Slope state prediction method and device based on physical constraint and electronic equipment

Abstract

The present disclosure provides a physical constraint-based slope state prediction method, device and electronic equipment, and relates to the technical field of slope monitoring. The method comprises: obtaining first initial feature data of a plurality of monitoring points on a slope according to monitoring information of the plurality of monitoring points; determining the mechanical similarity of a plurality of monitoring point pairs, and constructing a first adjacency matrix of the plurality of monitoring points according to the mechanical similarity; each monitoring point pair comprises two monitoring points; performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of a first monitoring point, and obtaining target feature data of the first monitoring point based on the first aggregated feature data; and processing the target feature data by using a sequence model to obtain state prediction data of the first monitoring point. The present disclosure improves the accuracy of slope state prediction.

Description

Technical Field

[0001] The present disclosure relates to the technical field of slope monitoring, and in particular to a slope state prediction method, device and electronic equipment based on physical constraints. Background Art

[0002] Slopes are defined as slopes or excavated surfaces with a certain inclination. These slopes are prone to geological disasters such as landslides and debris flows, which can cause serious consequences. Slope condition prediction technology can identify potential disaster risks, facilitate timely prevention and control measures, and protect people's lives and property.

[0003] Related technologies typically build data-driven models, using statistical learning and other methods to extract patterns and trends from slope monitoring data to predict slope conditions. However, this method has limited accuracy, especially when dealing with complex geological conditions, which can hinder disaster prevention and control efforts. Summary of the Invention

[0004] The present disclosure provides a slope state prediction method, device and electronic device based on physical constraints, so as to at least improve the problem of low accuracy of slope state prediction to a certain extent.

[0005] According to a first aspect of the present disclosure, a slope state prediction method based on physical constraints is provided, the method comprising: obtaining first initial feature data of multiple monitoring points on the slope based on monitoring information of the multiple monitoring points; determining the mechanical similarity of multiple monitoring point pairs, and constructing a first adjacency matrix of the multiple monitoring points based on the mechanical similarity; each monitoring point pair includes two monitoring points; performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and obtaining target feature data of the first monitoring point based on the first aggregated feature data; and processing the target feature data using a sequence model to obtain state prediction data of the first monitoring point.

[0006] According to a second aspect of the present disclosure, a physical constraint based slope state prediction device is provided, the device comprising: an initial feature data acquisition module configured to acquire first initial feature data of a plurality of monitoring points on a slope according to monitoring information of the plurality of monitoring points; an adjacency matrix construction module configured to determine mechanical similarity of a plurality of monitoring point pairs, and construct a first adjacency matrix of the plurality of monitoring points according to the mechanical similarity; each monitoring point pair comprising two monitoring points; a feature aggregation processing module configured to perform feature aggregation processing based on the first adjacency matrix and the first initial feature data, to obtain first aggregated feature data of a first monitoring point, and to obtain target feature data of the first monitoring point based on the first aggregated feature data; and a prediction processing module configured to process the target feature data using a sequence model to obtain state prediction data of the first monitoring point.

[0007] According to a third aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method of the first aspect and possible implementation manners thereof.

[0008] According to a fourth aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the method of the first aspect and possible implementation manners thereof by executing the executable instructions.

[0009] The technical solution of the present disclosure has the following beneficial effects:

[0010] On the one hand, for the monitoring points on the slope, the initial feature data is aggregated by constructing an adjacency matrix to obtain aggregated feature data, which can learn the correlation between different monitoring points, and the sequence model is used to process the target feature data, combining the spatio-temporal correlation of the monitoring points, to improve the accuracy of the slope state prediction result. On the other hand, the physical constraint is introduced to determine the mechanical similarity between different monitoring points and construct a first adjacency matrix, so that the mechanical modal correlation between the monitoring points can be captured when the feature aggregation processing is performed, so that the slope state prediction result conforms to the geotechnical mechanics criterion (such as the limitation of shear strength). Compared with the traditional data driven model, the present solution can reduce the non-physical bias in the prediction result, improve the physical consistency of the prediction result, and further improve the prediction accuracy, and is particularly suitable for processing complex geological conditions. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 A flowchart of a physical constraint based slope state prediction method in an embodiment of the present disclosure is shown.

[0012] Figure 2A flow chart of obtaining first aggregated feature data in an embodiment of the present disclosure is shown.

[0013] Figure 3 A flow chart of obtaining second aggregated feature data in an embodiment of the present disclosure is shown.

[0014] Figure 4 A flow chart of obtaining state prediction data using GRU in an embodiment of the present disclosure is shown.

[0015] Figure 5 A flow chart of training a model in an embodiment of the present disclosure is shown.

[0016] Figure 6 An architecture diagram of a slope state prediction method in an embodiment of the present disclosure is shown.

[0017] Figure 7 A schematic diagram of the calculation results of a first adjacency matrix and a second adjacency matrix in an embodiment of the present disclosure is shown.

[0018] Figure 8 A comparison diagram of total loss function convergence curves of three schemes in an embodiment of the present disclosure is shown.

[0019] Figure 9 A structure diagram of a slope state prediction device based on physical constraints in an embodiment of the present disclosure is shown.

[0020] Figure 10 A structure diagram of an electronic device in an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0021] Example embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings.

[0022] The accompanying drawings, which are included to provide a further understanding of the disclosure and are incorporated in and constitute a part of this application, illustrate certain non-limiting embodiments of the present disclosure. The illustrated embodiments are intended to explain the present disclosure and not to impose limitations on the scope of the present disclosure. In the drawings:

[0023] Slope state prediction technology is crucial for geological disaster warning. In related technologies, a data-driven model is usually constructed to mine the rules and trends of data development from the monitoring data of the slope by using statistical learning and other methods, and then the slope state is predicted.

[0024] The above method ignores the physical characteristics of the slope, and may predict results that exceed the actual mechanical constraints, such as displacement prediction results with shear strength exceeding the limit, so that there are non-physical deviations (i.e. deviations due to non-compliance with physical constraints) in the prediction results, affecting the prediction accuracy. Moreover, the slope monitoring data often has the characteristics of small sample and multi-modal, and the model is difficult to fully learn the internal rules of the data, and is prone to overfitting and other situations, which also leads to low prediction accuracy, especially in complex geological conditions, and affects the model training efficiency.

[0025] In view of one or more of the above problems, the embodiments of the present disclosure provide a slope state prediction method based on physical constraints to improve the accuracy of slope state prediction.

[0026] Figure 1 An exemplary flow of the slope state prediction method based on physical constraints is shown, including the following steps S110-S140:

[0027] Step S110, obtaining first initial feature data of a plurality of monitoring points according to monitoring information of the plurality of monitoring points on the slope;

[0028] Step S120, determining the mechanical similarity of a plurality of monitoring point pairs, and constructing a first adjacency matrix of the plurality of monitoring points according to the mechanical similarity; each monitoring point pair includes two monitoring points;

[0029] Step S130, performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and obtaining target feature data of the first monitoring point based on the first aggregated feature data;

[0030] Step S140, processing the target feature data by using a sequence model to obtain state prediction data of the first monitoring point.

[0031] Based on Figure 1The proposed method, on the one hand, aggregates the initial feature data for the monitoring points on the slope by constructing an adjacency matrix to obtain aggregated feature data. This allows the association between different monitoring points to be learned, and the target feature data to be processed using a sequence model. This combines the spatiotemporal associations of the monitoring points to improve the accuracy of the slope state prediction results. On the other hand, physical constraints are introduced to determine the mechanical similarity between different monitoring points and construct a first adjacency matrix. This allows the mechanical modal correlation between the monitoring points to be captured during feature aggregation, ensuring that the slope state prediction results meet geotechnical criteria (such as shear strength limits). Compared to traditional data-driven models, this solution can reduce non-physical deviations in the prediction results and enhance the physical consistency of the prediction results, thereby further improving the prediction accuracy. It is particularly suitable for dealing with complex geological conditions.

[0032] Below Figure 1 Provide detailed instructions for each step.

[0033] refer to Figure 1 In step S110, first initial feature data of multiple monitoring points are obtained based on monitoring information of multiple monitoring points on the slope.

[0034] In the disclosed embodiments, the slope status information that needs to be monitored or predicted includes, but is not limited to, displacement, slope, deformation or other structural conditions, rock and soil conditions, and so on. Sensors or monitoring objects are installed at specific points on the slope to monitor the relevant information at these points. These points are referred to as monitoring points. For example, displacement monitoring can be achieved by installing displacement sensors, GNSS (Global Navigation Satellite System) monitoring stations, or visual targets at these monitoring points to detect even tiny displacements.

[0035] Initial feature data refers to the feature data that is initially extracted. The monitoring information can be used as the initial feature data, or the monitoring information can be pre-processed to obtain the initial feature data. In addition, the initial feature data can also come from the basic information of the monitoring point. The basic information is information that does not need to be obtained through continuous monitoring. It can be information that does not change with time, environment, etc., such as the geological type, rock and soil type, and first monitoring time of the monitoring point. The first initial feature data may include initial feature data related to mechanical properties, such as stress, strain, etc. In one embodiment, the first initial feature data may include all the initial feature data.

[0036] For example, the initial feature data of multiple monitoring points can be expressed as ,in N is the number of monitoring points, d is the dimension of the initial feature data. The first initial feature data is represented as ,in d1 is the dimension of the first initial feature data.

[0037] Continue to refer Figure 1 In step S120, the mechanical similarity of multiple monitoring point pairs is determined, and a first adjacency matrix of the multiple monitoring points is constructed according to the mechanical similarity; each monitoring point pair includes two monitoring points.

[0038] The monitoring point pairs used to calculate mechanical similarity can include all monitoring point pairs, i.e., any two monitoring point pairs, or only some monitoring point pairs. In one embodiment, two monitoring points whose physical spatial distance is less than a specific value (or whose spatial correlation is greater than a spatial correlation threshold, which can be set based on experience or specific business requirements) can be combined into a monitoring point pair and mechanical similarity calculated. Mechanical similarity is not calculated for other monitoring point pairs.

[0039] The mechanical similarity of each monitoring point pair represents the degree of similarity in the mechanical parameters of the two monitoring points in the monitoring point pair. In one embodiment, the mechanical parameters of the monitoring points can be represented as vectors, and the similarity of the mechanical parameter vectors of the two monitoring points can be calculated to obtain the mechanical similarity of the two monitoring points, that is, the mechanical similarity of the monitoring point pair consisting of the two monitoring points.

[0040] In one embodiment, determining the mechanical similarity of multiple monitoring point pairs comprises the following steps:

[0041] The shear strength is determined based on the mechanical parameters of the two monitoring points in the monitoring point pair, and the mechanical similarity of the monitoring point pair is obtained based on the difference in the shear strength of the two monitoring points; the mechanical similarity is negatively correlated with the difference in shear strength.

[0042] That is, the closer the shear strengths of the two monitoring points in a monitoring point pair are, the higher their mechanical similarity is. For example, the mechanical similarity of a monitoring point pair can be calculated using the following formula (1):

[0043] ;

[0044] ; (1)

[0045] in, i 、 j represents the two monitoring points in the monitoring point pair, W mech (i,j) represents the mechanical similarity of the monitoring point pair, τ represents the shear strength, c Indicates cohesion, represents the internal friction angle, σ nRepresents normal stress. One or more of shear strength, cohesion, internal friction angle, and normal stress can be calculated based on monitoring information. These parameters can also be derived from monitoring information combined with other experimental or analytical data (such as finite element simulation data).

[0046] When determining the mechanical similarity, the mechanical similarity of two monitoring points can be used as the first adjacency weight of the two monitoring points, that is, the first adjacency weight is W mech (i,j) , which represents the correlation between the two monitoring points under the mechanical mode, and thus constructs the first adjacency matrix.

[0047] In one embodiment, monitoring points are used as nodes, and the edge weights between nodes are determined based on mechanical similarity to construct a graph convolutional network (GCN). Since the edge weight represents the correlation between two monitoring points under mechanical modalities, it can be specifically called a mechanical weight. The graph convolutional network is represented by the first adjacency matrix. Of course, in addition to the first adjacency matrix, the graph convolutional network can also include other parameters, such as the number of layers, the learnable parameters of each layer, etc. Referring to formula (1), it can be seen that the node i and j Difference in shear strength | τ i -τ j The smaller the mechanical weight W mech (i,j) The larger the value, the higher the correlation between monitoring points with similar mechanical properties. This allows us to transform information such as geotechnical parameters and stress states into weights within the graph convolutional network structure through physical principles. This allows us to dynamically adapt to the stability differences between different slope regions when predicting slope conditions, improving prediction accuracy.

[0048] In one embodiment, to facilitate the distinction between the graph convolutional network constructed based on spatial similarity below, the graph convolutional network constructed based on mechanical similarity can be referred to as the first graph convolutional network, and the graph convolutional network constructed based on spatial similarity can be referred to as the second graph convolutional network. Alternatively, the embodiments of the present disclosure can construct a dual-channel graph convolutional network, including a mechanical channel constructed based on mechanical similarity and a spatial channel constructed based on spatial correlation.

[0049] Continue to refer Figure 1 In step S130, feature aggregation processing is performed based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and target feature data of the first monitoring point is obtained based on the first aggregated feature data.

[0050] The first monitoring point is a monitoring point to be predicted, and can be any one or more monitoring points. For example, all monitoring points can be taken as the first monitoring point, so that the states of all monitoring points at a certain time can be obtained through one prediction, and then the overall state prediction result of the slope can be obtained.

[0051] The first adjacency matrix includes the mechanical modal correlation between different monitoring points, and the first initial feature data of the monitoring points related to the first monitoring point can be aggregated based on the first adjacency matrix. For example, the first initial feature data of the first monitoring point and the adjacent monitoring points can be aggregated based on the first adjacency matrix, such as weighting the first initial feature data of the first monitoring point and the adjacent monitoring points according to the mechanical weight in the first adjacency matrix to achieve the aggregation effect, and multi-layer aggregation can be performed to obtain the first aggregated feature data of the first monitoring point.

[0052] In an embodiment, referring to FIG. 1, the above feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain the first aggregated feature data of the first monitoring point includes the following steps S210 to S230: Figure 2

[0053] In step S210, a first initial feature matrix is generated according to the first initial feature data of the plurality of monitoring points.

[0054] In the first initial feature matrix, each row represents the first initial feature data of a monitoring point, and the first initial feature data of the plurality of monitoring points is organized into a matrix. X1 For example, the first initial feature data is reconstructed into a row-column form to obtain the first initial feature matrix, which is represented as wherein the subscript mech represents the mechanical property-related feature, and the superscript 0 represents the matrix formed by the initial feature.

[0055] In step S220, one or more layers of feature aggregation processing are performed based on the first adjacency matrix, taking the first initial feature matrix as the starting point, to obtain a first aggregated feature matrix.

[0056] For example, the feature aggregation processing can be performed layer by layer through the following formula (2):

[0057] (2)

[0058] wherein, W mech The first adjacency matrix is represented as Amech, is a degree matrix of the row-normalized first adjacency matrix, is the first aggregated feature matrix of the i-th layer, l is the mechanical channel of the graph convolution network, and is the i-th layer of the first aggregated feature matrix, l ​Layers have learnable parameters, l It can be any non-negative integer. ReLU (Rectified Linear Unit) is a nonlinear activation function. Of course, other activation functions can also be used in formula (2).

[0059] Referring to formula (2), we can see that when performing feature aggregation processing, first let l is 0, at this time is the first initial feature matrix, and the first aggregated feature matrix of the first layer is obtained by performing feature aggregation processing on the first layer using formula (2). Then let l If it is 1, the feature aggregation processing of the second layer is performed to obtain the first aggregated feature matrix of the second layer. . This makes l Add one layer at a time and perform feature aggregation until the number of layers set by the mechanical channel of the graph convolutional network is reached. L1 When , the final first aggregate feature matrix is ​​obtained, which is recorded as .

[0060] Step S230: extracting first aggregated feature data of the first monitoring point from the first aggregated feature matrix.

[0061] The first aggregated feature data is high-level feature data obtained by aggregating the initial feature data from the mechanical modal. The first aggregated feature matrix can be regarded as a collection of the first aggregated feature data of all monitoring points. For example, if the first monitoring point includes all monitoring points, the first aggregated feature matrix can be directly used as the first aggregated feature data of the first monitoring point. If the first monitoring point is a specific monitoring point, the first aggregated feature data of the first monitoring point can be extracted from the corresponding row in the first aggregated feature matrix. The first aggregated feature data realizes the monitoring point representation based on the mechanical modal.

[0062] Once the first aggregated feature data for the first monitoring point is obtained, the first aggregated feature data can be used as the target feature data, or the first aggregated feature data can be combined with other feature data (such as initial feature data other than the first initial feature data) to obtain the target feature data. The target feature data is feature data that can be processed by the sequence model and can include relatively comprehensive feature information for the first monitoring point. For example, if the target feature data includes information from the first aggregated feature data, it means that it carries information about the mechanical modal correlations between different monitoring points and can reflect the physical constraint effect.

[0063] In addition to the above-mentioned constraints based on mechanical modalities, the physical constraints in the embodiments of the present disclosure may also include other physical constraints. The following description will take spatial modal constraints as an example.

[0064] In one embodiment, reference Figure 3 As shown, the slope condition monitoring method further includes the following steps S310 to S330:

[0065] Step S310: obtaining second initial feature data of the plurality of monitoring points according to monitoring information of the plurality of monitoring points.

[0066] As described above, the monitoring information can be used as initial feature data, or the monitoring information can be pre-processed to obtain initial feature data. Then, the second initial feature data is extracted from the initial feature data. The second initial feature data may include initial feature data related to spatial characteristics, such as geological parameters, displacement history, etc. In one embodiment, the second initial feature data may include all initial feature data. For example, the second initial feature data is represented as ,in N is the number of monitoring points, d2 is the dimension of the second initial feature data.

[0067] Step S320: determining the spatial correlations of the plurality of monitoring point pairs, and constructing a second adjacency matrix of the plurality of monitoring points according to the spatial correlations.

[0068] The monitoring point pairs used to calculate spatial correlation can include all monitoring point pairs, i.e., any two monitoring point pairs, or only some monitoring point pairs. In one embodiment, two monitoring points whose mechanical similarity exceeds a mechanical similarity threshold (this threshold can be set based on experience or specific business needs) can be combined into a monitoring point pair, and spatial correlation is calculated. Spatial correlation is not calculated for other monitoring point pairs. Furthermore, the monitoring point pairs used to calculate spatial correlation can be the same as or different from the monitoring point pairs used to calculate mechanical similarity.

[0069] The spatial correlation of each monitoring point pair represents the degree of geographic spatial correlation between the two monitoring points in the pair. In one embodiment, the geospatial parameters of the monitoring points (e.g., location coordinates, slope, normal coordinates, accumulated historical displacement, distance from the slope center or edge, etc.) can be represented as vectors. The similarity of the geospatial parameter vectors of the two monitoring points can be calculated to obtain the spatial correlation of the two monitoring points, i.e., the spatial correlation of the monitoring point pair consisting of the two monitoring points.

[0070] In one embodiment, determining the spatial correlation of multiple monitoring point pairs includes the following steps:

[0071] The spatial correlation of the monitoring point pair is determined according to the distance between the two monitoring points in the monitoring point pair; the spatial correlation is negatively correlated with the distance between the two monitoring points.

[0072] That is, the closer the distance between the two monitoring points in a monitoring point pair is, the higher the spatial correlation is. For example, the spatial correlation of the monitoring point pair can be calculated by the following formula (3):

[0073] ; (3)

[0074] in, i 、 j represents the two monitoring points in the monitoring point pair, d ij Indicates monitoring point i and j distance, W spa (i,j) It represents the spatial correlation of the monitoring point pair, and σ is the Gaussian kernel parameter, which is used to control the speed at which the spatial correlation changes with distance. It can be set based on experience or specific business needs.

[0075] When determining the spatial correlation, the spatial correlation of two monitoring points can be used as the second adjacency weight of the two monitoring points, that is, the second adjacency weight is W spa (i,j) , which represents the correlation between two monitoring points in spatial mode, and thus constructs the second adjacency matrix.

[0076] In one embodiment, monitoring points are used as nodes, and the edge weights between nodes are determined based on spatial similarity to construct a second graph convolutional network or a spatial channel of a graph convolutional network. Since the edge weight represents the correlation between two monitoring points in the spatial modality, it can be specifically called a spatial weight. The second graph convolutional network or spatial channel is represented by a second adjacency matrix. Of course, in addition to the second adjacency matrix, the second graph convolutional network or spatial channel can also include other parameters, such as the number of layers, the learnable parameters of each layer, etc. Referring to formula (3), it can be seen that the node i and j The smaller the distance, the greater the spatial weight W spa (i,j) The larger the value, the higher the correlation between spatially close monitoring points. This allows the model to more fully learn the spatial relationship between monitoring points and the spatial structural characteristics of the slope when predicting slope conditions, which helps improve prediction accuracy.

[0077] Step S330: performing feature aggregation processing based on the second adjacency matrix and the second initial feature data to obtain second aggregated feature data of the first monitoring point.

[0078] The second adjacency matrix comprises spatial mode correlations between different monitoring points, according to which the second initial feature data of the monitoring points related to the first monitoring point can be subjected to feature aggregation processing. For example, the second initial feature data of the first monitoring point and the adjacent monitoring points can be aggregated based on the second adjacency matrix, such as being weighted according to the spatial weights in the second adjacency matrix to achieve the aggregation effect, and multi-layer aggregation can be performed to obtain the second aggregated feature data of the first monitoring point.

[0079] In an embodiment, the feature aggregation processing based on the second adjacency matrix and the second initial feature data to obtain the second aggregated feature data of the first monitoring point comprises the following steps:

[0080] generating a second initial feature matrix according to the second initial feature data of the plurality of monitoring points;

[0081] performing one or more layers of feature aggregation processing based on the second adjacency matrix, taking the second initial feature matrix as the starting point, to obtain a second aggregated feature matrix;

[0082] extracting the second aggregated feature data of the first monitoring point from the second aggregated feature matrix.

[0083] wherein the second initial feature data of the plurality of monitoring points is organized into a matrix, such as each row representing the second initial feature data of a monitoring point, forming a second initial feature matrix. For example, the second initial feature data X2 is reconstructed into a row-column form to obtain the second initial feature matrix, denoted as wherein the subscript spa represents the spatial characteristic-related features, and the superscript 0 represents the matrix formed by the initial features.

[0084] The feature aggregation processing can be performed layer by layer through the following formula (4):

[0085]

[0086]

[0087] wherein W spa denotes the second adjacency matrix, I is an identity matrix, used to add self-connection in W spa so that the monitoring point retains its own features when aggregating the information of the adjacent monitoring points. denotes a symmetric normalized degree matrix, composed of a diagonal matrix of the degrees (the number of connections of each node) of . is the second aggregated feature matrix of the first layer, l ​​​is the spatial channel of the graph convolutional network l Layers have learnable parameters, l Can be any non-negative integer. The ReLU activation function in formula (4) can be replaced by other activation functions.

[0088] The process of layer-by-layer feature aggregation in formula (4) can be referred to in the above formula (2) and will not be repeated here. It should be noted that the feature aggregation based on the first adjacency matrix and the first initial feature data is feature aggregation from the mechanical mode. The feature aggregation based on the second adjacency matrix and the second initial feature data is feature aggregation from the spatial mode. The two aggregation processes are similar, but the data processed and the parameters used are different.

[0089] The second aggregated feature data is high-level feature data obtained by aggregating the initial feature data through spatial modality. The second aggregated feature matrix can be regarded as a collection of the second aggregated feature data of all monitoring points. For example, if the first monitoring point includes all monitoring points, the second aggregated feature matrix can be directly used as the second aggregated feature data of the first monitoring point. If the first monitoring point is a specific monitoring point, the second aggregated feature data of the first monitoring point can be extracted from the corresponding row in the second aggregated feature matrix. The second aggregated feature data realizes the representation of monitoring points based on spatial modality.

[0090] This paper constructs a spatial-mechanical spatiotemporal dual-channel graph convolutional network (SMGCN) based on physical constraints. It generates first and second aggregated feature data through two channels. It can fuse the features of the two-channel inputs and combine them with the temporal dependencies of sequence models to achieve dynamic spatiotemporal feature fusion and temporal modeling.

[0091] In the feature aggregation processing of the mechanical channel and the spatial channel, the number of layers is involved. It can be set by human beings. In an embodiment, the maximum and minimum of the mechanical similarity degrees of the plurality of monitoring point pairs can be determined, and the number of layers of the mechanical channel (referred to as the first number of layers) can be determined according to the ratio of the maximum and minimum. For example, the mapping relationship between different ratios and different numbers of layers can be obtained by statistical history data, which can be a segmented mapping relationship. It reflects that under different ratios (i.e. different mechanical similarity degree distributions), a single monitoring point needs to pass through several layers of feature aggregation processing to match the size of its mechanical correlation range. Thus, the first number of layers is obtained by mapping according to the ratio of the maximum and minimum of the mechanical similarity degrees of the plurality of monitoring point pairs. Further, in the feature aggregation processing of the mechanical channel, the feature aggregation processing is performed according to the first number of layers. In an embodiment, the maximum and minimum of the spatial correlation degrees of the plurality of monitoring point pairs can be determined, and the number of layers of the spatial channel (referred to as the second number of layers) can be determined according to the ratio of the maximum and minimum. For example, the mapping relationship between different ratios and different numbers of layers can be obtained by statistical history data, which can be a segmented mapping relationship. It reflects that under different ratios (i.e. different spatial correlation degree distributions), a single monitoring point needs to pass through several layers of feature aggregation processing to match the size of its spatial correlation range. Thus, the second number of layers is obtained by mapping according to the ratio of the maximum and minimum of the spatial correlation degrees of the plurality of monitoring point pairs. Further, in the feature aggregation processing of the spatial channel, the feature aggregation processing is performed according to the second number of layers.

[0092] In an embodiment, the target feature data of the first monitoring point based on the first aggregated feature data comprises: the target feature data of the first monitoring point based on the first aggregated feature data and the second aggregated feature data. For example, the first aggregated feature data and the second aggregated feature data can be spliced to obtain the target feature data, which is represented as If the first initial feature data and the second initial feature data both include all the initial feature data, then d1+d2=2d .

[0093] In an embodiment, the first aggregated feature data and the second aggregated feature data can be weighted and fused to obtain the target feature data. The weights corresponding to the first aggregated feature data and the second aggregated feature data can be set by human beings. Alternatively, the weights can be determined according to the normalized first adjacency matrix and the normalized second adjacency matrix. For example, the sum of the values in the normalized first adjacency matrix is calculated to obtain a mechanical similarity quantization value, the sum of the values in the normalized second adjacency matrix is calculated to obtain a spatial correlation quantization value, and the mechanical similarity quantization value and the spatial correlation quantization value are converted into two weights with a sum of one, which are the weights corresponding to the first aggregated feature data and the second aggregated feature data. The first aggregated feature data and the second aggregated feature data are weighted and fused according to the weights to obtain the target feature data.

[0094] With reference to the foregoing Figure 1 In step S140, the target feature data is processed by using a sequence model to obtain state prediction data of the first monitoring point.

[0095] In the embodiments of the present disclosure, the sequence model includes but is not limited to a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), etc., which can learn sequence information in the data and predict the current or future state based on the sequence information.

[0096] The target feature data can include feature data of the first monitoring point at multiple time points, which is input into the sequence model, and the sequence model can learn the sequence information. Alternatively, the target feature data only includes feature data of the first monitoring point at one time point (such as the latest monitoring time point), the target feature data and the feature data at one or more previous time points are input into the sequence model, so that the sequence model learns the sequence information, or the sequence model has obtained the feature data at one or more previous time points in the previous processing process, the target feature data is input into the sequence model, and the sequence model learns the sequence information by combining the target feature data and the feature data at one or more previous time points.

[0097] The state prediction data of the first monitoring point can include displacement prediction data (such as a predicted displacement amount or a position after displacement) or prediction data of other aspects (such as deformation and geotechnical state).

[0098] In an embodiment, the sequence model includes a gated recurrent unit, and the gated recurrent unit includes an update gate and a reset gate. With reference to the foregoing Figure 4 As shown in the foregoing, the processing of the target feature data by using the sequence model to obtain the state prediction data of the first monitoring point includes the following steps S410 to S450:

[0099] In step S410, the target feature data is input into the gated recurrent unit;

[0100] In step S420, the target feature data and the previous time point hidden state corresponding to the target feature data are weighted by the update gate to obtain update information;

[0101] In step S430, the target feature data and the previous time point hidden state are weighted by the reset gate to obtain reset information;

[0102] In step S440, a candidate hidden state is determined according to the target feature data, the reset information, and the previous time point hidden state;

[0103] Step S450 , weighting the candidate hidden states and the previous hidden state based on the update information to obtain the target hidden state, and obtaining state prediction data based on the target hidden state.

[0104] For example, after inputting the target feature data into the gated recurrent unit, the update gate calculates the update information using the following formula: z t :

[0105] ; (5)

[0106] in, t Indicates the target moment, that is, the moment of the state to be predicted, such as the next moment of the current moment, , represents the input weight matrix of the update gate, which can map the target feature data to the hidden layer (dimension h ), U z To update the historical state weight matrix of the gate, h t-1 is the hidden state at the previous moment, and Sigmoid is an activation function that can compress the output to [0,1], indicating the information retention ratio.

[0107] The reset gate calculates the reset information using the following formula r t :

[0108] ; (6)

[0109] in, W r represents the input weight matrix of the reset gate, U r is the historical state weight matrix of the reset gate.

[0110] The candidate hidden state is then calculated by the following formula:

[0111] ; (7)

[0112] in, represents the candidate hidden state, W h is the input weight matrix of the candidate hidden state, U h Table 1 is the historical state weight matrix of the candidate hidden state, Represents Hadamard product, and tanh is an activation function that can compress the output to [0,1] and enhance nonlinear expression capabilities.

[0113] Finally, the hidden state is updated in a weighted manner and the hidden state at the target moment is output. Refer to the following formula:

[0114] ; (8)

[0115] in, ht The hidden state at the target moment can be used as the final state prediction data, or the hidden state at the target moment can be mapped to the output space through a fully connected layer or other structures to obtain the state prediction data.

[0116] In one embodiment, reference Figure 5 As shown, the slope state prediction method further includes the following steps S510 to S540:

[0117] Step S510: Acquire real sample data based on historical monitoring information of multiple monitoring points.

[0118] The real sample data comes from real data (i.e., actual historical monitoring information). Correspondingly, the synthetic sample data described below is not sample data directly generated from real data, but rather non-real sample data generated through synthesis. In slope monitoring scenarios, real sample data is usually scarce. For example, historical monitoring information for new slopes is scarce, and real sample data generated based on historical monitoring information is also scarce, which is usually insufficient to fully train models (such as convolutional networks and sequence models). Therefore, the disclosed embodiments generate synthetic sample data and train the model using a two-stage training method.

[0119] Step S520, using real sample data to perform the first stage training on the graph convolutional network and the sequence model; the graph convolutional network is a network used for feature aggregation processing.

[0120] The graph convolutional network can be the first or second graph convolutional network described above, or a graph convolutional network with two channels. The first phase of training the graph convolutional network and the sequence model can be performed using real sample data, or only the graph convolutional network or only the sequence model can be trained in the first phase. Due to the small number of real samples, fast training can be achieved.

[0121] In one embodiment, the first stage of training can be performed through meta-learning to preliminarily update parameters and enhance the model's generalization ability, allowing it to adapt quickly under a small number of samples while maintaining the physical consistency of mechanical laws.

[0122] Divide real sample data into multiple meta-tasks , where it represents the training dataset simulating a small sample scenario, Represents the target dataset for the same task.

[0123] In each meta-task based on Calculate the gradient and update the parameters of the model to get the adaptation i Local optimal parameters for each meta-task θ i ':

[0124] ; (9)

[0125] in, θ represents the initial parameters of the dual-channel graph convolutional network and the sequence model, α is the inner learning rate, Representation Model f θ In meta-task The training set The mean square error loss on Represents the loss function on the model parameters θ Optimize the initial parameters on the set of meta-tasks θ , making it suitable for fast training. Thus the first stage of training is achieved.

[0126] Step S530 : adding disturbance information to the real sample data according to the coefficient of variation of the geotechnical parameters of the slope to obtain synthetic sample data, and mixing the real sample data and the synthetic sample data to obtain mixed sample data.

[0127] The coefficient of variation of geotechnical parameters is used to measure the discreteness or uncertainty of the physical and mechanical parameters of the slope, and can include the coefficient of variation of cohesion and the coefficient of variation of the internal friction angle. The coefficient of variation of geotechnical parameters can be obtained by calculating the standard deviation of one or more physical and mechanical parameters, or the ratio of the standard deviation to the mean, based on historical monitoring information.

[0128] Perturbation information is added to the real sample data based on the coefficient of variation of the geotechnical parameters of the slope. For example, the perturbation range can be determined based on the coefficient of variation of the geotechnical parameters, and perturbation information is generated within the perturbation range. This ensures that the perturbation information matches the discreteness or uncertainty level of the real data. This results in synthetic sample data that conforms to the characteristics of the real data, ensuring the quality of the synthetic sample data.

[0129] In one embodiment, the above-mentioned adding disturbance information to the real sample data according to the coefficient of variation of the geotechnical parameters of the slope to obtain the synthetic sample data includes the following steps:

[0130] The following formula (10) is used to add disturbance information to the real sample data to obtain the synthetic sample data:

[0131] ; (10)

[0132] in, represents synthetic sample data, represents the real sample data, ε Represents the disturbance probability, which can be set based on experience or specific business needs. σ represents the standard deviation determined based on the coefficient of variation of geotechnical parameters; Z(0,σ 2 ) The mean is 0 and the variance is σ 2 Gaussian distribution. By using formula (10), perturbation information that satisfies the Gaussian distribution is generated and added to the real sample data to obtain synthetic sample data.

[0133] In one embodiment, after obtaining the synthetic sample data, the slope state prediction method further includes the following steps:

[0134] Determine the corresponding non-true shear strength based on the synthetic sample data;

[0135] If the untrue shear strength is greater than the true shear strength corresponding to the true sample data used to generate the synthetic sample data, the synthetic sample data is deleted.

[0136] Among them, when synthetic sample data is obtained, the shear strength can be calculated according to the mechanical parameters in the synthetic sample data, such as using the formula Perform calculations, τ syn represents the shear strength calculated based on the perturbed mechanical parameters in the synthetic sample data, c syn is the cohesion after disturbance, σ n,syn is the normal stress after disturbance, is the internal friction angle after disturbance. Since the shear strength is not derived from real data, it is called non-real shear strength. The non-real shear strength is compared with the corresponding real shear strength, as shown in formula (10). Calculate the non-true shear strength τ syn ,according to Corresponding Calculate the true shear strength τ real ,Should τ syn and τ real According to the feasible domain of the Mohr-Coulomb criterion, the two should satisfy τ syn ≤τ real ,If the relationship is not satisfied, the corresponding synthetic sample data is deleted and ,regenerated synthetic sample data.

[0137] In this way, the shear strength of the synthetic sample data is effectively limited, which can avoid the occurrence of situations in which the synthetic sample data does not conform to the mechanical properties, thereby ensuring the quality of the synthetic sample data.

[0138] It should be noted that since real sample data may contain multi-dimensional data, when adding perturbation information, perturbation information can be added to the data in all dimensions, or only to the data in some dimensions. For example, perturbation information can be added to variable data in the real sample data (such as data that can be monitored for real-time changes), while perturbation information is not added to non-variable data in the real sample data (such as some fixed geological parameters). The original data is retained, and the non-variable data is combined with the variable data after the perturbation information is added to form synthetic sample data.

[0139] Mix the real sample data and the synthetic sample data to obtain mixed sample data. For example, you can use the mixing ratio λ Mixing λ It is a value in the range of [0,1] and can be set based on experience or specific business needs. Refer to the following formula:

[0140] ; (11)

[0141] in, D real represents the real sample data, D syn represents synthetic sample data, D mix Represents mixed sample data.

[0142] Step S540: Perform the second phase of training on the graph convolutional network and the sequence model using the mixed sample data.

[0143] The second stage of training can update all parameters in the graph convolutional network and sequence model based on the first stage of training, or it can update only some of the parameters. For example, the second stage of training is used to fine-tune some parameters in the graph convolutional network, such as the spatial channel of the graph convolutional network. , the gated recurrent unit W z 、 W h .

[0144] The introduction of synthetic sample data increases the richness of the sample data, enabling the model to fully learn the inherent rules and correlations of the data, making up for the possible lack of training in the first stage.

[0145] In one embodiment, the second stage of training the graph convolutional network and the sequence model using mixed sample data includes the following steps:

[0146] Processing the sample input data in the mixed sample data through the graph convolutional network and the sequence model, and obtaining a first loss value based on the deviation between the sample output data and the label data;

[0147] Determine a second loss value based on the absolute values ​​of the parameters to be fine-tuned in the graph convolutional network and the sequence model;

[0148] The parameter to be fine-tuned is updated based on the first loss value and the second loss value.

[0149] Among them, the parameters to be fine-tuned refer to the parameters that need to be updated in the second stage, such as the spatial channels of the graph convolutional network. , the gated recurrent unit W z 、 W h The loss function used in the second stage can include a first loss term and a second loss term. The first loss term corresponds to the first loss value, which represents the deviation loss of the model prediction, and the second loss term corresponds to the second loss value, which represents the parameter complexity penalty. The loss function K can be referred to as follows:

[0150] ; (12)

[0151] Where N is the number of training samples, For the model i Sample output data for each sample output, For the i The label data of samples (i.e. the true state value); γ is the regularization coefficient, which can be set manually or adjusted by the system; is a set of parameters to be fine-tuned, such as 、 W z 、 W h wait; Represents the sum of the squares of the parameters to be fine-tuned. Fine-tuning the parameters in the model based on the above loss function makes the model output more accurate prediction data as a whole, and tends to simplify parameters, making the model calculation process simpler.

[0152] In one embodiment, a third loss term may be added to the above loss function, which corresponds to a third loss value, representing the mechanical similarity distribution of the monitoring points calculated based on the sample output data output by the model (denoted as Z mech,pred ), and the original mechanical similarity distribution (denoted as Zmech,orig ). Refer to the following formula:

[0153] ; (13)

[0154] in, β is the regularization coefficient, which can be set manually or adjusted by the system. By introducing a third loss term into the loss function, the original mechanical similarity distribution is maintained during the second stage of fine-tuning, avoiding large deviations in model predictions.

[0155] The disclosed embodiment uses a two-stage training method to accelerate the training process at each stage, thereby improving overall training efficiency, achieving better training results, and obtaining a model with better performance parameters.

[0156] Figure 6 The schematic diagram of the architecture of the slope state prediction method of the embodiment of the present disclosure is shown. First, the slope monitoring graph node modeling is performed, and the monitoring points can be used as nodes of the graph. The first adjacency matrix is ​​constructed based on mechanical similarity, and the second adjacency matrix is ​​constructed based on spatial correlation. Then, SMGCN is constructed and fused with GRU to generate a dual-channel time series fusion model (GMGCN-GRU). A two-stage training strategy based on small samples is constructed. For GMGCN-GRU, small sample training and fine-tuning are performed in two stages, so that the model converges quickly and achieves better performance. In the actual prediction process, the initial feature data of the monitoring point is input into GMGCN-GRU, and the slope state prediction result is output.

[0157] The following is an example of a large slope on the right bank of a hydropower station to further illustrate the slope state prediction method. The slope is affected by reservoir water level fluctuations and rainfall, and there is periodic displacement. A total of 12 GNSS surface displacement monitoring stations are deployed on the slope (that is, 12 monitoring points are set up, numbered N1~N12), recording data from 2020 to 2023. The initial characteristic data include six categories: cohesion (unit kPa), internal friction angle (unit degree), terrain slope (unit degree), daily rainfall (unit mm), groundwater level (unit m), cumulative displacement (unit mm), etc., which can be obtained d = 6. The monitoring time step is once a day, and a total of 730 days of data are obtained, of which the training set is the first 600 days of data and the test set is the last 130 days of data.

[0158] Based on the above information, the initial feature data, the first adjacency matrix, and the second adjacency matrix of each monitoring point are constructed. The calculation results of the first adjacency matrix and the second adjacency matrix can be referred to Figure 7 shown.

[0159] Based on the above calculation results, a meta-learning strategy is used to train the SMGCN-GRU model architecture composed of SMGCN and GRU. Table 1 shows some parameters (or hyperparameters) of the model architecture.

[0160] Table 1 Partial parameters of the model architecture

[0161]

[0162] We used all real sample data to train SMGCN-GRU (referred to as full-data SMGCN-GRU), and used real sample data and mixed sample data to train SMGCN-GRU in two stages (referred to as two-stage SMGCN-GRU). We compared the two training methods and compared them with the conventional GCN scheme. We used the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) Verify and compare the prediction accuracy of the above three schemes. The comparison results are shown in Table 2. The two-stage SMGCN-GRU uses 30 days of real sample data from 5 monitoring points.

[0163] Table 2 Comparison of prediction accuracy of three schemes

[0164]

[0165] Figure 8 The total loss function convergence curves of the three schemes are shown. Among them, the two-stage SMGCN-GRU has the fastest convergence speed and the smallest final total loss function, proving that the introduction of meta-learning to optimize the initial parameters can enable the model to converge quickly with a small amount of data, verifying the generalization ability of this training strategy and avoiding overfitting.

[0166] Table 2 shows that the RMSE (1.83 mm) of the two-stage SMGCN-GRU is 38% lower than that of the conventional GCN scheme, indicating that the introduction of physical constraints significantly reduces model error. This demonstrates that by constructing a first adjacency matrix based on mechanical correlations, the influence of geotechnical parameters (such as cohesion and internal friction angle) on displacement can be dynamically captured, reducing reliance on empirical assumptions. The RMSE (2.14 mm) of the two-stage SMGCN-GRU is 14% lower than that of the full-data SMGCN-GRU, and the MAE is 27% lower than that of the conventional GCN scheme. The R² of the small-sample SMGCN-GRU is 0.95, indicating that the model explains 95% of the displacement variation, with the remaining 5% likely due to unmonitored factors (such as local seepage paths). The R² of the conventional GCN scheme is 0.89, indicating its limited ability to model complex mechanical mechanisms.

[0167] In the disclosed embodiments, the two-stage SMGCN-GRU scheme can achieve high-precision displacement prediction of slopes with a small number of samples through meta-learning and physical enhancement, reduce dependence on dense monitoring networks, and maintain high generalization in data-scarce scenarios (such as only 5 monitoring points), providing a high-precision, low-dependence, and explainable intelligent prediction solution for slope status.

[0168] The embodiment of the present disclosure also provides a slope state prediction device based on physical constraints. Figure 9 As shown, the slope state prediction device 900 may include the following modules:

[0169] The initial feature data acquisition module 910 is configured to acquire first initial feature data of a plurality of monitoring points on the slope based on monitoring information of the plurality of monitoring points;

[0170] The adjacency matrix construction module 920 is configured to determine the mechanical similarity of a plurality of monitoring point pairs and construct a first adjacency matrix of the plurality of monitoring points according to the mechanical similarity; each monitoring point pair includes two monitoring points;

[0171] a feature aggregation processing module 930 configured to perform feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and to obtain target feature data of the first monitoring point based on the first aggregated feature data;

[0172] The prediction processing module 940 is configured to process the target feature data using a sequence model to obtain the state prediction data of the first monitoring point.

[0173] In one embodiment, determining the mechanical similarity of multiple monitoring point pairs includes: determining the shear strength based on the mechanical parameters of two monitoring points in the monitoring point pair, and obtaining the mechanical similarity of the monitoring point pair based on the difference in shear strength of the two monitoring points; the mechanical similarity is negatively correlated with the difference in shear strength.

[0174] In one embodiment, determining the shear strength based on the mechanical parameters of two monitoring points in the monitoring point pair, and obtaining the mechanical similarity of the monitoring point pair based on the difference in the shear strength of the two monitoring points, includes: calculating the mechanical similarity of the monitoring point pair using the following formula:

[0175] ;

[0176] ;

[0177] in, i 、 j represents the two monitoring points in the monitoring point pair, W mech (i,j) represents the mechanical similarity of the monitoring point pair, τ represents the shear strength, c Indicates cohesion, represents the internal friction angle, σ n represents the normal stress.

[0178] In one embodiment, the feature aggregation processing is performed based on the first adjacency matrix and the first initial feature data to obtain the first aggregated feature data of the first monitoring point, including: generating a first initial feature matrix based on the first initial feature data of the multiple monitoring points; starting with the first initial feature matrix, performing one or more layers of feature aggregation processing based on the first adjacency matrix to obtain a first aggregated feature matrix; and extracting the first aggregated feature data of the first monitoring point from the first aggregated feature matrix.

[0179] In one embodiment, the device is further configured to: obtain second initial feature data of the multiple monitoring points based on the monitoring information of the multiple monitoring points; determine the spatial correlation of multiple monitoring point pairs, and construct a second adjacency matrix of the multiple monitoring points based on the spatial correlation; perform feature aggregation processing based on the second adjacency matrix and the second initial feature data to obtain second aggregated feature data of the first monitoring point; obtaining the target feature data of the first monitoring point based on the first aggregated feature data includes: obtaining the target feature data of the first monitoring point based on the first aggregated feature data and the second aggregated feature data.

[0180] In one embodiment, determining the spatial correlation of the plurality of monitoring point pairs includes: referring to the following formula to determine the spatial correlation of the monitoring point pair according to the distance between two monitoring points in the monitoring point pair;

[0181] ;

[0182] in, i 、 j represents the two monitoring points in the monitoring point pair, d ij Indicates monitoring point i and j distance, W spa (i,j) represents the spatial correlation of the monitoring point pair, and σ is the Gaussian kernel parameter.

[0183] In one embodiment, the feature aggregation processing is performed based on the second adjacency matrix and the second initial feature data to obtain the second aggregated feature data of the first monitoring point, including: generating a second initial feature matrix based on the second initial feature data of the multiple monitoring points; starting with the second initial feature matrix, performing one or more layers of feature aggregation processing based on the second adjacency matrix to obtain a second aggregated feature matrix; and extracting the second aggregated feature data of the first monitoring point from the second aggregated feature matrix.

[0184] In one embodiment, the sequence model includes a gated recurrent unit, which includes an update gate and a reset gate; the processing of the target feature data using the sequence model to obtain state prediction data of the first monitoring point includes: inputting the target feature data into the gated recurrent unit; weighting the target feature data and the hidden state at the previous moment corresponding to the target feature data through the update gate to obtain update information; weighting the target feature data and the hidden state at the previous moment through the reset gate to obtain reset information; determining a candidate hidden state based on the target feature data, the reset information, and the hidden state at the previous moment; referring to the following formula, weighting the candidate hidden state and the hidden state at the previous moment based on the update information to obtain a target moment hidden state, and obtaining the state prediction data based on the target moment hidden state;

[0185] ;

[0186] in, h t Hidden state for the target moment, z t To update information, represents the candidate hidden state, h t-1 is the hidden state at the previous moment, represents the Hadamard product.

[0187] In one embodiment, the device is further configured to: obtain real sample data based on the historical monitoring information of the multiple monitoring points; use the real sample data to perform the first stage training of the graph convolution network and the sequence model; the graph convolution network is a network for performing feature aggregation processing; add disturbance information to the real sample data based on the coefficient of variation of the geotechnical parameters of the slope to obtain synthetic sample data, and mix the real sample data and the synthetic sample data to obtain mixed sample data; use the mixed sample data to perform the second stage training of the graph convolution network and the sequence model.

[0188] In one embodiment, adding disturbance information to the real sample data according to the coefficient of variation of the geotechnical parameters of the slope to obtain synthetic sample data includes: adding disturbance information to the real sample data according to the following formula to obtain synthetic sample data:

[0189] ;

[0190] in, represents synthetic sample data, represents the real sample data, ε represents the perturbation probability, σ represents the standard deviation determined based on the coefficient of variation of geotechnical parameters; Z(0,σ 2 ) The mean is 0 and the variance is σ 2 Gaussian distribution.

[0191] In one embodiment, the second stage training of the graph convolutional network and the sequence model using the mixed sample data includes: processing the sample input data in the mixed sample data through the graph convolutional network and the sequence model, and obtaining a first loss value based on the deviation between the sample output data and the label data; determining a second loss value based on the absolute values ​​of the parameters to be fine-tuned in the graph convolutional network and the sequence model; and updating the parameters to be fine-tuned based on the first loss value and the second loss value.

[0192] In one embodiment, after obtaining the synthetic sample data, the device is further configured to: determine the corresponding non-real shear strength based on the synthetic sample data; if the non-real shear strength is greater than the real shear strength corresponding to the real sample data used to generate the synthetic sample data, delete the synthetic sample data.

[0193] The specific details of each part of the above-mentioned device have been described in detail in the implementation method part. The undisclosed details can be found in the implementation method part, so they will not be repeated here.

[0194] It should be noted that although several modules or units of the device for action execution are mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be concretized in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be concretized.

[0195] The embodiments of the present disclosure further provide a computer program product, which includes a computer program, and implements the above method when the computer program is executed by a processor.

[0196] In one embodiment, a computer program product may be a tangible product containing a computer program, such as a computer-readable storage medium storing the computer program. The computer-readable storage medium may be a storage medium based on electrical, magnetic, optical, electromagnetic, infrared, or other signals, including but not limited to random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory, hard disk drive (HDD), solid state drive (SSD), and the like. Exemplarily, the computer program product may be implemented as a non-volatile storage medium storing the computer program, such as a read-only memory (ROM) or NAND flash memory.

[0197] In one embodiment, the computer program product may be an intangible product containing a computer program. For example, the computer program product may be implemented as a virtual digital product, such as an executable file or installation package storing the computer program.

[0198] The code of a computer program can be written in one or more programming languages, such as C, Java, C++, etc. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device can be connected to the user's computing device via any type of network, such as a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (e.g., via an internet connection provided by a carrier).

[0199] Computer programs can be carried or transmitted through electrical, magnetic, optical, electromagnetic, infrared, and other signals. Electronic devices can convert signals carrying computer programs into digital signals, thereby running the computer programs. When a computer program is run on an electronic device, its code is used to enable the electronic device to execute (more specifically, it can enable the processor of the electronic device to execute) the method steps of various exemplary embodiments of the present disclosure, for example: Step S110, obtaining first initial feature data of multiple monitoring points based on monitoring information of multiple monitoring points on the slope; Step S120, determining the mechanical similarity of multiple monitoring point pairs, and constructing a first adjacency matrix of multiple monitoring points based on the mechanical similarity; each monitoring point pair includes two monitoring points; Step S130, performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and obtaining target feature data of the first monitoring point based on the first aggregated feature data; Step S140, processing the target feature data using a sequence model to obtain state prediction data of the first monitoring point.

[0200] The above method is implemented based on a computer program. On the one hand, for monitoring points on the slope, the initial feature data is aggregated by constructing an adjacency matrix to obtain aggregated feature data. This allows the associations between different monitoring points to be learned. The target feature data is then processed using a sequence model, combining the temporal and spatial associations of the monitoring points to improve the accuracy of the slope state prediction results. On the other hand, physical constraints are introduced to determine the mechanical similarity between different monitoring points and construct a first adjacency matrix. This allows the mechanical modal correlations between the monitoring points to be captured during feature aggregation, ensuring that the slope state prediction results meet geotechnical criteria (such as shear strength limits). Compared to traditional data-driven models, this solution can reduce non-physical deviations in the prediction results and enhance the physical consistency of the prediction results, thereby further improving prediction accuracy. It is particularly suitable for dealing with complex geological conditions.

[0201] Embodiments of the present disclosure also provide an electronic device. The electronic device may include a processor and a memory. The memory stores executable instructions for the processor, such as a computer program. The processor executes the executable instructions to perform the method steps of various exemplary embodiments of the present disclosure.

[0202] Reference below Figure 10 , the electronic device is exemplarily described in the form of a general-purpose computing device. It should be understood that Figure 10 The electronic device 1000 shown is merely an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0203] like Figure 10As shown, the electronic device 1000 may include: a processor 1010 , a memory 1020 , a bus 1030 , an I / O (input / output) interface 1040 , and a network adapter 1050 .

[0204] Memory 1020 may include volatile memory, such as RAM 1021 and cache unit 1022, and non-volatile memory, such as ROM 1023. Memory 1020 may also include one or more program modules 1024. Such program modules 1024 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. For example, program modules 1024 may include the modules described above.

[0205] The processor 1010 may include one or more processing units. For example, the processor 1010 may include an AP (Application Processor), a modem processor, a GPU (Graphics Processing Unit), an ISP (Image Signal Processor), a controller, an encoder, a decoder, a DSP (Digital Signal Processor), a baseband processor and / or an NPU (Neural-Network Processing Unit) and other processing units.

[0206] The processor 1010 can be used to execute executable instructions stored in the memory 1020, which may include method steps of various exemplary embodiments of the present disclosure, for example: step S110, obtaining first initial feature data of multiple monitoring points based on monitoring information of multiple monitoring points on the slope; step S120, determining the mechanical similarity of multiple monitoring point pairs, and constructing a first adjacency matrix of multiple monitoring points based on the mechanical similarity; each monitoring point pair includes two monitoring points; step S130, performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and obtaining target feature data of the first monitoring point based on the first aggregated feature data; step S140, using a sequence model to process the target feature data to obtain state prediction data of the first monitoring point, which is particularly suitable for processing complex geological conditions.

[0207] Based on the execution of the above method by processor 1010, on the one hand, for the monitoring points on the slope, feature aggregation processing is performed on the initial feature data by constructing an adjacency matrix to obtain aggregated feature data. This can learn the associations between different monitoring points, and use a sequence model to process the target feature data, combining the spatiotemporal associations of the monitoring points to improve the accuracy of the slope state prediction results. On the other hand, physical constraints are introduced to determine the mechanical similarity between different monitoring points and construct a first adjacency matrix. In this way, the mechanical modal correlation between the monitoring points can be captured during feature aggregation processing, so that the slope state prediction results meet geotechnical criteria (such as shear strength limits). Compared with traditional data-driven models, this solution can reduce non-physical deviations in the prediction results and improve the physical consistency of the prediction results, thereby further improving the prediction accuracy.

[0208] The bus 1030 is used to realize the connection between different components of the electronic device 1000 and may include a data bus, an address bus, and a control bus.

[0209] The electronic device 1000 can communicate with one or more external devices 1100 (eg, a keyboard, a mouse, an external controller, etc.) through the I / O interface 1040 .

[0210] The electronic device 1000 can communicate with one or more networks via the network adapter 1050. For example, the network adapter 1050 can provide mobile communication solutions such as 3G / 4G / 5G, or wireless communication solutions such as wireless LAN, Bluetooth, and near-field communication. The network adapter 1050 can communicate with other modules of the electronic device 1000 via the bus 1030.

[0211] although Figure 10 Not shown, other hardware and / or software modules may also be provided in the electronic device 1000, including but not limited to: a display, microcode, device drivers, redundant processors, an external disk drive array, a tape drive, and a data backup storage system.

[0212] As can be seen from the above, the technical solutions of the present disclosure can be implemented as methods, devices, systems, computer program products, storage media, electronic devices, etc. Those skilled in the art will appreciate that various aspects of the present disclosure can be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation that combines hardware and software aspects, such as "circuits," "modules," or "systems," respectively.

[0213] It should be understood that the present disclosure is not limited to the specific method steps or structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. Those skilled in the art will easily think of other embodiments based on the specific embodiments provided by the present disclosure. Therefore, the specific embodiments provided by the present disclosure are merely exemplary, and the scope and spirit of the present disclosure are indicated by the claims, which should cover any variations, uses or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary technical means in the field of the present technology that are not disclosed in the present disclosure.

Claims

1. A slope state prediction method based on physical constraints, characterized in that: The method comprises: Acquire first initial feature data of the plurality of monitoring points on the slope according to monitoring information of the plurality of monitoring points; Determining mechanical similarities of a plurality of monitoring point pairs, and constructing a first adjacency matrix of the plurality of monitoring points according to the mechanical similarities, wherein each monitoring point pair includes two monitoring points; Performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and obtaining target feature data of the first monitoring point based on the first aggregated feature data; Processing the target feature data using a sequence model to obtain state prediction data of the first monitoring point; The step of performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point includes: generating a first initial feature matrix based on the first initial feature data of the plurality of monitoring points; starting with the first initial feature matrix, performing one or more layers of feature aggregation processing based on the first adjacency matrix to obtain a first aggregated feature matrix; and extracting the first aggregated feature data of the first monitoring point from the first aggregated feature matrix. The method further includes: obtaining second initial feature data of the plurality of monitoring points based on monitoring information of the plurality of monitoring points; determining spatial correlations of the plurality of monitoring point pairs, and constructing a second adjacency matrix of the plurality of monitoring points based on the spatial correlations; performing feature aggregation processing based on the second adjacency matrix and the second initial feature data to obtain second aggregated feature data of the first monitoring point; The obtaining the target feature data of the first monitoring point based on the first aggregate feature data includes: obtaining the target feature data of the first monitoring point based on the first aggregate feature data and the second aggregate feature data.

2. The method according to claim 1, characterized in that Determining the mechanical similarity of multiple monitoring point pairs includes: The shear strength is determined based on the mechanical parameters of the two monitoring points in the monitoring point pair, and the mechanical similarity of the monitoring point pair is obtained based on the difference in the shear strength of the two monitoring points; the mechanical similarity is negatively correlated with the difference in shear strength.

3. The method according to claim 2, characterized in that The determining of the shear strength according to the mechanical parameters of two monitoring points in the monitoring point pair, and obtaining the mechanical similarity of the monitoring point pair according to the difference in the shear strength of the two monitoring points, includes: The mechanical similarity of the monitoring point pair is calculated using the following formula: ； ； in, i 、 j represents the two monitoring points in the monitoring point pair, W mech (i,j) represents the mechanical similarity of the monitoring point pair, τ represents the shear strength, τ i , τ j Indicates monitoring point i , monitoring points j The shear strength, c represents cohesion, represents the internal friction angle, σ n represents the normal stress.

4. The method according to claim 1, wherein Determining the spatial correlation of multiple monitoring point pairs includes: Refer to the following formula to determine the spatial correlation of the monitoring point pair based on the distance between the two monitoring points in the monitoring point pair; ； in, i 、 j represents the two monitoring points in the monitoring point pair, d ij Indicates monitoring point i and j distance, W spa (i,j) represents the spatial correlation of the monitoring point pair, and σ is the Gaussian kernel parameter.

5. The method according to claim 1, wherein The performing feature aggregation processing based on the second adjacency matrix and the second initial feature data to obtain second aggregated feature data of the first monitoring point includes: generating a second initial feature matrix according to the second initial feature data of the plurality of monitoring points; Starting with the second initial feature matrix, performing one or more layers of feature aggregation processing based on the second adjacency matrix to obtain a second aggregated feature matrix; Second aggregated feature data of the first monitoring point is extracted from the second aggregated feature matrix.

6. The method according to claim 1, characterized in that The sequence model includes a gated recurrent unit, and the gated recurrent unit includes an update gate and a reset gate; and the target feature data is processed using the sequence model to obtain the state prediction data of the first monitoring point, including: Inputting the target feature data into the gated recurrent unit; Weighting the target feature data and the previous hidden state corresponding to the target feature data through the update gate to obtain update information; Weighting the target feature data and the hidden state at the previous moment through the reset gate to obtain reset information; Determining a candidate hidden state based on the target feature data, the reset information, and the previous hidden state; Referring to the following formula, weighting the candidate hidden state and the previous hidden state based on the update information to obtain a target hidden state, and obtaining the state prediction data based on the target hidden state; ； in, h t Hidden state for the target moment, z t To update information, it represents the candidate hidden state, is the hidden state at the previous moment, representing the Hadamard product.

7. The method according to any one of claims 1 to 3, 5 to 6, characterized in that The method further comprises: Acquiring real sample data based on historical monitoring information of the multiple monitoring points; Performing a first-stage training on a graph convolutional network and the sequence model using the real sample data; the graph convolutional network is a network used for feature aggregation processing; adding disturbance information to the real sample data according to the coefficient of variation of the geotechnical parameters of the slope to obtain synthetic sample data, and mixing the real sample data and the synthetic sample data to obtain mixed sample data; The mixed sample data is used to perform a second phase of training on the graph convolutional network and the sequence model.

8. The method according to claim 7, characterized in that Adding disturbance information to the real sample data according to the coefficient of variation of the geotechnical parameters of the slope to obtain synthetic sample data includes: The following formula is used to add disturbance information to the real sample data to obtain synthetic sample data: ； Among them, represents the synthetic sample data, represents the real sample data, ε represents the perturbation probability, σ represents the standard deviation determined based on the coefficient of variation of geotechnical parameters; Z(0,σ 2 ) The mean is 0 and the variance is σ 2 Gaussian distribution.

9. The method according to claim 7, characterized in that After obtaining the synthesized sample data, the method further includes: determining a corresponding non-real shear strength according to the synthetic sample data; If the non-real shear strength is greater than the real shear strength corresponding to the real sample data used to generate the synthetic sample data, the synthetic sample data is deleted.

10. The method according to claim 7, characterized in that The second stage training of the graph convolutional network and the sequence model using the mixed sample data includes: Processing the sample input data in the mixed sample data by using the graph convolutional network and the sequence model, and obtaining a first loss value according to a deviation between the sample output data and the label data; Determining a second loss value based on absolute values ​​of parameters to be fine-tuned in the graph convolutional network and the sequence model; The parameter to be fine-tuned is updated based on the first loss value and the second loss value.

11. A slope state prediction device based on physical constraints, characterized in that: The device comprises: An initial feature data acquisition module is configured to acquire first initial feature data of a plurality of monitoring points on the slope based on monitoring information of the plurality of monitoring points; an adjacency matrix construction module configured to determine mechanical similarities of a plurality of monitoring point pairs and construct a first adjacency matrix of the plurality of monitoring points according to the mechanical similarities; each monitoring point pair includes two monitoring points; a feature aggregation processing module configured to perform feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point, and obtain target feature data of the first monitoring point based on the first aggregated feature data; a prediction processing module configured to process the target feature data using a sequence model to obtain state prediction data of the first monitoring point; The step of performing feature aggregation processing based on the first adjacency matrix and the first initial feature data to obtain first aggregated feature data of the first monitoring point includes: generating a first initial feature matrix based on the first initial feature data of the plurality of monitoring points; starting with the first initial feature matrix, performing one or more layers of feature aggregation processing based on the first adjacency matrix to obtain a first aggregated feature matrix; and extracting the first aggregated feature data of the first monitoring point from the first aggregated feature matrix. The apparatus is further configured to: obtain second initial feature data of the plurality of monitoring points based on monitoring information of the plurality of monitoring points; determine spatial correlations of the plurality of monitoring point pairs, and construct a second adjacency matrix of the plurality of monitoring points based on the spatial correlations; perform feature aggregation processing based on the second adjacency matrix and the second initial feature data to obtain second aggregated feature data of the first monitoring point; The obtaining the target feature data of the first monitoring point based on the first aggregate feature data includes: obtaining the target feature data of the first monitoring point based on the first aggregate feature data and the second aggregate feature data.

12. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 10 when the computer program is executed by a processor.

13. An electronic device, characterized in that: include: processor; as well as a memory for storing executable instructions of the processor; The processor is configured to perform the method according to any one of claims 1 to 10 by executing the executable instructions.

Images