Dam structure health assessment method fused with deep learning

Abstract

The invention relates to the technical field of civil engineering structure health monitoring and evaluation, and discloses a dam structure health evaluation method fused with deep learning. Monitoring data such as displacement, osmotic pressure, water level and temperature are unified and standardized, and a dimensionless monitoring vector is constructed; forming a confidence domain in the space based on fault-free historical data, identifying the most unfavorable working condition, calculating a tolerance and generating a tolerance label set; designing a depth function mapper and non-probability loss, establishing mapping from a dimensionless monitoring vector to tolerance estimation, and determining parameters; in the operation period, a mapper is used for outputting monitoring variable tolerance estimation on a new sample in real time, a monitoring value and a tolerance ratio are mapped into a health state level, automatic conversion from monitoring data to a structure health assessment result is achieved, a confidence domain, labels and parameters are refreshed according to a period, tolerance mapping is adaptively updated along with the newest working condition, and the structure health assessment result is obtained. And working condition identification and model effectiveness are improved.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural health monitoring and assessment technology, specifically a method for assessing the structural health of dams that integrates deep learning. Background Technology

[0002] With the widespread construction and increasing service life of large hydraulic structures (such as reservoirs and dams), dam structural health monitoring and safety assessment have become crucial issues for ensuring the safe operation of these projects and public safety. Currently, the water conservancy industry commonly employs the deployment of various types of sensors (such as displacement gauges, piezometers, water level gauges, and thermometers) to monitor dam operation in real-time or periodically. The temporal changes in monitoring data and the identification of anomalies are fundamental to assessing dam safety risks and providing timely warnings. Therefore, how to rationally utilize multi-source monitoring data to achieve scientific analysis of dam operation and reliable early warning systems remains a core issue of ongoing concern in the industry.

[0003] In existing technologies, the methods for determining the limits of dam monitoring indicators mainly include empirical limit methods, probabilistic statistical methods, and numerical simulation methods. Empirical limit methods rely heavily on historical engineering experience or design specifications, assigning upper and lower thresholds based on monitoring values ​​from similar projects. However, this method is highly subjective, difficult to adapt to different dam types, operating conditions, and environmental changes, and lacks specificity. Probabilistic statistical methods typically assume that monitoring data follows a specific distribution (such as a normal distribution), inferring the data limit interval through parameters such as mean, standard deviation, and confidence interval. This method is applicable to scenarios with large amounts of monitoring data and stable distribution patterns. However, in actual engineering projects, the number of monitoring samples is often insufficient due to factors such as monitoring equipment placement, data acquisition cycle, and changes in operating conditions. Furthermore, the data distribution changes with operating conditions and seasons, leading to a decrease in the reliability of statistical methods. In addition, monitoring data often exhibits random fluctuations, abnormal discrete points, and even abnormal drift, further weakening the discriminative effect of traditional statistical methods. Numerical simulation methods rely on physical modeling and finite element analysis, utilizing dam structure and material parameters to calculate the response of key monitoring points under different loads and operating conditions, and determining the safety status by comparing it with measured data. Although simulation methods offer strong physical interpretability, the difficulty in obtaining model parameters, the complexity of setting boundary conditions, and the challenge of real-time dynamic updates to environmental changes during operation significantly limit their widespread application in daily monitoring and real-time early warning. More notably, existing methods generally lack full utilization of the time-varying correlations and coupling relationships between multi-source data, often focusing only on the limits of a single monitoring indicator, making it difficult to achieve collaborative analysis of multiple indicators at a global level.

[0004] Therefore, this study aims to propose a dam structure health assessment method integrating deep learning. After unifying the processing of multi-source monitoring data, a dimensionless monitoring vector sequence is constructed. Based on historical fault-free intervals, a high-confidence monitoring context domain is built, and the most unfavorable operating condition is determined. Through upper and lower limit deviation analysis and label construction, a limit label dataset corresponding to historical samples is formed. Based on this, a deep function mapper suitable for this dataset is designed, and the mapper parameters are iteratively solved using a non-probabilistic loss function. Newly sampled data is used to train the mapper to output limit estimates, and then the health status is graded through limit ratio analysis. Finally, a periodic refresh mechanism is set to continuously update the mapper and context domain, achieving dynamic and adaptive health assessment of the dam structure. Summary of the Invention

[0005] This invention provides a method for assessing the structural health of dams that integrates deep learning, thereby helping to solve the problems mentioned in the background art.

[0006] This invention provides the following technical solution: a method for assessing the structural health of dams by integrating deep learning, comprising:

[0007] Monitoring parameters were selected and multi-source monitoring variables were unified to construct a dimensionless multidimensional monitoring vector sequence that characterizes the dam structure's operational status.

[0008] By using fault-free historical monitoring data to construct a contextual confidence region and determine the most unfavorable operating conditions, a contextual confidence region reflecting the overall level of the monitored variables is formed.

[0009] Based on the context confidence region and historical monitoring data, the upper and lower limit difference vectors are calculated and the context labels are constructed to obtain the context difference label dataset corresponding to each historical monitoring sample.

[0010] Based on the dimensionless monitoring vector and contextual limit label dataset, a custom depth function mapper structure is designed to establish a depth function mapping relationship that maps dimensionless monitoring vectors to limit estimation vectors.

[0011] Design a nonprobabilistic loss function on a context-bounded label dataset and solve for the parameters of a deep function mapper to obtain a set of parameters that converge within the range of engineering accuracy requirements;

[0012] The depth function mapper is used to generate limit estimates for each monitoring variable on the newly collected monitoring samples;

[0013] Based on the ratio between the monitored value and the limit estimation result, a limit ratio analysis of the monitored value is carried out and a health status is classified to form the structural health status level of each monitored variable.

[0014] The context confidence domain is refreshed and the deep function mapper is updated according to a preset period to ensure that the context limit mapping relationship reflects the latest monitoring data.

[0015] Optionally, the step of selecting monitoring parameters and unifying the processing of multi-source monitoring variables to construct a dimensionless multidimensional monitoring vector sequence characterizing the dam structure's operational status specifically includes:

[0016] Four core monitoring indicators were selected: horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body. For each type of monitoring indicator, a time-arranged sequence of monitoring values ​​was formed at designated monitoring points on the dam.

[0017] Set a fixed data acquisition cycle, and collect the horizontal displacement, seepage pressure, upstream water level and internal temperature of each monitoring point according to the data acquisition cycle throughout the monitoring period. At each sampling time, construct an original monitoring vector containing the observation values ​​of four physical quantities.

[0018] The original monitoring vectors from all historical sampling times form a historical monitoring dataset. For each monitoring variable, the historical mean and historical standard deviation are calculated on all historical samples of the monitoring variable. The historical mean is the arithmetic mean of all sample observations, and the historical standard deviation is the square root of the arithmetic mean of the squares of the differences between all sample observations and the historical mean.

[0019] Normalization is performed for each sampling time and each monitoring variable. When the historical standard deviation of a monitoring variable is greater than zero, the original value of the monitoring variable is subtracted from the historical average value of the monitoring variable at the corresponding sampling time, and the historical standard deviation of the monitoring variable is used as the denominator for scaling to obtain the dimensionless normalized value of the monitoring variable at the corresponding sampling time. When the historical standard deviation is equal to zero, the normalization result of the monitoring variable at all sampling times is uniformly set to zero.

[0020] The normalized results of the four monitoring variables at the same sampling time are combined in a fixed order to form a four-dimensional dimensionless monitoring vector. The dimensionless vectors of all sampling times are arranged in chronological order to form a unified sequence of dimensionless monitoring vectors for multi-source monitoring variables.

[0021] Optionally, the step of constructing a contextual confidence region using fault-free historical monitoring data and determining the most unfavorable operating condition to form a contextual confidence region reflecting the overall level of the monitored variables specifically includes:

[0022] In the historical monitoring data, a fault-free operating interval is selected. Within this fault-free operating interval, the median time value of each of the four original monitoring variables—horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body—is calculated according to the time series. The four median time values ​​are then combined in a fixed order to form the central vector in the original monitoring space.

[0023] Based on the historical average and historical standard deviation of each monitoring variable, the same normalization process as the historical sample is performed on each component of the central vector. When the historical standard deviation of the corresponding monitoring variable is greater than zero, the dimensionless component of the corresponding monitoring variable in the central vector is obtained by subtracting the historical average and dividing by the historical standard deviation. When the historical standard deviation is equal to zero, the dimensionless component of the corresponding monitoring variable is set to zero, and then the components are combined to form the normalized central vector.

[0024] For each historical dimensionless monitoring vector, calculate the Euclidean distance between the historical dimensionless monitoring vector and the normalized center vector, and arrange all historical distance values ​​in ascending order to obtain an ordered distance sequence.

[0025] Based on the preset confidence level, the distance value located near 95% of the total number of samples in the ordered distance sequence is selected as the context confidence radius, and the hypersphere region formed in the four-dimensional dimensionless monitoring space with the normalized center vector as the geometric center and the context confidence radius as the radius is used as the context confidence domain.

[0026] After the contextual confidence domain is constructed, monitoring samples with a large distance from the normalized center vector are identified within the contextual confidence domain as reference contexts for evaluating the most unfavorable operating conditions.

[0027] Optionally, the step of calculating the upper and lower limit difference vectors and constructing context labels based on the context confidence region and historical monitoring data to obtain a context limit difference label dataset corresponding to each historical monitoring sample specifically includes:

[0028] For each historical sample, the difference between the monitoring vector and the center vector of the historical sample is calculated in the original physical space. The components of the difference in the four directions of horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body are used as the response offset of the historical sample in each monitoring variable direction.

[0029] Statistical analysis was performed on the response offsets of all historical samples. The maximum value was selected from the absolute values ​​of the offsets of all historical samples in each direction of the monitored variable. The maximum absolute offsets in the four directions of the monitored variable were combined in a fixed order to form a global offset reference vector.

[0030] For each historical sample, the confidence margin of the historical sample is calculated based on the distance between the historical sample and the normalized center vector calculated in the normalized space and the context confidence radius. When the distance of the historical sample is less than the context confidence radius, the distance of the historical sample is subtracted from the context confidence radius to obtain a positive confidence margin. When the distance of the historical sample is greater than or equal to the context confidence radius, the confidence margin of the historical sample is set to zero.

[0031] For each historical sample and each monitored variable, according to the proportional relationship between the response offset of the historical sample in the direction of the monitored variable and the global offset benchmark of the corresponding monitored variable, the response offset is proportionally superimposed with the confidence margin of the historical sample to obtain the context limit label value of the monitored variable under the context conditions of the historical sample. When the global offset benchmark of a monitored variable is equal to zero, the context limit label value of the monitored variable on all historical samples is directly set as the response offset of the monitored variable in the original space.

[0032] The contextual limit label values ​​of the four monitoring variables corresponding to each historical sample are combined into a limit label vector, and the dimensionless monitoring vectors of all historical samples are combined with the corresponding limit label vectors in a one-to-one correspondence relationship to form a contextual limit label dataset.

[0033] Optionally, the step of designing a custom depth function mapper structure based on the dimensionless monitoring vectors and contextual limit label dataset, and establishing a depth function mapping relationship that maps the dimensionless monitoring vectors to the limit estimation vectors, specifically includes:

[0034] A deep function mapper is constructed, which takes a four-dimensional dimensionless monitoring vector as input and a four-dimensional limit estimation vector as output. The nonlinear mapping relationship between the input and output is characterized by a set of trainable parameters.

[0035] The deep function mapper is configured as a three-layer structure, where the first layer is a feature stretching layer that receives a four-dimensional dimensionless monitoring vector and generates a first intermediate feature vector with a dimension greater than four dimensions through linear transformation and bias addition, thereby completing the linear stretching of the input features in the high-dimensional space.

[0036] The second layer is set as a nonlinear coupling layer. A fixed parameter vector is introduced into the first intermediate feature vector component by component. Multiplication is performed on each component. At the same time, the hyperbolic tangent function value is calculated for each component of the first intermediate feature vector. The component multiplication result and the hyperbolic tangent function result are added component by component on the same dimension to obtain the second intermediate feature vector containing high-order nonlinear features.

[0037] The third layer is set as the limiting output layer. The second intermediate feature vector is multiplied by the linear transformation matrix of the output layer and the output layer bias vector is added. The results are combined into a four-dimensional limiting estimation vector, forming a mapping relationship from the dimensionless monitoring vector to the limiting estimation vector under a given set of parameters.

[0038] The linear transformation matrices, bias vectors, and parameter vectors in the three-layer structure are used together as the parameter set of the deep function mapper, and this parameter set is updated during training.

[0039] Optionally, the step of designing a non-probabilistic loss function and solving for the parameters of the deep function mapper on the context-bounded label dataset to obtain a parameter set that converges within the engineering accuracy requirements specifically includes:

[0040] Based on the limit estimation vector and the context limit label vector, the normalized error component is calculated for each historical sample and each monitored variable. When the global offset benchmark of the corresponding monitored variable is greater than zero, the difference between the limit estimation value and the context limit label value is normalized according to the global offset benchmark. When the global offset benchmark of the corresponding monitored variable is equal to zero, the error component of the monitored variable is set to zero.

[0041] Construct a non-probabilistic loss function by summing the squares of the error components of all historical samples and all monitored variables according to preset weights in the sample dimension and variable dimension to form a scalar loss value that reflects the fitting bias of the deep function mapper on the context-bound label dataset.

[0042] For each linear transformation matrix and each bias vector in the parameter set, calculate the partial derivative of the loss function with respect to each parameter, and set each partial derivative to zero to form a system of equations for all parameters;

[0043] The system of equations is solved analytically, and the set of parameters that minimizes the loss function is searched. The current minimum loss value is recorded, and the corresponding set of parameters is used as the candidate optimal parameters.

[0044] Set a convergence threshold for the loss function and continuously monitor the change in the current value of the loss function during the solution process. When the decrease in the loss function is lower than the preset change threshold and the current loss value is not higher than the convergence threshold in several consecutive iterations, the parameter set at this time is recorded as the final optimal parameter set, and it is determined that the fitting accuracy of the deep function mapper on the context-bound label dataset meets the engineering requirements.

[0045] Optionally, the step of generating limit estimation results for each monitoring variable using a deep function mapper on newly collected monitoring samples specifically includes:

[0046] At the target moment when structural health assessment is required, raw monitoring values ​​of horizontal displacement, seepage pressure, upstream water level and internal temperature of the dam are collected at the dam monitoring points. The four physical quantities are combined in a fixed order to form the raw monitoring vector of the new input sample.

[0047] Based on the historical average and historical standard deviation of each monitoring variable, normalization is performed on each monitoring variable of the new input sample. When the historical standard deviation of a monitoring variable is greater than zero, the original monitoring value of the monitoring variable at the target time is subtracted from the historical average of the monitoring variable, and the historical standard deviation of the monitoring variable is used as the denominator for scaling. When the historical standard deviation of a monitoring variable is equal to zero, the normalization result of the monitoring variable at the target time is set to zero.

[0048] The normalized results of the four monitoring variables at the target time are combined in a fixed order to form a new four-dimensional dimensionless sample vector;

[0049] The dimensionless new sample vector is input into the deep function mapper that has been trained and obtained the final optimal parameter set. The mapper is calculated sequentially through a three-layer structure and outputs the corresponding four-dimensional limit estimation vector. The four components in the limit estimation vector are used as the limit estimates of horizontal displacement, seepage pressure, upstream water level and internal temperature of the dam at the target time.

[0050] Optionally, the step of conducting a monitoring value limit ratio analysis based on the ratio between the monitored value and the limit estimation result, and classifying health status to form a structural health status level for each monitored variable, specifically includes:

[0051] Based on the center vector calculated within the fault-free historical interval, the differences between the horizontal displacement, seepage pressure, upstream water level and dam internal temperature at the target time and the corresponding components of the center vector are calculated in the original physical space, and the absolute value of the difference is taken as the current deviation of each monitoring variable.

[0052] Based on the limit estimation vector, for each monitored variable, when the corresponding limit estimate is greater than zero, the current deviation is divided by the corresponding limit estimate to obtain the limit ratio of the monitored variable; when the corresponding limit estimate is not greater than zero, the limit ratio of the monitored variable is set to zero.

[0053] The limit ratios of the four monitored variables at the target time are combined into a limit ratio vector in a fixed order;

[0054] The structural health status of each monitored variable is classified according to the limit ratio. When the limit ratio is not greater than 0.6, the monitored variable is classified as normal. When the limit ratio is greater than 0.6 but not more than 0.9, the monitored variable is classified as warning. When the limit ratio is greater than 0.9 but not more than 1, the monitored variable is classified as critical danger. When the limit ratio is greater than 1, the monitored variable is classified as over-limit alarm.

[0055] Optionally, the step of refreshing the context confidence domain and updating the deep function mapper according to a preset period so that the context limit mapping relationship reflects the latest monitoring data specifically includes: setting the refresh period in days;

[0056] Before each refresh moment, new monitoring data since the last refresh is collected, and the new monitoring data is merged with historical fault-free monitoring data to form an updated monitoring dataset.

[0057] Based on the updated monitoring dataset, the historical mean and historical standard deviation of each monitoring variable are recalculated, the median of each monitoring variable within the fault-free historical interval is recalculated, the center vector in the original monitoring space and its central dimensionless vector in the normalized space are updated, the distance sequence between all historical samples and the normalized center vector in the normalized space is recalculated, and the context confidence radius is reselected according to the preset confidence level to construct a new context confidence domain.

[0058] On the updated monitoring dataset, the response offset vector of each sample in the original space is recalculated, the global offset baseline vector is recalculated, and the context limit label vector of each sample is reconstructed based on the new context confidence radius and the new global offset baseline vector to form a new context limit label dataset.

[0059] The nonprobabilistic loss function is reconstructed on the new contextual confidence domain and the parameter solving process is re-executed to obtain a new optimal parameter set. All parameters of the deep function mapper are updated so that the updated contextual confidence domain and the deep function mapper reflect the relationship of dam structural health assessment under the latest monitoring data.

[0060] The present invention has the following beneficial effects:

[0061] 1. Four heterogeneous monitoring indicators—horizontal displacement, seepage pressure, water level, and temperature—are standardized over time. Dimensionless normalization is achieved by calculating historical averages and standard deviations, eliminating the interference of dimensional differences and numerical scales on subsequent models. Unlike existing simple linear normalization or empirical scale conversion, this approach uses the standard deviation of all historical samples to identify null scenarios and uniformly sets variables with a standard deviation of zero to zero, ensuring the robustness of the processing flow. This truly automates and ensures the reliability of data preprocessing, providing accurate and consistent input for subsequent confidence region construction and depth mapping, and solving the challenge of standardizing multi-source heterogeneous data.

[0062] 2. By calculating the median value of each variable in the normalized space and using it to form a central vector, and then selecting a threshold based on the percentile of the historical sample distance distribution, a four-dimensional hypersphere confidence region is constructed. This effectively eliminates outlier data and automatically determines the most unfavorable operating conditions. Unlike traditional methods based on empirical thresholds or single-variable limit values, this scheme automatically generates a confidence region in a multi-dimensional monitoring space, taking into account the coupling relationships between variables, significantly improving the ability to capture extreme operating conditions. It solves the problem of difficulty in identifying implicit coupling patterns in historical data, providing a more objective and adaptive contextual boundary for deviation label construction.

[0063] 3. By statistically analyzing the offset of each historical sample from the center vector and combining it with the global maximum offset benchmark and confidence region margin, context-sensitive limit labels are generated. The innovation lies in combining the global statistical benchmark with the contextual margin to achieve personalized limit assessment at the sample level; simultaneously, special handling is applied to scenarios where the offset benchmark is zero, reverting the label value to the original offset to ensure label consistency and integrity. Compared to conventional unified limit methods, this method considers both global risk and local context, providing a precise supervision target for deep mapper training and improving the relevance and accuracy of limit estimation.

[0064] 4. A three-layer trainable network structure is proposed, comprising a linear feature stretching layer, a nonlinear coupling layer, and a limiting output layer. By combining linear transformations, element-wise multiplication, and hyperbolic tangent activation, it achieves both a full representation of high-dimensional features and a refined limiting output. Unlike general convolutional or fully connected networks, this mapper is structurally customized for the limiting estimation task, reducing network redundancy and improving training convergence speed and inference efficiency. This structure can capture high-order nonlinear coupling relationships in the data and output limiting estimates consistent with the semantics of physical biases, solving the problem that general networks are difficult to adapt to engineering limiting requirements.

[0065] 5. A non-probabilistic loss function based on a global offset benchmark and context labels was designed. A global fitting index is formed by weighted summation of the normalized errors of each sample and variable. The optimal parameter set is then solved analytically or numerically, and a convergence threshold is set. Unlike conventional losses based on mean squared error or cross-entropy, this loss function directly incorporates the offset benchmark and confidence labels from practical engineering needs, preserving both statistical significance and physical interpretability. This design addresses the problems of deep mappers lacking engineering semantic constraints and having unclear training objectives, ensuring that parameter optimization satisfies both model fitting accuracy and practical error limit determination requirements.

[0066] 6. During real-time evaluation, newly acquired raw monitoring vectors are rapidly preprocessed using historical normalization parameters, and dimensionless samples are input into a pre-trained deep function mapper, instantly outputting the limit estimation vector. The key innovation lies in the seamless integration of online new sample processing with offline deep network training, achieving extremely low latency and high accuracy in limit estimation. Unlike traditional methods that require retraining or manual threshold calculation, this method offers real-time response, addressing the challenge of both timeliness and accuracy in limit calculation for dam monitoring systems and enhancing the system's online early warning capabilities.

[0067] 7. The ratio between real-time monitored values ​​and the estimated limit values ​​is calculated. Based on the ratio interval, four states are categorized: normal, warning, critical danger, and over-limit alarm, and corresponding handling suggestions are provided. The innovation lies in the fact that the limit ratio not only reflects the degree of monitoring deviation but also dynamically adjusts the threshold in conjunction with the context of the limit, thus more accurately reflecting the risk level. Unlike fixed threshold grading or univariate assessment, this method has the advantages of adaptability and multivariate coupled judgment, solving the problems of false alarms, missed alarms, and unclear handling suggestions that are prone to occur under static thresholds, and providing more refined decision support for operation and maintenance.

[0068] 8. New monitoring data is collected according to a preset cycle. After merging with historical fault-free data, the normalized parameters, center vector, confidence radius, offset baseline, and context labels are recalculated. This allows for the re-solution of the mapper parameters, enabling online adaptive updates of the model. Its innovation lies in the high coupling between model maintenance and data acquisition processes, avoiding the risk of the model becoming disconnected from the real environment. Unlike traditional methods that rely solely on initial training or manual periodic updates, this method can automatically sense changes in data distribution and self-calibrate, ensuring that the limit mapping and health assessment maintain high accuracy and meet the continuous needs of long-term dam operation and safety management. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Example, refer to Figure 1 A method for assessing the structural health of dams that integrates deep learning, comprising:

[0072] Monitoring parameters were selected and multi-source monitoring variables were unified to construct a dimensionless multidimensional monitoring vector sequence that characterizes the dam structure's operational status.

[0073] By using fault-free historical monitoring data to construct a contextual confidence region and determine the most unfavorable operating conditions, a contextual confidence region reflecting the overall level of the monitored variables is formed.

[0074] Based on the context confidence region and historical monitoring data, the upper and lower limit difference vectors are calculated and the context labels are constructed to obtain the context difference label dataset corresponding to each historical monitoring sample.

[0075] Based on the dimensionless monitoring vector and contextual limit label dataset, a custom depth function mapper structure is designed to establish a depth function mapping relationship that maps dimensionless monitoring vectors to limit estimation vectors.

[0076] Design a nonprobabilistic loss function on a context-bounded label dataset and solve for the parameters of a deep function mapper to obtain a set of parameters that converge within the range of engineering accuracy requirements;

[0077] The depth function mapper is used to generate limit estimates for each monitoring variable on the newly collected monitoring samples;

[0078] Based on the ratio between the monitored value and the limit estimation result, a limit ratio analysis of the monitored value is carried out and a health status is classified to form the structural health status level of each monitored variable.

[0079] The context confidence domain is refreshed and the deep function mapper is updated according to a preset period to ensure that the context limit mapping relationship reflects the latest monitoring data.

[0080] First, the four types of monitoring data—displacement, seepage pressure, water level, and temperature—are unified into a dimensionless multidimensional vector, eliminating the influence of differences in dimensions and scales. Then, a confidence region is automatically constructed within the multidimensional space to determine the most unfavorable operating condition, avoiding the shortcomings of previous univariate or static threshold models in capturing extreme conditions. Limit labels are constructed based on the data context, allowing the deep mapper to consider both global scope and local bias during training, solving the problem of the model training objective being disconnected from actual monitoring needs. In real-time, the trained mapper is used to estimate the limit of new data, and the ratio of monitored values ​​to limit values ​​is mapped to predefined health status levels, achieving full automation from raw data to health assessment results. Finally, a periodic refresh mechanism is introduced to dynamically reconstruct the confidence region and label dataset based on the latest collected data and update the mapper parameters, effectively addressing the risk of model failure caused by changes in environment, operating conditions, and equipment status.

[0081] The process of selecting monitoring parameters and unifying multi-source monitoring variables to construct a dimensionless multidimensional monitoring vector sequence characterizing the dam structure's operational status specifically includes:

[0082] Four core monitoring indicators were selected: horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body. For each type of monitoring indicator, a time-arranged sequence of monitoring values ​​was formed at designated monitoring points on the dam.

[0083] Set a fixed data acquisition cycle, and collect the horizontal displacement, seepage pressure, upstream water level and internal temperature of each monitoring point according to the data acquisition cycle throughout the monitoring period. At each sampling time, construct an original monitoring vector containing the observation values ​​of four physical quantities.

[0084] The original monitoring vectors from all historical sampling times form a historical monitoring dataset. For each monitoring variable, the historical mean and historical standard deviation are calculated on all historical samples of the monitoring variable. The historical mean is the arithmetic mean of all sample observations, and the historical standard deviation is the square root of the arithmetic mean of the squares of the differences between all sample observations and the historical mean.

[0085] Normalization is performed for each sampling time and each monitoring variable. When the historical standard deviation of a monitoring variable is greater than zero, the original value of the monitoring variable is subtracted from the historical average value of the monitoring variable at the corresponding sampling time, and the historical standard deviation of the monitoring variable is used as the denominator for scaling to obtain the dimensionless normalized value of the monitoring variable at the corresponding sampling time. When the historical standard deviation is equal to zero, the normalization result of the monitoring variable at all sampling times is uniformly set to zero.

[0086] The normalized results of the four monitoring variables at the same sampling time are combined in a fixed order to form a four-dimensional dimensionless monitoring vector. The dimensionless vectors of all sampling times are arranged in chronological order to form a unified sequence of dimensionless monitoring vectors for multi-source monitoring variables.

[0087] Further specific implementation steps include:

[0088] Four core monitoring indicators were selected for assessing the structural health of the dam, including: horizontal displacement. osmotic pressure Upstream water level Temperature inside the dam body ; These correspond to the time mapping functions respectively: , , , ;in, For continuous time intervals; , , , At consecutive time points The horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam at a certain monitoring point at that time; Let be the set of all real numbers;

[0089] Set the collection period to every Once every minute;

[0090] At the point of time , Construct the first The original monitoring vector for this observation is:

[0091] ;in, This refers to the time interval for collecting monitoring data; To monitor the start time; For discrete-time indexing; For the first The continuous time corresponding to each sampling; This represents the total number of historical observations. In the first Second sampling time The raw monitoring data includes four physical quantities: displacement, seepage pressure, water level, and temperature. It is a 4-dimensional column vector of real numbers;

[0092] All historical observations Perform linear standardization, specifically:

[0093] S101. For each variable, calculate the historical mean and standard deviation:

[0094] , ;in, This indicates the index of the monitored variable within the vector. Corresponding horizontal displacement, Corresponding osmotic pressure, Corresponding to upstream water level, Corresponding to the internal temperature of the dam body; In the first Each sampling time The The original values ​​of each monitored variable; For the first Each monitoring variable in all Historical average over a sample; For the first Each monitoring variable in all Historical standard deviation over a sample;

[0095] S102, For each pair Perform normalization: And construct the normalized dimensionless monitoring vector:

[0096] ;in, For the first The second sampling, the first The normalized values ​​of the variables; For the first The 4-dimensional dimensionless monitoring vector after sampling and normalization.

[0097] The process of constructing a contextual confidence region using fault-free historical monitoring data and determining the most unfavorable operating condition to form a contextual confidence region reflecting the overall level of the monitored variables specifically includes:

[0098] In the historical monitoring data, a fault-free operating interval is selected. Within this fault-free operating interval, the median time value of each of the four original monitoring variables—horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body—is calculated according to the time series. The four median time values ​​are then combined in a fixed order to form the central vector in the original monitoring space.

[0099] Based on the historical average and historical standard deviation of each monitoring variable, the same normalization process as the historical sample is performed on each component of the central vector. When the historical standard deviation of the corresponding monitoring variable is greater than zero, the dimensionless component of the corresponding monitoring variable in the central vector is obtained by subtracting the historical average and dividing by the historical standard deviation. When the historical standard deviation is equal to zero, the dimensionless component of the corresponding monitoring variable is set to zero, and then the components are combined to form the normalized central vector.

[0100] For each historical dimensionless monitoring vector, calculate the Euclidean distance between the historical dimensionless monitoring vector and the normalized center vector, and arrange all historical distance values ​​in ascending order to obtain an ordered distance sequence.

[0101] Based on the preset confidence level, the distance value located near 95% of the total number of samples in the ordered distance sequence is selected as the context confidence radius, and the hypersphere region formed in the four-dimensional dimensionless monitoring space with the normalized center vector as the geometric center and the context confidence radius as the radius is used as the context confidence domain.

[0102] After the contextual confidence domain is constructed, monitoring samples with a large distance from the normalized center vector are identified within the contextual confidence domain as reference contexts for evaluating the most unfavorable operating conditions.

[0103] Further specific implementation steps include:

[0104] Within the fault-free historical interval, the median value of each original monitoring variable is calculated to construct a central vector in the original monitoring space, specifically as follows:

[0105] ;in, This is a 4-dimensional column vector representing the "center values" of each monitored variable within a fault-free historical period, where each component is the median of the corresponding variable over time. The median operator is used to take the middle value after sorting a given set of real numbers in ascending order. When the number of samples is even, the average of the two middle values ​​is taken.

[0106] At the same time, Mapping to the normalized space yields the normalized center vector. Specifically:

[0107] , ;in, Center vector The The component corresponds to the first component. The median value of each physical quantity; For the normalized center vector at the th Dimensionless components in the dimension; For the normalized central column vector, by composition;

[0108] Constructing the context confidence region in the normalized space is as follows:

[0109] ;in, It is the Euclidean norm; In order to monitor the space in a normalized manner, with the center Center of the sphere, radius is The four-dimensional hypersphere region is called the context confidence region; Let be any 4-dimensional dimensionless monitoring state vector; The context confidence radius in the normalized space;

[0110] For each historical sample, calculate the normalized distance: ,Will Sort in ascending order, and denote the sorted sequence as follows: ;in, For the first Each historical sample is located at the center in the normalized space. Euclidean distance; To all After sorting from smallest to largest, the number The ordered distance between individuals; For sorting position index;

[0111] Take the first Using ordered distances as the normalized context confidence radius, we obtain: .

[0112] The step of calculating upper and lower limit difference vectors based on context confidence regions and historical monitoring data, and constructing context labels to obtain a context limit difference label dataset corresponding to each historical monitoring sample, specifically includes:

[0113] For each historical sample, the difference between the monitoring vector and the center vector of the historical sample is calculated in the original physical space. The components of the difference in the four directions of horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body are used as the response offset of the historical sample in each monitoring variable direction.

[0114] Statistical analysis was performed on the response offsets of all historical samples. The maximum value was selected from the absolute values ​​of the offsets of all historical samples in each direction of the monitored variable. The maximum absolute offsets in the four directions of the monitored variable were combined in a fixed order to form a global offset reference vector.

[0115] For each historical sample, the confidence margin of the historical sample is calculated based on the distance between the historical sample and the normalized center vector calculated in the normalized space and the context confidence radius. When the distance of the historical sample is less than the context confidence radius, the distance of the historical sample is subtracted from the context confidence radius to obtain a positive confidence margin. When the distance of the historical sample is greater than or equal to the context confidence radius, the confidence margin of the historical sample is set to zero.

[0116] For each historical sample and each monitored variable, according to the proportional relationship between the response offset of the historical sample in the direction of the monitored variable and the global offset benchmark of the corresponding monitored variable, the response offset is proportionally superimposed with the confidence margin of the historical sample to obtain the context limit label value of the monitored variable under the context conditions of the historical sample. When the global offset benchmark of a monitored variable is equal to zero, the context limit label value of the monitored variable on all historical samples is directly set as the response offset of the monitored variable in the original space.

[0117] The contextual limit label values ​​of the four monitoring variables corresponding to each historical sample are combined into a limit label vector, and the dimensionless monitoring vectors of all historical samples are combined with the corresponding limit label vectors in a one-to-one correspondence relationship to form a contextual limit label dataset.

[0118] Further specific implementation steps include:

[0119] For each historical sample, calculate its response offset vector relative to the center value in the original physical space, specifically as follows:

[0120] , ;in, For the first Each sample is relative to the center vector in the original physical space. The offset; For the first The sample at the th Offset in the direction of each monitored variable;

[0121] For the offsets of all samples, calculate the historical maximum absolute deviation along the variable direction, and construct a global offset reference vector, specifically as follows:

[0122] , ;in, For the first The sample at the th The absolute value of the directional offset of each monitored variable; This means selecting the first sample from all samples. The maximum absolute value of the offset of each monitored variable; For the first The maximum absolute offset of a monitoring variable across all historical samples; For the reason The resulting column vector serves as the global offset reference vector;

[0123] Using the normalized distance and the global offset baseline vector, a context-dependent limiting label vector is constructed, specifically as follows:

[0124] S301. First, construct the confidence margin coefficients of the samples: ;in, Indicates the first The remaining distance of each sample from the confidence boundary in the normalized space;

[0125] S302. Then, calculate the contextual limit label value for each sample and each monitored variable:

[0126] ;in, In the first Given the context of the nth sample, the nth Contextual limit label values ​​for each monitored variable;

[0127] The contextual limit label dataset is constructed as follows: ;in, This is a context-bounded differential label dataset, where each element is a tuple. ; For the first A 4-dimensional limit label vector for each sample.

[0128] The design of a custom depth function mapper structure based on dimensionless monitoring vectors and contextual limit label datasets, establishing a depth function mapping relationship that maps dimensionless monitoring vectors to limit estimation vectors, specifically includes:

[0129] A deep function mapper is constructed, which takes a four-dimensional dimensionless monitoring vector as input and a four-dimensional limit estimation vector as output. The nonlinear mapping relationship between the input and output is characterized by a set of trainable parameters.

[0130] The deep function mapper is configured as a three-layer structure, where the first layer is a feature stretching layer that receives a four-dimensional dimensionless monitoring vector and generates a first intermediate feature vector with a dimension greater than four dimensions through linear transformation and bias addition, thereby completing the linear stretching of the input features in the high-dimensional space.

[0131] The second layer is set as a nonlinear coupling layer. A fixed parameter vector is introduced into the first intermediate feature vector component by component. Multiplication is performed on each component. At the same time, the hyperbolic tangent function value is calculated for each component of the first intermediate feature vector. The component multiplication result and the hyperbolic tangent function result are added component by component on the same dimension to obtain the second intermediate feature vector containing high-order nonlinear features.

[0132] The third layer is set as the limiting output layer. The second intermediate feature vector is multiplied by the linear transformation matrix of the output layer and the output layer bias vector is added. The results are combined into a four-dimensional limiting estimation vector, forming a mapping relationship from the dimensionless monitoring vector to the limiting estimation vector under a given set of parameters.

[0133] The linear transformation matrices, bias vectors, and parameter vectors in the three-layer structure are used together as the parameter set of the deep function mapper, and this parameter set is updated during training.

[0134] Further specific implementation steps include:

[0135] Construct a depth function mapper for Normalize the monitoring vector Mapped to limit vector estimate ;in, For the parameter set The determined depth function mapper takes a 4-dimensional normalized monitoring vector as input and outputs a 4-dimensional limit estimation vector. This is the set of all parameters that need to be solved. For the depth function mapper For the first Estimates of the limit of each sample;

[0136] The mapper structure adopts a three-layer structure, consisting of a feature stretching layer, a nonlinear coupling layer, and a limiting output layer, as follows:

[0137] S401, First Feature Extension Layer:

[0138] For any input ,set up ;in, This is the first-level linear transformation function, which linearly stretches the input features; The intermediate feature vector output from the first layer has a dimension of . ; Let be the linear mapping matrix of the first layer, with size . of; This is the bias vector for the first layer, with dimensions of... ; Let be the dimension of the intermediate feature space;

[0139] S402, Second Layer (Linear Coupling Layer):

[0140] Constructing nonlinear transformations: ;in, For the second layer of nonlinear coupling transformation, the linear features are... It transforms into more complex higher-order nonlinear characteristics; The intermediate features of the second layer output have a dimension of ; for The product is the element-wise multiplication of two vectors of the same dimension. To Apply the hyperbolic tangent function to each component;

[0141] For any real number The hyperbolic tangent function is defined as follows: ;in, is the base of the natural logarithm;

[0142] S403, Third Layer Limiting Output Layer:

[0143] Projecting nonlinear features onto the bounded space: ;in, The third layer output transformation maps the intermediate features into a 4-dimensional limit estimation vector; It is a column vector representing any input. The corresponding limit estimation results have a dimension of 4; This represents a linear mapping matrix from intermediate features to a 4D output, with size . ; This represents the output layer bias column vector, with a dimension of 4;

[0144] Combining the three layers yields a complete depth function approximator:

[0145] ;in, It is a combinatorial function, representing the sequential passing through... , , A composite mapping used to approximate data from the monitoring context. The mapping relationship to the limit difference vector; This is the parameter set, containing all the matrices and vectors that need to be determined through training.

[0146] The step of designing a non-probabilistic loss function and solving for the parameters of the deep function mapper on the context-bounded label dataset to obtain a parameter set that converges within the engineering accuracy requirements specifically includes:

[0147] Based on the limit estimation vector and the context limit label vector, the normalized error component is calculated for each historical sample and each monitored variable. When the global offset benchmark of the corresponding monitored variable is greater than zero, the difference between the limit estimation value and the context limit label value is normalized according to the global offset benchmark. When the global offset benchmark of the corresponding monitored variable is equal to zero, the error component of the monitored variable is set to zero.

[0148] Construct a non-probabilistic loss function by summing the squares of the error components of all historical samples and all monitored variables according to preset weights in the sample dimension and variable dimension to form a scalar loss value that reflects the fitting bias of the deep function mapper on the context-bound label dataset.

[0149] For each linear transformation matrix and each bias vector in the parameter set, calculate the partial derivative of the loss function with respect to each parameter, and set each partial derivative to zero to form a system of equations for all parameters;

[0150] The system of equations is solved analytically, and the set of parameters that minimizes the loss function is searched. The current minimum loss value is recorded, and the corresponding set of parameters is used as the candidate optimal parameters.

[0151] Set a convergence threshold for the loss function and continuously monitor the change in the current value of the loss function during the solution process. When the decrease in the loss function is lower than the preset change threshold and the current loss value is not higher than the convergence threshold in several consecutive iterations, the parameter set at this time is recorded as the final optimal parameter set, and it is determined that the fitting accuracy of the deep function mapper on the context-bound label dataset meets the engineering requirements.

[0152] Further specific implementation steps include:

[0153] Error components are constructed using the global offset reference vector, specifically as follows:

[0154] ;in, for The One component; for The One component; Global offset reference vector The One component;

[0155] Based on this, the loss function is constructed as follows: ;

[0156] right Take the partial derivatives of each matrix and vector in the equations and establish a system of equations:

[0157] , , , ;in, Represents the zero matrix or zero vector with the same dimension as the corresponding partial derivative;

[0158] By analytically solving the above system of equations, we obtain the following: The set of parameters that yields the minimum value: ;in, To make the loss function The set of parameters that reaches the minimum value is the final set of parameters obtained after training. Indicates that the function Get the minimum value The set of possible values;

[0159] Set the convergence threshold of the loss function to... ,For example ;

[0160] When the minimum loss value obtained during the solution process satisfies: At that time, the accuracy of the mapper's fit to the label dataset meets the engineering requirements.

[0161] The process of generating limit estimates for each monitoring variable using a depth function mapper on newly acquired monitoring samples specifically includes:

[0162] At the target moment when structural health assessment is required, raw monitoring values ​​of horizontal displacement, seepage pressure, upstream water level and internal temperature of the dam are collected at the dam monitoring points. The four physical quantities are combined in a fixed order to form the raw monitoring vector of the new input sample.

[0163] Based on the historical average and historical standard deviation of each monitoring variable, normalization is performed on each monitoring variable of the new input sample. When the historical standard deviation of a monitoring variable is greater than zero, the original monitoring value of the monitoring variable at the target time is subtracted from the historical average of the monitoring variable, and the historical standard deviation of the monitoring variable is used as the denominator for scaling. When the historical standard deviation of a monitoring variable is equal to zero, the normalization result of the monitoring variable at the target time is set to zero.

[0164] The normalized results of the four monitoring variables at the target time are combined in a fixed order to form a new four-dimensional dimensionless sample vector;

[0165] The dimensionless new sample vector is input into the deep function mapper that has been trained and obtained the final optimal parameter set. The mapper is calculated sequentially through a three-layer structure and outputs the corresponding four-dimensional limit estimation vector. The four components in the limit estimation vector are used as the limit estimates of horizontal displacement, seepage pressure, upstream water level and internal temperature of the dam at the target time.

[0166] Further specific implementation steps include:

[0167] For any new moment Collect input samples: ;in, For the current continuous time points that need to be evaluated; For the new moment The 4-dimensional vector of original monitoring values;

[0168] Based on historical mean and standard deviation, the new sample is normalized to construct... :

[0169] ;in, For a new moment The The original values ​​of the monitored variables; For the first in the new sample Dimensionless values ​​of each variable after normalization; Let be the 4-dimensional dimensionless vector of the new sample in the normalized space;

[0170] It can be written in vector form as follows: ;

[0171] Use the pre-trained mapper Output the corresponding limit estimation vector:

[0172] ;in, To use the pre-trained optimal parameter set Perform limit estimation on the new sample; For a new moment 4-dimensional limit estimation vector; For the first in the new sample Limit estimates for each variable.

[0173] The method of conducting limit ratio analysis of monitoring values ​​based on the ratio between the monitored values ​​and the limit estimation results, and classifying health status to form the structural health status level of each monitored variable, specifically includes:

[0174] Based on the center vector calculated within the fault-free historical interval, the differences between the horizontal displacement, seepage pressure, upstream water level and dam internal temperature at the target time and the corresponding components of the center vector are calculated in the original physical space, and the absolute value of the difference is taken as the current deviation of each monitoring variable.

[0175] Based on the limit estimation vector, for each monitored variable, when the corresponding limit estimate is greater than zero, the current deviation is divided by the corresponding limit estimate to obtain the limit ratio of the monitored variable; when the corresponding limit estimate is not greater than zero, the limit ratio of the monitored variable is set to zero.

[0176] The limit ratios of the four monitored variables at the target time are combined into a limit ratio vector in a fixed order;

[0177] The structural health status of each monitored variable is classified according to the limit ratio. When the limit ratio is not greater than 0.6, the monitored variable is classified as normal. When the limit ratio is greater than 0.6 but not more than 0.9, the monitored variable is classified as warning. When the limit ratio is greater than 0.9 but not more than 1, the monitored variable is classified as critical danger. When the limit ratio is greater than 1, the monitored variable is classified as over-limit alarm.

[0178] Further specific implementation steps include:

[0179] By comparing the ratio of the difference between the current monitored value and the center value to the limit estimate, a limit ratio vector for each indicator is constructed. Specifically:

[0180] ;in, For the current sample at the th The values ​​of each variable; Center vector The One component; This provides the limit estimate for the corresponding variable in the current context. For the first in the new sample The limit ratio of each variable;

[0181] Construct a vector from the limit ratios of the four variables in the new sample. ;

[0182] The structural health status of each variable is graded according to the magnitude of the limit ratio:

[0183] when When the monitoring value deviates slightly from the center and is within the safe range, it is considered to be in a normal state and no additional action is required.

[0184] when When the monitoring value deviates significantly but does not approach the limit boundary, it is determined to be in a warning state: the monitoring value deviates significantly but does not approach the limit boundary. It is recommended to appropriately increase the observation frequency of this indicator and conduct a cause analysis in conjunction with the operating conditions.

[0185] when When the monitored value is close to the context limit boundary, it is determined to be a critical dangerous state. It is recommended to immediately arrange on-site verification and structural safety assessment for the part or related working conditions, and prepare necessary control measures.

[0186] when When the monitored value exceeds the context limit, it is determined to be in an over-limit alarm state. This indicates that the indicator has exceeded the historical context allowable range. It is recommended to immediately activate the emergency response plan, including but not limited to measures such as load reduction, water level lowering, strengthening patrols, and implementing special testing.

[0187] The step of refreshing the context confidence domain and updating the depth function mapper according to a preset period, so that the context limit mapping relationship reflects the latest monitoring data, specifically includes: setting the refresh period, with the refresh period set in days;

[0188] Before each refresh moment, new monitoring data since the last refresh is collected, and the new monitoring data is merged with historical fault-free monitoring data to form an updated monitoring dataset.

[0189] Based on the updated monitoring dataset, the historical mean and historical standard deviation of each monitoring variable are recalculated, the median of each monitoring variable within the fault-free historical interval is recalculated, the center vector in the original monitoring space and its central dimensionless vector in the normalized space are updated, the distance sequence between all historical samples and the normalized center vector in the normalized space is recalculated, and the context confidence radius is reselected according to the preset confidence level to construct a new context confidence domain.

[0190] On the updated monitoring dataset, the response offset vector of each sample in the original space is recalculated, the global offset baseline vector is recalculated, and the context limit label vector of each sample is reconstructed based on the new context confidence radius and the new global offset baseline vector to form a new context limit label dataset.

[0191] The nonprobabilistic loss function is reconstructed on the new contextual confidence domain and the parameter solving process is re-executed to obtain a new optimal parameter set. All parameters of the deep function mapper are updated so that the updated contextual confidence domain and the deep function mapper reflect the relationship of dam structural health assessment under the latest monitoring data.

[0192] Further specific implementation steps include:

[0193] As dam service conditions change (such as material aging after years of operation, long-term temperature changes, etc.), the contextual confidence region and the limit mapping relationship need to be updated regularly.

[0194] Set the refresh cycle to For example, the sky This means it refreshes every 30 days;

[0195] Before each refresh time, collect the newly added monitoring data in the most recent cycle and merge it with the historical fault-free data to form a new dataset;

[0196] Recalculate the mean of each variable using the updated dataset. Standard deviation , center vector Normalized center vector and normalized distance sequence and the new contextual confidence radius And compute the new response offset vector on the updated dataset. With the new global offset reference vector And then according to the new , , Reconstruct contextual difference labels ;

[0197] In the new label dataset Reconstruct the loss function Solve for the updated optimal parameter set. This yields a new contextual limit mapping relationship.

[0198] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0199] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for assessing the structural health of dams by incorporating deep learning, characterized in that, include: Monitoring parameters were selected and multi-source monitoring variables were unified to construct a dimensionless multidimensional monitoring vector sequence that characterizes the dam structure's operational status. By using fault-free historical monitoring data to construct a contextual confidence region and determine the most unfavorable operating conditions, a contextual confidence region reflecting the overall level of the monitored variables is formed. Based on the context confidence region and historical monitoring data, the upper and lower limit difference vectors are calculated and the context labels are constructed to obtain the context difference label dataset corresponding to each historical monitoring sample. Based on the dimensionless monitoring vector and contextual limit label dataset, a custom depth function mapper structure is designed to establish a depth function mapping relationship that maps dimensionless monitoring vectors to limit estimation vectors. Design a nonprobabilistic loss function on a context-bounded label dataset and solve for the parameters of a deep function mapper to obtain a set of parameters that converge within the range of engineering accuracy requirements; The depth function mapper is used to generate limit estimates for each monitoring variable on the newly collected monitoring samples; Based on the ratio between the monitored value and the limit estimation result, a limit ratio analysis of the monitored value is carried out and a health status is classified to form the structural health status level of each monitored variable. The context confidence domain is refreshed and the deep function mapper is updated according to a preset period to ensure that the context limit mapping relationship reflects the latest monitoring data.

2. The dam structural health assessment method integrating deep learning according to claim 1, characterized in that, The process of selecting monitoring parameters and unifying multi-source monitoring variables to construct a dimensionless multidimensional monitoring vector sequence characterizing the dam structure's operational status specifically includes: Four core monitoring indicators were selected: horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body. For each type of monitoring indicator, a time-arranged sequence of monitoring values ​​was formed at designated monitoring points on the dam. Set a fixed data acquisition cycle, and collect the horizontal displacement, seepage pressure, upstream water level and internal temperature of each monitoring point according to the data acquisition cycle throughout the monitoring period. At each sampling time, construct an original monitoring vector containing the observation values ​​of four physical quantities. The original monitoring vectors from all historical sampling times form a historical monitoring dataset. For each monitoring variable, the historical mean and historical standard deviation are calculated on all historical samples of the monitoring variable. The historical mean is the arithmetic mean of all sample observations, and the historical standard deviation is the square root of the arithmetic mean of the squares of the differences between all sample observations and the historical mean. Normalization is performed for each sampling time and each monitoring variable. When the historical standard deviation of a monitoring variable is greater than zero, the original value of the monitoring variable is subtracted from the historical average value of the monitoring variable at the corresponding sampling time, and the historical standard deviation of the monitoring variable is used as the denominator for scaling to obtain the dimensionless normalized value of the monitoring variable at the corresponding sampling time. When the historical standard deviation is equal to zero, the normalization result of the monitoring variable at all sampling times is uniformly set to zero. The normalized results of the four monitoring variables at the same sampling time are combined in a fixed order to form a four-dimensional dimensionless monitoring vector. The dimensionless vectors of all sampling times are arranged in chronological order to form a unified sequence of dimensionless monitoring vectors for multi-source monitoring variables.

3. The method for assessing the structural health of a dam by incorporating deep learning according to claim 2, characterized in that, The process of constructing a contextual confidence region using fault-free historical monitoring data and determining the most unfavorable operating condition to form a contextual confidence region reflecting the overall level of the monitored variables specifically includes: In the historical monitoring data, a fault-free operating interval is selected. Within this fault-free operating interval, the median time value of each of the four original monitoring variables—horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body—is calculated according to the time series. The four median time values ​​are then combined in a fixed order to form the central vector in the original monitoring space. Based on the historical average and historical standard deviation of each monitoring variable, the same normalization process as the historical sample is performed on each component of the central vector. When the historical standard deviation of the corresponding monitoring variable is greater than zero, the dimensionless component of the corresponding monitoring variable in the central vector is obtained by subtracting the historical average and dividing by the historical standard deviation. When the historical standard deviation is equal to zero, the dimensionless component of the corresponding monitoring variable is set to zero, and then the components are combined to form the normalized central vector. For each historical dimensionless monitoring vector, calculate the Euclidean distance between the historical dimensionless monitoring vector and the normalized center vector, and arrange all historical distance values ​​in ascending order to obtain an ordered distance sequence. Based on the preset confidence level, the distance value located near 95% of the total number of samples in the ordered distance sequence is selected as the context confidence radius, and the hypersphere region formed in the four-dimensional dimensionless monitoring space with the normalized center vector as the geometric center and the context confidence radius as the radius is used as the context confidence domain. After the contextual confidence domain is constructed, monitoring samples with a large distance from the normalized center vector are identified within the contextual confidence domain as reference contexts for evaluating the most unfavorable operating conditions.

4. The dam structural health assessment method integrating deep learning according to claim 3, characterized in that, The step of calculating upper and lower limit difference vectors based on context confidence regions and historical monitoring data, and constructing context labels to obtain a context limit difference label dataset corresponding to each historical monitoring sample, specifically includes: For each historical sample, the difference between the monitoring vector and the center vector of the historical sample is calculated in the original physical space. The components of the difference in the four directions of horizontal displacement, seepage pressure, upstream water level, and internal temperature of the dam body are used as the response offset of the historical sample in each monitoring variable direction. Statistical analysis was performed on the response offsets of all historical samples. The maximum value was selected from the absolute values ​​of the offsets of all historical samples in each direction of the monitored variable. The maximum absolute offsets in the four directions of the monitored variable were combined in a fixed order to form a global offset reference vector. For each historical sample, the confidence margin of the historical sample is calculated based on the distance between the historical sample and the normalized center vector calculated in the normalized space and the context confidence radius. When the distance of the historical sample is less than the context confidence radius, the distance of the historical sample is subtracted from the context confidence radius to obtain a positive confidence margin. When the distance of the historical sample is greater than or equal to the context confidence radius, the confidence margin of the historical sample is set to zero. For each historical sample and each monitored variable, according to the proportional relationship between the response offset of the historical sample in the direction of the monitored variable and the global offset benchmark of the corresponding monitored variable, the response offset is proportionally superimposed with the confidence margin of the historical sample to obtain the context limit label value of the monitored variable under the context conditions of the historical sample. When the global offset benchmark of a monitored variable is equal to zero, the context limit label value of the monitored variable on all historical samples is directly set as the response offset of the monitored variable in the original space. The contextual limit label values ​​of the four monitoring variables corresponding to each historical sample are combined into a limit label vector, and the dimensionless monitoring vectors of all historical samples are combined with the corresponding limit label vectors in a one-to-one correspondence relationship to form a contextual limit label dataset.

5. The dam structural health assessment method integrating deep learning according to claim 4, characterized in that, The design of a custom depth function mapper structure based on dimensionless monitoring vectors and contextual limit label datasets, establishing a depth function mapping relationship that maps dimensionless monitoring vectors to limit estimation vectors, specifically includes: A deep function mapper is constructed, which takes a four-dimensional dimensionless monitoring vector as input and a four-dimensional limit estimation vector as output. The nonlinear mapping relationship between the input and output is characterized by a set of trainable parameters. The deep function mapper is configured as a three-layer structure, where the first layer is a feature stretching layer that receives a four-dimensional dimensionless monitoring vector and generates a first intermediate feature vector with a dimension greater than four dimensions through linear transformation and bias addition, thereby completing the linear stretching of the input features in the high-dimensional space. The second layer is set as a nonlinear coupling layer. A fixed parameter vector is introduced into the first intermediate feature vector component by component. Multiplication is performed on each component. At the same time, the hyperbolic tangent function value is calculated for each component of the first intermediate feature vector. The component multiplication result and the hyperbolic tangent function result are added component by component on the same dimension to obtain the second intermediate feature vector containing high-order nonlinear features. The third layer is set as the limiting output layer. The second intermediate feature vector is multiplied by the linear transformation matrix of the output layer and the output layer bias vector is added. The results are combined into a four-dimensional limiting estimation vector, forming a mapping relationship from the dimensionless monitoring vector to the limiting estimation vector under a given set of parameters. The linear transformation matrices, bias vectors, and parameter vectors in the three-layer structure are used together as the parameter set of the deep function mapper, and this parameter set is updated during training.

6. The dam structural health assessment method integrating deep learning according to claim 5, characterized in that, The step of designing a non-probabilistic loss function and solving for the parameters of the deep function mapper on the context-bounded label dataset to obtain a parameter set that converges within the engineering accuracy requirements specifically includes: Based on the limit estimation vector and the context limit label vector, the normalized error component is calculated for each historical sample and each monitored variable. When the global offset benchmark of the corresponding monitored variable is greater than zero, the difference between the limit estimation value and the context limit label value is normalized according to the global offset benchmark. When the global offset benchmark of the corresponding monitored variable is equal to zero, the error component of the monitored variable is set to zero. Construct a non-probabilistic loss function by summing the squares of the error components of all historical samples and all monitored variables according to preset weights in the sample dimension and variable dimension to form a scalar loss value that reflects the fitting bias of the deep function mapper on the context-bound label dataset. For each linear transformation matrix and each bias vector in the parameter set, calculate the partial derivative of the loss function with respect to each parameter, and set each partial derivative to zero to form a system of equations for all parameters; The system of equations is solved analytically, and the set of parameters that minimizes the loss function is searched. The current minimum loss value is recorded, and the corresponding set of parameters is used as the candidate optimal parameters. Set a convergence threshold for the loss function and continuously monitor the change in the current value of the loss function during the solution process. When the decrease in the loss function is lower than the preset change threshold and the current loss value is not higher than the convergence threshold in several consecutive iterations, the parameter set at this time is recorded as the final optimal parameter set, and it is determined that the fitting accuracy of the deep function mapper on the context-bound label dataset meets the engineering requirements.

7. The dam structural health assessment method integrating deep learning according to claim 6, characterized in that, The process of generating limit estimates for each monitoring variable using a depth function mapper on newly acquired monitoring samples specifically includes: At the target moment when structural health assessment is required, raw monitoring values ​​of horizontal displacement, seepage pressure, upstream water level and internal temperature of the dam are collected at the dam monitoring points. The four physical quantities are combined in a fixed order to form the raw monitoring vector of the new input sample. Based on the historical average and historical standard deviation of each monitoring variable, normalization is performed on each monitoring variable of the new input sample. When the historical standard deviation of a monitoring variable is greater than zero, the original monitoring value of the monitoring variable at the target time is subtracted from the historical average of the monitoring variable, and the historical standard deviation of the monitoring variable is used as the denominator for scaling. When the historical standard deviation of a monitoring variable is equal to zero, the normalization result of the monitoring variable at the target time is set to zero. The normalized results of the four monitoring variables at the target time are combined in a fixed order to form a new four-dimensional dimensionless sample vector; The dimensionless new sample vector is input into the deep function mapper that has been trained and obtained the final optimal parameter set. The mapper is calculated sequentially through a three-layer structure and outputs the corresponding four-dimensional limit estimation vector. The four components in the limit estimation vector are used as the limit estimates of horizontal displacement, seepage pressure, upstream water level and internal temperature of the dam at the target time.

8. The method for assessing the structural health of a dam by incorporating deep learning according to claim 7, characterized in that, The method of conducting limit ratio analysis of monitoring values ​​based on the ratio between the monitored values ​​and the limit estimation results, and classifying health status to form the structural health status level of each monitored variable, specifically includes: Based on the center vector calculated within the fault-free historical interval, the differences between the horizontal displacement, seepage pressure, upstream water level and dam internal temperature at the target time and the corresponding components of the center vector are calculated in the original physical space, and the absolute value of the difference is taken as the current deviation of each monitoring variable. Based on the limit estimation vector, for each monitored variable, when the corresponding limit estimate is greater than zero, the current deviation is divided by the corresponding limit estimate to obtain the limit ratio of the monitored variable; when the corresponding limit estimate is not greater than zero, the limit ratio of the monitored variable is set to zero. The limit ratios of the four monitored variables at the target time are combined into a limit ratio vector in a fixed order; The structural health status of each monitored variable is classified according to the limit ratio. When the limit ratio is not greater than 0.6, the monitored variable is classified as normal. When the limit ratio is greater than 0.6 but not more than 0.9, the monitored variable is classified as warning. When the limit ratio is greater than 0.9 but not more than 1, the monitored variable is classified as critical danger. When the limit ratio is greater than 1, the monitored variable is classified as over-limit alarm.

9. The method for assessing the structural health of a dam by incorporating deep learning as described in claim 8, characterized in that, The step of refreshing the context confidence domain and updating the depth function mapper according to a preset period, so that the context limit mapping relationship reflects the latest monitoring data, specifically includes: setting the refresh period, with the refresh period set in days; Before each refresh moment, new monitoring data since the last refresh is collected, and the new monitoring data is merged with historical fault-free monitoring data to form an updated monitoring dataset. Based on the updated monitoring dataset, the historical mean and historical standard deviation of each monitoring variable are recalculated, the median of each monitoring variable within the fault-free historical interval is recalculated, the center vector in the original monitoring space and its central dimensionless vector in the normalized space are updated, the distance sequence between all historical samples and the normalized center vector in the normalized space is recalculated, and the context confidence radius is reselected according to the preset confidence level to construct a new context confidence domain. On the updated monitoring dataset, the response offset vector of each sample in the original space is recalculated, the global offset baseline vector is recalculated, and the context limit label vector of each sample is reconstructed based on the new context confidence radius and the new global offset baseline vector to form a new context limit label dataset. The nonprobabilistic loss function is reconstructed on the new contextual confidence domain and the parameter solving process is re-executed to obtain a new optimal parameter set. All parameters of the deep function mapper are updated so that the updated contextual confidence domain and the deep function mapper reflect the relationship of dam structural health assessment under the latest monitoring data.

Images