Displacement prediction method of immersed tube in submarine immersed tube tunnel based on deep learning model

Abstract

The present invention discloses a method for predicting the displacement of a submarine immersed tube in a submarine immersed tube tunnel based on a deep learning model. First, monitoring data of the submarine immersed tube tunnel is obtained, and the data include temperature, humidity, temperature gradient, triaxial acceleration, horizontal differential displacement, vertical differential displacement and tension and closing amount to form a time series data set; then, data features such as the mean, median, mode, standard deviation and variance of each monitoring data are calculated, and the correlation features between each time point in the data are identified, and then the long-term trend, seasonal fluctuation and random residual component features of the data are obtained; the data set and the obtained data features, correlation features and component features are used to train an LSTM model and a Transformer model; after the model training is completed, the prediction results of the two models are fused by a weighted average method to generate a final immersed tube displacement prediction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent monitoring of engineering structures, and specifically relates to a method for predicting the displacement of a submerged tube of a submarine immersed tube tunnel based on a deep learning model. Background Art

[0002] In the construction of modern large-scale infrastructure, especially in the construction of submarine immersed tube tunnels, structural health monitoring is particularly important. Submarine immersed tube tunnels are usually faced with extremely complex environmental conditions, including high water pressure, salt water erosion, uneven geological pressure, ocean currents, etc. These factors will cause the tunnel structure to sink, shift, or even crack, which seriously threatens the long-term stability and safety of the tunnel. Traditional structural health monitoring methods mainly rely on regular manual inspections and limited monitoring point data. This method is not only inefficient and costly, but also unable to capture the dynamic changes of the tunnel structure in real time, making it difficult to detect and warn potential structural problems at an early stage. In addition, the complexity of environmental conditions during the construction of submarine immersed tube tunnels, such as drastic changes in humidity, temperature difference effects, and tidal effects, further increases the difficulty of monitoring, making it difficult for traditional monitoring methods to cope with these challenges. With the continuous expansion of the scale of the project and the improvement of the requirements for tunnel safety, monitoring based on traditional methods is obviously unable to meet the needs of modern submarine immersed tube tunnel construction. There is an urgent need for a real-time, accurate, and comprehensive monitoring technology to ensure the construction and operation safety of the tunnel.

[0003] In this context, intelligent structural health monitoring systems based on deep learning have gradually become an effective way to solve this problem. Deep learning technology, especially long short-term memory networks (LSTM) and Transformer models, have shown excellent performance in time series data analysis with their powerful data processing capabilities and ability to capture complex nonlinear relationships. The LSTM model is good at processing long-term dependencies in time series and can effectively deal with the long-term trends and historical dependencies in submarine immersed tube tunnel structure monitoring data. The Transformer model has significant advantages in processing multi-dimensional data through the self-attention mechanism, especially in capturing the complex interactive relationships between different environmental and dynamic parameters. By applying these two models to the submarine immersed tube tunnel structure health monitoring system, real-time monitoring of multiple environmental parameters such as temperature, humidity, temperature gradient, and triaxial acceleration can be achieved, and an accurate prediction model can be established by combining key structural parameters such as horizontal differential displacement, vertical differential displacement, and inter-section tension. Further, through data fusion technology, the prediction results of the LSTM and Transformer models are weighted averaged, which can significantly improve the prediction accuracy, thereby providing strong technical support for real-time safety monitoring and maintenance decision-making in tunnel construction. This technical background has laid a solid foundation for the proposal of the present invention and demonstrated the broad application prospects of deep learning-based intelligent monitoring systems in complex engineering environments. Summary of the invention

[0004] The present invention is mainly to overcome the shortcomings of the prior art and provides a method for predicting the displacement of a submerged tube in a submarine immersed tube tunnel based on a deep learning model.

[0005] The present invention is achieved through the following technical solutions:

[0006] A method for predicting the displacement of a subsea immersed tube tunnel based on a deep learning model comprises the following steps:

[0007] Step 1, obtaining monitoring data of the submarine immersed tube tunnel, including temperature, humidity, temperature gradient, triaxial acceleration, horizontal differential displacement, vertical differential displacement and tension;

[0008] Step 2: preprocess and normalize the monitoring data obtained in step 1 to form a time series data set;

[0009] Step 3, perform descriptive statistical analysis on the monitoring data to obtain the data characteristics of the mean, median, mode, standard deviation and variance of each monitoring data;

[0010] Step 4, perform autocorrelation function ACF and partial autocorrelation function PACF analysis on the monitoring data to identify the correlation characteristics between each time point in the data;

[0011] Step 5, performing Fourier transform on the monitoring data to convert the data into a frequency domain signal;

[0012] Step 6, according to the frequency domain signal obtained in step 5, the STL decomposition method is used to obtain the component characteristics of the long-term trend, seasonal fluctuation and random residual of the monitoring data;

[0013] Step 7, construct LSTM model and Transformer model, take the temperature, humidity, temperature gradient, triaxial acceleration, horizontal differential displacement, vertical differential displacement and tension in the previous time period as input variables, and take the horizontal differential displacement, vertical differential displacement and tension in the next time period as predicted output variables; use the data set in step 2 and the data features obtained in step 3, the correlation features obtained in step 4, and the long-term trend, seasonal fluctuation and random residual component features obtained in step 6 to train the LSTM model and Transformer model; after the model training is completed, the prediction results of the LSTM model and the Transformer model are fused by the weighted average method to generate the final prediction result.

[0014] In the above technical solution, in step 1, the sampling frequency of the data is dynamically adjusted based on the change rate of the historical monitoring data.

[0015] In the above technical solution, in step 2, preprocessing includes: missing value processing and outlier detection and correction.

[0016] In the above technical solution, in step 2, the missing value processing includes: 1) if the missing rate of monitoring data at a certain time point exceeds 30%, the data at that time point will be deleted; 2) if continuous missing values ​​appear within the sampling period, choose to use the first valid value in front or behind to fill the missing data.

[0017] In the above technical solution, in step 2, outliers in the monitoring data are identified based on the Z-score method, and the identified abnormal data are corrected.

[0018] In the above technical solution, in step 2, the normalization process is to scale all data to the interval [0, 1] to ensure that all features have the same scale.

[0019] In the above technical solution, in step 4, the interrelationship characteristics between the time points in the data include:

[0020] 1. Calculate the autocorrelation coefficient of the data at different lag orders through the autocorrelation function ACF;

[0021] 2. The partial autocorrelation function PACF is used to calculate the autocorrelation coefficient of the data after eliminating the indirect effect.

[0022] In the above technical solution, in step 7, when the prediction results of the LSTM model and the Transformer model are fused by the weighted average method, the weights of the two models are allocated according to the inverse proportion of the errors of the two models, and the model with smaller error is allocated a higher weight.

[0023] The advantages and beneficial effects of the present invention are:

[0024] The present invention deploys a high-precision sensor network in a large submarine immersed tube tunnel project to record and process multi-dimensional key data including temperature, humidity, temperature gradient, triaxial acceleration, horizontal differential displacement, vertical differential displacement, and tension and closing amount. These data are optimized through time series processing and adaptive sampling technology to ensure the continuity and representativeness of the data. In the data preprocessing stage, the quality and integrity of the data are improved by deleting data points with high missing rates, repairing a small amount of missing data with the forward and backward filling method, and detecting and correcting outliers with the Z-score method.

[0025] The present invention performs descriptive statistical analysis on the monitoring data to obtain the data characteristics of the mean, median, mode, standard deviation and variance of each monitoring data, which are used to provide a training basis for the immersed tube displacement prediction model; the present invention performs autocorrelation function ACF and partial autocorrelation function PACF analysis on the monitoring data to identify the characteristics of the relationship between each time point in the data, which are used to provide a training basis for the immersed tube displacement prediction model; in addition, the present invention performs Fourier transform on the monitoring data to convert the data into frequency domain signals, and then combines the STL decomposition method to decompose the time series data into the characteristics of three components, namely long-term trends, seasonal fluctuations and random residuals, so that the prediction model can more accurately capture the potential patterns in the data. Finally, the present invention achieves high-precision prediction of tunnel structure displacement by weighted averaging the prediction results of the two deep learning models, LSTM and Transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of the method for predicting the displacement of a subsea immersed tube tunnel based on a deep learning model of the present invention.

[0027] Figure 2 A schematic diagram of original data and predicted data of displacement deformation of an immersed tube tunnel provided in an embodiment of the present invention.

[0028] For ordinary technicians in this field, other relevant drawings can be obtained based on the above drawings without any creative work. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with specific embodiments.

[0030] The present invention designs a method for predicting the displacement of a submerged tube in a submarine immersed tube tunnel based on a deep learning model, and the specific steps are as follows.

[0031] Step 1: Deploy multiple high-precision sensor networks in the target submarine immersed tube tunnel project to monitor key environmental and structural data.

[0032] Monitoring data include:

[0033] Temperature (T) and temperature gradient ( ): Calculated by temperature sensor, the data range is 15-40°C;

[0034] Humidity (H): measured by humidity sensors with a humidity range of 50-100%. Humidity sensors are located on the top side wall of the carriageway of the immersed tube tunnel.

[0035] Three-axis acceleration ( A x ，A y ，A z ): obtained through an acceleration sensor with a value of 0.001-1.5g. The acceleration sensor is arranged on the side wall of the pipe segment of the immersed tube tunnel;

[0036] Horizontal difference displacement ( D h ): obtained through displacement sensor, the value range is 0-20mm; vertical difference displacement ( D v ): obtained through displacement sensor, the value range is 0-15mm; opening and closing amount ( D s ): obtained by structural strain sensors, with a value range of 0-5 mm. The displacement sensors and structural strain sensors are arranged at the small segment joints of the pipe segments of the immersed tube tunnel.

[0037] These monitoring data are stored in a structured manner in the database. The system processes these data in a time series and associates each record with a specific timestamp to form a highly structured data set. , the data monitored by each sensor is , which the system stores as:

[0038] ;

[0039] Furthermore, for the optimization of sampling intervals, adaptive sampling technology is used to dynamically adjust the sampling frequency of data based on the change rate of historical monitoring data, thereby determining the appropriate sampling interval for each type of monitoring data. This process can be expressed by the following formula: ;

[0040] in is the sampling frequency, is the sampling time interval, is the rate of change of the data.

[0041] Step 2: Preprocess and normalize the monitoring data obtained in step 1 to form a time series data set.

[0042] Preprocessing, including missing value processing and outlier detection and correction, to ensure data quality and integrity;

[0043] 1. Missing value processing: If the missing rate of monitoring data at a certain time point exceeds 30%, the data at that time point will be deleted to avoid significant impact on the overall data set. For a small number of consecutive missing data (such as missing 3-5 points), forward or backward filling methods are used to fill in. Specifically, if If the data at a certain moment is missing, you can use the previous valid data or filling:

[0044] or ;

[0045] 2. Outlier detection and correction: The Z-score-based method is used for outlier detection. The outlier judgment formula is: ;

[0046] in is the sample mean, is the sample standard deviation; if , then the point is considered as an outlier and is corrected.

[0047] Example: Assume that the original temperature data is: T(t)=[22.5, 23.0, 22.8, missing, 23.2, outlier], the processed data is: T processed (t)=[22.5, 23.0, 22.8, 22.9 (filled), 23.2, 23.0 (corrected)]. Through these processing, the original data are converted into a high-quality data set suitable for the training and analysis of the immersed tube displacement prediction model, ensuring the accuracy and reliability of the subsequent immersed tube displacement prediction model.

[0048] Normalization is to scale all data to interval to ensure that all data have the same scale, thereby improving the performance and stability of the immersed tube displacement prediction model.

[0049] Step 3: Perform descriptive statistical analysis on the monitoring data to obtain data characteristics such as the mean, median, mode, standard deviation and variance of each monitoring data, so as to quantify the central trend and dispersion characteristics of the data set, which can provide a training basis for the subsequent immersed tube displacement prediction model.

[0050] 1. Quantifying the central tendency includes the following data features:

[0051] 1. Calculate the mean: Calculate the average value of the data to reflect the central tendency of the data;

[0052] ;

[0053] in, Indicates i data, is the total amount of data;

[0054] 2. Calculate the median: represents the middle point of the data;

[0055] 3. Calculate the mode: the value that appears most frequently in the data.

[0056] 2. Quantify the degree of discreteness, including the following data characteristics:

[0057] 1. Calculate the standard deviation: calculate the volatility of the data;

[0058] ;

[0059] In the formula, is the mean;

[0060] 2. Calculate the variance: evaluate the difference between the data and the mean. The calculation formula is:

[0061] .

[0062] Step 4: Perform ACF (autocorrelation function) and PACF (partial autocorrelation function) analysis on the monitoring data to identify the correlation characteristics between each time point in the data, which provides a training basis for the subsequent immersed tube displacement prediction model.

[0063] Specifically, the mutual relationship characteristics include the following two types:

[0064] 1. Calculate the autocorrelation coefficient of the data at different lag orders through ACF (autocorrelation function);

[0065] ;

[0066] in, is the lag order, Represents time series data at a point in time The value of is the mean, Indicates that the time series data is lagging The observed value after time points, that is, the current Point Back The value at a time point. and is the deviation of the current time point and the lagged time point from the mean; N represents the total number of data points in the time series data. ACF can help identify the periodicity or trend of the data;

[0067] 2. Calculate the autocorrelation coefficient of the data after removing the indirect effect through PACF (partial autocorrelation function). .

[0068] Step 5: Perform Fourier transform on the monitoring data to convert the data into frequency domain signals.

[0069] The basic formula for Fourier transform is: ;

[0070] in, is the frequency domain signal, is the time domain signal, is the frequency.

[0071] Step 6: In the construction monitoring of submarine immersed tube tunnels, the monitoring data often have three main components: long-term trend, seasonal fluctuation and random residual. In order to better understand these data and improve the prediction accuracy of the immersed tube displacement prediction model, the STL decomposition method is used according to the frequency domain signal obtained in step 5 to obtain the three component characteristics of the monitoring data: long-term trend, seasonal fluctuation and random residual, which provide a training basis for the subsequent immersed tube displacement prediction model. The specific operations are as follows.

[0072] The long-term trend reflects the overall upward or downward trend of the data. For example, the temperature inside the tunnel gradually increases or decreases with the passage of seasons, and this change can be captured by the trend component. The trend component formula is expressed as: ;

[0073] in, LOESS is a local regression method, It is the trend value at time t, which reflects the changing trend of the data in the long term, such as a gradual increase or decrease.

[0074] Seasonality is used to capture periodic changes in data, such as seasonal fluctuations in temperature each year. Seasonality is usually reflected in patterns that repeat within a fixed period of the data. The seasonal component formula is expressed as: ;

[0075] in, Represents a time series at a point in time The seasonal component, which reflects patterns in the data that recur over a fixed period, such as annual, quarterly, or monthly fluctuations; Represents time series data at a point in time The original value of , i.e., the observed data; Represents a time series at a point in time The random residual components of are the unexplained random fluctuations in the data, usually the residuals after removing the trend and seasonal components.

[0076] The random residuals are the random changes remaining after removing the trend and seasonal fluctuations. This part of the change is usually unpredictable. The formula for the random residual component is: .

[0077] Step 7, construct LSTM model and Transformer model, take the temperature, humidity, temperature gradient, triaxial acceleration, horizontal differential displacement, vertical differential displacement and tension in the previous time period as input variables, and take the horizontal differential displacement, vertical differential displacement and tension in the next time period as predicted output variables; use the data set in step 2 and the data features obtained in step 3, the correlation features obtained in step 4, and the long-term trend, seasonal fluctuation and random residual component features obtained in step 6 to train the LSTM model and Transformer model; after the model training is completed, the prediction results of the LSTM model and the Transformer model are fused by the weighted average method to generate the final prediction result.

[0078] Specifically:

[0079] In this step, the LSTM network can well capture the long-term dependencies in the input data; the Transformer model can process multi-dimensional input features through the self-attention mechanism, combine the sequential information in the time series, and predict the future displacement.

[0080] During model training, the mean square error (MSE) is used as the loss function: ;in is the true value, is the predicted value.

[0081] After the model training is completed, the test set is used for evaluation, the mean square error (MSE) is calculated, and the parameters are continuously adjusted to ensure the accuracy of the prediction results.

[0082] Finally, the prediction results of the LSTM model and the Transformer model are combined, and the weighted average method is used to further optimize the final prediction output. First, the error performance of the LSTM model and the Transformer model on the validation set is evaluated. Common evaluation indicators include mean square error (MSE), root mean square error (RMSE) or mean absolute error (MAE). The evaluation results of the example are as follows:

[0083] The MSE of the LSTM model is: ;

[0084] The MSE of the Transformer model is: ;

[0085] The weights of the two models are assigned according to the inverse ratio of their errors, that is, the model with smaller error is assigned higher weight. Assume that the inverse ratio of MSE is used as the weight calculation method:

[0086] The weights of the LSTM model are: ;

[0087] The weights of the Transformer model are: ;

[0088] Substituting the values ​​in the above formula, we get:

[0089] ;

[0090] ;

[0091] Finally, the prediction results based on the LSTM and Transformer models and As the predicted value of future data, the final weighted average prediction result for:

[0092] .

[0093] See attached Figure 2 In this embodiment, combining the prediction results of the LSTM model and the Transformer model, the weighted average prediction results of the final vertical differential displacement, the amount of tension between pipe segments, and the horizontal differential displacement are approximately 80 mm, 72 mm, and 57 mm. Figure 2 In the timeline, 8 / 19-8 / 21 refers to the time period from August 19th to August 21st.

[0094] Furthermore, the trained prediction model is used for prediction, the proportion of outliers in the actual data and predicted data is analyzed, and the current and future health status of the immersed tube tunnel structure is evaluated in combination with the predefined health status threshold. Among them, the standard score method is used to calculate the standard score for the actual data and the predicted data respectively, and then whether the data point is an outlier is determined based on whether the absolute value of the standard score is greater than the threshold 3.

[0095] The specific health status assessment strategy is as follows.

[0096] (1) Calculate the proportion of outliers: For the actual and predicted data sets, calculate the proportion of outliers. This proportion reflects the volatility of the data and the severity of possible structural problems.

[0097] (2) Define status thresholds: The system predefines multiple health status thresholds, such as "excellent", "good", "fair", and "poor". These thresholds are based on the proportion of abnormal values. For example, an excellent status may correspond to an abnormal proportion of less than 5%, while a poor status may correspond to an abnormal proportion of more than 20%.

[0098] (3) Status determination: Based on the outlier ratio and the predefined threshold, the system will assign a health status to each data set.

[0099] (4) Comprehensive evaluation: Considering the health status of the actual data and the health status of the predicted data, the overall health status of the structure is comprehensively evaluated by weighted average method. The health status of the actual data is usually given a higher weight because they reflect the current actual condition of the structure.

[0100] The present invention is described above by way of example. It should be noted that, without departing from the core of the present invention, any simple deformation, modification or other equivalent replacement that can be made by those skilled in the art without inventive effort falls within the protection scope of the present invention.

Claims

1. A method for predicting the displacement of a submarine immersed tube tunnel based on a deep learning model, characterized in that: The steps include: Step 1, obtaining monitoring data of the submarine immersed tube tunnel, including temperature, humidity, temperature gradient, triaxial acceleration, horizontal differential displacement, vertical differential displacement and tension; Step 2: preprocess and normalize the monitoring data obtained in step 1 to form a time series data set; Step 3, perform descriptive statistical analysis on the monitoring data to obtain the data characteristics of the mean, median, mode, standard deviation and variance of each monitoring data; Step 4, perform autocorrelation function ACF and partial autocorrelation function PACF analysis on the monitoring data to identify the correlation characteristics between each time point in the data; Step 5, performing Fourier transform on the monitoring data to convert the data into a frequency domain signal; Step 6, according to the frequency domain signal obtained in step 5, the STL decomposition method is used to obtain the component characteristics of the long-term trend, seasonal fluctuation and random residual of the monitoring data; Step 7, constructing LSTM model and Transformer model, taking temperature, humidity, temperature gradient, triaxial acceleration, horizontal differential displacement, vertical differential displacement and tension and closing amount of the previous time period as input variables, and taking horizontal differential displacement, vertical differential displacement and tension and closing amount of the next time period as predicted output variables; using the data set of step 2 and the data features obtained in step 3, the correlation features obtained in step 4, and the long-term trend, seasonal fluctuation and random residual component features obtained in step 6, the LSTM model and Transformer model are trained; After the model training is completed, the prediction results of the LSTM model and the Transformer model are fused through the weighted average method to generate the final prediction result.

2. The method for predicting displacement of a subsea immersed tube tunnel based on a deep learning model according to claim 1 is characterized in that: In step 1, the sampling frequency of the data is dynamically adjusted based on the rate of change of the historical monitoring data.

3. The method for predicting displacement of a subsea immersed tube tunnel based on a deep learning model according to claim 1, characterized in that: In step 2, preprocessing includes: missing value processing and outlier detection and correction.

4. The method for predicting displacement of a subsea immersed tube tunnel based on a deep learning model according to claim 3 is characterized in that: In step 2, missing value processing includes: 1) if the missing rate of monitoring data at a certain time point exceeds 30%, the data at that time point will be deleted; 2) if there are consecutive missing values ​​within the sampling period, choose to use the first valid value before or after to fill the missing data.

5. The method for predicting displacement of a subsea immersed tube tunnel based on a deep learning model according to claim 3 is characterized in that: In step 2, outliers in the monitoring data are identified based on the Z-score method, and the identified abnormal data are corrected.

6. The method for predicting displacement of a subsea immersed tube tunnel based on a deep learning model according to claim 1, characterized in that: In step 2, the normalization process is to scale all data to the interval [0, 1].

7. The method for predicting displacement of a subsea immersed tube tunnel based on a deep learning model according to claim 1, characterized in that: In step 4, the correlation characteristics between each time point in the data include: (1) Calculate the autocorrelation coefficient of the data at different lag orders through the autocorrelation function ACF; (2) The partial autocorrelation function (PACF) is used to calculate the autocorrelation coefficient of the data after eliminating the indirect effect.

8. The method for predicting displacement of a subsea immersed tube tunnel based on a deep learning model according to claim 1, characterized in that: In step 7, when the prediction results of the LSTM model and the Transformer model are fused by the weighted average method, the weights of the two models are assigned according to the inverse ratio of the errors of the two models. The model with smaller error is assigned higher weight.

Images