A grey GM(1,1)-RBF neural network combined model composite foundation deformation dynamic prediction method

Abstract

The application discloses a kind of grey GM (1,1) -RBF neural network composite foundation deformation dynamic prediction method, obtains foundation settlement value;The measured value of composite foundation is input into the RBF neural network model obtained by training, and the corrected residual error is obtained;Based on the corrected residual error, the deep composite foundation prediction value is obtained.The application is based on the situation that measurement cannot reach absolute equal interval in actual application process, and the method of non-equal interval grey model and RBF neural network series connection is achieved by programming three times interpolation method based on MATLAB software for equitime interval processing, which is more reasonable and accurate than the traditional calculation method for predicting the settlement of composite foundation, and the feedback information can be used to guide engineering construction.

Description

Technical Field

[0001] This invention relates to a method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model, belonging to the field of geotechnical engineering technology. Background Technology

[0002] Accurate and dynamic prediction of settlement deformation in composite foundations is challenging due to the frequent interference from external factors such as construction work and soil properties. Therefore, obtaining accurate calculations for composite foundation settlement is complex. Currently, hyperbolic, logarithmic, and exponential curve fitting are used in civil engineering to predict composite foundation settlement, but each method has its limitations, resulting in relatively large errors in the predicted settlement. GM(1,1) grey prediction falls under grey control principles. Based on differential equations, grey theory can predict settlement using relatively little data, obtaining prediction residuals and constructing a residual sequence. An RBF neural network model is then created to correct the residuals, yielding the settlement corrected by the RBF neural network. Therefore, the grey GM(1,1)-RBF neural network combination is an effective approach to solving the settlement deformation problem in composite foundations.

[0003] Zhang Jianyi, Luo Yu, Wang Shuai, Zhang Yonghua, Feng Ziyi. A method for determining the surface deformation of CFG pile composite foundation under reverse fault displacement, Patent Publication No.: CN202410895291.9, 2024.10.29. Li Chenglin, Wang Jiading, Gu Tianfeng. Research and analysis on prediction model of settlement of high embankment foundation in Northwest China [J]. Northwest Geology, 2021, 12(55): 221-235. Wang Peiyu, Cao Rihong, Yan Fang, Tang Lizhong. Settlement of soft soil foundation based on modified optimization comprehensive prediction model. Research. [J]. Building Science, 2020, 1(37): 30-35. Zhao Zhenping; Liu Xiaoqiang; Chang Ben; Yao Ming; Zhu Jianzhong; Chai Shaoqiang; Li Jianxin: Zhang Yonghong. Calculation method, system and application based on lightweight cement-soil pile composite foundation, patent publication number: CN202211093118.4, 2023.01.23. Lan Zequan, Fu Benfu, Tian Dongmei, et al. Grey prediction and analysis of particularly serious accidents in coal mines [J]. Coal Technology, 2020, 39(7): 81-83. Li Yun, Liu Zheng, Du Xiuli. Improvement of Analytic Hierarchy Process Based on Grev Correlation Model and Its Engineering Application[J]. AscE-AsME Journal of Risk and Uncertainty in Engineering Systems, Part A: Civil Engineering, 2021, 7(2). Peng Liangquan. A design method for rigid-flexible pile composite foundation, Patent Publication No.: CN202211615044.6, 2024.04.14. Zhu Wenying, Wang Haiwen, Zhang Xiaheng. Synergyeval-uation model of container multimodal transDort based on BP neural network[J]. Neural computing and Applications, 2021 (prepublish). Wang Qi, Wang Lei, Xie Shengdong, Ni Shisong, Zhang Quan, Ma Yunpeng, Chen Mingxing. A non-equidistant gray BP neural network method for predicting deep foundation pit deformation, Patent Publication No.: CN 114611820 A, March 23, 2022. Summary of the Invention

[0004] To address the problem that the settlement of composite foundations is affected by multiple factors and it is difficult to establish a suitable prediction model, this invention provides a method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model.

[0005] To achieve the above objectives, the present invention provides a method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model, comprising: acquiring the settlement measurement value of the composite foundation;

[0006] Observe the time series and process it with equal time intervals;

[0007] Firstly, a GM(1,1) grey model is constructed to predict the change of settlement over time, and the predicted value of GM(1,1) is obtained.

[0008] Obtain the prediction residuals and construct the residual training;

[0009] First, construct an RBF neural network model;

[0010] The predicted residuals are input into the RBF neural network for residual correction.

[0011] Repeat the above steps until the required number of iterations is met, and output the final RBF neural network model.

[0012] Based on the corrected residuals, the predicted values ​​of the GM(1,1) model are superimposed with the predicted values ​​of the residuals to obtain the final settlement.

[0013] First, observe the time series and process it with equal time intervals, including:

[0014] Construct an equally spaced sequence using the collected measured values ​​of the composite foundation:

[0015]

[0016]

[0017] x (1) The production sequence Y with adjacent terms (1) ;

[0018] in

[0019] The basic form of the grey GM(1,1) model can be expressed as:

[0020]

[0021] In the formula, the effective interval is (2, 2), a is the development coefficient, and b is the grey effect quantity;

[0022] Generate the accumulated torch vibration B and the constant vector Y.

[0023]

[0024] The gray coefficient is calculated using the least squares method.

[0025]

[0026] Firstly, the gray GM(1,1) composite foundation prediction model is calculated:

[0027]

[0028] Original settlement sequence:

[0029]

[0030] Settlement simulation sequence:

[0031]

[0032] Firstly, based on the predicted sequence, the residual sequence is calculated:

[0033] residual

[0034] Mean of the original settlement sequence:

[0035]

[0036] variance is

[0037] residual mean

[0038] residual variance

[0039] Then the mean squared error

[0040] Small error probability test:

[0041]

[0042] Among these, the smaller the standard deviation C, the higher the prediction accuracy; conversely, the larger the probability of small error p, the higher the prediction accuracy. The grey model uses a level one accuracy setting, with a variance C ≤ 0.35 and a small error probability p ≥ 0.95.

[0043] Preferred, based on Gaussian function Construct an RBF neural network model;

[0044] Obtain the predicted values ​​of the GM(1,1) grey model:

[0045] Prior to this, based on the predicted sequence, the residual sequence is calculated, including:

[0046] The predicted residual is Construct residual training;

[0047] The predicted residuals are used as input samples and fed into the RBF neural network model.

[0048] As the output sample of the RBF neural network model, residual correction is performed;

[0049] Firstly, the predicted values ​​from the GM(1,1) model are superimposed with the predicted values ​​from the residuals to obtain the final settlement:

[0050]

[0051] Based on the MATLAB mathematical analysis software, the gray GM(1,1)-RBF neural network model was programmed, and the gray GM(1,1) and RBF neural network were combined in series to predict the final settlement of the composite foundation.

[0052] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of any of the methods described above.

[0053] The beneficial effects achieved by this invention are as follows:

[0054] (1) This invention mainly solves the problem that it is difficult to accurately predict the settlement and deformation of composite foundations. By combining the grey model GM(1,1) and RBF neural network, a dynamic prediction method for composite foundation deformation of grey GM(1,1)-RBF neural network combined model is established to solve the problem of predicting the settlement of composite foundations.

[0055] (2) In actual engineering projects, most of the measured values ​​of composite foundations are non-equal interval sequences. Therefore, it is of great practical significance to establish a measured settlement data sequence by using cubic interpolation to construct a sequence of non-equal interval settlement monitoring data. Therefore, the grey model GM(1,1) is used to obtain the prediction sequence, and the residual sequence is calculated from the prediction sequence. The residual sequence is used as the input sample in the RBF neural network model to correct the residual sequence and improve the prediction accuracy. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1This is a flowchart of the dynamic prediction method for composite foundation deformation based on the gray GM(1,1)-RBF neural network combined model of the present invention.

[0058] Figure 2 This invention is based on the framework diagram of an RBF neural network system. Detailed Implementation

[0059] The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0060] The gray GM(1,1)-RBF neural network combined model for dynamic prediction of composite foundation deformation of the present invention includes the following steps: First, observe the time series and process the time series with equal time intervals, including:

[0061] Construct an equally spaced sequence using the collected measured values ​​of the composite foundation:

[0062]

[0063]

[0064] x (1) The production sequence Y with adjacent terms (1) ;

[0065] in

[0066] The basic form of the grey GM(1,1) model can be expressed as:

[0067]

[0068] In the formula, the effective interval is (2, 2), a is the development coefficient, and b is the grey effect quantity;

[0069] Generate the accumulated torch vibration B and the constant vector Y.

[0070]

[0071] The gray coefficient is calculated using the least squares method.

[0072]

[0073] The gray GM(1,1) composite foundation prediction model is calculated as follows:

[0074]

[0075] Original settlement sequence:

[0076]

[0077] Settlement simulation sequence:

[0078]

[0079] Based on the predicted sequence, the residual sequence is calculated:

[0080] residual

[0081] Mean of the original settlement sequence:

[0082]

[0083] variance is

[0084] residual mean

[0085] residual variance

[0086] Then the mean squared error

[0087] Small error probability test:

[0088]

[0089] Among these, the smaller the standard deviation C, the higher the prediction accuracy; conversely, the larger the probability of small error p, the higher the prediction accuracy. The grey model uses a level one accuracy setting, with a variance C ≤ 0.35 and a small error probability p ≥ 0.95.

[0090] Based on Gaussian function Construct an RBF neural network model;

[0091] Obtain the predicted values ​​of the GM(1,1) grey model:

[0092] Based on the predicted sequence, the residual sequence is calculated, including:

[0093] The predicted residual is Construct residual training;

[0094] The predicted residuals are used as input samples and fed into the RBF neural network model.

[0095] As the output sample of the RBF neural network model, residual correction is performed;

[0096] The final settlement is obtained by superimposing the predicted values ​​from the GM(1,1) model with the predicted values ​​from the residuals.

[0097]

[0098] Implementation Cases

[0099] An 18-story building with a shear wall structure and a foundation treated with cement-fly ash-gravel piles was selected. The settlement of the cement-fly ash-gravel pile composite foundation was predicted using the method of this invention. The results are shown in Table 1. Table 1. Settlement Analysis of Cement-Fly Ash-Gravel Pile Composite Foundation Using Grey GM(1,1)-RBF Neural Network Model

[0100] The data in the table shows that, based on the MATLAB mathematical analysis software, the gray GM(1,1)-RBF neural network model is programmed and corrected by the RBF neural network. The gray GM(1,1) neural network is more accurate in predicting the settlement of cement fly ash gravel pile composite foundation.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model, characterized in that, Obtain the measurement values ​​of the composite foundation; The composite foundation measurement values ​​are input into the trained gray GM(1,1) network model to obtain the prediction residuals and construct the residual sequence. Create an RBF neural network model, perform residual correction, and obtain the settlement amount corrected by the RBF neural network.

2. The method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model according to claim 1, characterized in that... The RBF neural network model was obtained through training, including: Observe the time series and process it with equal time intervals; Obtain the measured data sequence and residual sequence; The residual sequence is input into the constructed RBF neural network model to obtain the corrected residual; Repeat the above steps until the required number of iterations is met, and output the final RBF neural network model. Based on the corrected residuals, the predicted values ​​of the GM(1,1) model are superimposed with the predicted values ​​of the residuals to obtain the final settlement.

3. The method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model according to claim 2, characterized in that, Obtain the measured data sequence and residual sequence, including: Non-equidistant settlement monitoring data were used to construct a measured settlement data sequence using a cubic interpolation method. t = 0, 1, 2, 3, 4…n; x (1) The production sequence Y with adjacent terms (1) ; in The residual sequence is calculated based on the predicted sequence.

4. The method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model according to claim 3, characterized in that, Obtaining the predicted sequence using a grey model includes: The basic form of the grey GM(1,1) model can be expressed as: The effective interval in the formula is (2, 2), where a is the development coefficient and b is the gray action quantity; the accumulated torch vibration B and the constant vector Y are generated. Solve using the least squares method to calculate the gray coefficient. Solve for the development coefficient a and the grey action quantity b; The gray GM(1,1) composite foundation prediction model is calculated as follows:

5. The method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model according to claim 4, characterized in that, Based on the grey GM(1,1) prediction sequence, the residual sequence is calculated, including: Original settlement sequence: Settlement simulation sequence: residual Mean of original settlement sequence variance is residual mean residual variance Then the mean squared error Small error probability test: Among these, the smaller the standard deviation C, the higher the prediction accuracy; conversely, the larger the probability of small error p, the higher the prediction accuracy. When the grey model accuracy is set to level one, the variance C ≤ 0.35, and the probability of small error p ≥ 0.

95.

6. The method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model according to claim 2, characterized in that, Set the parameters of the RBF neural network model; GM(1,1) Grey Model Predictions: The predicted residual is Construct residual training; Create an RBF neural network model, {ξ i } as input samples for the RBF neural network, As the output sample, perform residual correction; Finally, by superimposing the predicted values ​​from the GM(1,1) model with the residual predicted values, the settlement amount corrected by the RBF neural network can be obtained.

7. The method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model according to claim 6, characterized in that, The parameters consist of an input layer, a hidden layer, and an output layer. The mathematical model uses a Gaussian function, with a nonlinear function used as the mapping from the input layer to the hidden layer and a linear function used as the mapping from the hidden layer to the output layer.

8. A method for dynamic prediction of composite foundation deformation using a gray GM(1,1)-RBF neural network combined model, characterized in that, A GM(1,1) grey model is constructed to predict the change of settlement over time.

9. The RBF neural network can be residual corrected by using the residual values ​​of the data fitted by the obtained GM(1,1) model as "training samples".

10. The settlement amount corrected by the RBF neural network can be obtained by superimposing the predicted values ​​of the GM(1,1) model with the residual predicted values.

Images