An engineering structure deformation prediction method based on two-stage clustering-deep learning

Abstract

The present application relates to a kind of engineering structure deformation prediction methods based on two-stage clustering-deep learning, comprising the following steps: real-time collection different measuring point structure deformation time series data in actual engineering;The collected structure deformation time series data is preprocessed;Two-stage clustering method is used to cluster all monitoring points, and the two-stage clustering includes two stages of physical quantity clustering and measuring point clustering;According to the clustering result, the multi-channel time series data of the same kind monitoring point is input into the corresponding multi-channel multi-step prediction model trained, and the multi-channel future multi-step structure deformation prediction result is output.Compared with prior art, the prediction result of the present application makes full use of the correlation between different measuring points, has good generalization ability for actual engineering, and can effectively deal with complex data environment in actual engineering.

Description

Technical Field

[0001] This invention relates to the field of engineering structure deformation prediction, and in particular to an engineering structure deformation prediction method based on two-stage clustering-deep learning. Background Technology

[0002] Various engineering structures undergo deformations such as vibration and settlement during construction or service. Excessive deformation can lead to structural damage and failure, causing severe economic losses and even casualties. Dynamic prediction of structural deformation data allows for the timely detection of abnormal deformations that may pose serious threats to buildings, enabling timely warnings and preventative measures to be taken. This effectively reduces safety accidents and minimizes loss of life and property. Therefore, monitoring structural deformation and making predictions based on measured data is indispensable for forecasting and evaluating structural safety.

[0003] Currently, scholars both domestically and internationally have conducted extensive research on the prediction of deformation data in engineering structures, including the widespread application of neural networks and their improved algorithms, which has improved the accuracy of model predictions to some extent. Although there are high-performance models such as LSTM and Transformer that can fully utilize historical information to obtain time-series features, the prediction accuracy of time-series data is often determined by the quality and information contained in the input data. In practical engineering, the time-series data from single-point monitoring is limited in information, and the data itself exhibits nonlinear and non-stationary characteristics. For such data, the prediction performance obtained by directly applying neural network models is difficult to improve through data preprocessing and hyperparameter adjustment.

[0004] Building upon neural network models for prediction, some scholars have introduced clustering methods into the prediction process for engineering deformation data, enriching the information in the prediction input and improving the prediction effect to some extent. However, real-world engineering structural deformation data contains a wide variety of physical quantities, and the measuring points are widely distributed spatially with uneven density and complex relationships between them. For such complex monitoring time-series datasets, the performance improvement achieved by simply introducing clustering methods is quite limited. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the deformation of engineering structures based on two-stage clustering and deep learning. By using a two-stage clustering method to analyze the correlation of physical quantities, the spatiotemporal correlation of monitoring data at measurement points is fully explored, which improves the generalization ability and prediction accuracy of the model. It can effectively deal with the monitoring deformation data collected in various scenarios in actual engineering, and has certain reference value for practical engineering applications.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for predicting the deformation of engineering structures based on two-stage clustering and deep learning includes the following steps:

[0008] S1: Real-time acquisition of structural deformation time-series data at different measuring points in actual engineering projects;

[0009] S2: Perform data preprocessing on the collected structural deformation time series data;

[0010] S3: A two-stage clustering method is used to cluster all monitoring points. The two-stage clustering includes two stages: physical quantity clustering and measurement point clustering.

[0011] S4: Based on the clustering results, input the multi-channel time series data of the same type of monitoring points into the corresponding trained multi-channel multi-step prediction model, and output the multi-channel future multi-step structural deformation prediction results.

[0012] The data preprocessing includes outlier detection, data resampling, and data denoising, performed sequentially.

[0013] The anomaly detection specifically involves using an absolute median difference technique based on a sliding window to detect and replace outliers in the collected structural deformation time series data. Specifically, for the time series data of each monitoring point, a set of data within a preset fixed-length window before and after the point is selected. Based on the statistical characteristics of the data set, it is determined whether the data point is an outlier. If it is determined to be an outlier, the data point is replaced with the average of the two non-outlier data points before and after the outlier.

[0014] The data resampling specifically involves: using a downsampling strategy for data monitored multiple times within a day, selecting the average value of all monitored data for that day as the representative value; using an upsampling strategy for data monitored across days, adjusting the data monitoring frequency to once a day, and filling missing data using interpolation during the upsampling process; after resampling, all monitoring data have a uniform sampling frequency.

[0015] The data denoising process specifically involves: performing wavelet transform on the data after anomaly detection and data resampling to convert the signal to the wavelet domain, obtaining wavelet coefficients, and adjusting the wavelet coefficients according to a set threshold to remove noise; and then performing inverse wavelet transform on the denoised signal to obtain a denoised time-domain data sequence.

[0016] The physical quantity clustering includes the following steps:

[0017] S31: Based on the preprocessed data, calculate the DTW distance between the time series data of any two monitoring points to evaluate the similarity of the time series data of the two monitoring points;

[0018] S32: Based on the DTW distance between the time series data of each pair of monitoring points, a distance matrix between physical quantities is constructed using the bidirectional Hausdorff distance. The bidirectional Hausdorff distance H(A, B) between the pair of physical quantities is calculated as follows:

[0019] H(A,B)=max(h(A,B),h(B,A))

[0020]

[0021]

[0022] In the formula, a is the internal measurement point of physical quantity A, b is the internal measurement point of physical quantity B, and DTW(a, b) is the DTW distance between the time series data of measurement point a and measurement point b.

[0023] For each of the N physical quantities in the monitoring data, a bidirectional Hausdorff distance is calculated pairwise to obtain the physical quantity distance matrix.

[0024] S33: Based on the distance matrix between physical quantities, a hierarchical clustering algorithm is used to obtain a physical quantity clustering tree. Combined with the hierarchical clustering tree, the first stage of the two-stage clustering is completed, namely, physical quantity clustering.

[0025] The measurement point clustering specifically involves: based on the results of physical quantity clustering, all measurement points within the same physical quantity are clustered, and the clustering results are evaluated using the silhouette coefficient, CH index, and DBI index to determine the optimal number of clusters. Based on the optimal number of clusters, the second stage of the two-stage clustering, namely measurement point clustering, is completed. The data used in the measurement point clustering includes the spatial coordinate information of the monitoring points and the time-series data collected by the monitoring points.

[0026] S4 includes the following steps:

[0027] S41: Using the clustering results of measurement points, construct and train a multi-channel, multi-step prediction model for each type of measurement point.

[0028] S42: Based on the clustering results of the measurement points, reconstruct the time series data of the measurement points, and after performing linear normalization on the data, input the multi-channel data of the same type of measurement points into the trained multi-channel multi-step prediction model to obtain the future multi-step deformation prediction results of the monitoring points within the same category.

[0029] S43: Perform step S42 on all measuring point data to obtain the deformation prediction results for all monitoring points in the future multiple steps.

[0030] The multi-channel, multi-step prediction model is built on a deep learning network, which includes Transformer, LSTM, GRU, or CNN.

[0031] The multi-channel, multi-step prediction model evaluates the prediction results using root mean square error, normalized root mean square error, and coefficient of determination.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) The present invention uses a variety of data preprocessing methods to remove the adverse effects of abnormal data on subsequent analysis and improve the quality of the original time series data.

[0034] (2) The present invention uses a two-stage clustering method to analyze the correlation of physical quantities and fully explore the spatiotemporal correlation of monitoring data of measuring points, which helps to improve the prediction accuracy of subsequent prediction models.

[0035] (3) This invention combines clustering and deep learning neural networks to build a multi-physical quantity, multi-channel, multi-step prediction model, which can obtain more accurate prediction results.

[0036] (4) This invention has good generalization ability and can accurately predict complex measurement point datasets in different scenarios, which has certain reference value for practical engineering applications. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0038] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0039] This embodiment provides a method for predicting the deformation of engineering structures based on two-stage clustering and deep learning, such as... Figure 1 As shown, it includes the following steps:

[0040] S1: Real-time acquisition of structural deformation time-series data at different measuring points in actual engineering projects.

[0041] Specifically, this embodiment collected and organized multiple structural deformation and related data of foundation pit projects of multiple subway stations in a certain city. The specific monitored physical quantities include surface settlement, groundwater level, pipeline settlement, settlement of surrounding buildings, tilting of surrounding buildings, column settlement, lateral deformation of soil, axial force of supports, longitudinal displacement of wall top, horizontal displacement of wall top, tilting of walls and cracks in buildings. Each monitored physical quantity contains a different number of measuring points, the spatial density of different monitoring points is inconsistent, the start and end times of data for each monitoring point are not completely consistent, and the monitoring frequency of each monitoring point is not completely the same.

[0042] S2: Perform data preprocessing on the collected structural deformation time series data, specifically including anomaly detection, data resampling, and data denoising.

[0043] The characteristics of the dataset described in step S1 may lead to poor prediction results in the subsequent model prediction stage, or result in poor robustness and weak generalization ability of the model; in addition, data distortion may occur during the collection of the dataset, resulting in poor prediction performance. Therefore, in this embodiment, outlier replacement, data resampling, and data denoising operations are performed on all acquired original data sequences to ensure good data quality. Specifically, the following steps are included:

[0044] S21: Anomaly Detection

[0045] An absolute median difference (AMD) technique based on a sliding window is used to detect and replace outliers in the collected monitoring data. Specifically, for each monitoring point's time-series data, a data set within a fixed-length window preceding and following it is selected. The statistical characteristics of this data set are used to determine whether the data point is an outlier. If it is determined to be an outlier, it is replaced with the average of the two non-outlier data points before and after it, thus achieving data cleaning.

[0046] S22: Data Resampling

[0047] To achieve efficient input for the predictive model, the data after outlier processing is resampled to ensure all monitoring data have a uniform sampling frequency. Specifically, a downsampling strategy is used for data from multiple monitoring sessions within a day, selecting the average of all monitoring data for that day as the representative value. For data from cross-day monitoring, an upsampling strategy is used to ensure data is recorded once daily. During the upsampling process, missing data is filled using cubic spline interpolation to ensure data integrity and continuity.

[0048] S23: Data Denoising

[0049] To ensure data quality and analytical accuracy, denoising processing is performed on the data after outlier detection and replacement, as well as resampling. Specifically, wavelet transform is applied to the data, converting the signal to the wavelet domain to obtain wavelet coefficients. These coefficients are then adjusted according to a set threshold to remove noise. The denoised signal is then subjected to inverse wavelet transform to obtain the denoised time-domain data sequence. Throughout this process, selected evaluation metrics are used to assess the denoising effect, and the optimal wavelet basis functions and wavelet decomposition levels are determined based on the evaluation results.

[0050] In one implementation, the threshold and denoising evaluation index in step S23 specifically include:

[0051] (1) Threshold selection. The specific formula is as follows:

[0052]

[0053] Where, λ i σ represents the threshold value after wavelet decomposition at level i. i Let N be the noise estimation standard deviation for the corresponding layer, and N be the total length of the data sequence.

[0054] (2) Denoising evaluation metrics, including root mean square error (RMSE) and signal-to-noise ratio (SNR), specifically:

[0055] RMSE is calculated by taking the square root of the mean of the squared errors between the denoised data and the original data. The formula is as follows:

[0056]

[0057] SNR reflects the improvement in the signal-to-noise ratio. The calculation formula is as follows:

[0058]

[0059] In the two equations above, y i It is the true value, y′ i This is the data after noise reduction, where n is the data length.

[0060] S3: A two-stage clustering method is used to cluster all monitoring points. The two-stage clustering includes two stages: physical quantity clustering and measurement point clustering.

[0061] Specifically, S3 includes the following steps:

[0062] S31: For the preprocessed data, the similarity between data sequences of every two physical quantities is measured using the Dynamic Time Warping (DTW) method. For time series data from any two monitoring points, the DTW distance between them is calculated to assess the similarity of the time series data from the two monitoring points.

[0063] S32: Based on the DTW distance between pairwise monitoring point time series data, a distance matrix between physical quantities is constructed using the bidirectional Hausdorff distance. The bidirectional Hausdorff distance H(A, B) between pairwise physical quantities is calculated as follows:

[0064] H(A,B)=max(h(A,B),h(B,A))

[0065]

[0066]

[0067] In the formula, a is the internal measurement point of physical quantity A, b is the internal measurement point of physical quantity B, and DTW(a, b) is the DTW distance between the time series data of measurement point a and measurement point b.

[0068] For each of the N physical quantities in the monitoring data, a bidirectional Hausdorff distance is calculated pairwise to obtain the physical quantity distance matrix.

[0069] S33: Based on the distance matrix between physical quantities, a hierarchical clustering algorithm is used to obtain a physical quantity clustering tree. Combined with the hierarchical clustering tree, the first stage of the two-stage clustering is completed, namely, physical quantity clustering.

[0070] S34: Based on the results of physical quantity clustering, all measurement points of physical quantities within the same category are clustered using the K-means clustering method. The silhouette coefficient (SI), Calinski-Harabasz index (CH), and Davies-Bouldin index (DBI) are used to evaluate the clustering results to determine the optimal number of clusters. Based on the optimal number of clusters, the second stage of the two-stage clustering is completed, namely measurement point clustering. The data used for measurement point clustering includes the spatial coordinate information of the monitoring points and the time series data collected by the monitoring points.

[0071] S4: Based on the clustering results, input the multi-channel time series data of the same type of monitoring points into the corresponding trained multi-channel multi-step prediction model, and output the multi-channel future multi-step structural deformation prediction results.

[0072] Specifically, S4 includes the following steps:

[0073] S41: Utilizing the clustering results of the measurement points, construct and train a multi-channel, multi-step prediction model for each type of measurement point. That is, each type of monitoring point uses a multi-channel, multi-step prediction model, and by inputting multi-channel, multi-step historical data, obtains a multi-channel, multi-step future deformation prediction output.

[0074] The multi-channel, multi-step prediction model is built on a deep learning network, which can be, but is not limited to, Transformer, LSTM, GRU, and CNN.

[0075] In one implementation, the prediction model used in step S4 is an LSTM model, which specifically includes:

[0076] 1) Within a single cell of the LSTM network, for any single moment in the data sequence, calculate the forget gate f. t and input gate i t The value of is calculated using the following formula:

[0077] f t =σ(W f ·[h t-1 xt ]+b f )

[0078] i t =σ(W i ·[h t-1 x t ]+b i )

[0079] 2) Calculate the value of candidate memory cells using the following formula:

[0080] C′ t =tanh(W c ·[h t-1 x t ]+b c )

[0081] 3) Update the value of memory cells, the calculation formula is as follows:

[0082]

[0083] 4) Calculate the output gate o t The output value of this single unit is calculated using the following formula:

[0084] o t =σ(W o ·[h t-1 x t ]+b o )

[0085]

[0086] In the calculation formulas 1) to 4) above, x t The input data is the current timestamp, h t-1 It is the output data of the previous timestamp, C t-1 It is the value of the memory cell from the previous timestamp.

[0087] An LSTM model is constructed, consisting of two LSTM layers and one fully connected (Dense Layer). Based on the measurement point results obtained in step S3, the time-series data of the monitoring points is reconstructed, and the data is linearly normalized. The processed data is then divided into training, validation, and test sets according to an appropriate ratio. The multi-channel, multi-step prediction model is trained using the training set, validated using the validation set, and the prediction results of this invention are tested using the test set.

[0088] S42: Reconstruct the time series data of the measurement points based on the clustering results of the measurement points, and after performing linear normalization on the data, input the multi-channel data of the same type of measurement points into the trained multi-channel multi-step prediction model to obtain the future multi-step deformation prediction results of the monitoring points within the same category.

[0089] S43: Perform step S42 on all measuring point data to obtain the deformation prediction results for all monitoring points in the future multiple steps.

[0090] This embodiment utilizes the root mean square error, the normalized root-mean-square error, and the coefficient of determination (R²). 2 As an evaluation index, the predictive effect of the method of the present invention on data from different monitoring points is qualitatively evaluated.

[0091] (1) Root Mean Square Error

[0092] RMSE is calculated by taking the square root of the square of the differences between predicted and observed values. The formula is as follows:

[0093]

[0094] Where n is the number of input samples, y i It is the true value, y′ i This is a predicted value.

[0095] (2) Normalized Root-Mean-Square Error

[0096] The formula for calculating NRMSE is:

[0097] NRMSE = RMSE / (y max -y min )

[0098] Where y max y min These are the maximum and minimum values ​​of the input sample.

[0099] (3) Coefficient of Determination (R²) 2 )

[0100] R 2 It is a metric used to evaluate the goodness of fit of a regression model. It represents the proportion of variance the model can explain in the data and is typically used to compare the performance of different models. Its formula is as follows:

[0101]

[0102] Where y represents the actual value of the input sample, and y′ represents the predicted value. This represents the mean of all actual values.

[0103] Specifically, this embodiment uses the root mean square error (RMSE) to determine the prediction accuracy of the model; it uses the normalized root mean square error (NMRSE) to reflect the relative accuracy of the model's prediction, which can be used to compare the difference in prediction quality between different physical quantities and measurement points; and it uses the coefficient of determination (R²) to... 2 This value is used to evaluate the goodness of fit of the model and can be used to compare the performance of different models.

[0104] The dataset contains monitoring data for a certain station, including horizontal displacement of the wall top, ground settlement, axial force of the supports, groundwater level, wall tilt, and lateral deformation of the soil. This method was applied to this station for multi-step prediction, and the prediction results were evaluated using the above prediction indicators. The results are summarized according to the monitored quantities in Table 1. The RMSE evaluation index is related to the magnitude of the monitored values ​​themselves, while NRMSE can eliminate the influence of dimensions and numerical magnitude.

[0105] The prediction results show that this method exhibits good prediction performance for a variety of physical quantities. For the three prediction steps, the NRMSE can be controlled below 0.15, 0.17, and 0.20, respectively, with minimum values ​​reaching 0.048, 0.094, and 0.122; R... 2 The values ​​can be controlled to be above 0.77, 0.66, and 0.59, respectively, and are preferably 0.972, 0.912, and 0.845, respectively. The results of the examples demonstrate that the predicted deformation of the engineering structure by this method has a high degree of consistency with the actual deformation.

[0106] Table 1 Prediction Results

[0107]

[0108] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting the deformation of engineering structures based on two-stage clustering and deep learning, characterized in that, Includes the following steps: S1: Real-time acquisition of structural deformation time-series data at different measuring points in actual engineering projects; S2: Perform data preprocessing on the collected structural deformation time series data; S3: A two-stage clustering method is used to cluster all monitoring points. The two-stage clustering includes two stages: physical quantity clustering and measurement point clustering. S4: Based on the clustering results, input the multi-channel time series data of the same type of monitoring points into the corresponding trained multi-channel multi-step prediction model, and output the multi-channel future multi-step structural deformation prediction results; The physical quantity clustering includes the following steps: S31: Based on the preprocessed data, calculate the DTW distance between the time series data of any two monitoring points to evaluate the similarity of the time series data of the two monitoring points; S32: Based on the DTW distance between pairwise monitoring point time series data, a distance matrix between physical quantities is constructed using the bidirectional Hausdorff distance. The bidirectional Hausdorff distance H between the pairwise physical quantities is... A , B The calculation is as follows: In the formula a for A Internal measurement points of physical quantities b for B Internal measurement points of physical quantities, DTW ( a, b )Right now a measuring points and b DTW distance of time series data at measurement points; Within the monitoring data N For each physical quantity data point, perform bidirectional Hausdorff distance calculations pairwise to obtain the physical quantity distance matrix. D ∈ N×N ; S33: Based on the distance matrix between physical quantities, a hierarchical clustering algorithm is used to obtain a physical quantity clustering tree. Combined with the hierarchical clustering tree, the first stage of the two-stage clustering is completed, namely, physical quantity clustering. The measurement point clustering specifically involves: based on the results of physical quantity clustering, all measurement points within the same physical quantity are clustered, and the clustering results are evaluated using the silhouette coefficient, CH index, and DBI index to determine the optimal number of clusters. Based on the optimal number of clusters, the second stage of the two-stage clustering, namely measurement point clustering, is completed. The data used in the measurement point clustering includes the spatial coordinate information of the monitoring points and the time-series data collected from the monitoring points.

2. The method for predicting engineering structure deformation based on two-stage clustering-deep learning according to claim 1, characterized in that, The data preprocessing includes outlier detection, data resampling, and data denoising, performed sequentially.

3. The method for predicting engineering structure deformation based on two-stage clustering-deep learning according to claim 2, characterized in that, The anomaly detection specifically involves using an absolute median difference technique based on a sliding window to detect and replace outliers in the collected structural deformation time series data. Specifically, for the time series data of each monitoring point, a set of data within a preset fixed-length window before and after the point is selected. Based on the statistical characteristics of the data set, it is determined whether the data point is an outlier. If it is determined to be an outlier, the data point is replaced with the average of the two non-outlier data points before and after the outlier.

4. The method for predicting engineering structure deformation based on two-stage clustering-deep learning according to claim 2, characterized in that, The data resampling specifically involves: a downsampling strategy for data monitored multiple times within a day, selecting the average value of all monitored data for that day as the representative value; an upsampling strategy for data monitored across days, adjusting the data monitoring frequency to once a day, and filling missing data using interpolation during the upsampling process; after resampling, all monitoring data have a uniform sampling frequency.

5. The method for predicting engineering structure deformation based on two-stage clustering-deep learning according to claim 2, characterized in that, The data denoising process specifically involves: performing wavelet transform on the data after anomaly detection and data resampling to convert the signal to the wavelet domain, obtaining wavelet coefficients, and adjusting the wavelet coefficients according to a set threshold to remove noise; and then performing inverse wavelet transform on the denoised signal to obtain a denoised time-domain data sequence.

6. The method for predicting engineering structure deformation based on two-stage clustering-deep learning according to claim 1, characterized in that, S4 includes the following steps: S41: Using the clustering results of measurement points, construct and train a multi-channel, multi-step prediction model for each type of measurement point. S42: Based on the clustering results of the measurement points, reconstruct the time series data of the measurement points, and after performing linear normalization on the data, input the multi-channel data of the same type of measurement points into the trained multi-channel multi-step prediction model to obtain the future multi-step deformation prediction results of the monitoring points within the same category. S43: Perform step S42 on all measuring point data to obtain the deformation prediction results for all monitoring points in the future multiple steps.

7. The method for predicting engineering structure deformation based on two-stage clustering-deep learning according to claim 1, characterized in that, The multi-channel, multi-step prediction model is built on a deep learning network, which includes Transformer, LSTM, GRU, or CNN.

8. The method for predicting engineering structure deformation based on two-stage clustering-deep learning according to claim 1, characterized in that, The multi-channel, multi-step prediction model evaluates the prediction results using root mean square error, normalized root mean square error, and coefficient of determination.

Images