Slope safety coefficient calculation method based on transfer learning and GRU
By using transfer learning and GRU layer methods in the calculation of slope safety coefficient, the problems of high computational complexity and performance requirements in the prior art are solved, and more efficient slope safety coefficient calculation is achieved.
Patent Information
- Application Number
- CN202411428672.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2025-03-07
AI Technical Summary
The existing methods of calculating slope safety coefficients take a lot of time and require high computer performance. The complexity makes it difficult to model and calculate software.
The slope safety coefficient calculation method based on transfer learning and GRU is adopted, and the slope physical and mechanical parameters are explored, and the safety coefficient calculation model is built and trained. The data processing and feature extraction are used for transfer learning and GRU layer, and the safety coefficient is output through the full connection layer.
It improves the efficiency of obtaining slope safety factors, reduces the computational complexity and requirements for computer performance, and achieves faster and more efficient slope safety factors calculation.
Smart Images
Figure CN119578202A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of civil engineering, and in particular to a method for calculating slope safety factor based on transfer learning and GRU. Background Art
[0002] The slope safety factor is the ratio of the sliding force to the anti-sliding force in the slope rock and soil mass, and is an important indicator for evaluating slope stability.
[0003] Existing methods for calculating slope safety factors include the limit equilibrium method and the finite element method. The limit equilibrium method requires searching all possible sliding surfaces to find the sliding surface with the smallest safety factor. The finite element method uses a reduction factor to reduce the physical and mechanical parameters of the slope rock and soil until the calculated displacement no longer converges to a certain value, thereby deriving the safety factor. However, the above methods all involve software modeling and complex calculations, which take a lot of time and have high requirements on computer performance. Summary of the invention
[0004] The present invention provides a slope safety factor calculation method based on transfer learning and GRU to solve the problems raised by the above background technology.
[0005] To achieve the above object, the present invention provides the following technical solution: a slope safety factor calculation method based on transfer learning and GRU, comprising the following steps:
[0006] Step S1: survey the slope whose safety factor is to be calculated and obtain its physical and mechanical parameters;
[0007] Step S2: Building and training a safety factor calculation model;
[0008] Step S3: Calculate the slope safety factor;
[0009] Step S4: transfer learning;
[0010] Step S5: Model training.
[0011] As a preferred technical solution of the present invention, step S1 specifically includes the following steps: slope rock and soil mechanical parameters, slope shape parameters; the slope rock and soil mechanical parameters include soil physical and mechanical indicators, rock and soil properties; the slope shape parameters include slope shape and size, structural surface properties; the soil physical and mechanical indicators are obtained through laboratory testing and field measurement, including the natural bulk density, saturated bulk density, floating bulk density, porosity and head gradient of the soil, which are used to evaluate the physical properties and mechanical properties of the soil; the rock and soil properties are obtained through drilling, exploration wells, exploration trenches and geophysical exploration, including the type, origin, properties, thickness of the covering layer, shape and slope of the bedrock surface, rock weathering and integrity.
[0012] As a preferred technical solution of the present invention, the slope shape and size are obtained by determining the height, slope and shape data of the slope through geological surveying and mapping, which are used to evaluate the stability of the slope. The structural surface properties are obtained by obtaining the type, grade, occurrence, degree of development and extension of the main structural surfaces through exploration and testing, which are used to analyze the stability of the structural surface on the slope.
[0013] As a preferred technical solution of the present invention, step S2 specifically includes the following steps: taking the various physical and mechanical parameters of the slope obtained in step S1 as the input layer and the corresponding safety factor as the output layer, and building and training a slope safety factor calculation model based on GRU;
[0014] In the above, GRU processes data information as a GRU layer, and the basic architecture of the GRU layer is as follows:
[0015] Update gate formula: z t =g(W (x) x t +U (x) h t-1 )
[0016] Reset gate formula: r t =g(W (N) x t +U (r) h t-1 )
[0017] Current memory content: h' i =tanh(Wx i +r i ) i-1 )
[0018] Current time final memory: h t =z t ⊙h t-1 +(1-z t )⊙h;
[0019] The input layer inputs data and then passes through two one-dimensional convolutional layers. Both of the one-dimensional convolutional layers are composed of 16 convolution kernels of size 4, with a step size of 1 and a linear activation function. After the two convolutional layers extract features from the data, the data passes through two GRU layers, which are also the core of the model, and finally passes through two fully connected layers. The first fully connected layer integrates all the features of the upper layer, and finally outputs the decomposition result through the second fully connected layer of N=1.
[0020] As a preferred technical solution of the present invention, step S3 specifically includes the following steps: for the labeled safety factors corresponding to the physical and mechanical parameters of the slope rock and soil mass obtained in the survey of step S1, a sample set A is constructed with the parameters of each slope and the safety factor of each slope as a sample, and the slope safety factor calculation model is trained through the sample set A to obtain the trained slope safety factor calculation model.
[0021] As a preferred technical solution of the present invention, step S4 specifically includes the following steps: data set preparation, data and model loading, source domain model construction, migration model construction, hyperparameter adjustment, and testing; the data set preparation includes two data sets, one is the source domain data set for training the basic model, and the other is the target domain data set for training the migration model; the data and model loading includes loading the pre-trained network model and the target domain data set; the source domain model construction includes constructing the source domain data set and training the learning model; the migration model construction is used to set the remaining network layers as migration layers and add new network layers on the basis of the source domain model construction, except for the fully connected layer, to adapt to the data of the target domain; the hyperparameter adjustment is used to fine-tune the model by reducing the initial learning rate of the source domain model to slow down its learning rate while increasing the learning rate of the fully connected layer; the test includes testing the above model using unlabeled sample data of the target domain, and evaluating the accuracy of the model on the target domain data through multiple tests, so as to obtain the recognition result.
[0022] As a preferred technical solution of the present invention, step S5 specifically includes the following steps: the training step includes:
[0023] a. Parameter initialization: set weights and learning rate parameters, and apply the above parameters to the model;
[0024] b. Establish the loss function and set the stopping condition: The loss function is an important learning criterion used to measure the performance of the network. The loss function is related to the optimization amount, and the mean square error loss function is used to calculate the mean square error. The calculation formula is as follows:
[0025]
[0026] Where N is the length of the input sequence, is the predicted value of the model, y i is the true value;
[0027] c. Optimize the weight and bias parameters in the model by applying the Adam optimizer;
[0028] d. By adjusting the parameters and functions, an effective network training process is achieved. That is, according to the pre-set number of training times, the parameters of the model are adjusted until a certain threshold is reached. Therefore, when the expected number of training times is not reached, the model can stop running until the threshold is met. After 3 epochs, if the loss function has not been significantly improved, the deep learning model will terminate.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] 1. The slope safety factor calculation method based on transfer learning and GRU takes the physical and mechanical parameters of the slope as the input layer, passes through two one-dimensional convolution layers to extract features of the data, then passes through two GRU layers, and finally passes through two fully connected layers to output the corresponding safety factor, thereby improving the efficiency of obtaining the safety factor.
[0031] 2. The slope safety factor calculation method based on transfer learning and GRU takes the parameters of each slope and the safety factor of each slope as a sample, trains the slope safety factor calculation model through the sample, and obtains the trained slope safety factor calculation model. The learning rate of the source domain model is slowed down by reducing its initial learning rate while increasing the learning rate of the fully connected layer, and multiple tests are performed to evaluate the accuracy of the model on the target domain data, thereby obtaining the recognition result, and adjusting the model parameters according to the pre-set number of training times, thereby assisting the deep learning of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 This is a schematic diagram of the transfer learning principle of the present invention;
[0033] Figure 2 This is a block diagram of the GRU layer structure of the present invention. DETAILED DESCRIPTION
[0034] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.
[0035] The present invention provides a technical solution: a slope safety factor calculation method based on transfer learning and GRU, comprising the following steps:
[0036] Step S1: survey the slope whose safety factor is to be calculated and obtain its physical and mechanical parameters;
[0037] Step S2: Building and training a safety factor calculation model;
[0038] Step S3: Calculate the slope safety factor;
[0039] Step S4: transfer learning;
[0040] Step S5: Model training.
[0041] In a preferred embodiment, step S1 specifically includes the following steps: mechanical parameters of the rock and soil mass of the slope, and the shape parameters of the slope; the mechanical parameters of the rock and soil mass of the slope include the physical and mechanical indicators of the soil mass and the properties of the rock and soil mass; the shape parameters of the slope include the shape and size of the slope, and the properties of the structural surface; the physical and mechanical indicators of the soil mass are obtained through laboratory testing and field measurements, including the natural bulk density, saturated bulk density, floating bulk density, porosity and head gradient of the soil, which are used to evaluate the physical properties and mechanical properties of the soil mass; the properties of the rock and soil mass are obtained through drilling, exploration wells, exploration trenches and geophysical exploration, including the type, origin, properties, thickness of the covering layer, shape and slope of the bedrock surface, rock weathering and integrity, and the mechanical parameters of the rock and soil mass of the slope refer to the data of the physical and mechanical properties used to describe and analyze the stability of the slope. These parameters include the natural bulk density, saturated bulk density, cohesion, internal friction angle, shear strength, etc. of the soil. These parameters play a vital role in slope engineering design, stability analysis and treatment engineering, and can help evaluate the stability and safety of the slope. Slope shape parameters refer to data describing the geometric shape and size of the slope, such as the height, slope, width, etc. These parameters have important guiding significance for the design and construction of the slope, and can help determine the reasonable shape and size of the slope to ensure its stability and safety.
[0042] In a preferred embodiment, the slope shape and size are obtained by geological surveying to determine the height, slope, and shape data of the slope, which are used to evaluate the stability of the slope. The structural surface properties are obtained by exploration and testing to obtain the type, grade, occurrence, degree of development, and extension of the main structural surfaces, which are used to analyze the structural surface effect on the slope stability.
[0043] In a preferred embodiment, step S2 specifically includes the following steps: using the various physical and mechanical parameters of the slope obtained in step S1 as the input layer and the corresponding safety factor as the output layer, and building and training a slope safety factor calculation model based on GRU;
[0044] In the above, GRU processes data information as a GRU layer. The basic architecture of the GRU layer is:
[0045] Update gate formula: z t =g(W (x) x t +U (x) h t-1 )
[0046] Reset gate formula: r t =g(W (N) x t +U (r) h t-1 )
[0047] Current memory content: h' i =tanh(Wx i +r i ) i-1 )
[0048] Current time final memory: h t =z t ⊙h t-1 +(1-z t )⊙h;
[0049] The input layer inputs data, and then passes through two one-dimensional convolutional layers. Both one-dimensional convolutional layers are composed of 16 convolution kernels of size 4, with a step size of 1 and a linear activation function. After the two convolutional layers extract features from the data, it passes through two GRU layers, which are also the core of the model, and finally passes through two fully connected layers. The first fully connected layer combines all the features of the upper layer, and finally outputs the decomposition result through the second fully connected layer of N=1. GRU (Gated Recurrent Unit) is a special recurrent neural network that controls the flow and update of information by introducing reset gates and update gates. These two gating mechanisms are the core features of GRU, which are used to improve the efficiency and effectiveness of the model when processing sequence data. The main function of the update gate is to determine how much of the new state is composed of copies of the old state. It calculates the output through the sigmoid function, which controls how much of the new state retains the information of the old state. If the output of the update gate is connected If the output is close to 1, the new state will be composed entirely of copies of the old state, and the new input information will be completely ignored; if the output is close to 0, the new state is independent of the old state15; the reset gate is used to determine the degree of forgetting previous information. It controls how the hidden state of the previous time step flows into the candidate hidden state of the current time step. The reset gate helps capture short-term dependencies in time series because it allows the model to reconsider previous state information at each time step and make appropriate adjustments to new inputs to better capture these short-term dependencies; the one-dimensional convolution layer is a special convolution layer that is mainly used to process one-dimensional data, such as time series data. It performs multiplication and addition operations on the input signal by sliding the convolution kernel to achieve feature extraction. The size and number of convolution kernels can be adjusted to meet different feature extraction requirements; the convolution kernel is the core component in the one-dimensional convolution layer. It is a weight vector used to slide on the input signal and perform multiplication and addition operations. The size of the convolution kernel determines the range of local features it can capture, while the number of convolution kernels determines the dimension of the output feature map. Linear activation functions (such as ReLU functions) are usually used after one-dimensional convolution layers to enhance the nonlinear expression ability of the model. The linear activation function sets the negative values in the input signal to zero and retains the positive values, thereby increasing the nonlinearity of the model and helping to learn more complex features.
[0050] In a preferred embodiment, step S3 specifically includes the following steps: for the annotated safety factors corresponding to the physical and mechanical parameters of the slope rock and soil mass obtained in step S1, a sample set A is constructed with the parameters of each slope and the safety factor of each slope as a sample, and the slope safety factor calculation model is trained through the sample set A to obtain the trained slope safety factor calculation model.
[0051] In a preferred embodiment, step S4 specifically includes the following steps: data set preparation, data and model loading, source domain model construction, migration model construction, hyperparameter adjustment, and testing; data set preparation includes two data sets, one is the source domain data set for training the basic model, and the other is the target domain data set for training the migration model; loading data and models includes loading a pre-trained network model and a target domain data set; source domain model construction includes constructing a source domain data set and training a learning model; migration model construction is used to set the remaining network layers as migration layers and add new network layers on the basis of source domain model construction, except for the fully connected layer, to adapt to the data of the target domain; hyperparameter adjustment is used to fine-tune the model by reducing the initial learning rate of the source domain model to slow down its learning rate while increasing the learning rate of the fully connected layer; testing includes testing the above model using unlabeled sample data of the target domain, and evaluating the accuracy of the model on the target domain data through multiple tests to obtain recognition results.
[0052] In a preferred embodiment, step S5 specifically includes the following steps:
[0053] a. Parameter initialization: set weights and learning rate parameters, and apply the above parameters to the model;
[0054] b. Establish the loss function and set the stopping condition: The loss function is an important learning criterion used to measure the performance of the network. The loss function is related to the optimization amount, and the mean square error loss function is used to calculate the mean square error. The calculation formula is as follows:
[0055]
[0056] Where N is the length of the input sequence, is the predicted value of the model, y i is the true value;
[0057] c. Optimize the weight and bias parameters in the model by applying the Adam optimizer;
[0058] d. By adjusting the parameters and functions, an effective network training process is achieved. That is, according to the pre-set number of training times, the parameters of the model are adjusted until a certain threshold is reached. Therefore, when the expected number of training times is not reached, the model can stop running until the threshold is met. After 3 epochs, if the loss function has not been significantly improved, the deep learning model will terminate.
[0059] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0060] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A slope safety factor calculation method based on transfer learning and GRU, characterized in that: The following steps are involved: Step S1: survey the slope whose safety factor is to be calculated and obtain its physical and mechanical parameters; Step S2: Building and training a safety factor calculation model; Step S3: Calculate the slope safety factor; Step S4: transfer learning; Step S5: Model training.
2. The slope safety factor calculation method based on transfer learning and GRU according to claim 1 is characterized in that: Step S1 specifically includes the following steps: slope rock and soil mechanical parameters, slope shape parameters; the slope rock and soil mechanical parameters include soil physical and mechanical indicators, rock and soil properties; the slope shape parameters include slope shape and size, structural surface properties; the soil physical and mechanical indicators are obtained through laboratory testing and field measurement, including the natural bulk density, saturated bulk density, floating bulk density, porosity and head gradient of the soil, which are used to evaluate the physical properties and mechanical properties of the soil; the rock and soil properties are obtained through drilling, exploration wells, exploration trenches and geophysical exploration, including the type, origin, properties, thickness of the covering layer, shape and slope of the bedrock surface, rock weathering and integrity.
3. The slope safety factor calculation method based on transfer learning and GRU according to claim 2 is characterized in that: The slope shape and size are obtained by determining the height, slope and shape data of the slope through geological surveying and mapping, which are used to evaluate the stability of the slope. The structural surface properties are obtained by obtaining the type, grade, occurrence, degree of development and extension of the main structural surfaces through exploration and testing, which are used to analyze the stability of the structural surface on the slope.
4. The slope safety factor calculation method based on transfer learning and GRU according to claim 1 is characterized in that: Step S2 specifically includes the following steps: using the various physical and mechanical parameters of the slope obtained in step S1 as the input layer and the corresponding safety factor as the output layer, and building and training a slope safety factor calculation model based on GRU; In the above, GRU processes data information as a GRU layer, and the basic architecture of the GRU layer is as follows: Update gate formula: z t =g(W (x) x t +U (x) h t-1 ) Reset gate formula: r t =g(W (N) x t +U (r) h t-1 ) Current memory content: h' i =tanh(Wx i +r i ) i-1 ) Current time final memory: h t =z t ⊙h t-1 +(1-z t )⊙h; The input layer inputs data and then passes through two one-dimensional convolutional layers. Both of the one-dimensional convolutional layers are composed of 16 convolution kernels of size 4, with a step size of 1 and a linear activation function. After the two convolutional layers extract features from the data, the data passes through two GRU layers, which are also the core of the model, and finally passes through two fully connected layers. The first fully connected layer integrates all the features of the upper layer, and finally outputs the decomposition result through the second fully connected layer of N=1.
5. The slope safety factor calculation method based on transfer learning and GRU according to claim 1 is characterized in that: Step S3 specifically includes the following steps: for the annotated safety factors corresponding to the physical and mechanical parameters of the slope rock and soil mass obtained in step S1, a sample set A is constructed with the parameters of each slope and the safety factor of each slope as a sample, and a slope safety factor calculation model is trained through the sample set A to obtain a trained slope safety factor calculation model.
6. The slope safety factor calculation method based on transfer learning and GRU according to claim 1 is characterized in that: Step S4 specifically includes the following steps: data set preparation, data and model loading, source domain model construction, migration model construction, hyperparameter adjustment, and testing; the data set preparation includes two data sets, one is the source domain data set for training the basic model, and the other is the target domain data set for training the migration model; the data and model loading includes loading the pre-trained network model and the target domain data set; the source domain model construction includes constructing the source domain data set and training the learning model; the migration model construction is used to set the remaining network layers as migration layers and add new network layers on the basis of the source domain model construction, except for the fully connected layer, to adapt to the data of the target domain; the hyperparameter adjustment is used to fine-tune the model by reducing the initial learning rate of the source domain model to slow down its learning rate while increasing the learning rate of the fully connected layer; the test includes testing the above model using unlabeled sample data of the target domain, and evaluating the accuracy of the model on the target domain data through multiple tests, so as to obtain the recognition result.
7. The slope safety factor calculation method based on transfer learning and GRU according to claim 1 is characterized in that: Step S5 specifically includes the following steps: a. Parameter initialization: set weights and learning rate parameters, and apply the above parameters to the model; b. Establish the loss function and set the stopping condition: The loss function is an important learning criterion used to measure the performance of the network. The loss function is related to the optimization amount, and the mean square error loss function is used to calculate the mean square error. The calculation formula is as follows: Where N is the length of the input sequence, is the predicted value of the model, y i is the true value; c. Optimize the weight and bias parameters in the model by applying the Adam optimizer; d. By adjusting the parameters and functions, an effective network training process is achieved. That is, according to the pre-set number of training times, the parameters of the model are adjusted until a certain threshold is reached. Therefore, when the expected number of training times is not reached, the model can stop running until the threshold is met. After 3 epochs, if the loss function has not been significantly improved, the deep learning model will terminate.