A multi-section tunnel geological risk federated learning prediction method and system

Abstract

The application discloses a kind of multi-section tunnel geological risk federated learning prediction method and system, specifically related to tunnel engineering safety monitoring technical field, for solving the feature conflict and model precision decline problem caused by ignoring geological space heterogeneity in existing federated learning;By obtaining the geological risk feature vector and three-dimensional geological coordinates output by each section local model;Calculate the cosine similarity and spatial Euclidean distance of the feature vectors of any two sections;The cosine similarity and the reciprocal of spatial Euclidean distance are exponentially attenuated weight weighted summation to generate geological correlation weight;Based on the space coherence coefficient of event cluster generated by clustering the time and space coordinates of surrounding rock microseismic events;When the coefficient is lower than the threshold, isolate abnormal section, and generate a global model by weighting and aggregating the model parameters of non-isolated sections using the geological correlation weight;Realize the collaborative modeling of geological space structure and dynamic risk, improve the risk prediction accuracy in multi-section complex geological environment under the premise of protecting data privacy.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering safety monitoring technology, and more specifically, to a federated learning method and system for predicting geological risks in multi-section tunnels. Background Technology

[0002] In multi-section tunnel construction, geological risk prediction requires the construction of a global model by integrating geological data from each section. Due to privacy and commercial confidentiality requirements regarding data between sections (such as geological survey data and TBM real-time monitoring data), contractors generally adopt a federated learning framework to enable each section to collaboratively train the risk prediction model without sharing raw data. Existing technologies achieve data privacy protection through local risk prediction model training and parameter aggregation, meeting the collaborative needs across construction entities.

[0003] However, the geological structures of different sections in the tunnel project have significant spatial heterogeneity (such as fault zones, water-rich areas, rockburst areas, etc.), which leads to the non-independent homogeneity of the feature distribution extracted by the local risk prediction model. The standard aggregation mechanism of federated learning ignores the spatial structural differences of geological features, which causes feature conflicts during the aggregation process, causing the global model to converge to a suboptimal solution, thus weakening the model's risk prediction accuracy in complex geological environments with multiple sections. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides a federated learning prediction method and system for geological risks of multi-section tunnels to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A federated learning method for predicting geological risks in multi-section tunnels includes the following steps:

[0007] S1. Obtain the geological risk feature vectors output by the local risk prediction model for each section and the corresponding three-dimensional geological coordinates of the section.

[0008] S2. Calculate the cosine similarity between the geological risk feature vectors of any two sections;

[0009] S3. Calculate the spatial Euclidean distance between corresponding sections based on the three-dimensional geological coordinates;

[0010] S4. Generate the geological association weights between the sections by weighted summation of cosine similarity and the reciprocal of spatial Euclidean distance;

[0011] S5. Generate the spatial coherence coefficient of event clusters based on the spatiotemporal coordinates of microseismic events in the surrounding rock of each section;

[0012] S6. When the spatial coherence coefficient is lower than the preset threshold, the geological risk feature vector of the corresponding section is isolated, and the parameters of the local risk prediction model of the unisolated section are weighted and aggregated using geological correlation weights to generate a global geological risk prediction model.

[0013] Furthermore, the geological risk feature vectors output by the local risk prediction model for each contract section and the corresponding three-dimensional geological coordinates of the contract section are obtained, including:

[0014] The raw geological data is processed by the local risk prediction model deployed in each section, and a geological risk feature vector is output.

[0015] The longitude, latitude, and altitude values ​​of the center point of each section are extracted from the engineering geological survey report to form a three-dimensional geological coordinate system.

[0016] Establish and store the mapping relationship between geological risk feature vectors and three-dimensional geological coordinates according to the section number.

[0017] Furthermore, the cosine similarity between the geological risk feature vectors of any two contract sections is calculated, including:

[0018] Obtain the geological risk feature vectors of the first and second sections from the stored mapping relationships;

[0019] Calculate the dot product of the geological risk feature vector of the first section and the geological risk feature vector of the second section;

[0020] Calculate the magnitude of the geological risk characteristic vector for the first contract section;

[0021] Calculate the magnitude of the geological risk characteristic vector for the second section;

[0022] Dividing the dot product by the product of the magnitudes of the geological risk feature vectors of the first and second sections yields the cosine similarity between the geological risk feature vectors of the two sections.

[0023] Furthermore, the spatial Euclidean distance between corresponding sections is calculated based on three-dimensional geological coordinates, including:

[0024] Obtain the three-dimensional geological coordinates of the first section and the two sections from the stored mapping relationship;

[0025] Calculate the difference between the longitude values ​​in the three-dimensional geological coordinates of the first section and the longitude values ​​in the three-dimensional geological coordinates of the second section;

[0026] Calculate the difference between the latitude value in the three-dimensional geological coordinates of the first section and the latitude value in the three-dimensional geological coordinates of the second section;

[0027] Calculate the difference between the elevation values ​​in the three-dimensional geological coordinates of the first section and the elevation values ​​in the three-dimensional geological coordinates of the second section;

[0028] Add the squares of the longitude difference, the latitude difference, and the altitude difference;

[0029] Take the square root of the sum to obtain the spatial Euclidean distance between the two sections.

[0030] Furthermore, the cosine similarity and the reciprocal of the spatial Euclidean distance are weighted and summed to generate the geological association weights between the sections, including:

[0031] Obtain the cosine similarity between the first and second contract segments, and obtain the spatial Euclidean distance between the first and second contract segments;

[0032] Calculate the reciprocal of the Euclidean distance in space;

[0033] Calculate the weighting coefficients for spatial Euclidean distance;

[0034] Multiply the cosine similarity by a fixed weight coefficient;

[0035] Multiply the reciprocal of the spatial Euclidean distance by the weighting coefficient of the spatial Euclidean distance;

[0036] The geological association weight between the first and second sections is generated by adding the product of the cosine similarity multiplied by the fixed weight coefficient and the product of the inverse of the spatial Euclidean distance multiplied by the weight coefficient of the spatial Euclidean distance.

[0037] Furthermore, the weighting coefficient of the spatial Euclidean distance is an exponential function with the natural constant e as the base, and the power of the exponential function is a negative decay coefficient multiplied by the spatial Euclidean distance.

[0038] Furthermore, the weighting coefficient of the spatial Euclidean distance decreases exponentially as the spatial Euclidean distance increases.

[0039] Furthermore, based on the spatiotemporal coordinates of microseismic events in the surrounding rock of each section, the spatial coherence coefficients of event clusters are generated, including:

[0040] Obtain the spatial coordinates of all microseismic events in the surrounding rock within the current section;

[0041] Cluster analysis is performed on the spatial coordinates of all surrounding rock microseismic events within the current section to form at least one event cluster;

[0042] Calculate the centroid location of the spatial coordinates of all surrounding rock microseismic events within the event cluster;

[0043] Calculate the spatial Euclidean distance from the spatial coordinates of each microseismic event in the surrounding rock within the event cluster to the centroid location;

[0044] Calculate the average of all spatial Euclidean distances within the event cluster;

[0045] The reciprocal of the average spatial Euclidean distance within an event cluster is taken as the spatial coherence coefficient of the event cluster.

[0046] Furthermore, when the spatial coherence coefficient is lower than a preset threshold, the geological risk feature vector of the corresponding contract segment is isolated, and the parameters of the local risk prediction model of the unisolated contract segment are weighted and aggregated using geological correlation weights to generate a global geological risk prediction model, including:

[0047] Obtain the spatial coherence coefficient of the event cluster in the current segment, and compare the spatial coherence coefficient with a preset threshold.

[0048] When the spatial coherence coefficient is less than a preset threshold, the geological risk feature vector of the corresponding section is prohibited from participating in the aggregation operation;

[0049] Obtain the local risk prediction model parameters corresponding to the geological risk feature vectors of all unisolated sections from the stored mapping relationship;

[0050] Obtain the geological correlation weights between all pairs of unisolated sections;

[0051] Calculate the weighted average of the local risk prediction model parameters for all unisolated sections, where the weight of each section is the sum of the geological association weights of all sections that have geological association weights with the corresponding section;

[0052] The weighted average value is used as a parameter in the global geological risk prediction model.

[0053] On the other hand, the present invention provides a federated learning prediction system for geological risks of multi-section tunnels, comprising the following modules:

[0054] The geological acquisition module is used to obtain the geological risk feature vectors output by the local risk prediction model for each section and the corresponding three-dimensional geological coordinates of the section.

[0055] The similarity calculation module is used to calculate the cosine similarity between the geological risk feature vectors of any two sections.

[0056] The spatial distance module is used to calculate the spatial Euclidean distance between corresponding sections based on three-dimensional geological coordinates.

[0057] The weight generation module is used to generate geological association weights between sections by weighted summation of cosine similarity and the reciprocal of spatial Euclidean distance. The weight coefficient of spatial Euclidean distance decreases exponentially as spatial Euclidean distance increases.

[0058] The transmission processing module is used to generate the spatial coherence coefficient of the event cluster based on the spatiotemporal coordinates of the surrounding rock microseismic events in each section;

[0059] The model aggregation module is used to isolate the geological risk feature vector of the corresponding section when the spatial coherence coefficient is lower than a preset threshold, and to use geological correlation weights to weight and aggregate the local risk prediction model parameters of the unisolated section to generate a global geological risk prediction model.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. By using spatial correlation modeling of geological risk feature vectors and dynamic isolation mechanism of abnormal sections, the accuracy of geological risk prediction for multi-section tunnels is effectively improved. First, based on the dual constraints of spatial Euclidean distance of three-dimensional geological coordinates and cosine similarity of geological risk feature vectors, a geological correlation weight between sections is constructed. This weight dynamically adjusts the contribution of adjacent sections through an exponential decay function of spatial distance, accurately quantifies the spatial continuity characteristics of geological structures, and makes the federated aggregation process reflect both the proximity of the physical location of the rock mass and capture the similarity of geological risk characteristics, fundamentally alleviating the problem of standard federated learning feature conflict caused by spatial heterogeneity such as fault zones and rockburst areas.

[0062] 2. The spatial coherence coefficient of the surrounding rock microseismic event cluster is introduced as a dynamic monitoring indicator of geological stability. When the spatial coherence coefficient is lower than the threshold, abnormal section data is automatically isolated to realize intelligent filtering driven by geological risk events. This not only retains the collaborative learning benefits of normal rock mass areas, but also blocks the noise transmission of local unstable areas, so that the global model always maintains high robustness in complex geological environments. Compared with traditional federated learning, it improves the accuracy of multi-section tunnel risk prediction and accelerates the convergence speed while ensuring data privacy. Attached Figure Description

[0063] Figure 1 This is a flowchart of a federated learning method for predicting geological risks in multi-section tunnels according to the present invention.

[0064] Figure 2 This is a schematic diagram of the structure of a federated learning prediction system for geological risks in multi-section tunnels according to the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1: Figure 1 This invention presents a federated learning method for predicting geological risks in multi-section tunnels, comprising the following steps:

[0067] S1. Obtain the geological risk feature vectors output by the local risk prediction model for each section and the corresponding three-dimensional geological coordinates of the section.

[0068] S2. Calculate the cosine similarity between the geological risk feature vectors of any two sections;

[0069] S3. Calculate the spatial Euclidean distance between corresponding sections based on the three-dimensional geological coordinates;

[0070] S4. Generate the geological association weights between the sections by weighted summation of cosine similarity and the reciprocal of spatial Euclidean distance;

[0071] S5. Generate the spatial coherence coefficient of event clusters based on the spatiotemporal coordinates of microseismic events in the surrounding rock of each section;

[0072] S6. When the spatial coherence coefficient is lower than the preset threshold, the geological risk feature vector of the corresponding section is isolated, and the parameters of the local risk prediction model of the unisolated section are weighted and aggregated using geological correlation weights to generate a global geological risk prediction model.

[0073] In the specific implementation process, the geological risk feature vectors output by the local risk prediction models for each section and the corresponding three-dimensional geological coordinates of the section are obtained. First, the raw geological data is processed using the local risk prediction models deployed in each section. The local risk prediction models adopt a convolutional neural network architecture, which contains three convolutional layers and two fully connected layers. The raw geological data received by the input layer includes the uniaxial compressive strength of rock from borehole cores, rock integrity coefficient, joint and fracture development degree, groundwater permeability coefficient, and geostress measurements. The raw geological data for each section is collected in real time by multi-parameter geological monitoring sensors deployed at the construction site of that section. Sensor types include rock mass acoustic wave testers, borehole camera equipment, and pore water pressure gauges. After standardizing and preprocessing the raw geological data, the local risk prediction models input the data into the convolutional layers. After feature extraction and dimensionality reduction operations, a geological risk feature vector with a dimension of 256 is output at the fully connected layers.

[0074] Subsequently, the longitude, latitude, and altitude values ​​of the center point of each section were extracted from the engineering geological survey reports of each section. The engineering geological survey reports were prepared by a Class A qualified surveying unit based on field geological mapping and exploration operations. The reports clearly marked the planar location range and elevation information of each section. The method for determining the center point of each section was as follows: first, the planar coordinates of the inflection points of the section boundaries were obtained; then, the arithmetic mean of the longitude values ​​of all inflection points was calculated as the longitude value of the center point; the arithmetic mean of the latitude values ​​of all inflection points was calculated as the latitude value of the center point; finally, the median of the highest and lowest altitudes within the section was taken as the altitude value of the center point. For example, if the coordinates of the boundary inflection points of a certain section are: point A, longitude 112.35 degrees, latitude 28.41 degrees; point B, longitude 112.38 degrees, latitude 28.43 degrees; then the longitude of the center point is 112.365 degrees, and the latitude is 28.42 degrees. If the elevation range of this section is between 85 meters and 105 meters, then the elevation of the center point is determined to be 95 meters. This forms a three-dimensional geological coordinate system containing longitude, latitude, and elevation values.

[0075] Finally, the geological risk feature vectors and 3D geological coordinates are mapped and stored according to the section number. This mapping is achieved by creating a hash table data structure. The key of the hash table is the section number string, which follows the unified engineering naming convention and consists of the project code plus a three-digit serial number. The hash table value is set as a tuple containing two elements: the first element stores a floating-point array of the geological risk feature vectors, and the second element stores a floating-point array of the 3D geological coordinates. The specific storage process is as follows: after the local risk prediction model completes the calculation of the geological risk feature vectors, the system automatically retrieves the electronic document of the engineering geological survey report corresponding to the section, parses the coordinate data table marked in the report PDF file, extracts the 3D geological coordinate values, and then generates key-value pairs to write to the hash table. This hash table is persistently stored in the geological feature table of a relational database. The database table structure includes a section number field, a binary feature vector field, and a coordinate value field. The binary feature vector field uses a BLOB type to store a 256-dimensional floating-point array, and the coordinate value field is divided into three floating-point columns: longitude, latitude, and altitude.

[0076] For the deployment of local risk prediction models, edge computing servers are configured at each construction site. These servers are equipped with NVIDIA Jetson AGX Orin computing modules, running the Ubuntu operating system and the PyTorch deep learning framework. The raw geological data acquisition frequency is set to once every 30 minutes, with each acquisition including rock mass parameter measurements from 10 borehole points. The generation process of the geological risk feature vector is as follows: the edge server receives raw data in JSON format transmitted from the sensor network, checks whether each parameter value is within the valid physical range, then calls the prediction interface of the local risk prediction model, outputs the geological risk feature vector, and records the timestamp. The three-dimensional geological coordinate update mechanism is as follows: when the construction scope of a section changes, the engineering management personnel upload a new report to the survey report management system, and the system automatically triggers the coordinate resolution service to update the corresponding coordinate values ​​in the database. A hash table caches the latest data in memory and persists it to the database hourly to ensure the real-time nature and consistency of the mapping relationship. This implementation method meets the real-time requirements of the engineering site while ensuring the accurate association between geological feature data and spatial coordinates, providing a complete and reliable data foundation for subsequent steps.

[0077] The equipment used throughout the implementation process are all general-purpose equipment in the field of engineering monitoring. The local risk prediction model is modified from the ResNet-18 network structure. The database system uses the open-source PostgreSQL relational database, and the hash table implements the dictionary data type using the Python standard library.

[0078] In the specific implementation process, when calculating the cosine similarity between the geological risk feature vectors of any two sections, the geological risk feature vectors of the first and second sections are first obtained from the stored mapping relationship. The mapping relationship is stored in the geological feature table of the relational database, which contains a section number field, a binary feature vector field, and a coordinate value field. The retrieval operation is performed using a structured query language. Based on the input first and second section number strings, the corresponding record's binary feature vector field is queried. The byte stream stored in the binary feature vector field is parsed into a 256-dimensional floating-point array. The parsing process uses the struct module of the Python language to perform byte unpacking operations, converting every 4 bytes of data into single-precision floating-point numbers, and finally reconstructing the complete geological risk feature vector. For example, when the input section number is "TBM-015", the database query returns a binary object of length 1024 bytes, which, after parsing, forms an array containing 256 floating-point values, each floating-point number representing one dimension component of the geological risk feature vector.

[0079] After obtaining the geological risk feature vectors of the first and second sections, their dot product is calculated. The dot product calculation is achieved by iterating through each dimension of the two feature vectors. Specifically: the accumulator variable is initialized to 0.0; the loop index ranges from 0 to 255, and in each loop, the floating-point value of the geological risk feature vector of the first section at the index position is retrieved, along with the floating-point value of the geological risk feature vector of the second section at the same index position; the two floating-point numbers are multiplied to obtain the product; and the product is added to the accumulator variable. The value of the accumulator variable after the loop ends is the dot product result. This process requires ensuring that the dimensions of the two geological risk feature vectors are strictly equal. In implementation, a pre-check mechanism is used to verify this: before the loop begins, the length attributes of the two arrays are compared; if the lengths are not equal, the calculation is terminated and a dimension exception error is thrown. The dot product calculation uses double-precision floating-point operations throughout to ensure numerical accuracy and avoid precision loss due to large-scale accumulation.

[0080] The modulus of the geological risk feature vector for the first section is then calculated. The modulus calculation uses the Euclidean norm method, with the following steps: initialize the modulus square accumulator to 0.0; iterate through the 256 dimensions of the geological risk feature vector for the first section; in each iteration, extract the floating-point value of the current dimension; square this value and accumulate it into the modulus square accumulator; after the iteration, take the square root of the modulus square accumulator. The square root operation is implemented using standard functions from a mathematical library, employing Newton's iteration method for numerical approximation, with the iteration termination condition set to the difference between two adjacent iterations being less than 0.0001. To handle the special case of zero vectors, the modulus square accumulator value is checked before taking the square root: if the value is less than the machine's minimum floating-point threshold, the modulus is directly returned as 0.0, and an exception log is recorded. The modulus of the geological risk feature vector for the second section is calculated independently using the same process.

[0081] After completing the above calculations, the dot product is divided by the product of the modulus of the geological risk feature vectors of the first and second sections. A denominator verification mechanism is implemented before this division operation: the absolute value of the product of the two modulus lengths is calculated; if it is less than the numerical stability threshold, a preset invalid value is returned to avoid division by zero errors. Specifically, the numerical stability threshold is set to 10 to the power of -10. When the denominator is less than this threshold, the two geological risk feature vectors are determined to be orthogonal, and the cosine similarity is directly returned as 0.0. If the denominator is valid, a precise division operation is performed: the dot product value is converted to a double-precision floating-point number and divided by the product of the modulus lengths as represented by the double-precision floating-point number. The division result retains 8 significant decimal places, and excess digits are truncated using rounding rules. The final result is the cosine similarity between the geological risk feature vectors of the two sections.

[0082] The hardware resources involved in the computation process are configured as follows: the computing nodes deployed on the central server are equipped with Intel Xeon Gold processors, supporting AVX-512 instruction set acceleration for floating-point operations. Vector computation adopts multi-threaded parallel optimization, specifically implemented by dividing the 256-dimensional vector into 16 consecutive data blocks, with each thread processing the local dot product or modulus square calculation of the 16-dimensional data, and finally summing the results of each thread. The numerical computation library is linked to the Intel Mathematics Core Library, with automatic parallel compilation enabled. The exception handling mechanism includes: starting high-precision calculation mode when the dot product overflows; switching to logarithmic space operation when the modulus calculation exceeds the safe range. All intermediate and final results are written to a distributed file system, stored in JSON object format, containing fields: first segment number, second segment number, dot product value, first modulus value, second modulus value, modulus product value, cosine similarity value, and calculation timestamp. For example, a calculation record shows: the dot product of segment A and segment B is 158.72, the modulus A is 25.83, the modulus B is 32.17, the product of modulus lengths is 830.91, and the final cosine similarity is 0.191.

[0083] The entire implementation process established a calculation correctness verification workflow: after every 100 calculations, 3 sets of results were randomly selected for manual review. The review method was as follows: the original feature vector data was exported to the Matlab environment, the built-in cosine similarity function was used to calculate the result, and the difference was compared with the result of this system. An allowable error threshold of 0.001 was set, and when the difference exceeded the threshold, the calculation process audit was triggered. A periodic calibration mechanism was adopted during implementation to ensure numerical consistency: benchmark tests were run monthly using a standard test dataset. The standard dataset contained feature vector pairs with known cosine similarity to verify the deviation range between the system output and the theoretical value.

[0084] In the specific implementation process, the spatial Euclidean distance between corresponding sections is calculated based on the three-dimensional geological coordinates. First, the three-dimensional geological coordinates of the first and second sections are obtained from the stored mapping relationship. The mapping relationship is stored in the geological feature table of a relational database, which contains a section number field and a coordinate value field. The retrieval operation is performed through a database query interface. Based on the input first and second section number strings, the coordinate value field of the corresponding record is retrieved. The coordinate value field contains three independent floating-point columns: the longitude column stores decimal longitude values ​​in degrees, the latitude column stores decimal latitude values ​​in degrees, and the altitude column stores altitude values ​​in meters. The coordinate data returned by the database query is encapsulated as a tuple structure containing three elements: the first element is the floating-point longitude value, the second element is the floating-point latitude value, and the third element is the floating-point altitude value. During implementation, a data validity verification mechanism should be established: check whether the longitude value is within the range of -180 degrees to +180 degrees, whether the latitude value is within the range of -90 degrees to +90 degrees, and whether the altitude value is within the reasonable engineering range of -1000 meters to +9000 meters.

[0085] After obtaining the three-dimensional geological coordinates, the difference between the longitude values ​​of the first and second sections is calculated. The difference calculation uses double-precision floating-point arithmetic. Specifically, the longitude difference is obtained by subtracting the longitude value of the second section from the longitude value of the first section. This process considers the characteristics of Earth's spherical coordinates. When the two sections are located in East and West longitude regions respectively, a longitude sign adjustment mechanism is implemented: if the longitude value of the first section is positive and the longitude value of the second section is negative, then 360 degrees is added to the longitude value of the second section before calculating the difference; otherwise, the same processing is performed on the longitude value of the first section. The difference result is retained to 10 decimal places, and the unit of longitude difference is degrees. The same logic is independently applied to the latitude difference calculation: the latitude difference is obtained by subtracting the latitude value of the second section from the latitude value of the first section, and the unit of latitude difference is also degrees. For calculating the altitude difference, the altitude value of the first section is directly subtracted from the altitude value of the second section, and the unit of the difference is kept in meters.

[0086] After calculating the differences, the longitude, latitude, and altitude differences are squared. Before squaring, unit standardization is performed: the longitude and latitude differences are converted to metric units. The conversion method is based on the Earth ellipsoid model, specifically: the metric value of the longitude difference equals the degree of longitude difference multiplied by the length of the meridian at that latitude, calculated as 111319.488 multiplied by the cosine of the latitude value; the metric value of the latitude difference equals the degree of latitude difference multiplied by the constant 111319.488 for the length of the latitude at that latitude. The altitude difference is directly calculated in meters. The squaring operation uses the power function from the math library, with an exponent parameter of 2. The results are the squares of the longitude, latitude, and altitude differences, all in square meters.

[0087] The squares of the longitude difference, latitude difference, and altitude difference are then added together. The addition operation uses a double-precision floating-point accumulator, with the following order: first, the squares of the longitude difference are loaded into the accumulator; then, the squares of the latitude difference are added; and finally, the squares of the altitude difference are added. An overflow protection mechanism is implemented during the accumulation process: when the accumulator value exceeds 10 to the power of 15, it automatically switches to high-precision decimal arithmetic mode. The addition result is stored as a temporary variable representing the square of the straight-line distance in three-dimensional space, in square meters.

[0088] Finally, the square root of the sum is taken to obtain the spatial Euclidean distance between the two segments. The square root operation uses the Newton-Raphson iterative algorithm, with the initial estimate set to half of the sum. The iteration terminates when the difference between two consecutive iterations is less than 0.0001 meters. To handle special numerical cases, a pre-check rule is implemented: if the sum is less than zero (due to floating-point rounding error), the absolute value is taken before taking the square root; if the sum is equal to zero, the spatial Euclidean distance is directly returned as zero. The calculation result is converted to a single-precision floating-point number, retaining three decimal places, and the unit is uniformly set to meters.

[0089] A complete error control system was established for the calculation process: the coordinate reference system was calibrated before each calculation, and the WGS-84 geodetic coordinate system was uniformly adopted. At the hardware level, the processor's SIMD instruction set was used to perform the squaring operations of the three dimensions in parallel, improving computational efficiency. Result verification adopted the field measurement comparison method: 5% of the sections were randomly selected, and the spatial distance was measured using a total station and compared with the system's calculation results. An allowable error threshold of 0.1% of the actual distance was set, and a coordinate data verification process was triggered when the error exceeded. All calculation results were stored in a distance matrix table, the table structure of which included fields for the first section number, the second section number, longitude difference, latitude difference, altitude difference, sum of squares, and spatial Euclidean distance. For example, a calculation record shows that the longitude difference between section C and section D is 0.025 degrees, the latitude difference is 0.018 degrees, and the altitude difference is 15.3 meters. After conversion, the square of the longitude difference is 278.9 square meters, the square of the latitude difference is 358.2 square meters, the square of the altitude difference is 234.09 square meters, the sum of the squares is 871.19 square meters, and the final spatial Euclidean distance is 29.517 meters.

[0090] A dynamic compensation mechanism is established during implementation: the weighting coefficient of the altitude difference is automatically adjusted based on the altitude characteristics of the project section. When the project section is located in a plain area (altitude difference less than 100 meters), the square of the altitude difference is multiplied by a correction factor of 0.8; when located in a mountainous area (altitude difference greater than 300 meters), the square of the altitude difference is multiplied by a correction factor of 1.2. This correction factor is determined through statistical analysis of historical data and stored in the parameter configuration table.

[0091] In the specific implementation process, the cosine similarity and the reciprocal of the spatial Euclidean distance are weighted and summed to generate the geological association weights between the contract sections. First, the cosine similarity between the first and second contract sections is obtained. This cosine similarity data comes from the similarity matrix table of a relational database. By querying the joint primary key composed of the first and second contract section numbers, the pre-stored cosine similarity field value in the table is retrieved. The field value is stored as a single-precision floating-point number, with a value range between -1.0 and +1.0. Simultaneously, the spatial Euclidean distance between the first and second contract sections is obtained. This data comes from the spatial distance matrix table. The spatial Euclidean distance field value is queried by the same contract section number. The field value is in meters, and the precision is retained to three decimal places. The data acquisition process implements validity verification: when the cosine similarity value exceeds the theoretical range, it is automatically corrected to the boundary value; when the spatial Euclidean distance value is less than or equal to zero, an anomaly handling process is triggered.

[0092] After obtaining the spatial Euclidean distance, its reciprocal is calculated. The reciprocal calculation employs double-precision floating-point division: the constant 1.0 is divided by the spatial Euclidean distance value. A division-by-zero protection mechanism is implemented: when the spatial Euclidean distance value is less than a preset minimum distance threshold, the spatial Euclidean distance value is forcibly set to that threshold. The minimum distance threshold is set to 0.5 meters based on engineering accuracy requirements to avoid calculation instability caused by microscopic positional fluctuations. The reciprocal calculation result is stored as a temporary variable, representing a spatial proximity metric with units per meter. Numerical stability control is enabled during the calculation process: when the spatial Euclidean distance value is greater than 100 kilometers, a logarithmic spatial transformation technique is used to prevent the loss of significant digits.

[0093] The weighting coefficient for the spatial Euclidean distance is then calculated. This weighting coefficient is an exponential function with the natural constant e as its base. The power of the exponential function is the negative attenuation coefficient multiplied by the spatial Euclidean distance. The attenuation coefficient is determined through historical data analysis, ranging from 0.01 to 0.05, with the specific value dynamically configured based on the geological conditions of the engineering area. The exponential function calculation is implemented using the standard exponential function from a mathematical library, with the input parameter being the product of the negative attenuation coefficient and the spatial Euclidean distance. A numerical range check is performed during the calculation process: if the product value is less than -100, the weighting coefficient is returned as zero; if the product value is greater than zero, the weighting coefficient is returned as 1.0. The weighting coefficient calculation result is rounded to four decimal places and its physical meaning is the distance attenuation factor, which is dimensionless.

[0094] Next, the cosine similarity is multiplied by a fixed weighting coefficient. This fixed weighting coefficient represents the contribution weight of geological feature similarity, ranging from 0.3 to 0.7, with a default value of 0.5. This coefficient is determined through grid search optimization: in typical engineering cases, the impact of different coefficients on prediction accuracy is tested with a step size of 0.05, and the value with the highest accuracy on the validation set is finally selected. The multiplication operation uses double-precision floating-point calculation, and the result is temporarily stored as the first weighting term. Simultaneously, the reciprocal of the spatial Euclidean distance is multiplied by the weighting coefficient of the spatial Euclidean distance to obtain the second weighting term. Overflow detection is implemented for both multiplication operations: when the absolute value of the product exceeds 10 to the power of 10, a high-precision calculation mode is activated.

[0095] After completing the above calculations, the product of cosine similarity multiplied by a fixed weight coefficient is added to the product of the reciprocal of spatial Euclidean distance multiplied by the weight coefficient of spatial Euclidean distance. Before addition, a dimensional consistency check is performed: confirming that the first weighted term is dimensionless and the second weighted term is per meter, the second weighted term is then multiplied by a dimensional conversion factor. The conversion factor is set according to the engineering scale, with a typical value of 1000 (corresponding to a kilometer scale), unifying the two dimensions to dimensionless. The sum is the geological association weight between the first and second sections, theoretically ranging from 0 to positive infinity, but compressed to the 0-1 range using the Sigmoid function in practice. The final result is written into the geological association weight matrix table, which contains a complete chain of fields including the first section number, the second section number, cosine similarity value, spatial Euclidean distance value, reciprocal calculated value, weight coefficient value, first weighted term value, second weighted term value, and geological association weight.

[0096] The implementation process establishes a parameter calibration mechanism: the attenuation coefficient and fixed weight coefficient are updated quarterly. The calibration method involves collecting geological hazard records from the past three months and optimizing the coefficients in reverse to minimize the weight deviation in high-risk areas. GPU acceleration is enabled for computational hardware: tens of thousands of pairs of reference segments are processed in parallel using CUDA kernel functions, and the exponential function calculation is mapped to texture memory to optimize access patterns. The anomaly monitoring system includes: marking a point as an anomaly when the geological association weight exceeds 3.0, automatically checking the quality of the input data; and triggering sparse matrix compression storage when the weight is below 0.01. For example, a calculation result record might be: the cosine similarity of reference segment EF is 0.65, the spatial Euclidean distance is 3250 meters, the reciprocal value is 0.000307, ​​the attenuation coefficient is 0.03, the weight coefficient is 0.381, the fixed weight coefficient is 0.5, the first weighting term is 0.325, the second weighting term is 0.000117, and after dimensional conversion and addition, the geological association weight is 0.325.

[0097] The key parameters were set as follows: the initial value of the attenuation coefficient was determined based on the average spacing between sections, with an attenuation of 0.01 for every 10 kilometers of spacing; the fixed weight coefficient was set based on geological feature dimensions, increasing by 0.1 for every 50 additional feature dimensions. The method for determining the dimension conversion coefficient was: taking the reciprocal of the maximum spacing between project sections multiplied by 1000 as the baseline value. All intermediate variables were recorded in the calculation log, which included audit information such as timestamps, operator IDs, and parameter version numbers.

[0098] In the specific implementation process, the spatial coherence coefficient of the event cluster is generated based on the spatiotemporal coordinates of the surrounding rock microseismic events in each section. First, the spatial coordinates of all surrounding rock microseismic events within the current section are obtained. The surrounding rock microseismic event data comes from a distributed monitoring system, which deploys a four-channel microseismic sensor array every 50 meters behind the tunnel face. The sensors are connected to a local data acquisition instrument via shielded cables. The spatial coordinate analysis method is as follows: when the sensor array detects a microseismic event, the three-dimensional position coordinates of the triggering sensor and the event arrival time difference are recorded; a time-difference positioning algorithm is used to calculate the event source coordinates, with the sensor position coordinates and the propagation speed of sound waves in the rock mass as inputs. The propagation speed is determined based on field calibration tests: 5500 m / s for granite strata and 3800 m / s for shale strata. The obtained spatial coordinates are stored as floating-point tuples containing longitude, latitude, and altitude values. The coordinate system is converted to the local rectangular coordinate system of the project, with the origin at the starting station of the section.

[0099] After obtaining the spatial coordinates, cluster analysis is performed on the spatial coordinates of all surrounding rock microseismic events within the current section. The cluster analysis employs a density-based spatial clustering algorithm, which includes three key stages: The first stage constructs a spatial index, dividing the section space into 1-meter-side cubic grids and establishing a mapping between the grids and events. The second stage determines core events by searching for adjacent events within a 5-meter radius of each event. If the number of adjacent events exceeds a density threshold, it is marked as a core event. The density threshold is dynamically adjusted according to the section size: 15 events are selected for a 100-meter section, and 50 events for a 100-meter section. The third stage expands the clusters by recursively connecting all events with achievable density from the core events to form event clusters. A noise filtering mechanism is implemented during the clustering process: isolated events not assigned to any cluster are directly discarded and do not participate in subsequent calculations.

[0100] After clustering, the centroid positions of all surrounding rock microseismic events within an event cluster are calculated. The centroid calculation uses an arithmetic mean method: three accumulators are initialized to store the sum of longitude, latitude, and altitude values, respectively; each surrounding rock microseismic event within the event cluster is traversed, with its longitude value added to the longitude accumulator, its latitude value to the latitude accumulator, and its altitude value to the altitude accumulator; after traversal, the sum of longitude is divided by the number of events in the event cluster to obtain the centroid longitude value, the sum of latitude is divided by the number of events to obtain the centroid latitude value, and the sum of altitude is divided by the number of events to obtain the centroid altitude value. Weighted corrections are applied for special terrain features: when an event cluster crosses different rock layers, the coordinates are weighted according to the rock mass integrity coefficient, with rock layers having lower coefficients receiving higher weights.

[0101] Subsequently, the spatial Euclidean distance from the spatial coordinates of each microseismic event within the event cluster to the centroid was calculated. The distance calculation employed a three-dimensional linear distance formula. The specific procedure was as follows: For each event within the cluster, the longitude value was subtracted from the centroid's longitude value (longitude difference), the latitude value was subtracted from the centroid's latitude value (latitude difference), and the altitude value was subtracted from the centroid's altitude value (altitude difference). The longitude, latitude, and altitude differences were squared and summed. The square root of the sum was then taken to obtain the spatial Euclidean distance from the single event to the centroid. Batch processing optimization was implemented for the distance calculation: the processor's SIMD instructions were used to simultaneously calculate the squared differences of multiple events, and a lookup table was used to accelerate the square root calculation.

[0102] Next, the average spatial Euclidean distances within the event cluster are calculated. A robust statistical strategy is employed for average calculation: first, all distance values ​​are sorted, and the top 5% of maximum and minimum values ​​are removed; the remaining distance values ​​are summed and then divided by the number of valid events. The summation process uses the Kahan accumulation algorithm to reduce floating-point errors, and the division operation retains six decimal places of precision. An exception handling mechanism is implemented: if there are fewer than three valid events within an event cluster, a preset invalid value is returned and an exception log is recorded.

[0103] The spatial coherence coefficient of an event cluster is ultimately calculated as the reciprocal of the average spatial Euclidean distances within that cluster. A denominator check is performed before the reciprocal calculation: if the average distance is less than the minimum effective distance threshold, the average distance is set as the threshold; this threshold is determined to be 0.1 meters based on the sensor's positioning accuracy. The reciprocal result is multiplied by a scaling factor of 1000 to convert it to a dimensionless value, ensuring the coefficient falls within the commonly used range in engineering. The calculation results are written into a spatial coherence record table, which includes fields such as section number, event cluster number, number of events within the cluster, centroid coordinates, average distance, and spatial coherence coefficient. For example, the calculation results for an event cluster in a granite section are: 38 events within the cluster, average distance 2.35 meters, and spatial coherence coefficient 425.53.

[0104] A quality control system was established during implementation: Positioning accuracy was calibrated quarterly using a standard test ball. The test ball was used to generate acoustic signals at known locations, verifying that the system's positioning error was less than 0.3 meters. Clustering parameter optimization method: Clustering quality was evaluated using the silhouette coefficient, and the search radius was automatically adjusted to ensure the silhouette coefficient was greater than 0.6. All computing modules were deployed on edge computing nodes, equipped with FPGA accelerators to process 2000 micro-seismic events per second in real time.

[0105] In the specific implementation process, when the spatial coherence coefficient is lower than the preset threshold, geological risk isolation and global model aggregation operations are performed. First, the spatial coherence coefficient of the event cluster in the current section is obtained. This coefficient comes from the calculation result field of the spatial coherence record table, and the spatial coherence coefficient value in the latest monitoring period is retrieved by section number. The preset threshold is determined by collecting spatial coherence coefficient data of rock mass instability cases in historical projects before they occurred, and taking the fifth percentile value as the initial threshold; then, it is corrected according to the rock mass type of the current project, increasing the threshold by 20% for granite strata and decreasing the threshold by 15% for soft rock strata. The comparison operation uses double-precision floating-point comparison, and when the spatial coherence coefficient value is less than the preset threshold, the isolation flag is updated.

[0106] Once the isolation condition is triggered, the geological risk feature vectors of the corresponding section are prohibited from participating in the aggregation operation. The prohibition mechanism is implemented through a status flag: an isolation status field is created in the section status table, and this field is set to a prohibited state when the isolation flag is true. The system adds status verification logic to the subsequent data access layer; any aggregation operation request must check the section status field, and if it is in a prohibited state, the geological risk feature vector data will be refused. The isolation range control strategy is as follows: permanent isolation is implemented when the spatial coherence coefficient of a section is below a threshold for three consecutive monitoring periods; if it exceeds the threshold in a single instance, temporary isolation is implemented, and it is automatically lifted after the coefficient recovers to 120% of the threshold.

[0107] After completing the isolation settings, the local risk prediction model parameters corresponding to the geological risk feature vectors of all unisolated sections are retrieved from the stored mapping relationships. The retrieval operation undergoes a multi-level verification process: first, the set of section numbers in the section status table that are not in a prohibited state is filtered; then, the corresponding parameter file is retrieved from the model parameter library based on the section number; the parameter file is stored in Protobuf binary format and contains the weight matrix and bias vector of the fully connected layers. After file parsing, it is reconstructed into a parameter tensor data structure, with the tensor dimensions strictly matching the model architecture. A version management mechanism is implemented: only the parameter version updated in the most recent model iteration is retrieved to avoid interference from historical data.

[0108] Subsequently, the geological association weights between all pairs of unisolated sections were obtained. The weight data originated from the geological association weight matrix table, and a submatrix was generated using the set of unisolated section numbers. The submatrix construction employed sparse storage optimization: weight values ​​less than 0.05 were not loaded, reducing memory usage. Data preprocessing included: symmetry checks to ensure matrix symmetry; missing values ​​were filled using the moving average of adjacent section weights. After loading the weight matrix, a fast index was created, with both row and column indices corresponding to the unisolated section number sequence.

[0109] Next, the weighted average of the local risk prediction model parameters for all unisolated sections is calculated. The weight allocation rule is: the aggregate weight of each section is equal to the sum of all geological association weights related to that section in the weight matrix. The specific calculation process consists of three steps: First, the normalization coefficient is calculated by traversing each unisolated section and summing all weight values ​​in the row vector corresponding to that section to obtain the total weight of that section; second, a weight mapping table is established, with the section number as the key and the total weight floating-point number as the value; third, parameter aggregation is performed, and the parameters of each layer of the model are processed independently: an all-zero target tensor is initialized; each unisolated section is traversed, and the parameter tensor of that section is multiplied by the corresponding total weight value in the mapping table and then accumulated to the target tensor; finally, the target tensor is divided by the sum of the total weights of all sections. The aggregation process implements distributed computing optimization: the model parameters are divided into layers, and the parameters of different layers are processed in parallel by different computing nodes.

[0110] The weighted average value is ultimately used as the parameter for the global geological risk prediction model. Parameter injection employs an incremental update strategy: 30% of the previous cycle's global model parameters are retained, while the remaining 70% are new. The model structure remains fixed, and integrity checks are performed after parameter updates: verifying that parameter values ​​are within physically reasonable ranges and confirming that the model output dimensions are consistent with design specifications. The updated global model parameters are persistently stored in the central model repository, and a version snapshot is generated simultaneously. The system automatically pushes the new parameters to the edge servers of each contract section, and the edge nodes receive and load them into their local risk prediction models to complete the update.

[0111] A closed-loop verification system was established during implementation: the global model performance was evaluated monthly using an independent test set, and a rollback to the previous version was performed when the accuracy dropped by more than 5%. An isolation decision audit mechanism recorded all segment isolation events, including trigger coefficient values, threshold setting basis, and release conditions. For example, an update record showed that segment G was isolated because its spatial coherence coefficient of 182.4 was lower than the threshold of 200; the remaining 8 unisolated segments participated in aggregation; the total weight of segment H was calculated to be 3.85 (1.2 from segment J, 0.8 from segment K, and 1.85 from segment L); the bias vector of the first layer of the fully connected layer was updated from [-0.12, 0.34] to [-0.08, 0.29] after parameter aggregation.

[0112] Key parameter settings are based on: historical data showing that when the spatial coherence coefficient is below 200, the probability of rock mass instability increases fivefold; weight normalization avoids small sample segments dominating the global model. The system is deployed on a Kubernetes container cloud platform, and the parameter aggregation service is configured for automatic elastic scaling, supporting the processing of aggregation requests for 50 segments per second.

[0113] Example 2: Figure 2A schematic diagram of a federated learning prediction system for geological risks in multi-section tunnels according to the present invention is provided. This federated learning prediction system for geological risks in multi-section tunnels includes the following modules:

[0114] The geological acquisition module is used to obtain the geological risk feature vectors output by the local risk prediction model for each section and the corresponding three-dimensional geological coordinates of the section.

[0115] The similarity calculation module is used to calculate the cosine similarity between the geological risk feature vectors of any two sections.

[0116] The spatial distance module is used to calculate the spatial Euclidean distance between corresponding sections based on three-dimensional geological coordinates.

[0117] The weight generation module is used to generate geological association weights between sections by weighted summation of cosine similarity and the reciprocal of spatial Euclidean distance. The weight coefficient of spatial Euclidean distance decreases exponentially as spatial Euclidean distance increases.

[0118] The transmission processing module is used to generate the spatial coherence coefficient of the event cluster based on the spatiotemporal coordinates of the surrounding rock microseismic events in each section;

[0119] The model aggregation module is used to isolate the geological risk feature vector of the corresponding section when the spatial coherence coefficient is lower than a preset threshold, and to use geological correlation weights to weight and aggregate the local risk prediction model parameters of the unisolated section to generate a global geological risk prediction model.

[0120] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0121] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0122] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0123] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0126] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A federated learning method for predicting geological risks in multi-section tunnels, characterized in that, Includes the following steps: S1. Obtain the geological risk feature vectors output by the local risk prediction model for each section and the corresponding three-dimensional geological coordinates of the section. S2. Calculate the cosine similarity between the geological risk feature vectors of any two sections; S3. Calculate the spatial Euclidean distance between corresponding sections based on the three-dimensional geological coordinates; S4. Generate the geological association weights between the sections by weighted summation of cosine similarity and the reciprocal of spatial Euclidean distance; S5. Based on the spatiotemporal coordinates of acoustic emission events in the surrounding rock of each section, generate the spatial coherence coefficients of event clusters, including: Obtain the spatial coordinates of all acoustic emission events in the surrounding rock within the current section; Cluster analysis is performed on the spatial coordinates of all acoustic emission events in the current section to form at least one event cluster. Calculate the centroid position of the spatial coordinates of all acoustic emission events in the surrounding rock within the event cluster; Calculate the spatial Euclidean distance from the spatial coordinates of each acoustic emission event in the surrounding rock within the event cluster to the centroid location; Calculate the average of all spatial Euclidean distances within the event cluster; The reciprocal of the average spatial Euclidean distance within an event cluster is taken as the spatial coherence coefficient of the event cluster; S6. When the spatial coherence coefficient is lower than a preset threshold, the geological risk feature vector of the corresponding section is isolated, and the parameters of the local risk prediction model of the unisolated section are weighted and aggregated using geological correlation weights to generate a global geological risk prediction model, including: Obtain the spatial coherence coefficient of the event cluster in the current segment, and compare the spatial coherence coefficient with a preset threshold. When the spatial coherence coefficient is less than a preset threshold, the geological risk feature vector of the corresponding section is prohibited from participating in the aggregation operation; Obtain the local risk prediction model parameters corresponding to the geological risk feature vectors of all unisolated sections from the stored mapping relationship; Obtain the geological correlation weights between all pairs of unisolated sections; Calculate the weighted average of the local risk prediction model parameters for all unisolated sections, where the weight of each section is the sum of the geological association weights of all sections that have geological association weights with the corresponding section; The weighted average value is used as a parameter in the global geological risk prediction model.

2. The federated learning prediction method for geological risks of multi-section tunnels according to claim 1, characterized in that, Obtain the geological risk feature vectors output by the local risk prediction model for each contract section and the corresponding three-dimensional geological coordinates of the contract section, including: The raw geological data is processed by the local risk prediction model deployed in each section, and a geological risk feature vector is output. The longitude, latitude, and altitude values ​​of the center point of each section are extracted from the engineering geological survey report to form a three-dimensional geological coordinate system. Establish and store the mapping relationship between geological risk feature vectors and three-dimensional geological coordinates according to the section number.

3. The federated learning prediction method for geological risks of multi-section tunnels according to claim 2, characterized in that, Calculate the cosine similarity between the geological risk feature vectors of any two contract sections, including: Obtain the geological risk feature vectors of the first and second sections from the stored mapping relationships; Calculate the dot product of the geological risk feature vector of the first section and the geological risk feature vector of the second section; Calculate the magnitude of the geological risk characteristic vector for the first contract section; Calculate the magnitude of the geological risk characteristic vector for the second section; Dividing the dot product by the product of the magnitudes of the geological risk feature vectors of the first and second sections yields the cosine similarity between the geological risk feature vectors of the two sections.

4. The federated learning prediction method for geological risks of multi-section tunnels according to claim 3, characterized in that, The spatial Euclidean distance between corresponding sections is calculated based on three-dimensional geological coordinates, including: Obtain the three-dimensional geological coordinates of the first section and the two sections from the stored mapping relationship; Calculate the difference between the longitude values ​​in the three-dimensional geological coordinates of the first section and the longitude values ​​in the three-dimensional geological coordinates of the second section; Calculate the difference between the latitude value in the three-dimensional geological coordinates of the first section and the latitude value in the three-dimensional geological coordinates of the second section; Calculate the difference between the elevation values ​​in the three-dimensional geological coordinates of the first section and the elevation values ​​in the three-dimensional geological coordinates of the second section; Add the squares of the longitude difference, the latitude difference, and the altitude difference; Take the square root of the sum to obtain the spatial Euclidean distance between the two sections.

5. The federated learning prediction method for geological risks of multi-section tunnels according to claim 4, characterized in that, The geological association weights between sections are generated by weighting and summing the cosine similarity with the reciprocal of the spatial Euclidean distance, including: Obtain the cosine similarity between the first and second contract segments, and obtain the spatial Euclidean distance between the first and second contract segments; Calculate the reciprocal of the Euclidean distance in space; Calculate the weighting coefficients for spatial Euclidean distance; Multiply the cosine similarity by a fixed weight coefficient; Multiply the reciprocal of the spatial Euclidean distance by the weighting coefficient of the spatial Euclidean distance; The geological association weight between the first and second sections is generated by adding the product of the cosine similarity multiplied by the fixed weight coefficient and the product of the inverse of the spatial Euclidean distance multiplied by the weight coefficient of the spatial Euclidean distance.

6. The federated learning prediction method for geological risks of multi-section tunnels according to claim 5, characterized in that, The weighting coefficient of spatial Euclidean distance is an exponential function with the natural constant e as the base, and the power of the exponential function is a negative decay coefficient multiplied by the spatial Euclidean distance.

7. The federated learning prediction method for geological risks of multi-section tunnels according to claim 5, characterized in that, The weighting coefficient of spatial Euclidean distance decreases exponentially as spatial Euclidean distance increases.

8. A federated learning prediction system for geological risks of multi-section tunnels, used to implement the federated learning prediction method for geological risks of multi-section tunnels as described in any one of claims 1-7, characterized in that, Includes the following modules: The geological acquisition module is used to obtain the geological risk feature vectors output by the local risk prediction model for each section and the corresponding three-dimensional geological coordinates of the section. The similarity calculation module is used to calculate the cosine similarity between the geological risk feature vectors of any two sections. The spatial distance module is used to calculate the spatial Euclidean distance between corresponding sections based on three-dimensional geological coordinates. The weight generation module is used to generate geological association weights between sections by weighted summation of cosine similarity and the reciprocal of spatial Euclidean distance. The weight coefficient of spatial Euclidean distance decreases exponentially as spatial Euclidean distance increases. The emission processing module is used to generate the spatial coherence coefficient of the event cluster based on the spatiotemporal coordinates of acoustic emission events in the surrounding rock of each section; The model aggregation module is used to isolate the geological risk feature vector of the corresponding section when the spatial coherence coefficient is lower than a preset threshold, and to use geological correlation weights to weight and aggregate the local risk prediction model parameters of the unisolated section to generate a global geological risk prediction model.

Images