A Migration Method for Slope Displacement Prediction Model

Abstract

The present application discloses a migration method for a slope displacement prediction model, comprising the following steps: real-time acquisition of monitoring data of a group of slopes in a region; selecting slope monitoring data with relatively stable geological characteristics as source domain data, and selecting slope monitoring data with different geological characteristics as target domain data; selecting a basic model for slope displacement prediction, and training the basic model with source domain data; adversarially training the source domain data and the target domain data based on a DANN algorithm to blur the domain distinction of homogeneous features between the source domain data and the target domain data. The beneficial effects of the present application: by studying how to combine multiple source domain data for more complex migration learning, the generalization ability of the model is improved. By inputting real-time data streams into the DANN algorithm, using its adversarial feature extraction capabilities, the model parameters are dynamically adjusted to adapt to real-time changes in slope conditions.

Description

Technical Field

[0001] The present application relates to the technical field of slope safety monitoring, and in particular to a migration method of a slope displacement prediction model. Background Art

[0002] At present, the migration method of slope displacement prediction model is mainly based on model construction of data of the same slope, and the adaptability of the model on different slopes is achieved by identifying the common characteristics between similar slopes. For example, by establishing a relationship model between horizontal displacement data and soil sample data, a relationship model between soil strength parameters and slope safety factor can be constructed using a neural network algorithm to predict the stability of the slope. For example, a deep learning prediction model of Newmark permanent displacement is established, and the model is optimized through transfer learning. This method uses numerical simulation results to train a new model, thereby saving a lot of training time and computing resources.

[0003] Although the existing migration method of slope displacement prediction model has its advantages, it still has many limitations and disadvantages:

[0004] (1) The effectiveness of the model is highly dependent on the quality and quantity of the data used. If the training data is insufficient or noisy, the prediction results may be inaccurate.

[0005] (2) Although transfer learning can improve the performance of the model in new environments, if the difference between the source domain and the target domain is large, the generalization ability of the model may be limited, thereby affecting the prediction accuracy.

[0006] (3) Deep learning models generally require higher computing resources, especially when processing large-scale data, which may lead to longer training time and increased computing costs.

[0007] (4) Deep learning models such as neural networks are often viewed as “black boxes” and their decision-making process is difficult to explain, which may limit their acceptance in some engineering applications.

[0008] (5) During the transfer learning process, local features of certain specific areas may be ignored, thus affecting the application effect of the model in specific scenarios.

[0009] (6) As new data and conditions emerge, the model may need to be updated frequently, otherwise it may not reflect the latest slope status and safety. Summary of the invention

[0010] One of the objectives of the present application is to provide a migration method of a slope displacement prediction model that can solve at least one of the defects in the above-mentioned background technology.

[0011] In order to achieve at least one of the above purposes, the technical solution adopted by the present application is: a migration method of a slope displacement prediction model, comprising the following steps:

[0012] S100: Real-time acquisition of monitoring data of slope groups in the area, including displacement data and environmental data;

[0013] S200: Selecting slope monitoring data with relatively stable geological characteristics as source domain data, and selecting slope monitoring data with different geological characteristics as target domain data;

[0014] S300: selecting a basic model for slope displacement prediction and training the basic model using source domain data;

[0015] S400: Perform adversarial training on the source domain data and the target domain data based on the DANN algorithm to blur the domain distinction of the same features between the source domain data and the target domain data; and then perform model migration on the basic model according to the feature data after the adversarial training of the DANN algorithm.

[0016] Preferably, step S400 includes the following process:

[0017] S410: Input the source domain data and the target domain data into the feature extractor at the same time, generate a shared feature representation and send it to the task classifier and the domain classifier;

[0018] S420: The task classifier performs back propagation to update parameters of the feature extractor and the task classifier;

[0019] S430: When the domain classifier performs back propagation, a negative gradient is applied to the feature extractor, so that the features learned by the feature extractor are indistinguishable between the source domain data and the target domain data.

[0020] Preferably, the task classifier is implemented by the task loss function L task Perform back propagation, and the domain classifier passes the domain loss function L domain Perform back propagation; the total loss function of the DANN algorithm is L=L task +λL domain ; Task loss function L task And the domain loss function L domain The expression is:

[0021] ;

[0022] ;

[0023] Among them, λ represents the hyperparameter that controls the domain loss weight, N represents the total number of slope samples in the region, and y i represents the measured displacement value of the i-th sample in the region, represents the displacement value predicted by the basic model for the i-th sample, d i represents the target label of the i-th sample. If the sample data comes from the source domain data, then d i =1, if the sample data comes from the target domain data, then d i =0, Represents the predicted value of the domain classifier for the i-th sample.

[0024] Preferably, when performing adversarial training of the DANN algorithm in step S400, it is necessary to iterate the model parameter θ in the basic model f(x; θ) for multiple times until the model converges; the model parameter θ is suitable for iterative updating by stochastic gradient descent and Adam optimizer; the rule for updating the model parameter θ based on stochastic gradient descent is: The model parameters θ are updated by back propagation; the rule for updating the model parameters θ based on the Adam optimizer is: The model parameters θ are updated by back propagation; where x represents the input data, η represents the learning rate, represents the gradient of the loss function L(θ) with respect to the model parameter θ, A constant representing a small value, represents the second-order moment v relative to time step t t The bias correction, Represents the first-order moment m relative to time step t t Bias correction.

[0025] Preferably, and The expression is:

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] Among them, β1 and β2 are hyperparameters, with values ​​of 0.9 and 0.999 respectively; represents the gradient, .

[0031] Preferably, the slope groups in the region are divided into at least one local region according to the geographical location, and each local region includes multiple slopes; model migration is performed in each local region through steps S200 to S400 to obtain a migration model network corresponding to each local region; and a domain adaptation model representing the slope groups in the region is constructed according to the migration model network of each local region.

[0032] Preferably, the domain adaptation model includes an input layer, a shared feature layer, a dedicated feature layer, a target domain feature layer, a domain adversarial layer and an output layer; the input layer is used to receive source domain data and target domain data; the shared feature layer is used to extract common features of the source domain data and the target domain data; the dedicated feature layer is used to generate feature representations of the source domain data, and the target domain feature layer is used to generate and optimize feature representations of the target domain data; the domain adversarial layer minimizes the feature distribution distance between the source domain data and the target domain data by generating an adversarial network or by adversarial training; the output layer generates a final prediction result based on the output of the dedicated feature layer.

[0033] Preferably, the shared feature layer includes a convolution layer, a pooling layer and an activation function; the convolution layer is used to capture spatial features to extract slope features, the pooling layer is used to reduce feature dimensions, and the activation function is used to increase the nonlinear expression ability of the model.

[0034] Preferably, in the process of constructing the domain adaptation model, the model is updated through a loss function, and the loss function includes a source domain loss function and a target domain loss function; the source domain loss function uses cross entropy loss or mean square error to optimize the source domain data, and the target domain loss function uses adversarial loss to reduce the difference between the source domain data and the target domain data.

[0035] Preferably, after the domain adaptation model is constructed, the model is evaluated and optimized, including a validation and testing process and a hyperparameter tuning process; wherein, the validation and testing process: using independent validation and test data sets to evaluate model performance to ensure the generalization ability of the model on unknown data; the hyperparameter tuning process: continuously adjusting hyperparameters according to the performance of the model.

[0036] Compared with the prior art, the beneficial effects of this application are:

[0037] (1) By studying how to combine data from multiple source domains to perform more complex transfer learning, the generalization ability of the model can be improved.

[0038] (2) Combine real-time monitoring data with the DANN algorithm to achieve dynamic monitoring and prediction of slope deformation. By inputting the real-time data stream into the DANN algorithm and utilizing its adversarial feature extraction capability, the model parameters are dynamically adjusted to adapt to the real-time changes in the slope state. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the overall workflow of the migration method of the slope displacement prediction model in this application. DETAILED DESCRIPTION

[0040] Below, the present application is further described in conjunction with specific implementation methods. It should be noted that in the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms should not be understood as necessarily being directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0041] In the description of the present application, it should be noted that directional words, such as the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc., indicating directions and positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of narrating the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and cannot be understood as limiting the specific scope of protection of the present application.

[0042] It should be noted that the terms "first", "second", etc. in the description and claims of the present application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0043] In this application, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be connected, detachably connected, or integrated; it can be mechanically connected or electrically connected; it can be directly connected or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0044] In the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0045] The terms "including" and "having" and any variations thereof in the specification and claims of the present application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0046] One aspect of the present application provides a migration method for a slope displacement prediction model, such as Figure 1 As shown, one of the preferred embodiments includes the following steps:

[0047] S100: Acquire monitoring data of the slope groups in the area in real time, including displacement data and environmental data.

[0048] S200: Selecting slope monitoring data with relatively stable geological characteristics as source domain data, and selecting slope monitoring data with different geological characteristics as target domain data.

[0049] S300: Select a basic model for slope displacement prediction, and train the basic model using source domain data.

[0050] S400: Perform adversarial training on source domain data and target domain data based on the DANN algorithm to blur the domain distinction of the same features between the source domain data and the target domain data; then the basic model can perform model migration based on the feature data after the adversarial training of the DANN algorithm.

[0051] It is understandable that the present application can perform adversarial training between source domain data and target domain data by using the DANN (Deep Adversarial Domain Adaptation Network) algorithm, thereby minimizing the feature differences between the source domain and the target domain to ensure that the model can effectively learn the common features of different slopes. That is, by using the DANN algorithm, the basic model can learn a more universal feature representation, so that the corresponding target domain model can be obtained by fine-tuning the basic model on the target domain data.

[0052] It can also be understood that the slope corresponding to the source domain data is the key slope data in the slope group in the region, while the slope corresponding to the target domain data is the general slope data in the slope group in the region; the source domain corresponds to the key slope, and the target domain corresponds to the general slope. In terms of data volume, the amount of data corresponding to the source domain data is far less than the amount of data corresponding to the target domain data. Based on the DANN algorithm, the model can be fine-tuned with a small amount of data from the source domain to improve the prediction ability of the model in the target domain.

[0053] It should be known that in step S100, the environmental data of the slope group in the region includes rainfall, temperature, humidity, etc., which can be obtained through sensors, monitoring equipment or historical data. In step S200, after obtaining the source domain data and the target domain data, the collected data can be cleaned, normalized, and feature selected, etc., to ensure the quality and consistency of the data. The source domain data is mainly used for the subsequent training of the basic model, and the target domain data can be used to test the performance of the basic model after migration.

[0054] In step S300, the process of constructing the basic model for slope displacement prediction is a well-known technology for those skilled in the art. For ease of understanding, a brief description can be given below. A feature selection random forest can be used to select features in the source domain data that have a significant impact on slope displacement to reduce redundant features. Then, based on the classified data, a basic model for slope displacement prediction is constructed using machine learning algorithms such as support vector machines, decision trees, and deep learning models. Finally, according to specific needs, such as slope displacement prediction, the most suitable model architecture is selected, and the k-fold cross-validation method is used to evaluate the basic model's performance in displacement prediction to ensure the generalization ability of the basic model on different data sets.

[0055] In this embodiment, the architecture of the DANN algorithm includes a feature extractor, a task classifier, and a domain classifier. The specific process based on the DANN algorithm architecture in step S400 is as follows:

[0056] S410: Input the source domain data and the target domain data into the feature extractor at the same time, generate a shared feature representation and send it to the task classifier and the domain classifier.

[0057] S420: The task classifier performs back propagation to update parameters of the feature extractor and the task classifier.

[0058] S430: When the domain classifier performs back propagation, a negative gradient is applied to the feature extractor, so that the features learned by the feature extractor are indistinguishable between the source domain data and the target domain data.

[0059] It should be known that the feature extractor is responsible for mapping the input samples to a shared feature space; the samples are source domain data and target domain data, and the samples can be images, texts, or other forms of data. The feature extractor can extract the features of the samples through a convolutional layer or a fully connected layer. The task classifier is responsible for further processing the features extracted by the feature extractor to complete specific tasks, such as slope displacement prediction. The task classifier usually shares some parameters with the feature extractor to improve the learning ability of the task. The domain classifier is responsible for determining whether the features come from the source domain data or the target domain data. The goal of the domain classifier is to maximize the classification accuracy so that the features learned by the feature extractor are not distinguishable to the domain classifier.

[0060] Specifically, the source domain data and the target domain data are simultaneously input into the feature extractor, and the feature extractor can generate a shared feature representation and feed it into the task classifier and the domain classifier. Then, the loss of the task classifier is calculated to evaluate the accuracy of the displacement prediction; at the same time, the loss of the domain classifier is calculated to evaluate the model's ability to distinguish domains. For the task classifier, after the loss calculation of the task classifier is completed, standard back propagation can be performed to update the parameters of the feature extractor and the task classifier. For the domain classifier, after the domain loss calculation is completed, back propagation can also be performed to update the parameters, but an adversarial training mechanism is introduced during the back propagation process, that is, a negative gradient is applied to the feature extractor, so that the features learned by the feature extractor are indistinguishable between the source domain and the target domain.

[0061] It is understandable that task loss is used to measure the performance of the basic model in completing specific tasks. Domain loss is used to measure the domain discrimination ability of the basic model between source domain data and target domain data. The DANN algorithm distinguishes samples from the source domain and the target domain through a domain classifier, aiming to make the features extracted by the feature extractor indistinguishable between the source domain and the target domain. The goal of the domain classifier is to maximize its classification accuracy through adversarial training, while the feature extractor attempts to confuse the domain classifier so that it cannot accurately distinguish between the source domain data and the target domain data.

[0062] In this embodiment, the task classifier uses the task loss function L task Perform back propagation, and the domain classifier passes the domain loss function L domain Perform back propagation; the total loss function of the DANN algorithm is L=L task +λL domain ; Task loss function L task And the domain loss function L domain The expression is:

[0063] .

[0064] .

[0065] Among them, λ represents the hyperparameter that controls the domain loss weight, N represents the total number of slope samples in the region, and y i represents the measured displacement value of the i-th sample in the region, represents the displacement value predicted by the basic model for the i-th sample, d i represents the target label of the i-th sample. If the sample data comes from the source domain data, then d i =1, if the sample data comes from the target domain data, then d i =0, Represents the predicted value of the domain classifier for the i-th sample.

[0066] In this embodiment, when performing adversarial training of the DANN algorithm in step S400, it is necessary to iterate the model parameter θ in the basic model f(x; θ) for multiple times until the model converges; the model parameter θ can be iteratively updated by stochastic gradient descent (SGD) and Adam optimizer; wherein x represents input data, i.e., source domain data and target domain data.

[0067] The rule for updating the model parameters θ based on stochastic gradient descent is: The model parameters θ are updated by back propagation. Among them, η represents the learning rate, which is a hyperparameter that can control the size of the update step; It represents the gradient of the loss function L(θ) with respect to the model parameters θ.

[0068] The rule for updating the model parameter θ based on the Adam optimizer is: The model parameters θ are updated by back propagation. A constant representing a small value, such as 10 -8 ; represents the second-order moment v relative to time step t t The bias correction, Represents the first-order moment m relative to time step t t Bias correction.

[0069] It should be known that there are many ways to calculate the loss of the basic model f(x; θ). In this embodiment, the cross entropy loss is preferably used as the loss function. The loss function is For the first-order moment m t Bias correction , and the second-order moment v t Bias correction The derivation process is as follows:

[0070] First initialize the first-order moment estimate m and the second-order moment estimate v to 0.

[0071] Then at each time step t, we calculate the gradient .

[0072] Then according to the gradient Update first-order moment , the second moment . Among them, β1 and β2 are hyperparameters, usually with values ​​of 0.9 and 0.999 respectively; m t-1 and v t-1 They represent the first-order moment and second-order moment corresponding to time step t-1 respectively.

[0073] Finally, you can get and The expression is:

[0074] ; .

[0075] It should be noted that during the iteration process, the loss function L(θ) and other performance indicators of the model, such as accuracy, precision, and recall, are monitored; if the loss function L(θ) no longer decreases significantly or does not improve within a certain number of iterations, the model can be considered to have converged.

[0076] It is understandable that stochastic gradient descent (SGD) is relatively simple in operation and easy to understand and debug. In terms of memory efficiency, only the gradient of the current sample is used to update the parameters, so the memory overhead is relatively small. In terms of robustness to noise, since it uses random selection of a single sample, it can jump out of the local minimum to a certain extent, which helps to find a better solution.

[0077] In terms of adaptive learning, the Adam optimizer can adjust the learning rate according to the historical gradient of each parameter. It has strong adaptability and can be updated efficiently at different training stages. Because the Adam optimizer combines the idea of ​​momentum, it performs well when facing sparse gradients and converges quickly, which is especially suitable for large-scale data sets; and the Adam optimizer can handle very large data sets and parameter spaces, which is especially effective for complex models such as deep neural networks. The Adam optimizer is less sensitive to hyperparameters such as learning rate, so in this embodiment, stochastic gradient descent and the Adam optimizer can be combined to iteratively update the model.

[0078] In this embodiment, when the slope group in the region contains a lot of slope data, if the relatively concentrated source domain data is still used for transfer learning of the target domain data, the number of migration layers will increase, and then the exponential growth of the migration data will be caused. Therefore, when there are a lot of slope groups in the region, the slope groups in the region can be divided into multiple regions according to the geographical location, and then the model migration is performed in each region through the aforementioned steps S200 to S400 to obtain the migration model network corresponding to each local region; finally, the domain adaptation model representing the slope group in the region is constructed according to the migration model network of each local region. In this way, the amount of migration data of the slopes in each divided region in the region can be effectively reduced to achieve the purpose of rapid migration.

[0079] It is understandable that the use of DANN algorithm for transfer learning aims to achieve full coverage of slope monitoring and early warning, highlighting key areas while taking into account the monitoring needs of general areas. By building a domain adaptation model with key and comprehensive monitoring capabilities based on multi-source domain data, the characteristics of different types of slopes are comprehensively considered. Comprehensive monitoring and early warning of slope groups in the region can be achieved, ensuring timely identification and response to risk areas, and improving the effectiveness of slope management and safety.

[0080] In this embodiment, the domain adaptation model still adopts the DANN algorithm architecture, which specifically includes an input layer, a shared feature layer, a dedicated feature layer, a target domain feature layer, a domain adversarial layer, and an output layer. The input layer is used to receive source domain data and target domain data; the shared feature layer is used to extract common features of source domain data and target domain data. The dedicated feature layer is used to generate feature representations of source domain data, and the target domain feature layer is used to generate and optimize feature representations of target domain data. The domain adversarial layer minimizes the feature distribution distance between source domain data and target domain data by generating an adversarial network or using adversarial training; the output layer generates the final prediction result based on the output of the dedicated feature layer.

[0081] It should be known that the input data received by the input layer can be images, sensor data, or other types of data. The dedicated feature layer may consist of additional convolutional layers, fully connected layers, and other processing components suitable for the source domain features. The design of the target domain feature layer is similar to the dedicated feature layer, but it is optimized for the characteristics of the target domain. Adaptive adjustment mechanisms such as batch normalization or layer normalization can be used to increase the expressiveness of the target domain features. The shared feature layer is mainly composed of multiple convolutional layers or feature extraction layers, which are designed to extract common features of the source domain data and the target domain data. For ease of understanding, the specific structure of the shared feature layer can be described below through a specific example.

[0082] Specifically, the shared feature layer includes convolutional layer, pooling layer and activation function. The convolutional layer is used to capture spatial features to extract slope features; the pooling layer is used to reduce feature dimensions to increase the robustness of the model; the activation function is used to increase the nonlinear expression ability of the model. Commonly used activation functions include linear rectification function (ReLU) and rectified linear rectification function (LeakyReLU). The output of the shared feature layer will be used in the dedicated feature layer to ensure the consistency of features learned in different fields.

[0083] It should also be known that the domain adversarial layer is specifically added to achieve model adaptation between the source domain and the target domain, which can reduce the distribution difference between the source domain and the target domain. The core advantage of the domain adversarial layer is that it allows the features learned by the model from the source domain to perform well in the target domain. Through adversarial training, the features learned by the feature extractor are not limited to a specific domain, but can be generalized on multiple data distributions. This allows the performance of the model in the target domain to be significantly improved, especially when there is less data in the source domain.

[0084] Specifically, traditional deep learning models usually assume that the source domain and the target domain have the same data distribution. However, in practical applications, there is often a large distribution difference between the source domain and the target domain. The domain adversarial layer introduces an adversarial training mechanism to force the model to learn features that are not dependent on the specific domain, thereby reducing the distribution difference between the source domain and the target domain, making the model more adaptable between different data distributions. At the same time, the adversarial mechanism of domain adversarial training makes the features learned by the feature extractor effective not only in the source domain, but also in the target domain. This helps to improve the feature representation ability of the model, enabling the model to better extract useful information from the input data.

[0085] The domain adversarial layer can be applied to unsupervised domain adaptation tasks, that is, the labeled data of the target domain is completely missing. Through adversarial training, the model only relies on the labeled data of the source domain and the unlabeled data of the target domain, thus enabling cross-domain adaptive learning. This capability enables the model to learn effectively even when the labeled data of the target domain is very scarce or unavailable.

[0086] The domain adversarial layer can force the model to extract features that are insensitive to the domain classifier through the gradient reversal mechanism. In other words, the model will learn some general features that are not related to a specific domain, so that it can better perform tasks on multiple domains. This adaptive feature learning is very important for cross-domain tasks, especially when the target domain and the source domain have a large difference in distribution. The domain adversarial layer can also be used in combination with other deep learning methods (such as convolutional neural networks, recurrent neural networks, etc.) and flexibly embedded into existing models. It can not only solve simple classification problems, but also be extended to complex tasks such as generation models, sequence labeling, etc.

[0087] In this embodiment, when constructing a domain adaptation model, the model can be updated by a loss function. The loss function includes a source domain loss function and a target domain loss function; the source domain loss function uses cross entropy loss or mean square error to optimize the source domain data, and the target domain loss function uses adversarial loss to reduce the difference between the source domain data and the target domain data. For the specific expressions and update process of the source domain loss function and the target domain loss function, reference can be made to the iterative update process of the aforementioned DANN algorithm.

[0088] In this embodiment, after the construction of the domain adaptation model is completed, the model is evaluated and optimized, including a verification and testing process and a hyperparameter tuning process; wherein, the verification and testing process: using independent verification and test data sets to evaluate the model performance to ensure the generalization ability of the model on unknown data; the hyperparameter tuning process: continuously adjusting the hyperparameters according to the performance of the model, and the adjusted hyperparameters include the learning rate, batch size, and regularization parameters.

[0089] Another aspect of the present application provides a slope safety early warning system, which can migrate the slope displacement model within the region through the aforementioned migration method, thereby monitoring and predicting the displacement data of the slope group in the entire region, thereby achieving timely early warning of the safety of the slope group in the region.

[0090] Specifically, for the slope groups in the area, real-time monitoring of key hidden danger slopes can be carried out, and IoT sensors such as displacement sensors, groundwater pressure gauges and rain gauges can be integrated to collect real-time data on the physical state of the slopes, and groundwater level and rainfall data can be collected for general hidden danger slopes. These data are sent to the central processing system via wireless communication. The slope displacement prediction model of key hidden danger slopes is constructed using the historical data and real-time data of the slope group to predict the behavior and displacement trend of key hidden danger slopes; at the same time, the constructed slope displacement prediction model is migrated to general hidden danger slopes through the DANN algorithm, and real-time prediction is performed in combination with real-time data.

[0091] It should be known that when conducting safety warning of the slope, it is necessary to determine the warning threshold and warning level; different safety behaviors correspond to different warning thresholds and warning levels. By analyzing historical data, the preliminary warning threshold can be determined, and the warning threshold can be specifically set according to statistical indicators such as the standard deviation of the edge displacement and the maximum displacement. The warning level can be set according to the displacement of the slope, such as normal, attention, warning and emergency, and each level corresponds to a different specific threshold.

[0092] In the process of slope safety warning, real-time sensor data and slope history records are used as model input to extract relevant features, such as current displacement, displacement change rate, historical change trend, etc. The machine learning model is trained based on historical data to learn the relationship between slope behavior and warning response under different circumstances. The warning threshold is adjusted dynamically according to real-time data and model prediction results. For example, if displacement acceleration is detected, the warning threshold can be lowered to ensure that a warning can be issued at an early stage. The system continuously monitors the slope status and immediately triggers the warning mechanism once it detects that the displacement exceeds the warning threshold. The system has a variety of ways to issue warnings, such as sending text messages to relevant responsible persons and managers, pushing warning notifications on the mobile devices of relevant personnel, sending emails to inform the latest situation, and updating the warning status in real time on the visual dashboard of the monitoring center.

[0093] Establish a feedback mechanism to collect feedback from all parties after the warning event and evaluate the effectiveness and accuracy of the warning. Regularly evaluate and optimize the prediction model and retrain it with new monitoring data to further improve the accuracy of the prediction and the responsiveness of the system. This ensures that the system is always in the best condition, so as to effectively respond to changes in the slope environment. This continuous optimization process makes the system have good adaptability and can cope with different slope conditions and emergencies.

[0094] The above describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and the specification only describe the principles of the present application. The present application may have various changes and improvements without departing from the spirit and scope of the present application, and these changes and improvements fall within the scope of the present application for which protection is sought. The scope of protection claimed by the present application is defined by the attached claims and their equivalents.

Claims

1. A migration method for a slope displacement prediction model, characterized in that: The steps include: S100: Real-time acquisition of monitoring data of slope groups in the area, including displacement data and environmental data; S200: Selecting slope monitoring data with relatively stable geological characteristics as source domain data, and selecting slope monitoring data with different geological characteristics as target domain data; S3 00: Select the basic model for slope displacement prediction and train the basic model with source domain data; S400: performing adversarial training on the source domain data and the target domain data based on the DANN algorithm to blur the domain distinction of the same characteristics between the source domain data and the target domain data; and then performing model migration on the basic model according to the characteristic data after the adversarial training of the DANN algorithm; Step S400 includes the following process: S410: Input the source domain data and the target domain data into the feature extractor at the same time, generate a shared feature representation and send it to the task classifier and the domain classifier; S420: The task classifier performs back propagation to update parameters of the feature extractor and the task classifier; S430: applying a negative gradient to the feature extractor when the domain classifier performs back propagation, so that the features learned by the feature extractor are indistinguishable between the source domain data and the target domain data; When performing adversarial training of the DANN algorithm in step S400 , it is necessary to iterate the model parameter θ in the basic model f(x; θ) multiple times until the model converges; The model parameters θ are suitable for iterative updates via stochastic gradient descent and Adam optimizer; The rule for updating the model parameter θ based on stochastic gradient descent is: update the model parameter θ through back propagation of θ-η▽L(θ); The rule for updating the model parameter θ based on the Adam optimizer is: Back propagation updates the model parameters θ; Where x represents the input data, η represents the learning rate, ▽L(θ) represents the gradient of the loss function L(θ) with respect to the model parameter θ, ∈ represents a small constant, represents the second-order moment v relative to time step t t The bias correction, Represents the first-order moment relative to time step t Bias correction.

2. The migration method of the slope displacement prediction model according to claim 1, characterized in that: The task classifier is calculated through the task loss function L task Perform back propagation, and the domain classifier passes the domain loss function L domain Perform back propagation; the total loss function of the DANN algorithm is L=L task +λL domain ; Task loss function L task And the domain loss function L domain The expression is: ； ； Among them, λ represents the hyperparameter that controls the domain loss weight, N represents the total number of slope samples in the region, and y i represents the measured displacement value of the i-th sample in the region, represents the displacement value predicted by the basic model for the i-th sample, d i represents the target label of the i-th sample. If the sample data comes from the source domain data, then d i =1, if the sample data comes from the target domain data, then d i =0, Represents the predicted value of the domain classifier for the i-th sample.

3. The migration method of the slope displacement prediction model according to claim 1, characterized in that: and The expression is: ； ； ； ； Among them, β1 and β2 are hyperparameters, with values ​​of 0.9 and 0.999 respectively; represents the gradient, .

4. The migration method of the slope displacement prediction model according to any one of claims 1 to 3, characterized in that: The slope groups in the region are divided into at least one local region according to the geographical location, and each local region includes multiple slopes; In each local area, model migration is performed through steps S200 to S400 to obtain a migration model network corresponding to each local area; and a domain adaptation model representing the slope group in the area is constructed according to the migration model network of each local area.

5. The migration method of the slope displacement prediction model according to claim 4, characterized in that: Domain adaptation models include: Input layer: The input layer is used to receive source domain data and target domain data; Shared feature layer: The shared feature layer is used to extract common features of source domain data and target domain data; Dedicated feature layer: The dedicated feature layer is used to generate feature representations of source domain data; Target domain feature layer: The target domain feature layer is used to generate and optimize the feature representation of target domain data; Domain adversarial layer; the domain adversarial layer minimizes the feature distribution distance between the source domain data and the target domain data by generating an adversarial network or using adversarial training; and Output layer: The output layer generates the final prediction results based on the output of the dedicated feature layer.

6. The migration method of the slope displacement prediction model according to claim 5, characterized in that: The shared feature layers include convolutional layers, pooling layers, and activation functions; The convolution layer is used to capture spatial features to extract slope features, the pooling layer is used to reduce the feature dimension, and the activation function is used to increase the nonlinear expression ability of the model.

7. The migration method of the slope displacement prediction model according to claim 5, characterized in that: In the process of building the domain adaptation model, the model is updated through the loss function, which includes the source domain loss function and the target domain loss function; The source domain loss function uses cross entropy loss or mean square error to optimize the source domain data, and the target domain loss function uses adversarial loss to reduce the difference between the source domain data and the target domain data.

8. The migration method of the slope displacement prediction model according to claim 5, characterized in that: After the domain adaptation model is built, the model is evaluated and optimized, including the verification and testing process and the hyperparameter tuning process; in Verification and testing process: Use independent verification and test data sets to evaluate model performance and ensure the model's generalization ability on unknown data; hyperparameter tuning process: Continuously adjust hyperparameters based on model performance.

Images