A long-term settlement prediction method integrating machine learning and creep law

Abstract

The present invention discloses a long-term settlement prediction method integrating machine learning and creep law, which relates to the field of information technology, including a long-term settlement prediction method integrating machine learning and creep law, preprocessing settlement data by time-effect weight coefficient, fitting settlement data by neural network, and then realizing long-term settlement prediction by creep formula, which specifically includes the following steps: S1: acquisition of settlement data and inversion of creep parameters; S2: building a machine learning model; S3: training strategy and designing evaluation indicators; S4: long-term settlement prediction. The present invention adopts the above-mentioned long-term settlement prediction method integrating machine learning and creep law, which has higher accuracy, wider application scope, more complete theoretical support and simpler calculation process than the existing method, and can better provide a scientific basis for the prevention of settlement disasters.

Description

Technical Field

[0001] The present invention relates to the field of radio station communication technology, and in particular to a long-term settlement prediction method integrating machine learning with creep law. Background Art

[0002] Under load, the development of foundation settlement can be divided into two stages: short-term settlement and long-term settlement. Long-term settlement of foundation is mainly caused by the creep characteristics of soil. When the geological conditions are good and the soil layer is hard, the proportion of long-term settlement in the total settlement is small. It is often ignored in engineering calculations or estimated by a certain empirical formula. When the geological conditions are poor or the settlement control requirements in the project are high, such as the construction of tunnels and subways in soft soil layers in coastal areas, high-fill airports in mountainous areas, and the construction of reclaimed islands and reefs, it is necessary to establish a more accurate calculation method to predict the long-term settlement caused by creep and prevent the occurrence of settlement disasters.

[0003] At present, there are three main methods for predicting long-term settlement: First, the mechanical properties of soil are obtained through indoor tests, the creep constitutive model of soil is selected, and a finite element model for calculating long-term foundation settlement is established. However, since the parameter calibration of the creep constitutive model is difficult, the initial stress of the foundation is often difficult to determine reliably, the finite element method has a large workload, and the quality requirements of engineering technicians are very high, so this method has not been widely used in engineering practice. It is currently mainly used for long-term settlement prediction of important projects and key areas, as well as long-term settlement mechanism analysis and research by research institutions such as universities.

[0004] Second, the measured data is processed, and the settlement law of the settlement is predicted by data fitting. This prediction method is a settlement prediction based on relatively sufficient settlement data. It can comprehensively consider the influence of various complex factors on the properties of the soil. Moreover, this method is relatively simple to use and is more convenient and quick to put into practical application. Currently, the commonly used methods include hyperbola method, exponential curve method, Poisson curve method, shallow hill method, Xingye method, three-point method, settlement rate method, and inverse analysis method. The study found that the stability of the prediction results of the curve fitting settlement method is closely related to the fitting parameters of the regression formula. The hyperbola method and the Xingye method are less affected by the fitting starting point and have good stability; the shallow hill method is greatly affected by the fitting starting point and time interval. When the time interval is different, the predicted settlement results show instability. The inverse analysis method uses the settlement data measured during the construction process to invert and determine the physical and mechanical model parameters of the foundation soil, and then substitutes the inverted parameters back into the forward analysis model for forward analysis to calculate the foundation settlement. These foundation settlement prediction methods are generally not accurate enough to meet the expected requirements due to their simple structure, and the errors will accumulate as the future prediction interval increases, resulting in poor results in long-term settlement prediction.

[0005] The third is the machine learning algorithm. The machine learning algorithm has the characteristics of high fault tolerance, adaptability and strong nonlinearity. It can not only handle the coupling of multiple parameters, but also obtain prediction results in a short time. Currently, the commonly used algorithms are BP algorithm, LSTM algorithm and SVM algorithm. The BP algorithm has a simple structure and fast calculation, but it does not have the ability to remember long-term information; the SVM method has a better mathematical model, but the training cost is high and the training process is more complicated; the LSTM network has long-term memory ability and performs better than the BP network. It has been widely used in the field of settlement prediction, such as shield machine settlement prediction, roadbed settlement prediction, airport settlement prediction, etc. However, the existing machine learning methods often take into account the creep law of soil when predicting foundation settlement. The physical meaning is not clear enough. There is a lack of machine learning methods based on creep models, and it is impossible to effectively realize the accurate prediction of long-term foundation settlement under complex geological conditions.

[0006] Therefore, it is necessary to provide a long-term settlement prediction method that integrates machine learning and creep laws to solve the above problems. Summary of the invention

[0007] The purpose of the present invention is to provide a long-term settlement prediction method that integrates machine learning and creep laws. The data is preprocessed by a time weight coefficient, and while taking into account both timeliness and regularity, the settlement data is fitted with an LSTM neural network. The long-term settlement prediction is then realized through the creep formula. The method has higher accuracy, a wider range of applications, more complete theoretical support and a simpler calculation process, and can better provide a scientific basis for the prevention of settlement disasters.

[0008] To achieve the above object, the present invention provides a long-term settlement prediction method integrating machine learning and creep law, including a long-term settlement prediction method integrating machine learning and creep law, preprocessing settlement data by time-effect weight coefficient, fitting settlement data by neural network, and then realizing long-term settlement prediction by creep formula, specifically including the following steps:

[0009] S1: Acquisition of settlement data and inversion of creep parameters;

[0010] S2: Build a machine learning model;

[0011] S3: Training strategies and design evaluation metrics;

[0012] S4: Long-term subsidence prediction.

[0013] Preferably, an aging weight coefficient x is proposed, and an aging prediction interval theory is proposed according to the creep law. The aging weight coefficient x takes into account both the timeliness and regularity of the settlement data. The smaller the proportion of x in the settlement data, the better the timeliness of the settlement data; the larger the proportion of x in the settlement data, the better the regularity of the settlement data.

[0014] The weight coefficient x is set according to the engineering MSE mean square error limit. The settlement data corresponding to x are fitted and the error is calculated. x is increased until the maximum x within the allowable error range is obtained. max .

[0015] Preferably, in step S1, the settlement data is obtained by obtaining the settlement data of a certain period of time through a settlement monitoring device, the settlement monitoring device is dispersedly deployed at multiple monitoring points to reduce errors, and at the same time, using reference points to assist, through carrier phase difference, settlement and adjustment to obtain accurate measurement data, and the measurement data is stored in a database;

[0016] The settlement data is inverted according to the creep formula, which is:

[0017] s / s0=c t lg(t / t0+A)-h t

[0018]

[0019] where c t is a dimensionless parameter reflecting the secondary consolidation coefficient; h t It is a dimensionless parameter reflecting the degree of overconsolidation; s0 and t0 refer to unit settlement and unit time respectively;

[0020] Substitute the settlement data into the above formula and invert the parameter c according to the least squares method. t 、h t , and the c corresponding to each sedimentation time period t 、h t Together with the sedimentation value, it is used as input to train the machine learning model.

[0021] Preferably, in step S2, the machine learning model uses LSTM as a framework for multi-input training, and introduces a creep parameter c t 、h t , using the memory-forget mechanism, using the memory and forget gate structure, using the sigmoid function as the activation function, and the tanh function to calculate the activation gate, m, h, c represent the input value, output value, and memory value respectively;

[0022] In order to avoid the impact of sudden changes in measurement data on the prediction results, the formula is as follows:

[0023] C t =tanh(W c ·[h t-1 ,m t ]+b c )

[0024] i t =σ(Wi ·[h t-1 ,m t ]+b i

[0025] The LSTM framework structure adopts a Bi-LSTM dual structure.

[0026] Preferably, in step S3, the parameters are set as timestep=12, sample=50, epoch=50 during training; the evaluation index adopts the mean square error index of the loss function;

[0027] The measured data is divided into a training set and a test set according to the timeliness weight coefficient x, with the data with a proportion of 1-x as the training set and the remaining data as the test set;

[0028] The data corresponding to different x are input into LSTM for training and fitting and the mse error is calculated. The error limit is set to 1%-5%, and the maximum x is obtained within the error limit. max , set the increment of x to 0.1, input the first 10% of the data into LSTM for fitting, design the mse index according to the engineering requirements, if the accuracy meets the engineering requirements, increase x to 0.2 and repeat the operation until the accuracy does not meet the requirements;

[0029] Under the condition of meeting the accuracy requirements, the larger the x is, the longer the measured data interval length t is. m Increase, the time prediction interval t e The larger the value, the better the regularity of the data. After determining x, the corresponding measurement data is input into the trained LSTM to perform fitting optimization on the discrete measured data. The predicted data obtained by LSTM is used to repair abnormal points and supplement missing points of the measured data to generate a continuous and smooth measurement data curve.

[0030] Preferably, in step S4, according to the best prediction result x obtained in step S3 max Obtain the continuous curve of the measured data after processing and optimization, and substitute it into the creep formula to calculate the t 、h t The long-term settlement prediction curve was obtained by parameter fitting.

[0031] Therefore, the present invention adopts the above-mentioned long-term settlement prediction method integrating machine learning and creep law, which has the following beneficial effects:

[0032] (1) The design concept of the Bi-LSTM model of the present invention is to make the feature data obtained at time t have information between the past and the future at the same time. This neural network structure model has better feature extraction efficiency and performance than a single LSTM structure model.

[0033] (2) The present invention deploys the long-term prediction results to the online settlement monitoring and early warning platform for visual display, which is convenient for intuitive real-time monitoring of the settlement status.

[0034] (3) The present invention can achieve better results and wider versatility on data of different working conditions by introducing machine learning strategies and a time weight coefficient x that takes into account timeliness and regularity, and obtain higher prediction accuracy while using existing data to predict the future.

[0035] (4) The present invention adopts a method that integrates machine learning and creep law. Compared with other methods, it has clearer physical meaning and simpler operation process. It is not only simple to calculate, but also ensures the scientificity and objectivity of the method, providing a scientific basis for the long-term settlement disaster prevention decision-making of relevant departments.

[0036] (5) The present invention achieves the integration of machine learning strategies and physical models to solve practical problems, that is, it retains the characteristics of high fitting accuracy and strong versatility of machine learning methods, and combines the advantages of rigorous, scientific and reliable logic of physical models.

[0037] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of a long-term settlement prediction method integrating machine learning and creep law of the present invention;

[0039] Figure 2 It is a structural diagram of LSTM in the present invention;

[0040] Figure 3 It is a principle diagram of the time-effect prediction interval in the present invention;

[0041] Figure 4 It is a principle diagram of the aging weight coefficient in the present invention. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0043] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0044] The words "include" or "comprises" and the like used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms "inside", "outside", "upper", "lower", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly specified and limited, the terms "attachment" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral body; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0045] like Figure 1 As shown, the present invention provides a long-term settlement prediction method integrating machine learning and creep law, including a long-term settlement prediction method integrating machine learning and creep law, preprocessing settlement data through time-effectiveness weight coefficient, fitting settlement data with neural network, and then realizing long-term settlement prediction through creep formula, and proposing time-effectiveness weight coefficient x. The principle of time-effectiveness weight coefficient is as follows Figure 4 As shown in the figure, the time-effectiveness prediction interval theory is proposed according to the creep law. The time-effectiveness weight coefficient x takes into account both the time-effectiveness and regularity of the settlement data. The smaller the proportion of x in the settlement data, the better the time-effectiveness of the settlement data; the larger the proportion of x in the settlement data, the better the regularity of the settlement data.

[0046] The weight coefficient x is set according to the engineering MSE mean square error limit. The settlement data corresponding to x are fitted and the error is calculated. x is increased until the maximum x within the allowable error range is obtained. max .

[0047] The specific steps include:

[0048] S1: Acquisition of settlement data and inversion of creep parameters; In step S1, the settlement data is acquired by using a settlement monitoring device to acquire settlement data for a certain period of time. The settlement monitoring device is dispersedly deployed at multiple monitoring points to reduce errors. At the same time, the reference point is used to assist, and accurate measurement data is obtained through carrier phase difference, settlement and adjustment, and the measurement data is stored in the database;

[0049] The settlement data is inverted according to the creep formula, which is:

[0050] s / s0=c tlg(t / t0+A)-h t

[0051]

[0052] where c t is a dimensionless parameter reflecting the secondary consolidation coefficient; h t It is a dimensionless parameter reflecting the degree of overconsolidation; s0 and t0 refer to unit settlement and unit time respectively;

[0053] Substitute the settlement data into the above formula and invert the parameter c according to the least squares method. t 、h t , and the c corresponding to each sedimentation time period t 、h t Together with the sedimentation value, it is used as input to train the machine learning model.

[0054] S2: Build a machine learning model; in step S2, the machine learning model uses LSTM as a framework for multi-input training and introduces a creep parameter c t 、h t , using the memory-forget mechanism, using the memory and forget gate structure, using the sigmoid function as the activation function, and the tanh function to calculate the activation gate, m, h, c represent the input value, output value, and memory value respectively;

[0055] In order to avoid the impact of sudden changes in measurement data on prediction results caused by environmental factors such as heavy rain and construction, the formula is as follows:

[0056] C t =tanh(W c ·[h t-1 ,x t ]+b c )

[0057] i t =σ(W i ·[h t-1 ,x t ]+b i

[0058] The LSTM framework adopts a Bi-LSTM dual structure. The neural network model structure is divided into two independent structures. The input sequence is input into two neural networks in forward and reverse order for feature extraction. The vector formed by concatenating the two output vectors (i.e. the extracted feature vectors) is used as the final feature expression.

[0059] LSTM is different from general feedforward neural networks. LSTM can use time series to analyze input. When using feedforward neural network BP, the neural network will think that the content of our input at time t is completely unrelated to the content of the input at time t+1. However, for scenarios such as long-term settlement prediction that require processing long-term continuous time data, it is obvious that the input information can be used more reasonably by using the input at time t or before to process the time t+1, so as to obtain better training results. The LSTM method with better effect on continuous time data is adopted. The principle diagram of the method is shown in the figure below. Figure 2 shown.

[0060] S3: Training strategy and design evaluation indicators; In step S3, the parameters are set as timestep=12, sample=50, epoch=50 during training; the evaluation indicator is the mean square error indicator of the loss function;

[0061] The measured data is divided into a training set and a test set according to the timeliness weight coefficient x, with the data with a proportion of 1-x as the training set and the remaining data as the test set;

[0062] The data corresponding to different x are input into LSTM for training and fitting and the mse error is calculated. The error limit is set to 1%-5%, and the maximum x is obtained within the error limit. max , set the increment of x to 0.1, input the first 10% of the data into LSTM for fitting, design the mse index according to the engineering requirements, if the accuracy meets the engineering requirements, increase x to 0.2 and repeat the operation until the accuracy does not meet the requirements, and output x max ; If the error exceeds the error limit when x is 0.6, but is still within the error range when x is 0.5, then x is finally determined to be 0.5.

[0063] Under the condition of meeting the accuracy requirements, the larger the x is, the longer the measured data interval length t is. m Increase, the time prediction interval t e The larger the value, the better the regularity of the data. After x is determined, the measured data of the interval corresponding to x is input into the trained LSTM to perform fitting optimization on the discrete measured data. The predicted data obtained by LSTM is used to repair abnormal points and supplement missing points of the measured data to generate a continuous and smooth measured data curve. Figure 3 Shown

[0064] S4: Long-term settlement prediction. In step S4, according to the best prediction result x obtained in step S3, max Obtain the continuous curve of the measured data after processing and optimization, and substitute it into the creep formula to calculate the t 、h tThe long-term settlement prediction curve is obtained by parameter fitting. Finally, the prediction results can be deployed to the online settlement monitoring and early warning platform for visual display, which is convenient for intuitive and real-time monitoring of settlement conditions.

[0065] Therefore, the present invention adopts the above-mentioned long-term settlement prediction method that integrates machine learning and creep law. By introducing machine learning strategies and a time weight coefficient x that takes into account timeliness and regularity, better effects and wider versatility can be obtained on data under different working conditions, and higher prediction accuracy can be obtained while using existing data to predict the future.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A long-term settlement prediction method integrating machine learning and creep law, characterized by: The long-term settlement prediction method includes the fusion of machine learning and creep law. The settlement data is preprocessed by the time-effect weight coefficient, the settlement data is fitted by the neural network, and the long-term settlement prediction is realized by the creep formula. Specifically, the following steps are included: S1: Acquisition of settlement data and inversion of creep parameters; S2: Build a machine learning model; In step S2, the machine learning model uses LSTM as a framework for multi-input training and introduces creep parameters , using the memory-forget mechanism, using the memory and forget gate structure, and using the function As the activation function, The function calculates the opening state of the activation gate, Respectively represent input value, output value, and memory value; In order to avoid the impact of sudden changes in measurement data on the prediction results, the formula is as follows: ； ）； The frame structure adopts Double structure; S3: Training strategies and design evaluation metrics; In step S3, the training parameters are set as =12 =50, =50; the evaluation index adopts the mean square error index of the loss function; According to the time weight coefficient The measured data is divided into training set and test set, with the proportion of The data is used as the training set, and the remaining data is used as the test set. Different time weight coefficients Corresponding data input Perform training fitting and calculate Error, set the error limit to , within the error limit, the maximum , set the time weight coefficient The increment is 0.1, and the first 10% of the data input Fitting is performed and the mse index is designed according to the engineering requirements. If the accuracy meets the engineering requirements, the time-efficiency weight coefficient is Increase to 0.2 and repeat the operation until the accuracy does not meet the requirements; Under the condition of meeting the accuracy requirements, the time weight coefficient The larger the value, the longer the measured data interval. Increase, time prediction interval The larger the value, the better the regularity of the data; determine the timeliness weight coefficient Then, the measurement data of the corresponding interval is input into the trained The discrete measured data are fitted and optimized. The obtained prediction data is used to repair abnormal points and supplement missing points of the measured data to generate a continuous and smooth measured data curve; S4: Long-term subsidence prediction.

2. The long-term settlement prediction method integrating machine learning and creep law according to claim 1 is characterized by: Propose time weight coefficient According to the creep law, the aging prediction interval theory is proposed, and the aging weight coefficient Taking into account both the timeliness and regularity of settlement data, the timeliness weight coefficient in settlement data The smaller the corresponding ratio, the better the timeliness of the settlement data; the timeliness weight coefficient in the settlement data The larger the corresponding ratio, the better the regularity of the settlement data; Time weight coefficient According to the project settings Mean square error limit, for time-effectiveness weight coefficient The corresponding settlement data are fitted and the error is calculated, and the time-effect weight coefficient is Increase gradually until the maximum time-efficiency weight coefficient within the allowable error range is obtained. .

3. The long-term settlement prediction method integrating machine learning and creep law according to claim 1 is characterized by: In step S1, the settlement data is obtained by using a settlement monitoring device to obtain settlement data for a certain period of time. The settlement monitoring device is dispersedly deployed at multiple monitoring points to reduce errors. At the same time, the reference point is used to assist, and accurate measurement data is obtained through carrier phase difference, settlement and adjustment, and the measurement data is stored in a database; The settlement data is inverted according to the creep formula, which is: ； ； in It is a dimensionless parameter reflecting the secondary consolidation coefficient; It is a dimensionless parameter reflecting the degree of overconsolidation; They refer to unit settlement and unit time respectively; Substitute the settlement data into the above formula and invert the parameters using the least squares method. , and the corresponding Together with the sedimentation value, it is used as input to train the machine learning model.

4. The long-term settlement prediction method integrating machine learning and creep law according to claim 1 is characterized by: In step S4, according to the best prediction result obtained in step S3 Obtain the continuous curve of the measured data after processing and optimization, and substitute it into the creep formula to perform The long-term settlement prediction curve was obtained by parameter fitting.

Images