A slope monitoring method and system based on deep learning

Abstract

The present invention discloses a slope monitoring method and system based on deep learning, including S1: obtaining original slope displacement data, and performing seasonal division and resampling on the data to obtain pre-processed slope displacement data; S2: seasonal decomposition of the pre-processed slope displacement data to extract seasonal features; S3: constructing a slope monitoring network integrating seasonal features, extracting slope feature vectors, and inputting the slope feature vectors into a fully connected layer to obtain slope displacement data prediction values; S4: setting the optimization target of the slope monitoring network integrating seasonal features; S5: obtaining the parameters of the slope monitoring network integrating seasonal features, and optimizing the parameters of the slope monitoring network integrating seasonal features based on an adaptive learning rate optimization algorithm. The present invention solves the problem of insufficient accuracy of slope monitoring due to ignoring the influence of seasonal changes.

Description

Technical Field

[0001] The present invention relates to the field of slope engineering technology, and in particular to a slope monitoring method and system based on deep learning. Background Art

[0002] Landslides are natural disasters that threaten people's lives and property. After the slope is constructed, its geological conditions are still in a dynamic state of change. It is necessary to monitor the slope and take remedial measures in time to ensure the safety of the slope.

[0003] Traditional deep learning-based slope monitoring methods rely on sensor networks and data processing techniques, and although they have made significant progress in improving accuracy and automation, they still fail to consider the impact of seasonal changes on slope stability.

[0004] Seasonal changes have an important impact on slope monitoring, and slope displacement data show different change characteristics in different seasons. In the rainy season, soil water absorption increases, shear strength decreases, and slope displacement shows a clear upward trend; in winter, the soil freezes and thaws, and the slope displacement data fluctuates with freezing and thawing; in summer, the vegetation is lush, the soil stability is relatively improved, and the slope displacement data shows a relatively stable state. Seasonal changes have a significant impact on the characteristics of slope displacement data, and comprehensive considerations are needed to more accurately evaluate the stability of the slope. Summary of the invention

[0005] In view of this, the present invention provides a slope monitoring method and system based on deep learning, aiming to solve the problem of insufficient slope monitoring accuracy caused by ignoring the influence of seasonal changes in traditional deep learning methods. The present invention models seasonal characteristics, explicitly considers the influence of seasonal characteristics on slope monitoring, and integrates them into the time series slope displacement data, thereby improving the accuracy of slope monitoring.

[0006] To achieve the above object, the present invention provides a slope monitoring method based on deep learning, comprising the following steps:

[0007] S1: Obtain the original slope displacement data, divide it into seasons and resample it to obtain the preprocessed slope displacement data; specifically, it includes:

[0008] S11: dividing the original data of slope displacement by season to obtain data sets of four seasons: spring, summer, autumn and winter, wherein the original data of slope displacement is in days;

[0009] S12: resampling the original slope displacement data of each season, the sampling method is: taking the date corresponding to the first data of each season as the sampling start date, taking one data every 7 days until 13 data are taken; then pushing the sampling start date back one day, and performing the next round of sampling; repeating the above process 7 times, and finally obtaining 7 groups of the pre-processed slope displacement data;

[0010] S2: performing seasonal decomposition on the preprocessed slope displacement data to extract seasonal characteristics;

[0011] S21: Calculate the trend component using a local smoothing method, the calculation method is:

[0012] ;

[0013] in, for The trend component of the moment, is the target time, i is the target time index, n is the total number of data points, For data points For the weight of the target data point, for The preprocessed slope displacement data at time, is another moment, j is the index of another moment, and the weight is calculated as follows:

[0014] ;

[0015] Where, e is the natural constant, h is the bandwidth, is the kth moment, k is the counting index;

[0016] S22: Calculate the seasonal component, the calculation method is:

[0017] ;

[0018] in, For at time point The seasonal components of For at time point Preprocessed slope displacement data at ;

[0019] S23: Calculate the residual component, the calculation method is:

[0020] ;

[0021] in, For at time point The residual component at ;

[0022] The trend component , the seasonal components , the residual component ,by The splicing of forms is the seasonal feature of this time point; among them, For splicing operation;

[0023] S3: construct a slope monitoring network integrating seasonal features, extract seasonal feature vectors and time series feature vectors respectively according to the seasonal features and the preprocessed slope displacement data, and then perform feature fusion to obtain a slope feature vector; finally, input the slope feature vector into a fully connected layer to obtain a predicted value of the slope displacement data;

[0024] S4: setting the optimization target of the slope monitoring network integrating seasonal characteristics;

[0025] S5: Obtain parameters of the slope monitoring network integrating seasonal characteristics, and optimize the parameters of the slope monitoring network integrating seasonal characteristics based on an adaptive learning rate optimization algorithm.

[0026] For slope displacement data, the displacement phenomenon becomes obvious as the time scale increases. At too fine a time scale, such as daily or hourly, the slope displacement will basically not change. Using data directly for model training on such a fine time scale will lead to a waste of information, especially when extracting seasonal variation patterns. The present invention resamples the slope data in units of weeks, captures the evolution of slope displacement at more reasonable time intervals, and improves data utilization efficiency.

[0027] Compared with the traditional method of directly using daily data, this sampling method realizes multiple sampling within the same time period, which is equivalent to introducing more data points within the same time range, expanding the data that could only be used for one training to data that can be used for 7 training times, providing more learning opportunities for deep learning models.

[0028] In addition, this resampling method also takes into account the periodicity of each season, making it easier for the model to learn the slope displacement patterns in different seasons, thereby improving the ability to perceive seasonal changes.

[0029] In the present invention, the current time point is used Subtract another time point , is to calculate the time distance between two time points, this distance will affect the weight The size of is a time point in the data. j can be understood as a loop variable that traverses all data points. The bandwidth h is used to control the smoothness of local smoothing.

[0030] In the present invention, the larger the bandwidth h is, the more data points are used to calculate the trend component, which can better reflect the overall change of the time series; the smaller the bandwidth h is, the fewer data points are used to calculate the trend component, which can better capture the subtle fluctuations of the time series. Local smoothing can adjust the weight of each data point according to the time distance. The closer the data point is, the more important it is, and the farther the data point is, the less important it is, so as to better reflect the trend change of the time series.

[0031] In the present invention, this step decomposes a time series into three components: trend, season and residual. The trend component represents the long-term change of the time series, the season component represents the periodic change of the time series, and the residual component represents the random fluctuation of the time series. By calculating the three components, it is helpful to decompose the signals from different sources in the slope displacement data more finely and comprehensively, increase the interpretability of the model, and at the same time, provide a more comprehensive and rich input for the subsequent slope monitoring network, which helps the subsequent network to more accurately capture and learn the complex change pattern of displacement.

[0032] Furthermore, the S3 step constructs a slope monitoring network integrating seasonal features, extracts seasonal feature vectors and time series feature vectors according to the seasonal features and the preprocessed slope displacement data, and then performs feature fusion to obtain a slope feature vector; finally, the slope feature vector is input into a fully connected layer to obtain a predicted value of the slope displacement data, including:

[0033] S31: Construct a slope monitoring network integrating seasonal features, which consists of 2 GRUs and a feature fusion part;

[0034] S32: Input the seasonal features of each time point in each season into the first GRU, and take the last hidden state of the first GRU output as the seasonal feature vector, recorded as , GRU is calculated as:

[0035] ;

[0036] in, is the update gate at the current moment, t is the current moment, is the sigmoid function, To update the gate weight matrix, is the hidden state at the previous moment, is the preprocessed slope displacement data at the current moment, is the reset gate at the current moment, To reset the gate weight matrix, is the candidate hidden state at the current moment, tanh() is the tanh function, is the hyperbolic sine weight matrix, is element-wise multiplication;

[0037] S33: For the preprocessed slope displacement data, the second GRU is used to extract the time series feature vector, and the calculation method is:

[0038] ;

[0039] in, is the time series feature vector, For the GRU network, is the preprocessed slope displacement data;

[0040] S34: The seasonal feature vector is concatenated with the time series feature vector and input into a fully connected layer. The sigmoid function is used for activation and the seasonal weight is constructed and recorded as , the time series weight is recorded as , calculated as:

[0041] ;

[0042] in, is the weight matrix, is the bias vector;

[0043] S35: dot-multiply the seasonal weight with the seasonal feature vector, dot-multiply the temporal weight with the temporal feature vector, and then sum them up to perform feature fusion to obtain the slope feature vector. The calculation method is:

[0044] ;

[0045] in, is the slope characteristic vector, is element-wise multiplication;

[0046] S36: Input the slope feature vector into the fully connected layer to obtain the predicted value of the slope displacement data. The calculation method is:

[0047] ;

[0048] in, is the predicted value of slope displacement data, c is the moment count index, is the weight matrix, is the bias vector.

[0049] In the present invention, the sigmoid function is used as the activation function for seasonal weight calculation, so that the seasonal weight can more flexibly model the complex and nonlinear relationship between seasonal characteristics and temporal characteristics; at the same time, the model can have the ability to flexibly adjust the relative contributions of seasonal characteristics and temporal characteristics in the model output when calculating the seasonal weight; and the use of a gating mechanism helps the model to more accurately control the flow of information and reduce interference from irrelevant information, thereby improving the network's effective use of seasonal characteristics and temporal characteristics.

[0050] The advantage of the seasonal weight calculation method lies in its dynamic adaptability. The model can automatically adjust the weights according to the different situations of the input seasonal characteristics and time series characteristics, thus realizing the adaptive fusion of seasonal characteristics and time series characteristics.

[0051] The feature fusion is performed by element-by-element multiplication, which is relatively simple and efficient. This helps to reduce the computational complexity of the model, while retaining the weight information of the seasonal feature vector and the time series feature vector, improving the interpretability of the model.

[0052] Through feature fusion calculation, the model can perform personalized learning of seasonal features and time series features at each time point, and the weights in feature fusion calculation and It is dynamically adaptive. By considering the weight relationship under different seasons and time series conditions, it helps the model to more accurately fuse information from different features. At the same time, the element-by-element multiplication method is used, and the feature fusion process is relatively simple, which avoids excessive information loss and helps to retain the model's sensitivity to seasonal and time series information during the feature fusion process, thereby improving the accuracy of the prediction.

[0053] Furthermore, in step S4, setting the optimization target of the slope monitoring network integrating seasonal characteristics includes:

[0054] S41: The optimization target of the slope monitoring network integrating seasonal characteristics consists of two parts, the first part optimizes the ability of the slope monitoring network to model seasonal characteristics, and the second part optimizes the ability of the slope monitoring network to model temporal characteristics;

[0055] The first part optimizes the ability of the slope monitoring network to model seasonal characteristics, that is, the loss function in the first part is calculated as;

[0056] ;

[0057] Where C is the sequence length, is the actual value of slope displacement, is the natural logarithm, is the hyperbolic cosine function;

[0058] S42: The second part optimizes the ability of the slope monitoring network modeling time series characteristics, that is, the calculation method of the loss function in the second part is:

[0059] ;

[0060] The optimization objective of the slope monitoring network integrating seasonal characteristics is expressed as:

[0061] ;

[0062] in, is the optimization objective of the slope monitoring network integrating seasonal characteristics.

[0063] In the present invention, the Log-Cosh loss function measures the prediction error by the logarithmic function of the hyperbolic cosine, which can prevent the gradient explosion problem when the prediction error is large, and can also avoid the gradient vanishing problem when the prediction error is close to zero. The Log-Cosh loss function is a smooth and continuous function. This smoothness helps the optimization algorithm to converge better, making the training process more stable and improving the training efficiency of the network. At the same time, since the logarithmic function of the hyperbolic cosine grows slowly when the prediction error is large, it has less impact on outliers than MSE.

[0064] In the present invention, the optimization target of the slope monitoring network integrating seasonal characteristics adopts a combination of two loss functions, Log-Cosh and mean square error (MSE), which combines their respective advantages. Log-Cosh has good performance in robustness to outliers and gradient stability, while MSE has advantages in accurate measurement of small errors. The use of two different forms of loss functions makes the optimization target of the slope monitoring network integrating seasonal characteristics more flexible and adaptable in training. At different stages of model training, the influence of the loss function can be better adjusted to adapt to different training requirements and data characteristics, thereby improving the generalization performance of the model.

[0065] Furthermore, in step S5, the parameters of the slope monitoring network integrating seasonal characteristics are optimized based on an adaptive learning rate optimization algorithm, including:

[0066] S51: The adaptive learning rate optimization algorithm is used to optimize the parameters of the slope monitoring network integrating seasonal characteristics, and the calculation method is:

[0067] ;

[0068] in, is the first-order moment estimate, is the second-order moment estimate, is the first-order moment estimate of the previous iteration, is the gradient of the current iteration, is the second-order moment estimate of the previous iteration, and are the first and second attenuation coefficients, respectively. and is the deviation correction for the first and second order moments, are the parameters of the slope monitoring network integrating seasonal characteristics after iteration, that is, the parameters of the slope monitoring network integrating seasonal characteristics optimized based on the adaptive learning rate optimization algorithm, is the parameter of the slope monitoring network integrating seasonal characteristics, is the learning rate, a is the number of iterations, is a constant, for a to the power of, for a to the power of;

[0069] S52: After completing the training of the parameters of the slope monitoring network integrating seasonal characteristics, the slope monitoring data of the road section to be predicted is input, and the slope monitoring network integrating seasonal characteristics outputs the slope displacement at the next moment.

[0070] The present invention also provides a slope monitoring system based on deep learning, comprising:

[0071] Data acquisition and preprocessing module: obtain the original slope displacement data, divide it into seasons and resample it to obtain the preprocessed slope displacement data;

[0072] Seasonal feature extraction module: seasonal decomposition of preprocessed slope displacement data to extract seasonal features;

[0073] Network construction module: construct a slope monitoring network integrating seasonal features, extract seasonal feature vectors and time series feature vectors according to the seasonal features and the preprocessed slope displacement data, and then perform feature fusion to obtain slope feature vectors; finally, input the slope feature vectors into the fully connected layer to obtain the predicted value of slope displacement data;

[0074] Network training module: setting the optimization target of the slope monitoring network integrating seasonal characteristics;

[0075] Network optimization module: Optimize the parameters of the slope monitoring network integrating seasonal characteristics based on the adaptive learning rate optimization algorithm;

[0076] To realize the above-mentioned slope monitoring method based on deep learning.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] (1) In terms of the overall scheme, the present invention fully considers the potential impact of seasonal characteristics on slope displacement changes, organically integrates seasonal characteristics with time-series slope displacement data, realizes more sensitive modeling of seasonal changes, improves the model's ability to capture seasonal changes in slope displacement data, and thus improves the accuracy of slope displacement prediction;

[0079] (2) In terms of data processing, this paper proposes a new data resampling method, which introduces more data points within the same time range, providing more learning opportunities for the deep learning model. At the same time, this sampling method also takes into account the obviousness of the slope displacement, making it easier for the model to learn the slope displacement patterns in different seasons, thereby improving the perception of seasonal changes in slope displacement;

[0080] (3) In terms of algorithm improvement, the present invention proposes a method for extracting and fusing seasonal features of slope displacement data, combines seasonal decomposition with deep learning to extract seasonal features, and proposes seasonal weights to adaptively fuse the seasonal features and time series features of slope displacement data, ultimately improving the accuracy of slope displacement prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 A schematic diagram of the process of the slope monitoring method based on deep learning provided by the present invention;

[0082] Figure 2 A schematic flow chart of a resampling method is provided for the present invention;

[0083] Figure 3 A schematic diagram of an algorithm flow of a slope monitoring method based on deep learning is provided for the present invention. DETAILED DESCRIPTION

[0084] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention. Example

[0085] Embodiment 1:

[0086] A slope monitoring method based on deep learning, such as Figure 1 As shown, the following steps are included:

[0087] S1: Obtain the original slope displacement data, divide it into seasons and resample it to obtain the preprocessed slope displacement data;

[0088] S11: The original data of slope displacement are divided by season to obtain data sets of four seasons: spring, summer, autumn and winter. The original data of slope displacement is in days.

[0089] S12: resampling the original slope displacement data of each season, the sampling method is: taking the date corresponding to the first data of each season as the sampling start date, taking one data every 7 days until 13 data are taken; then pushing the sampling start date back one day, and performing the next round of sampling; repeating the above process 7 times, finally obtaining 7 groups of the pre-processed slope displacement data, such as Figure 2 shown.

[0090] For slope displacement data, the displacement phenomenon becomes obvious as the time scale increases. At too fine a time scale, such as daily or hourly, the slope displacement will basically not change. Using data directly for model training on such a fine time scale will lead to a waste of information, especially when extracting seasonal variation patterns. The present invention resamples the slope data in units of weeks, captures the evolution of slope displacement at more reasonable time intervals, and improves data utilization efficiency.

[0091] Compared with the traditional method of directly using daily data, this sampling method realizes multiple sampling within the same time period, which is equivalent to introducing more data points within the same time range, expanding the data that could only be used for one training to data that can be used for 7 training times, providing more learning opportunities for deep learning models.

[0092] In addition, this resampling method also takes into account the periodicity of each season, making it easier for the model to learn the slope displacement patterns in different seasons, thereby improving the ability to perceive seasonal changes.

[0093] For example, suppose there is a period of spring (March, April and May) of original slope displacement data in days, which before resampling is simplified to the following example and input into the model:

[0094] ;

[0095] After resampling, a data set is taken every 7 days, and the sampling start date is the first day of each season. The first set of sampled data can be obtained:

[0096] ;

[0097] Similarly, the second set of sampled data can be obtained:

[0098] ;

[0099] By analogy, a total of 7 sets of sampled data can be obtained, which is equivalent to expanding the data that can only be used for one training to data that can be used for 7 trainings, providing more learning opportunities;

[0100] In each season, this periodic resampling takes into account the significance of slope displacement changes, making it easier for the model to learn slope displacement patterns in different seasons and improving its ability to perceive seasonal changes.

[0101] S2: seasonal decomposition of the preprocessed slope displacement data to extract seasonal characteristics; specifically including: calculating the trend component, seasonal component and residual component of the preprocessed slope displacement data by local smoothing method; splicing the trend component, seasonal component and residual component to extract seasonal characteristics;

[0102] S21: Calculate the trend component using the local smoothing method. The calculation method is:

[0103] ;

[0104] in, for The trend component of the moment, is the target time, i is the target time index, n is the total number of data points, For data points For the weight of the target data point, for The preprocessed slope displacement data at time, is another moment, j is the index of another moment, and the weight is calculated as follows:

[0105] ;

[0106] Where, e is the natural constant, h is the bandwidth, is the kth moment, k is the counting index;

[0107] In the present invention, the bandwidth h is used to control the degree of smoothing of the local smoothing method. The larger the bandwidth h, the more data points are used to calculate the trend component, which can better reflect the overall change of the time series; the smaller the bandwidth h, the fewer data points are used to calculate the trend component, which can better capture the subtle fluctuations of the time series. Local smoothing can adjust the weight of each data point according to the time distance. The closer the data point, the more important it is, and the farther the data point, the less important it is, so as to better reflect the trend change of the time series.

[0108] Use current time point Subtract another time point , is to calculate the time distance between two time points, this distance will affect the weight The size of It is a certain time point in the data. j can be understood as a loop variable, which will traverse all data points.

[0109] For example, suppose The sequence is from the 1st to the 30th of this month, with one data point per day, then The sequence is also from the 1st to the 30th of this month. Assuming that you need to calculate the trend component of 2023-12-15, you need to Set it to 2023-12-15, and then for each Calculate the weight once , that is, we need to calculate , the sum of these weights is 1. Then, these weights are multiplied by the corresponding data points , and then sum them up to get the trend component of 2023-12-15 .

[0110] S22: Calculate the seasonal component, the calculation method is:

[0111] ;

[0112] in, For at time point The seasonal components of For at time point Preprocessed slope displacement data at ;

[0113] S23: Calculate the residual component, the calculation method is:

[0114] ;

[0115] in, For at time point The residual component at ;

[0116] The trend component , the seasonal components , the residual component ,by The splicing of forms is the seasonal feature of this time point; among them, For splicing operation.

[0117] This step decomposes a time series into three components: trend, season, and residual. The trend component represents the long-term change of the time series, the seasonal component represents the periodic change of the time series, and the residual component represents the random fluctuation of the time series. By calculating the three components, it is helpful to decompose the signals from different sources in the slope displacement data more finely and comprehensively, increase the interpretability of the model, and at the same time, provide more comprehensive and rich input for the subsequent slope monitoring network, which helps the subsequent network to more accurately capture and learn the complex change pattern of displacement.

[0118] S3: Construct a slope monitoring network integrating seasonal characteristics and model the seasonal and temporal characteristics of slope displacement, such as Figure 3 As shown;

[0119] S31: Construct a slope monitoring network integrating seasonal features. The network consists of two GRUs and a feature fusion part, including:

[0120] S32: Input the seasonal features of each time point in each season into the first GRU, and take the last hidden state of the first GRU output as the seasonal feature vector, recorded as , GRU is calculated as:

[0121] ;

[0122] in, is the update gate at the current moment, t is the current moment, is the sigmoid function, To update the gate weight matrix, is the hidden state at the previous moment, is the preprocessed slope displacement data at the current moment, is the reset gate at the current moment, To reset the gate weight matrix, is the candidate hidden state at the current moment, tanh() is the tanh function, is the hyperbolic sine weight matrix, is element-wise multiplication;

[0123] S33: For the preprocessed slope displacement data, the second GRU is used to extract the time series feature vector, and the calculation method is:

[0124] ;

[0125] in, is the time series feature vector, For the GRU network, is the preprocessed slope displacement data;

[0126] S34: The seasonal feature vector is concatenated with the time series feature vector and input into a fully connected layer. The sigmoid function is used for activation and the seasonal weight is constructed and recorded as , the time series weight is recorded as , calculated as:

[0127] ;

[0128] in, is the weight matrix, is the bias vector;

[0129] In particular, when the accuracy of slope monitoring is required to be higher, the following method can be used to calculate the new time series weight instead of the above time series weight :

[0130] ;

[0131] in, is the new timing weight, is the time series weight matrix, is the temporal weight bias vector;

[0132] The sigmoid function is used as the activation function for seasonal weight calculation, so that the seasonal weight can more flexibly model the complex and nonlinear relationship between seasonal characteristics and temporal characteristics; at the same time, the model can flexibly adjust the relative contribution of seasonal characteristics and temporal characteristics in the model output when calculating seasonal weights; and the gating mechanism helps the model to more accurately control the flow of information and reduce the interference of irrelevant information, thereby improving the network's effective use of seasonal characteristics and temporal characteristics;

[0133] The advantage of the seasonal weight calculation method is its dynamic adaptability. The model can automatically adjust the weight according to the different conditions of the input seasonal characteristics and time series characteristics, thus realizing the adaptive fusion of seasonal characteristics and time series characteristics.

[0134] For example, assuming that the dimension D of the hidden state output by the GRU is 10, two seasonal feature vectors have been obtained: and , and the corresponding weight matrix and the bias vector ,in:

[0135] for: ;

[0136] for: ;

[0137] for: ;

[0138] for: ;

[0139] Substituting these parameters into the formula we get:

[0140] ;

[0141] This means that in the fusion of seasonal characteristics and time series characteristics of this set of data, the relative contribution of seasonal characteristics is larger, while the relative contribution of time series characteristics is smaller. This dynamic adaptive calculation method enables the model to automatically adjust the weights according to the input data, more flexibly integrate seasonal characteristics and time series characteristics, and help improve the model's perception of seasonal changes.

[0142] S35: dot-multiply the seasonal weight with the seasonal feature vector, dot-multiply the temporal weight with the temporal feature vector, and then sum them up to perform feature fusion to obtain the slope feature vector. The calculation method is:

[0143] ;

[0144] in, is the slope characteristic vector, is element-wise multiplication;

[0145] The feature fusion is performed by element-by-element multiplication, which is relatively simple and efficient. This helps to reduce the computational complexity of the model, while retaining the weight information of the seasonal feature vector and the time series feature vector, improving the interpretability of the model;

[0146] Through feature fusion calculation, the model can perform personalized learning of seasonal features and time series features at each time point, and the weights in feature fusion calculation and It is dynamically adaptive. By considering the weight relationship under different seasons and time series conditions, it helps the model to more accurately fuse information from different features. At the same time, the feature fusion process is relatively simple by adopting the element-by-element multiplication method, which avoids excessive information loss and helps to retain the model's sensitivity to seasonal and time series information during the feature fusion process, thereby improving the accuracy of the prediction.

[0147] S36: Input the slope feature vector into the fully connected layer to obtain the predicted value of the slope displacement data. The calculation method is:

[0148] ;

[0149] in, is the predicted value of slope displacement data, c is the moment count index, is the weight matrix, is the bias vector.

[0150] S4: setting the optimization target of the slope monitoring network integrating seasonal characteristics;

[0151] S41: The optimization target of the slope monitoring network integrating seasonal characteristics consists of two parts, the first part optimizes the ability of the slope monitoring network to model seasonal characteristics, and the second part optimizes the ability of the slope monitoring network to model temporal characteristics;

[0152] The first part optimizes the ability of the slope monitoring network to model seasonal characteristics, that is, the loss function in the first part is calculated as;

[0153] ;

[0154] Where C is the sequence length, is the actual value of slope displacement, is the natural logarithm, is the hyperbolic cosine function;

[0155] The Log-Cosh loss function measures the prediction error through the logarithmic function of the hyperbolic cosine. It can prevent the gradient explosion problem when the prediction error is large, and it can also avoid the gradient vanishing problem when the prediction error is close to zero. The Log-Cosh loss function is a smooth and continuous function. This smoothness helps the optimization algorithm to converge better, making the training process more stable and improving the training efficiency of the network. At the same time, since the logarithmic function of the hyperbolic cosine grows more slowly when the prediction error is large, it has less impact on outliers than MSE.

[0156] S42: The second part optimizes the ability of the slope monitoring network modeling time series characteristics, that is, the calculation method of the loss function in the second part is:

[0157] ;

[0158] The optimization objective of the slope monitoring network integrating seasonal characteristics is expressed as:

[0159] ;

[0160] in, is the optimization objective of the slope monitoring network integrating seasonal characteristics.

[0161] The optimization target of the slope monitoring network integrating seasonal characteristics adopts a combination of two loss functions, Log-Cosh and mean square error (MSE), which combines their respective advantages. Log-Cosh has good performance in robustness to outliers and gradient stability, while MSE has advantages in accurate measurement of small errors. The use of two different forms of loss functions makes the optimization target of the slope monitoring network integrating seasonal characteristics more flexible and adaptable in training. At different stages of model training, the influence of the loss function can be better adjusted to adapt to different training requirements and data characteristics, thereby improving the generalization performance of the model.

[0162] S5: Optimizing the parameters of the slope monitoring network integrating seasonal characteristics based on an adaptive learning rate optimization algorithm;

[0163] S51: The adaptive learning rate optimization algorithm is used to optimize the parameters of the slope monitoring network integrating seasonal characteristics, and the calculation method is:

[0164] ;

[0165] in, is the first-order moment estimate, is the second-order moment estimate, is the first-order moment estimate of the previous iteration, is the gradient of the current iteration, is the second-order moment estimate of the previous iteration, and are the first and second attenuation coefficients, respectively. and is the deviation correction for the first and second order moments, are the parameters of the slope monitoring network integrating seasonal characteristics after iteration, that is, the parameters of the slope monitoring network integrating seasonal characteristics optimized based on the adaptive learning rate optimization algorithm, is the parameter of the slope monitoring network integrating seasonal characteristics, is the learning rate, a is the number of iterations, is a constant, for a to the power of, for a to the power of;

[0166] S52: After completing the training of the parameters of the slope monitoring network integrating seasonal characteristics, the slope monitoring data of the road section to be predicted is input, and the slope monitoring network integrating seasonal characteristics outputs the slope displacement at the next moment.

[0167] Embodiment 2:

[0168] A slope monitoring system based on deep learning, comprising:

[0169] Data acquisition and preprocessing module: obtain the original slope displacement data, divide it into seasons and resample it to obtain the preprocessed slope displacement data;

[0170] Seasonal feature extraction module: seasonal decomposition of preprocessed slope displacement data to extract seasonal features;

[0171] Network construction module: construct a slope monitoring network integrating seasonal features, extract seasonal feature vectors and time series feature vectors according to the seasonal features and the preprocessed slope displacement data, and then perform feature fusion to obtain slope feature vectors; finally, input the slope feature vectors into the fully connected layer to obtain the predicted value of slope displacement data;

[0172] Network training module: setting the optimization target of the slope monitoring network integrating seasonal characteristics;

[0173] Network optimization module: Optimize the parameters of the slope monitoring network integrating seasonal characteristics based on an adaptive learning rate optimization algorithm. This implementation scheme is used to implement the above-mentioned slope monitoring method based on deep learning.

[0174] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method 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, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0175] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0176] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A slope monitoring method based on deep learning, characterized in that: The following steps are involved: S1: Obtain the original slope displacement data, divide it into seasons and resample it to obtain the preprocessed slope displacement data; specifically including: S11: dividing the original data of slope displacement by season to obtain data sets of four seasons: spring, summer, autumn and winter, wherein the original data of slope displacement is in days; S12: resampling the original slope displacement data of each season, the sampling method is: taking the date corresponding to the first data of each season as the sampling start date, taking one data every 7 days until 13 data are taken; then pushing the sampling start date back one day, and performing the next round of sampling; repeating the above process 7 times, and finally obtaining 7 groups of the pre-processed slope displacement data; S2: seasonally decomposing the preprocessed slope displacement data to extract seasonal characteristics; including calculating trend components, seasonal components and residual components of the preprocessed slope displacement data by a local smoothing method; splicing the trend components, seasonal components and residual components to extract seasonal characteristics; S21: Calculate the trend component using a local smoothing method, the calculation method is: ； in, for The trend component of the moment, is the target time, i is the target time index, n is the total number of data points, For data points For the weight of the target data point, for The preprocessed slope displacement data at time, is another moment, j is the index of another moment, and the weight is calculated as follows: ； Where, e is the natural constant, h is the bandwidth, is the kth moment, k is the counting index; S22: Calculate the seasonal component, the calculation method is: ； in, For at time point The seasonal components of For at time point Preprocessed slope displacement data at ; S23: Calculate the residual component, the calculation method is: ； in, For at time point The residual component at ; The trend component , the seasonal components , the residual component ,by The splicing of forms is the seasonal feature of this time point; among them, For splicing operation; S3: construct a slope monitoring network integrating seasonal features, extract seasonal feature vectors and time series feature vectors respectively according to the seasonal features and the preprocessed slope displacement data, and then perform feature fusion to obtain a slope feature vector; finally, input the slope feature vector into a fully connected layer to obtain a predicted value of the slope displacement data; S4: setting the optimization target of the slope monitoring network integrating seasonal characteristics; S5: Obtain parameters of the slope monitoring network integrating seasonal characteristics, and optimize the parameters of the slope monitoring network integrating seasonal characteristics based on an adaptive learning rate optimization algorithm.

2. The slope monitoring method based on deep learning according to claim 1 is characterized in that: The step S3 comprises the following steps: S31: Construct a slope monitoring network integrating seasonal features, which consists of 2 GRUs and a feature fusion part; S32: Input the seasonal features of each time point in each season into the first GRU, and take the last hidden state of the first GRU output as the seasonal feature vector, recorded as , GRU is calculated as: ； in, is the update gate at the current moment, t is the current moment, is the sigmoid function, To update the gate weight matrix, is the hidden state at the previous moment, is the preprocessed slope displacement data at the current moment, is the reset gate at the current moment, To reset the gate weight matrix, is the candidate hidden state at the current moment, tanh() is the tanh function, is the hyperbolic sine weight matrix, is element-wise multiplication; S33: For the preprocessed slope displacement data, the second GRU is used to extract the time series feature vector, and the calculation method is: ； in, is the time series feature vector, For the GRU network, is the preprocessed slope displacement data; S34: The seasonal feature vector is concatenated with the time series feature vector and input into a fully connected layer. The sigmoid function is used for activation and the seasonal weight is constructed and recorded as , the time series weight is recorded as , calculated as: ； in, is the weight matrix, is the bias vector; S35: dot-multiply the seasonal weight with the seasonal feature vector, dot-multiply the temporal weight with the temporal feature vector, and then sum them up to perform feature fusion to obtain the slope feature vector. The calculation method is: ； in, is the slope characteristic vector, is element-wise multiplication; S36: Input the slope feature vector into the fully connected layer to obtain the predicted value of the slope displacement data. The calculation method is: ； in, is the predicted value of slope displacement data, c is the moment count index, is the weight matrix, is the bias vector.

3. The slope monitoring method based on deep learning according to claim 2 is characterized in that: The step S4 comprises the following steps: S41: The optimization target of the slope monitoring network integrating seasonal characteristics consists of two parts, the first part optimizes the ability of the slope monitoring network to model seasonal characteristics, and the second part optimizes the ability of the slope monitoring network to model temporal characteristics; The first part optimizes the ability of the slope monitoring network to model seasonal characteristics, that is, the loss function in the first part is calculated as; ； Where C is the sequence length, is the actual value of slope displacement, is the natural logarithm, is the hyperbolic cosine function; S42: The second part optimizes the ability of the slope monitoring network modeling time series characteristics, that is, the calculation method of the loss function in the second part is: ； The optimization objective of the slope monitoring network integrating seasonal characteristics is expressed as: ； in, is the optimization objective of the slope monitoring network integrating seasonal characteristics.

4. The slope monitoring method based on deep learning according to claim 3 is characterized in that: The step S5 comprises the following steps: S51: The adaptive learning rate optimization algorithm is used to optimize the parameters of the slope monitoring network integrating seasonal characteristics, and the calculation method is: ； in, is the first-order moment estimate, is the second-order moment estimate, is the first-order moment estimate of the previous iteration, is the gradient of the current iteration, is the second-order moment estimate of the previous iteration, and are the first and second attenuation coefficients, respectively. and is the deviation correction for the first and second order moments, are the parameters of the slope monitoring network integrating seasonal characteristics after iteration, that is, the parameters of the slope monitoring network integrating seasonal characteristics optimized based on the adaptive learning rate optimization algorithm, is the parameter of the slope monitoring network integrating seasonal characteristics, is the learning rate, a is the number of iterations, is a constant, for a to the power of, for a to the power of; S52: After completing the training of the parameters of the slope monitoring network integrating seasonal characteristics, the slope monitoring data of the road section to be predicted is input, and the slope monitoring network integrating seasonal characteristics outputs the slope displacement at the next moment.

5. A slope monitoring system based on deep learning, characterized in that: include: Data acquisition and preprocessing module: obtain the original slope displacement data, divide it into seasons and resample it to obtain the preprocessed slope displacement data; Seasonal feature extraction module: seasonal decomposition of preprocessed slope displacement data to extract seasonal features; Network construction module: construct a slope monitoring network integrating seasonal features, extract seasonal feature vectors and time series feature vectors according to the seasonal features and the preprocessed slope displacement data, and then perform feature fusion to obtain slope feature vectors; finally, input the slope feature vectors into the fully connected layer to obtain the predicted value of slope displacement data; Network training module: setting the optimization target of the slope monitoring network integrating seasonal characteristics; Network optimization module: Optimize the parameters of the slope monitoring network integrating seasonal characteristics based on the adaptive learning rate optimization algorithm; To realize the slope monitoring method based on deep learning as described in any one of claims 1-4.

Images