement point response data are standardized to a 10-minute sampling period; the current response value is determined using the median of the three most recent valid samples. Missing, out-of-limit, or quality-inspection-failed samples are not included as valid samples in the median calculation.

[0126] 4.2, S2: Current operating condition migration identification

[0127] The operating condition migration identification module calculates the relationship between the current operating condition characteristics and each stable sample domain hourly.

[0128] When the current operating condition feature is within a certain stable sample domain for a preset duration, the corresponding stable operating condition is recorded as the current stable operating condition. When the current operating condition feature leaves the stable sample domain, the departure time is recorded as the candidate migration starting point.

[0129] If the current operating condition features re-enter the original stable sample domain within the hysteresis time, the candidate migration starting point is cancelled; if they do not re-enter, the current operating condition migration is confirmed, the departure time is determined as the starting point of the operating condition migration window, and the stable operating condition before the migration is determined as the source stable operating condition.

[0130] During the migration process, candidate target stable operating conditions are determined based on the changing trajectory of the current operating condition characteristics. Once the current operating condition characteristics enter the stable sample domain corresponding to the candidate target stable operating condition and reach the preset duration, the moment of first entry or the confirmation moment of reaching the duration is determined as the end point of the operating condition migration window, and the candidate target stable operating condition is determined as the target stable operating condition.

[0131] The migration direction can be determined by the following formula:

[0132]

[0133] In the formula, This is the migration direction vector from the source stable condition to the target stable condition; The characteristic center of the source is the stable operating condition; The characteristic center is the target stable operating condition.

[0134] When the target stable operating condition has not yet been definitively confirmed, candidate target stable operating conditions can be determined based on the changing trends of the current operating condition characteristics, and the process can be run with candidate labels. When the subsequently entered stable sample domain is inconsistent with the candidate target stable operating condition, the target stable operating condition and migration direction are updated, and historical migration segments are re-screened.

[0135] If the current operating condition features remain outside each stable sample domain and exceed the maximum migration duration, the process is marked as an unidentified migration operating condition. Historical migration segments with incorrect matching are not used. The component model baseline can be maintained temporarily and a working condition verification mark can be generated.

[0136] 4.3, S3: Formation of historical baseline and component model baseline under the same working conditions

[0137] 4.3.1 Formation of historical baseline under the same working conditions

[0138] The dual-baseline construction module filters historical migration segments from those with acceptable quality status, selecting those with source stable conditions, target stable conditions, and migration directions that match the current condition. Further requirements include consistent measurement point identification, or that historical measurement points and the measurement points to be warned belong to the same calibrated type.

[0139] Since the duration of different historical migration segments varies, they cannot be directly averaged based on absolute time. Therefore, it is necessary to first determine the stages of change for each historical migration segment.

[0140] The stages of change nodes include the migration start point, the extreme point of the response change rate, and the migration end point. The response change rate can be determined by local linear fitting or adjacent window differencing. For segments with multiple local extrema, the main extreme point that remains significant within a preset duration is selected; when the extreme point of the response change rate cannot be stably identified, the extreme point of the environmental load change rate can be used as a substitute node.

[0141] The migration start point, the extreme point of the response change rate, and the migration end point are sequentially mapped. Resampling is performed between adjacent stage change nodes according to their relative migration progress. The relative migration progress of each historical migration segment between adjacent stage change nodes is:

[0142]

[0143] In the formula, For the first A moment in a historical migration fragment The relative migration progress; This is the starting point of the current stage; This is the termination point of the current stage; This refers to the monitoring point located between two stages of change.

[0144] when When, it corresponds to the starting node of the current stage; when At that time, it corresponds to the termination node of the current stage. Each segment can be resampled to a uniform relative migration progress position, such as 0, 0.05, 0.10 to 1.00.

[0145] At each alignment position, a historical baseline value is determined based on the median of the historical measurement point responses from multiple historical migration segments, and the historical fluctuation range is determined based on preset upper and lower quantiles.

[0146]

[0147] In the formula, relative migration progress Historical benchmark value; For the first A historical migration segment in relative migration progress The response value at the location; The number of historical migration fragments used in the calculation; This represents the historical fluctuation range. The lower quantile of the response values ​​for each historical migration segment; The upper quantile; This is the quantile parameter.

[0148] For example, It can be set to 0.05 to make the historical fluctuation range correspond to the 5th percentile to the 95th percentile. When the number of historical migration segments is small, the historical fluctuation range can be expanded, or the confidence of the historical baseline under the same working conditions can be marked as insufficient, and its fusion weight can be reduced.

[0149] 4.3.2 Component Model Baseline Formation

[0150] The dual baseline construction module inputs the current environmental load data and time-dependent data into the constructed component model, calculates the environmental load influence component and the time-dependent influence component respectively, and determines the composite value of each influence component as the model prediction value.

[0151] Select the corresponding historical validation sample residual distribution based on the current operating conditions, add the model prediction value to the residual quantile, and obtain the model uncertainty range.

[0152] When environmental load data is missing but the missing duration does not exceed the allowable duration, the most recent valid value or short-time interpolation value can be used for calculation, and the reliability level of the component model baseline is reduced. When the missing duration exceeds the allowable duration, the component model baseline output is paused, the dual baseline competition state is changed to the data shortage state, and the complete dual baseline judgment is not directly replaced by a single historical baseline under the same working condition.

[0153] like Figure 2 As shown, after identifying the current operating condition migration, the system forms two types of early warning baselines in parallel: On the one hand, it filters similar historical migration segments based on the source stable operating condition, the target stable operating condition, and the migration direction, and forms a historical baseline for the same operating condition through stage change node alignment and migration progress resampling; on the other hand, it forms a component model baseline based on the current environmental load data, monitoring time information, and the environmental load component sub-model and the time-dependent component sub-model. Subsequently, the system calculates the degree of divergence between the two types of baselines. When the degree of divergence does not exceed the divergence threshold, the two types of baselines are merged to form a dynamic early warning band for subsequent early warning judgment; when the degree of divergence exceeds the divergence threshold, the selected baseline and the mismatched baseline are determined based on the positional relationship of the current response value relative to the historical fluctuation range and the model uncertainty range, either maintaining the parallel operation of the two baselines or switching to anomaly judgment. Thus, false alarms caused by single baseline mismatch are avoided during operating condition migration, and a basis is provided for subsequent controlled baseline updates.

[0154] In one specific implementation, the component model baseline is formed using a model structure that superimposes environmental load components and time-dependent components, and is not limited to a specific neural network model. Specifically, the model prediction value of the component model baseline can be expressed as: In the formula, This represents the model prediction value at time t. This refers to the upstream water level. This refers to the downstream water level. This refers to air temperature or dam body temperature. (.) represents the environmental load component of the quantum model, used to characterize the influence of water level, temperature and their combination on the response of the measuring point; It is a time-dependent quantum model used to characterize long-term changes caused by operating time, material creep, or foundation deformation; This is a constant correction term. The environmental load component sub-model can be established using multinomial regression, spline function regression, or a generalized additive model, while the time-dependent component sub-model can be established using linear, logarithmic, or piecewise functions. Model parameters are obtained by fitting historical normal monitoring samples, and the model residual distribution is calculated using validation samples that did not participate in the fitting. The model uncertainty range is determined based on the residual quantiles. Therefore, the component model baseline includes at least the model predicted values ​​and their model uncertainty range, used to calculate the degree of divergence with historical baselines under the same working conditions.

[0155] 4.4, S4: Determination of baseline divergence and handling of baseline divergence

[0156] 4.4.1 Determination of the degree of baseline divergence

[0157] Let the historical baseline under the same operating conditions be:

[0158]

[0159] The baseline of the component model is:

[0160]

[0161] In the formula, This serves as the historical baseline for the same operating conditions. This is a historical benchmark value; and These are the lower and upper boundaries of the historical fluctuation range, respectively. The baseline for the component model; These are the model's predicted values; and These represent the lower and upper boundaries of the model's uncertainty range, respectively.

[0162] The central difference item is:

[0163]

[0164] In the formula, The central difference term; This represents the absolute difference between the historical baseline value and the model prediction value. This represents the width of the historical fluctuation range; This represents the width of the model's uncertainty range; To prevent positive numbers with a denominator of zero.

[0165] The range overlap coefficient is:

[0166]

[0167] In the formula, The numerator is the range overlap coefficient; the numerator is the intersection length of the historical fluctuation range and the model uncertainty range; the denominator is the union length of the two ranges; the meanings of the remaining parameters are the same as those described above.

[0168] The degree of baseline divergence is:

[0169]

[0170] In the formula, The degree of baseline divergence; Weights for the central difference terms; Weights for items with non-overlapping ranges; ,and and All are not less than zero.

[0171] 4.4.2 Handling when baseline divergence does not exceed the divergence threshold

[0172] When the degree of baseline divergence does not exceed the divergence threshold, it indicates that there is no significant conflict between the two baselines at the current determination time. The baseline competition module determines the fusion weight based on the recent error and range coverage of the two baselines.

[0173] The fused baseline values ​​and corresponding boundaries are as follows:

[0174]

[0175] In the formula, This serves as the baseline value for fusion. and These are the lower and upper boundaries of the dynamic early warning zone, respectively. Weights are integrated with historical baselines under the same operating conditions; Baseline fusion weights for component models; Furthermore, both fusion weights are not less than zero.

[0176] Baselines with low recent errors and coverage meeting preset requirements receive higher fusion weights. To avoid frequent weight jumps caused by short-term errors, a maximum single change in the fusion weight can be set.

[0177] When the current response value is within the dynamic warning zone, normal monitoring is maintained. When the current response value exceeds the dynamic warning zone but remains within at least one original baseline, fusion operation is maintained and continuous observation is initiated, without immediately generating a structural anomaly warning.

[0178] Only when the current response value continuously exceeds the dynamic warning zone, historical fluctuation range, and model uncertainty range in the same direction at multiple consecutive judgment times will the process proceed to the related measurement point response judgment. "Exceeding in the same direction" means that the current response value is continuously higher than the upper boundary of each range, or continuously lower than the lower boundary of each range.

[0179] 4.4.3 Handling when baseline divergence exceeds the divergence threshold

[0180] When the baseline divergence exceeds the divergence threshold, the following mutually exclusive processing is performed based on the position of the current response value relative to the historical fluctuation range and the model uncertainty range:

[0181] (1) When the current response value is within the historical fluctuation range and outside the model uncertainty range, the historical baseline under the same working condition is determined as the selected baseline, and the component model baseline is determined as the mismatch baseline;

[0182] (2) When the current response value is within the uncertainty range of the model and outside the historical fluctuation range, the component model baseline is determined as the selected baseline, and the historical baseline under the same working condition is determined as the mismatch baseline;

[0183] (3) When the current response value is within two ranges at the same time, the two baselines are kept running in parallel. If the selected baseline and the mismatched baseline are uncertain, the baseline update is not started.

[0184] (4) When the current response value is outside two ranges at the same time, the selected baseline and the mismatched baseline are uncertain, and the process is switched to anomaly judgment.

[0185] Range boundary values ​​are treated as falling within their respective ranges. Since the two ranges may partially overlap, the current response value falling within both ranges does not necessarily mean that the baseline divergence has not exceeded the divergence threshold; for example, the situation described above may still occur when the two ranges are wide but the baseline values ​​differ significantly.

[0186] 4.5, S5: Anomaly Detection and Baseline Update

[0187] 4.5.1 Data quality check and anomaly detection

[0188] Before performing structural anomaly determination, the environmental load data, the response data of the monitoring points to be warned, and the response data of the associated monitoring points are checked for quality.

[0189] The integrity check determines whether the proportion of valid data within the current judgment time window reaches a preset proportion. The time synchronization check determines whether the timestamp differences between different data sources do not exceed a preset synchronization error. The outlier check identifies data that exceeds the sensor's range, experiences abrupt changes, maintains a constant value for a long time, has duplicate timestamps, or does not conform to the device's status.

[0190] If any quality check fails, an anomaly flag is generated and structural anomaly assessment is paused. Data can continue to be saved during the pause, but the corresponding sample will not be added to the baseline update sample set.

[0191] When the abnormal candidate conditions are met, the response of the associated measuring points is checked within a preset association determination time window. Let the change in the response of the measuring point to be warned be:

[0192]

[0193] No. The response change of each associated measurement point after direction normalization is:

[0194]

[0195] In the formula, The change in the response of the monitoring point to be warned; and These are the response values ​​of the monitoring point to be warned at the current time and the previous comparison time, respectively; For the first The change in response of each associated measurement point after directional normalization; For the first The direction transformation coefficient of each associated measuring point can take the value of 1 or -1; and For the first The response values ​​of each associated measuring point at the corresponding time. This is the time interval for comparison.

[0196] like Figure 3 As shown, when the current response value exceeds the corresponding warning range and meets the anomaly candidate conditions, the system first performs a quality check on the environmental load data, the response data of the monitoring point to be warned, and the response data of related monitoring points, including integrity, time synchronization, and outlier checks. If the quality check fails, an acquisition anomaly marker is generated, and the structural anomaly determination is paused; if the quality check passes, the response of related monitoring points is further checked within a preset correlation determination time window. When a preset number of related monitoring points match the monitoring point to be warned in terms of response direction, occurrence time, and change amplitude, a structural anomaly warning is generated; otherwise, a verification anomaly marker is generated to prompt manual review, sensor retesting, or extended observation.

[0197] When a preset number of associated measurement points meet the following conditions, the associated measurement points are determined to have a matching response:

[0198] (1) and Same direction;

[0199] (2) The time difference between the time of occurrence of the response of the associated measuring point and the time of occurrence of the response of the measuring point to be warned does not exceed the preset time difference threshold;

[0200] (3) Reaching the The threshold for the change of each associated measurement point.

[0201] When the associated measuring points show a consistent response, a structural anomaly warning is generated; when the associated measuring points do not show a consistent response, a pending verification anomaly marker is generated. Pending verification anomalies can trigger manual review, sensor retesting, or extended continuous observation time, but are not directly recognized as confirmed structural anomalies.

[0202] 4.5.2. Baseline Formation to be Verified

[0203] After the operating condition migration window ends and the selected baseline and mismatch baseline have been determined, the early warning update module reviews the samples saved within this operating condition migration window and continues to collect samples from the initial stage of the target stable operating condition.

[0204] Select samples that simultaneously meet the following conditions to form an updated sample set:

[0205] (1) The environmental load data and the measurement point response data have passed the quality inspection;

[0206] (2) Not identified as a structural anomaly or an anomaly to be verified;

[0207] (3) The current response value remains within the range corresponding to the selected baseline;

[0208] (4) The data was not affected by equipment maintenance, manual loading or monitoring system switching.

[0209] "Continuously located" can be set to a preset number of consecutive judgment times within the selected baseline range, or the proportion of samples within the selected baseline range within a preset time period reaches a preset proportion.

[0210] When the mismatch baseline is the historical baseline under the same operating conditions, the updated samples are aligned according to the current migration process, and a new historical benchmark value and historical fluctuation range are calculated together with the retained historical migration segments with acceptable quality status, forming the historical updated baseline under the same operating conditions to be verified. To avoid a single migration completely covering historical statistical results, a maximum participation weight can be set for the new samples.

[0211] When the mismatch baseline is the component model baseline, the updated samples and the retained historical training samples are used together to update the environmental load component sub-model and the time-dependent component sub-model. The model residual distribution is recalculated based on the updated validation samples not used for parameter fitting, forming the updated baseline of the component model to be validated.

[0212] like Figure 4 As shown, after the operating condition transition window ends and the selected baseline and mismatched baseline have been determined, the system selects samples from the current transition window and the initial samples of the target stable operating condition that are of acceptable quality, not judged as abnormal, and continuously within the selected baseline range, forming an updated sample set. Subsequently, an updated baseline to be verified is formed according to the type of mismatched baseline, and it is run in parallel with the original mismatched baseline on an independent verification sample set. If the updated baseline to be verified meets the preset verification conditions in terms of error, range coverage, and false alarm rate, the original mismatched baseline is replaced; if the verification fails, the original baseline is retained, and the updated baseline to be verified is not enabled. This avoids unverified update results directly affecting the structural anomaly judgment.

[0213] 4.5.3. Verification and replacement of the updated baseline to be verified

[0214] The updated baseline to be validated and the mismatched baseline are run in parallel on independent validation sample sets. Both baselines receive the same input data, but the output of the updated baseline to be validated is only written to the validation record and is not used in the structural anomaly determination.

[0215] Baseline error can be expressed as mean absolute error:

[0216]

[0217] In the formula, Baseline error; Number of independent validation samples; For the first Measured response values ​​of an independent validation sample; This refers to the baseline or predicted value corresponding to the baseline.

[0218] The coverage rate is:

[0219]

[0220] In the formula, For range coverage; and The first The lower and upper boundaries of the range corresponding to each independent validation sample; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. The meanings of the other parameters are the same as described above.

[0221] The false alarm rate is the proportion of the number of samples in the normal operation verification samples that were identified as abnormal candidates by the baseline to the total number of normal operation verification samples.

[0222] When the baseline to be verified and updated simultaneously meets the following conditions: error is lower than the mismatched baseline, range coverage is not lower than the mismatched baseline, and false alarm rate is not higher than the mismatched baseline, the mismatched baseline to be verified and updated shall replace the mismatched baseline. During replacement, the baseline version number, activation time, and applicable operating conditions shall be updated.

[0223] If verification fails, the mismatched baseline is retained, the eligibility for updating the baseline to be verified is deleted or frozen, and subsequent valid samples continue to be accumulated. Baseline replacement only changes the baseline version corresponding to the mismatched baseline, does not change the history of the selected baseline, and does not retroactively modify already generated alert results.

[0224] V. Examples

[0225] 5.1 Example 1: Dynamic Early Warning of Horizontal Displacement of Dam Body

[0226] This embodiment uses a horizontal displacement monitoring system for a concrete gravity dam as an example. The monitoring point to be monitored is the vertical line monitoring point P0 in the middle section of the dam, and the associated monitoring points are the displacement monitoring points P1, P2 and P3 at the same elevation in adjacent dam sections.

[0227] In this embodiment, the environmental load data and the measurement point response data are uniformly sampled for 10 minutes; the current response value is determined by the median of the three most recent valid sample values.

[0228] The construction phase utilizes historical normal data from five consecutive operating years. Data from the first three operating years is used to construct the stable sample domain, historical migration segments, and component model parameters. Data from the fourth operating year is used to determine the divergence threshold and model uncertainty range. Data from the fifth operating year serves as retained validation data. This retained validation data is not used for component model parameter fitting, determining model residual distribution, or forming the updated baseline to be validated; it is only used for independent validation of the updated baseline and the original component model baseline.

[0229] Based on upstream water level and temperature characteristics, three stable sample domains were formed: low water level, medium water level, and high water level. The process of leaving the medium water level stable sample domain and entering the high water level stable sample domain was defined as the migration from medium water level to high water level. The migration direction was the direction of water level rise.

[0230] The historical migration segment library contains 14 migration segments ranging from medium to high water levels. After integrity checks, time synchronization checks, outlier checks, and operational record verification, two segments with long data gaps and one segment where monitoring equipment maintenance occurred were excluded. The remaining 11 historical migration segments were used to form a historical baseline under the same operating conditions.

[0231] With the current migration progress at 0.60, the dual baseline output results are shown in Table 1.

[0232] Table 1. Dual baseline output and divergence calculation data

[0233]

[0234] The intersection length of the two ranges is 0.27 mm, and the union length is 1.33 mm. Therefore, the range overlap coefficient is approximately 0.203. The central difference term is approximately 0.331. Setting the weights for both the central difference term and the range non-overlap term to 0.5, the baseline divergence is approximately 0.564.

[0235] In this embodiment, the divergence threshold is set to 0.45, therefore the baseline divergence exceeds the divergence threshold.

[0236] The current response value is 4.88 mm, which is within the historical fluctuation range but below the lower boundary of the model uncertainty range of 4.93 mm. Therefore, the historical baseline under the same operating conditions is determined as the selected baseline, and the component model baseline is determined as the mismatch baseline.

[0237] Since the current response value falls within the historical fluctuation range corresponding to the selected baseline, no anomaly candidates are generated. The data processing module continues to perform quality checks and save monitoring records.

[0238] Once the current operating condition characteristics enter the high-water-level stable sample domain and remain there for 48 hours, the operating condition migration window is determined to have ended. The early warning update module selects samples from the completed migration window and the initial samples of the high-water-level stable operating condition that have passed quality inspection, have not been identified as structurally abnormal or have anomalies pending verification, and remain within the historical fluctuation range to form an updated sample set.

[0239] The updated component model parameters and residual distribution are updated based on the updated sample set to form the updated baseline of the component model to be validated. The updated baseline of the component model to be validated and the original component model baseline are run in parallel on the retained validation data, and the results are shown in Table 2.

[0240] Table 2 Independent validation results before and after component model update

[0241]

[0242] As shown in Table 2, the error of the updated baseline of the component model to be verified is lower than that of the original component model baseline, the range coverage is not lower than that of the original component model baseline, and the false alarm rate is not higher than that of the original component model baseline, thus meeting the update conditions. Therefore, the updated baseline of the component model to be verified replaces the original component model baseline.

[0243] Before the replacement is completed, the updated baseline of the component model to be verified only outputs verification records. The current warning is still executed based on the selected historical baseline under the same working conditions and the original warning rules, thereby avoiding the unverified update results from directly changing the structural anomaly judgment.

[0244] The data shown in Table 2 is provided as an implementation example to verify the above calculations and judgments, and does not represent a safety assessment conclusion for any specific reservoir dam. In actual engineering applications, verified monitoring data of the corresponding reservoir dam should be used to determine each indicator.

[0245] 5.2 Example 2: Fusion operation when baseline divergence does not exceed the threshold

[0246] At another determination point, the historical baseline value for the same working condition was 3.10 mm, with a historical fluctuation range of 2.78–3.42 mm; the model prediction value for the component model baseline was 3.18 mm, with a model uncertainty range of 2.80–3.56 mm.

[0247] The calculated central difference term is approximately 0.057, and the range overlap coefficient is approximately 0.795. When both divergence terms have a weight of 0.5, the baseline divergence is approximately 0.131, which does not exceed the divergence threshold of 0.45.

[0248] Based on the recent errors and coverage of the two baselines, the fusion weight of the historical baseline under the same working condition was determined to be 0.56, and the fusion weight of the component model baseline was determined to be 0.44. After fusion, the fusion reference value of the dynamic warning zone is approximately 3.14 mm, the lower boundary is approximately 2.79 mm, and the upper boundary is approximately 3.48 mm.

[0249] When the current response value is 3.32 mm, it is within the dynamic warning zone, and the system maintains normal monitoring status.

[0250] When the current response value is 3.51 mm, it exceeds the dynamic warning zone but remains within the uncertainty range of the component model. The system then enters continuous observation mode and does not immediately generate a structural anomaly warning. The continuous observation mode is deactivated when the current response value subsequently re-enters the dynamic warning zone.

[0251] 5.3 Example 3: Associated Measurement Points in Structural Anomaly Determination

[0252] Continue using the dual baseline data from Example 1. If the current response value becomes 6.14 mm, then the current response value is simultaneously higher than both the upper boundary of the historical fluctuation range (5.20 mm) and the upper boundary of the model uncertainty range (5.77 mm), and the process proceeds to anomaly detection.

[0253] The quality inspection results show that the environmental load data and the response data from P0 to P3 are complete, with a valid sample ratio of no less than 95%; the time synchronization error between different data sources does not exceed 5 minutes; and none of the response values ​​exceed the sensor's range, nor are there any long-term constant values ​​or single-point jumps. Therefore, the quality inspection is qualified.

[0254] The correlation determination time window was set to 30 minutes, the preset number of correlation measurement points was 2, and the preset time difference threshold was 20 minutes. After direction normalization, the response change directions of P1 and P2 were the same as those of P0, with response time differences of 10 minutes and 15 minutes respectively, and the response change amounts of both reached their respective change amount thresholds; P3 did not reach its corresponding change amount threshold.

[0255] If a preset number of associated measuring points respond in agreement, the system generates a structural anomaly warning. If only P1 responds in agreement, and the preset number is not reached, an anomaly marker to be verified is generated, prompting manual retesting, equipment inspection, or extended observation, rather than directly confirming a structural anomaly.

[0256] VI. Other Implementation Methods

[0257] 6.1 System Implementation Method

[0258] A reservoir dam dynamic early warning system can consist of a server, data storage device, communication interface, and computer program.

[0259] The data processing module receives environmental load data of the reservoir dam, monitoring time information, response data of monitoring points awaiting early warning, and response data of related monitoring points from the monitoring database. It forms a unified monitoring record according to the monitoring point identifier and timestamp, and determines the current response value.

[0260] The operating condition migration identification module identifies the current operating condition migration based on the relationship between the operating condition sequence and the stable sample domain, and generates a current operating condition migration record that includes the operating condition migration window, the source stable operating condition, the target stable operating condition, and the migration direction.

[0261] The dual baseline construction module filters historical migration fragments based on the current working condition migration records, performs phase change node alignment and relative migration progress resampling to form the same working condition historical baseline; at the same time, it calls the environmental load component sub-model and the time-dependent component sub-model to form the component model baseline.

[0262] The baseline competition module calculates the baseline divergence level. When the baseline divergence level does not exceed the divergence threshold, a dynamic warning band is output; when the baseline divergence level exceeds the divergence threshold, based on the current response value and the positional relationship between the two ranges, the selected baseline, mismatched baseline, dual baselines in parallel, or an abnormal judgment state is output.

[0263] After the quality check is passed, the early warning update module generates a structural anomaly warning or an anomaly mark to be verified based on the current response value's range exceeding the state and the response of the associated measuring points. When the working condition migration window ends and the selected baseline and mismatched baseline have been determined, the module forms and verifies the updated baseline to be verified, and replaces the mismatched baseline after the update conditions are met.

[0264] When the server's processor executes the computer program in memory, it can complete S1 to S5 sequentially. Stable sample domains, historical migration fragments, model parameters, baseline versions, early warning results, and updated verification results can be stored in relational databases, time-series databases, or file storage systems.

[0265] Through the above implementation method, when a reservoir dam undergoes changes in operating conditions such as water storage, water discharge, and seasonal temperature variations, the system simultaneously utilizes historical baselines matching the historical migration process and component model baselines reflecting environmental loads and long-term effects for judgment. When the two baselines are consistent, they are fused; when there is a significant divergence between the two baselines, the supported baseline is determined based on the current response value; if structural anomalies cannot be ruled out by a single point response, further judgment is made by combining associated measurement points. Updated versions of mismatched baselines must be independently verified and are not included in structural anomaly judgment before replacement, thereby achieving continuous processing between operating condition migration identification, dual-baseline competition, associated anomaly judgment, and controlled baseline updates.

[0266] 6.2 Implementation Methods of Computer Program Products

[0267] This invention also provides a computer program product, which includes a computer program. The computer program can be stored in a computer-readable storage medium or provided to a reservoir dam dynamic early warning system via a wired or wireless network. When the computer program is executed by a processor, the processor performs each step of the above-described reservoir dam dynamic early warning method, including acquiring monitoring data and forming a working condition sequence, identifying current working condition migration, forming historical baselines and component model baselines for the same working condition, performing dual baseline competition and anomaly determination, and forming, verifying, and replacing the updated baseline to be verified. The computer program can also perform further limitations corresponding to the method steps in the above embodiments.

Claims

1. A dynamic early warning method for reservoir dams based on working condition migration and dual baseline competition, characterized in that, include: S1: Obtain environmental load data, monitoring time information, response data of monitoring points to be warned and related monitoring points of the reservoir dam, form a working condition sequence and determine the current response value; S2, Identify the current operating condition migration based on the changes in the relatively stable sample domain of the operating condition sequence, and determine the operating condition migration window, source stable operating condition, target stable operating condition, and migration direction; S3, align the historical migration segments that match the current working condition migration according to the migration process to form a historical baseline of the same working condition that includes historical benchmark values ​​and historical fluctuation ranges, and form a component model baseline that includes model prediction values ​​and model uncertainty ranges through component models based on environmental load data and time data. S4. Determine the degree of baseline divergence based on the difference in the baseline values ​​and the degree of overlap between the two baselines. If the divergence threshold is not exceeded, merge the two baselines to form a dynamic warning zone. If the divergence threshold is exceeded, if the current response value is only in one of the two ranges, then the corresponding baseline is determined as the selected baseline and the other baseline is determined as the mismatched baseline. If it is in both ranges at the same time, then keep the two baselines running in parallel. If it is outside both ranges at the same time, then switch to anomaly judgment. S5. After the data quality check is passed, determine the structural abnormality or the abnormality to be verified based on the degree of baseline divergence, the current response value's range of deviation, and the response of related measurement points. When the working condition migration window ends and the selected baseline and mismatched baseline have been determined, an updated baseline to be verified is formed based on samples that are free of anomalies and conform to the selected baseline. Once the verification is successful, the mismatched baseline is replaced.

2. The dynamic early warning method for reservoir dams according to claim 1, characterized in that, In S2, the stable sample domain is formed by the historical environmental load characteristics during the normal operation of the reservoir dam; the moment when the current operating condition feature leaves the stable sample domain corresponding to the source stable operating condition is determined as the starting point of the operating condition migration window, and the moment when the current operating condition feature enters the stable sample domain corresponding to the target stable operating condition and reaches the preset duration is determined as the ending point of the operating condition migration window. The migration direction is determined based on the direction of change between the characteristic centers of the source stable condition and the target stable condition.

3. The dynamic early warning method for reservoir dams according to claim 1, characterized in that, In S3, each historical migration segment is associated with a source stable condition, a target stable condition, a migration direction, a measurement point identifier, and a quality status. From the historical migration segments with qualified quality status, historical migration segments whose source stable condition, target stable condition, and migration direction match the current condition migration are selected. Using the migration start point, the extreme point of the response change rate, and the migration end point as stage change nodes, the stage change nodes of the selected historical migration segments are aligned sequentially, and resampling is performed between adjacent stage change nodes according to the relative migration progress. The historical reference value is determined based on the median of the historical measurement point response at each aligned position, and the historical fluctuation range is determined based on the preset upper and lower quantiles.

4. The dynamic early warning method for reservoir dams according to claim 1, characterized in that, The current response value is the response value of the monitoring point to be warned at the current judgment time, or it is the statistical characteristic value of the response data of the monitoring point to be warned within the current judgment time window; The time-dependent data is determined based on the time interval between the current monitoring time and the preset reference time. The component model includes an environmental load component sub-model and a time-dependent component sub-model. The environmental load component sub-model and the time-dependent component sub-model calculate the environmental load influence component and the time-dependent influence component, respectively. The composite value of each influence component is determined as the model prediction value. The model uncertainty range is determined based on the model residual distribution of historical verification samples that match the current working conditions.

5. The dynamic early warning method for reservoir dams according to claim 1, characterized in that, In S4, determining the baseline divergence level includes: dividing the absolute difference between the historical baseline value and the model prediction value by the sum of the historical fluctuation range width and the model uncertainty range width to obtain the central difference term; determining the ratio of the intersection length and union length of the historical fluctuation range and the model uncertainty range as the range overlap coefficient; weighting the central difference term with the result obtained by subtracting the range overlap coefficient to obtain the baseline divergence level; when forming a dynamic early warning zone, determining the fusion weight based on the recent error and range coverage of the two baselines, and using the fusion weight to fuse the historical baseline value and the model prediction value, as well as the corresponding boundaries of the historical fluctuation range and the model uncertainty range.

6. The dynamic early warning method for reservoir dams according to claim 5, characterized in that, When the baseline divergence does not exceed the divergence threshold, and the current response value at multiple consecutive judgment times continuously exceeds the dynamic warning zone, historical fluctuation range, and model uncertainty range in the same direction, the structural anomaly or the anomaly to be verified is determined based on the response of the associated measurement points. When the baseline divergence exceeds the divergence threshold and the current response value is outside both the historical fluctuation range and the model uncertainty range, the structural anomaly or the anomaly to be verified is determined based on the response of the associated measurement points. When the baseline divergence exceeds the divergence threshold, and the current response value is within the historical fluctuation range but outside the model uncertainty range, the historical baseline under the same operating condition is determined as the selected baseline, and the component model baseline is determined as the mismatched baseline; when the current response value is within the model uncertainty range but outside the historical fluctuation range, the component model baseline is determined as the selected baseline, and the historical baseline under the same operating condition is determined as the mismatched baseline; when the current response value is simultaneously within the historical fluctuation range and the model uncertainty range, both baselines are kept running in parallel.

7. The dynamic early warning method for reservoir dams according to claim 1, characterized in that, In S5, the quality inspection includes integrity checks, time synchronization checks, and outlier checks on environmental load data, response data of monitoring points to be warned, and response data of associated monitoring points. If any check fails, an abnormal acquisition mark is generated and the structural anomaly determination is suspended. Within the preset correlation determination time window, when the response change direction of the preset number of correlated measurement points and the measurement point to be warned is the same, the response time difference does not exceed the preset time difference threshold, and the response change amount reaches the corresponding change amount threshold, it is determined that the correlated measurement point has a consistent response and a structural anomaly warning is generated; otherwise, an anomaly mark to be verified is generated.

8. The dynamic early warning method for reservoir dams according to claim 1, characterized in that, In S5, forming and verifying the updated baseline to be verified includes: after the end of the operating condition transition window and the selected baseline and mismatched baseline have been determined, selecting samples that have passed quality inspection, have not been judged as structural anomalies or anomalies to be verified, and whose current response values ​​are continuously within the range corresponding to the selected baseline, to form an updated sample set; when the mismatched baseline is a historical baseline of the same operating condition, forming a historical updated baseline of the same operating condition to be verified based on the updated sample set; when the mismatched baseline is a component model baseline, updating the component model and its model residual distribution based on the updated sample set, forming a component model updated baseline to be verified; determining the historical updated baseline of the same operating condition to be verified or the component model updated baseline to be verified as the updated baseline to be verified, and running it in parallel with the mismatched baseline on an independent verification sample set; when the error of the updated baseline to be verified is lower than that of the mismatched baseline, the range coverage is not lower than that of the mismatched baseline, and the false alarm rate is not higher than that of the mismatched baseline, replacing the mismatched baseline with the updated baseline to be verified, otherwise retaining the mismatched baseline; the updated baseline to be verified does not participate in the structural anomaly judgment before replacing the mismatched baseline.

9. A dynamic early warning system for reservoir dams, characterized in that, This system is used to implement the dynamic early warning method for reservoir dams as described in any one of claims 1-8, comprising: The data processing module is used to determine the monitoring points to be warned and their associated monitoring points, acquire environmental load data, monitoring time information, response data of the monitoring points to be warned and response data of associated monitoring points, form a working condition sequence, and determine the current response value based on the response data of the monitoring points to be warned; The working condition migration identification module is used to identify the current working condition migration based on the change of the working condition sequence relative to the stable sample domain, and to determine the working condition migration window, source stable working condition, target stable working condition and migration direction corresponding to the current working condition migration. The dual baseline construction module is used to align historical migration segments that match the current working condition migration according to the migration process to form a historical baseline of the same working condition, and to form a component model baseline through component model based on environmental load data and timeliness data determined by monitoring time information; the baseline competition module is used to determine the degree of baseline divergence, to form a dynamic early warning zone when the degree of baseline divergence does not exceed the divergence threshold, and to determine the selected baseline and the mismatched baseline when the degree of baseline divergence exceeds the divergence threshold and the current response value is only within one of the historical fluctuation range and the model uncertainty range; The early warning update module is used to determine structural anomalies or anomalies to be verified based on the degree of baseline divergence, the current response value's range exceeding the limit, and the responses of related measurement points after the environmental load data and the measurement point response data have passed the quality check. When the working condition migration window ends and the selected baseline and mismatched baseline have been determined, a verification update baseline is formed. The verification update baseline and the mismatched baseline are verified in parallel. When the update conditions are met, the verification update baseline replaces the mismatched baseline. The verification update baseline does not participate in the structural anomaly determination before replacing the mismatched baseline.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the dynamic early warning method for reservoir dams as described in any one of claims 1-8.

Images