GNSS displacement prediction method and system based on multi-source feature fusion and attention mechanism

Abstract

The application discloses a GNSS displacement prediction method and system based on multi-source feature fusion and attention mechanism, and aims to solve the defects of feature dimension redundancy interference, phase lag prediction, poor fitting effect of extreme deformation and the like in the prior art.The method comprises the following steps: 1) collecting GNSS observation data and completing feature extraction and classification; 2) reducing the dimension of redundant auxiliary features by a PCA algorithm, and splicing the features to construct a composite feature vector; 3) performing standardization processing on the vector, and constructing a time sequence sample in the aid of a sliding window; 4) building a prediction model with a fusion core reasoning layer, a self-attention mechanism layer and a fully connected output layer; 5) designing a WMSE loss function to complete model training; and 6) checking the prediction effect of the model by using a test set.The application is suitable for complex high-frequency dynamic displacement monitoring and prediction scenes, has the characteristics of accurate extreme value capture, fast mutation response and excellent noise resistance robustness, and can effectively improve the amplitude reduction problem in extreme deformation prediction.

Description

Technical Field

[0001] This invention belongs to the field of spatial information processing and geological disaster early warning technology, and further relates to GNSS temporal deformation data deep learning prediction technology. Specifically, it is a GNSS displacement prediction method and system based on multi-source feature fusion and attention mechanism, which can be used for deformation monitoring and early warning of engineering structures such as slopes, landslides, dams, and bridges. Background Technology

[0002] With the comprehensive networking of Global Navigation Satellite Systems (GNSS) and the increasing maturity of high-precision positioning algorithms, leveraging spatial positioning technology to empower safety monitoring and early warning of geological disasters (such as landslides) and major infrastructure (such as dams and bridges) has become an industry consensus. In the entire disaster prevention and emergency response system, spatial displacement and location status are the most intuitive core indicators for judging the evolution trend of hazards. Therefore, advanced and accurate time-series prediction of structural deformation and displacement trends is a core element in building a disaster early warning defense line.

[0003] In recent years, deep learning frameworks, represented by recurrent neural networks (RNNs) and their variant, long short-term memory networks (LSTMs), have been widely applied to the dynamic evolution modeling of displacement monitoring sequences due to their advantages in processing time series data. However, when faced with the complex and variable nonstationarity and nonlinear characteristics of high-frequency GNSS monitoring data, existing displacement prediction technologies and systems still have the following significant defects and shortcomings:

[0004] In processing multi-source monitoring data, existing technologies are prone to the "curse of dimensionality" and are susceptible to multipath noise interference. Existing high-precision positioning solutions typically output heterogeneous data containing dozens of dimensions, including three-dimensional coordinates, velocity components, and positioning quality indicators. If redundant features with strong spatial collinearity (such as ECEF coordinates, geodetic coordinates, etc.) are fully input into the deep network, it will not only greatly increase the computational complexity of the model, leading to the "curse of dimensionality," but also introduce severe environmental multipath noise into the prediction model, interfering with the network's judgment of the true trend of subtle deformations.

[0005] In terms of temporal feature extraction, existing recurrent neural networks suffer from a lack of global context and "prediction lag." Traditional unidirectional LSTM networks can only utilize data from historical moments, ignoring the complete contextual relationships inherent in the evolutionary trend of temporal signals. More seriously, in actual RTK (Real-Time Kinematic) monitoring sequences, stations often experience high-frequency, sudden shifts due to sudden external environmental disturbances. Traditional LSTMs, relying on the smooth propagation mechanism of "memory gates," cannot quickly focus on key information when faced with such abrupt extreme signals, resulting in significant phase lag in the prediction curve.

[0006] In terms of model optimization strategies, conventional loss functions are prone to causing "amplitude attenuation" in prediction results. Existing time series regression prediction models generally use mean squared error (MSE) as the loss function. Due to the gradient smoothing effect of MSE, the model tends to fit the average evolution level of the sequence during the optimization process. When processing GNSS high-frequency monitoring data, this global smoothing tendency causes the model to predict amplitudes that are often lower than the actual observed values ​​when faced with real displacement abrupt changes (peaks or troughs) with instantaneous impact characteristics, resulting in severe "amplitude collapse." This lack of sensitivity to extreme deformation can easily lead to underreporting of geological disasters in actual engineering deployments, posing serious safety hazards.

[0007] Therefore, there is an urgent need in this field for a GNSS high-precision displacement prediction method that can effectively filter multi-source high-frequency noise, keenly capture abrupt signals, and accurately restore extreme deformation amplitudes.

[0008] Patent application CN202610234278.8 discloses a method and system for adaptive adjustment of foundation pit support strength based on real-time monitoring data. This method involves real-time acquisition of foundation pit monitoring data, standardized preprocessing, and then inputting the data into a pre-trained unidirectional LSTM prediction model to output a sequence of predicted displacement values. Finally, a multi-objective optimization model is constructed based on these predicted values ​​to output the prestress adjustment amount. The proposed method for adjusting the support strength of the foundation pit has three main technical limitations: First, the basic unidirectional LSTM model it uses lacks bidirectional global context association and the ability to focus on abrupt feature weights, which makes it prone to prediction phase lag when processing time-series data with drastic changes; second, the system does not optimize the underlying loss function for extreme deformations, and the gradient smoothing effect of the conventional mean square error (MSE) will cause severe prediction amplitude reduction, which is very likely to lead to missed safety warnings in high-risk extreme displacement scenarios; finally, the method lacks deep physical space dimensionality reduction and decoupling processing for multi-source high-dimensional heterogeneous monitoring data. Directly inputting redundant features not only easily causes the "curse of dimensionality" but also introduces environmental multipath effect noise, which greatly reduces the model's prediction robustness in complex and harsh environments. Patent application CN202610131090.0 proposes a bridge early warning method based on Physical Information Neural Network (PINN). Its core lies in fusing environmental and load data, and introducing residuals from elasticity and dynamic equations, along with boundary conditions, to construct a total loss function. This outputs a dynamic adaptive early warning interval under physical constraints to eliminate environmental interference. However, this method has significant limitations when processing complex, high-frequency monitoring data: First, it only uses basic normalization and lacks deep spatial dimensionality reduction and decoupling for multi-source, high-dimensional redundant data, easily introducing environmental noise and causing the "curse of dimensionality." Second, the network lacks a bidirectional temporal feature extraction and weight focusing mechanism for abrupt key steps, easily resulting in prediction phase lag when facing instantaneous vehicle impacts. Third, its loss function does not introduce deviation perception weighting for extreme deformations; the smoothness of the physical equations themselves severely exacerbates the "amplitude reduction" and missed detection risk during sudden high-risk responses. Finally, the multi-source physical constraint loss is prone to severe "gradient conflicts" during joint training, making model convergence extremely difficult and the parameter tuning cost for deployment extremely high. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by proposing a GNSS displacement prediction method and system based on multi-source feature fusion and an attention mechanism. This addresses problems in processing high-frequency dynamic GNSS monitoring data, such as dimensional redundancy interference, prediction phase lag due to abrupt signal changes, and amplitude reduction in extreme deformations caused by the smoothing effect of conventional loss functions. Principal component analysis (PCA) is used to reduce the dimensionality of redundant features to filter environmental noise. A bidirectional long short-term memory network (BiLSTM) and a self-attention mechanism are combined to dynamically focus weights on sudden critical time steps. Furthermore, a weighted mean square error (WMSE) loss function is designed to overcome the extreme value collapse problem from the algorithm's underlying layer.

[0010] To achieve the above objectives, the technical solution of the present invention includes the following:

[0011] A GNSS displacement prediction method based on multi-source feature fusion and attention mechanism includes the following steps:

[0012] (1) Obtain the original data of GNSS high-precision positioning observation, extract and classify it into the displacement features of the target to be predicted, the core positioning quality index features and the redundant auxiliary features;

[0013] (2) Principal component analysis (PCA) is used to perform deep physical space dimensionality reduction and decoupling of redundant auxiliary features, and the first principal component is extracted. Then, it is horizontally spliced ​​with the displacement features of the target to be predicted and the core positioning quality index features to construct a composite feature vector.

[0014] (3) Perform data standardization preprocessing on the composite feature vector to eliminate the difference in dimensions. Then, use the sliding window algorithm to construct three-dimensional tensor samples with historical time context and divide them into model training set, validation set and test set according to time order.

[0015] (4) Construct a deep displacement prediction network model PCA-Attention-BiLSTM, which includes a core inference layer, a self-attention mechanism layer and a fully connected output layer. The core inference layer uses a two-layer deep stacked bidirectional long short-term memory network BiLSTM. The self-attention mechanism layer receives the output features of BiLSTM at each time step and generates a unified context representation. The fully connected output layer consists of a random deactivation layer Dropout and a fully connected layer Dense, which are used to perform high-dimensional mapping of the context feature vector and further restore it to obtain the displacement prediction value.

[0016] (5) Construct a weighted mean square error (WMSE) loss function based on bidirectional deviation perception. Using the training set and validation set, the adaptive moment estimation Adam optimization algorithm is used to iteratively optimize the deep displacement prediction network model PCA-Attention-BiLSTM to obtain the trained displacement prediction model.

[0017] (6) Test the trained displacement prediction model using the test set, and use root mean square error, mean absolute error, coefficient of determination and maximum absolute error as evaluation indicators to evaluate the prediction performance of the model.

[0018] Meanwhile, the present invention also proposes a GNSS displacement prediction system for implementing the above method, including: a feature extraction and dimensionality reduction module, a data construction module, a model construction and training module, and a displacement prediction and evaluation module;

[0019] The feature extraction and dimensionality reduction module is used to receive raw GNSS high-precision positioning observation data, classify it, use principal component analysis (PCA) to extract core principal components to eliminate dimensional redundancy, obtain composite feature vectors, and output them to the data construction module.

[0020] The data construction module is used to horizontally concatenate and standardize the extracted composite feature vectors, and construct a three-dimensional tensor sample set with historical time series constraints using a sliding window algorithm; after dividing the sample set into a training set and a test set, it is transmitted to the model construction and training module.

[0021] The model building and training module is used to build a prediction network that includes a deep bidirectional long short-term memory network BiLSTM and a self-attention mechanism layer MyAttention, extract bidirectional temporal features of training set samples, and dynamically focus on extreme time steps of the self-attention mechanism; and use the weighted mean square error (WMSE) loss function based on bidirectional deviation perception for adaptive optimization training, and then transmit the trained prediction model to the displacement prediction and evaluation module.

[0022] The displacement prediction and evaluation module is used to input real-time high-frequency data into the trained prediction model for online inference to predict deformation, restore the network output results to the physical displacement prediction value after inverse normalization, and evaluate the model's early warning robustness using statistical indicators including RMSE, MAE, R^2, and MaxE.

[0023] Compared with the prior art, the present invention has the following advantages:

[0024] First, because this invention employs a multi-source feature decoupling strategy based on principal component analysis (PCA) combined with prior environmental constraints, it extracts the first principal component (PC1) containing strongly collinear redundant coordinates and velocity features, and horizontally concatenates it with target displacement features and positioning solution quality indicators at the data and processing layer. This effectively reduces and compresses high-dimensional auxiliary data, while introducing real-time observation environment and solution reliability constraints to the prediction network. This greatly reduces the computational complexity of the model to avoid the curse of dimensionality and enables the model to accurately distinguish between real physical deformation and pseudo-fluctuation errors caused by signal degradation. This significantly enhances the robustness and generalization ability of high-frequency monitoring models in complex environments.

[0025] Secondly, because this invention uses a stacked two-layer bidirectional long short-term memory network (BiLSTM) architecture as the core inference layer, its forward hidden layer extracts the displacement growth trend in chronological order, and the backward hidden layer learns deformation backtracking features from the end of the sequence backward. This enables the model to capture the global contextual dependence of deformation across a longer time step from a complete bidirectional time window perspective, thereby greatly enhancing the model's fitting accuracy for complex nonlinear displacement evolution patterns under multi-source composite feature inputs and significantly improving the ability to identify and warn of small deformation trajectories and complex dynamic responses.

[0026] Third, because this invention adopts a bidirectional deviation-aware WMSE weighted loss function oriented towards extreme value perception, by setting a dynamic threshold, a high penalty weight coefficient is applied to the transient extreme value impact point that deviates from the reference zero point during backpropagation. This forces the center of gravity of the neural network gradient optimization to tilt towards the displacement peak and trough region, thereby effectively overcoming the problem of amplitude reduction at abrupt change caused by the global smoothing tendency of conventional loss functions from the bottom layer of the algorithm. This significantly reduces the maximum absolute error (MaxE) at extreme value jumps and improves the response sensitivity of the early warning system to extreme transient deformations. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall architecture of the depth displacement prediction network model PCA-Attention-BiLSTM constructed in this invention.

[0028] Figure 2 This is a comparison chart of the predicted displacement curve of the eastward position POSE dimension during the model training phase and the actual sequence in this invention.

[0029] Figure 3 This is a comparison chart of the predicted displacement curve of the eastward position POSE dimension during the model testing phase of this invention and the actual sequence.

[0030] Figure 4 This is a diagram showing the full sequence inference results of novel, unknown data in the POSU dimension of elevation location in this invention. Detailed Implementation

[0031] The present invention will now be further described with reference to the accompanying drawings.

[0032] Example 1: Refer to Appendix Figure 1 This invention proposes a GNSS displacement prediction method based on multi-source feature fusion and attention mechanism, comprising the following steps:

[0033] Step 1) Obtain the raw GNSS high-precision positioning observation data, extract and classify it into target displacement features, core positioning quality index features, and redundant auxiliary features; in this embodiment, the F-dimensional feature space constructed from the raw GNSS high-precision positioning observation data completes the hierarchical extraction and category classification, and the implementation steps are as follows:

[0034] (1.1) Extract the displacement components of the carrier in the east, north and sky directions in the coordinate system of the ENU station in the northeast and divide them into 3D displacement features of the target to be predicted;

[0035] (1.2) Extract GNSS positioning quality information features, including multiple positioning quality indicators such as positioning solution confidence, number of visible satellites and standard deviation of three-axis displacement, and combine them into G-dimensional core positioning quality indicator features to introduce prior environmental quality constraints;

[0036] (1.3) All remaining features, including geodetic coordinates, geocentric coordinate system coordinates, and multi-axis motion velocity components, are classified into F-3-G dimension redundant auxiliary features.

[0037] Step 2) Use Principal Component Analysis (PCA) to perform deep physical space dimensionality reduction and decoupling on redundant auxiliary features, and extract the first principal component. Then, horizontally concatenate it with the displacement features of the target to be predicted and the core positioning quality index features to construct a composite feature vector.

[0038] In this embodiment, Principal Component Analysis (PCA) is used to perform deep physical space dimensionality reduction and decoupling on redundant auxiliary features, and the first principal component is extracted. The specific execution steps are as follows:

[0039] (2.1) Decentralize the redundant auxiliary feature matrix;

[0040] (2.2) Solve for the covariance matrix of the centered matrix and perform eigenvalue decomposition on it. Determine the eigenvector corresponding to the largest eigenvalue as the first principal component. The linear combination expression of the first principal component is as follows:

[0041] ,

[0042] in, Original redundant auxiliary features For the corresponding load factor, =F-3-G represents the total number of dimensions of redundant auxiliary features.

[0043] Step 3) Perform data standardization preprocessing on the composite feature vectors to eliminate dimensional differences. Then, use the sliding window algorithm to construct three-dimensional tensor samples with historical temporal context, and divide them into model training set, validation set, and test set according to time order. The implementation steps are as follows:

[0044] (3.1) The min-max normalizer MinMaxScaler is used to linearly scale the composite feature vector after fusion to the interval [0.1, 0.9].

[0045] (3.2) Set the length of the sliding window to The sliding step size is ,and , ; The continuous time-series data is segmented to construct a data structure with shape (batch_size, ... A three-dimensional tensor of (3+G+1);

[0046] (3.3) Divide the dataset according to the time sequence: In this embodiment, it is preferable to select the first 80% of the continuous samples on the time axis as the initial training set, and extract 10% of the data from the training set as the cross-validation set, and divide the remaining 20% ​​of the continuous samples on the time axis into the test set.

[0047] Step 4) Construct a deep displacement prediction network model PCA-Attention-BiLSTM, which includes a core inference layer, a self-attention mechanism layer, and a fully connected output layer. The core inference layer uses a two-layer deep stacked bidirectional long short-term memory network (BiLSTM). The self-attention mechanism layer receives the output features from each time step of the BiLSTM to generate a unified contextual representation. The fully connected output layer consists of a random deactivation layer (Dropout) and a fully connected layer (Dense), used to perform high-dimensional mapping of the contextual feature vector and further restore it to obtain the displacement prediction value. This embodiment specifically constructs the above prediction network model according to the following steps.

[0048] (4.1) A two-layer deep-stacked bidirectional long short-term memory network (BiLSTM) is used to construct the core inference layer, which is used to deeply mine the bidirectional temporal context dependencies of the shift sequence within the sliding window from both forward and reverse directions; in the two-layer deep-stacked bidirectional long short-term memory network (BiLSTM), the first layer contains A hidden unit, the second layer contains One hidden unit; and , ;

[0049] (4.2) Following the core inference layer is the attention mechanism layer MyAttention, which scores the output features of BiLSTM using a trainable weight matrix and obtains the attention weights using the Softmax function. These weights are then aggregated with the hidden states to generate a unified contextual representation. The implementation steps include:

[0050] (4.2.1) The output features of BiLSTM at each time step are scored using a trainable weight matrix, as shown in the following formula:

[0051] ,

[0052] in, The hidden state features of the current time step output by BiLSTM. For a trainable weight matrix, For bias vectors, The original attention score corresponding to the current time step;

[0053] (4.2.2) Using the Softmax function, the scores are mapped to attention weights in the form of a probability distribution according to the following formula:

[0054] ,

[0055] in, The attention weights at the current time step after normalization. The total number of time steps. The original attention score for the i-th time step, i=1,2,...,T;

[0056] (4.2.3) Multiply the latent state features and attention weights element-wise, and then aggregate them along the time dimension to obtain the aggregated context feature vector. :

[0057] .

[0058] (4.3) The fully connected output layer is constructed by the random deactivation layer Dropout and the fully connected layer Dense. The fully connected layer Dense is configured with a linear activation function to perform high-dimensional mapping of the context feature vector and restore it to the displacement prediction value with actual physical meaning through the inverse normalization operation.

[0059] Step 5) Construct a weighted mean square error (WMSE) loss function based on bidirectional deviation perception. Using the training set and validation set, iteratively optimize the deep displacement prediction network model PCA-Attention-BiLSTM using the Adam optimization algorithm of adaptive moment estimation to obtain the trained displacement prediction model.

[0060] In this embodiment, the weighted mean square error (WMSE) loss function based on bidirectional deviation sensing is calculated as follows:

[0061] ,

[0062] in, The number of samples within a single batch. These are actual displacement observations. These are the predicted displacement values ​​for the model. This is a dynamic weight adjustment factor;

[0063] The The triggering logic depends on the standardized true displacement. Absolute distance from the center point of the normalized interval by 0.5:

[0064] when hour, ;

[0065] when hour, ;

[0066] The above-mentioned adaptive moment estimation Adam optimization algorithm is used to iteratively optimize the model. The training parameter configuration includes:

[0067] Set the batch size for a single-input network to [value]. With the objective of minimizing the weighted mean square error (WMSE) loss function, the global maximum number of iterations is set to... The algorithm iterates through rounds and monitors the loss using a validation set at the end of each round to prevent overfitting.

[0068] Step 6) Test the trained displacement prediction model using the test set. Use root mean square error, mean absolute error, coefficient of determination, and maximum absolute error as evaluation metrics to assess the model's predictive performance.

[0069] like Figure 1As shown, the model constructed in this invention adopts a hierarchical topology logic of parallel feature preprocessing, deep temporal extraction, and weight focusing in its architecture design. First, it processes multi-source heterogeneous data through parallel branches: target displacement features and positioning quality indicators directly enter the fusion layer, while environmental redundancy features are input into the PCA dimensionality reduction module. Principal component analysis is used to extract the first principal component with the highest contribution, achieving data dimensionality reduction from the input end and effectively filtering environmental noise interference. Subsequently, the three features are horizontally concatenated in the fusion layer to construct a 9-dimensional composite input vector containing physical mechanism constraints. In the core inference layer, the model stacks two layers of Bi... LSTM delves into the bidirectional temporal dependencies of displacement sequences within a sliding window. To address the issue of lag fitting of deformation signals at abrupt changes, the model is further integrated with a self-attention layer, which achieves precise focusing on key time steps through a self-learning weight matrix. Finally, after Dropout regularization and nonlinear mapping of the fully connected layer, the model outputs the predicted physical displacement value restored by inverse normalization. During this process, a WMSE weighted loss function is introduced to construct an extreme value penalty mechanism. In the backpropagation optimization stage of the model, a high penalty weight is applied to transient impact points that deviate from the baseline, effectively solving the amplitude collapse problem of the prediction results.

[0070] Example 2: This example proposes a GNSS displacement prediction system for implementing the method in Example 1, including a feature extraction and dimensionality reduction module, a data construction module, a model construction and training module, and a displacement prediction and evaluation module. The functions and transmission relationships of each module are as follows:

[0071] The feature extraction and dimensionality reduction module is used to receive the raw data of GNSS high-precision positioning observation, separate the target displacement features, positioning quality indicators and environmental redundancy features from them, and then use principal component analysis (PCA) to reduce the dimensionality of the redundant features to eliminate dimensionality redundancy. After extracting its core principal components, the three are spliced ​​and fused into a composite feature vector and output to the data construction module.

[0072] The data construction module is used to horizontally concatenate and standardize the extracted composite feature vectors, and construct a three-dimensional tensor sample set with historical time series constraints using the sliding window algorithm. That is, based on the sliding window algorithm, the continuous time series data is divided into a three-dimensional tensor of fixed length. Then, the sample set is divided into a training set and a test set and transmitted to the model construction and training module.

[0073] The model building and training module is used to build a prediction network that includes a deep bidirectional long short-term memory network BiLSTM and a self-attention mechanism layer MyAttention. It extracts bidirectional temporal features of the training set samples and dynamically focuses on the extreme time steps of the self-attention mechanism. It then uses the weighted mean square error (WMSE) loss function based on bidirectional deviation perception for adaptive optimization training and transmits the trained prediction model to the displacement prediction and evaluation module.

[0074] The displacement prediction and evaluation module is used to input the time-series test samples to be predicted or unknown real-time high-frequency data into the trained prediction model for online inference to predict deformation. After the network output results are denormalized, they are restored to the physical displacement prediction value. The module also uses statistical indicators including RMSE, MAE, R^2, and MaxE to evaluate the model's early warning robustness.

[0075] This embodiment also provides a displacement prediction device, including: a memory for storing a computer program implementing the displacement prediction method, and a processor for implementing the displacement prediction when executing the computer program. Additionally, a computer-readable storage medium is provided, storing a computer program that, when executed by the processor, can implement the above-described displacement prediction method.

[0076] Example 3: The overall implementation steps of the GNSS displacement prediction method based on multi-source feature fusion and attention mechanism proposed in this example are the same as those in Example 1. The following is a detailed description in conjunction with the appendix. Figure 1-4 Specific parameter selections are given, and the implementation process of the method of the present invention is further described in detail with examples.

[0077] Step 1: Acquire raw GNSS high-precision positioning observation data. In this embodiment, the data comes from real continuous high-frequency observation stations in the monitoring area. A base station and a monitoring station are set up in a semi-open area. The solution mode is selected as a combination of GPS+BDS+GALILEO+GLONASS mode; the data update rate is set to 1Hz. Extract the raw feature space and divide it into the displacement features of the target to be predicted, the core positioning quality index features, and redundant auxiliary features:

[0078] Step 1a: Obtain the raw data of continuous observation results generated by the GNSS high-precision positioning algorithm and store it in CSV format. It covers observation fields of multiple dimensions such as time identifier (week, weekSec, timesstr, etc.), satellite status, positioning solution quality (age, ratio, nsv, etc.), spatial coordinates (lat, lon, ecefx, etc.), displacement observation values ​​(pose, posn, posu) and velocity information, and constructs a raw feature space containing a total of 35 dimensions.

[0079] Step 1b: Based on the physical deformation characteristics of GNSS and the principle of eliminating collinearity redundancy in multi-source data, the original data contains ECEF coordinates, geodetic coordinates, and triaxial displacements (pose / posn / posu). These fields are highly correlated in spatial geometry. Retaining redundant coordinate fields will cause collinearity problems, interfering with the neural network's judgment of the true deformation trend. Therefore, this patent only extracts the target features pose, posn, and posu that directly reflect deformation to reduce the sparsity of the input feature space.

[0080] Step 1c: To introduce prior environmental quality constraints to identify pseudo-fluctuation noise caused by multipath effects or satellite loss of lock, core fields reflecting the reliability of real-time observation environment and measurement solution are selected, including solution confidence (ratio), number of visible satellites (nsv), and three-axis positioning standard deviation (posestd, posnstd, posustd), which are divided into core positioning quality index features (5 dimensions).

[0081] Step 1d: The remaining auxiliary fields in the original feature space, including geodetic coordinates (lat, lon, alt), geocentric and geofixed coordinates (ecefx, ecefy, ecefz), and multi-axis velocity components (velx, vely, velz, etc.), are uniformly classified into redundant auxiliary features (27 dimensions).

[0082] Step 2: Perform deep physical space dimensionality reduction and decoupling on the redundant auxiliary features identified in Step 1. Use Principal Component Analysis (PCA) to extract the first principal component, and concatenate it with the displacement features of the target to be predicted and the core positioning quality index features to construct a composite feature vector.

[0083] Step 2a: The redundant auxiliary feature matrix containing 27 dimensions, including geodetic coordinates and multi-axis velocities, is centered (i.e., the mean of each dimension is subtracted). Then, the covariance matrix of the centered matrix is ​​calculated, and its eigenvalues ​​are decomposed to obtain the eigenvector with the largest eigenvalue as the first principal component (PC1). Its linear combination is expressed as follows: ,in This represents the original redundant features. The corresponding load coefficients are shown. This process successfully compressed the 27-dimensional strongly collinear auxiliary features into a 1-dimensional comprehensive index, effectively filtering out high-frequency fluctuation noise introduced by environmental multipath effects and achieving a significant reduction in the load on the model input.

[0084] Step 2b: The 1D PC1 feature extracted in Step 2.1 is horizontally concatenated with the 3D target displacement features (pose, posn, posu) and the 5D core positioning quality index features (ratio, nsv, posestd, posnstd, posustd) extracted in Step 1. This step integrates multi-source heterogeneous data to construct a 9D composite input vector that includes both the core deformation trajectory and the physical mechanism constraints of the positioning environment reliability.

[0085] Step 3: Perform data standardization preprocessing on the composite feature vectors obtained in Step 2 to eliminate dimensional differences. Then, use the sliding window algorithm to construct three-dimensional tensor samples with historical time context and divide them into model training set, validation set and test set according to time order.

[0086] Step 3a: Use MinMaxScaler to linearly scale the 9-dimensional composite input vector to a uniform size. Interval. Its mapping formula is: This scaling interval not only eliminates the interference of numerical differences between heterogeneous features (such as the millimeter-scale displacement versus the single-digit-scale satellite number) and prevents gradient vanishing caused by the activation function falling into the saturation region, but more importantly, it provides a basis for the subsequent WMSE loss function to "scale down" to achieve the desired effect. It provides a reference mapping space for the "relative physical zero point".

[0087] Step 3b: Apply the sliding window algorithm to the standardized feature sequence to construct time series samples. Set a fixed sliding window length of 10 (i.e., `look_back=10`) and a sliding step size of 1. The algorithm extracts 10 consecutive historical time-series features along the time axis as input each time, and uses the target displacement component of the next time-series moment as the prediction label, thereby segmenting the two-dimensional time series and constructing a time series with the shape of... The three-dimensional tensor is used as a supervised learning sample.

[0088] Step 3c: Given the strong autocorrelation of high-frequency GNSS time-series monitoring data, to avoid "data leakage" affecting the objectivity of model evaluation, the 3D tensor samples are strictly segmented according to chronological order. The first 80% of the continuous samples on the timeline are divided into the training set (of which 10% is randomly separated into a cross-validation set for loss monitoring and hyperparameter validation at the end of each epoch). The remaining 20% ​​of the continuous samples on the second epoch are used as a completely independent test set to test the model's generalization ability to future unknown data.

[0089] Step 4: Establish the depth displacement prediction network model PCA-Attention-BiLSTM.

[0090] The constructed PCA-Attention-BiLSTM model adopts a hierarchical topology logic of "parallel feature preprocessing - deep temporal extraction - weight focusing", specifically including the following network layers:

[0091] Step 4a: Deep bidirectional temporal feature extraction: Extract the shape constructed in step 3 as... The core inference layer takes 3D tensor samples as input. This layer employs a two-layer deep stacked bidirectional long short-term memory (BiLSTM) network. The first layer contains 50 hidden units, responsible for the initial temporal-dimensional spatial mapping of the 9-dimensional composite features; the second layer expands to 100 hidden units to further mine complex nonlinear displacement patterns. BiLSTM extracts displacement growth trends in temporal order and learns deformation backtracking features from the end of the sequence forward through two independent forward and backward hidden layers. These bidirectional hidden states are then concatenated to deeply mine the bidirectional temporal global contextual dependencies of the displacement sequence within the sliding window.

[0092] Step 4b: Weight Focusing at Key Time Steps: A custom self-attention mechanism layer (MyAttention) is connected after the BiLSTM network layer. For each time step feature output by the BiLSTM... First, by introducing a trainable weight matrix With bias vector Scoring is calculated using a nonlinear activation function. Subsequently, the Softmax function is used to map the unbounded scores into attention weights in the form of a strict probability distribution. Finally, the hidden state With the corresponding weights Element-wise multiplication and aggregation along the time dimension outputs a context feature vector. This mechanism enables the network to autonomously amplify the weight fraction of the "critical time step" where high-frequency abrupt signals occur, breaking the smoothing hysteresis effect of traditional RNNs and achieving precise focusing on sudden extreme value features.

[0093] Step 4c: Nonlinear mapping and physical quantity restoration: Aggregating the context feature vector after the attention mechanism layer The input is regularized to a Dropout layer with a dropout rate of 0.2 to break local collinear dependencies between features and prevent overfitting. Then, a fully connected (Dense) layer maps high-dimensional features to a single predicted value. To ensure physical mapping constraints, a linear activation function (LinearActivation) is explicitly specified in the output layer. This ensures that the model can accurately output negative displacements representing the opposite physical motion direction without being forcibly truncated; finally, the network output values ​​are denormalized to restore them to displacement prediction values ​​(unit: meters) with actual physical meaning.

[0094] Step 5: Construct a weighted loss function (WMSE) based on "bidirectional deviation perception", and use the adaptive moment estimation (Adam) optimization algorithm for iterative optimization to obtain the trained PCA-Attention-BiLSTM displacement prediction model.

[0095] Step 5a: Constructing the Weighted Mean Squared Error (WMSE) Loss Function: To address the "amplitude reduction" problem in conventional mean squared error (MSE) predictions due to gradient smoothing effects in extreme deformation predictions, a WMSE loss function is constructed. Its core logic is: In Within the normalized mapping space, with the center point This serves as the physical relative zero point (i.e., the stable background region) of the displacement sequence. The actual displacement is calculated. Relative to the reference point The degree of deviation is used to dynamically allocate error penalty weights. When the actual displacement deviates When the absolute distance is greater than the preset fluctuation threshold of 0.2 (i.e., the original normalized data falls into the range of 0.2), or (High-risk extreme value range), indicating that the station has experienced significant unidirectional subsidence or uplift, at which point a higher penalty coefficient is applied to the corresponding sample point; if the deviation distance is less than or equal to the preset fluctuation threshold, the weight... This mechanism forces the network to shift the focus of gradient optimization towards the peak and trough regions during backpropagation, as shown in the following formula:

[0096]

[0097] In the formula, This refers to the number of samples within a single iteration batch. and These are the actual observed values ​​and the model predicted values, respectively.

[0098] Step 5b: Iterative Model Training: The model is iteratively trained using the training and validation sets defined in Step 3. Considering both hardware memory overhead and the stability of gradient updates for long-sequence data, the batch size is set to 128. The Adam algorithm is selected, which dynamically adjusts the learning rate for different network weights under multi-source heterogeneous inputs by calculating the first and second moments of the gradient, quickly escaping local optima. With the goal of minimizing the WMSE loss function, global iterative parameter updates are performed for 200 epochs. At the end of each epoch, the loss is monitored using the cross-validation set until the loss function converges.

[0099] Step 6: Use the test set obtained in Step 3 to test the model trained in Step 5, and use multidimensional statistical indicators to evaluate the model's generalization performance and mutation response ability.

[0100] Using a 20% independent test set (not involved in model training) as input to a PCA-Attention-BiLSTM model with fixed parameters, the model outputs predicted test sequences. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (CDO) are then used as the statistical measures. A comprehensive evaluation is conducted using four indicators: maximum absolute error (MaxE), maximum absolute error (MaxE), and maximum absolute error (MaxE). The specific calculation formulas and physical meanings are as follows:

[0101] Root mean square error (RMSE) is a sample standard deviation used to reflect the deviation between predicted and true values, and it is highly sensitive to large errors.

[0102] ,

[0103] Mean absolute error (MAE) reflects the actual overall level of error in predicted values ​​and is unaffected by the direction of the error.

[0104] ,

[0105] Coefficient of determination ( This is used to measure the goodness of fit between the predicted curve and the actual observed trajectory; the closer the value is to 1, the better the fit.

[0106] ,

[0107] Maximum absolute error (MaxE) is specifically used to evaluate the extreme value deviation of a model at the jump points (peaks / troughs) in the displacement sequence. It is a core indicator for quantitatively verifying whether the model has successfully overcome "amplitude reduction".

[0108] ,

[0109] in, The total number of samples in the test set. These are the actual displacement observations from GNSS. To predict the output value for the model, This is the sample mean of the actual observed values.

[0110] The effects of the present invention will be further explained below with reference to experiments.

[0111] 1. Experimental conditions:

[0112] The simulation experiments of this invention were conducted in a hardware environment with high-performance computing capabilities. Specifically, the CPU used was a 12th generation Intel Core i9-12900K processor to ensure the efficiency of concurrent preprocessing and sliding window sample segmentation of large-scale heterogeneous monitoring data; the GPU was configured with dual NVIDIA GeForce RTX 3080Ti high-performance graphics accelerator cards to support large-scale deep learning tensor operations and rapid backpropagation of gradients; and 64GB of system memory (RAM) was provided to support high-speed residence and training batch scheduling of high-frequency long-time-series GNSS data streams.

[0113] In terms of software environment, this experiment was deployed on a stable version of the Ubuntu 20.04.6 LTS operating system to provide robust low-level system calls and multi-GPU parallel computing power scheduling. The core prediction algorithm is based on the Python language environment and relies on the TensorFlow and Keras advanced deep learning frameworks for construction and operator implementation, thereby successfully completing the network topology connection, calculation of the custom weighted loss function, and iterative training and online inference evaluation tasks of the prediction model described in this invention.

[0114] See Table 1 for specific configuration details:

[0115] Table 1

[0116]

[0117] 2. Experiment Content:

[0118] Using a real, high-precision, long-term GNSS static monitoring dataset, which initially encompassed 35 dimensions of raw observation information including triaxial displacement, positioning solution quality indicators, and environmental redundancy features (see Table 2), a principal component analysis (PCA) dimensionality reduction strategy was first introduced for the environmental redundancy features exhibiting strong collinearity. The first principal component was extracted and horizontally concatenated with the triaxial target displacement components and the five-dimensional core quality constraint indicators to construct a 9-dimensional composite feature vector with clear physical meaning. Subsequently, the max-min scaling method was used to linearly map the composite features to a normalized interval of [0.1, 0.9]. Based on a sliding window algorithm with a time step of 10, the continuous monitoring sequence was divided into a three-dimensional temporal supervised learning tensor, with 80% and 20% allocated to the training and test sets, respectively. A separate long-span data segment without any parameter updates was reserved for independent generalization inference experiments.

[0119] Table 2

[0120]

[0121] The PCA-WMSE-Attention-BiLSTM prediction architecture constructed in this invention was used to perform deep training, ablation comparison, and online inference verification on the constructed time-series samples. In the specific operation of model training, the Adam optimization algorithm was adopted, with a batch size of 128 and 200 global iterations. During the backpropagation optimization stage, a WMSE weighted loss function with extremum awareness was applied to impose a high penalty on extrema points falling into the high-amplitude range. To rigorously demonstrate the effectiveness of the various technical features of this invention, under the condition that the hidden layer neuron dimensions, iteration rounds, and linear output activation functions were completely consistent, four sets of ablation experiments were simultaneously conducted based on a single LSTM, BiLSTM, Attention-BiLSTM, and the complete model of this invention. The prediction performance of each model was quantitatively evaluated from multiple dimensions using root mean square error, mean absolute error, coefficient of determination, and maximum absolute error.

[0122] Furthermore, to visually examine the model's dynamic focusing response capability to transient displacement jumps under actual high-frequency noise interference, and its global trend tracking performance in the face of unknown complex deformations, the experiment extracted typical displacement component prediction outputs from the model during the training set fitting, test set validation, and independent unknown data full sequence inference stages, and then performed a qualitative comparison by visual overlay with the actual observed trajectories, as shown in the following figures. Figure 2 , Figure 3 and Figure 4 As shown.

[0123] 3. Experimental Results:

[0124] To verify the effectiveness and advancement of the prediction system proposed in this invention, this embodiment compares the PCA-WMSE-Attention-BiLSTM model of this invention with existing conventional solutions from two dimensions: qualitative visualization and quantitative ablation experiments.

[0125] Figure 2 This is a comparison of the POSE dimension displacement prediction curve and the actual sequence during the training phase of this invention. It clearly shows a typical step-change signal, with the orange dashed predicted value and the blue solid observed value highly overlapping. The original GNSS observation sequence not only contains high-frequency spike noise but also experiences step jumps near epochs 1100 and 3600. Traditional recurrent neural networks are prone to severe lag and amplitude collapse in such strong interference scenarios, while the model of this invention accurately captures these two abrupt inflection points. The prediction curve not only shows no obvious phase delay but also achieves amplitude wrapping in the extreme value range after the jump. This intuitively and powerfully demonstrates the fitting ability of the self-attention mechanism and the WMSE weighted loss function in focusing on extreme value features and overcoming amplitude reduction.

[0126] Secondly, to verify whether the model has fallen into the overfitting trap, Figure 3 The model's prediction performance is shown on a test set where it was not involved in parameter updates.

[0127] Figure 3 This is a comparison chart of the POSE-dimensional displacement prediction curve and the actual sequence during the testing phase of this invention. Even with previously unseen, highly noisy data, the green dashed line in the test set shows that the predicted values ​​still closely track the actual displacement evolution trend. Faced with entirely new input high-frequency test data, the model's output prediction curve (green dashed line) still exhibits extremely high robustness. Despite the presence of complex local oscillations and multipath effects in the test data, the model not only did not diverge but also stably eliminated invalid noise, closely tracking the low-frequency main trend of the actual displacement.

[0128] To further examine the reliability of the model in actual engineering early warning deployment, this invention uses a training model with fixed parameters to perform full-sequence inference on a completely independent new GNSS time series data that has not participated in any training or verification stages. Figure 4 It presents the complete reasoning for the most complex deformation direction in elevation.

[0129] exist Figure 4In this new data series, the time span is tens of thousands of epochs and includes several dramatic step deformations and pullbacks. Faced with these extremely challenging and drastic fluctuations, the prediction model of this invention (orange dashed line) can still achieve high-precision dynamic wrapping of the actual trajectory (blue solid line) without any obvious phase lag or amplitude collapse, demonstrating the model's dynamic fitting and generalization ability when dealing with long-span, multi-abrupt sequence.

[0130] To rigorously demonstrate the independent contribution of each improved module in the model of this invention, four ablation experiments were designed under the condition that the dimensions of neurons in the hidden layers of the network, the number of iteration rounds, and the physical mapping constraints were completely consistent: the performance of single LSTM, BiLSTM, Attention-BiLSTM, and the final model of this invention, PCA-WMSE-Attention-BiLSTM. The experimental results on the test set are shown in Table 3:

[0131] Table 3

[0132]

[0133] As shown in Table 3, the technical solution of this invention achieves the best overall prediction performance: Firstly, the model utilizes a bidirectional BiLSTM mechanism to introduce reverse temporal sequence to construct a global state context, resulting in a substantial decrease in RMSE and MAE compared to traditional unidirectional networks, significantly improving prediction stability. Secondly, a self-attention module is superimposed, which, through dynamic allocation of weights at different time steps, successfully filters out high-frequency irrelevant noise and accurately focuses on local deformation signals, further reducing the MaxE index. Finally, after fully integrating PCA feature reduction and WMSE weighted loss mechanisms, the model of this invention achieves a significant breakthrough in the most challenging aspect of extreme value error control. Taking the POSE dimension as an example, its MaxE decreases significantly by 9.2% to 0.004339 compared to the baseline LSTM, and the coefficient of determination... The value jumped to a maximum of 0.1811, and the overall RMSE and MAE remained at the lowest optimal level, which fully demonstrates the comprehensive leading advantage of this architecture in global trend fitting and abrupt extreme value capture.

[0134] To highlight the beneficial effects of the present invention, the following description is provided in conjunction with the table below:

[0135] Table 4

[0136]

[0137] Based on the model training hyperparameter configurations shown in the table above, this invention significantly improves the overall prediction performance and robustness of the system through refined low-level parameter coordination. Specifically, a custom WMSE weighted loss function and a Linear output activation function are used to force the network to focus on extreme value jump regions from the bottom layer of the algorithm and avoid the forced truncation of physical displacement values, effectively overcoming the core pain point of "amplitude reduction" at abrupt changes. At the same time, a rigorous dual anti-overfitting mechanism is constructed by setting a hidden layer dropout rate of 0.2 and a cross-validation set ratio of 10%, which effectively removes pseudo-jump noise caused by multi-path effects and greatly enhances the model's spatial generalization ability in unknown and complex environments. In addition, the comprehensive application of simplified input features (InputDim) to 9 dimensions, a sliding window (Lookback) of length 10, and the Adam adaptive optimization algorithm successfully avoids the curse of dimensionality and memory overload, while taking into account the efficient extraction of bidirectional temporal continuous features and the fast and stable convergence of the non-convex optimization space, thus providing solid parameterized implementation support for achieving high-precision, low-latency dynamic displacement early warning.

[0138] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0139] The above simulation analysis proves the correctness and effectiveness of the method proposed in this invention.

[0140] The displacement prediction method proposed in this invention, through deep fusion of physical space feature dimensionality reduction, bidirectional temporal context extraction, key abrupt change feature focusing, and bidirectional deviation perception weighted loss mechanism, can achieve accurate and lag-free extreme displacement prediction. This effectively overcomes the risk of missed reports in existing early warning systems and provides crucial intelligent decision support and technical assurance for the following fields: In disaster early warning and emergency response, it rapidly and without delay captures instantaneous jumps and extreme deformations of structural displacement, providing highly reliable underlying data support for disaster prevention and mitigation early warning platforms at all levels. This helps emergency management departments secure a valuable "golden window" before disasters (such as the acceleration phase of landslides or precursors to piping in foundation pits) occur, enabling timely evacuation and traffic control, and maximizing the protection of life and property. In the health monitoring and operation and maintenance of large-scale infrastructure, specifically for cross-sea bridges, super high-rise buildings, and water conservancy dams, this invention can effectively isolate environmental noise interference such as wind load, temperature, and instantaneous vehicle impact, accurately tracking the true dynamic response trajectory of the structure under extreme loads. This provides a solid scientific basis for fatigue damage assessment, targeted repair and reinforcement, and safe operation throughout the entire life cycle of infrastructure. In the construction of deep foundation pits and complex underground engineering projects, this model can keenly identify minute abrupt changes and instability trends in the support structure during deep foundation pit excavation or tunnel boring in densely populated urban areas. It dynamically guides the adaptive adjustment of the foundation pit support strength, effectively preventing collapse accidents and ensuring the absolute safety of surrounding buildings and underground pipelines. In the area of ​​geological disaster prevention and control, and regional governance, facing landslide and debris flow hazard areas with complex terrain and significant multi-path effects, this method can accurately extract the true creep law and deformation inflection point of the sliding surface from high-noise monitoring signals. This provides high-precision data verification for refined prevention and control of regional geological disasters, risk micro-zoning, and comprehensive governance project acceptance. In scientific research and the evolution of digital twins, the high-fidelity, disturbance-resistant time-series prediction results output by this system can provide high-quality sample accumulation for structural dynamics mechanism analysis and geological evolution law exploration, thereby promoting the construction of the digital twin foundation for civil engineering and the continuous iterative optimization of intelligent construction prediction models.

[0141] The parts of this invention not described in detail are common knowledge to those skilled in the art.

[0142] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, those skilled in the art, after understanding the content and principle of the present invention, may make various modifications and changes in form and detail without departing from the principle and structure of the present invention. However, these modifications and changes based on the concept of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A GNSS displacement prediction method based on multi-source feature fusion and attention mechanism, characterized in that, Includes the following steps: (1) Obtain the original data of GNSS high-precision positioning observation, extract and classify it into the displacement features of the target to be predicted, the core positioning quality index features and the redundant auxiliary features; (2) Principal component analysis (PCA) is used to perform deep physical space dimensionality reduction and decoupling of redundant auxiliary features, and the first principal component is extracted. Then, it is horizontally spliced ​​with the displacement features of the target to be predicted and the core positioning quality index features to construct a composite feature vector. (3) Perform data standardization preprocessing on the composite feature vector to eliminate the difference in dimensions. Then, use the sliding window algorithm to construct three-dimensional tensor samples with historical time context and divide them into model training set, validation set and test set according to time order. (4) Construct a deep displacement prediction network model PCA-Attention-BiLSTM, which includes a core inference layer, a self-attention mechanism layer and a fully connected output layer. The core inference layer uses a two-layer deep stacked bidirectional long short-term memory network BiLSTM. The self-attention mechanism layer receives the output features of BiLSTM at each time step and generates a unified context representation. The fully connected output layer consists of a random deactivation layer Dropout and a fully connected layer Dense, which are used to perform high-dimensional mapping of the context feature vector and further restore it to obtain the displacement prediction value. (5) Construct a weighted mean square error (WMSE) loss function based on bidirectional deviation perception, and use the training set and validation set to iteratively optimize the deep displacement prediction network model PCA-Attention-BiLSTM using the Adam optimization algorithm of adaptive moment estimation to obtain the trained displacement prediction model. (6) Test the trained displacement prediction model using the test set, and use root mean square error, mean absolute error, coefficient of determination and maximum absolute error as evaluation indicators to evaluate the prediction performance of the model.

2. The method according to claim 1, characterized in that: The target displacement features, core positioning quality index features, and redundant auxiliary features in step (1) are obtained by hierarchical extraction and classification of the F-dimensional feature space constructed from the original GNSS high-precision positioning observation data. The implementation steps are as follows: (1.1) Extract the displacement components of the carrier in the east, north and sky directions in the coordinate system of the ENU station in the northeast and divide them into 3D displacement features of the target to be predicted; (1.2) Extract GNSS positioning quality information features, including multiple positioning quality indicators such as positioning solution confidence, number of visible satellites and standard deviation of three-axis displacement, and combine them into G-dimensional core positioning quality indicator features to introduce prior environmental quality constraints; (1.3) All remaining features, including geodetic coordinates, geocentric coordinate system coordinates, and multi-axis motion velocity components, are classified into F-3-G dimension redundant auxiliary features.

3. The method according to claim 2, characterized in that: In step (2), principal component analysis (PCA) is used to perform deep physical space dimensionality reduction and decoupling on redundant auxiliary features, and the first principal component is extracted. The specific execution steps are as follows: (2.1) Decentralize the redundant auxiliary feature matrix; (2.2) Solve for the covariance matrix of the centered matrix and perform eigenvalue decomposition on it. Determine the eigenvector corresponding to the largest eigenvalue as the first principal component. The linear combination expression of the first principal component is as follows: ， in, Original redundant auxiliary features For the corresponding load factor, =F-3-G represents the total number of dimensions of redundant auxiliary features.

4. The method according to claim 3, characterized in that: The model training set, validation set, and test set mentioned in step (3) are obtained specifically according to the following steps: (3.1) The min-max normalizer MinMaxScaler is used to linearly scale the composite feature vector after fusion to the interval [0.1, 0.9]. (3.2) Set the length of the sliding window to The sliding step size is ,and , ; The continuous time-series data is segmented to construct a data structure with shape (batch_size, ...). A three-dimensional tensor of (3+G+1); (3.3) Divide the dataset according to the time sequence: Select the first 80% of the continuous samples on the time axis as the initial training set, and extract 10% of the data from the training set as the cross-validation set. The remaining 20% ​​of the continuous samples on the time axis are divided into the test set.

5. The method according to claim 1, characterized in that: The depth displacement prediction network model PCA-Attention-BiLSTM mentioned in step (4) is specifically constructed according to the following steps; (4.1) A two-layer deep-stacked bidirectional long short-term memory network (BiLSTM) is used to construct the core inference layer, which is used to deeply mine the bidirectional temporal context dependencies of the shift sequence within the sliding window from both forward and reverse directions; in the two-layer deep-stacked bidirectional long short-term memory network (BiLSTM), the first layer contains A hidden unit, the second layer contains One hidden unit; and , ; (4.2) The MyAttention layer is connected after the core inference layer. This layer scores the output features of BiLSTM through a trainable weight matrix and obtains the attention weights using the Softmax function. Then, it aggregates the attention weights with the hidden state to generate a unified context representation. (4.3) The fully connected output layer is constructed by the random deactivation layer Dropout and the fully connected layer Dense. The fully connected layer Dense is configured with a linear activation function to perform high-dimensional mapping of the context feature vector and restore it to the displacement prediction value with actual physical meaning through the inverse normalization operation.

6. The method according to claim 5, characterized in that: The self-attention mechanism layer MyAttention described in step (4.2) generates a contextual representation according to the following steps: (4.2.1) The output features of BiLSTM at each time step are scored using a trainable weight matrix, as shown in the following formula: ， in, The hidden state features of the current time step output by BiLSTM. For a trainable weight matrix, For bias vectors, The original attention score corresponding to the current time step; (4.2.2) Using the Softmax function, the scores are mapped to attention weights in the form of a probability distribution according to the following formula: ， in, The attention weights at the current time step are normalized. The total number of time steps. The original attention score for the i-th time step, i=1,2,...,T; (4.2.3) Multiply the latent state features and attention weights element-wise, and then aggregate them along the time dimension to obtain the aggregated context feature vector. : 。 7. The method according to claim 1, characterized in that: The weighted mean square error (WMSE) loss function based on bidirectional deviation sensing, mentioned in step (5), is calculated as follows: ， in, The number of samples within a single batch. These are actual displacement observations. These are the predicted displacement values ​​for the model. This is a dynamic weight adjustment factor; The The triggering logic depends on the standardized true displacement. Absolute distance from the center point of the normalized interval by 0.5: when hour, ; when hour, ; The adaptive moment estimation Adam optimization algorithm is used to iteratively optimize the model, and its training parameter configuration includes: Set the batch size for a single-input network to [value]. With the objective of minimizing the weighted mean square error (WMSE) loss function, the global maximum number of iterations is set to... The algorithm iterates through rounds and monitors the loss using a validation set at the end of each round to prevent overfitting.

8. A GNSS displacement prediction system for implementing any one of the methods of claims 1-7, characterized in that, include: The module includes feature extraction and dimensionality reduction, data construction, model construction and training, and displacement prediction and evaluation. The feature extraction and dimensionality reduction module is used to receive raw GNSS high-precision positioning observation data, classify it, use principal component analysis (PCA) to extract core principal components to eliminate dimensional redundancy, obtain composite feature vectors, and output them to the data construction module. The data construction module is used to horizontally concatenate and standardize the extracted composite feature vectors, and construct a three-dimensional tensor sample set with historical time series constraints using a sliding window algorithm; after dividing the sample set into a training set and a test set, it is transmitted to the model construction and training module. The model building and training module is used to build a prediction network that includes a deep bidirectional long short-term memory network BiLSTM and a self-attention mechanism layer MyAttention, extract bidirectional temporal features of training set samples, and dynamically focus on extreme time steps of the self-attention mechanism; and use the weighted mean square error (WMSE) loss function based on bidirectional deviation perception for adaptive optimization training, and then transmit the trained prediction model to the displacement prediction and evaluation module. The displacement prediction and evaluation module is used to input real-time high-frequency data into the trained prediction model for online inference to predict deformation, restore the network output results to the physical displacement prediction value after inverse normalization, and evaluate the model's early warning robustness using statistical indicators including RMSE, MAE, R^2, and MaxE.

9. A displacement prediction device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the GNSS displacement prediction method based on multi-source feature fusion and attention mechanism as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the GNSS displacement prediction method based on multi-source feature fusion and attention mechanism as described in any one of claims 1 to 7.

Images