Slope stability evaluation method based on neural network model

Abstract

The invention relates to the technical field of slope stability, in particular to a slope stability evaluation method based on a neural network model. According to the technical scheme, the method comprises the steps of constructing an original database, building a multi-layer perceptron neural network model, training an optimization model and achieving dynamic updating and early warning of the neural network model. According to the method, comprehensive data support is provided for the neural network model through multi-dimensional slope engineering data, the slope stability is objectively quantified, and the evaluation consistency and accuracy are ensured; meanwhile, the multi-layer perceptron neural network model can automatically learn complex features in input data, so that the model can more accurately judge the slope stability under different working conditions; according to the method, the adaptive moment estimation algorithm is combined with the early stop method to train the optimization model, efficient and reliable model training is ensured through cooperation of the two, and good performance is achieved under finite time and data resources; according to the method, the early warning credibility is improved through the Bayesian theory in combination with historical data and a model prediction result.

Description

Technical Field

[0001] The invention relates to the technical field of slope stability, and in particular to a slope stability evaluation method based on a neural network model. Background Art

[0002] Slope stability evaluation is crucial for many engineering fields, such as road construction, mining, water conservancy and hydropower engineering. Accurately assessing whether the slope is stable can effectively prevent geological disasters such as landslides, protect people's lives and property safety and the smooth progress of the project. Although traditional slope stability evaluation methods have their own advantages, their limitations are becoming more and more prominent as engineering needs become increasingly complex.

[0003] Among the existing slope stability evaluation methods, the engineering analogy method and the expert scoring method rely more on the experience of professionals, lack a unified objective standard, and are difficult to accurately quantify the degree of stability. The limit equilibrium law assumes that the slope rock and soil are ideal rigid bodies and ignores the complex internal deformation characteristics, which is inconsistent with the actual slope mechanical behavior, resulting in a large deviation in the calculated stability coefficient and low utilization of monitoring data. Therefore, we propose a slope stability evaluation method based on a neural network model. Summary of the invention

[0004] The purpose of the present invention is to propose a slope stability evaluation method based on a neural network model to address the problems that the existing methods in the background technology are difficult to accurately quantify the degree of stability, the stability coefficient has a large deviation and the monitoring data utilization rate is low.

[0005] The technical solution of the present invention is a slope stability evaluation method based on a neural network model, comprising the following steps: constructing an original database, wherein the data source of the original database is slope engineering example data under different working conditions;

[0006] Building a multi-layer perceptron neural network model to pre-process and divide samples of the data collected from the original database;

[0007] Adopting adaptive moment estimation algorithm combined with early stopping method to train and optimize the neural network model;

[0008] The dynamic update and early warning of the neural network model are achieved through incremental learning and Bayesian theory.

[0009] Optionally, the slope engineering example data includes rock and soil parameters, slope geometric characteristics, external loads and meteorological elements, and the original database is constructed in multiple dimensions;

[0010] The rock mass parameters include internal friction angle Cohesion c, gravity γ, the slope geometric characteristics include height h, slope α, the external load includes earthquake force F e , groundwater pressure Pw The meteorological elements include rainfall R and rainfall duration t.

[0011] Optionally, the building of a multi-layer perceptron neural network model includes the following steps:

[0012] The collected factors affecting slope stability are used to determine the number of neurons in the input layer n, so that each factor corresponds to a neuron to complete the construction of the input layer;

[0013] According to the design of the hidden layer decreasing structure, the number of neurons in the first hidden layer is m1 = 2n, and the second hidden layer is The third layer has m3=n, and the rectified linear unit is selected as the hidden layer neuron activation function. The activation function formula is as follows:

[0014] f(x)=max(0,x)

[0015] The output layer form is determined according to the evaluation task. If the Softmax function is introduced for the classification task, the Softmax function formula is expressed as follows:

[0016]

[0017] Among them, z is the output layer input vector, j represents the category index, and K=3 is the total number of categories;

[0018] Suppose 3 neurons output three state probabilities, namely the probabilities of stable, potentially unstable, and unstable states. If it is to predict the value of the stability coefficient, then set 1 neuron to directly output the weighted sum result.

[0019] Optionally, preprocessing and sample dividing the data collected in the original database includes the following steps:

[0020] For each influencing factor data series, the box plot method is used to screen outliers, and the lower quartile Q1, median Q2, and upper quartile Q3 are calculated. The interquartile range is used to determine outliers, and the mean of the adjacent normal data is used as a substitute. The interquartile range formula is IQR=Q3-Q1;

[0021] A combined feature is constructed based on the original factors, wherein the combined feature includes the product of the internal friction angle and the cohesion Severity and slope height ratio And perform Fourier transform on time series data to extract spectrum features;

[0022] The processed data is divided into a training set, a validation set, and a test set in a ratio of 80%, 10%, and 10%. The training set is used for learning the neural network model parameters, the validation set is used to adjust the model to prevent overfitting, and the test set is used to evaluate the generalization performance of the model.

[0023] Optionally, the specific steps of performing Fourier transform on the time series data to extract spectrum features include:

[0024] The time series data is formed into a discrete time series data set, denoted as x[n], where n represents the discrete time index;

[0025] The discrete Fourier transform formula is used to calculate the time series data. The formula is as follows:

[0026]

[0027] Where N is the total number of data points, k is the frequency index, ranging from 0 to N-1, and j is the imaginary unit. The above formula converts the time domain data x[n] into the frequency domain data X[k], where X[k] represents the amplitude and phase information of different frequency components.

[0028] Calculate the spectrum amplitude using the following formula:

[0029]

[0030] Where Re(X[k]) is the real part of X[k], and Im(X[k]) is the imaginary part of X[k];

[0031] The obtained spectrum amplitude X[k] is analyzed to observe the energy distribution in different frequency bands, and the frequency intervals are set, including low frequency band, medium frequency band, and high frequency band. The sum or average value of the spectrum amplitude in each interval is counted as a feature and input into the neural network model.

[0032] Optionally, the training and optimizing of the adaptive moment estimation algorithm in the neural network model includes setting the parameter vector to be optimized as θ, and the formula for updating the gradient first-order moment estimation during training is as follows:

[0033]

[0034] in, is the gradient of the parameter θ at time t, m t is the gradient first-order moment estimate, and β1 is used to control the first-order moment estimate m t The decay rate of β1 = 0.9;

[0035] The formula for updating the gradient second-order moment estimate is as follows:

[0036]

[0037] in, is the square of the gradient, v t is the gradient second moment estimate, β2 = 0.999;

[0038] Correction of deviation The parameter update follows the following formula:

[0039]

[0040] Where α is the initial learning rate, ∈=10 -8 ;

[0041] The early stopping method monitors the model performance index in the validation set every 10 rounds. Once the following formula is met for 5 consecutive times, the training is terminated immediately and the optimal model is retained. The formula is as follows:

[0042]

[0043] Where k = 5, L val (w) is the validation set loss function, and w is the model weight.

[0044] Optionally, the training and optimizing the neural network model further includes optimizing the learning rate by adopting a learning rate warm-up strategy, and the model adapts to a larger parameter update step size during the warm-up process, including the following steps:

[0045] Set a warm-up period, which is the first N rounds of training, where N is an integer greater than 0. During the warm-up period, the model performs preliminary learning of the data and stable adjustment of parameters;

[0046] Determine the initial learning rate during the warm-up period and set the initial learning rate during the warm-up period to the minimum value β. During the warm-up period, the learning rate η for each round of training increases linearly according to the following formula:

[0047]

[0048] Where n is the current training round number, N is the total number of warm-up rounds, α is the initial learning rate of normal training, α = 0.001.

[0049] Optionally, the value range of the initial learning rate β during the warm-up period is (10 -8 , 10 -4 ), when the model converges too slowly or becomes unstable during training, re-evaluate the warm-up period settings, initial learning rate, and learning rate increment method, and optimize the model training effect by adjusting the values ​​of N, β, and α.

[0050] Optionally, the dynamic update of the neural network model through incremental learning is specifically:

[0051] As new monitoring data and engineering cases accumulate, based on the existing training results, the weights of the previous layer that have tended to converge stably are locked, and only the weights of the l=2 layers close to the output layer are fine-tuned. The original model weight matrix W=[W1,W2,…,W L], where L is the total number of layers, and the gradient is calculated based on the new data And the partial weights are updated by the adaptive moment estimation algorithm.

[0052] Optionally, the early warning of the neural network model is realized by using Bayesian theory, and the steps are as follows:

[0053] The Bayesian formula is used to improve the credibility of the warning. The Bayesian formula is as follows:

[0054]

[0055] Among them, the probability of slope instability predicted by the model is recorded as P(S=unstable|X), X is the input data, and the posterior probability is calculated as the quantified value of the warning credibility based on the historical data statistics prior probability P(S=unstable) and the likelihood function P(X|S=nustable). When the credibility is greater than the threshold of 0.8, the warning is issued.

[0056] In summary, the present application includes at least one of the following beneficial technical effects:

[0057] 1. The present invention provides comprehensive data support for the neural network model through multi-dimensional slope engineering data, and uses data-driven model construction to objectively quantify slope stability. Different operators can obtain the same and reliable results, ensuring the consistency and accuracy of the evaluation. At the same time, the multi-layer perceptron neural network model adopts a hidden layer decreasing structure and a modified linear unit activation function, which can automatically learn the complex features in the input data, effectively capture the nonlinear interaction between factors, and fit the physical and mechanical process of the slope, so that the model can make more accurate judgments on the slope stability under different working conditions.

[0058] 2. The present invention adopts an adaptive moment estimation algorithm combined with an early stopping method to train an optimization model. The adaptive moment estimation algorithm provides an adaptive learning rate adjustment mechanism for each parameter of the neural network to accelerate the convergence of the model to a better solution. The early stopping method monitors the model performance on the validation set regularly, terminates the training in time to prevent overfitting, and retains the optimal model with strong generalization ability. The two work together to ensure that the model training is efficient and reliable, and achieves good performance under limited time and data resources.

[0059] 3. The present invention improves the credibility of early warning by combining historical data with model prediction results through Bayesian theory. By calculating the posterior probability as a quantitative indicator for early warning, it can effectively avoid false alarms or missed alarms compared to traditional early warning methods, accurately realize the advanced prevention and control of slope instability risks, gain sufficient response time for engineering personnel, and effectively ensure the safety of slope projects and surrounding environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1A flow chart of a slope stability evaluation method based on a neural network model is provided in the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] Example

[0063] The present invention proposes a slope stability evaluation method based on a neural network model, such as Figure 1 As shown, the following steps are included:

[0064] Construct an original database, the data source of which is slope engineering case data under different working conditions; build a multi-layer perceptron neural network model to pre-process and sample the data collected in the original database; use an adaptive moment estimation algorithm combined with early stopping method to train and optimize the neural network model; and realize dynamic update and early warning of the neural network model through incremental learning and Bayesian theory. The details are as follows:

[0065] 1. Data Collection and Database Construction

[0066] A survey team was formed to investigate the slopes along the route and collect data. Slope engineering data covering multi-dimensional information was collected to provide comprehensive and representative materials for subsequent model training, so that the model can learn the characteristic laws of slope stability under various actual scenarios. Slope engineering example data included rock and soil parameters, slope geometric characteristics, external loads and meteorological elements. The original database was constructed in multiple dimensions. The equipment used to collect data included high-precision total stations, electronic levels, professional slope meters, seismic monitoring sensors and groundwater pressure monitoring equipment to ensure the accuracy and reliability of data collection.

[0067] Geotechnical parameter collection: In-situ direct shear test was carried out at the selected slope location to measure the internal friction angle. The internal friction angle of a soil slope was measured to be 30° by combining direct shear test and triaxial test. The cohesion c was obtained by unconfined compressive strength test in geotechnical test. The measured value of the cohesion of the slope was 20 kPa. The rock and soil samples were collected at different depths and positions by ring knife method. The specific gravity γ was measured and the value was 18 kN / m 3 ;

[0068] Slope geometric feature collection: Use a total station to start from the foot of the slope, set measurement points at a certain distance along the contour line of the slope, measure the height h, and measure the height of the high slope up to 50m. Use a slope meter to measure multiple times at different parts of the slope. Each slope surface is measured at least 5 points, and the average value is taken to determine the slope α. The average slope of a rock slope is measured to be 45°;

[0069] External load data collection: Estimate the earthquake force F based on the local earthquake monitoring network data, earthquake history data and seismic parameter zoning map, combined with the requirements of highway seismic fortification. e It is estimated that the maximum horizontal seismic force that the slope may withstand under the peak acceleration of the earthquake in this area is 100kN. Groundwater pressure monitoring equipment is installed in different locations of the slope to monitor the groundwater level and pressure for a long time. The groundwater pressure P of the slope is measured. w 30kPa;

[0070] Meteorological data acquisition: The historical data of rainfall R in the area along the route were obtained through the local meteorological station. The maximum rainfall in a certain period of the past year reached 500 mm, and the rainfall duration t was recorded. The longest continuous rainfall was 7 days.

[0071] The various types of data collected were summarized to build an original database, covering slope instance information under different geological conditions and working conditions. The data were preliminarily sorted and labeled, including information such as the time, location, and measurement method of data collection to ensure the traceability of the data. At the same time, for some obviously erroneous or abnormal data, on-site review and re-measurement were carried out. Finally, a total of 500 sets of valid data were collected.

[0072] The present invention provides comprehensive data support for the neural network model through multi-dimensional slope engineering data, and uses data-driven model construction to objectively quantify slope stability. Different operators can obtain the same and reliable results, ensuring the consistency and accuracy of the evaluation.

[0073] 2. Neural Network Model Construction

[0074] The number of neurons in the input layer is determined based on the collected factors affecting slope stability, including 3 rock and soil parameters, 2 slope geometric characteristics, 2 external loads, and 2 meteorological elements, a total of 9 factors. Therefore, 9 neurons are set in the input layer, and each neuron corresponds to a specific factor to ensure that the data can be accurately input into the model;

[0075] According to the design of the hidden layer decreasing structure: the number of neurons in the first hidden layer is m1=2×9=18. The first layer is used to extract the complex features of the input data, and the input information is preliminarily processed and combined through a large number of neurons. The number of neurons is rounded to 14. The second time is used to further screen and integrate the features extracted from the first layer. Reducing the number of neurons can avoid overfitting of the model while retaining key feature information. The third layer has m3=9. The third time, the features processed by the first two layers are finally refined to make them more in line with the requirements of slope stability evaluation. The corrected linear unit is selected as the hidden layer neuron activation function. The activation function formula is expressed as follows:

[0076] f(x)=max(0,x)

[0077] When the input x is greater than 0, the output is equal to the input x. At this time, the neuron is activated and the signal is transmitted smoothly, which can effectively avoid the gradient vanishing problem, accelerate the gradient propagation during model training, and allow the model to quickly learn complex features; when the input x is less than or equal to 0, the output is 0, which means that the neuron is temporarily in an unactivated state, which has the effect of sparse activation and reduces the amount of calculation.

[0078] Determine the output layer form according to the evaluation task: This project needs to judge the stability of the slope, adopt the classification task mode, introduce the Softmax function, and the Softmax function formula is as follows:

[0079]

[0080] Among them, z is the input vector of the output layer, j represents the category index, K = 3 is the total number of categories, and the output layer has 3 neurons, corresponding to the three state probabilities of stability, potential instability, and instability. The Softmax function can convert the output value of the output layer neurons into the distribution representing the probability of each classification, making the model output easy to understand and explain. If it is to predict the value of the stability coefficient, one neuron is set to directly output the weighted sum result, and directly give the quantitative numerical index of slope stability.

[0081] The multilayer perceptron neural network adopts a hidden layer decreasing structure and a modified linear unit activation function. It can automatically learn the complex features in the input data, effectively capture the nonlinear interaction between factors, fit the physical and mechanical process of the slope, overcome the problem of over-simplification of traditional mechanical analysis, and make the model more accurate in judging the stability of the slope under different working conditions.

[0082] 3. Data Preprocessing and Sample Division

[0083] Data preprocessing: The box plot method is used to screen outliers for the data series of each influencing factor. Taking rainfall data as an example, the lower quartile Q1 = 100 mm, the median Q2 = 200 mm, the upper quartile Q3 = 300 mm, and the interquartile range IQR = 300-100 = 200 mm are calculated. Data greater than Q3 + 1.5 × IQR = 600 mm and less than Q1 + 1.5 × IQR = -200 mm are judged as outliers (there is actually no rainfall less than 0, this is only an example of calculation). For outliers, the cause of the outliers is determined by consulting meteorological data and on-site investigation. If the outliers are caused by a failure of the measuring equipment, the mean of the adjacent normal data is used as a substitute; if the outliers are caused by special meteorological events, they are explained in the data annotation and the data are retained so that the model can learn the slope stability characteristics under special circumstances;

[0084] Structural combination characteristics: Calculation of the product of internal friction angle and cohesion The product term reflects the strength of the synergistic effect of friction and adhesion in the anti-sliding ability of rock and soil. The value of this term for the soil slope mentioned above is 30°×20kPa=600. This ratio reflects the weight borne by the slope per unit height. The ratio of this high slope is The combined features can capture the interaction between factors, and then Fourier transform is performed on the time series data to extract the spectrum features. In this embodiment, the time series data takes rainfall data as an example. The rainfall time series data of one year is discretized according to a fixed sampling interval to form a discrete time series data set, which is recorded as x[n], where n represents the discrete time index. The spectrum data is calculated using the discrete Fourier transform formula, which is as follows:

[0085]

[0086] Where N is the total number of data points, k is the frequency index, ranging from 0 to N-1, and j is the imaginary unit. The above formula is used to convert the time domain data x[n] into frequency domain data X[k], where X[k] represents the amplitude and phase information of different frequency components, and calculate the spectrum amplitude. The calculation formula is as follows:

[0087]

[0088] Where Re(X[k]) is the real part of X[k], and Im(X[k]) is the imaginary part of X[k]. The obtained spectrum amplitude X[k] is analyzed to observe the energy distribution in different frequency bands, and analyze the energy distribution of low-frequency bands, medium-frequency bands, and high-frequency bands. The low-frequency band is set to seasonal changes, such as 0-0.1 cycles / day, the medium-frequency band is monthly changes, 0.1-0.5 cycles / day, and the high-frequency band is set to short-term rainfall fluctuations, such as 0.5-1 cycles / day. The sum or average of the spectrum amplitudes in each interval is counted as features. These features will be input into the neural network model to help it capture the slope stability-related information contained in the time series data, such as the cumulative effect of rainfall periodicity changes on the long-term stability of the slope;

[0089] Sample division: The processed data are divided into 400 training sets, 50 validation sets, and 50 test sets in the ratio of 80%, 10%, and 10%. During the division process, the stratified sampling method is adopted to ensure that each subset contains slope data under different geological conditions and working conditions, so as to ensure that the model can be exposed to comprehensive samples during the training, validation and testing process, and improve the generalization ability of the model. The training set is used for neural network model parameter learning, so that the model can adjust its own weights and other parameters based on a large amount of sample data to fit the slope stability law. The validation set is used to adjust the model to prevent overfitting. During the model training process, the model hyperparameters are adjusted in time through the performance feedback on the validation set to ensure that the model has good generalization ability. The test set is used to evaluate the generalization performance of the model. After the model training is completed, the test set is used to test the model's prediction ability for unseen data to ensure the reliability of the model in actual application scenarios.

[0090] In the data preprocessing stage, on the one hand, the box plot method is used to screen and reasonably handle outliers, combined with the construction of combined features and the implementation of Fourier transform on time series data to extract spectral features, enrich the model input feature dimensions so that the model can fully learn the effective rules in the data. On the other hand, the training set, validation set, and test set are divided to ensure the model training effect and generalization performance.

[0091] 4. Model Training and Optimization

[0092] The steps of training the optimization model using the adaptive moment estimation algorithm combined with the early stopping method are as follows:

[0093] Adaptive moment estimation algorithm: Set the parameter vector to be optimized as θ, which contains all the parameters that need to be learned and adjusted in the neural network, such as the connection weights and thresholds of neurons in each layer of the neural network. During training, the initial learning rate α=0.001, ∈=10 -8 , taking a certain weight parameter as an example, assuming that its gradient first-order moment estimate m t Initially it is 0. At the first iteration, according to the formula β1=0.9, combined with the current gradient Update m t ,here is the gradient of the parameter θ at time t, that is, the function J(θ t ) The partial derivative vector of θ reflects the changing trend of the parameter θ at the current time t. β1 is used to control the decay rate of the first-order moment estimate, so that the algorithm pays more attention to the recent gradient information. In the first iteration, the current gradient is calculated Then m1=0.9×0+(1-0.9)×0.1=0.01; Similarly, through the formula β2 = 0.999, update the gradient second-order moment estimate v t , β2 controls the decay rate of the second-order moment estimate, which is used to measure the square of the gradient and reflect the change in the step size of the parameter update. In the first iteration, v1 = 0.999 × 0 + (1-0.999) × (0.1) 2 =0.0001; then correct the deviation to get In the first iteration, Finally, according to the formula: Update the weight parameters. In the first iteration, set θ1 = 0.5, then

[0094] Early stopping method: monitor the model performance indicators in the validation set every 10 rounds, using accuracy as the indicator. During the initial training, the accuracy of the model in the validation set gradually increases. When the 50th round of training is reached, the accuracy of the validation set stops improving for 5 consecutive times, satisfying Where k = 5, L val (w) is 1-accuracy, w is the model weight, the training is terminated immediately, and the current optimal model is retained.

[0095] The adaptive moment estimation algorithm is combined with the early stopping method to train the optimization model. The adaptive moment estimation algorithm provides an adaptive learning rate adjustment mechanism for each parameter of the neural network model, and dynamically updates the parameters based on the gradient first-order moment and second-order moment estimation. Different parameters are optimized at different speeds according to their own gradient conditions, thereby accelerating the model to converge to a better solution. The early stopping method monitors the model performance on the validation set regularly, terminates the training in time to prevent overfitting, and retains the optimal model with strong generalization ability. The two work together to ensure that the model training is efficient and reliable, and achieves good performance with limited time and data resources.

[0096] Training and optimizing the neural network model also includes optimizing the learning rate using a learning rate warm-up strategy. At the beginning of model training, since the parameters are randomly initialized, if the learning rate is large, the step size of the parameter update will be too large. The learning rate warm-up strategy allows the model to slowly learn the basic characteristics of the data in the initial stage, so that the model gradually adapts to a larger parameter update step size during the warm-up process. The learning rate warm-up strategy includes the following steps:

[0097] Set a warm-up period, which is the first N rounds of training. N is an integer greater than 0. During the warm-up period, the model performs preliminary learning of the data and stabilizes the parameters.

[0098] Determine the initial learning rate during the warm-up period and set the initial learning rate during the warm-up period to the minimum value β, β is 10 -6 During the warm-up period, the learning rate η is linearly increased according to the following formula in each round of training:

[0099]

[0100] Where n is the current training round number, and 1≤n≤N, N is the total round number of the warm-up period, α is the initial learning rate of normal training, α=0.001, and the total round number N of the warm-up period is adaptively adjusted according to different slope engineering data characteristics, model complexity and training resource factors to achieve the best training effect;

[0101] In addition, the value range of the initial learning rate β during the warm-up period is (10 -8 , 10 -4 ), the specific value is optimized based on the sensitivity of the model to the initial parameter update step and the fluctuation of the data to ensure that the model can stably learn the data features in the early stage of training. When the model converges too slowly or is unstable during training, re-evaluate the warm-up period settings, initial learning rate, and learning rate increment method, and optimize the model training effect by adjusting the values ​​of N, β, and α. In addition, the learning rate increment formula can be changed to a nonlinear increment, such as an exponential increment. To optimize the model training effect.

[0102] The present invention uses a learning rate warm-up strategy to enable the model to slowly learn the basic characteristics of the data in the initial stage, avoiding the model from deviating from the optimal solution due to excessive parameter update step size, thereby improving the stability of model training. As the training proceeds, the learning rate gradually increases, which accelerates the convergence speed of the model and improves the training efficiency, thereby ensuring that the training time and consumption of computing resources are reduced while ensuring the accuracy of the model, while enhancing the adaptability and generalization ability of the model.

[0103] V. Dynamic Update and Warning

[0104] Dynamic update: After the project is completed, monitoring equipment is installed at key locations on the slope to continuously collect data. For example, when the slope detects a significant increase in rainfall and a rise in groundwater levels in the subsequent rainy season, the new monitoring data and new engineering maintenance cases are accumulated to 50 sets of new data, and incremental learning is started. Based on the existing training results, the weights of the previous layer that have tended to converge stably are locked, and only the weights of the l=2 layer close to the output layer are fine-tuned, and the gradient is calculated based on the new data. And refresh these weights according to the above weight updating steps through the adaptive moment estimation algorithm to make the model adapt to the new working conditions;

[0105] By using an incremental learning strategy, as new monitoring data and engineering cases accumulate, the key weights close to the output layer are fine-tuned and the model is quickly updated based on existing training results, so that it can dynamically adapt to changes in slope conditions, avoid repeated large-scale training, save computing resources, and ensure that the model always fits the actual situation and continuously outputs accurate stability evaluations.

[0106] Early warning: Bayesian theory is used to achieve early warning. According to historical data statistics, the probability of instability of the slope in this area under similar working conditions in the past is 0.2. The model inputs new monitoring data and predicts the probability of slope instability P(S=unstable|X). Combined with the likelihood function P(X|S=unstable), the likelihood function estimates the frequency of unstable slopes in historical data similar to the input data, and then uses the Bayesian formula The posterior probability is calculated as the quantified value of the warning credibility. When the credibility is greater than the threshold of 0.8, an early warning is immediately issued to the highway maintenance department to notify them to take emergency measures.

[0107] By combining Bayesian theory with historical data and model prediction results, the credibility of early warnings can be improved. By calculating the posterior probability as a quantitative indicator for early warning, false alarms or missed alarms can be effectively avoided, and proactive prevention and control of slope instability risks can be achieved. This will buy sufficient response time for engineering personnel, ensure the safety of slope projects and surrounding environments, and reduce the risk of casualties and property losses caused by geological disasters such as landslides.

[0108] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A slope stability evaluation method based on a neural network model, characterized in that: The following steps are involved: Constructing an original database, wherein the data source of the original database is slope engineering example data under different working conditions; Building a multi-layer perceptron neural network model to pre-process and divide samples of the data collected from the original database; Adopting adaptive moment estimation algorithm combined with early stopping method to train and optimize the neural network model; The dynamic update and early warning of the neural network model are achieved through incremental learning and Bayesian theory.

2. A slope stability evaluation method based on a neural network model according to claim 1, characterized in that: The slope engineering example data includes rock and soil parameters, slope geometric characteristics, external loads and meteorological elements, and the original database is constructed in multiple dimensions; The rock mass parameters include internal friction angle Cohesion c, gravity γ, the slope geometric characteristics include height h, slope α, the external load includes earthquake force F e , groundwater pressure P w The meteorological elements include rainfall R and rainfall duration t.

3. A slope stability evaluation method based on a neural network model according to claim 1, characterized in that: The construction of the multi-layer perceptron neural network model includes the following steps: The collected factors affecting slope stability are used to determine the number of neurons in the input layer n, so that each factor corresponds to a neuron to complete the construction of the input layer; According to the design of the hidden layer decreasing structure, the number of neurons in the first hidden layer is m1 = 2n, and the second hidden layer is The third layer has m3=n, and the rectified linear unit is selected as the hidden layer neuron activation function. The activation function formula is as follows: f(x)=max(0,x) The output layer form is determined according to the evaluation task. If the Softmax function is introduced for the classification task, the Softmax function formula is expressed as follows: Among them, z is the output layer input vector, j represents the category index, and K=3 is the total number of categories; Suppose 3 neurons output three state probabilities, namely the probabilities of stable, potentially unstable, and unstable states. If it is to predict the value of the stability coefficient, then set 1 neuron to directly output the weighted sum result.

4. The slope stability evaluation method based on a neural network model according to claim 1 is characterized in that: Preprocessing and sample division of the data collected in the original database includes the following steps: For each influencing factor data series, the box plot method is used to screen outliers, and the lower quartile Q1, median Q2, and upper quartile Q3 are calculated. The interquartile range is used to determine outliers, and the mean of the adjacent normal data is used as a substitute. The interquartile range formula is IQR=Q3-Q1; A combined feature is constructed based on the original factors, wherein the combined feature includes the product of the internal friction angle and the cohesion Severity and slope height ratio And perform Fourier transform on time series data to extract spectrum features; The processed data is divided into a training set, a validation set, and a test set in a ratio of 80%, 10%, and 10%. The training set is used for learning the neural network model parameters, the validation set is used to adjust the model to prevent overfitting, and the test set is used to evaluate the generalization performance of the model.

5. A slope stability evaluation method based on a neural network model according to claim 4, characterized in that: The specific steps of performing Fourier transform on the time series data to extract spectrum features include: The time series data is formed into a discrete time series data set, denoted as x[n], where n represents the discrete time index; The discrete Fourier transform formula is used to calculate the time series data. The formula is as follows: Where N is the total number of data points, k is the frequency index, ranging from 0 to N-1, and j is the imaginary unit. The above formula converts the time domain data x[n] into the frequency domain data X[k], where X[k] represents the amplitude and phase information of different frequency components. Calculate the spectrum amplitude using the following formula: Where Re(X[k]) is the real part of X[k], and Im(X[k]) is the imaginary part of X[k]; The obtained spectrum amplitude X[k] is analyzed to observe the energy distribution in different frequency bands, and the frequency intervals are set, including low frequency band, medium frequency band, and high frequency band. The sum or average value of the spectrum amplitude in each interval is counted as a feature and input into the neural network model.

6. The slope stability evaluation method based on a neural network model according to claim 1 is characterized in that: The training and optimization of the adaptive moment estimation algorithm in the neural network model includes setting the parameter vector to be optimized as θ, and the formula for updating the gradient first-order moment estimation during training is as follows: in, is the gradient of the parameter θ at time t, m t is the gradient first-order moment estimate, and β1 is used to control the first-order moment estimate m t The decay rate of β1 = 0.9; The formula for updating the gradient second-order moment estimate is as follows: in, is the square of the gradient, v t is the gradient second moment estimate, β2 = 0.999; Correction of deviation The parameter update follows the following formula: Where α is the initial learning rate, ∈=10 -8 ; The early stopping method monitors the model performance index in the validation set every 10 rounds. Once the following formula is met for 5 consecutive times, the training is terminated immediately and the optimal model is retained. The formula is as follows: Where k = 5, L val (w) is the validation set loss function, and w is the model weight.

7. The slope stability evaluation method based on a neural network model according to claim 1 is characterized in that: The training and optimizing of the neural network model further includes optimizing the learning rate by using a learning rate warm-up strategy, wherein the model adapts to a larger parameter update step size during the warm-up process, including the following steps: Set a warm-up period, which is the first N rounds of training, where N is an integer greater than 0. During the warm-up period, the model performs preliminary learning of the data and stable adjustment of parameters; Determine the initial learning rate during the warm-up period and set the initial learning rate during the warm-up period to the minimum value β. During the warm-up period, the learning rate η for each round of training increases linearly according to the following formula: Where n is the current training round number, N is the total number of warm-up rounds, α is the initial learning rate of normal training, α = 0.

001.

8. The slope stability evaluation method based on a neural network model according to claim 1 is characterized in that: The value range of the initial learning rate β during the warm-up period is (10 -8 ,10 -4 ), when the model converges too slowly or becomes unstable during training, re-evaluate the warm-up period settings, initial learning rate, and learning rate increment method, and optimize the model training effect by adjusting the values ​​of N, β, and α.

9. The slope stability evaluation method based on a neural network model according to claim 1 is characterized in that: The dynamic update of the neural network model through incremental learning is specifically: As new monitoring data and engineering cases accumulate, based on the existing training results, the weights of the previous layer that have tended to converge stably are locked, and only the weights of the l=2 layers close to the output layer are fine-tuned. The original model weight matrix W=[W1,W2,…,W L ], where L is the total number of layers, and the gradient is calculated based on the new data And the partial weights are updated by the adaptive moment estimation algorithm.

10. The slope stability evaluation method based on a neural network model according to claim 1, characterized in that: The steps of implementing the early warning of the neural network model through Bayesian theory are as follows: The Bayesian formula is used to improve the credibility of the warning. The Bayesian formula is as follows: Among them, the probability of slope instability predicted by the model is recorded as P(S=unstable|X), X is the input data, and the posterior probability is calculated as the quantified value of the warning credibility based on the historical data statistics prior probability P(S=unstable) and the likelihood function P(X|S=unstable). When the credibility is greater than the threshold of 0.8, the warning is issued.

Images