A Deep Learning Method for Predicting Dam Deformation by Integrating Lag HST and Interpretable Hybrid Convolutional Attention Mechanism

Abstract

This invention proposes a deep learning prediction method for dam deformation that integrates delayed HST (Hyper-Self-Attention) and an interpretable hybrid convolutional attention mechanism, belonging to the field of hydraulic structure health monitoring. The HST model categorizes dam deformation influencing factors into three types: water level, air temperature, and time. It also introduces phase difference analysis and related compensation techniques, significantly reducing the number of model parameters from the traditional nine periodic terms to four core components, thus improving modeling efficiency and accuracy. Based on this, a hybrid deep learning model combining 1D convolution, LSTM, 2D convolution, multi-head self-attention mechanism, and the interpretability method SHAP is constructed to achieve efficient prediction of dam deformation. This model possesses both long-term and short-term response modeling capabilities and long-term sequence stability, and exhibits strong interpretability, making it suitable for dam structural health assessment and early warning in complex environments.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of hydraulic structure health monitoring, and more particularly relates to a dam deformation deep learning prediction method fusing a lag HST and an interpretable hybrid convolution attention mechanism. BACKGROUND

[0002] In recent years, with the continuous development of China's economy and society and the continuous improvement of water resource regulation demand, large-scale hydraulic structures such as concrete dams play an increasingly important role in flood control, water supply, and water energy utilization. As a key infrastructure related to people's life and property safety, the operation safety of a dam directly affects the flood control and disaster reduction of a river basin and the stability of the society and economy. Therefore, accurately and timely grasping the deformation state of a dam and improving the intelligent level of dam health monitoring and early warning have become important scientific and engineering problems to be solved in the field of water conservancy engineering.

[0003] In actual engineering, dam deformation is affected by various factors, including water level fluctuation, temperature change, material aging, and long-term operation effect, and these influences often have complex spatio-temporal coupling and lag characteristics. Traditional deformation prediction methods are mainly based on empirical statistics or physical mechanisms, such as the HST (hydrostatic, temperature, and time effect) model that decomposes displacement, which fits the deformation of a dam by linearly superimposing water level, temperature, and time factors. This kind of method has strong physical interpretability and can reflect the action components of each main control factor, but it is difficult to fully excavate the high-order nonlinear relationships and complex dynamic lag characteristics in the monitoring data, especially in the scenes of severe environmental changes or multi-source heterogeneous data fusion, the prediction accuracy and adaptability are limited.

[0004] During the long-term service of a concrete dam, deformation occurs under the joint action of various environmental factors, including water level, temperature, and time effect. The HST model is an empirical statistical model widely used in the fitting of monitoring data and the decomposition of influencing factors of dam deformation. Its basic idea is to simulate the periodic deformation process through sine and cosine functions. However, the traditional HST model needs to introduce multiple periodic functions (usually up to 9 terms) to cover various lag responses, which has many parameters and poor interpretability, making it difficult to adapt to high-dimensional and strongly nonlinear engineering data characteristics. On the other hand, in recent years, deep learning models based on LSTM have been tried for dam deformation prediction, but traditional LSTM faces problems such as information dilution and prediction fluctuation when dealing with long sequences, and the model lacks physical interpretability, which is not conducive to the promotion of engineering application.

[0005] Therefore, how to effectively fuse the explainability of physical mechanism modeling and the strong expression ability of deep learning, on the basis of fully considering the lag effect of key environmental factors, to realize the accurate prediction of dam deformation and the identification of main control factors, is the research frontier and technical difficulty in the field of dam health monitoring and intelligent early warning. In view of the above problems, a new type of deformation prediction method which can consider feature explainability, data driven learning and lag effect modeling is needed to help the development of dam safety management towards intelligence and scientization. SUMMARY

[0006] The application aims to solve the technical problems that the existing dam deformation prediction method cannot simultaneously consider the explainability of physical mechanism and the modeling ability of deep learning, and the lag effect of key environmental factors is not fully captured, and the prediction accuracy and engineering applicability need to be improved. 1. An improved HST model introducing a phase difference lag modeling mechanism; 2. A hybrid deep learning prediction model combining 1D convolution, LSTM, 2D convolution and attention mechanism based on the improved model.

[0007] In order to achieve the above purpose, the application is realized by adopting the following technical scheme:

[0008] The method comprises:

[0009] The dam monitoring data is preprocessed, including linear interpolation of key column data such as temperature, water level and river direction displacement to fill in missing values, identification and removal of outliers by quartile range method, smoothing of displacement data by Kalman filter, and normalization operation to keep the scale of each feature data consistent;

[0010] The phase difference analysis is performed on the river direction displacement and air temperature, and the river direction displacement and water level, respectively, the lag step corresponding to the maximum correlation point is obtained by cross-correlation method, and the phase compensation is performed on the correlation signal to enhance the correlation between the features and the deformation index;

[0011] Based on the compensated data, a sliding window method is used to construct a feature sequence sample, and four HST features including upstream water level, air temperature, linear time and logarithmic time are selected as model inputs;

[0012] A deep learning model is designed by stacking one-dimensional convolution, long short-term memory network, multi-head attention mechanism, two-dimensional convolution and fully connected regression layer in sequence, wherein the one-dimensional convolution module is used to extract local time sequence features, the long short-term memory network is used to model global time sequence dependence, the multi-head attention mechanism is used to enhance key time segment features, the two-dimensional convolution module is used to fuse and extract high-order spatial patterns, and the fully connected regression module outputs the target prediction value;

[0013] The training set, the validation set and the test set divided in time sequence are used for model training and evaluation, the mean square error is used as the loss function, the Adam optimizer is used to optimize the parameters, the early stopping strategy is introduced to prevent overfitting, and the model performance is comprehensively evaluated by the mean square error, the root mean square error, the mean absolute error and the determination coefficient and other indicators;

[0014] After training, the contribution of different features and time steps to displacement prediction is analyzed according to the model attention weight and the convolution channel response result, and the explainability analysis is realized to assist dam safety monitoring and risk diagnosis.

[0015] In one scheme, linear interpolation is used in the data preprocessing step to fill in the missing values of temperature, water level and river direction displacement, ensuring data continuity and laying a data foundation for subsequent feature extraction and model training, and improving the adaptability of the model to the actual operation scene.

[0016] In one scheme, the identification of abnormal values uses the interquartile range method, which retains valid data points within the upper and lower threshold intervals when performing distribution statistics on temperature, water level, displacement and other data, effectively filtering abnormal noise and measurement errors, and ensuring the stability and accuracy of the input data.

[0017] In one scheme, the Kalman filter smoothing processing is performed on the interpolated or abnormal value removed river direction displacement sequence to output a smoother and clearer displacement curve, which can effectively reduce the interference of random disturbance on the deformation characteristics and highlight the real deformation trend.

[0018] In one scheme, the normalization processing adopts interval normalization method to uniformly transform the key feature values such as water level, air temperature and displacement to a specified standard interval, so that the input features have a unified scale under different dimensions and absolute sizes, providing a good data foundation for deep learning model to efficiently learn multi-source heterogeneous data.

[0019] In one scheme, the phase difference analysis specifically includes quantifying the synchronicity and causal delay relationship between river direction displacement and air temperature, or river direction displacement and water level, determining the time lag between them, and completing phase compensation by adjusting the time position of the feature data, enhancing the physical causal explanation ability and model feature expression ability.

[0020] In one scheme, the sliding window sample construction method is to cut the normalized and compensated monitoring data into fixed length sequences in a fixed step, ensuring that each input sample contains four key variables driven by physical mechanisms, i.e. upstream water level, air temperature, linear time and logarithmic time, realizing the standardization of model input and the explicitness of feature physical meaning.

[0021] In one scheme, the one-dimensional convolution module in the multi-module deep learning prediction model is composed of three one-dimensional convolution structures, which can effectively capture short-term trends and local fluctuation rules, the long short-term memory network module is used to mine and retain long-distance dependence and lag response in the sequence, the multi-head attention mechanism module can automatically assign weights to different time segments, focus on feature windows that contribute significantly to the final prediction, the two-dimensional convolution module further captures high-order correlations across features and time, and the full-connection regression module realizes the fusion of multi-layer features and the output of the final value, and the above series structure significantly improves the prediction accuracy of complex time series deformation data and the explainability of engineering application.

[0022] In one scheme, in the training and evaluation process, time-sequential non-overlapping data sets are used, and early stopping strategy and multi-index comprehensive evaluation method are introduced to realize the best balance of model generalization ability and engineering applicability, ensure that the model does not fail in actual dam health monitoring due to overtraining and data fitting, and improve the reliability of model results in engineering scenarios.

[0023] The present application has the following beneficial effects:

[0024] The present application significantly improves the accuracy and stability of dam deformation prediction by fusing lag HST physical characteristics and explainable hybrid convolution attention mechanism. The method can not only fully mine the complex dynamic lag relationship between dam deformation and main control factors such as temperature, water level, etc., but also automatically extract multi-level spatio-temporal features based on deep learning model to realize high-order representation of deformation process. Through the introduction of explainability analysis, the contribution of input features and time window to the prediction result can be intuitively revealed, which helps engineering and technical personnel to deeply understand the cause mechanism and risk process of dam deformation. In addition, the data preprocessing process of the present application is perfect, has good robustness and wide engineering applicability, can provide intelligent and scientific decision support for dam safety monitoring and operation and management, and promote the development of hydraulic structure health monitoring towards intelligence and high reliability. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 : normalized air temperature and time-varying curve of riverwise displacement value;

[0026] Figure 2 : normalized upstream water level and time-varying curve of riverwise displacement value;

[0027] Figure 3 : feature correlation analysis diagram before phase compensation;

[0028] Figure 4 : feature correlation heat map before phase compensation;

[0029] Figure 5 : feature correlation analysis diagram after phase compensation;

[0030] Figure 6 : correlation heat map of features after phase compensation

[0031] Figure 7 : correlation analysis graph of selected feature items based on fusion lag HST model

[0032] Figure 8 : correlation heat map of selected feature items based on fusion lag HST model

[0033] Figure 9 : mixed convolution-LSTM-attention model structure diagram

[0034] Figure 10 : comparison diagram of prediction results of training set model and actual monitoring data

[0035] Figure 11 : comparison diagram of prediction results of verification set and actual monitoring data

[0036] Figure 12 : comparison diagram of prediction results of test set and actual monitoring data

[0037] Figure 13 : explainable SHAP summary diagram DETAILED DESCRIPTION

[0038] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The drawings show typical embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described in the present application. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0039] Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used in the present application in the specification of the present application are only for the purpose of describing the specific embodiments and are not intended to limit the present application. In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The drawings show typical embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described in the present application. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0040] A dam deformation deep learning prediction method combining lag HST and explainable mixed convolution attention mechanism, the specific process is as follows:

[0041] Step 1, data preprocessing:

[0042] (1) Missing value interpolation: Linear interpolation method interpolate(method='linear') is used to fill in the missing values of temperature, water level and river direction displacement columns to ensure data continuity.

[0043] (2) Abnormal value detection and filtering: The interquartile range (IQR) method is used to identify abnormal values. Normal data points within the Q1-1.5IQR to Q3+1.5IQR interval are retained, which can effectively handle noise or error data that may occur during measurement.

[0044] (3) Kalman filter smoothing and noise reduction processing: Initialize the single-variable Kalman filter to smooth the interpolated displacement data and generate a smoother displacement curve. Kalman filtering can effectively reduce random fluctuations and highlight the true displacement trend.

[0045] (4) Data normalization: Most algorithms are sensitive to feature scales. The original data time series of each index has large differences in fluctuation amplitude due to different dimensions. Normalization can improve model performance. By normalizing the hydrological and structural data (water level, air temperature, displacement) with different dimensions to the same scale, cross-index trend comparison and preparation for deep learning are facilitated, and it is intuitive to observe whether the changes of different factors are synchronized (e.g., whether water level rise is accompanied by displacement increase). In this scheme, MinMaxScaler is used to normalize the specified columns (water level, air temperature, displacement) to the [0, 1] interval. After the above (1) ~ (4) data preprocessing steps, the normalized data changes with the date as shown in Figures 1-2 .

[0046] (5) Phase difference and correlation analysis: The essence of phase difference analysis is to analyze the relationship between signals from the "time offset" dimension, and to convert the apparent correlation into a causal chain or dynamic response process with physical meaning, which plays an irreplaceable role in engineering monitoring, system analysis and prediction decision-making. In order to quantify the delayed impact of environmental factors on displacement and provide time parameters for safety evaluation, this method performs phase difference analysis on the river direction displacement and air temperature, and the river direction displacement and water level: uses cross-correlation method to calculate the cross-correlation function of two signals, finds the lag value corresponding to the maximum correlation point (specific method: traverse the lag range (-n / / 4 to n / / 4, n is the data length); for each lag, calculate the correlation coefficient of x col and y col after shifting; record the lag corresponding to the maximum correlation coefficient as the optimal lag); use Fourier transform method to perform FFT transform on the signals to calculate the phase difference in the frequency domain. Through analysis, it is found that the phase difference between the upstream water level and the river direction displacement is about -168.77 degrees, with a lag time of 0 days and a correlation coefficient of 0.848; the phase difference between air temperature and river direction displacement is about -25.09 degrees, with a river direction displacement lag time of 220 days and a correlation coefficient of 0.643.

[0047] (6) Phase compensation and correlation verification: Based on the correlation analysis results obtained in (5), the input temperature signal was shifted along the time axis by a specified lag of 220 days to achieve phase compensation, and the phase difference after phase compensation was recalculated. The calculated phase difference result after compensation was 0, thus preliminarily verifying the accuracy of the phase difference analysis results. In order to identify the key environmental factors affecting displacement (such as water level as the dominant factor), exploratory data analysis (EDA) was conducted on the dam monitoring data (such as displacement along the river, upstream water level, temperature, etc.) before and after phase compensation. The correlation between variables was revealed through visualization: the sns.pairplot function was used to draw the characteristic correlation analysis diagram to show the correlation between multiple variables, and the sns.heatmap function was used to draw the characteristic correlation heatmap to show the correlation strength between water level, temperature and displacement. Before phase compensation, the characteristic correlation analysis diagram and the characteristic correlation heatmap were as follows: Figures 3-4 As shown, the correlation between downstream displacement and temperature is weak, with a correlation coefficient of -0.18, while the correlation coefficient between downstream displacement and upstream water level is 0.85; Figures 5-6 As shown, after phase compensation, the correlation between downstream displacement and temperature is greatly enhanced, with a correlation coefficient of 0.65.

[0048] After data preprocessing, it is convenient to use for subsequent deep learning processing. At the same time, it can be found that after phase compensation, there is a strong correlation between air temperature and upstream water level and downstream displacement, which is convenient for building a new HST statistical model.

[0049] Step 2: Constructing the HST model considering temperature and water level lag effects:

[0050] Under the combined action of water pressure, temperature, and other loads, a point on a concrete dam will experience a displacement y at time t. t In the traditional HST model, this displacement y t Decomposed into hydrostatic pressure component y H Temperature component y T Time-effect component y θ And the constant term ε, whose expression is as follows:

[0051] y t =y H +y T +y θ +ε(1) where ε represents the zero point of displacement measurement. y H y T y θ The expressions are shown in (2) to (4):

[0052]

[0053]

[0054] y θ = c1θ + c2lnθ (4)

[0055] where a i , b 1i , b 2i , c1 and c2 are model coefficients, which can be solved by multiple linear regression or stepwise regression method. Combining equations (1) to (4), the complete mathematical expression of HST model is as follows:

[0056]

[0057] According to the description of equation (5), when performing deep learning prediction of dam deformation, the input vector contains three influencing factors: water level factor (H, H 2 , H 3 ), temperature factor (sin(ωt), cos(ωt), sin(2ωt), cos(2ωt)) and time effect factor (θ, lnθ). These 9 influencing factors jointly act on the dam displacement, constituting the input variables of the prediction model, which are summarized as follows:

[0058] Input = (x1, x2, x3, x4, x5, x6, x7, x8, x9) = (H, H 2 , H 3 , sin(ωt), cos(ωt), sin(2ωt), cos(2ωt), θ, lnθ)

[0059] Formula (5) does not consider the hysteresis effect of actual air temperature and water level on the riverwise displacement, and the influencing factors considered are numerous and not physically interpretable. Through phase period and correlation analysis in data preprocessing, it is found that the upstream water level and air temperature have strong periodic correlation with the riverwise displacement. After phase compensation, the upstream water level and air temperature at the same step have strong positive correlation with the riverwise displacement, therefore, the direct positive contribution of the upstream water level and air temperature at the current step after phase compensation to the riverwise displacement can be directly considered, and the complete mathematical expression of the HST model integrating the hysteresis is shown in formula (6), which is composed of the phase-compensated hydrostatic pressure component and temperature component, time effect component and constant term. According to formula (6), the feature input of deep learning can be selected more conveniently, which is a total of 4 items, i.e. the phase-compensated hydrostatic pressure, air temperature and 2 time effect items.

[0060] y t = y H-comp + y T-comp + y θ+ epsilon (6)

[0061] Input = (x1, x2, x3, x4) = (H comp , T comp , theta, ln theta)

[0062] The sns.pairplot function is used to draw a feature correlation analysis diagram of the input quantity and the target quantity river direction displacement, and the correlation between the four feature items and the target item is shown. Figure 7 It can be seen that the target item and the four feature items have strong correlation; the sns.heatmap function is used to draw a feature correlation heat map, and the correlation strength between the target item and the feature item is shown. Figure 8 It can be seen that the correlation of upstream water level, air temperature, linear time and logarithmic time with river direction displacement is 0.847, 0.649, 0.483 and 0.488 respectively.

[0063] Step 3, deep learning model training:

[0064] The present application designs and trains a deep learning prediction model integrating convolution, cycle and attention mechanism for the demand of concrete dam deformation prediction, and the model architecture is as shown in Figure 9 , and the specific design is as follows:

[0065] (1) Network structure design:

[0066] The deep learning model is composed of the following modules in sequence:

[0067] One-dimensional convolution module: containing three one-dimensional convolution (1D CNN), the first layer of convolution kernel number is 16, the second layer is 32, and the convolution kernel size of each layer is 3, and appropriate edge padding is adopted to maintain the time step length; the module is used to extract the local time sequence features of the input sequence.

[0068] Cyclic neural network module: containing a layer of long short-term memory network (LSTM), the number of hidden units is 16, which is used to capture the global time sequence dependence relationship of time sequence.

[0069] Multi-head attention mechanism: introduce multi-head attention module (Multi-head Self-attention), strengthen the attention to key time segment features, and improve the modeling ability of the model to the lag effect.

[0070] Two-dimensional convolution module: containing two two-dimensional convolution (2D CNN), the first layer of convolution kernel number is 8, the second layer of convolution kernel number is 12, and the convolution kernel size is 3, which is mainly used to fuse the time sequence features of LSTM output and further extract spatial patterns.

[0071] Fully connected regression module: After the two-dimensional convolutional features are flattened, they are sequentially passed through three fully connected networks (with ReLU activation), and finally a single value prediction is output.

[0072] In the deep learning model training process for concrete dam deformation prediction, a multi-module fusion neural network prediction framework is constructed according to the results of the previous HST feature reconstruction and data preprocessing. The input of the model is the normalized time series sample of five steps and four features obtained after sliding window processing, which corresponds to the upstream water level, temperature, linear time, and log time variables selected by physical mechanism and phase lag compensation. Each input sample is in the form of

[0073] In the network structure design, the model first extracts local features from the input sequence layer by layer through a three-layer one-dimensional convolution module (1D CNN). Let the input be X0, the feature map generated by the first layer convolution be X1 = ReLU(Conv1D k＝3,C＝16 (X0)); the second and third layers are similar, respectively X2 = ReLU(Conv1D k＝3,C＝32 (X1)), X3 = ReLU(Conv1D k＝3,C＝32 (X2)). Each layer of convolution uses appropriate edge padding to keep the time step length unchanged, ensuring that the feature information is not lost. The convolution module can extract local change patterns related to dam deformation, enhancing the robustness to short-term feature disturbances.

[0074] Then, the convolutional feature sequence is input to a one-layer LSTM (Long Short-Term Memory Network) module, with 16 hidden units. Let the input sequence be S = X3, the LSTM recursive calculation is as follows: h t ,c t = LSTM(s t ,h t-1 ,c t-1 )

[0075] Where h t is the hidden state at the current time, c t is the cell state, and LSTM can effectively capture the long-term dependence and lag response characteristics of the dam deformation time series.

[0076] The LSTM output is then input to the multi-head self-attention mechanism (Multi-head Self-attention), which assigns importance weights to each time step feature adaptively, highlighting key influence moments and lag periods. The attention output is:

[0077]

[0078] where (Q, K, V) are query, key, and value vectors, respectively, output by LSTM and obtained by different linear transformations, d k is the dimension scaling factor. The multi-head attention structure enables different sub-space features in the behavior sequence, enhancing the model's ability to model lag-driven and multi-variable synergistic effects.

[0079] The time series features weighted by the attention mechanism are reshaped into a two-dimensional structure and then input into two layers of two-dimensional convolution (2D CNN) in turn. The first layer has 8 convolutional kernels, and the second layer has 12 convolutional kernels, both with a size of 3x3. The two-dimensional convolution operation is represented as:

[0080]

[0081] where Y0 is the reshaped input feature, C l is the number of convolutional kernels in the lth layer. Two-dimensional convolution achieves high-order spatial feature fusion by extracting the interaction mode of time series and features, capturing complex dynamic relationships between variables.

[0082] The two-dimensional convolution output is flattened and then input into a three-layer fully connected network, with the first two layers using ReLU activation and the last layer being a linear output. The mathematical expression of the fully connected regression is:

[0083]

[0084] where z is the flattened feature vector, W i ,b i are the weights and biases of the fully connected layer. The process finally outputs the prediction of the dam's riverwise displacement in the next period

[0085] (2) Input dimension:

[0086] The model input is the HST (Hydro-Seasonal-Temperature) feature sequence extracted by the sliding window. Each input sample has a shape of [window length x feature number], where the window length is 5 and the feature number is 4 (corresponding to upstream water level, air temperature, linear time, and logarithmic time, respectively).

[0087] (3) Output dimension:

[0088] The model output is the prediction of the displacement in the next period (i.e., the next time), which is a single scalar.

[0089] (4) Training and validation:

[0090] The model is trained and evaluated using training, validation, and test sets divided in chronological order. The loss function uses mean squared error (MSE), and the optimizer is Adam. Early stopping strategy is introduced during training to prevent overfitting. Model performance is evaluated on different datasets using indicators such as MSE, RMSE, MAE, and determination coefficient (R 2 ) etc. The comparison results of true value and predicted value of training set, validation set, and test set of river direction displacement are shown in Figures 10-12

[0091] The model training uses training, validation, and test sets divided in chronological order. The training objective function is mean squared error loss (MSE):

[0092]

[0093] where y i is the true value, is the predicted value. The optimizer is Adam, and early stopping (EarlyStopping) strategy is introduced during training to prevent overfitting. Model performance evaluation uses indicators such as MSE, root mean squared error (RMSE), mean absolute error (MAE), and determination coefficient (R 2 ) to comprehensively reflect the generalization ability and prediction accuracy of the model at each stage.

[0094] To achieve the interpretability of the model results, the weight distribution of the multi-head attention layer and the response intensity of the convolution feature channel are visualized after training. The attention distribution function can be represented by the following formula:

[0095]

[0096] where α t is the attention score of each time step, W a ,b a ,v are trainable parameters. By analyzing the weights of each time step and each input feature under the attention mechanism, we can reveal the relative contribution of upstream water level, temperature, and time to the prediction of dam deformation. The experimental results show that the upstream water level feature has the highest weight, indicating its dominant role in the change of dam displacement in river direction, followed by linear time and temperature, and the influence of logarithmic time is relatively weak. The above interpretability analysis not only provides physical evidence for the model prediction results, but also provides effective decision-making reference for the subsequent safety monitoring and risk diagnosis of the dam.

[0097] (5) Interpretability analysis:

[0098] ​After the training is completed, the attention layer weight and the importance of the convolution feature channel in the model are extracted, and the contribution of different time steps and feature channels to the displacement prediction is visualized and explained to assist the engineering application of dam safety monitoring and risk diagnosis. Figure 13 It can be seen that the upstream water level of the second characteristic term has the greatest impact on the dam displacement along the river, followed by the linear time term, and the temperature has a greater impact than the logarithmic time effect.

[0099] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing relevant hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of each method. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), or the like.

[0100] It should be understood that the above detailed description of the technical solutions of the present application by means of preferred embodiments is illustrative rather than limiting. Those skilled in the art can modify the technical solutions recorded in each embodiment on the basis of the present application, or make equivalent substitutions for part of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A dam deformation deep learning prediction method fusing hysteresis HST and an interpretable hybrid convolution attention mechanism, characterized in that: The method comprises: The dam monitoring data is preprocessed, including linear interpolation of temperature, water level and river direction displacement key column data to fill in missing values, identification and removal of abnormal values by quartile range method, smoothing of displacement data by Kalman filter, and normalization to keep the scales of each feature data consistent; Phase difference analysis is performed on river direction displacement and air temperature, and river direction displacement and water level, respectively, the maximum correlation point corresponding to the lag step number is obtained by cross-correlation method, and the phase compensation is performed on the correlation signal to enhance the correlation between the features and the deformation indicators; Based on the compensated data, a sliding window method is used to construct feature sequence samples, and four HST features including upstream water level, air temperature, linear time and logarithmic time are selected as model inputs; A deep learning model is designed by stacking one-dimensional convolution, long short-term memory network, multi-head attention mechanism, two-dimensional convolution and fully connected regression layer in sequence, wherein the one-dimensional convolution module is used to extract local time sequence features, the long short-term memory network is used to model global time sequence dependence, the multi-head attention mechanism is used to enhance key time segment features, the two-dimensional convolution module is used to fuse and extract high-order spatial patterns, and the fully connected regression module outputs the target prediction value; The training set, the validation set and the test set divided in time sequence are used for model training and evaluation, the mean square error is used as the loss function, the Adam optimizer is used to optimize the parameters, the early stopping strategy is introduced to prevent overfitting, and the model performance is evaluated by mean square error, root mean square error, mean absolute error and determination coefficient indicators; After training, the contribution of different features and time steps to displacement prediction is analyzed according to the model attention weight and convolution channel response result.

2. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: The linear interpolation in the data preprocessing is used to fill in the missing values of temperature, water level and river direction displacement, ensuring data continuity.

3. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: The identification of abnormal values adopts the quartile range method, which retains the effective data points within the upper and lower threshold intervals when distributing and counting temperature, water level and displacement data.

4. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: The Kalman filter smoothing processing is used for smoothing and noise reduction of the river direction displacement sequence after interpolation or removal of abnormal values, and a smoother and clearer displacement curve is output.

5. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: The normalization processing adopts interval normalization method to uniformly transform the key feature values of water level, air temperature and displacement to a specified standard interval, so that the input features have uniform scale under different dimensions and absolute sizes.

6. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: The phase difference analysis specifically includes quantifying the synchronicity and causal delay relationship between river direction displacement and air temperature, or river direction displacement and water level, determining the time lag between them, and completing phase compensation by adjusting the time position of the feature data.

7. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: The sliding window sample construction method is to cut the normalized and compensated monitoring data into fixed length sequences in a fixed step, ensuring that each input sample contains four key variables driven by physical mechanisms, i.e. upstream water level, air temperature, linear time and logarithmic time.

8. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: The one-dimensional convolution module in the multi-module deep learning prediction model is composed of three one-dimensional convolution structures, which can effectively capture short-term trends and local fluctuation rules. The long short-term memory network module is used to mine and retain long-distance dependence and lag response in the sequence. The multi-head attention mechanism module can automatically assign weights to different time segments and focus on feature windows that contribute significantly to the final prediction. The two-dimensional convolution module further captures high-order correlations across features and time. The fully connected regression module realizes the fusion of multi-layer features and the output of the final value.

9. The dam deformation deep learning prediction method of fusing hysteresis HST and interpretable hybrid convolution attention mechanism according to claim 1, characterized in that: In the training and evaluation process, time-sequential non-overlapping data sets are used, and early stopping strategies and multi-index comprehensive evaluation methods are introduced.

Images