A geological data analysis road slope deformation early warning monitoring system

Abstract

This invention discloses a road slope deformation early warning and monitoring system based on geological data analysis, comprising a parameter acquisition module, a parameter association mapping module, a feature extraction and enhancement module, a model training and inference module, an early warning threshold determination module, and an early warning response module. The parameter acquisition module acquires data such as slope displacement and transmits it to the parameter association mapping module. After parsing the parameter coupling relationship, the module transmits the feature tensor to the feature extraction and enhancement module. This module decomposes and fuses the features and transmits them to the model training and inference module. The model training and inference module calculates the anomaly degree using an optimized multi-head attention mechanism and a road slope local outlier factor model, and transmits this calculation to the early warning threshold determination module to match the early warning level. The early warning response module then triggers the corresponding response. The feature extraction and enhancement module and the model training and inference module each contain multiple units that function independently. This system comprehensively captures slope deformation characteristics, improving the accuracy and reliability of early warnings.

Description

Technical Field

[0001] This invention relates to the field of road slope deformation early warning, and in particular to a road slope deformation early warning and monitoring system based on geological data analysis. Background Technology

[0002] During the construction and operation of road engineering projects, road slopes are prone to deformation and even instability due to the combined effects of multiple factors such as geological structure, soil and rock properties, hydrological conditions, and external loads. This poses a serious threat to road traffic safety, surrounding infrastructure, and people's lives and property. With the continuous expansion of transportation networks, the mileage of roads with high and steep slopes and complex geological conditions continues to increase. The concealment, suddenness, and destructiveness of slope deformation are becoming increasingly significant. There is an urgent need to build a precise and efficient deformation early warning and monitoring system. Through real-time capture, in-depth analysis, and anomaly identification of slope deformation-related parameters, this system can provide technical support for slope safety management.

[0003] Existing technologies for monitoring road slope deformation have two significant shortcomings. Firstly, the processing of slope deformation parameters often employs a single-dimensional analysis approach, failing to fully exploit the nonlinear coupling relationships between multiple parameters such as slope displacement, soil moisture content, and pore water pressure. This results in the loss of parameter correlation information and an inability to reflect the overall dynamic characteristics of slope deformation. Secondly, in the abnormal deformation identification stage, traditional models are poorly adaptable to slope deformation data with uneven spatiotemporal distribution, unable to effectively distinguish between normal fluctuations and abnormal deformation. Furthermore, the models lack the ability to capture subtle deformation features in localized areas of the slope, easily leading to delayed or misjudged anomalies. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a road slope deformation early warning and monitoring system based on geological data analysis.

[0005] The technical solution adopted in this invention is a road slope deformation early warning and monitoring system based on geological data analysis, comprising:

[0006] Parameter acquisition module: Through the deployment of multi-type sensor arrays, the slope displacement, soil moisture content, pore water pressure, slope top settlement, deep horizontal displacement and slope surface temperature field distribution data of the road slope are acquired in real time. The collected parameters are transmitted to the parameter association mapping module in the form of time-series data stream.

[0007] Parameter association mapping module: Receives the time-series data stream output by the parameter acquisition module, and analyzes the nonlinear coupling relationship between slope displacement and soil moisture content, pore water pressure and deep horizontal displacement, and slope top settlement and temperature field distribution through the constructed dynamic parameter association matrix. The analysis results are sent to the feature extraction and enhancement module in the form of feature tensors.

[0008] Feature extraction enhancement module: Receives the feature tensor output by the parameter association mapping module, uses a multi-scale feature decomposition algorithm to split the feature tensor into dimensions, obtains slope deformation feature sub-tensors at different scales, and transmits each sub-tensor to the model training and inference module through the feature fusion interface.

[0009] Model training and inference module: Receives the feature sub-tensor output by the feature extraction and enhancement module, calls the built-in optimized multi-head attention mechanism to perform weight allocation and association learning on the feature sub-tensor, generates a high-dimensional feature vector, and then inputs the high-dimensional feature vector into the local outlier factor model of the road slope for anomaly calculation. The calculation result is sent to the early warning threshold determination module.

[0010] Warning threshold determination module: Receives the anomaly calculation results output by the model training and inference module, combines them with the preset multi-level warning threshold range, performs interval matching on the anomaly values, and transmits the matching results to the warning response module in the form of a signal;

[0011] Early warning response module: Receives the matching signal output by the early warning threshold determination module, triggers the corresponding audible and visual alarm device and data storage instructions according to the early warning level corresponding to the signal, and transmits the early warning level information to the remote monitoring terminal through the communication link.

[0012] Furthermore, the parameter acquisition module also includes an adaptive sampling unit, which dynamically samples and adjusts the acquired parameters based on an optimized multi-head attention mechanism. The adjustment model formula is as follows:

[0013] S t+1 =S t ·exp(α·MHAttn(P t P t-1 P t-2 )-β·LOF s (P t ))

[0014] Among them, S t+1 S is the sampling frequency at time t+1. t Let be the sampling frequency at time t, α be the attention weight adjustment coefficient, β be the local outlier factor influence coefficient, and MHAttn(P) be the sampling frequency at time t. t P t-1 P t-2 ) is the parameter P collected based on times t, t-1, and t-2. t Slope displacement, P t-1 Water content of soil and rock, P t-2 Optimized multi-head attention value for pore water pressure, LOF s (P t () is based on the slope displacement P tThe road slope local outlier sampling correction value; the association strength calculation unit in the parameter association mapping module adopts an optimized multi-head attention mechanism combined with the road slope local outlier model to construct the parameter association strength calculation model:

[0015]

[0016] Among them, R ij P represents the association strength between the i-th type of parameter and the j-th type of parameter, where n is the number of parameter samples collected. ik For the k-th sample value of the i-th type parameter, P jk For the k-th sample value of the j-th class parameter, MHAttn(P) ik P jk ) represents the multi-head attention association value between the i-th class and the k-th sample of the j-th class, LOF. r (P ik P jk ) is the local outlier association factor between the i-th class and the j-th class of parameters for the k-th sample.

[0017] Furthermore, the feature extraction enhancement module includes a feature dimension expansion unit, which utilizes an optimized multi-head attention mechanism to increase the dimension of the feature sub-tensor. The expansion model formula is as follows:

[0018] F e =Concat(MHAttn(F s1 F s2 F s3 ),MHAttn(F s2 F s3 F s4 ))·W e +b e

[0019] Among them, F e For the expanded high-dimensional feature tensor, F s1 F s2 F s3 F s4 For feature subtensors of different scales, Concat is the feature concatenation function, and W... e To expand the weight matrix by dimension, b e The bias vector is used for dimensional expansion, and MHAtn is an optimized multi-head attention computation function. Simultaneously, the feature enhancement unit of this module employs a roadside slope local outlier factor model to enhance features; the model formula is as follows:

[0020] F en =F e ·(1+γ·LOF f (F e ))

[0021] Among them, F en The enhanced feature tensor, γ is the enhancement coefficient, and LOF is the feature tensor. f (F e ) is based on the feature tensor F e The local outlier feature enhancement value.

[0022] Furthermore, the attention weight update unit in the model training and inference module dynamically adjusts the weights through an optimized multi-head attention mechanism, and the model formula is as follows:

[0023]

[0024] Among them, W t+1 Let W be the attention weight matrix at time t+1. t Let be the attention weight matrix at time t, and λ be the learning rate. Regarding W t The gradient operator, where M is the number of attention heads, Q... m K m V m Let d be the query matrix, key matrix, and value matrix of the m-th attention head. k The key vector dimension is 1, Softmax() is the softmax activation function, and LOF is 1. w (W t () is based on the weight matrix W t The local outlier factor weight correction value; the anomaly inference unit of this module uses the local outlier factor model of road slope for calculation, and the formula is:

[0025]

[0026] Where OD is the overall outlier value, and N is the total number of samples. k (i) is the k-nearest neighbor set of the i-th sample, dist(X) i X j ) is the sample X i With X j The distance between them, X i X j For the input feature samples, MHAtn(X) i X j ) is the sample X i With X j Multi-head attention correlation value.

[0027] Furthermore, the threshold dynamic adjustment unit of the warning threshold determination module constructs a threshold adjustment model by combining an optimized multi-head attention mechanism, with the following formula:

[0028] Tnew =T old ·Sigmoid(MHAttn(H,T) old ,ΔD)+θ·LOF t (T old ))

[0029] Among them, T new For the new warning threshold, T old The original warning threshold is defined by sigmoid(), the sigmoid activation function is defined by H, the historical warning data matrix is ​​defined by ΔD, the recent slope displacement change is defined by θ, and the threshold adjustment coefficient is defined by LOF. t (T old () is based on the original threshold T old The local outlier threshold correction value; the interval matching unit of this module is matched using the following formula:

[0030]

[0031] Where Match is the matching result vector, L is the number of warning levels, and T is the number of warning levels. l Here, `round` is the level I warning threshold, `oneHot()` is the rounding function, `OD` is the one-hot encoding function, and `MHAttn(OD, T)` represents the anomaly score. l ) represents the multi-head attention matching degree between the anomaly score and the level I threshold, LOF m (OD) is the local outlier factor matching correction value based on the outlier degree OD.

[0032] Furthermore, the alarm signal generation unit of the early warning response module generates signals using an optimized multi-head attention mechanism, with the following formula:

[0033] S=Concat(MHAttn(A1,OD,T),MHAttn(A2,OD,T),...,MHAttn(A L ,OD,T))·W s +b s

[0034] Where S is the alarm signal vector, A l Let OD be the level I alarm mode vector, T be the anomaly value, and W be the warning threshold matrix. s Generate a weight matrix for the signal, b s A bias vector is generated for the signal; Concat() is the vector concatenation function. The data storage control unit of this module adjusts the storage strategy through the local outlier factor model of the road slope, with the formula: Store = MHAttn(D, OD, T) · LOF s(D) + μ·sign(OD-T), where Store is the storage control instruction, D is the data matrix to be stored, μ is the control coefficient, sign() is the sign function, and LOF... s (D) is the local outlier factor storage correction value based on data D, and MHTtn(D, OD, T) is the multi-head attention association value of data, outlier degree, and threshold.

[0035] Furthermore, the feature extraction enhancement module includes the following units:

[0036] Multi-scale feature decomposition unit: Receives the feature tensor output by the parameter association mapping module, and decomposes the feature tensor into layers according to preset different scale division rules to obtain multiple feature sub-tensors with different dimensions and resolutions. Each feature sub-tensor corresponds to the preset spatial or temporal scale features of road slope deformation. The feature sub-tensors are transmitted to the feature fusion interface through the internal data bus.

[0037] Feature association mining unit: Receives the feature sub-tensors output by the multi-scale feature decomposition unit, deeply mines the potential correlations between slope displacement and soil moisture content parameters contained in different sub-tensors, and transforms the mined correlation information into a correlation coefficient matrix by constructing a feature association map, which is then transmitted to the feature enhancement unit.

[0038] Feature enhancement unit: Receives the correlation coefficient matrix output by the feature association mining unit and the feature sub-tensor output by the multi-scale feature decomposition unit. Based on the correlation coefficient matrix, it dynamically adjusts the weight of each feature sub-tensor, strengthens the feature components that are strongly correlated with slope deformation, and weakens irrelevant or weakly correlated components. The processed feature tensor is sent to the model training and inference module.

[0039] Feature Dimension Adaptation Unit: Receives the enhanced feature tensor output by the feature enhancement unit, performs dimension transformation and adaptation processing on the feature tensor according to the requirements of the model training and inference module for the input feature dimension, ensures that the dimension of the output feature tensor is completely matched with the dimension of the model input, and transmits the adapted feature tensor to the model training and inference module through the interface protocol.

[0040] Furthermore, the model training and inference module includes the following units:

[0041] Attention Mechanism Configuration Unit: Receives the feature tensor output by the feature extraction enhancement module, and based on the feature attributes of the slope deformation parameters, initializes the number of attention heads, the dimension configuration of each head, and the attention calculation method of the optimized multi-head attention mechanism, generates the basic configuration parameters of the attention mechanism, and passes them to the attention weight update unit;

[0042] Attention Weight Update Unit: Receives the basic configuration parameters output by the attention mechanism configuration unit and the feature tensor output by the feature extraction enhancement module. Based on the variation law of slope displacement and pore water pressure parameters in the input feature tensor, it iteratively updates the weight matrix of each attention head through the gradient descent algorithm, so that the attention weight can better capture the important correlation between parameters. The updated weight matrix is ​​sent to the anomaly inference unit.

[0043] Local outlier factor model construction unit: Based on historical monitoring data of road slope deformation, the basic structure of the local outlier factor model of road slope is constructed, the parameters of the number of nearest neighbors and the distance calculation method in the model are determined, the initial local outlier factor model is formed, and it is passed to the anomaly inference unit.

[0044] Anomaly inference unit: Receives the attention weight matrix output by the attention weight update unit, the initial model output by the local outlier factor model construction unit, and the feature tensor output by the feature extraction enhancement module. It uses the attention weight matrix to weight the feature tensor, inputs the processing result into the local outlier factor model for calculation, obtains the anomaly value of road slope deformation, and transmits it to the early warning threshold determination module.

[0045] Furthermore, the early warning threshold determination module includes the following units:

[0046] Threshold interval division unit: Based on the engineering geological conditions, historical deformation data and safety level requirements of the road slope, the slope deformation anomaly is divided into multiple continuous and non-overlapping threshold intervals. Each interval corresponds to a preset warning level. The division results are stored in the form of interval boundary values ​​and transmitted to the anomaly matching unit.

[0047] Anomaly matching unit: Receives the anomaly value output by the model training and inference module and the interval boundary value output by the threshold interval division unit, compares the anomaly value with each threshold interval one by one, determines the specific interval to which the anomaly value belongs, and generates an interval matching signal to send to the threshold dynamic adjustment unit.

[0048] Threshold dynamic adjustment unit: Receives the interval matching signal and historical early warning data output by the anomaly matching unit, analyzes the distribution pattern of anomaly values ​​and slope deformation trends in long-term monitoring, dynamically corrects the boundary values ​​of each threshold interval, so that the threshold interval can adapt to the long-term changing characteristics of slope deformation, and feeds back the adjusted interval boundary values ​​to the threshold interval division unit.

[0049] Warning signal generation unit: Receives the interval matching signal output by the anomaly degree matching unit, generates the corresponding warning signal code according to the matched warning level, the code contains the warning level and anomaly degree value range information, and sends the warning signal code to the warning response module through the signal transmission channel.

[0050] Beneficial Effects: This invention proposes a road slope deformation early warning and monitoring system based on geological data analysis. The system acquires various types of slope deformation parameters through a parameter acquisition module, and analyzes the nonlinear coupling relationships of multiple parameters through a parameter association mapping module, overcoming the problem of lost parameter association information caused by single-dimensional analysis. It strengthens key features through a feature extraction enhancement module, and combines optimized multi-head attention mechanisms in the model training and inference module for deep learning of multi-parameter dynamic associations. This, along with a local outlier model of road slopes, accurately identifies abnormal deformations, solving the problems of weak adaptability to spatiotemporally uneven data and insufficient capture of subtle local features in traditional models. The early warning threshold determination module dynamically adjusts the threshold range, and the early warning response module provides tiered responses. The synergistic effect of these modules enhances the depth of parameter association analysis and the accuracy of anomaly identification. Multi-unit processing by the feature extraction enhancement module strengthens feature expression, and unit optimization of attention weights and inference logic in the model training and inference module enables the system to effectively mine the overall dynamic characteristics of parameters, distinguish between normal fluctuations and abnormal deformations, reduce early warning lag and misjudgments, and achieve efficient early warning and monitoring of road slope deformation. Attached Figure Description

[0051] Figure 1 This is a diagram showing the system module composition of the present invention;

[0052] Figure 2 This is a flowchart illustrating the system operation of the present invention. Detailed Implementation

[0053] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown, a road slope deformation early warning and monitoring system based on geological data analysis includes the following modules:

[0055] Parameter acquisition module: Through the deployment of multi-type sensor arrays, the slope displacement, soil moisture content, pore water pressure, slope top settlement, deep horizontal displacement and slope surface temperature field distribution data of the road slope are acquired in real time. The collected parameters are transmitted to the parameter association mapping module in the form of time-series data stream.

[0056] Specifically, the parameter acquisition module is the foundation for the system to obtain raw data. Its technical parameters cover the sensor's measurement range, accuracy, and sampling frequency. The measurement range needs to cover the potential slope displacement (0-500mm), soil moisture content (0-50%), pore water pressure (0-200kPa), slope crest settlement (0-300mm), deep horizontal displacement (0-1000mm), and slope surface temperature (-20-60℃). The accuracy is controlled within ±0.1mm, ±1%, ±1kPa, ±0.05mm, ±0.5mm, and ±0.5℃, respectively. The sampling frequency can be adjusted within the range of 1-60 minutes / time according to actual conditions. The significance of this module is to provide comprehensive, accurate, and real-time raw data support for subsequent data analysis and early warning, ensuring that the system can analyze and judge based on the actual slope condition.

[0057] In the specific implementation process, multiple types of sensor arrays are deployed at different locations on the roadside slope. Slope displacement is acquired using guy-wire displacement sensors and GPS displacement monitoring stations. Guy-wire displacement sensors are deployed along different elevations on the slope, while GPS displacement monitoring stations are deployed at the slope top and toe in stable areas. Soil and rock moisture content is measured using a time-domain reflectometer (TDAR) with sensors embedded in boreholes at different depths on the slope. Pore water pressure is measured using pore water pressure sensors embedded near the potential sliding surface of the slope. Slope top settlement is monitored using a precision level in conjunction with settlement observation points. Deep horizontal displacement is measured at different depths in boreholes using an inclinometer. The slope surface temperature field distribution is collected by scanning with an infrared thermometer array. All data collected by the sensors is transmitted in real-time to the parameter association and mapping module via wired or wireless transmission.

[0058] Parameter association mapping module: Receives the time-series data stream output by the parameter acquisition module, and analyzes the nonlinear coupling relationship between slope displacement and soil moisture content, pore water pressure and deep horizontal displacement, and slope top settlement and temperature field distribution through the constructed dynamic parameter association matrix. The analysis results are sent to the feature extraction and enhancement module in the form of feature tensors.

[0059] Specifically, the technical parameters of the parameter association mapping module mainly include data processing capability and association resolution accuracy. The data processing capability must be able to simultaneously process time-series data streams from at least 100 sensors, with a processing delay of no more than 1 second for each stream. The association resolution accuracy must ensure that the resolution error for the nonlinear coupling relationship between slope displacement and other parameters does not exceed 5%. The significance of this module lies in associating and integrating scattered and independent sensor data, uncovering the intrinsic connections between parameters, and providing a data foundation with correlational significance for subsequent feature extraction and model training, thus avoiding the one-sidedness of analysis caused by isolated data.

[0060] In the specific implementation process, after receiving time-series data streams such as slope displacement, soil moisture content, and pore water pressure from the parameter acquisition module, the data is first standardized to convert data from different sensors into a unified data format and timestamp. Then, a dynamic parameter correlation matrix is ​​constructed, where rows and columns correspond to different parameter types, and matrix elements represent the correlation strength between corresponding parameters. Initial values ​​for matrix elements are determined through learning and analysis of historical data, and the matrix elements are dynamically updated based on real-time data. The nonlinear coupling relationships between multiple sets of parameters, such as slope displacement and soil moisture content, pore water pressure and deep horizontal displacement, and slope top settlement and temperature field distribution, are analyzed. The analyzed correlation features are converted into feature tensors and transmitted to the feature extraction and enhancement module through a preset interface protocol.

[0061] Feature extraction enhancement module: Receives the feature tensor output by the parameter association mapping module, uses a multi-scale feature decomposition algorithm to split the feature tensor into dimensions, obtains slope deformation feature sub-tensors at different scales, and transmits each sub-tensor to the model training and inference module through the feature fusion interface.

[0062] Specifically, the technical parameters of the feature extraction enhancement module involve the scale range of feature decomposition and the intensity coefficient of feature enhancement. The scale range of feature decomposition can be adjusted within a scale level of 1-100 to adapt to the feature extraction needs of different resolutions; the intensity coefficient of feature enhancement is dynamically adjusted between 0.1 and 2.0 according to the correlation between the parameters and slope deformation. The significance of this module lies in extracting feature information that is of great significance for slope deformation early warning from the feature tensor output by the parameter correlation mapping module, and highlighting key features and suppressing irrelevant features through enhancement processing, thereby improving the efficiency and accuracy of subsequent model training and inference.

[0063] In the specific implementation process, after receiving the feature tensor output by the parameter association mapping module, the multi-scale feature decomposition unit decomposes the feature tensor into multiple feature sub-tensors of different scales according to preset scale division rules. Each feature sub-tensor corresponds to a specific spatial or temporal scale feature of road slope deformation, such as hourly or day-level temporal scale features, or spatial scale features at the top, middle, and bottom of the slope. The feature association mining unit analyzes these feature sub-tensors to mine the correlation between slope displacement and soil moisture content at different scales, generating a correlation coefficient matrix. The feature enhancement unit dynamically adjusts the weights of each feature sub-tensor based on the correlation coefficient matrix, enhancing features strongly correlated with slope deformation, such as slope displacement and pore water pressure, while weakening features in the temperature field distribution that are weakly correlated with deformation. The feature dimension adaptation unit converts the enhanced feature tensor into a form that matches the input dimension of the model training and inference module and transmits it to the model training and inference module.

[0064] Model training and inference module: Receives the feature sub-tensor output by the feature extraction and enhancement module, calls the built-in optimized multi-head attention mechanism to perform weight allocation and association learning on the feature sub-tensor, generates a high-dimensional feature vector, and then inputs the high-dimensional feature vector into the local outlier factor model of the road slope for anomaly calculation. The calculation result is sent to the early warning threshold determination module.

[0065] Specifically, the technical parameters of the model training and inference module include the number of attention heads and the number of nearest neighbors for calculating local outliers. The number of attention heads can be set from 4 to 16, and the dimension of each attention head can be adjusted from 32 to 128 dimensions; the number of nearest neighbors for calculating local outliers can be selected from 5 to 50 depending on the number of samples. The significance of this module lies in using an optimized multi-head attention mechanism and a road slope local outlier model to perform in-depth analysis and inference on the extracted features, achieving accurate calculation of slope deformation anomalies and providing a reliable basis for determining early warning thresholds.

[0066] In the specific implementation process, the attention mechanism configuration unit sets the number of attention heads to 8, with each head having a dimension of 64, based on the characteristics of the slope deformation parameters, and determines the attention calculation method. After receiving the feature tensor output by the feature extraction enhancement module, the attention weight update unit updates the weight matrix of each attention head iteratively based on the changing patterns of parameters such as slope displacement and pore water pressure in the feature tensor, making the attention mechanism focus more on parameters that significantly affect deformation. The local outlier model construction unit determines the number of nearest neighbors to 10 based on historical monitoring data and constructs an initial local outlier model. The anomaly inference unit inputs the high-dimensional feature vector obtained after the attention mechanism processing into the local outlier model to calculate the anomaly value of the road slope deformation. In this calculation, specific parameters such as the rate of change of slope displacement and the gradient change of soil and rock water content are combined, and the calculated anomaly value is transmitted to the early warning threshold determination module.

[0067] Warning threshold determination module: Receives the anomaly calculation results output by the model training and inference module, combines them with the preset multi-level warning threshold range, performs interval matching on the anomaly values, and transmits the matching results to the warning response module in the form of a signal;

[0068] Specifically, the technical parameters of the early warning threshold determination module include the number of threshold intervals and the step size for threshold adjustment. The number of threshold intervals can be set to 3-5, corresponding to different early warning levels, such as low, medium, high, and emergency. The step size for threshold adjustment is set between 0.01 and 0.1 based on the fluctuation of historical data. The significance of this module is to determine the corresponding early warning level based on the anomaly value output by the model training and inference module, combined with the preset threshold intervals, providing a clear trigger signal for early warning response and ensuring the timeliness and accuracy of early warnings.

[0069] In the specific implementation process, the threshold interval division unit divides the anomaly values ​​into 5 threshold intervals based on the engineering geological conditions and historical deformation data of the road slope, corresponding to 5 warning levels. The interval boundary values ​​are determined according to the safe allowable deformation range of the slope, such as a low warning level corresponding to anomalies of 0-0.2, a medium warning level corresponding to 0.2-0.4, etc. The anomaly matching unit receives the anomaly values ​​output by the model training and inference module, compares them with each threshold interval, and determines the interval to which the value belongs. For example, when the anomaly value is 0.3, it is determined to belong to the medium warning level interval. The threshold dynamic adjustment unit dynamically corrects the boundary values ​​of the threshold intervals based on historical warning data and recent slope deformation trends, such as a continuous increase in slope displacement or a sudden change in pore water pressure. If the recent overall anomaly values ​​are found to be high, the boundary values ​​of each interval are appropriately increased. The warning signal generation unit generates a corresponding warning signal code based on the matched warning level, containing information such as the warning level and the anomaly value range, and sends it to the warning response module.

[0070] Early warning response module: Receives the matching signal output by the early warning threshold determination module, triggers the corresponding audible and visual alarm device and data storage instructions according to the early warning level corresponding to the signal, and transmits the early warning level information to the remote monitoring terminal through the communication link.

[0071] Specifically, the technical parameters of the early warning response module include the response time of the alarm device, the capacity and rate of data storage, etc. The response time of the alarm device must be controlled within 1-3 seconds to ensure a rapid alarm response upon receiving the warning signal; the data storage capacity should be no less than 10TB, and the storage rate no less than 100MB / s to meet the storage needs of large amounts of monitoring data and early warning information. The significance of this module is that upon receiving the early warning signal output by the early warning threshold determination module, it triggers the corresponding alarm device and data storage instructions to promptly remind relevant personnel to take measures and save relevant data for subsequent analysis, ensuring the safety of roadside slopes.

[0072] In the specific implementation process, after receiving the warning signal code output by the warning threshold determination module, the code is first parsed to determine the corresponding warning level. Depending on the warning level, corresponding audible and visual alarm devices are triggered; for example, a low warning level triggers a yellow warning light and a low-frequency alarm sound, while a high warning level triggers a red warning light and a high-frequency alarm sound. The alarm devices are deployed in both the on-site duty room at the slope and the remote monitoring center. Simultaneously, a data storage program is initiated to store current monitoring data such as slope displacement, soil moisture content, and pore water pressure, as well as information such as the warning level and warning time, in a local database and on a remote server. Data encryption technology is used during storage to ensure data security. The warning level information is transmitted in real-time to the remote monitoring terminal via a communication link using the TCP / IP protocol to ensure the stability and reliability of information transmission, enabling remote personnel to promptly understand the slope deformation warning situation.

[0073] Preferably, the parameter acquisition module further includes an adaptive sampling unit, which dynamically samples and adjusts the acquired parameters based on an optimized multi-head attention mechanism. The adjustment model formula is as follows:

[0074] S t+1 =S t ·exp(α·MHAttn(P t P t-1 P t-2 )-β·LOF s (P t ))

[0075] Among them, S t+1 S is the sampling frequency at time t+1. t Let be the sampling frequency at time t, α be the attention weight adjustment coefficient, β be the local outlier factor influence coefficient, and MHAttn(P) be the sampling frequency at time t. t P t-1 P t-2 ) is the parameter P collected based on times t, t-1, and t-2. t (slope displacement), P t-1 (Water content of soil and rock mass), P t-2 Optimized multi-head attention value for (pore water pressure), LOF s (P t () is based on the slope displacement P t The road slope local outlier sampling correction value; the association strength calculation unit in the parameter association mapping module adopts an optimized multi-head attention mechanism combined with the road slope local outlier model to construct the parameter association strength calculation model:

[0076]

[0077] Among them, Rij P represents the association strength between the i-th type of parameter and the j-th type of parameter, where n is the number of parameter samples collected. ik For the k-th sample value of the i-th type parameter, P jk For the k-th sample value of the j-th class parameter, MHAttn(P) ik P jk ) represents the multi-head attention association value between the i-th class and the k-th sample of the j-th class, LOF. r (P ik P jk ) is the local outlier association factor between the i-th class and the j-th class of parameters for the k-th sample.

[0078] Specifically, the technical parameters of the adaptive sampling unit include the sampling frequency adjustment range (1-60 minutes / time), the attention weight adjustment coefficient (0.1-1.0), and the local outlier factor influence coefficient (0.05-0.5). This allows the sampling frequency to be dynamically adjusted according to parameter changes, improving data acquisition efficiency and relevance. During implementation, real-time parameters are received, and the sampling frequency for the next moment is determined through model calculations, ensuring data density during critical deformation stages. The technical parameters of the correlation strength calculation unit include the sample size (no less than 1000 sets) and the local outlier correlation factor threshold (0-1). This is significant for accurately quantifying the correlation between parameters. During implementation, multi-head attention correlation values ​​and local outlier correlation factors are calculated on the received parameter samples, and the correlation strength value is obtained through the model, providing a basis for subsequent analysis.

[0079] Preferably, the feature extraction enhancement module includes a feature dimensionality expansion unit, which utilizes an optimized multi-head attention mechanism to increase the dimensionality of the feature sub-tensor. The expansion model formula is as follows:

[0080] F e =Concat(MHAttn(F s1 F s2 F s3 ),MHAttn(F s2 F s3 F s4 ))·W e +b e

[0081] Among them, F e For the expanded high-dimensional feature tensor, F s1 F s2 F s3 F s4 For feature subtensors of different scales, Concat is the feature concatenation function, and W... e To expand the weight matrix by dimension, b eThe bias vector is used for dimensional expansion, and MHAtn is an optimized multi-head attention computation function. Simultaneously, the feature enhancement unit of this module employs a roadside slope local outlier factor model to enhance features; the model formula is as follows:

[0082] F en =F e ·(1+γ·LOF f (F e ))

[0083] Among them, F en The enhanced feature tensor, γ is the enhancement coefficient, and LOF is the feature tensor. f (F e ) is based on the feature tensor F e The local outlier feature enhancement value.

[0084] Specifically, the technical parameters of the feature dimension expansion unit include the number of feature sub-tensors (4-8), the dimension of the dimension expansion weight matrix (matching the dimension of the feature sub-tensors), and the dimension of the bias vector (same as the output dimension of the weight matrix). Its purpose is to improve feature representation capabilities. During implementation, multi-head attention is calculated on the multi-scale feature sub-tensors and then concatenated. The high-dimensional feature tensor is obtained after processing with the weight matrix and bias vector. The technical parameters of the feature enhancement unit include the enhancement coefficient (0.2-1.5) and the range of local outlier factor feature enhancement values ​​(0-2). Its purpose is to strengthen key features. During implementation, the feature enhancement value is calculated using a local outlier model, and the high-dimensional feature tensor is adjusted in conjunction with the enhancement coefficient. The enhanced feature tensor is then output to the model training and inference module.

[0085] Preferably, the attention weight update unit in the model training and inference module dynamically adjusts the weights through an optimized multi-head attention mechanism, and the model formula is as follows:

[0086]

[0087] Among them, W t+1 Let W be the attention weight matrix at time t+1. t Let be the attention weight matrix at time t, and λ be the learning rate. Regarding W t The gradient operator, where M is the number of attention heads, Q... m K m V m Let d be the query matrix, key matrix, and value matrix of the m-th attention head. k The key vector dimension is 1, Softmax() is the softmax activation function, and LOF is 1. w (W t () is based on the weight matrix W tThe local outlier factor weight correction value; the anomaly inference unit of this module uses the local outlier factor model of road slope for calculation, and the formula is:

[0088]

[0089] Where OD is the overall outlier value, and N is the total number of samples. k (i) is the k-nearest neighbor set of the i-th sample, dist(X) i X j ) is the sample X i With X j The distance between them, X i X j For the input feature samples, MHAtn(X) i X j ) is the sample X i With X j Multi-head attention correlation value.

[0090] Specifically, the technical parameters of the attention weight update unit include learning rate (0.001-0.01), number of attention heads (4-16), key vector dimension (32-128), and local outlier factor weight correction value range (0.5-1.5). The purpose is to dynamically optimize attention weights and improve feature focus. During implementation, the gradient is calculated based on the feature tensor, and the weight matrix is ​​updated by combining the learning rate and local outlier factor weight correction value, making the attention heads focus more on the correlation of important parameters. The technical parameters of the anomaly inference unit include the total number of samples (no less than 500) and the number of nearest neighbors (5-50). The purpose is to accurately calculate the overall anomaly degree. During implementation, the distance between samples and the multi-head attention correlation value are calculated, and the anomaly degree value is obtained through the model by combining the nearest neighbor set, providing a basis for early warning.

[0091] Preferably, the threshold dynamic adjustment unit of the early warning threshold determination module constructs a threshold adjustment model in conjunction with an optimized multi-head attention mechanism, and the formula is:

[0092] T new =T old ·Sigmoid(MHAttn(H,T) old ,ΔD)+θ·LOF t (T old ))

[0093] Among them, T new For the new warning threshold, T old The original warning threshold is defined by sigmoid(), the sigmoid activation function is defined by H, the historical warning data matrix is ​​defined by ΔD, the recent slope displacement change is defined by θ, and the threshold adjustment coefficient is defined by LOF. t (T old () is based on the original threshold Told The local outlier threshold correction value; the interval matching unit of this module is matched using the following formula:

[0094]

[0095] Where Match is the matching result vector, L is the number of warning levels, and T is the number of warning levels. l The threshold for Level I warning is defined by `round()`, which is the rounding function, `OneHot()` is the one-hot encoding function, and `OD` is the anomaly value. `MHAttn(OD, T)` represents the threshold for Level I warning. l ) represents the multi-head attention matching degree between the anomaly score and the level I threshold, LOF m (OD) is the local outlier factor matching correction value based on the outlier degree OD.

[0096] Specifically, the technical parameters of the threshold dynamic adjustment unit include the sigmoid activation function parameter (default setting), threshold adjustment coefficient (0.1-0.5), and local outlier factor threshold correction value range (0-1). This is to adapt the warning threshold to slope changes. During implementation, historical data and displacement changes are received, and after multi-head attention calculation and local outlier factor correction, a new warning threshold is obtained through the sigmoid function. The technical parameters of the interval matching unit include the number of warning levels (3-5) and the local outlier factor matching correction value range (0-1). This is to improve the accuracy of warning level matching. During implementation, the multi-head attention matching degree is calculated between the abnormal threshold and each level of threshold. Combined with the local outlier factor correction value and one-hot encoding, a matching result vector is generated to ensure the accuracy of the warning level.

[0097] Preferably, the alarm signal generation unit of the early warning response module generates signals using an optimized multi-head attention mechanism, with the following formula:

[0098] S=Concat(MHAttn(A1,OD,T),MHAttn(A2,OD,T),...,MHAttn(A L ,OD,T))·W s +b s

[0099] Where S is the alarm signal vector, A l Let OD be the level I alarm mode vector, T be the anomaly value, and W be the warning threshold matrix. s Generate a weight matrix for the signal, b s A bias vector is generated for the signal; Concat() is the vector concatenation function. The data storage control unit of this module adjusts the storage strategy through the local outlier factor model of the road slope, with the formula: Store = MHAttn(D, OD, T) · LOF s(D) + μ·sign(OD-T), where Store is the storage control instruction, D is the data matrix to be stored, μ is the control coefficient, sign() is the sign function, and LOF... s (D) is the local outlier factor storage correction value based on data D, and MHTtn(D, OD, T) is the multi-head attention association value of data, outlier degree, and threshold.

[0100] Specifically, the technical parameters of the alarm signal generation unit include the alarm mode vector dimension (corresponding to the number of warning levels), the signal generation weight matrix dimension (matching the dimensions of the concatenated vector and the alarm signal vector), and the bias vector dimension (same as the alarm signal vector). This is to generate accurate alarm signals. During implementation, multi-head attention calculations are performed on the alarm mode vector, anomaly value, and threshold matrix, and then concatenated. The alarm signal vector is obtained after processing with the weight matrix and bias vector, triggering the corresponding audible and visual alarm. The technical parameters of the data storage control unit include the control coefficient (0.3-0.8) and the local outlier factor storage correction value range (0-1). This is to optimize the data storage strategy. During implementation, the multi-head attention correlation value of data, anomaly, and threshold is calculated, combined with the local outlier factor storage correction value and the sign function, to generate storage control instructions and control the data storage to local and remote servers.

[0101] Preferably, the feature extraction enhancement module includes the following units:

[0102] Multi-scale feature decomposition unit: Receives the feature tensor output by the parameter association mapping module, and decomposes the feature tensor into layers according to preset different scale division rules to obtain multiple feature sub-tensors with different dimensions and resolutions. Each feature sub-tensor corresponds to the preset spatial or temporal scale features of road slope deformation. The feature sub-tensors are transmitted to the feature fusion interface through the internal data bus.

[0103] Feature association mining unit: Receives the feature sub-tensors output by the multi-scale feature decomposition unit, deeply mines the potential correlations between slope displacement and soil moisture content parameters contained in different sub-tensors, and transforms the mined correlation information into a correlation coefficient matrix by constructing a feature association map, which is then transmitted to the feature enhancement unit.

[0104] Feature enhancement unit: Receives the correlation coefficient matrix output by the feature association mining unit and the feature sub-tensor output by the multi-scale feature decomposition unit. Based on the correlation coefficient matrix, it dynamically adjusts the weight of each feature sub-tensor, strengthens the feature components that are strongly correlated with slope deformation, and weakens irrelevant or weakly correlated components. The processed feature tensor is sent to the model training and inference module.

[0105] Feature Dimension Adaptation Unit: Receives the enhanced feature tensor output by the feature enhancement unit, performs dimension transformation and adaptation processing on the feature tensor according to the requirements of the model training and inference module for the input feature dimension, ensures that the dimension of the output feature tensor is completely matched with the dimension of the model input, and transmits the adapted feature tensor to the model training and inference module through the interface protocol.

[0106] Preferably, the model training and inference module includes the following units:

[0107] Attention Mechanism Configuration Unit: Receives the feature tensor output by the feature extraction enhancement module, and based on the feature attributes of the slope deformation parameters, initializes the number of attention heads, the dimension configuration of each head, and the attention calculation method of the optimized multi-head attention mechanism, generates the basic configuration parameters of the attention mechanism, and passes them to the attention weight update unit;

[0108] Attention Weight Update Unit: Receives the basic configuration parameters output by the attention mechanism configuration unit and the feature tensor output by the feature extraction enhancement module. Based on the variation law of slope displacement and pore water pressure parameters in the input feature tensor, it iteratively updates the weight matrix of each attention head through the gradient descent algorithm, so that the attention weight can better capture the important correlation between parameters. The updated weight matrix is ​​sent to the anomaly inference unit.

[0109] Local outlier factor model construction unit: Based on historical monitoring data of road slope deformation, the basic structure of the local outlier factor model of road slope is constructed, the parameters of the number of nearest neighbors and the distance calculation method in the model are determined, the initial local outlier factor model is formed, and it is passed to the anomaly inference unit.

[0110] Anomaly inference unit: Receives the attention weight matrix output by the attention weight update unit, the initial model output by the local outlier factor model construction unit, and the feature tensor output by the feature extraction enhancement module. It uses the attention weight matrix to weight the feature tensor, inputs the processing result into the local outlier factor model for calculation, obtains the anomaly value of road slope deformation, and transmits it to the early warning threshold determination module.

[0111] Preferably, the early warning threshold determination module includes the following units:

[0112] Threshold interval division unit: Based on the engineering geological conditions, historical deformation data and safety level requirements of the road slope, the slope deformation anomaly is divided into multiple continuous and non-overlapping threshold intervals. Each interval corresponds to a preset warning level. The division results are stored in the form of interval boundary values ​​and transmitted to the anomaly matching unit.

[0113] Anomaly matching unit: Receives the anomaly value output by the model training and inference module and the interval boundary value output by the threshold interval division unit, compares the anomaly value with each threshold interval one by one, determines the specific interval to which the anomaly value belongs, and generates an interval matching signal to send to the threshold dynamic adjustment unit.

[0114] Threshold dynamic adjustment unit: Receives the interval matching signal and historical early warning data output by the anomaly matching unit, analyzes the distribution pattern of anomaly values ​​and slope deformation trends in long-term monitoring, dynamically corrects the boundary values ​​of each threshold interval, so that the threshold interval can adapt to the long-term changing characteristics of slope deformation, and feeds back the adjusted interval boundary values ​​to the threshold interval division unit.

[0115] Warning signal generation unit: Receives the interval matching signal output by the anomaly degree matching unit, generates the corresponding warning signal code according to the matched warning level, the code contains the warning level and anomaly degree value range information, and sends the warning signal code to the warning response module through the signal transmission channel.

[0116] like Figure 2 As shown, a road slope deformation early warning and monitoring system based on geological data analysis is described. The system operation includes the following steps:

[0117] Step S1: Start the multi-type sensor array deployed by the parameter acquisition module, and continuously collect data on the slope displacement, soil moisture content, pore water pressure, slope top settlement, deep horizontal displacement and slope surface temperature field distribution of the road slope according to the preset sampling period, forming a time-series parameter data stream, and send the data stream to the parameter association mapping module through the data transmission link;

[0118] Step S2: After receiving the time series parameter data stream, the parameter association mapping module calls the built-in dynamic parameter association matrix construction algorithm to analyze the nonlinear coupling relationship between the slope displacement and other parameters, and converts the analyzed association features into feature tensors, which are then transmitted to the feature extraction and enhancement module through the interface.

[0119] Step S3: After receiving the feature tensor, the feature extraction and enhancement module uses a multi-scale feature decomposition algorithm to decompose the feature tensor into layers, obtaining feature sub-tensors of different scales. After feature fusion processing, the weights of each feature sub-tensor are dynamically adjusted to strengthen the feature components. The processed feature tensor is then transmitted to the model training and inference module.

[0120] Step S4: After receiving the feature tensor, the model training and inference module performs weight allocation and association learning on the feature tensor based on the optimized multi-head attention mechanism to generate a high-dimensional feature vector. Then, the high-dimensional feature vector is input into the local outlier factor model of the road slope to calculate the anomaly degree, obtain the anomaly degree value of the slope deformation, and send it to the early warning threshold judgment module.

[0121] Step S5: After receiving the abnormality value, the early warning threshold determination module matches it with the preset multi-level early warning threshold range to determine the corresponding early warning level, generates a corresponding early warning signal according to the early warning level, and sends the early warning signal to the early warning response module through the signal transmission channel.

[0122] Step S6: After receiving the warning signal, the warning response module triggers the corresponding audible and visual alarm device according to the level of the warning signal. At the same time, it starts the data storage program to store the current monitoring data and warning information, and transmits the warning level information to the remote monitoring terminal in real time through the communication link.

[0123] A road slope deformation early warning and monitoring system based on geological data analysis demonstrates significant advantages in parameter processing through multi-module collaborative operation. The parameter acquisition module acquires various data types, including slope displacement and soil moisture content. The parameter correlation mapping module analyzes the nonlinear coupling relationships between these parameters, overcoming the limitations of traditional single-dimensional analysis and preventing the loss of parameter correlation information. By leveraging the feature extraction enhancement module to decompose and fuse multi-scale feature sub-tensors, the system strengthens feature components strongly correlated with slope deformation, enabling it to more comprehensively capture the overall dynamic characteristics of slope deformation and effectively compensate for the shortcomings of existing technologies in parameter correlation analysis.

[0124] In terms of model application, the system innovatively integrates and optimizes a multi-head attention mechanism and a local outlier factor model for road slopes. In the model training and inference module, the attention mechanism performs deep learning on the dynamic correlation of multiple parameters, optimizing the allocation of attention weights; the local outlier factor model accurately calculates anomalies, improving adaptability to data with uneven spatiotemporal distribution. The combination of these two mechanisms enhances the ability to capture subtle local deformation features of slopes, solving the problem that traditional models struggle to distinguish between normal fluctuations and abnormal deformations, and reducing early warning delays and misjudgments.

[0125] The subdivided units of each module in the system further enhance its technological advantages. The feature extraction enhancement module employs multi-unit processing, progressively optimizing feature representation from decomposition and correlation mining to enhancement and adaptation. The model training and inference module continuously optimizes inference logic by configuring and updating attention mechanisms and constructing and running local outlier factor models. The early warning threshold determination module dynamically adjusts threshold ranges, and the early warning response module triggers responses in stages. These designs make the system more efficient in mining overall parameter features and accurately identifying abnormal deformations, comprehensively overcoming the shortcomings of previous technologies and achieving reliable early warning monitoring of road slope deformation.

[0126] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0127] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A road slope deformation pre-warning monitoring system for geological data analysis, characterized in that, include: The parameter acquisition module, through a multi-type sensor array, acquires real-time data on slope displacement, soil and rock moisture content, pore water pressure, slope top settlement, deep horizontal displacement, and slope surface temperature field distribution. The acquired parameters are transmitted to the parameter association and mapping module in the form of a time-series data stream. The parameter acquisition module also includes an adaptive sampling unit, which dynamically adjusts the acquired parameters based on an optimized multi-head attention mechanism. The adjustment model formula is as follows: ； in, for Sampling frequency at time , Let be the sampling frequency at time t. This is the attention weight adjustment coefficient. This represents the influence coefficient of the local outlier factor. Based on Parameters are collected at all times Slope displacement Soil and rock water content Optimization of multi-head attention value for pore water pressure. Based on slope displacement The road slope local outlier sampling correction value; the association strength calculation unit in the parameter association mapping module adopts an optimized multi-head attention mechanism combined with the road slope local outlier model to construct the parameter association strength calculation model: ； in, Let be the association strength value between the i-th type of parameter and the j-th type of parameter. To collect sample size for parameters, For the k-th sample value of the i-th type parameter, For the k-th sample value of the j-th class parameter, The multi-head attention association value between the i-th class and the k-th sample of the j-th class is given. The local outlier association factor between the i-th class and the k-th sample with parameters of the j-th class; The parameter association mapping module receives the time-series data stream output by the parameter acquisition module. Through the constructed dynamic parameter association matrix, it analyzes the nonlinear coupling relationship between slope displacement and soil moisture content, pore water pressure and deep horizontal displacement, and slope top settlement and temperature field distribution. The analysis results are sent to the feature extraction and enhancement module in the form of feature tensors. The feature extraction and enhancement module receives the feature tensor output by the parameter association mapping module, uses a multi-scale feature decomposition algorithm to split the feature tensor into dimensions, and obtains slope deformation feature sub-tensors at different scales. Each sub-tensor is then transmitted to the model training and inference module through the feature fusion interface. The model training and inference module receives the feature sub-tensors output by the feature extraction and enhancement module, calls the built-in optimized multi-head attention mechanism to perform weight allocation and association learning on the feature sub-tensors, generates high-dimensional feature vectors, and then inputs the high-dimensional feature vectors into the local outlier factor model of the road slope for anomaly calculation. The calculation results are then sent to the early warning threshold determination module.

2. The system according to claim 1, characterized in that, The system also includes: Warning threshold determination module: Receives the anomaly calculation results output by the model training and inference module, combines them with the preset multi-level warning threshold range, performs interval matching on the anomaly values, and transmits the matching results to the warning response module in the form of a signal; Early warning response module: Receives the matching signal output by the early warning threshold determination module, triggers the corresponding audible and visual alarm device and data storage instructions according to the early warning level corresponding to the signal, and transmits the early warning level information to the remote monitoring terminal through the communication link.

3. The system according to claim 1, characterized in that, The feature extraction enhancement module includes a feature dimension expansion unit, which utilizes an optimized multi-head attention mechanism to increase the dimension of the feature sub-tensor. The expansion model formula is as follows: ； in, For the extended high-dimensional feature tensor, For feature subtensors of different scales, Concat is the feature concatenation function. Expand the weight matrix to accommodate the dimensions. The bias vector is used for dimensional expansion, and MHAtn is an optimized multi-head attention computation function. Simultaneously, the feature enhancement unit of this module employs a roadside slope local outlier factor model to enhance features; the model formula is as follows: ； in, For the enhanced feature tensor, To enhance the coefficient, For feature tensors The local outlier feature enhancement value.

4. The system of claim 1, wherein, The attention weight update unit in the model training and inference module dynamically adjusts the weights through an optimized multi-head attention mechanism. The model formula is as follows: ； in, for Attention weight matrix at time step Let be the attention weight matrix at time t. For learning rate, For about gradient operator, To focus on the number of heads, Let be the query matrix, key matrix, and value matrix of the m-th attention head, respectively. The dimension of the key vector. The softmax activation function is used. For weight matrix based The local outlier factor weight correction value; the anomaly inference unit of this module uses the local outlier factor model of road slope for calculation, and the formula is: ； Wherein, OD represents the overall outlier value. The total number of samples, Let k be the set of the k nearest neighbors of the i-th sample. For the sample and The distance between them For the input feature samples, For the sample and Multi-head attention correlation value.

5. The system of claim 2, wherein, The threshold dynamic adjustment unit of the early warning threshold determination module constructs a threshold adjustment model by combining an optimized multi-head attention mechanism, and the formula is: ； in, The new warning threshold, The original warning threshold, It is the sigmoid activation function. This is a historical early warning data matrix. This represents the recent change in slope displacement. This is the threshold adjustment coefficient. Based on the original threshold The local outlier threshold correction value; the interval matching unit of this module is matched using the following formula: ； Where Match is the matching result vector. The number of warning levels, This represents the Level I warning threshold, where `round` is the rounding function. Here, OD is the one-hot coding function, and OD is the outlier value. The multi-head attention matching degree between the anomaly score and the level I threshold. This is the local outlier factor matching correction value based on the outlier degree OD.

6. The system of claim 2, wherein, The alarm signal generation unit of the early warning response module generates signals using an optimized multi-head attention mechanism, with the following formula: ； in, This is the alarm signal vector. This represents the Level I alarm mode vector, where OD is the anomaly value. This is the early warning threshold matrix. Generate a weight matrix for the signal. A bias vector is generated for the signal; Concat() is the vector concatenation function. The data storage control unit of this module adjusts the storage strategy through the local outlier factor model of the road slope, as shown in the formula: Store Where Store is the storage control command. The data matrix to be stored For control coefficients, For symbolic functions, For data-based Local outlier storage correction value, This represents the multi-head attention correlation value for data, anomaly rate, and threshold.

7. The system according to claim 1, characterized in that, The feature extraction and enhancement module includes the following units: The multi-scale feature decomposition unit receives the feature tensor output by the parameter association mapping module, and decomposes the feature tensor into layers according to the preset different scale division rules to obtain multiple feature sub-tensors with different dimensions and resolutions. Each feature sub-tensor corresponds to the preset spatial or temporal scale features of road slope deformation. The feature sub-tensors are transmitted to the feature fusion interface through the internal data bus. The feature association mining unit receives the feature sub-tensors output by the multi-scale feature decomposition unit, and deeply mines the potential correlations between slope displacement and soil moisture content parameters contained in different sub-tensors. By constructing a feature association map, the mined association information is converted into an association coefficient matrix and transmitted to the feature enhancement unit. The feature enhancement unit receives the correlation coefficient matrix output by the feature association mining unit and the feature sub-tensor output by the multi-scale feature decomposition unit. Based on the correlation coefficient matrix, it dynamically adjusts the weight of each feature sub-tensor, strengthens the feature components that are strongly correlated with slope deformation, and weakens the irrelevant or weakly correlated components. The processed feature tensor is then sent to the model training and inference module. The feature dimension adaptation unit receives the enhanced feature tensor output by the feature enhancement unit, performs dimension transformation and adaptation processing on the feature tensor according to the requirements of the model training and inference module for the input feature dimension, and ensures that the dimension of the output feature tensor is completely matched with the dimension of the model input. The adapted feature tensor is then transmitted to the model training and inference module through the interface protocol.

8. The system of claim 1, wherein, The model training and inference module includes the following units: The attention mechanism configuration unit receives the feature tensor output by the feature extraction enhancement module, and sets the number of attention heads, the dimension configuration of each head, and the attention calculation method of the optimized multi-head attention mechanism according to the feature attributes of the slope deformation parameters. It generates the basic configuration parameters of the attention mechanism and passes them to the attention weight update unit. The attention weight update unit receives the basic configuration parameters output by the attention mechanism configuration unit and the feature tensor output by the feature extraction enhancement module. Based on the variation law of slope displacement and pore water pressure parameters in the input feature tensor, it iteratively updates the weight matrix of each attention head through the gradient descent algorithm, so that the attention weight can better capture the important correlation between parameters. The updated weight matrix is ​​sent to the anomaly inference unit. The local outlier factor model construction unit, based on historical monitoring data of road slope deformation, constructs the basic structure of the local outlier factor model of road slope, determines the parameters of the number of nearest neighbors and the distance calculation method in the model, forms the initial local outlier factor model, and passes it to the anomaly inference unit; The anomaly inference unit receives the attention weight matrix output by the attention weight update unit, the initial model output by the local outlier factor model construction unit, and the feature tensor output by the feature extraction enhancement module. It uses the attention weight matrix to weight the feature tensor, inputs the processing result into the local outlier factor model for calculation, obtains the anomaly value of the road slope deformation, and transmits it to the early warning threshold determination module.

9. The system of claim 2, wherein, The warning threshold determination module includes the following units: The threshold interval division unit divides the slope deformation anomaly into multiple continuous and non-overlapping threshold intervals based on the engineering geological conditions, historical deformation data and safety level requirements of the road slope. Each interval corresponds to a preset warning level. The division results are stored in the form of interval boundary values ​​and transmitted to the anomaly matching unit. The anomaly matching unit receives the anomaly value output by the model training and inference module and the interval boundary value output by the threshold interval division unit. It compares the anomaly value with each threshold interval one by one to determine the specific interval to which the anomaly value belongs, and generates an interval matching signal to send to the threshold dynamic adjustment unit. The threshold dynamic adjustment unit receives the interval matching signal and historical early warning data output by the anomaly matching unit, analyzes the distribution pattern of anomaly values ​​and slope deformation trends in long-term monitoring, dynamically corrects the boundary values ​​of each threshold interval, so that the threshold interval can adapt to the long-term changing characteristics of slope deformation, and feeds back the adjusted interval boundary values ​​to the threshold interval division unit. The early warning signal generation unit receives the interval matching signal output by the anomaly degree matching unit, generates the corresponding early warning signal code according to the matched early warning level, and the code contains the early warning level and anomaly degree value range information. The early warning signal code is sent to the early warning response module through the signal transmission channel.

Images