A method for predicting the compressive strength of anti-slide pile

Abstract

The present application relates to the technical field of artificial intelligence and geotechnical engineering, in particular to a kind of anti-slide pile compressive strength prediction method, specifically as follows: the relevant data of anti-slide pile are collected and labeled and then pretreated, the collected data are modulated, denoised and compressed in turn, then input into the anti-slide pile compressive strength prediction model constructed to predict the compressive strength grade and the probability of geological risk, and the model is iteratively trained and parameter updated, to generate the trained model, which is mainly composed of a geological attention feature reweighting module, a physically driven residual gate module and a multi-task interference output layer;The newly collected data are input into the trained model after preprocessing to generate the final prediction result.The present application lays a solid foundation for model analysis and prediction by preprocessing the collected data, and through the multi-module fusion anti-slide pile compressive strength prediction model, accurate prediction and real-time early warning of compressive strength can be realized.

Description

A method for predicting the compressive strength of anti-slide piles Technical Field

[0001] This invention relates to the fields of artificial intelligence and geotechnical engineering, and in particular to a method for predicting the compressive strength of anti-slide piles. Background Technology

[0002] In landslide prevention and slope stabilization engineering, anti-slide piles are one of the most commonly used support structures, and their compressive strength is a key indicator for evaluating the pile's bearing capacity and the project's safety. Traditional methods for assessing the compressive strength of anti-slide piles mainly rely on on-site static load tests or empirical formula calculations. These methods are not only costly and time-consuming, but also difficult to obtain continuously distributed strain response information in complex geological environments. With the development of distributed fiber optic sensing technology, high-density strain monitoring can be achieved along the entire length of the pile, providing a data foundation for real-time assessment of the stress state of anti-slide piles.

[0003] Existing technologies still have many shortcomings in practical applications: most existing technologies rely solely on raw monitoring data from strain sensors for modeling, failing to establish a quantitative correlation between geological parameters and mechanical responses, making it difficult for models to identify the true strain characteristics of soil abrupt change zones; conventional wavelet or filtering denoising methods use a uniform threshold, ignoring the differences in hardness and noise characteristics of different soil layers, easily leaving noise in soft soil layers and losing effective micro-strain characteristics in hard soil layers, affecting data quality; existing feature reduction or principal component analysis methods do not consider geological constraints, weakening key information in geologically sensitive areas during compression, resulting in insufficient robustness of prediction models to complex strata; conventional neural networks rely solely on data fitting, lacking physical and geological constraints, easily learning non-physical features, leading to inconsistencies between prediction results and actual mechanical responses, and unstable performance in areas with abrupt changes or anomalies in geological conditions.

[0004] Therefore, this invention proposes a method for predicting the compressive strength of anti-slide piles to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention develops a method for predicting the compressive strength of anti-slide piles. This invention lays a solid foundation for model analysis and prediction by preprocessing the collected data. Through a multi-module integrated anti-slide pile compressive strength prediction model, accurate prediction and real-time early warning of compressive strength can be achieved.

[0006] The technical solution of this invention to solve the technical problem is a method for predicting the compressive strength of anti-slide piles, comprising the following steps:

[0007] S1. Collect axial strain data of anti-slide piles under load through sensors, and mark the collected data with compressive strength grade and geological risk. At the same time, obtain the soil layer type information of each anti-slide pile, and measure the soil friction angle and cohesion parameters to form a discrete geological parameter sequence.

[0008] S2. Construct geophysical equations through geological parameter sequences, and then modulate the collected data using geophysical equations to generate modulated strain data;

[0009] S3. Perform wavelet packet denoising on the modulated strain data and adaptively adjust the wavelet packet denoising threshold according to the soil layer type to obtain denoised strain data.

[0010] S4. The denoised strain data is compressed using the constructed autoencoder network structure to generate latent features. Then, a total optimization objective function is constructed. The autoencoder network parameters are trained by minimizing this function. The autoencoder network parameters are iteratively trained and updated to obtain the optimized autoencoder network.

[0011] S5. Construct a prediction model for the compressive strength of anti-slide piles. Sequentially construct a geological attention feature reweighting module, a physics-driven residual gating module, and a multi-task interference output layer. Input latent features and output the predicted probabilities of compressive strength level and geological risk. Then calculate the multi-task joint loss function and iteratively train and update the parameters of the prediction model for the compressive strength of anti-slide piles until a well-trained model is obtained.

[0012] S6. After the newly collected data has been preprocessed in steps S2 to S4, it is input into the trained model to generate the final prediction result.

[0013] S1 is as follows:

[0014] Strain sensor monitoring points are arranged at equal intervals along the depth direction of the anti-slide pile, and data is collected in real time using distributed fiber optic sensing technology.

[0015] By conducting borehole exploration along the corresponding locations of the anti-slide pile depth, information on the soil layer type at each pile depth was obtained, and the soil friction angle and cohesion parameters at each depth point were determined through indoor geotechnical tests.

[0016] The collected data are labeled, and the compressive strength grade is labeled based on the ultimate compressive strength of the pile body measured by the on-site static load test. The geological risk label is labeled by the geotechnical engineer based on the comprehensive judgment results of the degree of fracture development in the borehole core, the abrupt change gradient of cohesion, and historical landslide records.

[0017] S2 is as follows:

[0018] Geophysical equations include the calculation of geological modulation vectors and geological gradient vectors;

[0019] Specifically, based on the soil friction angle and cohesion parameters, the geological modulation vector is calculated through the normalized deviation of the friction angle and the hyperbolic tangent function of the cohesion; the geological gradient vector is calculated through the absolute rate of change of cohesion at adjacent pile depths.

[0020] The collected data is then modulated using geophysical equations. Specifically, the collected strain data and the geological modulation vector are subjected to the Hadamard product, and the modulated strain data is obtained by combining the scaled geological gradient vector.

[0021] S3 is as follows:

[0022] First, based on the acquired soil type information, the soil types are pre-classified, and the modulated strain data is divided into sub-vectors corresponding to different soil types. At the same time, the wavelet packet denoising threshold is dynamically calculated based on the cohesion parameters of each type of soil.

[0023] Then, Symlet wavelet packet decomposition is performed on the strain vector of each soil layer. Specifically, a continuously differentiable threshold function is used to process the wavelet coefficients to obtain the representation of the threshold processing function.

[0024] Finally, the wavelet coefficients processed by the threshold processing function are reconstructed using wavelet packets to recover the denoised strain data of each soil layer. Then, the sub-vectors are integrated into complete denoised strain data through the soil layer indicator function to obtain the denoised strain data.

[0025] S4 is as follows:

[0026] First, the constructed autoencoder network structure is an encoder-decoder. The encoder compresses the denoised strain data and incorporates the geological modulation vector and geological gradient vector contained in the geophysical equation into the feature compression process. The encoder outputs the latent features. At the same time, a geological regularization loss term is constructed based on the deviation between the soil friction angle parameter and the global mean, which forces the encoder to reduce the weight in the friction angle anomalous region.

[0027] Then, the overall optimization objective function of the autoencoder is constructed by combining the reconstruction error, sparsity constraint, and geological regularization term. The parameters of the autoencoder network are trained by minimizing this function. Here, the reconstruction error is the absolute value of the difference between the latent features output by the encoder and the reconstructed latent features output by the decoder; the sparsity constraint refers to the L1 norm of the latent features.

[0028] Finally, the autoencoder network parameters are iteratively trained and updated. The specific process is as follows: after initializing the weight matrix and bias vector of the autoencoder network, the denoised strain data vector is input into the encoder. After being enhanced by the geological modulation vector and geological gradient vector, a 128-dimensional latent feature vector is generated, and the decoder reconstructs the output vector. In each iteration, the total optimization objective function is calculated, and the Adam optimizer is used for gradient descent to update the weight matrix and bias parameters. An initial learning rate is set, and the learning rate is set to decay by 10% every 100 rounds. At the same time, an iteration stopping condition is set until the condition is met to end the training, and the optimized autoencoder network is obtained.

[0029] The specific operation of the geological attention feature reweighting module is as follows:

[0030] The geological attention feature reweighting module adopts the geological perception attention mechanism. First, it constructs a geological query vector and generates a geological query vector by using the global mean of the geological modulation vector and the global maximum value of the geological gradient vector through linear transformation and ReLU activation function.

[0031] Then, the key vector and attention weights are calculated. The compressed latent features are transformed linearly to obtain the key vector, which is then scaled and multiplied by the geological query vector. Finally, the attention weight distribution is obtained by Softmax normalization. The key vector is calculated from the trainable key weight matrix and latent features.

[0032] Finally, the compressed features are reweighted using attention weights, normalized geological gradient vectors, and normalized geological modulation vectors to obtain geological attention-enhanced features.

[0033] The operation of the physical-driven residual gating module is as follows:

[0034] The physical-driven residual gating module adopts a physical gating mechanism based on the Mohr-Coulomb criterion, embedding the geotechnical yield condition into the residual connection;

[0035] Specifically, the Mohr-Coulomb yield function value is calculated based on the estimated stress state, friction angle, and cohesion. The Mohr-Coulomb yield function value represents the yielding degree of the soil and rock mass under the current stress state. Then, the yield function value, the global mean of the geological attention enhancement feature vector, and the global maximum value of the geological gradient vector are mapped through linear transformation and the Sigmoid activation function to obtain the physical gating vector. Finally, the physical gating vector is introduced on the basis of the standard residual connection to perform element-wise weighting on the residual function output, thereby updating the features output by the residual neural network layer by layer.

[0036] The specific operations of the multi-task interference output layer are as follows:

[0037] The multi-task interference output layer adopts a multi-task joint learning framework to simultaneously predict compressive strength level and geological risk. An adaptive weight adjustment mechanism is used to strengthen risk monitoring in areas of geological abrupt changes, as detailed below:

[0038] (1) The final layer features of the residual neural network in the physical-driven residual gating module are concatenated with the global mean of the geological attention enhancement features and the geological gradient vector, and the geological risk probability is predicted through a fully connected layer and the Sigmoid function.

[0039] (2) The final layer features and geological risk prediction probability of the residual neural network in the physical-driven residual gating module are predicted by the fully connected layer and the Softmax function to predict the probability distribution of compressive strength level;

[0040] (3) Based on the magnitude of the geological gradient vector, the adaptive weight of the geological regularization loss term is calculated by the Sigmoid function.

[0041] The calculation process of the joint loss function for multiple tasks is as follows:

[0042] By combining the cross-entropy loss from intensity level prediction and the geological regularization loss term, a multi-task joint loss function is constructed using adaptive weights.

[0043] The cross-entropy loss of the strength grade prediction is calculated using standard cross-entropy based on the predicted probability distribution of the compressive strength grade and the actual compressive strength grade label.

[0044] S6 is detailed below:

[0045] After preprocessing, the newly collected monitoring data of anti-slide piles is input into the trained anti-slide pile compressive strength prediction model, which outputs the probability distribution of compressive strength levels. The level with the highest probability is taken as the final strength prediction value. At the same time, the geological risk probability is output. When the probability value exceeds the preset risk threshold, a sliding risk warning is triggered.

[0046] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0047] This invention introduces a geophysical modulation mechanism into the prediction of compressive strength of anti-slide piles. By constructing a geological modulation vector and a geological gradient vector through friction angle and cohesion, it achieves physical coupling between geological parameters and strain data, enabling the model to generate directional feature amplification in soil abrupt change zones, thereby improving the identification ability at key geological interfaces. The multi-scale geologically correlated wavelet packet denoising method proposed in this invention adaptively adjusts the wavelet threshold based on soil cohesion, establishing a negative correlation between the threshold and soil hardness. This achieves a dynamic balance between noise suppression in soft soil layers and preservation of micro-features in hard soil layers, solving the problem of over-smoothing features caused by conventional uniform threshold denoising. The geologically constrained sparse autoencoder network constructed in this invention introduces friction angle deviation and geological regularization terms in the feature compression stage, enabling potential features to simultaneously satisfy data reconstructibility and geological sensitivity, prioritizing the preservation of features in key geological areas, thereby effectively enhancing the model's adaptability to complex geological environments. This invention achieves dual embedding of geological constraints and mechanical laws through a geologically perceptive attention mechanism and a physically gated residual network driven by the Moore-Coulomb criterion. It dynamically controls the update amplitude during feature propagation, ensuring that the prediction results are both highly accurate and conform to engineering physical rationality. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0049] Figure 1 is a schematic diagram of the method flow of the present invention.

[0050] Figure 2 is a heat map showing the distribution of the collected raw strain data.

[0051] Figure 3 is a heat map showing the effect of geological modulation vectors on enhancing the characteristics of soil interface.

[0052] Figure 4 is a heat map showing the effect of geological gradient vector on the enhancement of soil interface features.

[0053] Figure 5 is a heat map of the distribution of strain data after modulation.

[0054] Figure 6 is a comparison of the denoising effects of the denoising method of the present invention and the conventional wavelet denoising method on the backfill layer.

[0055] Figure 7 is a comparison of the denoising effects of the denoising method of the present invention and the conventional wavelet denoising method on silty clay layers.

[0056] Figure 8 is a comparison of the denoising effects of the denoising method of the present invention and the conventional wavelet denoising method on strongly weathered rock strata.

[0057] Figure 9 shows a comparison of the average prediction error of the method of the present invention and conventional neural networks at different intensity levels.

[0058] Figure 10 is a comparison of the accuracy of the method of the present invention and conventional neural networks in identifying different geological risks. Detailed Implementation

[0059] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0060] Example 1

[0061] A method for predicting the compressive strength of anti-slide piles includes the following steps:

[0062] S1. Collect axial strain data of anti-slide piles under load through sensors, and mark the collected data with compressive strength grade and geological risk. At the same time, obtain the soil layer type information of each anti-slide pile, and measure the soil friction angle and cohesion parameters to form a discrete geological parameter sequence.

[0063] S2. Construct geophysical equations through geological parameter sequences, and then modulate the collected data using geophysical equations to generate modulated strain data;

[0064] S3. Perform wavelet packet denoising on the modulated strain data and adaptively adjust the wavelet packet denoising threshold according to the soil layer type to obtain denoised strain data.

[0065] S4. The denoised strain data is compressed using the constructed autoencoder network structure to generate latent features. Then, a total optimization objective function is constructed. The autoencoder network parameters are trained by minimizing this function. The autoencoder network parameters are iteratively trained and updated to obtain the optimized autoencoder network.

[0066] S5. Construct a prediction model for the compressive strength of anti-slide piles. Sequentially construct a geological attention feature reweighting module, a physics-driven residual gating module, and a multi-task interference output layer. Input latent features and output the predicted probabilities of compressive strength level and geological risk. Then calculate the multi-task joint loss function and iteratively train and update the parameters of the prediction model for the compressive strength of anti-slide piles until a well-trained model is obtained.

[0067] S6. After the newly collected data has been preprocessed in steps S2 to S4, it is input into the trained model to generate the final prediction result.

[0068] In a specific implementation, S1 is as follows:

[0069] Strain sensor monitoring points are arranged at equal intervals along the depth direction of the anti-slide pile. The axial strain data of the anti-slide pile under load is collected in real time through distributed optical fiber sensing technology to form the original strain data vector.

[0070] Simultaneously, borehole exploration was carried out at the corresponding locations along the pile depth to obtain information on the soil layer type at each pile depth. The soil friction angle and cohesion parameters at each depth point were determined through indoor geotechnical tests to form a discrete geological parameter sequence.

[0071] In practice, distributed optical fiber sensing technology is used to collect data from 4,096 monitoring points. By utilizing the continuous spatial sampling characteristics of optical fiber, the pile body is equivalently discretized into 4,096 virtual measuring points.

[0072] The collected data is labeled, and the data labeling includes two categories:

[0073] The compressive strength grade label is based on the ultimate compressive strength of the pile body measured by on-site static load test, and is classified into C10-C50 concrete strength grades according to engineering specifications.

[0074] Geological risk labels are determined by geotechnical engineers based on the degree of fracture development in borehole cores, abrupt changes in cohesion gradient, and historical landslide records, indicating whether the pile body has a risk of sliding (0 / 1 binary classification).

[0075] It should be noted that, during data acquisition, if the implementation conditions are limited, a high-resolution data sequence can be established in the direction of pile depth by numerical simulation or high-density interpolation reconstruction of limited sensor data, thereby discretizing the pile into 4096 data points.

[0076] In a specific implementation, S2 is as follows:

[0077] In the task of predicting the compressive strength of anti-slide piles, strain monitoring data is affected by high-frequency noise, and conventional methods often ignore the physical correlation between geological parameters such as soil friction angle and cohesion and strain response, failing to establish the coupling relationship between geology and mechanics. This easily leads to the prediction model failing to fully utilize geological information to enhance the characteristic response. Therefore, this invention constructs a modulation matrix through geophysical equations, encoding geological parameters as physical constraints, thereby enhancing the characteristic response of abrupt soil changes. The specific steps are as follows:

[0078] 1) Calculate the geological modulation vector

[0079] Based on the soil friction angle and cohesion parameters, a geological modulation vector is calculated using the normalized deviation of the friction angle and the hyperbolic tangent function of the cohesion. This achieves physical constraint modulation of the original strain data, enhancing the characteristic response of abrupt soil transition zones, and is expressed as:

[0080]

[0081] In the formula, The geological modulation vector is represented in the first... The feature values ​​at each location are used to modulate the strain data, corresponding to the sampling point at the j-th pile depth;

[0082] The soil friction angle at the j-th pile depth represents the soil shear strength parameter.

[0083] This represents the global mean of the friction angles of all pile locations in the training set;

[0084] This represents the standard deviation of the friction angles of all pile locations in the training set;

[0085] j is the index of the sampling point in the pile depth direction, with a value range from 1 to 4096, corresponding to the location of monitoring points evenly distributed along the pile depth direction;

[0086] Represents an exponential function;

[0087] Represents the hyperbolic tangent function;

[0088] This represents the soil cohesion at the j-th pile depth, a parameter characterizing the soil's bond strength.

[0089] In practical implementation, the soil friction angle at the j-th pile depth is... Cohesion of the soil layer at pile depth j The discrete geological parameter sequence along the pile depth was obtained through borehole exploration and laboratory geotechnical tests, and the global mean of the friction angle of all pile locations in the training set was also obtained. and standard deviation The angle of soil friction of all piles in the training set was obtained through global statistics.

[0090] It should be noted that, The term is used to map the friction angle deviation to weights, and the friction angle deviation is mapped to weights through an exponential decay function. near The time weight approaches 1, preserving the response of typical soil layers, when Deviation Time-weighted decay suppresses interference from atypical soil layers and highlights geological anomaly areas; nonlinear amplification of the difference between soft and hard soil layers; the hyperbolic tangent function has a large gradient near zero. The term is used to compress the cohesion value to The range is sensitive to small changes in cohesion, enhances the characteristics of abrupt changes in soil layers, avoids instability caused by excessively large values, and retains the positive and negative sign information of cohesion, making it easier to distinguish the bonding characteristics of soil during modulation and enhancing the effectiveness of physical constraints.

[0091] 2) Calculate the geological gradient vector

[0092] The geological gradient vector, calculated based on cohesion parameters, is used as the absolute rate of change of cohesion at adjacent pile depths to enhance the characteristic response of abrupt soil changes, expressed as:

[0093]

[0094] In the formula, The geological gradient vector is represented in the first position. The eigenvalues ​​at each location, padded with zeros at the ends, characterize the absolute rate of change of cohesion between adjacent points, and physically reflect the abrupt change in shear strength at the soil interface.

[0095] Indicates the first Cohesion of the soil layer at the depth of the pile.

[0096] 3) Calculate the modulated strain data

[0097] The original strain data and the geological modulation vector are subjected to the Hadamard product operation, and combined with the scaled geological gradient vector to obtain the modulated strain data vector, thus achieving the fusion of geological features and strain data, represented as:

[0098]

[0099] In the formula, This represents a modulated strain data vector with a dimension of 4096;

[0100] This represents the original strain data vector, with a dimension of 4096. It is the original strain data vector collected by the strain sensor, corresponding to the strain measurement values ​​at different depths of the pile. It contains high-frequency noise and the true strain response, and has spatiotemporal continuity but is affected by environmental interference.

[0101] This represents the Hadamard product, which is the product of elements.

[0102] This represents a geological modulation vector with a dimension of 4096, and its first... The feature values ​​are ;

[0103] This represents the geological gradient coefficient, which controls the weight of the influence of the geological gradient. The preferred value is 0.05.

[0104] This represents the geological gradient vector, with dimensions 4096, and its i-th... The feature values ​​are .

[0105] It should be noted that, through The exponential decay term establishes a probabilistic relationship between the statistical distribution of the friction angle and the strain response, while The project introduces nonlinear geological constraints based on cohesion, and simultaneously... This condition causes strain data to be enhanced in the typical friction angle region and attenuated in the abnormal region, while The feature enhancer injects gradient information from the soil interface and then converts geological parameters into strain data through physical modulation. This generates directional feature amplification in areas of abrupt cohesion change, such as the soil-rock interface, enabling subsequent models to capture critical slip signals that are ignored by conventional methods.

[0106] In a specific implementation, S3 is as follows:

[0107] The modulated strain data still contains high-frequency noise from the sensor. This invention adaptively adjusts the wavelet packet denoising threshold according to soil type to achieve differentiated denoising based on geological attributes. The specific steps are as follows:

[0108] 1) Segmenting strain data by soil layer type

[0109] Based on pre-defined soil layer types, the modulated strain data is segmented into sub-vectors corresponding to different soil layers, providing a data basis for differential denoising, as shown below:

[0110]

[0111] In the formula, Indicates the first The modulated strain data subvector corresponding to the soil layer contains the th layer. Data from all strain monitoring points within a soil depth range;

[0112] This indicates the total number of soil layer types, determined based on the geological survey report;

[0113] This represents a set of data subvectors, where each subvector is arranged consecutively in the original data.

[0114] Indexes representing soil layer types, from 1 to .

[0115] In the specific implementation process, the total number of soil layer types for slope engineering is set. , Corresponding fill layer Silty clay layer Corresponding to strongly weathered rock strata.

[0116] 2) Calculate the adaptive threshold of the soil layer

[0117] The wavelet packet denoising threshold is dynamically calculated based on the cohesion parameters of each soil layer. The threshold is negatively correlated with the soil hardness. A smaller threshold is used for hard soil layers to retain more detailed features, while a larger threshold is used for soft soil layers to enhance the denoising effect. This is expressed as follows:

[0118]

[0119] In the formula, Indicates the first The wavelet packet denoising threshold for each soil layer is used to control the denoising intensity of that soil layer.

[0120] This represents the basic threshold, preferably 0.1, which serves as the basis for threshold calculation.

[0121] This represents the average soil cohesion value of the k-th soil layer;

[0122] This represents the cohesion sensitivity coefficient, with a preferred value of 0.3, which controls the degree of influence of cohesion on the threshold.

[0123] 3) Perform wavelet packet decomposition and thresholding.

[0124] Symlet wavelet packet decomposition is performed on the strain vector of each soil layer. A continuously differentiable threshold function is used to process the wavelet coefficients, achieving a smooth transition near the threshold and avoiding the Gibbs phenomenon caused by conventional hard thresholding. This is expressed as:

[0125]

[0126] In the formula, This represents the threshold processing function, which performs nonlinear compression on the wavelet coefficients;

[0127] The wavelet coefficients after wavelet packet decomposition contain strain information from different frequency bands, and are calculated as follows: , It is the absolute value of the wavelet coefficients after wavelet packet decomposition, which characterizes the intensity of the signal in that frequency band;

[0128] This represents the Symlet wavelet packet decomposition operator, which transforms a time-domain signal into the wavelet domain.

[0129] Represents the natural constant.

[0130] It should be noted that the strain signal of soil and rock exhibits a continuous and gradual change characteristic at the yield critical point. The term represents an exponentially decaying term, avoiding a hard threshold. The truncation effect at the threshold is used to preserve the weak but physically real plastic deformation characteristics near the threshold and suppress Gibbs oscillation artifacts. This indicates the boundary of the threshold processing range. When the absolute value of the wavelet coefficient exceeds this value, it is considered significant noise and is then set to zero.

[0131] 4) Wavelet reconstruction and data integration

[0132] The wavelet coefficients after thresholding are reconstructed using wavelet packets to recover the denoised strain data of each soil layer. Then, the sub-vectors are integrated into complete denoised strain data using the soil layer indicator function, preserving the abrupt change characteristics of the soil layer interface, as shown below:

[0133]

[0134] In the formula, This represents a denoised strain data vector with a dimension of 4096, preserving the abrupt change characteristics of the soil interface.

[0135] Indicates the first The indicator function for each soil layer takes a value of 1 at the corresponding location of that soil layer and 0 at other locations;

[0136] This indicates the Symlet wavelet packet reconstruction operator, which restores the processed wavelet coefficients to the time-domain signal.

[0137] In the specific implementation process, the first The soil layer is a fill layer, corresponding to a pile depth of 0-5 meters, and the indicator function for this soil layer... The value of is 1, and the rest are 0, thus achieving seamless splicing after the soil sub-vector reconstruction.

[0138] It should be noted that the core innovation of this step is the coupling of geological properties and frequency domain noise reduction. A negative correlation mechanism between the threshold and cohesion is established through wavelet packet denoising thresholding. A small threshold is used for high-cohesion hard soil layers to protect the effective high-frequency signal, while a large threshold is used for low-cohesion soft soil layers to suppress random noise. The threshold processing function... The continuous differentiability of the wavelet characteristic matches the physical nature of continuous yielding of soil and rock strain, avoiding the loss of micro-strain features in hard soil layers by conventional wavelet denoising. The adaptive threshold based on geological hardness allows the differential preservation of the micro-fracture signal of the rock layer and the plastic deformation characteristics of the soil layer; the soil and rock strain signal exhibits a continuous and gradual change characteristic at the yield critical point. The term performs thresholding on the wavelet coefficients of the k-th soil layer strain data, based on... Adaptively adjust the noise reduction intensity, in High-value hard soil layers retain weak strain characteristics, Low-value soft soil layers suppress noise by coupling geological properties with frequency domain noise reduction, and differentiatedly retain effective features to avoid over-smoothing or noise residue caused by uniform threshold.

[0139] In a specific implementation, S4 is as follows:

[0140] S4.1. A geologically regularized sparse autoencoder is adopted. By introducing geological constraint terms and sparse constraint terms, the feature preservation of key geological areas is enhanced while reducing data dimensionality. The specific steps are as follows:

[0141] 1) Constructing an autoencoder network structure

[0142] An autoencoder network with an encoder-decoder structure is constructed. The encoder compresses the input data into latent features and incorporates geophysical constraints into the feature compression process by combining geological modulation vectors and geological gradient vectors, thereby enhancing the feature extraction capability for abrupt soil layer changes. This is represented as:

[0143]

[0144] In the formula, The latent feature vector output by the encoder has a dimension of 128 and is a compressed representation of the original strain data.

[0145] This represents the Sigmoid activation function, which maps the linear transformation result to the [0,1] interval;

[0146] This represents the encoder's weight matrix, with dimensions of . , are trainable parameters used for feature transformation;

[0147] This represents the encoder's bias vector, which has a dimension of 128 and is a trainable parameter.

[0148] 2) Calculate the geological regularization loss term

[0149] A geological regularization loss term is constructed based on the deviation between the soil friction angle parameter and the global mean. This forces the encoder to reduce the weight in areas of abnormal friction angles, thereby improving the feature preservation ability of key geological areas. This is expressed as:

[0150]

[0151] In the formula, This represents the geological regularization loss term, used to constrain the consistency between encoder weights and geological parameters;

[0152] Represents the encoder weight matrix The List;

[0153] This represents the L2 norm.

[0154] It should be noted that, The project will use geological modulation vectors and geological gradient vector As prior knowledge, this forces the network to preferentially retain strain characteristics in soil abrupt change zones during the feature compression stage, transforming the physical properties of geological anomalies into feature enhancement signals, thus improving the latent feature vector output by the encoder. Implicit geomechanical coupling information; geological regularization loss term encoder weight matrix The OK The L2 norm is tied to the local deviation of the friction angle when When the value deviates from the global mean, the feature extraction weight of the encoder at that location is penalized, which suppresses redundant features of atypical soil layers such as interlayers or boulders, while increasing the feature contribution of typical friction angle areas. This achieves local sparsification guided by geological attributes, making the compressed features more focused on key geological areas that are strongly correlated with compressive strength.

[0155] S4.2. Construct the overall optimization objective function of the autoencoder by combining reconstruction error, sparsity constraints, and geological regularization term. Train the network parameters by minimizing this function to achieve joint optimization of feature compression and geological constraints, expressed as:

[0156]

[0157] In the formula, This represents all trainable parameters of the autoencoder, including weights and biases;

[0158] Represents the L1 norm;

[0159] This represents the sparsity coefficient, with a preferred value of 0.01, which controls the weight of the sparsity of the latent features.

[0160] This represents the geological regularization coefficient, with a preferred value of 0.2, which controls the weight of the geological constraint term.

[0161] The output vector reconstructed by the decoder is specifically obtained by decoding the latent feature vector Z output by the encoder through the decoder of the autoencoder network. It has a dimension of 4096 and represents the ability of the autoencoder to compress and reconstruct strain information of key geological areas.

[0162] This indicates that the optimization is performed by minimizing all trainable parameters of the autoencoder.

[0163] In practice, the objective function is minimized by the gradient descent algorithm, and all trainable parameters of the autoencoder are updated to achieve joint optimization of feature compression and geological constraints.

[0164] It should be noted that, The L2 norm represents the reconstruction error and measures the decoder output. With input The difference ensures that the main strain characteristics are preserved after dimensionality reduction, avoiding information loss; it should also be noted that... Represents the latent feature vector The L1 norm promotes the output of the latent feature vector of the latent encoder. Sparsity reduces redundancy and highlights key modes related to compressive strength.

[0165] S4.3 After initializing the weight matrix and bias vector of the autoencoder network, the denoised strain data vector is input into the encoder. After being enhanced by the geological modulation vector and the geological gradient vector, a 128-dimensional latent feature vector is generated, and the decoder reconstructs the output vector.

[0166] Each iteration calculates the overall optimization objective function, which includes the reconstruction error L2 norm, latent feature L1 sparsity constraints, and geological regularization loss term.

[0167] The Adam optimizer is used for gradient descent to update the weight matrix and bias parameters. The learning rate is initially set to 0.001 and is set to decay by 10% every 100 rounds.

[0168] The validation set reconstruction error is monitored during training. If the validation loss does not decrease for 5 consecutive rounds, the early stop mechanism is activated, or training is terminated when the maximum number of iterations of 1000 rounds is reached.

[0169] In a specific implementation, S5 is as follows:

[0170] S5.1. Employing a geological perception attention mechanism, geological parameters are transformed into query vectors. Attention weight allocation is used to enhance the feature response of geological anomaly areas. The specific steps are as follows:

[0171] 1) Constructing geological query vectors

[0172] The global mean of the geological modulation vector and the global maximum of the geological gradient vector are used to generate a geological query vector through linear transformation and the ReLU activation function. Geophysical information is embedded in the attention calculation process, as shown below:

[0173]

[0174] In the formula, This represents a geological query vector with 128 dimensions, used to characterize geological features in the attention mechanism;

[0175] This indicates the operation of taking the average. Represents the geological modulation vector The global average;

[0176] This indicates the operation of retrieving the maximum value. Represents the geological gradient vector The global maximum value;

[0177] This indicates that the activation function of the linear unit is modified by introducing a nonlinear transformation;

[0178] This represents the geological query weight matrix, with dimensions of . , are trainable parameters used to map 2D geological parameters to a 128-dimensional query space.

[0179] 2) Calculate the key vector and attention weights

[0180] The compressed features are transformed linearly to obtain the key vector, which is then scaled and multiplied by the geological query vector. Softmax normalization is then applied to obtain the attention weight distribution, and a geological gradient modulation term is introduced to enhance feature focusing at soil interfaces. This is represented as:

[0181]

[0182] In the formula, This represents the geological attention weight vector, which has a dimension of 128, with each element having a value in the range [0,1] and a sum of 1.

[0183] This represents the Softmax normalization function, which transforms the input vector into a probability distribution;

[0184] This is the geological gradient modulation coefficient, preferably set to 0.1, used for adjustment. The contribution of the term to the attention weight enhances feature focusing at the soil interface;

[0185] Represents a geological query vector Transpose of;

[0186] The key vector is represented as follows: The distribution of importance of character features;

[0187] This is the key weight matrix, with dimensions [missing information]. , are trainable parameters;

[0188] This represents the scaling factor, with a preferred value of 128, to prevent the gradient from vanishing due to an excessively large dot product result.

[0189] 3) Perform feature reweighting

[0190] The compressed features are reweighted using attention weights, normalized geological gradient vectors, and normalized geological modulation vectors to obtain a geological attention-enhanced feature vector. This vector, incorporating geophysical constraints, is expressed as:

[0191]

[0192] In the formula, This represents the geological attention-enhanced feature vector, with a dimension of 128, characterizing the depth features that incorporate geophysical constraints;

[0193] It is the geological enhancement coefficient, with a preferred value of 0.1, which controls the influence weight of the gradient and the modulation vector;

[0194] It is a geological gradient vector The normalized vector is calculated as follows: This is used to amplify the characteristic response of regions where cohesion abruptly changes;

[0195] It is a geological modulation vector The normalized vector is calculated as follows: , used to introduce friction angle constraints;

[0196] The first weighted linear transformation weight matrix has dimension . , are trainable parameters used to train on Perform a linear transformation to 128 dimensions;

[0197] The first weighted linear transformation weight matrix has dimension . , are trainable parameters used to train on Perform a linear transformation to 128 dimensions.

[0198] It should be noted that the geological query vector Extracting the geological modulation vector reflecting the overall distribution of friction angle The global mean and the gradient vector reflecting the abrupt change in maximum cohesion. The global maximum value, through Mapping two-dimensional geostatistics to a 128-dimensional query space, making the geological query vector By carrying physical information about soil hardness distribution and the intensity of interface abrupt changes, geological priors can be transformed into attention-guided signals, forcing the network to focus on key areas of engineering risk, such as high friction angle differences or strong cohesion gradient zones.

[0199] S5.2. A physical gating mechanism based on the Mohr-Coulomb criterion is adopted to embed the soil mechanics yield condition into the residual connection. The feature update amplitude is dynamically controlled through the physical criterion. The specific steps are as follows:

[0200] 1) Calculate the Mohr-Coulomb yield function value

[0201] The Mohr-Coulomb yield function value is calculated based on the estimated stress state, friction angle, and cohesion to characterize the yielding degree of the soil and rock mass under the current stress state, providing a mechanical basis for physical gating, and is expressed as follows:

[0202]

[0203] In the formula, This represents the Mohr-Coulomb yield function, used to quantify the distance between the stress state and the yield surface;

[0204] The estimated stress value is calculated using elasticity principles and is expressed as follows: ;

[0205] This represents the global mean of the denoised strain data vector, used to estimate the average stress.

[0206] This represents the average cohesion of all soil layers at the current pile, serving as a representative value for global cohesion.

[0207] This represents the average friction angle of all soil layers at the current pile, serving as the representative value for the global friction angle.

[0208] This refers to the elastic modulus of the pile material, a parameter of engineering materials. For example, the elastic modulus of C30 concrete pile material is... ;

[0209] express The sine value of represents the internal friction characteristics of the soil, and is calculated as follows: ;

[0210] express The cosine value of represents the soil shear resistance, and is calculated as follows: .

[0211] It should be noted that the elastic modulus refers to the stress divided by the strain in a uniaxial stress state. During the elastic deformation stage of a material, the stress and strain obey Hooke's Law, and the proportionality coefficient is called the elastic modulus, with the unit being Pascal.

[0212] 2) Calculate the physical gating vector

[0213] The yield function value, the global mean of the geological attention-enhanced feature vector, and the global maximum of the geological gradient vector are mapped to a physical gating vector through a linear transformation and a sigmoid activation function. This enables continuous adjustment of the yield degree to feature updates, as expressed in:

[0214]

[0215] In the formula, This represents the physical gate vector, whose elements range from [0,1] and are used to control the magnitude of residual feature updates.

[0216] It is a 128-dimensional vector of all ones, used to project the physical gate vector onto a 128-dimensional vector space;

[0217] Represents the geological attention-enhanced feature vector The global mean;

[0218] Represents the geological gradient vector The global maximum value;

[0219] Indicates will , and Concatenate them into a vector;

[0220] This represents the gating weight vector, which is a trainable parameter used to adjust the influence of the yield function value on the gating.

[0221] It should be noted that, , and After the three parts are spliced ​​together, Learning weights makes the physical gate vector Simultaneously responding to mechanical state, characteristic intensity, and geological abrupt changes, when When, it indicates that the rock and soil mass is close to yielding, or when When large, it indicates the presence of strong interfacial abrupt changes, and the soil... Approaching 0 to suppress updates of non-physical features; and make Simultaneously, it senses the distribution of features and geological abrupt changes, such as at the interface between strongly weathered rock layers and soft soil. Even if the stress state does not reach yield, the gating will still moderately attenuate the feature update, while in homogeneous hard soil layers, Smaller sizes fully preserve characteristics, and multi-factor collaborative decision-making enables the gating mechanism to combine mechanical rigor with geological adaptability, overcoming the limitation that a single physical model cannot cover complex stratigraphic conditions.

[0222] 3) Perform feature updates based on physical constraints

[0223] By introducing a physical gate vector on top of the standard residual connection, and performing element-wise weighting on the residual function output, feature updates driven by mechanical criteria are achieved, as follows:

[0224]

[0225] In the formula, Represents the residual neural network of the th The layer's output feature vector has a dimension of 128;

[0226] Represents the residual neural network of the th The layer's output feature vector has a dimension of 128, and the initial output feature vector... Geological attention enhancement features ,Right now ;

[0227] The residual function is represented by a two-layer fully connected network with an output dimension of 128. The mapping relationship is expressed as follows: ;

[0228] is the first weight matrix of the residual function, and are trainable parameters;

[0229] The second weight matrix of the residual function is a set of trainable parameters;

[0230] The first bias term of the residual function is a trainable parameter;

[0231] The second bias term of the residual function is a trainable parameter.

[0232] S5.3. A multi-task joint learning framework is adopted to simultaneously predict compressive strength level and geological risk. An adaptive weight adjustment mechanism is used to strengthen risk monitoring in areas of geological abrupt change. The specific steps are as follows:

[0233] 1) Geological risk prediction

[0234] The final layer feature vector is concatenated with the global mean of the geological attention enhancement feature and the geological gradient vector. The geological risk probability is then predicted using a fully connected layer and a sigmoid function, as follows:

[0235]

[0236] In the formula, This represents the predicted probability of geological risk, with a value range of [value missing]. A higher value indicates a higher geological risk.

[0237] This represents the geological risk classification weight vector, with dimension 1. , are trainable parameters;

[0238] This represents a vector concatenation operation. Item representation will , and By concatenating the feature dimensions, a 257-dimensional feature vector is obtained.

[0239] Represents the geological gradient vector The global mean;

[0240] Represents the residual neural network of the th The layer's output feature vector has a dimension of 128;

[0241] This represents the total number of layers in the residual neural network.

[0242] 2) Intensity level prediction

[0243] Based on the final layer feature vector and the predicted probability of geological risk, the probability distribution of compressive strength level is predicted through a fully connected layer and a softmax function to achieve multi-class classification task output, as shown below:

[0244]

[0245] In the formula, This represents the predicted probability distribution of compressive strength grades, with the dimension being the number of categories, and each element representing the probability of the corresponding strength grade.

[0246] This represents the intensity classification weight matrix, with dimension 1. , are trainable parameters;

[0247] Number of intensity level categories;

[0248] Characterization will and The vectors are concatenated into a 129-dimensional vector.

[0249] 3) Adaptive loss weight calculation

[0250] Based on the magnitude of the geological gradient vector, an adaptive weight for the geological regularization loss term is calculated using the Sigmoid function. This automatically strengthens risk monitoring in areas of geological abrupt change, as expressed below:

[0251]

[0252] In the formula, The adaptive weights of the geological regularization loss term, with values ​​ranging from 1 to 2. This is used to adjust the proportion of the geological regularization loss term in the total loss;

[0253] This represents the scaling factor, with a preferred value of 0.1, ensuring that the adaptive weight of the geological regularization loss term approaches 0.5 in the geologically homogeneous region and approaches 1 in the abrupt change region.

[0254] S5.4. Combining the intensity classification cross-entropy loss and the geological risk binary cross-entropy loss, a multi-task joint loss function is constructed through adaptive weights to achieve dual-task collaborative optimization, expressed as:

[0255]

[0256] In the formula, This represents the joint loss function for multiple tasks, which serves as the overall optimization objective for model training.

[0257] The cross-entropy loss represents the prediction probability distribution of the compressive strength level. The predicted strength distribution is calculated using the standard cross-entropy method, along with the actual compressive strength grade label, to measure the difference between the predicted strength distribution and the actual label.

[0258] It should be noted that the denominator The term ensures that the weight value increases with increasing geological gradient, especially in areas of abrupt geological changes. Approaching 1, in the uniform region, Approaching 0.5, the geological regularization loss term With reduced weights, the model focuses on the primary task of intensity prediction, especially in areas of geological abrupt change. big, Significantly improved The loss contribution forces the model to strengthen the identification of interface slip risk, and uses the geological gradient as a loss weight adjuster to achieve multi-task balance of geological complexity perception. In high-risk areas, the robustness of intensity prediction is improved by strengthening risk supervision, and in stable areas, overfitting of risk labels is avoided.

[0259] S5.5 Initialize the residual gating network and multi-task output layer parameters, input the geological attention-enhanced feature vector into the L-layer residual network, perform physical gating feature update in each round of forward propagation, and simultaneously calculate the intensity level probability distribution and geological risk probability.

[0260] The contribution of the geological regularization loss term is dynamically adjusted by adaptive loss weights, thereby increasing the risk monitoring intensity when the mean of the geological gradient increases.

[0261] During the training of the anti-slide pile compressive strength prediction model, the stochastic gradient descent method was used to optimize the parameters, with a batch size of 32 and a learning rate of 0.005.

[0262] The training termination condition for the anti-slide pile compressive strength prediction model is:

[0263] 1) The joint loss function for multiple tasks did not decrease for 10 consecutive rounds on the validation set;

[0264] 2) Reach the preset 2000 training rounds.

[0265] In a specific implementation, S6 is as follows:

[0266] For newly acquired anti-slide pile monitoring data, firstly, 4096-dimensional original strain data and corresponding pile depth friction angle and cohesion sequence are obtained according to the S1 method. Then, the following steps are performed in sequence: S2 geological feature fusion strain data modulation (generating modulated strain vector), S3 multi-scale geological correlation wavelet packet denoising (outputting denoised strain vector), and S4 geological constraint sparse autoencoder feature compression (generating 128-dimensional potential features).

[0267] The compressed features are input into the trained prediction model, which then undergoes reweighting to enhance the response in abrupt change zones via S5.1 geological attention features. The mechanically constrained features are then transferred through the S5.2 physics-driven residual gating module. Finally, the S5.3 multi-task output layer synchronously generates two predictions:

[0268] The probability distribution of compressive strength grades is used, and the grade with the highest probability is taken as the final strength prediction value.

[0269] Geological risk probability: When the probability value exceeds the threshold of 0.7, a slip risk warning is triggered.

[0270] Example 2

[0271] As shown in Figures 2 to 5, a thermographic analysis of the modulation effect on the soil interface features was conducted to verify the feature enhancement effect of modulation on the original strain data at the soil interface. The experiment compared and demonstrated the spatial variation characteristics of the original strain distribution, geological modulation vector distribution, geological gradient vector distribution, and modulated strain data by simulating strain monitoring data in the depth direction of the pile.

[0272] In the experimental setup, the changes in three typical soil layers were simulated: fill, silty clay, and strongly weathered rock. A transition zone with abrupt changes in geological parameters was set at the soil layer interface, near the red dashed line, to reflect the continuous characteristics of soil layer changes in actual engineering. As shown in the thermogram in Figure 2, the original strain data exhibits a continuous but noise-affected distribution across the entire pile depth. The strain values ​​change slightly near the soil layer interface. The distribution of the geological modulation vector in Figure 3 shows a significant change in the modulation coefficient in the transition region at the soil layer interface, reflecting the impact of abrupt changes in geological parameters on the modulation effect. The geological gradient vector in Figure 4 directly characterizes the intensity of the abrupt change in cohesion parameters, forming a significant peak response at the interface. Most importantly, the distribution of the modulated strain data is shown in Figure 5. It is evident that the characteristics at the soil layer interface are effectively enhanced, the strain changes at the interface are more prominent, and noise in other areas is suppressed to a certain extent.

[0273] This experiment visually demonstrates, through the spatial distribution characteristics of heatmaps, how geological modulation technology transforms geological parameters into physical constraints, effectively enhancing the characteristic response of abrupt soil layer changes. Experimental results show that the technology of this invention can generate directional feature amplification in areas with complex geological conditions, enabling subsequent models to better capture critical slip signals that are easily overlooked by conventional methods, thus providing a more reliable data foundation for accurately predicting the compressive strength of anti-slide piles.

[0274] Example 3

[0275] As shown in Figures 6 to 8, the denoising effects of multi-scale geological correlation wavelet packets were compared to verify the adaptability advantage of the proposed multi-scale geological correlation wavelet packet denoising method under different soil layer conditions. Three typical soil layer environments were configured in the experiment, including low-cohesion fill layer, medium-cohesion silty clay layer, and high-cohesion strongly weathered rock layer. Noise interference consistent with its geological characteristics was applied to each soil layer, and the denoising effects of conventional wavelet denoising and the method of the present invention were compared.

[0276] As can be observed from the comparison images of the three soil layers, in the fill layer shown in Figure 6, due to the soft soil and strong noise interference, the conventional wavelet denoising method, while able to remove some noise, also over-smooths the effective signal, resulting in the loss of some key strain features. The method of this invention, by adaptively adjusting the denoising threshold, effectively suppresses noise while better preserving the true strain response of the soil layer. In the silty clay layer shown in Figure 7, the difference between the two methods is even more pronounced. The conventional method exhibits the phenomenon of some effective features being mistakenly removed when processing medium-hard soil layers, while the method of this invention achieves a good balance between noise suppression and feature preservation. The most significant contrast is seen in the strongly weathered rock layer shown in Figure 8. Since the strain signal of hard rock layers is usually weak but contains important micro-fracture information, the conventional wavelet denoising method, using a uniform threshold, causes these weak but crucial strain features to be over-smoothed and lost. The method of this invention, based on the high cohesion characteristics of the rock layer, automatically adopts a smaller denoising threshold, successfully preserving these micro-strain features.

[0277] In summary, the signal-to-noise ratio improvement data shown in Figures 6 to 8 further quantify the advantages of the method of this invention. It achieves a denoising effect superior to conventional methods under all soil layer conditions. This experiment demonstrates the technological innovation value of denoising based on geological attribute differentiation. By establishing a negative correlation mechanism between the denoising threshold and soil hardness, an intelligent balance is achieved between the protection of weak features in hard soil layers and the suppression of strong noise in soft soil layers. This can provide higher quality strain data for subsequent feature extraction and intensity prediction.

[0278] Example 4

[0279] As shown in Figures 9 and 10, the overall performance of the method of the present invention is comprehensively evaluated by multi-dimensional indicators. The conventional method used in the comparison method refers to the technical route of feature compression based on principal component analysis combined with conventional neural networks for prediction. It only performs linear dimensionality reduction without introducing geological and physical constraints and lacks targeted preservation of key geological features.

[0280] As shown in Figure 9, the prediction errors under different concrete strength grades are compared. The horizontal axis represents the strength grade from C10 to C50, and the vertical axis represents the prediction error in megapascals. The method of the present invention shows lower prediction errors at all strength grades, and its advantages are particularly prominent in the prediction of high-strength concrete.

[0281] As shown in Figure 10, a geological risk identification performance test was conducted. The horizontal axis represents different geological risk conditions, and the vertical axis represents the percentage of risk identification accuracy. The method of this invention maintains a high identification accuracy under various risk conditions, indicating the synergistic effect of the geological attention mechanism and the multi-task learning framework. According to the experimental results shown in Figure 10, it can also be seen that conventional methods treat all strain data equally during feature compression and cannot prioritize the retention of key information in soil abrupt change zones, which leads to a decline in prediction performance under complex geological conditions.

[0282] In summary, the method of the present invention outperforms existing conventional technologies in both the accuracy of predicting concrete of different strength grades and the accuracy of geological risk identification, thus the overall performance of the present invention is better.

[0283] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for predicting the compressive strength of anti-slide piles, characterized in that, Includes the following steps: S1. Axial strain data of anti-slide piles under load is collected using sensors. The collected data is then labeled with compressive strength grade and geological risk. Simultaneously, soil layer type information for each anti-slide pile is obtained, and soil friction angle and cohesion parameters are measured to form a discrete geological parameter sequence. S2. Geophysical equations are constructed using the geological parameter sequence. The collected data is then modulated using the geophysical equations to generate modulated strain data. S3. Wavelet packet denoising is performed on the modulated strain data, and the wavelet packet denoising threshold is adaptively adjusted according to the soil layer type to obtain denoised strain data. S4. The denoised strain data is compressed using the constructed autoencoder network structure to generate latent features. Then, a total optimization objective function is constructed, and the autoencoder network parameters are trained by minimizing this function. The autoencoder network parameters are iteratively trained and updated to obtain the optimized autoencoder network. Specifically, S4 is as follows: First, the constructed autoencoder network structure is an encoder-decoder. The encoder compresses the denoised strain data and incorporates geological modulation vectors and geological gradient vectors from the geophysical equations into the feature compression process, outputting latent features. Simultaneously, a geological regularization loss term is constructed based on the deviation between the soil friction angle parameter and the global mean, forcing the encoder to reduce weights in areas of abnormal friction angles. Then, the total optimization objective function of the autoencoder is constructed by combining reconstruction error, sparsity constraints, and the geological regularization term. The autoencoder network parameters are trained by minimizing this function. The reconstruction error is the absolute value of the difference between the latent features output by the encoder and the reconstructed latent features output by the decoder; the sparsity constraint refers to the L1 norm of the latent features. Finally, the autoencoder network parameters are optimized. The iterative training and parameter update process involves initializing the autoencoder network weight matrix and bias vector, inputting the denoised strain data vector into the encoder, enhancing it with geological modulation and geological gradient vectors to generate a 128-dimensional latent feature vector, and then reconstructing the output vector by the decoder. In each iteration, the overall optimization objective function is calculated, and the Adam optimizer is used for gradient descent to update the weight matrix and bias parameters. An initial learning rate is set, decaying by 10% every 100 iterations, and an iteration stopping condition is set until the condition is met, resulting in an optimized autoencoder network. S5: A landslide pile compressive strength prediction model is constructed, sequentially building a geological attention feature reweighting module, a physics-driven residual gating module, and a multi-task interference output layer. Latent features are input, and the predicted probabilities of compressive strength level and geological risk are output. The multi-task joint loss function is then calculated, and the landslide pile compressive strength prediction model is iteratively trained and its parameters updated until a well-trained model is obtained. S6: Newly acquired data, after preprocessing in steps S2 to S4, is input into the trained model to generate the final prediction result.

2. The method for predicting the compressive strength of anti-slide piles according to claim 1, characterized in that, S2 is as follows: The geophysical equations include the calculation of the geological modulation vector and the geological gradient vector; specifically, based on the soil friction angle and cohesion parameters, the geological modulation vector is calculated through the normalized deviation of the friction angle and the hyperbolic tangent function of the cohesion; the geological gradient vector is calculated through the absolute rate of change of cohesion at adjacent pile depths; then, the collected data is modulated through the geophysical equations, specifically by performing the Hadamard product operation between the collected strain data and the geological modulation vector, and combining it with the scaled geological gradient vector to obtain the modulated strain data.

3. The method for predicting the compressive strength of anti-slide piles according to claim 1, characterized in that, S3 is as follows: First, based on the acquired soil layer type information, the soil layer type is pre-classified, and the modulated strain data is divided into sub-vectors corresponding to different soil layers. At the same time, the wavelet packet denoising threshold is dynamically calculated based on the cohesion parameters of each type of soil layer. Then, Symlet wavelet packet decomposition is performed on the strain sub-vector of each soil layer. Specifically, a continuously differentiable threshold function is used to process the wavelet coefficients to obtain the representation of the threshold processing function. Finally, the wavelet coefficients processed by the threshold processing function are reconstructed using wavelet packets to recover the denoised strain data of each soil layer. Then, the sub-vectors are integrated into complete denoised strain data through the soil layer indicator function to obtain the denoised strain data.

4. The method for predicting the compressive strength of anti-slide piles according to claim 1, characterized in that, The specific operation of the geological attention feature reweighting module is as follows: The geological attention feature reweighting module adopts the geological perception attention mechanism. First, a geological query vector is constructed. The global mean of the geological modulation vector and the global maximum value of the geological gradient vector are transformed by linear transformation and ReLU activation function to generate the geological query vector. Then, the key vector and attention weights are calculated. The compressed latent features are transformed by linear transformation to obtain the key vector, and then scaled dot product operation is performed with the geological query vector. Finally, the attention weight distribution is obtained by Softmax normalization. The key vector is calculated from the trainable key weight matrix and latent features; Finally, the compressed features are reweighted using attention weights, normalized geological gradient vectors, and normalized geological modulation vectors to obtain geological attention-enhanced features.

5. The method for predicting the compressive strength of anti-slide piles according to claim 4, characterized in that, The operation of the physics-driven residual gating module is as follows: The physics-driven residual gating module adopts a physical gating mechanism based on the Mohr-Coulomb criterion, embedding the yield condition of geotechnical mechanics into the residual connection; specifically, it calculates the Mohr-Coulomb yield function value based on the estimated stress state, friction angle, and cohesion, and uses the Mohr-Coulomb yield function value to represent the yield degree of the geotechnical mass under the current stress state; then, it maps the yield function value, the global mean of the geological attention enhancement feature vector, and the global maximum value of the geological gradient vector through linear transformation and the Sigmoid activation function to obtain the physical gating vector; finally, it introduces the physical gating vector on the basis of the standard residual connection, performs element-wise weighting on the residual function output, and then updates the features output by the residual neural network layer by layer.

6. The method for predicting the compressive strength of anti-slide piles according to claim 5, characterized in that, The operation of the multi-task interference output layer is as follows: The multi-task interference output layer adopts a multi-task joint learning framework to simultaneously predict the compressive strength level and geological risk. It strengthens risk supervision in the geological mutation zone through an adaptive weight adjustment mechanism, as follows: (1) The final layer features of the residual neural network in the physical-driven residual gating module are concatenated with the global mean of the geological attention enhancement features and the geological gradient vector. The geological risk probability is predicted through a fully connected layer and the Sigmoid function; (2) Based on the final layer features of the residual neural network in the physical-driven residual gating module and the predicted probability of geological risk, the compressive strength level probability distribution is predicted through a fully connected layer and the Softmax function; (3) Based on the magnitude of the geological gradient vector, the adaptive weight of the geological regularization loss term is calculated through the Sigmoid function.

7. The method for predicting the compressive strength of anti-slide piles according to claim 6, characterized in that, The calculation process of the multi-task joint loss function is as follows: Combining the cross-entropy loss of intensity level prediction and the geological regularization loss term, the multi-task joint loss function is constructed through adaptive weights; the cross-entropy loss of intensity level prediction is calculated using standard cross-entropy based on the predicted probability distribution of compressive strength level and the actual compressive strength level label.

8. The method for predicting the compressive strength of anti-slide piles according to claim 1, characterized in that, S6 is as follows: After preprocessing the newly collected anti-slide pile monitoring data, it is input into the trained anti-slide pile compressive strength prediction model, outputs the compressive strength level probability distribution, takes the level corresponding to the highest probability as the final strength prediction value, and outputs the geological risk probability. When the probability value exceeds the preset risk threshold, a sliding risk warning is triggered.

9. The method for predicting the compressive strength of anti-slide piles according to claim 1, characterized in that, S1 is as follows: Strain sensor monitoring points are arranged at equal intervals along the depth of the anti-slide piles, and data is collected in real time through distributed fiber optic sensing technology; borehole exploration is carried out along the corresponding position of the anti-slide pile depth to obtain soil layer type information at each pile depth, and soil friction angle and cohesion parameters at each depth point are determined through indoor geotechnical tests; the collected data are labeled, and the compressive strength grade label is marked according to the ultimate compressive strength of the pile body measured by the on-site static load test; the geological risk label is marked by the geotechnical engineer based on the comprehensive judgment results of the borehole core fracture development degree, cohesion abrupt change gradient and historical landslide records.

Images