Rock-soil mass deformation identification method and device based on neural network and electronic equipment

Abstract

The application provides a rock-soil deformation identification method and device based on a neural network and electronic equipment, and relates to the technical field of artificial intelligence. The method realizes double protection of physical consistency and data authenticity by pre-constructing a multi-source rock-soil fusion tensor, and realizes feature identification with physical and mechanical consistency and real field data by matching the corresponding multi-source rock-soil fusion tensor according to the environmental factor data corresponding to the monitoring data of the rock-soil body. Moreover, the method forms a complete feature system covering the mechanical nature, real-time response and time-space evolution by extracting the branch features, strain sensitive features and multi-scale space-time features of the data, and comprehensively captures the nature of rock-soil deformation. Based on the self-attention mechanism, the above features are classified in cooperation, which can accurately determine the deformation mode in the deformation identification scene of different geological conditions, and realize rock-soil deformation analysis with the advantages of physical consistency and data driving.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a rock-soil mass deformation identification method and device based on a neural network and an electronic device. BACKGROUND

[0002] In the field of geotechnical engineering, the deformation and failure laws of rock-soil mass are highly dependent on engineering activities such as slope stability, tunnel excavation, foundation bearing, and underground space development. Traditional research methods mainly rely on numerical simulation and field monitoring data for analysis, but both have significant limitations. Numerical simulation can describe the overall deformation process of rock-soil mass from a theoretical and mechanical perspective, but its parameters are often idealized and difficult to fully reflect the actual situation under complex environmental conditions. While monitoring data can accurately reflect the on-site conditions, they often have problems such as locality, noise interference, and uneven sampling. How to organically integrate the two and form a rock-soil mass deformation analysis that combines the advantages of physical consistency and data-driven is a key scientific and engineering problem that needs to be solved. SUMMARY

[0003] The present application relates to the technical field of artificial intelligence, and in particular to a rock-soil mass deformation identification method and device based on a neural network and an electronic device.

[0004] In the first aspect, the present application provides a rock-soil mass deformation identification method based on a neural network, which includes: performing data monitoring on a target rock-soil mass to obtain monitoring data of the target rock-soil mass and environmental factor data of the environment in which the monitoring data is located; determining a multi-source rock-soil fusion tensor corresponding to the monitoring data from a pre-constructed multi-source rock-soil fusion tensor set based on the environmental factor data; the multi-source rock-soil fusion tensor set is constructed based on monitoring data, simulation data, and environmental factor data of a pre-set rock-soil mass; extracting branch features, strain-sensitive features, and multi-scale spatio-temporal features corresponding to the multi-source rock-soil fusion tensor; the branch features include elastic features and plastic features; based on a pre-constructed self-attention mechanism, the branch features, strain-sensitive features, and multi-scale spatio-temporal features are classified in coordination to determine a deformation mode corresponding to the target rock-soil mass; and performing deformation analysis on the target rock-soil mass based on the deformation mode.

[0005] With reference to the first aspect, in a first implementation form of the first aspect, the method further comprises: obtaining a numerical simulation tensor, a monitoring tensor and a corresponding environmental factor tensor pre-collected for the preset rock-soil mass; performing linear combination on the numerical simulation tensor, the monitoring tensor and the environmental factor tensor to generate an interaction tensor of the preset rock-soil mass; performing dimension unification on the numerical simulation tensor, the monitoring tensor and the environmental factor tensor, and performing data fusion on the dimension-unified numerical simulation tensor, monitoring tensor and environmental factor tensor and the interaction tensor to generate a multi-source rock-soil fusion tensor set of the preset rock-soil mass.

[0006] With reference to the first aspect, in a second implementation form of the first aspect, the step of extracting the branch feature of the multi-source rock-soil fusion tensor comprises: performing first convolution processing on the multi-source rock-soil fusion tensor by using a preset neural network model to extract an elastic feature tensor corresponding to the multi-source rock-soil fusion tensor; performing second convolution processing on the monitoring data by using the preset neural network model to determine a gating weight corresponding to the multi-source rock-soil fusion tensor; performing range compression on the elastic feature tensor, and performing gating selection on the range-compressed elastic feature tensor based on the gating weight to determine a plastic feature tensor of the multi-source rock-soil fusion tensor; and performing feature splicing on the elastic feature tensor and the plastic feature tensor to determine the branch feature of the multi-source rock-soil fusion tensor. With reference to the first aspect, in a third implementation form of the first aspect, the step of extracting the strain-sensitive feature of the multi-source rock-soil fusion tensor comprises: extracting a strain component of the multi-source rock-soil fusion tensor, determining a gradient of the strain component in a spatial direction and a time dimension to generate a strain gradient vector; performing importance identification on a position feature of the strain gradient vector to determine a strain-sensitive weight corresponding to the position feature; and performing feature weighting on the multi-source rock-soil fusion tensor by using the strain-sensitive weight to determine a strain-sensitive feature of the multi-source rock-soil fusion tensor. With reference to the first aspect, in a fourth implementation form of the first aspect, the step of extracting the multi-scale spatio-temporal feature of the multi-source rock-soil fusion tensor comprises: performing a cavity convolution operation on the multi-source rock-soil fusion tensor by using a plurality of preset cavity convolution branches in parallel and adopting different expansion rates respectively to capture spatio-temporal features of the multi-source rock-soil fusion tensor at different scales; and performing summation processing on the spatio-temporal features to obtain a multi-scale spatio-temporal feature of the multi-source rock-soil fusion tensor.

[0007] With reference to the first aspect, in a fifth implementation form of the first aspect, in the step of determining the deformation mode corresponding to the target rock-soil mass based on the pre-constructed self-attention mechanism, the branch feature, the strain-sensitive feature and the multi-scale space-time feature are cooperatively classified, the step includes: converting the branch feature, the strain-sensitive feature and the multi-scale space-time feature into a global feature vector, mapping the global feature vector to a query space to generate a query representation matrix; performing dot product processing on the query representation matrix by using a scaled dot-product attention mechanism to determine a mode correlation graph corresponding to the global feature vector; the mode correlation graph includes deformation mode similarity between samples; performing matrix multiplication processing on the mode correlation graph and the global feature vector to generate a fusion output, and determining the deformation mode corresponding to the target rock-soil mass based on the fusion output.

[0008] With reference to the first aspect, in a sixth implementation form of the first aspect, the method further includes: extracting the branch feature, the strain-sensitive feature and the multi-scale space-time feature corresponding to the multi-source rock-soil fusion tensor by using a preset neural network model; the preset neural network model is trained based on a preset physical information hybrid loss function; the method further includes: training the preset neural network model by using a preset training sample set to determine an initial prediction output; calculating a cross-entropy loss of the preset neural network model based on the initial prediction output, and generating a dynamic stiffness matrix according to the initial prediction output; the dynamic stiffness matrix is determined based on a material constitutive relationship corresponding to the training sample and a finite element discretization manner of the preset neural network model; constructing a physical constraint residual loss of the preset neural network model based on the stiffness matrix; performing weighted fusion on the physical constraint residual loss and the cross-entropy loss to generate the physical information hybrid loss function. With reference to the first aspect, in a seventh implementation form of the first aspect, the method further includes: obtaining a pre-constructed multi-source rock-soil fusion tensor training sample; performing physical constraint on the multi-source rock-soil fusion tensor training sample based on a linear physical response of a rock-soil mass corresponding to the multi-source rock-soil fusion tensor training sample to generate a physically constrained tensor; constructing the training sample set based on the physically constrained tensor.

[0009] In a second aspect, an embodiment of the present application provides a rock-soil deformation identification device based on a neural network, which comprises: a data monitoring module configured to monitor a target rock-soil body to obtain monitoring data of the target rock-soil body and environmental factor data of an environment in which the monitoring data is located; a data processing module configured to determine a multi-source rock-soil fusion tensor corresponding to the monitoring data from a pre-constructed multi-source rock-soil fusion tensor set based on the environmental factor data; the multi-source rock-soil fusion tensor set is constructed based on monitoring data, simulation data, and environmental factor data of a preset rock-soil body; a feature extraction module configured to extract branch features, strain-sensitive features, and multi-scale spatio-temporal features corresponding to the multi-source rock-soil fusion tensor; the branch features include elastic features and plastic features; an execution module configured to perform collaborative classification on the branch features, the strain-sensitive features, and the multi-scale spatio-temporal features based on a pre-constructed self-attention mechanism to determine a deformation mode corresponding to the target rock-soil body; and an output module configured to perform deformation analysis on the target rock-soil body based on the deformation mode.

[0010] In a third aspect, an embodiment of the present application provides an electronic device, which comprises a processor and a memory, the memory storing machine executable instructions capable of being executed by the processor, and the processor executes the machine executable instructions to implement the method of any of the above embodiments.

[0011] The embodiments of the present application have the following beneficial effects: The embodiments of the present application provide a rock-soil deformation identification method and device based on a neural network and an electronic device, pre-constructed multi-source rock-soil fusion tensors are used to realize double protection of physical consistency and data authenticity, and corresponding multi-source rock-soil fusion tensors are matched according to environmental factor data corresponding to monitoring data of a rock-soil body to realize feature identification with physical and mechanical consistency and real field data. In addition, branch features, strain-sensitive features, and multi-scale spatio-temporal features of the data are extracted to form a complete feature system covering the mechanical essence, real-time response, and spatio-temporal evolution, the deformation essence of the rock-soil body is comprehensively captured, and the deformation essence is comprehensively described. Collaborative classification of the features based on a self-attention mechanism can accurately determine the deformation mode in the deformation identification scene of different geological conditions (clay, rock mass, and sand pebble), and realize rock-soil deformation analysis with the advantages of physical consistency and data driving.

[0012] Other features and advantages of the present application will be described in the following description, or can be inferred from the description or determined without doubt, or can be known by implementing the above-mentioned technologies of the present application. In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the drawings required to be used in the specific embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0014] Figure 1 A flow chart of a rock-soil body deformation identification method based on a neural network provided by an embodiment of the present application is shown in the figure.

[0015] Figure 2 A flow chart of another rock-soil body deformation identification method based on a neural network provided by an embodiment of the present application is shown in the figure.

[0016] Figure 3 A comparison diagram of identification accuracies of different methods provided by an embodiment of the present application is shown in the figure.

[0017] Figure 4 A comparison diagram of prediction errors provided by an embodiment of the present application is shown in the figure.

[0018] Figure 5 A diagram of prediction error distribution of different methods in each strain component provided by an embodiment of the present application is shown in the figure.

[0019] Figure 6 A diagram of the influence of physical constraint strength on identification performance provided by an embodiment of the present application is shown in the figure.

[0020] Figure 7 A diagram of performance changes of different methods under gradually increasing noise levels provided by an embodiment of the present application is shown in the figure.

[0021] Figure 8 A structural diagram of a rock-soil body deformation identification device based on a neural network provided by an embodiment of the present application is shown in the figure.

[0022] Figure 9 A structural diagram of an electronic device provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0023] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0024] The embodiments of the present application provide a rock-soil mass deformation identification method and device based on a neural network, which can realize rock-soil mass deformation analysis with physical consistency and data-driven advantages.

[0025] For ease of understanding, first, a rock-soil mass deformation identification method based on a neural network provided by the embodiments of the present application is described with reference to Figure 1 The method comprises the following steps:

[0026] In step S102, data monitoring is performed on the target rock-soil mass to obtain monitoring data of the target rock-soil mass and environmental factor data of the environment in which the monitoring data is located.

[0027] By monitoring the target rock-soil mass for which deformation needs to be identified, the real state of the rock-soil mass can be reflected through the monitoring data, and then the spatial displacement and deformation degree of the rock-soil mass can be determined. In addition, the external influence data corresponding to the environment in which the rock-soil mass is located, i.e., the environmental factor data, is also determined synchronously, and the spatio-temporal matching between the environmental factor data and the monitoring data is established, which facilitates accurate classification of deformation patterns under deformation factors. The monitoring data can be collected by sensors deployed on site, including the orientation angle and burial depth of a strain gauge (determined by installation records), temperature (measured by a temperature sensor), moisture content (measured by a humidity sensor), strain value (x-direction normal strain, y-direction normal strain, z-direction normal strain, xy shear strain), and displacement (x-direction displacement, y-direction displacement, z-direction displacement). These data are obtained by real-time monitoring of strain gauges and displacement sensors. The environmental factor data is obtained from external observation equipment, including rainfall (recorded by a weather station), groundwater level (obtained from water level monitoring well data), and vibration intensity (measured by a vibration sensor). The attributes of the above-mentioned multi-source data can be processed by a normalization method to eliminate dimensions, or the values can be collected directly according to specified units, and then the units are removed uniformly to maintain dimensional consistency.

[0028] Step S104, based on the environmental factor data, determine the multi-source rock-soil fusion tensor corresponding to the monitoring data from the pre-constructed multi-source rock-soil fusion tensor set.

[0029] The conventional method identifies the deformation of the rock-soil body by establishing theoretical numerical simulation data or by field monitoring data, and often cannot accurately analyze the deformation mode due to differences in environmental conditions. The embodiment of the present application pre-establishes the organic fusion data of theoretical simulation data, field monitoring data and environmental influence data (i.e. multi-source rock-soil fusion tensor set), fuses the field monitoring data with the theory and historical field data in a unified environmental dimension, retains the authenticity of the monitoring data, and makes up for the locality of the monitoring through the simulation data, solving the defects of idealization or localization of single data.

[0030] Further, the environmental factor data is used as the matching basis to screen out the fusion tensor highly consistent with the field environment (such as actual rainfall, groundwater level depth) from the tensor set, so as to perform mode recognition under the real deformation mechanism based on the multi-source rock-soil fusion tensor covering the mechanical nature, external driving and field performance of the deformation, and improve the efficiency and accuracy of the deformation cause tracing and mode recognition. The numerical simulation data can be generated by setting the numerical simulation software (such as finite element analysis tool), including grid size, element type, elastic modulus, Poisson's ratio, cohesion, internal friction angle, permeability coefficient and load step length, etc. These properties are derived from material testing or literature values, and the rock-soil body response under different working conditions is simulated by software running. Further, the collected data can be used to construct a database, which can be used for model training. Each sample corresponds to an independent rock-soil body case, including N samples and T time step time series data. Each sample can be assigned a deformation mode class label, including elastic deformation, elastic-plastic deformation, creep and shear failure four categories, which correspond to the typical deformation behavior of rock-soil body, ensuring that the training sample set can support the subsequent model recognition task of deformation mode. The data labeling process can be based on expert knowledge or historical deformation records.

[0031] Step S106, extract the branch feature, strain sensitive feature and multi-scale spatio-temporal feature corresponding to the multi-source rock-soil fusion tensor.

[0032] After determining the multi-source rock-soil fusion tensor corresponding to the current monitoring data, multi-dimensional feature extraction can be performed based on the current data to comprehensively capture the deformation influencing factors from the aspects of mechanical properties (such as branch features, including elastic features and plastic features), deformation correlation (such as strain sensitive features), and spatio-temporal distribution (such as multi-scale spatio-temporal features). By screening key features with mechanical correlation and trend indication for rock-soil deformation from massive data, redundant data interference can be avoided. Among them, the elastic feature is used to represent the reversible deformation capability and elastic mechanical property of the rock-soil mass, the plastic feature is used to represent the permanent deformation capability and damage accumulation characteristics of the rock-soil mass, and is used to represent the evolution process of the rock-soil mass from reversible elastic deformation to irreversible plastic deformation. Based on the rock-soil mechanics theory, two types of features can be extracted from the attribute dimension through tensor decomposition algorithm (such as PARAFAC decomposition). The strain sensitive feature is used to represent the real-time degree, development trend and instability risk of the rock-soil mass deformation, and focuses on the strain parameters directly related to the deformation. The maximum shear strain of each spatio-temporal point in the tensor can be calculated to judge the shear deformation risk, the axial strain rate can reflect the deformation velocity, and the sensitive data points with strain values exceeding the specification threshold can be screened to strengthen the trend indication of the feature. The multi-scale spatio-temporal feature is used to represent the evolution rhythm in different time dimensions and the distribution difference in different spatial dimensions of the rock-soil mass deformation, and the deformation nature can be fully presented through cross-scale analysis. A multi-resolution analysis method can be used to decompose short-term (1-7 days), medium-term (1-3 months), and long-term (3-12 months) deformation features in the time dimension through wavelet transform; in the spatial dimension, through downsampling (taking 1 representative point every 5 monitoring points) and interpolation (Kriging interpolation to supplement the blank area), the spatial deformation features of local points (within 10m), regions (within 100m), and the whole (the entire engineering region) are obtained, and finally a multi-scale spatio-temporal feature matrix is formed.

[0033] In step S108, based on the pre-constructed self-attention mechanism, the branch features, strain sensitive features and multi-scale spatio-temporal features are classified cooperatively to determine the deformation mode corresponding to the target rock-soil mass.

[0034] In step S110, based on the deformation mode, the deformation of the target rock-soil mass is analyzed.

[0035] The collaborative classification is realized through the self-attention mechanism, so as to determine the deformation mode of the target rock-soil body, the internal relationship of the three types of features can be used to realize the collaborative fusion of the mechanical properties, deformation response and spatial and temporal distribution characteristics, and the one-sidedness caused by single feature analysis can be avoided. For example, the combination feature of sudden strain rate increase + local displacement concentration at the slope foot is given a higher weight. Moreover, the core direction of the deformation mode can be anchored by focusing on the corresponding key information, for example, if the classification result is the creep deformation dominated by long-term time characteristics, the aging characteristics (such as the viscous coefficient) of the rock-soil body and the long-term environmental action (such as seasonal groundwater level change) are preferentially analyzed, so as to realize the accurate positioning from experience judgment to data driving, and realize the accurate deformation analysis of the rock-soil body with the advantages of physical consistency and data driving.

[0036] The embodiment of the present application pre-constructs a multi-source rock-soil fusion tensor, realizes the dual guarantee of physical consistency and data authenticity, and matches the corresponding multi-source rock-soil fusion tensor according to the environmental factor data corresponding to the monitoring data of the rock-soil body, realizes the feature recognition with physical and mechanical consistency and real field data. Moreover, by extracting the branch features, strain sensitive features and multi-scale space-time features of the data, a complete feature system covering the mechanical essence, real-time response and space-time evolution is formed, the deformation essence of the rock-soil body is comprehensively captured, and the deformation essence is comprehensively described. And based on the self-attention mechanism, the above-mentioned features are collaboratively classified, which can accurately determine the deformation mode in the deformation recognition scene of different geological conditions (clay, rock mass, sand and gravel), and realize the rock-soil body deformation analysis with the advantages of physical consistency and data driving.

[0037] Further, with the development of deep learning, more and more researches try to apply neural networks to rock-soil body deformation prediction and disaster identification. However, most of the existing methods are limited to the use of single data source, and usually use simple feature splicing to process multi-source information, resulting in the neglect of the potential physical coupling relationship between cross-source data. The embodiment of the present application combines projection matrix for unified dimension mapping, adaptively captures the nonlinear correlation between cross-source attributes, and solves the feature mismatch problem caused by traditional simple splicing. In addition, the elastic and plastic deformation of rock-soil body has high nonlinearity and time variability, and the existing methods are difficult to effectively distinguish the two types of features, resulting in feature coupling and information loss, especially in the early stage of damage, the weak local high strain signal is often covered, and it is difficult to realize timely identification. The embodiment of the present application effectively separates the elastic and plastic features by using a double-branch decoupling encoder combined with a gating mechanism. At the same time, the embodiment of the present application also introduces a strain sensitive weight matrix, which automatically amplifies the importance of the local high strain area based on the strain gradient, so as to highlight the key deformation signal and solve the feature confusion and weak signal capture problem.

[0038] On the basis of the above-mentioned embodiments, the embodiment of the present application also provides another rock-soil body deformation identification method based on neural network, Figure 2A flow chart of an embodiment of the present application is shown, referring to Figure 2 The method comprises the following steps:

[0039] In step S202, data of the target rock-soil mass is monitored to obtain monitoring data of the target rock-soil mass and environmental factor data of the environment in which the monitoring data is located.

[0040] In step S204, based on the environmental factor data, a multi-source rock-soil fusion tensor corresponding to the monitoring data is determined from a pre-constructed multi-source rock-soil fusion tensor set.

[0041] In combination with the above embodiment, the rock-soil data has multi-scale and heterogeneous characteristics, wherein the numerical simulation data contains high-dimensional field variables, and the monitoring data has sparsity in space and time. The conventional processing method usually adopts a simple splicing method, but this method is prone to cause feature mismatch and is difficult to establish effective association across data sources, thereby affecting the accuracy of deformation identification. The present application constructs numerical simulation tensors, monitoring tensors and environmental factor tensors, and uses a learnable projection matrix and an interaction tensor for data fusion, so as to map the multi-source data to a unified dimension, enhance the association across data sources, and solve the feature mismatch problem. The specific steps are as follows:

[0042] 1) Obtain the numerical simulation tensors, monitoring tensors and corresponding environmental factor tensors pre-collected for the preset rock-soil mass.

[0043] The numerical simulation data, monitoring data and environmental factors are respectively organized into three-dimensional tensors, including: constructing a numerical simulation tensor , the dimension of which is , is the number of numerical simulation attributes, The numerical simulation attributes include: grid size, element type, elastic modulus, Poisson's ratio, cohesion, internal friction angle, permeability coefficient and load step; constructing a monitoring tensor , the dimension of which is , is the number of monitoring attributes, The monitoring attributes include: strain gauge azimuth angle, burial depth, temperature, water content, strain value, displacement amount, settlement rate and inclination change, wherein the strain value includes x-direction normal strain, y-direction normal strain, z-direction normal strain and xy shear strain, and the displacement amount includes x-direction displacement, y-direction displacement and z-direction displacement; and constructing an environmental factor tensor , the dimension of which is , is the number of environmental factor attributes, The environmental factor attributes are: rainfall, groundwater level and vibration intensity; is the number of samples, is the time step.

[0044] In practical implementation, the data sources, attributes, collection methods, significance, and units / dimensions of multi-source data are introduced as follows:

[0045]

[0046]

[0047] 2) Linearly combine the numerical simulation tensor, monitoring tensor, and environmental factor tensor to generate the interaction tensor of the preset soil and rock mass.

[0048] By linearly combining all attribute vectors of the numerical simulation tensor, monitoring tensor, and environmental factor tensor at each sample and time step through fusion weight vectors, and then applying a hyperbolic tangent activation function for nonlinear transformation, an interaction tensor is generated. The elements of the interaction tensor represent the cross-source interaction strength, thereby enhancing the correlation between different data sources, as expressed below:

[0049]

[0050] In the formula, The interaction tensor represents the feature value of the j-th attribute of the i-th sample at time t, which is used to characterize the cross-source interaction strength; Let J be the fusion weight vector of the j-th attribute, with dimension J. , is a trainable parameter used for linear combination of multi-source features; This represents a vector concatenation operation that combines three components into a single vector. This is the hyperbolic tangent activation function, used to implement nonlinear transformations and limit the output range; Let represent the vector of all attributes at time t for the i-th sample, i.e. Represents numerical simulation tensor The vector of all attributes of the i-th sample at time t. Represents the monitoring tensor The vector of all attributes of the i-th sample at time t. Represents environmental factor tensors The vector of all attributes of the i-th sample at time t; i is the first dimension index of the tensor, which also represents the row index; j is the second dimension index of the tensor, which also represents the column index; t is the third dimension index of the tensor, which also represents the time step index.

[0051] It should be noted that the interaction tensor Each element By dynamically quantifying the correlation strength between different data source attributes through nonlinear activation, such as the coupling relationship between elastic modulus and displacement sensors, the implicit influence of environmental factors on material parameters can be adaptively corrected. For example, when the environmental factor tensor... When groundwater levels change abruptly, the interaction tensor automatically enhances the correlation weight between the permeability coefficient and the strain value, whereas conventional methods require manual setting of correction coefficients, which cannot achieve the same effect.

[0052] 3) Unify the dimensions of the numerical simulation tensor, monitoring tensor, and environmental factor tensor, and then fuse the unified numerical simulation tensor, monitoring tensor, and environmental factor tensor with the interaction tensor to generate a multi-source soil-rock fusion tensor set for the preset soil-rock body.

[0053] In practical implementation, a projection matrix can be used to map the numerical simulation tensor, monitoring tensor, and environmental factor tensor along the attribute dimension to a unified dimension. Then, tensor-matrix multiplication is used to add the mapped tensor to the interaction tensor to obtain a fused tensor, capturing the nonlinear interactions between data, represented as:

[0054]

[0055] In the formula, This represents a fusion tensor used to integrate information from multiple sources. The modulo product of a tensor and a matrix (Mode-2-Product), also known as the 2-modulo product, is a tensor-matrix multiplication performed along the attribute dimension. The projection matrix of the numerical simulation data has dimensions of . Mapping simulated data to a unified dimension provides trainable parameters. To unify attribute dimensions, It is the sum of the number of numerical simulation attributes and the number of monitored attributes; The projection matrix of the monitoring data has dimensions of . Mapping monitoring data to a unified dimension provides trainable parameters. This is the projection matrix of the environmental factor data, with dimensions of . Mapping environmental factor data to a unified dimension provides trainable parameters. For the interaction tensor, the dimension is It is used to enhance cross-source correlations and capture non-linear interactions between data. It is an interaction tensor In the The sample index, the first The index of the attribute and the first The element value at each time step index, i.e., the interaction tensor. It is made by all The tensor formed by this. It should be noted that the projection matrix... , , The dimension gap problem can be solved, and the attribute dimension is mapped to a unified space while retaining the interpretability of the original physical quantity, for example, the "xy shear strain" in the monitoring tensor is projected through a projection matrix The mapping still maintains the mechanical correlation with the "cohesion" of the numerical simulation in the unified dimension, while the conventional feature extraction destroys this physical correspondence.

[0056] Further, the embodiments of the present application also compare the recognition accuracy of different methods on various deformation modes, and verify the recognition performance advantage of the present technology on different rock-soil deformation modes (elastic deformation, elastic-plastic deformation, creep, shear failure). Figure 3 The recognition accuracy comparison diagram corresponding to different methods is shown, and the embodiments of the present application highlight the overall superiority of the present technology in multi-mode recognition through comparison with three conventional technologies (multi-source splicing method, single-branch convolutional neural network, and physical constraint network). The vertical axis unit is the recognition accuracy (%), and the higher the value, the better the performance. The same data set (containing four kinds of deformation mode samples) is used in the experiment, and the four kinds of methods are compared under the same training conditions. In the experiment, the present technology realizes multi-source tensor fusion, physical constraint embedding and double-branch decoupling coding, the multi-source splicing method uses simple splicing of numerical simulation / monitoring / environmental data and then inputs a single-branch network, the single-branch convolutional neural network uses a basic convolutional network trained only using monitoring data, and the physical constraint network only embeds strain compatibility constraints but does not realize multi-source fusion. The experimental results show that the recognition accuracy of the present technology (blue column) on the four types of deformation modes is significantly higher than that of other methods, and the conventional methods perform the weakest (the column height is the lowest) in creep and shear failure recognition. Because these modes require multi-source data for collaborative analysis, the tensor fusion mechanism of the present technology effectively solves the feature mismatch problem. The physical constraint network (purple column) performs well in elastic deformation, because the physical constraint is effective for simple modes, but the effect drops sharply in nonlinear deformation (such as creep), which highlights the necessity of the "physical constraint + multi-source fusion" dual mechanism of the present technology.

[0057] Step S206, extract the branch features, strain sensitive features and multi-scale spatio-temporal features corresponding to the multi-source rock-soil fusion tensor.

[0058] 1) Extract branch features: rock-soil deformation features have strong coupling characteristics, such as elastic deformation and plastic flow often occurring simultaneously and influencing each other. Conventional convolutional neural networks use a single branch structure, which cannot effectively separate these coupled features, easily leading to feature representation confusion and reducing classification accuracy. The present application uses a double-branch structure to extract two types of features (such as elastic features and plastic features), and combines a gating mechanism to realize feature separation, decouple coupled features, and improve the clarity of feature representation.

[0059] The specific steps are as follows:

[0060] Step S1, a first convolutional processing of the multi-source rock-soil fusion tensor is performed by a preset neural network model to extract an elastic feature tensor corresponding to the multi-source rock-soil fusion tensor.

[0061] A three-dimensional convolution operation is performed on the physical constraint tensor using an elastic convolution kernel weight tensor and a bias vector, and then a linear rectifier activation function is applied to extract an elastic feature tensor for representing a recoverable elastic deformation mode, which is represented as:

[0062]

[0063] In the formula, represents the elastic feature tensor, representing the recoverable elastic deformation mode; is a linear rectifier activation function, used to realize a nonlinear transformation and maintain sparsity; is an elastic convolution kernel weight tensor, with a dimension of , which is specially used to capture the recoverable elastic deformation feature and is a trainable parameter; represents a three-dimensional convolution operation, used to extract a spatio-temporal feature; is a bias vector of the elastic branch, used to adjust the reference value of the output feature, which is a trainable parameter; is the size of the elastic convolution kernel, such as , which controls the receptive field size; is the number of channels of the elastic feature output, which controls the feature richness, such as setting .

[0064] Step S2, a second convolutional processing of the monitoring data is performed by a preset neural network model to determine a gating weight corresponding to the multi-source rock-soil fusion tensor.

[0065] Step S3, a range compression is performed on the elastic feature tensor, and based on the gating weight, a gated selection is performed on the range-compressed elastic feature tensor to determine a plastic feature tensor of the multi-source rock-soil fusion tensor.

[0066] A three-dimensional convolution operation is performed on the physical constraint tensor using a plastic convolution kernel weight tensor and a bias vector, and then a Sigmoid activation function is applied to output the gating weight, while the elastic feature tensor is compressed in range by a hyperbolic tangent activation function, and finally a gated selection is realized by a Hadamard product to extract a plastic feature tensor for representing a permanent plastic deformation mode, which is represented as:

[0067]

[0068] In the formula, represents the plastic feature tensor, representing the permanent plastic deformation mode; is a Sigmoid activation function, used to output the gating weight; The plastic convolution kernel weight tensor has dimension . Specifically designed to capture features of permanent plastic deformation, it is a trainable parameter; The bias vector of the plastic branch is used to adjust the baseline value of the output feature and is a trainable parameter. This represents the Hadamard product, used to implement gating selection of features; It is a hyperbolic tangent activation function used to compress the range of elastic eigenvalues ​​to adapt to gating; The size of the plastic convolution kernel, such as, Controlling the size of the receptive field; The number of channels output for plastic features controls feature richness; for example, setting... .

[0069] Step S4: Perform feature splicing on the elastic feature tensor and the plastic feature tensor to determine the branch characteristics of the multi-source soil-rock fusion tensor.

[0070] The elastic and plastic feature tensors are concatenated along the channel dimension, represented as follows: This leads to the formation of decoupled coding features. The decoupled coding features simultaneously contain both elastic and plastic feature information; among them, To decouple the encoded features, the number of channels is It characterizes the output features of the dual-branch decoupled encoder, while also containing elastic and plastic feature information; This represents a vector concatenation operation that combines two components into a single vector. It should be noted that while conventional dual-branch networks extract features in simple parallel processes, the strong coupling of the elasto-plastic deformation of soil and rock masses leads to feature confusion. This invention addresses this by... The range of compressive elastic eigenvalues ​​is up to Then, a Hadamard product is performed with the plastic gating weights to achieve a subtraction operation in the feature space. When elastic deformation dominates, such as in the early stages of loading, the gating value approaches 1, fully preserving the plastic characteristics. When plastic flow intensifies, such as in the yielding stage, the gating value automatically decays in the shear band region, suppressing the interference of elastic characteristics on the plastic branches. Based on this, it can be found that creep and elastic recovery can be accurately separated. Under cyclic loading, the gating mechanism follows... The oscillating synchronous switching of the term uses a plastic channel to simulate the Bauschinger effect. It should also be noted that the elastic branch uses the convolution kernel weight tensor. and bias vector It is specifically designed to capture recoverable elastic deformation; the plastic branch uses convolution kernel weight tensors. and bias vector This architecture is specifically designed to capture irreversible plastic deformation. These parameters are updated independently during training via backpropagation and gradient descent, allowing the network to learn to respond to different types of deformation patterns. Furthermore, different activation functions guide the network to learn different types of responses. The elastic branch uses a linear rectified activation function, suitable for capturing non-negative, sparse, and recoverable elastic responses, consistent with the physical characteristics of elastic deformation. The plastic branch uses the Sigmoid function as gate weights, which are then multiplied by the elastic features compressed by the hyperbolic tangent activation function using the Hadamard product. The Sigmoid function's output is a gate value between 0 and 1, used to suppress or enhance the plastic deformation-related parts of the elastic features, achieving selective extraction of plastic deformation. Simultaneously, it makes the plastic branch dependent on the output of the elastic branch, thus achieving elastoplastic decoupling in the feature space.

[0071] 2) Addressing Strain-Sensitive Features: During the deformation of soil and rock masses, the contribution of strain at different spatial locations and time points to the overall deformation varies significantly. Conventional average pooling or max pooling methods cannot distinguish these differences, leading to the dilution of strain information in key areas and the loss of important features, thus affecting classification accuracy. This invention adaptively weights features by calculating the strain gradient tensor and constructing a strain-sensitive weight matrix, highlighting the feature importance of high strain gradient regions, thereby enhancing the feature representation of key areas. The specific steps are as follows:

[0072] Step S1: Extract the strain components of the multi-source soil-rock fusion tensor, determine the gradient of the strain components in the spatial direction and time dimension, and generate the strain gradient vector.

[0073] Strain components are extracted from the input data, and their gradients in the spatial and temporal dimensions are calculated to form a strain gradient vector, which characterizes the intensity and direction of strain change, and is expressed as:

[0074]

[0075] In the formula, Indicates the first The strain gradient vector at position t at time t is used to characterize the intensity and direction of strain change; It represents the spatial gradient of strain in the x-direction. The gradient is calculated using discrete difference methods, such as central difference, in both spatial and temporal dimensions, reflecting the rate of change of strain in the horizontal direction. It represents the spatial gradient of strain in the y-direction. The gradient is calculated using discrete difference methods, such as central difference, in both spatial and temporal dimensions, reflecting the rate of change of strain in the vertical direction. It represents the gradient of strain in the time dimension. The gradient calculation uses discrete difference methods, such as central difference, to perform the calculation in both spatial and temporal dimensions, reflecting the rate of change of strain over time. This represents a vector concatenation operation that combines three components into a single vector. It should be noted that... Item and The term characterizes the spatial gradient and can identify potential slip surfaces. The term characterizes the time gradient and is associated with rate-sensitive properties, such as clay creep.

[0076] Step S2: Identify the importance of the positional features of the strain gradient vector and determine the strain sensitivity weights corresponding to the positional features.

[0077] Specifically, based on the L2 norm of the strain gradient vector, the strain sensitivity weights at each location and time can be calculated using exponential and softmax functions to quantify the importance of the feature at that location. Regions with larger strain gradients receive higher weights, as expressed below:

[0078]

[0079] In the formula, Indicates the first The strain-sensitive weight at position t is used to quantify the importance of the feature at that position; For sensitivity factors, such as, It is used to adjust the sharpness of the weight distribution and control the degree of attention concentration; Represents the L2 norm; This represents an exponential function, used to amplify gradient differences; Indicates the first Position No. The strain gradient vector at time t; Indicates a spatial location index; Indicates difference from Spatial location index; This is to distinguish it from the time step index t.

[0080] Step S3: Use strain-sensitive weights to perform feature weighting on the multi-source soil-rock fusion tensor to determine the strain-sensitive features of the multi-source soil-rock fusion tensor.

[0081] The strain-sensitive weight tensor is element-wise multiplied with the feature tensor (such as the multi-source soil-rock fusion tensor mentioned above, or alternatively, the branched features mentioned above) to achieve adaptive weighting of the features, resulting in a weighted feature tensor that highlights the characteristics of high strain gradient regions. The following formula is for branched features. (That is, the decoupling coding feature) is explained as follows:

[0082]

[0083] In the formula, This represents the weighted feature tensor, highlighting the characteristics of the high strain gradient region; For the strain-sensitive weight tensor, the dimension is , It is a strain-sensitive weight tensor In the Position No. The element value at time t, i.e., the strain-sensitive weight tensor. It is made by all The tensor that makes up the tensor; The spatial height of the weighted feature tensor; The spatial width of the weighted feature tensor; It is the total time step, which is determined based on the length of the time series of the input data, such as the number of time points in monitoring data or numerical simulation data; This represents element-wise multiplication of tensors. It should be noted that, through... The quantitative strain change rate corresponds physically to the localization intensity of deformation, and is an exponential function. Items and Sensitivity Factors The combination of these factors produces a superlinear amplification effect. When the strain gradient at a certain point exceeds a critical value, such as... Its weight rises sharply to become dominant, which can solve the problem of capturing weak deformation signals in the early stage of creep and achieve automatic focusing on the accelerated creep point.

[0084] 3) Addressing Multi-Scale Spatiotemporal Characteristics: Soil and rock deformation exhibits significant multi-scale spatiotemporal characteristics, such as the coexistence of instantaneous settlement and long-term creep, and the interaction of local shearing and global slip. Conventional single-scale convolutional neural networks cannot simultaneously capture these features at different scales, resulting in insufficient ability to identify complex deformation patterns. This invention constructs a multi-scale convolutional structure and performs multi-scale convolution operations, using hollow convolutional kernels with different expansion rates to extract multi-scale spatiotemporal features in parallel, and then fuses these features to comprehensively characterize the complex deformation behavior of soil and rock masses. The specific steps are as follows:

[0085] Step S1 involves performing a dilated convolution operation on the multi-source soil-rock fusion tensor using multiple pre-set dilated convolution branches in parallel with different expansion rates, thereby capturing the spatiotemporal characteristics of the multi-source soil-rock fusion tensor at different scales.

[0086] This can be achieved by employing three parallel dilated convolution branches, each using a different dilation rate to perform dilated convolution operations, in order to capture spatiotemporal features at different scales, as shown below. .in, Indicates the first The output of the dilated convolution operation performed by each dilated convolution branch; For the first The dilation rate of each dilated convolution branch takes a value of [value]. These correspond to three receptive field scales: small, medium, and large. For the convolution kernel size, such as, , three branches keep the same size but have different receptive fields; is the branch index of the empty convolution, taking values , wherein the above multi-source rock-soil fusion tensor can be processed, and the strain-sensitive feature can be further extracted, and the above formula is based on the strain-sensitive feature , that is, the above weighted feature tensor) is explained. Wherein the receptive field is expanded by the empty convolution, corresponds to the grain size, captures the rearrangement of clay particles, matches the joint spacing, monitors the opening of the rock mass structure surface, covers the engineering scale, such as tunnel convergence deformation.

[0087] Step S2, summing the spatio-temporal features to obtain the multi-scale spatio-temporal features of the multi-source rock-soil fusion tensor.

[0088] The outputs of each empty convolution branch are summed to obtain a multi-scale fusion feature tensor that integrates multi-scale information and comprehensively represents the complex deformation behavior of the rock-soil mass, expressed as:

[0089]

[0090] In the formula, , the multi-scale fusion feature tensor integrates multi-scale information, and the dimension is ; The number of channels of the multi-scale convolution output. It should be noted that the "point sampling" feature of the empty convolution can match the non-uniformity of the rock-soil medium, and in the soil-rock mixture, The convolution kernel of the rock-soil medium automatically skips the block stone area and focuses on the soil matrix, which significantly improves the feature signal-to-noise ratio compared with the conventional uniform convolution, and The uniform kernel size of ensures the balance of the parameter amount of different branches, preventing the dominance of large-scale branches in training. Further, the above step S108 is further executed. Wherein, there is a physical correlation between the deformation types of the rock-soil mass, such as elastic deformation, elastic-plastic deformation, creep and shear failure, and the conventional single-label classification method regards each category as independent, ignoring the cooperation and repulsion relationship between the modes, resulting in the decline of the classification performance of the complex composite deformation mode. The present application uses the self-attention mechanism to model the correlation between the deformation modes and performs collaborative classification through feature global pooling and query mapping, mode association graph construction and collaborative classification prediction, thereby improving the recognition ability of the composite deformation mode. For details, see the following steps S208-S212.

[0091] Step S208, convert the branch features, strain-sensitive features and multi-scale spatio-temporal features into global feature vectors, map the global feature vectors to the query space, and generate a query representation matrix.

[0092] In one implementation, feature transformation can be performed on the three types of features (i.e., branch features, strain-sensitive features, and multi-scale spatiotemporal features) separately. Alternatively, feature transformation can be performed after further determining the strain-sensitive features among the branch features and the multi-scale spatiotemporal features. In a specific implementation, high-dimensional spatiotemporal features are compressed into sample-level global feature vectors by performing global average pooling on the corresponding data. Then, the global feature vectors are multiplied by a trainable query projection matrix to map to the query space, resulting in a query representation matrix.

[0093] Multi-scale fusion feature tensor To explain, global average pooling calculates the average value of all spatial locations and time steps for each channel, resulting in a global pooled feature vector. , dimension Furthermore, the global pooling feature vectors are... With dimension Query projection matrix Multiply to obtain the query representation matrix. , dimension This allows for the mapping of global features of each sample to the query space; where, To query the projection matrix, which is a trainable parameter with dimension 1. This maps features to the query space. To query the dimensions of the space, such as, .

[0094] Step S210: The query representation matrix is ​​processed by a scaling dot product attention mechanism to determine the pattern association graph corresponding to the global feature vector; the pattern association graph includes the similarity of deformed patterns between samples.

[0095] Combining the above steps, the similarity matrix between samples is calculated using the query representation matrix. Normalized attention weights are obtained through a scaled dot product attention mechanism and a softmax function. A pattern association graph is then constructed to express the association strength between different samples, as follows: In the formula, Query representation matrix transpose; This is a pattern association graph with dimension 1. The first pattern association graph Line number Column elements Indicates the first The sample and the first The strength of the association between individual samples; The row index of the pattern association graph is the sample index; The column index for the schema association graph is the sample index; is a normalized exponential function, ensuring that the sum of all elements in each row is 1, forming a probability distribution. It is noted that the query representation matrix maps the samples to the physical semantic space, and the dot product calculates the similarity of deformation modes between samples, divided by The scaling operation of the item can balance the numerical drift problem of similarity in high-dimensional space, so that stably reflects the physical nature, such as similar unstable modes of two slopes.

[0096] In step S212, the pattern association graph and the global feature vector are subjected to matrix multiplication processing to generate a fusion output, and the target rock mass corresponding to the deformation mode is determined based on the fusion output.

[0097] In the specific implementation, the pattern association graph and the global pooling feature are subjected to matrix multiplication to realize information propagation and fusion of the feature based on the pattern association. Further, the fused feature can be input into a multilayer perceptron classifier, represented as , and an output classification output matrix is output. Wherein, is a multilayer perceptron classifier; is a classification output matrix, with a dimension of , each row of which represents the predicted probability of a sample belonging to each deformation mode; is the number of deformation modes, which is 4. In an embodiment, the multilayer perceptron classifier adopts a two-layer perceptron structure, with an input layer dimension of , a hidden layer dimension of 64, and an output layer dimension of , corresponding to four deformation modes of elasticity, elastoplasticity, creep, and shear failure, and the activation function adopts a ReLU activation function.

[0098] It is noted that the item can realize knowledge transfer, and when a new sample lacks deformation data of a certain type, the association graph automatically refers to the features of similar samples, such as predicting long-term deformation of an unknown sample through a known creep sample.

[0099] In step S214, the target rock mass is subjected to deformation analysis based on the deformation mode.

[0100] Further, the present embodiment also compares displacement prediction errors at different time steps to verify the stability of the present technique in rock mass displacement time series prediction. Figure 4 Fig. 4 shows a prediction error comparison diagram of the present embodiment, and the details are described with reference to Figure 4, the vertical axis is the displacement prediction error (millimeters), the lower the value, the higher the accuracy. The experiment compares the error of the multi-source splicing method within 100 hours to reflect the improvement effect of physical constraint embedding on long-term prediction accuracy. The experiment selects a typical slope case and monitors its z-direction displacement (unit: meters) within 100 hours. The absolute error (unit: millimeters) of the prediction results of the two methods and the true displacement is compared, and the time step is 5 hours (the horizontal axis unit is hours). The experimental results show that the proposed technology (blue line) maintains low error fluctuation throughout (the curve is close to the horizontal axis), while the error of the multi-source splicing method (orange line) increases significantly with time and fluctuates sharply. In 60-80 hours (potential instability stage), the error of the proposed technology is still stable, while the comparative method has a sharp error peak, proving that the physical stiffness matrix effectively suppresses the prediction bias in the instability state. In addition, the area under the curve of the proposed technology (blue filled area) is much smaller than that of the comparative method, reflecting that it reduces the error accumulation effect through strain coordination constraints.

[0101] Furthermore, this embodiment of the invention extracts branch features, strain-sensitive features, and multi-scale spatiotemporal features corresponding to the multi-source soil-rock fusion tensor through a preset neural network model, achieving a fusion of data-driven and model-driven capabilities. The data-driven capability, jointly realized by multi-source tensor fusion, dual-branch decoupled encoding, multi-scale convolution, and self-attention mechanism, ensures the model's learning ability and fitting of complex soil-rock monitoring data patterns. The physical-driven capability, jointly realized by strain coordination constraints, physical stiffness matrix, and physical information hybrid loss, ensures the physical consistency and reliability of the prediction results. The combination of data-driven and model-driven capabilities avoids the "black box" problem of purely data-driven approaches and breaks through the idealized assumptions of purely physical modeling, enabling robust and reliable identification of soil-rock deformation in real engineering environments. Traditional model loss function designs are mostly based on classification or regression errors, without incorporating prior knowledge of soil-rock instability mechanisms. When soil-rock enters the instability stage, mechanical properties such as stiffness degradation and yield expansion are crucial to system stability, but existing methods often fail to reflect this sensitivity to critical states in the loss function, thus weakening the ability to predict catastrophic critical points. This invention proposes a hybrid form of classification loss and physical constraint residual loss, trained based on a pre-set physical information hybrid loss function. In this embodiment, a stiffness matrix is ​​dynamically generated based on the predicted deformation pattern, achieving instability reinforcement constraints under abrupt stiffness changes, thereby improving the model's ability to identify and warn of critical failure points. In the task of classifying and identifying deformation in soil and rock masses, the conventional cross-entropy loss function only focuses on the consistency between the predicted and true labels, ignoring the physical laws that the deformation process of soil and rock masses must follow, such as equilibrium equations and constitutive relations. This may lead to the model predicting physically unreliable deformation patterns, reducing its reliability and generalization ability in real-world scenarios. This invention calculates cross-entropy classification loss and constructs physical constraint residual loss, then forms a hybrid loss function through weighted fusion, ensuring that the model's prediction results conform to physical laws. The specific steps are as follows:

[0102] Step S10: Train the preset neural network model using the preset training sample set to determine the initial prediction output.

[0103] Calculate the cross-entropy loss between the probability distribution of the predicted deformation patterns from the network and the true labels. This is used to ensure the basic accuracy of the model's classification prediction, and is expressed as:

[0104]

[0105] In the formula, Cross-entropy represents the classification loss, used to measure the difference between the predicted probability and the true label; For the first The one-hot encoded value of the true label of a sample in the c-th category is a scalar of 0 or 1; The model predicts the first The probability value of a sample belonging to the c-th category of deformation pattern is a scalar between 0 and 1. It is a classification output matrix The element in the i-th row and c-th column represents the element predicted by the model. The probability that a sample belongs to the c-th category; It is a logarithmic function, with the default base being the natural constant.

[0106] Step S20: Calculate the cross-entropy loss of the preset neural network model based on the initial prediction output, and generate a dynamic stiffness matrix according to the initial prediction output. The dynamic stiffness matrix is ​​determined based on the material constitutive relation corresponding to the training samples and the finite element discretization method of the preset neural network model.

[0107] Step S30: Construct the physical constraint residual loss of the preset neural network model based on the stiffness matrix.

[0108] 2) Constructing physical constraint residual loss

[0109] The stiffness matrix is ​​dynamically generated based on the deformation mode categories predicted by the model. Then, the L2 norm squared of the residual between the predicted displacement vector and the load vector is calculated as the physical constraint residual loss to ensure that the prediction conforms to physical laws. This is expressed as:

[0110]

[0111] In the formula, This represents the physical constraint residual loss, used to ensure that the prediction conforms to physical laws; According to the first The stiffness matrix generated by the predicted deformation mode categories of each sample has dimensions of [dimensional value missing]. Its construction depends on the material constitutive relation and the finite element discretization method; For the first The displacement vector corresponding to each sample is derived from the monitoring tensor; For the first The load vector corresponding to each sample is derived from the numerical simulation tensor. The dimension of the displacement vector depends on the number of degrees of freedom of the finite element model. Assume the model has... There are nodes, each node has One degree of freedom, such as in a 3D model ,but In one implementation, the stiffness matrix Dynamically generated based on the deformation pattern categories predicted by the model; specifically, from the classification output matrix. Get the first Predicted probability distribution of each sample , respectively corresponding to elasticity, elastoplasticity, creep, shear failure; the category with the maximum probability is taken as the prediction mode of the first sample ; the prediction mode of the first sample material parameters are selected and the equivalent elastic modulus is calculated , specifically:

[0112]

[0113] based on the equivalent elastic modulus and Poisson's ratio a physical stiffness matrix is generated, represented as: ; wherein, is the elastic modulus, derived from the numerical simulation tensor, representing the material's ability to resist elastic deformation; is the Poisson's ratio, derived from the numerical simulation tensor, representing the ratio of transverse deformation to longitudinal deformation of the material; is the first Kronecker symbol, with a value of 1 when , otherwise 0, used to indicate the position of matrix elements, such as, , ; is the second Kronecker symbol, for fixed index , because is always equal to itself, then , such as, when , when ; is the third Kronecker symbol, for fixed index , because is always equal to itself, then , such as, when , when . It should be noted that conventional physical loss usually uses fixed equations, but geotechnical failure involves stiffness catastrophe, and the present application dynamically generates stiffness matrix based on the prediction mode when predicting shear failure, singularizes the stiffness matrix , forcing to sharply increase in the instability state, and the instability hardening mechanism can improve the sensitivity of the model to the catastrophe turning point, which cannot be achieved by constant weight loss.

[0114] In step S40, the physical constraint residual loss and the cross-entropy loss are weighted and fused to generate a physical information hybrid loss function.

[0115] ​​The cross-entropy classification loss and the physical constraint residual loss are linearly superimposed according to a preset physical constraint intensity coefficient to form a final hybrid loss function for guiding model training, and is expressed as:

[0116]

[0117] In the formula, The hybrid loss function is used to optimize the model parameters. The physical constraint intensity coefficient is used to adjust the relative importance of the physical constraint term in the total loss, such as In summary, the model iteration training and parameter updating process is based on the hybrid loss function to achieve optimization and convergence of the model parameters. Specifically, the model is trained for multiple rounds using the training data set. In each iteration, a batch of sample data is input, and the fusion tensor, the physical constraint tensor, the decoupled encoded feature, the weighted feature tensor, and the multi-scale fusion feature tensor are calculated through forward propagation. Finally, the classification prediction result is output. Then, the total loss value is calculated according to the hybrid loss function, and the gradients of the loss with respect to each trainable parameter are calculated through the backpropagation algorithm. Meanwhile, the Adam optimizer is used to update these parameters, adjusting the step size to minimize the loss function. The parameter updating process continues until the model converges. The stopping conditions for model iteration include: when the hybrid loss function does not significantly decrease (with a change rate less than a preset threshold, e.g., 0.001) on the validation set for multiple consecutive iteration rounds (e.g., 10 rounds), or reaches a preset maximum number of iterations (e.g., 1000 rounds), the training is terminated to prevent overfitting and ensure generalization performance.

[0118] Further, the deformation of the rock-soil mass needs to satisfy the strain compatibility equation, but there are noises and errors in the actual monitoring data, leading to physical inconsistency. Traditional neural networks often only pursue prediction accuracy, lack consideration of strain compatibility, constitutive relationship, and mechanical constraints, and may make the prediction results violate the laws of mechanics, affecting the reliability of the model. The embodiments of the present application also generate a physical stiffness matrix and construct a strain compatibility constraint term, embed the strain compatibility equation and the physical stiffness matrix as physical constraints into the training sample (such as the multi-source rock-soil fusion tensor described above), correct the prediction results through the strain compatibility constraint and the curl operator, and perform physical consistency correction on the fusion tensor to ensure that the strain satisfies the compatibility equation. The trained model not only guarantees data-driven learning ability, but also strictly complies with the laws of continuum mechanics, improving the physical consistency and reliability of the prediction. In specific implementation, the following steps are included:

[0119] Step S1, obtaining a pre-constructed multi-source rock-soil fusion tensor training sample.

[0120] Step S2, based on the linear physical response of the rock-soil mass corresponding to the multi-source rock-soil fusion tensor training sample, performing physical constraint on the multi-source rock-soil fusion tensor training sample to generate a physical constraint tensor.

[0121] In one embodiment, the physical stiffness matrix can be dynamically generated according to the elastic modulus and Poisson's ratio, the elements of which are calculated by the elastic modulus, Poisson's ratio and the Kronecker symbol, for building the constitutive relationship of the material, representing the linear physical response of the material, expressed as:

[0122]

[0123] wherein, represents the element of the physical stiffness matrix in the i-th row and the j-th column, for building the constitutive relationship of the material; represents the physical stiffness matrix, with a dimension of , for representing the linear physical response of the material, reflecting the constraint of the material parameters on the deformation; is the elastic modulus, derived from the numerical simulation tensor, representing the ability of the material to resist elastic deformation; is the Poisson's ratio, derived from the numerical simulation tensor, representing the ratio of transverse deformation to longitudinal deformation of the material; is the first Kronecker symbol, with a value of 1 when , otherwise 0, for indicating the position of the matrix element, such as, , ; is the second Kronecker symbol, for fixed index , because is always equal to itself, then , such as, when , when ; is the third Kronecker symbol, for fixed index , because is always equal to itself, then , such as, when , when . It should be noted that, and the double Kronecker symbol combination not only simplifies the matrix expression of three-dimensional Hooke's law, but more importantly, through the term, the constraint strength of normal stress and shear stress can be distinguished, such as the strengthening constraint of Poisson's ratio on transverse deformation.

[0124] ​​Further, the strain tensor can be extracted from the multi-source rock-soil fusion tensor training sample and reorganized into a matrix form, then the vectorized strain tensor is linearly transformed using the physical stiffness matrix, and a strain compatibility constraint function is calculated in combination with a strain compatibility penalty term, wherein the strain compatibility penalty term is calculated based on the curl and transpose of the strain field Frobenius norm, the strain compatibility constraint function is used for physical consistency correction of the strain, and the strain is ensured to satisfy the compatibility equation, which is expressed as:

[0125]

[0126] In the formula, represents the strain compatibility constraint function, which is used for physical consistency correction of the input strain; is the strain tensor extracted from the fusion tensor , with a dimension of , representing the local deformation state of the rock-soil body; represents the vectorization operation, which expands the matrix column by column into a column vector, facilitating matrix operation; is a constraint strength coefficient, such as , used to adjust the importance of the strain compatibility penalty term and control the strictness of the physical constraint; represents the Frobenius norm of the matrix, which is used to measure the size of the strain compatibility deviation; represents the curl operator, which is approximated by central difference in the discrete case, and is used to evaluate the compatibility of the strain field; represents the transpose operation.

[0127] In specific implementation, the strain tensor is indexed from the attribute dimension of the fusion tensor to correspond to the strain components, including , , , , and reorganized into a matrix, wherein is the x-direction normal strain, is the y-direction normal strain, is the z-direction normal strain, is the xy shear strain. It should be noted that the strain compatibility equation is a constraint condition for the strain tensor , and the strain tensor is a second-order symmetric tensor in continuum mechanics, expressed as a matrix, so as to calculate the curl and transpose operation, therefore, the strain components scattered in the attribute dimension need to be extracted and reorganized into a matrix form, and the rock-soil body deformation needs to satisfy the strain compatibility equation The strain compatibility equation is embedded into the network as a physical constraint, the strain compatibility deviation is calculated as a penalty term, the predicted strain satisfies the physical law, and the physical consistency and reliability of the prediction are improved. It should be noted that the conventional physical constraint only corrects the static strain with the stiffness matrix, while the curl operator can capture the dynamic propagation characteristics of strain compatibility. When the local shear zone such as landslide surface is formed, the curl operator will amplify the spatial discontinuity of , so that the penalty term sharply increases in the instability area, so that the model weakens the constraint in the early stage of creep when the strain gradient is small, and strengthens the constraint at the damage critical point of the strain gradient mutation.

[0128] Step S3, based on the physical constraint tensor, a training sample set is constructed.

[0129] In specific implementation, the output of the strain constraint function is projected back to the original data space through a linear mapping matrix, and the physical constraint tensor is obtained by combining the fusion tensor, so as to improve the physical consistency of the data, and is expressed as:

[0130]

[0131] In the formula, is the physical constraint tensor, which represents the tensor after applying the physical constraint (which can be used as a training sample for model training). is a linear mapping matrix with a dimension of , which is used to transform the dimension of the term to be consistent with the numerical simulation tensor , and is a trainable parameter. It should be noted that the linear mapping matrix learns to ignore the feature dimensions unrelated to mechanics in the training, and projects the 9-dimensional term back to the high-dimensional feature space through the linear mapping matrix , so as to automatically suppress the pseudo-strain compatibility caused by environmental noise, such as sensor drift caused by temperature change.

[0132] Further, the embodiments of the present application also use a box plot to analyze the prediction error distribution of different methods on each strain component. Figure 5 The corresponding schematic diagram is shown, and the box plot is shown in Figure 5The horizontal axis represents four strain components: normal strain in the x-direction, normal strain in the y-direction, normal strain in the z-direction, and xy shear strain. The vertical axis represents the prediction error, using absolute error values; smaller values ​​indicate more accurate predictions. This evaluates the accuracy of each method in predicting specific physical quantities. The median line of the box plot shows that the method of this invention has the lowest prediction error across all strain components, and the error distribution is the most concentrated, indicating more stable and reliable prediction results. This advantage is particularly evident in the shear strain component, demonstrating the special handling of shear deformation by the strain field physical constraint embedding mechanism. Other methods have larger error distribution ranges and higher median positions, indicating poorer prediction stability. Experiments demonstrate that the method of this invention, through physical constraints and feature decoupling, not only improves classification accuracy but also provides more accurate and reliable predictions of specific physical quantities.

[0133] Furthermore, this embodiment of the invention also analyzes the impact of physical constraint strength on recognition performance, explores the optimization range of the physical constraint strength coefficient, and analyzes its balancing effect on recognition accuracy and physical consistency. Figure 6 A corresponding schematic diagram is shown. In practical implementation, other parameters are fixed through experiments, the physical constraint strength coefficient is adjusted, and two key indicators are measured: recognition accuracy (left vertical axis, unit: %) and physical residual (right vertical axis, dimensionless), reflecting the degree to which the prediction results violate physical laws. Experimental results show that when A value of 0 indicates no physical constraints, resulting in the highest physical residual (peak of the red curve), proving that ignoring physical laws leads to prediction distortion. At a value of 0.6 (gray dashed line), the recognition accuracy (blue curve) reaches its peak, while the physical residual (red curve) drops below 2, verifying the rationality of the parameter settings.

[0134] Furthermore, the embodiments of the present invention also analyze the performance changes of different methods under gradually increasing noise levels, and evaluate the robustness of each method against noise interference in real-world environments. Figure 7 A corresponding schematic diagram is shown. (Refer to...) Figure 7 The horizontal axis represents the noise level, gradually increasing from 0 to 0.5, representing the intensity of noise introduced into the data. The vertical axis represents the recognition accuracy, ranging from 0.4 to 1. Experimental results show that the performance of all methods decreases with increasing noise levels, but the method of this invention exhibits the smallest decrease, maintaining a relatively high accuracy even at high noise levels. Experiments demonstrate that the method of this invention, through strain-sensitive feature weighting and multi-scale feature extraction, can effectively suppress noise interference and extract robust feature representations. In contrast, traditional numerical simulation methods are most sensitive to noise and exhibit the most significant performance degradation, illustrating the limitations of a single data source in real-world environments.

[0135] Further, the embodiment of the present application also provides a rock-soil deformation identification device based on a neural network, Figure 8 The structural schematic diagram of the embodiment of the present application is shown, and the embodiment of the present application is described in detail below with reference to Figure 8 The device comprises: a data monitoring module 100, configured to monitor data of a target rock-soil body, to obtain monitoring data of the target rock-soil body and environmental factor data of an environment in which the monitoring data is located; a data processing module 200, configured to determine a multi-source rock-soil fusion tensor corresponding to the monitoring data from a pre-constructed multi-source rock-soil fusion tensor set based on the environmental factor data; the multi-source rock-soil fusion tensor set is constructed based on monitoring data, simulation data and environmental factor data of a preset rock-soil body; a feature extraction module 300, configured to extract branch features, strain sensitive features and multi-scale space-time features corresponding to the multi-source rock-soil fusion tensor; the branch features comprise elastic features and plastic features; an execution module 400, configured to perform collaborative classification on the branch features, the strain sensitive features and the multi-scale space-time features based on a pre-constructed self-attention mechanism, to determine a deformation mode corresponding to the target rock-soil body; and an output module 500, configured to perform deformation analysis on the target rock-soil body based on the deformation mode.

[0136] The rock-soil deformation identification device based on the neural network provided by the embodiment of the present application has the same implementation principle and generated technical effects as the method embodiment, and for brevity of description, the part of the device embodiment not mentioned can be referred to the corresponding content in the foregoing method embodiment.

[0137] The embodiment of the present application also provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the method of any of the foregoing embodiments when executing the computer program. Figures 1 to 2 The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by the processor to perform the steps of the method of any of the foregoing embodiments. Figures 1 to 2 The embodiment of the present application also provides a structural schematic diagram of an electronic device, as shown in Figure 9 The electronic device comprises a processor 91 and a memory 90, the memory 90 stores computer executable instructions executable by the processor 91, and the processor 91 executes the computer executable instructions to implement the method of any of the foregoing embodiments. Figures 1 to 2 The embodiment of the present application also provides a structural schematic diagram of an electronic device, as shown in Figure 9In the illustrated embodiment, the electronic device further includes a bus 92 and a communication interface 93, wherein the processor 91, the communication interface 93 and the memory 90 are connected through the bus 92. The memory 90 can include a high-speed random access memory (RAM) and can also include a non-volatile memory such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 93 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used. The bus 92 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture) bus, wherein AMBA defines three types of buses, including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus, and an AXI (Advanced eXtensible Interface) bus. The bus 92 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9Only one bidirectional arrow is used to represent multiple buses or other forms of electrical connections that exist among various buses. The processor 91 can be a general purpose integrated circuit chip with a processing capability. In the implementation, each step of the above method can be completed by integrated logic circuit of hardware in the processor 91 or by instructions in the form of software. The processor 91 described above can be a general purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware decoding processor for execution, or be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a random access memory, a flash memory, a read only memory, a programmable read only memory, an electrically erasable programmable memory, a register, or other mature storage medium in the art. The storage medium is located in the memory, and the processor 91 reads information in the memory, and combines the hardware to complete the foregoing steps. Figures 1 to 2Any of the methods described above. The embodiments of the present application provide a rock and soil deformation identification method and device based on a neural network and a computer program product of an electronic device. The computer program product includes a computer readable storage medium storing program codes. The program codes include instructions for executing the method described in the foregoing method embodiments. The specific implementation can be referred to the method embodiments, and will not be described here. Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiments, and will not be described here. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes. In the description of the present application, it should be noted that the terms "first", "second", "third" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance. Finally, it should be noted that: the above embodiments are only specific embodiments of the present application, used to illustrate the technical solutions of the present application, and not to limit it. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent replacements to some technical features. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A neural network-based geotechnical deformation identification method, characterized in that, The method comprises: monitoring data of a target rock-soil mass to obtain monitoring data of the target rock-soil mass and environmental factor data of an environment in which the monitoring data is located; determining a multi-source rock-soil fusion tensor corresponding to the monitoring data from a pre-constructed multi-source rock-soil fusion tensor set based on the environmental factor data; the multi-source rock-soil fusion tensor set is constructed based on monitoring data, simulation data and environmental factor data of a pre-set rock-soil mass; extracting branch features, strain sensitive features and multi-scale space-time features corresponding to the multi-source rock-soil fusion tensor; the branch features include elastic features and plastic features; wherein the elastic features and the plastic features are extracted from the multi-source rock-soil fusion tensor based on rock-soil mechanics theory through a tensor decomposition algorithm; the strain sensitive features are determined based on shear strain of each space-time point in the multi-source rock-soil fusion tensor; and the multi-scale space-time features are determined by using a multi-resolution analysis method on the multi-source rock-soil fusion tensor; cooperatively classifying the branch features, the strain sensitive features and the multi-scale space-time features based on a pre-constructed self-attention mechanism to determine a deformation mode corresponding to the target rock-soil mass; performing deformation analysis on the target rock-soil mass based on the deformation mode; the step of cooperatively classifying the branch features, the strain sensitive features and the multi-scale space-time features based on the pre-constructed self-attention mechanism to determine the deformation mode corresponding to the target rock-soil mass comprises: converting the branch features, the strain sensitive features and the multi-scale space-time features into a global feature vector, mapping the global feature vector to a query space to generate a query representation matrix; performing dot product processing on the query representation matrix using a scaled dot product attention mechanism to determine a mode correlation graph corresponding to the global feature vector; the mode correlation graph includes deformation mode similarity between samples; performing matrix multiplication processing on the mode correlation graph and the global feature vector to generate a fusion output, and determining the deformation mode corresponding to the target rock-soil mass based on the fusion output.

2. The method of claim 1, wherein, The method further comprises: acquiring a numerical simulation tensor, a monitoring tensor and a corresponding environmental factor tensor pre-collected for a pre-set rock-soil mass; performing linear combination on the numerical simulation tensor, the monitoring tensor and the environmental factor tensor to generate an interaction tensor of the pre-set rock-soil mass; performing dimension unification on the numerical simulation tensor, the monitoring tensor and the environmental factor tensor, and performing data fusion on the dimension-unified numerical simulation tensor, monitoring tensor and environmental factor tensor and the interaction tensor to generate a multi-source rock-soil fusion tensor set of the pre-set rock-soil mass.

3. The method of claim 1, wherein, The step of extracting branch features of the multi-source rock-soil fusion tensor comprises: performing first convolution processing on the multi-source rock-soil fusion tensor through a pre-set neural network model to extract an elastic feature tensor corresponding to the multi-source rock-soil fusion tensor; and performing second convolution processing on the monitoring data through the pre-set neural network model to determine a gating weight corresponding to the multi-source rock-soil fusion tensor; range compressing the elastic feature tensor, gated selecting the range compressed elastic feature tensor based on the gating weight, determining a plastic feature tensor of the multi-source rock-soil fusion tensor; concatenating the elastic feature tensor and the plastic feature tensor to determine a branch feature of the multi-source rock-soil fusion tensor.

4. The method of claim 1, wherein, The step of extracting the strain sensitive feature of the multi-source rock-soil fusion tensor comprises: extracting a strain component of the multi-source rock-soil fusion tensor, determining a gradient of the strain component in a spatial direction and a time dimension, and generating a strain gradient vector; performing importance identification on a position feature of the strain gradient vector to determine a strain sensitive weight corresponding to the position feature; performing feature weighting on the multi-source rock-soil fusion tensor using the strain sensitive weight to determine a strain sensitive feature of the multi-source rock-soil fusion tensor.

5. The method of claim 1, wherein, The step of extracting the multi-scale spatio-temporal feature of the multi-source rock-soil fusion tensor comprises: performing a dilated convolution operation on the multi-source rock-soil fusion tensor by using different dilation rates in parallel through a plurality of preset dilated convolution branches to capture spatio-temporal features of the multi-source rock-soil fusion tensor at different scales; summing the spatio-temporal features to obtain a multi-scale spatio-temporal feature of the multi-source rock-soil fusion tensor.

6. The method of claim 1, wherein, The method further comprises: extracting the branch feature, the strain sensitive feature and the multi-scale spatio-temporal feature corresponding to the multi-source rock-soil fusion tensor through a preset neural network model; wherein the preset neural network model is trained based on a preset physical information hybrid loss function; The method further comprises: training the preset neural network model using a preset training sample set to determine an initial prediction output; calculating a cross-entropy loss of the preset neural network model based on the initial prediction output, and generating a dynamic stiffness matrix according to the initial prediction output; the dynamic stiffness matrix is determined based on a material constitutive relationship corresponding to the training sample and a finite element discretization manner of the preset neural network model; constructing a physical constraint residual loss of the preset neural network model based on the stiffness matrix; performing weighted fusion on the physical constraint residual loss and the cross-entropy loss to generate a physical information hybrid loss function.

7. The method of claim 6, wherein, The method further comprises: obtaining a pre-constructed multi-source rock-soil fusion tensor training sample; physically constraining the multi-source rock-soil fusion tensor training sample based on a linear physical response of a rock-soil body corresponding to the multi-source rock-soil fusion tensor training sample to generate a physically constrained tensor; constructing a training sample set based on the physically constrained tensor.

8. A neural network-based geotechnical body deformation identification device, characterized by, The device comprises: a data monitoring module configured to monitor a target rock-soil body to obtain monitoring data of the target rock-soil body and environmental factor data of an environment in which the monitoring data is located; a data processing module configured to determine a multi-source rock-soil fusion tensor corresponding to the monitoring data from a pre-constructed multi-source rock-soil fusion tensor set based on the environmental factor data; the multi-source rock-soil fusion tensor set is constructed based on monitoring data, simulation data and environmental factor data of a preset rock-soil body. The feature extraction module is configured to extract branch features, strain-sensitive features, and multi-scale space-time features corresponding to the multi-source rock-soil fusion tensor; the branch features include elastic features and plastic features; the elastic features and the plastic features are extracted from the multi-source rock-soil fusion tensor based on a rock-soil body mechanics theory and a tensor decomposition algorithm; the strain-sensitive features are determined based on shear strain of each space-time point in the multi-source rock-soil fusion tensor; and the multi-scale space-time features are determined by using a multi-resolution analysis method on the multi-source rock-soil fusion tensor. The execution module is configured to perform collaborative classification on the branch features, the strain-sensitive features, and the multi-scale space-time features based on a pre-constructed self-attention mechanism, and determine a deformation mode corresponding to the target rock-soil body. The output module is configured to perform deformation analysis on the target rock-soil body based on the deformation mode. The execution module is further configured to convert the branch features, the strain-sensitive features, and the multi-scale space-time features into a global feature vector, map the global feature vector to a query space to generate a query representation matrix, perform dot product processing on the query representation matrix by using a scaled dot product attention mechanism, determine a mode correlation graph corresponding to the global feature vector, and generate a fusion output by performing matrix multiplication on the mode correlation graph and the global feature vector, and determine the deformation mode corresponding to the target rock-soil body based on the fusion output.

9. An electronic device, comprising: The electronic device includes a processor and a memory, the memory stores machine executable instructions that can be executed by the processor, and the processor executes the machine executable instructions to implement the rock-soil body deformation recognition method based on a neural network according to any one of claims 1 to 7.

Images