Dam deformation prediction method based on dynamic fusion of multi-source data and related products
By combining goal-oriented dynamic weight fusion and modal aliasing criterion-driven selective decomposition with a hybrid optimization strategy, the problem of insufficient model accuracy and stability in dam deformation prediction is solved, achieving high-precision and robust prediction results that are adaptable to deployment in different computing resource environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN SHUIFA SURVEY DESIGN & RES CO LTD
- Filing Date
- 2026-03-03
- Publication Date
- 2026-05-19
AI Technical Summary
In existing dam deformation prediction methods, static fusion of multi-source data and decomposition of single signals result in low signal-to-noise ratio of model input, masking the dynamic importance of key influencing factors, limiting the accuracy and stability of prediction models, making neural network parameter optimization prone to getting trapped in local optima, and lacking intelligent triggering and adaptive adjustment capabilities.
A goal-oriented dynamic weight fusion, selective quadratic mode decomposition based on the mode mixing criterion, and a hybrid optimization strategy triggered by fitness stagnation state are adopted to construct a high-precision and robust dam deformation prediction system. This system includes dynamic feature weight matrix update, ICEEMDAN algorithm decomposition, and condition-triggered hybrid optimization neural network parameter optimization.
It improves the accuracy and stability of dam deformation prediction. By extracting high-purity feature components through dynamic weight fusion and selective decomposition, it ensures the global optimality of model parameters, reduces prediction errors, and adapts to the deployment needs of different computing environments.
Smart Images
Figure CN121765290B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water conservancy project safety monitoring and data analysis technology, specifically to a method for predicting dam deformation based on dynamic fusion of multi-source data and related products. Background Technology
[0002] As the core structure of a water conservancy project, the safe operation of a dam is of paramount importance. In dam safety monitoring systems, deformation monitoring data is the most direct and crucial data for detecting the dam's structural behavior. With the development of science and technology, current dam deformation prediction mainly relies on various monitoring data (such as reservoir water level, temperature, and displacement) and data analysis techniques. The mainstream solutions and their corresponding shortcomings are summarized below:
[0003] Often, multi-source data (external displacement, internal stress, environmental factors) are simply spliced together or fused with equal weights. Although weighting based on correlation analysis is occasionally used, it fails to dynamically and non-linearly reflect the differentiated contributions of different factors to deformation, especially neglecting the lag effect of key factors such as reservoir water level. Equal weighting or static weighting fusion results in a low signal-to-noise ratio of the model input, masks the dynamic importance of key influencing factors, and leads to insufficient quality of source information in prediction.
[0004] Single-signal decomposition methods, such as Empirical Mode Decomposition (EMD) or Variational Mode Decomposition (VMD), are commonly used. EMD suffers from mode aliasing, while VMD is sensitive to preset parameters (such as the number of modes K). Single methods are insufficient to adaptively obtain trend, periodic, and residual components with clear physical meaning and high purity. The modal components (IMFs) generated by single decomposition methods are often mixed with noise or signals of different scales, causing the residuals to fail to meet the white noise assumption, which directly affects the accuracy and stability of subsequent prediction models.
[0005] Single models (ARIMA, SVM) or "decomposition-single prediction" combined models (such as EMD+BP neural network) are often used. Although BP neural networks are commonly used, their inherent slow convergence and susceptibility to local optima have not been fundamentally solved, and the model is not very specific to different characteristic components. The parameter optimization of prediction models (especially neural networks) is prone to getting trapped in local optima, and existing hybrid optimization strategies lack intelligent triggering and adaptive adjustment capabilities, thus limiting optimization efficiency and final performance.
[0006] To improve backpropagation (BP) networks, particle swarm optimization (PSO) or genetic algorithms (GA) are often introduced for parameter optimization. However, traditional PSO is prone to premature convergence, GA has low global search efficiency, and is mostly a fixed serial structure, lacking a mechanism to dynamically switch strategies according to the optimization process, making it difficult to balance global exploration and local development. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention provides a dam deformation prediction method and related products based on dynamic fusion of multi-source data. It realizes a high-precision, robust, and engineering-deployable dam deformation prediction system by using goal-oriented dynamic weight fusion as input, selective quadratic mode decomposition driven by aliasing discrimination as the core signal processing method, and a hybrid optimization strategy triggered by fitness stagnation state as the key parameter optimization mechanism.
[0008] This invention is achieved through the following technical solution:
[0009] A method for predicting dam deformation based on dynamic fusion of multi-source data includes the following steps:
[0010] Acquire multi-source monitoring data during dam operation, and construct an extended feature set after preprocessing;
[0011] Based on the correlation and time-varying characteristics between each feature in the extended feature set and the dam deformation response, a target-oriented feature weight matrix is constructed and dynamically updated. A weighted fusion input sequence is generated based on the updated weight matrix.
[0012] The weighted fusion input sequence is subjected to initial mode decomposition. The mode aliasing criterion is used to identify the aliased mode components. The identified aliased mode components are subjected to secondary mode decomposition. All the components obtained after decomposition are recombined into three types of components with different statistical properties.
[0013] Based on the statistical characteristics of the three types of components, corresponding prediction models are matched to perform prediction modeling to obtain prediction results for each type of component.
[0014] The prediction results of various components are fused together to output the dam deformation prediction results.
[0015] Optionally, methods for constructing extended feature sets specifically include:
[0016] Obtain the reservoir water level sequence, ambient temperature sequence, time factor sequence and their corresponding lag term sequence of the dam, and construct the original feature vector;
[0017] The original feature vector is subjected to anomaly detection and repair using statistical criteria and Grubbs test; for mild outliers and missing values, cubic spline interpolation is used for repair; for severe outliers, manual confirmation or trend correction is performed.
[0018] A second-order interactive feature is constructed based on the repaired original feature vector, and the second-order interactive feature is combined with the original feature vector to form an extended feature set.
[0019] Optionally, the method for constructing and dynamically updating the goal-oriented feature weight matrix includes:
[0020] Calculate the correlation coefficients between each feature in the extended feature set and the dam deformation response sequence: if the data satisfies the normal distribution assumption, use the Pearson correlation coefficient; otherwise, use the Spearman rank correlation coefficient.
[0021] Based on the calculated correlation coefficient, a judgment matrix is constructed using the analytic hierarchy process. The eigenvector corresponding to the largest eigenvalue is calculated, and after satisfying the consistency constraint, normalization is performed to obtain the initial weight vector.
[0022] A sliding time window mechanism was established to calculate the change in the correlation coefficient between each feature and the dam deformation response sequence within adjacent time windows;
[0023] The change in the correlation coefficient is compared with a preset change threshold: when the change in the correlation coefficient exceeds the preset change threshold, the weight of the feature is incrementally updated using the learning rate and the change in the correlation coefficient; when the change in the correlation coefficient does not exceed the preset change threshold, the current weight of the feature remains unchanged.
[0024] Optionally, the specific steps for performing secondary mode decomposition include:
[0025] The weighted fusion input sequence is initially decomposed using the ICEEMDAN algorithm to obtain several intrinsic mode components.
[0026] Calculate the sample entropy and bandwidth of each intrinsic mode component;
[0027] Establish a joint rule for mode aliasing: when the sample entropy value of a certain intrinsic mode component is greater than a preset entropy threshold and its bandwidth is greater than the center frequency of the intrinsic mode component, the intrinsic mode component is determined to be a mode component with aliasing.
[0028] The VMD algorithm is used to perform secondary decomposition on the modal components that are determined to have aliasing, and the optimal number of modes is determined by the minimum envelope entropy criterion.
[0029] Optionally, the sub-mode components obtained from the second decomposition and the intrinsic mode components that have not aliased are recombined into three types of components with different statistical properties based on the characteristic of frequency from low to high: trend components, periodic components and residual components.
[0030] For the aforementioned trend components, a regression-based prediction model is matched;
[0031] For the periodic component, a time series prediction model is matched;
[0032] For the residual components, a matching neural network prediction model is used.
[0033] Optionally, a conditionally triggered hybrid optimization strategy based on fitness stagnation states is used to optimize the parameters of the neural network prediction model, specifically including:
[0034] Step A: Initialize the population for the particle swarm optimization algorithm and construct the fitness function;
[0035] Step B: Perform particle swarm optimization iteration and calculate the global optimal fitness value for the current iteration based on the fitness function;
[0036] Step C: If the change in the global optimal fitness value is lower than a preset threshold in a preset number of consecutive iterations, the optimization process is determined to have entered a fitness stagnation state.
[0037] Step D: When the fitness stagnation state is determined, the Levy flight perturbation mechanism is triggered to update the position of the current global best particle;
[0038] Step E: If the fitness remains stagnant after triggering the Levi flight perturbation mechanism, trigger the genetic mutation mechanism;
[0039] Step F: Repeat steps B to E until the preset maximum number of iterations is reached or the convergence condition is met, and output the neural network prediction model parameters corresponding to the optimal particle position.
[0040] Optionally, the neural network prediction model is: ,in, The dimension of the input feature. Let i be the i-th component of the input feature vector. The number of neurons in the hidden layer. The weights are used to connect the i-th feature of the input layer to the j-th neuron of the first hidden layer. This is the bias term for the j-th neuron in the first hidden layer. The activation function for the hidden layer is... The weights connecting the j-th neuron in the second hidden layer to the output neuron. This is the bias term for the output neuron. The activation function for the output layer;
[0041] The fitness function includes a mean squared error term and an L2 regularization term, specifically: ,in, The number of training samples, For the true value, For predicted values, For model parameters, The regularization coefficient is used.
[0042] Criteria for determining fitness stagnation: Record the first... The global optimal fitness value in the next iteration Calculate the change between adjacent iterations. When continuous The next iteration satisfies When the time comes, it is determined that the system has entered a standstill state, where, The preset threshold;
[0043] The specific mechanism for triggering the Lévy flight perturbation is as follows: a global perturbation is performed using an improved Lévy distribution random step size to update the particle position. The update formula is: ,in, Let be the current position of the particle at time t. This represents the current position of the globally optimal particle. Let be a random vector that follows a Lévy distribution. This is the inertia weighting coefficient. For the number of iterations The decay step size, The initial step size, As a social learning factor, This represents the maximum number of iterations.
[0044] The specific mechanism for triggering genetic mutation is as follows: performing a Gaussian mutation operation on some individual parameters in the population, with the mutation formula being: ,in, These are the mutated parameters. With a mean of 0 and a variance of Gaussian distributed random noise.
[0045] Optionally, depending on the computing resources and accuracy requirements of the actual application scenario, the fusion mode of the prediction results includes:
[0046] Mode 1: High-precision fusion mode, which adopts a stacked ensemble learning strategy, uses the prediction results of each component as meta-features, and trains a meta-learner to perform nonlinear fusion, outputting the dam deformation prediction results.
[0047] Mode 2: High-efficiency fusion mode, which adopts a dynamic weighted linear fusion strategy. Based on the fusion weight of the prediction model of each component, the prediction results of each component are linearly weighted and summed to output the dam deformation prediction result.
[0048] A computer-readable storage medium storing a computer program that, when executed by a processor, implements a dam deformation prediction method based on dynamic fusion of multi-source data as described above.
[0049] A computer program product includes a computer program / instructions that, when executed by a processor, implement a dam deformation prediction method based on dynamic fusion of multi-source data as described above.
[0050] Compared with the prior art, the present invention has the following features and beneficial effects:
[0051] This invention solves the problem that traditional static weighting cannot reflect the actual load state of a dam by constructing a goal-oriented dynamic update mechanism for the feature weight matrix. By adopting a selective quadratic decomposition strategy based on the modal aliasing criterion, it effectively eliminates modal aliasing and extracts high-purity feature components while avoiding redundant calculations of non-aliased components. By performing hybrid optimization on the neural network parameters, it balances convergence speed and population diversity, ensuring the global optimality of model parameters and reducing prediction errors. Attached Figure Description
[0052] The accompanying drawings illustrate exemplary embodiments of the present invention and, together with the description thereof, serve to explain the principles of the invention. These drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, but do not constitute a limitation on the embodiments of the present invention.
[0053] Figure 1 This is a flowchart illustrating the dam deformation prediction method based on dynamic fusion of multi-source data according to the present invention.
[0054] Figure 2 This is a schematic diagram of the process for constructing and dynamically updating the target-oriented feature weight matrix in Embodiment 2 of the present invention.
[0055] Figure 3 This is a flowchart illustrating the selective quadratic mode decomposition based on the modal aliasing criterion in Embodiment 2 of the present invention.
[0056] Figure 4 This is a schematic diagram of the process of differential prediction modeling based on statistical characteristics in Embodiment 2 of the present invention.
[0057] Figure 5 This is a schematic diagram of the parameter optimization process for a neural network prediction model according to Embodiment 2 of the present invention.
[0058] Figure 6 This is a schematic diagram of the prediction result fusion process in Embodiment 2 of the present invention. Detailed Implementation
[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.
[0060] It should also be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.
[0061] Where there is no conflict, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0062] Example 1
[0063] like Figure 1 As shown, this embodiment provides a dam deformation prediction method based on dynamic fusion of multi-source data, following the approach of "data perception—dynamic fusion—refined decomposition—component modeling—global integration", specifically including the following steps:
[0064] Step 1: Acquisition of multi-source monitoring data and construction of extended feature sets.
[0065] Multi-source monitoring data during dam operation were acquired and preprocessed to construct an extended feature set.
[0066] Multi-source monitoring data typically includes environmental quantities (such as water level and temperature) and effect quantities (such as deformation and displacement). After acquiring the raw data, necessary preprocessing (such as cleaning and noise reduction) is performed, and then the data is expanded to obtain an extended feature set.
[0067] Step 2: Construction and dynamic updating of the goal-oriented feature weight matrix.
[0068] Based on the correlation and time-varying characteristics between each feature in the extended feature set and the dam deformation response, a target-oriented feature weight matrix is constructed and dynamically updated. A weighted fusion input sequence is then generated based on the updated weight matrix.
[0069] Correlation analysis refers to analyzing the degree of correlation between each feature in the extended feature set and the predicted target (i.e., dam deformation response).
[0070] Time-varying characteristics refer to the fact that the contribution of various influencing factors to the deformation of the dam varies in different seasons or operating stages (for example, the impact of water level on deformation during the flood season may be different from that during the dry season).
[0071] Based on the above analysis, a "feature weight matrix" that can dynamically adjust weights over time is constructed. This matrix is used to weight the input data, generating a "weighted fusion input sequence," thereby highlighting key factors at critical moments and suppressing interference from irrelevant information.
[0072] Step 3: Selective quadratic mode decomposition based on mode aliasing criterion.
[0073] The weighted fusion input sequence is subjected to an initial mode decomposition, which decomposes it into several mode components. Then, the mode aliasing criterion is used to identify the mode components with aliasing. The identified mode components with aliasing are subjected to a second mode decomposition. All the components obtained after decomposition are recombined into three types of components with different statistical properties.
[0074] The presence of aliasing indicates that the modal components are still mixed with signals of different scales. Then, a second mode decomposition is performed on these signals, and the pure components are retained.
[0075] All processed components (including the un-aliased primary components and the sub-components obtained from the secondary decomposition) are recombined into three types of components according to their frequency or statistical regularity. These typically include: low-frequency stable trend components, mid-frequency periodic components with significant periodicity, and high-frequency random residual components.
[0076] Step 4: Differential prediction modeling based on statistical characteristics.
[0077] Based on the statistical characteristics of the three types of components, corresponding prediction models are matched to perform prediction modeling to obtain prediction results for each type of component.
[0078] The sub-mode components obtained from the second decomposition and the intrinsic mode components that have not aliased are recombined into three types of components with different statistical properties based on the characteristic of frequency from low to high: trend components, periodic components and residual components.
[0079] For the aforementioned trend component, the degree of nonlinearity is relatively low, matching the regression-type prediction model; the preferred model is the autoregressive integral moving average (ARIMA) model or the linear regression model, which is good at capturing linear trends and low-order nonlinear changes and can accurately fit the long-term trend of dam deformation.
[0080] For the aforementioned periodic components, which exhibit seasonal patterns, a time series forecasting model is suitable. The preferred model is the seasonal autoregressive moving average (SARIMA) model, which introduces a seasonal difference term based on ARIMA, effectively solving the problem that ordinary regression models cannot handle periodic structures.
[0081] For the residual components, which contain high-frequency noise and nonlinear dynamic characteristics, a neural network prediction model is matched. Through the nonlinear mapping and adaptive capabilities of the neural network, the prediction error is reduced.
[0082] Step 5: Fusion of prediction results.
[0083] The prediction results of various components are fused to output the dam deformation prediction result. The result fusion can adopt any of the following modes: a) High-precision fusion mode: nonlinear fusion is performed using stacked ensemble learning; b) Efficient fusion mode: dynamic weighted linear fusion is performed.
[0084] Example 2
[0085] This embodiment provides a detailed description of the method in Embodiment 1.
[0086] Step 1: Acquisition of multi-source monitoring data and construction of extended feature sets, such as... Figure 2 As shown.
[0087] In order to comprehensively capture the influence of various environmental loads and structural characteristics on dam deformation, the reservoir water level sequence, ambient temperature sequence, time factor sequence and their corresponding lag term sequence of the dam are obtained to construct the original feature vector; ,in, Representing the Several monitoring variables (such as current water level, previous period's water level, etc.) (water level, current temperature, etc.) The total dimension of the original features.
[0088] The original monitoring data contained outliers and missing data. Statistical criteria and Grubbs tests are used to detect and repair anomalies in the original feature vector.
[0089] When data points satisfy However, when data deviates from the normal range or is directly missing, it is judged as a slight anomaly / missing value. For the detected slight anomalies and missing values, cubic spline interpolation is used for repair. That is, piecewise cubic polynomials are used to approximate the data curve to ensure the continuity of the first and second derivatives at the repair points. The repair formula is expressed as: ,in, This is the data value after repair.
[0090] When data points satisfy When a severe anomaly is detected, it is determined to be a serious anomaly. For the detected serious anomaly values, manual confirmation or trend correction is performed. The system marks such data and submits it for manual confirmation. If it is confirmed as an error, correction is performed based on the changing trend of neighboring data.
[0091] To capture the interactions between variables, second-order interaction features are constructed based on the repaired original feature vectors. ,in, and These are the components in the original feature vector.
[0092] The second-order interactive features are combined with the original feature vectors to form an extended feature set.
[0093] Step 2: Construction and dynamic updating of the goal-oriented feature weight matrix, such as... Figure 2 As shown.
[0094] Considering that different environmental factors have different degrees of influence on dam deformation, a feature weight matrix that can adaptively adjust over time is constructed to generate a weighted fusion input sequence that reflects the real-time loading state of the dam.
[0095] To quantify the correlation between each feature and dam deformation, the correlation coefficients between each feature in the extended feature set and the dam deformation response sequence were calculated:
[0096] Perform a normality test on the data series. If the data satisfies the normal distribution assumption, use the Pearson correlation coefficient. ,in, Features With deformation response covariance, and These are the standard deviations of the two values, respectively.
[0097] If the data does not satisfy the normal distribution assumption, use the Spearman rank correlation coefficient. ,in, This represents the rank sequence after sorting the data sequence. This is a correlation calculation function.
[0098] Based on the calculated correlation coefficients, the initial weight vectors of each feature are determined using the analytic hierarchy process (AHP). Construct a judgment matrix Matrix elements Reflects characteristics Relative to features The importance of .
[0099] Calculate the judgment matrix Maximum eigenvalue They and their corresponding feature vectors are analyzed, and a consistency check is performed to calculate the consistency index. and consistency ratio : , ,in, For the number of features, This is the average random consistency index.
[0100] when At this point, the judgment matrix is determined to satisfy the consistency constraint. Then, the eigenvectors are normalized to obtain the initial weights for each feature. : ,in, The first eigenvector Each component.
[0101] A sliding time window mechanism is introduced to dynamically update the weights online, with the length of the sliding time window set to [value missing]. At every moment Calculate each feature within the current window With deformation response Local correlation coefficient .
[0102] Calculate the change in correlation coefficient between adjacent time windows. : .
[0103] Set preset change threshold ,Will and A comparison is performed to determine whether to update the weights.
[0104] When the change in the correlation coefficient exceeds the preset change threshold... The weights of this feature are incrementally updated using the learning rate and the change in the correlation coefficient: ,in, This is the sign function, used to determine the direction of weight adjustment; The learning rate parameter is used to control the update step size.
[0105] When the change in the correlation coefficient does not exceed the preset change threshold Keep the current weight of this feature unchanged, that is .
[0106] Based on the dynamic weight matrix obtained from the above steps, the feature components in the extended feature set are weighted and combined to construct the final weighted fusion sequence used as the model input. : ,in, To expand the dimensionality of the feature set, To expand the first feature set Each feature component.
[0107] Step 3: Selective quadratic mode decomposition based on mode aliasing criterion.
[0108] like Figure 3 As shown, the ICEEMDAN algorithm is used to perform an initial decomposition of the weighted fused input sequence. ICEEMDAN obtains several intrinsic mode components by adaptively adding white noise and calculating the local mean of the residuals during the decomposition process; the original sequence is expressed as... One intrinsic mode function (IMF) and one residual term Linear superposition: ,in, Representing the Each intrinsic mode component.
[0109] To identify the aliased modes that require secondary decomposition, the sample entropy and bandwidth of each intrinsic mode component are calculated. The sample entropy is used to quantify the complexity of the time series, and the bandwidth is used to measure the dispersion of the signal frequency components.
[0110] Set sample entropy threshold For each initial decomposition obtained A joint rule for determining modal aliasing is established: when the sample entropy value of an intrinsic modal component is greater than a preset entropy threshold, and its bandwidth is greater than the center frequency of that intrinsic modal component, the intrinsic modal component is determined to be an aliased modal component; that is, and ,in, Let be the center frequency of the intrinsic mode component. A component is considered to exceed the limit if and only if both its complexity and bandwidth exceed the limit. For modal components that exhibit aliasing.
[0111] The VMD algorithm is used to perform a secondary decomposition on the modal components identified as having aliasing, and the optimal number of modes is determined by the minimum envelope entropy criterion. VMD aims to find... Modal functions To minimize the sum of the estimated bandwidths for each mode, the constraint variational equation is as follows: .
[0112] The constraint is that the sum of all submodes equals the aliasing component to be decomposed. : ,in, The center frequencies of each mode are... For the Dirac function, This represents the convolution operation.
[0113] To achieve adaptive decomposition, this embodiment uses the minimum envelope entropy criterion to determine the optimal value. value, ,in, This represents the normalized probability distribution of the modal component envelope signal. By traversing different... Choose the value that minimizes the envelope entropy. As the final decomposition level.
[0114] Step 4: Differential prediction modeling based on statistical characteristics.
[0115] like Figure 4 As shown, based on their frequency characteristics (from low to high) and physical meaning, they are classified and reorganized, and matched with corresponding prediction models:
[0116] Trend component: Composed of low-frequency components, exhibiting a stable long-term trend. A linear regression model or an autoregressive integral moving average (ARIMA) model is used.
[0117] Periodic component: Reconstructed from mid-frequency components, exhibiting significant periodic oscillation characteristics. A seasonal autoregressive moving average (SARIMA) model is employed.
[0118] Residual components: composed of high-frequency components, exhibiting randomness and nonlinearity. A neural network prediction model is employed.
[0119] For the residual components, the output expression of the constructed neural network prediction model is: ,in, The dimension of the input feature. Let i be the i-th component of the input feature vector. The number of neurons in the hidden layer. The weights are used to connect the i-th feature of the input layer to the j-th neuron of the first hidden layer. This is the bias term for the j-th neuron in the first hidden layer. The activation function for the hidden layer is... The weights connecting the j-th neuron in the second hidden layer to the output neuron. This is the bias term for the output neuron. The activation function for the output layer;
[0120] like Figure 5 As shown, this embodiment provides a method for parameter optimization of a neural network prediction model based on a conditionally triggered hybrid optimization strategy using fitness stagnation states, specifically including:
[0121] Step A: Initialize the population for the particle swarm optimization algorithm and construct the fitness function; the fitness function includes a mean squared error term and an L2 regularization term, specifically: ,in, The number of training samples, For the true value, For predicted values, For model parameters, This is the regularization coefficient.
[0122] Step B: Perform particle swarm optimization iteration, calculate the global optimal fitness value of the current iteration based on the fitness function, and update the individual extreme values and the global extreme values.
[0123] Step C: In order to determine whether the algorithm is trapped in a local optimum, a stagnation criterion is established; if the change in the global optimal fitness value is lower than a preset threshold in a preset number of consecutive iterations, the optimization process is determined to have entered a fitness stagnation state.
[0124] Criteria for determining fitness stagnation: Record the first... The global optimal fitness value in the next iteration Calculate the change between adjacent iterations. When continuous The next iteration satisfies When the time comes, it is determined that the system has entered a standstill state, where, This is a preset threshold.
[0125] Step D: When the fitness stagnation state is determined, the Levy flight perturbation mechanism is triggered to update the position of the current global best particle;
[0126] When it is determined that a stagnant state has been entered (e.g., continuous...), (Second stagnation), triggering the Lévy flight perturbation mechanism, updates the position of the currently globally optimal particle. This involves using an improved Lévy distribution with a random step size for global perturbation, updating the particle position using the following formula: ,in, Let be the current position of the particle at time t. This represents the current position of the globally optimal particle. Let be a random vector that follows a Lévy distribution. This is the inertia weighting coefficient. For the number of iterations The decay step size, The initial step size, As a social learning factor, This represents the maximum number of iterations.
[0127] Step E: If the fitness remains stagnant after triggering the Levi flight perturbation mechanism, trigger the genetic mutation mechanism.
[0128] If the Lévy flight perturbation is triggered, the algorithm remains in a fitness stagnation state in subsequent iterations (e.g., cumulative continuous perturbation). (secondary stagnation), triggering the genetic mutation mechanism, performs Gaussian mutation on some individual parameters in the population. The mutation formula is: ,in, These are the mutated parameters. With a mean of 0 and a variance of Gaussian distributed random noise.
[0129] Step F: Repeat steps B to E until the preset maximum number of iterations is reached or the convergence condition is met, and output the neural network prediction model parameters corresponding to the optimal particle position.
[0130] Step 5: Fusion of prediction results.
[0131] To adapt to diverse computing environments in dam safety monitoring systems (from high-performance cloud servers to resource-constrained edge terminals), this embodiment designs two selectable result fusion modes. The system can choose to execute one based on the current computing resource configuration and accuracy requirements. For example... Figure 6 As shown.
[0132] Mode 1: High-precision fusion mode, which adopts a stacked ensemble learning strategy, uses the prediction results of each component as meta-features, and trains a meta-learner (such as a logistic regression model or support vector machine SVM) to perform nonlinear fusion and output the dam deformation prediction results.
[0133] Mode 2: High-efficiency fusion mode, which adopts a dynamic weighted linear fusion strategy. Based on the fusion weight of the prediction model of each component, the prediction results of each component are linearly weighted and summed to output the dam deformation prediction result.
[0134] Combining the above technical solutions, the following beneficial effects can be achieved:
[0135] By using goal-oriented dynamic weighted fusion, the model input can more accurately reflect the contribution of key factors; through secondary mode decomposition, signal components with clearer physical meaning and higher purity are obtained, laying the foundation for subsequent accurate modeling, thus ensuring the final prediction accuracy from both the source and intermediate processing aspects.
[0136] For neural network models with high-frequency residual components, a conditionally triggered hybrid optimization strategy is adopted. This strategy can intelligently trigger Lévy flight or genetic mutation when the PSO gets stuck in a local optimum, effectively escaping the stagnant region and ensuring the global convergence capability of parameter optimization, thereby improving the prediction stability of the model under complex and nonlinear conditions.
[0137] A selective quadratic decomposition is employed, performing VMD only on aliased modes to avoid unnecessary computational overhead. Adaptive model matching assigns appropriate lightweight models to components with different characteristics. The overall framework is clearly modular, with some modules capable of parallel computation, making it easier to deploy on edge computing devices and meeting real-time engineering requirements.
[0138] It provides dual-mode deployment capability of cloud-edge collaboration, ensuring high accuracy of core monitoring nodes while meeting the needs of rapid response of edge terminals.
[0139] Example 3
[0140] This embodiment proposes a system optimization and deployment strategy for complex engineering environments.
[0141] Deploy high-precision fusion models in the cloud / services, leveraging the ample computing resources of the cloud to pursue the ultimate in prediction accuracy.
[0142] Deploying the efficient fusion mode on edge terminals can reduce computational latency by directly using a dynamic weighted summation method for fusion. The weights can be periodically updated based on the determination coefficients of each component model on recent data.
[0143] Dam monitoring data is often affected by environmental noise, and the data distribution changes slowly over time. Therefore, this embodiment introduces the following mechanism:
[0144] Robust training: During the model training phase, a small amount of synthetic noise or anomalous samples are actively injected into the training data. Simultaneously, robust loss functions such as Huber loss are used when training the neural network to enhance the model's tolerance to data anomalies.
[0145] Online drift detection and update: After deployment, changes in prediction error or input data distribution are monitored in real time. When significant concept drift or data drift is detected, a model update mechanism is triggered. Updates can use incremental learning to fine-tune model parameters; if the drift is severe, local retraining is initiated.
[0146] To increase engineers' trust in the AI model and to adapt to the hardware limitations of edge devices, this embodiment implements the following optimizations:
[0147] Interpretation of prediction results: Provides global feature importance analysis (such as permutation importance or SHAP value) and local prediction interpretation (such as LIME) to enhance engineers' understanding and trust in model decisions.
[0148] Edge deployment optimization: To deploy the model to resource-constrained edge monitoring devices, techniques such as model quantization (e.g., quantizing 32-bit floating-point parameters into 8-bit integers), network pruning, and knowledge distillation can be used to compress and accelerate the prediction model (especially the neural network part), and a dedicated inference framework can be used for efficient deployment.
[0149] Example 4
[0150] To verify the engineering applicability of the method proposed in this invention, this embodiment selects a concrete gravity dam as the research object and conducts a horizontal displacement prediction experiment on the dam crest.
[0151] Daily automated monitoring data of the dam was collected for five years, from 2018 to 2022, including upstream reservoir water level. ), daily average temperature ( ) and timeliness factor ( ).
[0152] Construct an extended feature set containing original items and interaction items. : .
[0153] Set the sliding window length in months. The Spearman correlation coefficient was used to dynamically calculate the relationship between various features and displacements. The correlation between them. Set a change threshold. During the sliding process, only when the correlation coefficient of a certain feature changes... Only when the time condition is met will the weight update be triggered according to the formula described in Example 2; otherwise, the weights from the previous time step will be retained. This yields a weighted fusion sequence reflecting the weights of the real-time operating conditions. .
[0154] For the weighted fused displacement sequence The ICEEMDAN algorithm was used to decompose the intrinsic mode components (IMFs), resulting in eight IMFs. The sample entropy of each IMF was calculated, and an entropy threshold was set. Calculations revealed that... The sample entropy is 0.62. The sample entropy is 0.58, which is greater than the threshold of 0.5; and the bandwidth of both components is greater than their center frequency. Based on the joint criterion, it is determined that... and Modal aliasing exists.
[0155] Only for those identified and Perform a second-order VMD decomposition. Adaptively determine the number of decomposition levels using the envelope entropy minimization criterion, and calculate the optimal number of modes. The un-overlapping IMF and the sub-modes obtained from the second decomposition are reordered and combined, and finally reorganized into: 1 trend component, 2 periodic components and 1 residual component.
[0156] The trend term uses an ARIMA(1,1,1) model. The periodic term uses a SARIMA(1,0,1)(1,1,1)365 model. The residual term is modeled using a neural network (structure 10-8-1), and the BP network is trained using conditionally triggered hybrid optimization (PSO + Lévy flight + GA mutation). ); ; ; .
[0157] By superimposing the predictions from various models for the first 100 days of 2023, the final dam displacement prediction curve is obtained. By comparison, the mean absolute error (MAE) is reduced by approximately 42% compared to the traditional EMD-BP-PSO method.
[0158] Example 5
[0159] A computer-readable storage medium storing a computer program that, when executed by a processor, implements a dam deformation prediction method based on dynamic fusion of multi-source data as described above.
[0160] Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instruction data structures, program modules, or other data. Computer storage media includes RAM, ROM, EPROM, EEPROM, flash memory or other solid-state storage technologies, CD-ROM, DVD or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The aforementioned system memories and mass storage devices can be collectively referred to as memory.
[0161] A computer program product includes a computer program / instructions that, when executed by a processor, implement a dam deformation prediction method based on dynamic fusion of multi-source data as described above.
[0162] Computer program products include computer programs or instruction sets used to perform specific tasks or achieve specific functions. These programs or instructions are designed to be executed by a processor to implement a series of predefined steps or operations. The program product may be stored in various forms of computer storage media, such as memory, hard disks, solid-state drives, optical discs, or other forms of digital storage devices. It may exist in the form of compiled binary code or in the form of scripts or bytecode that can be executed by an interpreter. Through carefully designed algorithms and logical instructions, the program product enables the processor to process data in a specific order and manner, performing various functions such as data analysis, user interaction, and device control.
[0163] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.
[0164] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0165] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above invention, and these changes or modifications still fall within the scope of the present invention.
Claims
1. A method for predicting dam deformation based on dynamic fusion of multi-source data, characterized in that, Includes the following steps: Acquire multi-source monitoring data during dam operation, and construct an extended feature set after preprocessing; Based on the correlation and time-varying characteristics between each feature in the extended feature set and the dam deformation response, a target-oriented feature weight matrix is constructed and dynamically updated. A weighted fusion input sequence is generated based on the updated weight matrix. The weighted fusion input sequence is subjected to initial mode decomposition. The mode aliasing criterion is used to identify the aliased mode components. The identified aliased mode components are subjected to secondary mode decomposition. All the components obtained after decomposition are recombined into three types of components with different statistical properties. Based on the statistical characteristics of the three types of components, corresponding prediction models are matched to perform prediction modeling to obtain prediction results for each type of component. The prediction results of various components are fused together to output the dam deformation prediction results; The specific methods for constructing and dynamically updating the goal-oriented feature weight matrix include: Calculate the correlation coefficients between each feature in the extended feature set and the dam deformation response sequence: if the data satisfies the normal distribution assumption, use the Pearson correlation coefficient; otherwise, use the Spearman rank correlation coefficient. Based on the calculated correlation coefficient, a judgment matrix is constructed using the analytic hierarchy process. The eigenvector corresponding to the largest eigenvalue is calculated, and after satisfying the consistency constraint, normalization is performed to obtain the initial weight vector. A sliding time window mechanism was established to calculate the change in the correlation coefficient between each feature and the dam deformation response sequence within adjacent time windows; The change in the correlation coefficient is compared with a preset change threshold: when the change in the correlation coefficient exceeds the preset change threshold, the weight of the feature is incrementally updated using the learning rate and the change in the correlation coefficient; when the change in the correlation coefficient does not exceed the preset change threshold, the current weight of the feature remains unchanged. The specific steps for performing secondary mode decomposition include: The weighted fusion input sequence is initially decomposed using the ICEEMDAN algorithm to obtain several intrinsic mode components. Calculate the sample entropy and bandwidth of each intrinsic mode component; Establish a joint rule for mode aliasing: when the sample entropy value of a certain intrinsic mode component is greater than a preset entropy threshold and its bandwidth is greater than the center frequency of the intrinsic mode component, the intrinsic mode component is determined to be a mode component with aliasing. The VMD algorithm is used to perform secondary decomposition on the modal components that are determined to have aliasing, and the optimal number of modes is determined by the minimum envelope entropy criterion.
2. The dam deformation prediction method based on dynamic fusion of multi-source data according to claim 1, characterized in that, The specific methods for constructing extended feature sets include: Obtain the reservoir water level sequence, ambient temperature sequence, time factor sequence and their corresponding lag term sequence of the dam, and construct the original feature vector; The original feature vector is subjected to anomaly detection and repair using statistical criteria and Grubbs test; for mild outliers and missing values, cubic spline interpolation is used for repair; for severe outliers, manual confirmation or trend correction is performed. A second-order interactive feature is constructed based on the repaired original feature vector, and the second-order interactive feature is combined with the original feature vector to form an extended feature set.
3. The dam deformation prediction method based on dynamic fusion of multi-source data according to claim 1, characterized in that, The sub-mode components obtained from the second decomposition and the intrinsic mode components that have not aliased are recombined into three types of components with different statistical properties based on the characteristic of frequency from low to high: trend components, periodic components and residual components. For the aforementioned trend components, a regression-based prediction model is matched; For the periodic component, a time series prediction model is matched; For the residual components, a matching neural network prediction model is used.
4. The dam deformation prediction method based on dynamic fusion of multi-source data according to claim 3, characterized in that, A conditionally triggered hybrid optimization strategy based on fitness stagnation states is used to optimize parameters in neural network prediction models, specifically including: Step A: Initialize the population for the particle swarm optimization algorithm and construct the fitness function; Step B: Perform particle swarm optimization iteration and calculate the global optimal fitness value for the current iteration based on the fitness function; Step C: If the change in the global optimal fitness value is lower than a preset threshold in a preset number of consecutive iterations, the optimization process is determined to have entered a fitness stagnation state. Step D: When the fitness stagnation state is determined, the Levy flight perturbation mechanism is triggered to update the position of the current global best particle; Step E: If the fitness remains stagnant after triggering the Levy flight perturbation mechanism, trigger the genetic mutation mechanism; Step F: Repeat steps B to E until the preset maximum number of iterations is reached or the convergence condition is met, and output the neural network prediction model parameters corresponding to the optimal particle position.
5. The dam deformation prediction method based on dynamic fusion of multi-source data according to claim 4, characterized in that, The neural network prediction model is as follows: ,in, The dimension of the input feature. Let i be the i-th component of the input feature vector. The number of neurons in the hidden layer. The weights are used to connect the i-th feature of the input layer to the j-th neuron of the first hidden layer. This is the bias term for the j-th neuron in the first hidden layer. The activation function for the hidden layer is... The weights connecting the j-th neuron in the second hidden layer to the output neuron. This is the bias term for the output neuron. The activation function for the output layer; The fitness function includes a mean squared error term and an L2 regularization term, specifically: ,in, The number of training samples, For the true value, For predicted values, For model parameters, The regularization coefficient is used. Criteria for determining fitness stagnation: Record the first... The global optimal fitness value in the next iteration Calculate the change between adjacent iterations. When continuous The next iteration satisfies When the time comes, it is determined that the system has entered a standstill state, where, The preset threshold; The specific mechanism for triggering the Lévy flight perturbation is as follows: a global perturbation is performed using an improved Lévy distribution random step size to update the particle position. The update formula is: ,in, Let be the current position of the particle at time t. This represents the current position of the globally optimal particle. Let be a random vector that follows a Lévy distribution. This is the inertia weighting coefficient. For the number of iterations The decay step size, The initial step size, As a social learning factor, This represents the maximum number of iterations. The specific mechanism for triggering genetic mutation is as follows: performing a Gaussian mutation operation on some individual parameters in the population, with the mutation formula being: ,in, These are the mutated parameters. With a mean of 0 and a variance of Gaussian distributed random noise.
6. The dam deformation prediction method based on dynamic fusion of multi-source data according to claim 1, characterized in that, Based on the computing resources and accuracy requirements of actual application scenarios, the fusion modes of prediction results include: Mode 1: High-precision fusion mode, which adopts a stacked ensemble learning strategy, uses the prediction results of each component as meta-features, and trains a meta-learner to perform nonlinear fusion, outputting the dam deformation prediction results. Mode 2: High-efficiency fusion mode, which adopts a dynamic weighted linear fusion strategy. Based on the fusion weight of the prediction model of each component, the prediction results of each component are linearly weighted and summed to output the dam deformation prediction result.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a dam deformation prediction method based on dynamic fusion of multi-source data as described in any one of claims 1-6.
8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements a dam deformation prediction method based on dynamic fusion of multi-source data as described in any one of claims 1-6.