Foundation processing depth intelligent recommendation method based on neural network

Abstract

The invention discloses a foundation processing depth intelligent recommendation method based on a neural network, and the method comprises the following steps: collecting goaf region data, and carrying out the preprocessing of the data to generate a geological structure topological graph; executing depth-first traversal, constructing a structure path, extracting node structure parameters, and generating a structure path feature tensor set; inputting the feature tensor set into the improved Tree-LSTM network into which the lithologic modulation vector is introduced, and executing structure state updating calculation; a lithology modulation vector is introduced into the forgetting gate, and differential modulation is carried out on the memory state of the child nodes; calculating a settlement rate and an acceleration, constructing a stability score and taking the score as an attention weight fusion hidden state; generating a stability focusing representation, and outputting a processing thickness and a mode number by an input processing suggestion generation unit; and binding a space coordinate generation structure to process a suggestion data record set. According to the invention, foundation processing intelligent matching and structure differentiation identification are realized, and the evaluation precision and the processing suggestion pertinence are improved.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering analysis technology, and in particular to a method for intelligent recommendation of foundation treatment depth based on neural networks. Background Technology

[0002] Current research on stability analysis and foundation treatment measures for goaf sites has formed a technical system mainly based on mechanical analysis, numerical simulation, and fuzzy comprehensive evaluation methods, which are widely used in the site selection and construction of infrastructure such as power transmission and transformation projects, photovoltaic power stations, and high-speed transportation. Traditional methods mostly rely on static analysis of single or combined factors such as geology, mining, and structural loads. Although settlement prediction and stratum response simulation can be achieved through Knothe influence function, probability integral method, or FLAC3D numerical model, there are still problems such as abstract simplification and insufficient characterization of response mechanisms in structural path identification, overburden evolution process modeling, and stress transfer mechanism expression. Especially in the context of complex structural evolution with multiple paths and multiple sections, existing models cannot fully integrate the spatiotemporal continuity, lithological heterogeneity, and response sensitivity characteristics in the structural path of the goaf, and lack the ability to express the stability of key structural points with high resolution. In addition, treatment suggestions are mostly derived from engineering experience templates, lacking a structural state-driven intelligent recommendation mechanism, and cannot achieve differentiated treatment responses for different structural states along different paths. Existing technologies have not yet established a complete technical process for joint modeling of structural topology and geological lithology, and for achieving path state perception and ground treatment suggestion linkage output through neural networks. Therefore, how to provide a deep intelligent recommendation method for ground treatment based on neural networks is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0003] One objective of this invention is to propose a neural network-based intelligent recommendation method for foundation processing depth. This invention fully utilizes geological structure topology modeling and lithological parameter modulation mechanisms to construct a detailed process for updating structural path states and generating processing suggestions. It has the advantages of high recognition accuracy, strong processing diversity, and strong adaptability to complex geological structures.

[0004] The ground-based intelligent recommendation method for ground treatment based on neural networks according to embodiments of the present invention includes the following steps:

[0005] Collect data from the goaf area and preprocess it to generate a geological structure topology map;

[0006] A depth-first traversal is performed on the geological structure topology map to construct several structural paths. The structural parameters corresponding to each node in each path are arranged in the path order to form a set of structural path feature tensors.

[0007] The set of structural path feature tensors is input into an improved Tree-LSTM network that incorporates lithology modulation vectors for structural state update calculation. Specifically, a lithology modulation vector consisting of rock type, porosity, and water content is added during the forget gate calculation process to differentially modulate the memory state of each sub-node.

[0008] The settlement rate and settlement acceleration of each node in each structural path are standardized and calculated to construct a node stability score. During the path aggregation stage, attention weighting coefficients are calculated based on the node stability scores, and the hidden states of all child nodes are weighted and fused to generate a stability-focused representation of the current path.

[0009] The stability focus representation is input into the treatment suggestion generation unit, multi-level linear mapping and parameter regression operations are performed, and the foundation treatment thickness value and treatment method number of the corresponding path are output.

[0010] Bind the processing thickness value and the processing method number to the spatial coordinate label corresponding to the path to generate a structural processing suggestion data record;

[0011] Output a set of data records suggesting structural processing, and complete the identification of site stability and recommendation of foundation treatment in the goaf area.

[0012] Optionally, the data collected from the goaf area includes coal pillar size, coal seam depth, roof lithology, fault distribution, and settlement monitoring sequence. The preprocessing process includes performing spatial coordinate registration, numerical normalization, and time series completion operations on various structural parameters to generate a geological structure topology map.

[0013] Optionally, the process of generating the set of structural path feature tensors includes the following steps:

[0014] A depth-first traversal operation is performed on the geological structure topology map. The starting point of the traversal is selected as the node located at the boundary between the surface and the main mining area. During the traversal, the numbers of all nodes passed in the path are recorded, and the structural path set is established in the order of access.

[0015] For all nodes in each structural path, extract the structural parameters corresponding to the nodes in the path order. The structural parameters include coal pillar size, burial depth, roof lithology code, fault boundary distance, and settlement monitoring point number corresponding to the node.

[0016] Among them, the lithology code of the roof is mapped to a fixed-dimensional lithology category vector using the unique thermal coding method, and the fault boundary distance is calculated numerically as the shortest spatial distance to the center point of the structural path and normalized to a floating-point value in the range of 0 to 1.

[0017] The various structural parameters of a node are concatenated along the same node dimension to form a node feature vector. All node feature vectors are arranged sequentially along the structural path to form the input tensor of a structural path. After performing the above steps on all structural paths, the input tensors of all structural paths are summarized to form a set of structural path feature tensors.

[0018] During tensor construction, zero-padding is performed on structural paths whose length is less than the preset minimum number of nodes, and truncation is performed on structural paths whose length exceeds the upper limit. All structural path tensors have the same dimension.

[0019] Optionally, the process of inputting the set of structural path feature tensors into the improved Tree-LSTM network for structural state update calculation includes:

[0020] For each sub-node in each structural path, extract the rock layer type, porosity, and water content corresponding to the sub-node. Convert the rock layer type into a fixed-length category vector using a unique thermal encoding method. Convert the porosity and water content into floating-point values ​​by normalizing the maximum and minimum values ​​respectively. Concatenate the three in order to form a lithology modulation vector. Concatenate the lithology modulation vector with the hidden state vector of the sub-node according to the dimension to form a new extended node vector input.

[0021] When performing the forget gate calculation, the extended node vector is called as input, and the retention coefficient of the child node in the current propagation cycle is generated by weighting it with the current input vector of the parent node.

[0022] The calculation process of the retention coefficient is directly affected by the values ​​of each element in the lithology modulation vector. Different rock strata, porosity levels and water content will change the value of the retention coefficient, thereby changing the amount of structural state information transmitted by the sub-node in the path.

[0023] Subsequently, element-wise multiplication is performed using the retention coefficient and the memory state vector of the child nodes to obtain the weighted memory state result under lithological modulation. The weighted memory state results of all child nodes are summed element-wise and used as the update value of the memory state vector of the parent node, thus completing the progressive propagation of the structural state.

[0024] Optionally, the process of generating the stability focus representation of the current path includes:

[0025] For each structural node in each structural path, extract the settlement monitoring time series corresponding to the node, calculate the difference between adjacent time steps to obtain the settlement rate, and calculate the settlement acceleration based on the rate change.

[0026] The settling velocity and settling acceleration of each node are normalized to the interval between 0 and 1. A stability score is constructed using a weighted average function. The stability score of each child node is used as the attention input weight and multiplied element-wise with the hidden state vector of that child node to generate a weighted state vector. The weighted state vectors of all child nodes are accumulated and fused according to their dimensions to generate an aggregated state representation. The aggregated state representation is used as the stability focus representation of the current path to express the dominant influence of unstable structures on the path state. It also participates in the propagation of the improved Tree-LSTM network as the hidden state of the parent node.

[0027] Optionally, the processing suggestion generation unit includes the following components:

[0028] The input interface layer is used to receive the stability focus representation corresponding to the structure path;

[0029] The feature transformation subunit includes at least two layers of linear mapping structure, which are used to perform dimensional transformation and nonlinear activation processing on the input features and output intermediate feature representations for prediction.

[0030] The thickness regression sub-unit is used to perform parametric regression operations based on intermediate feature representations to generate predicted values ​​for the foundation treatment thickness of the structure path.

[0031] The processing method classification subunit is used to perform multi-class classification operations based on intermediate feature representations and outputs the processing method type number.

[0032] The output combination layer is used to combine the predicted thickness value and the processing method type number to form a processing suggestion vector, which is then bound to the corresponding structure path label to output the structure processing suggestion record.

[0033] Optionally, the process of outputting the foundation treatment thickness value and treatment method number for the corresponding path includes:

[0034] The input interface layer receives the stability-focused representation corresponding to the structural path, which serves as the model input feature.

[0035] The feature transformation subunit first feeds the input vector into the first linear mapping layer, performs a linear affine transformation with a bias term, completes the activation through the ReLU activation function, and then feeds it into the second linear mapping layer to compress the dimension and generate an intermediate feature representation;

[0036] The thickness regression sub-unit receives intermediate feature representations, calls a linear regression function to generate floating-point predicted values, which are used as the foundation treatment thickness corresponding to the structural path.

[0037] The processing method classification subunit receives the same intermediate feature representation, performs a linear mapping and then a softmax function, outputs the probability distribution corresponding to each processing method, and selects the category index corresponding to the maximum probability as the processing method number of the structure path.

[0038] Optionally, the structural treatment suggestion data record includes spatial coordinate labels of the structural path, predicted values ​​of foundation treatment thickness, and treatment method number. This set of structural treatment suggestion data records is used to complete the identification of site stability in the goaf and to recommend foundation treatment suggestions, specifically including:

[0039] The spatial coordinate labels of each structural path are bound to the corresponding predicted values ​​of foundation treatment thickness and treatment method number to form a structural treatment suggestion data record. All suggestions are recorded on the geological structure topology map for spatial mapping.

[0040] Based on the range of processed thickness values, the paths are divided into three categories of path records: those with thickness less than the set threshold, those in the middle range, and those exceeding the threshold.

[0041] Based on the treatment method number identifier, match the corresponding treatment process parameter table for each path, including grouting parameters, filling material type and interlayer spacing requirements;

[0042] For paths with a thickness less than the threshold, it is recommended to set up a single borehole grouting point in the corresponding area of ​​the path, use a low-pressure injection method, and the injection depth should not exceed the bottom interface of the path control layer.

[0043] For paths with thickness in the middle range, it is recommended to perform segmented grouting, set two or more grouting hole positions, control the grouting spacing to be no more than 3 meters, adopt medium-pressure quantitative injection method, and distribute the cumulative grouting volume linearly according to the treatment thickness.

[0044] For paths where the thickness exceeds the upper limit threshold, it is recommended to perform layer-by-layer filling operations in the target area, using a layered compaction method of crushed stone and silty clay, controlling the thickness of each layer to not exceed 30 cm, and conducting a standard penetration test after each layer is filled to confirm the bearing capacity. After passing the test, the next layer of construction can continue.

[0045] For records with path numbers indicating combined treatment methods, a combined grouting and filling process is performed, with grouting done first and then filling. Settlement monitoring is conducted after the first phase of construction is completed, and subsequent layer construction begins once the settlement stabilizes.

[0046] Finally, all specific operational parameters are bound to the corresponding path spatial coordinates to form a set of structural treatment suggestion data records, thus completing the stability identification of the goaf area and the recommendation of foundation treatment based on structural feature differences.

[0047] The beneficial effects of this invention are:

[0048] (1) This invention constructs a set of structural path feature tensors and introduces an improved Tree-LSTM network. It combines the lithological modulation vector composed of rock layer type, porosity and water content to perform differentiated modeling of structural state, improve the stability identification accuracy under complex geological conditions of multiple paths, and effectively solve the problem of insufficient path structure response capability in traditional methods.

[0049] (2) In the path aggregation stage, the present invention introduces a stability scoring mechanism based on settlement rate and acceleration, and combines an attention weighting strategy to express the path status in a high-risk focus, thereby enhancing the ability to locate areas of settlement anomalies and improving the risk sensitivity and regional difference expression ability of structural path status representation.

[0050] (3) The present invention establishes a processing suggestion generation unit, which focuses on the stability of the output of the structural path state and intelligently generates processing thickness and processing method suggestions through linear mapping and parameter regression. This realizes the linkage recommendation of processing strategy and structural state, which has high adaptability and engineering practical value, and avoids the problems of template-based and delayed response of traditional processing methods. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is an overall flowchart of the neural network-based ground-based intelligent recommendation method proposed in this invention.

[0053] Figure 2 This is a schematic diagram of the structure of the improved Tree-LSTM network proposed in this invention, which injects the lithology modulation vector into a forget gate and performs differential modulation. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] refer to Figure 1 and Figure 2 A neural network-based ground-based intelligent recommendation method includes the following steps:

[0056] Collect data from the goaf area and preprocess it to generate a geological structure topology map;

[0057] A depth-first traversal is performed on the geological structure topology map to construct several structural paths. The structural parameters corresponding to each node in each path are arranged in the path order to form a set of structural path feature tensors.

[0058] The set of structural path feature tensors is input into an improved Tree-LSTM network that incorporates lithology modulation vectors for structural state update calculation. Specifically, a lithology modulation vector consisting of rock type, porosity, and water content is added during the forget gate calculation process to differentially modulate the memory state of each sub-node.

[0059] The settlement rate and settlement acceleration of each node in each structural path are standardized and calculated to construct a node stability score. During the path aggregation stage, attention weighting coefficients are calculated based on the node stability scores, and the hidden states of all child nodes are weighted and fused to generate a stability-focused representation of the current path.

[0060] The stability focus representation is input into the treatment suggestion generation unit, multi-level linear mapping and parameter regression operations are performed, and the foundation treatment thickness value and treatment method number of the corresponding path are output.

[0061] Bind the processing thickness value and the processing method number to the spatial coordinate label corresponding to the path to generate a structural processing suggestion data record;

[0062] Output a set of data records suggesting structural processing, and complete the identification of site stability and recommendation of foundation treatment in the goaf area.

[0063] In this embodiment, the data collected in the goaf area includes coal pillar size, coal seam depth, roof lithology, fault distribution, and settlement monitoring sequence. The preprocessing process includes performing spatial coordinate registration, numerical normalization, and time series completion operations on various structural parameters to generate a geological structure topology map.

[0064] In this embodiment, the process of generating the structural path feature tensor set includes the following steps:

[0065] A depth-first traversal operation is performed on the geological structure topology map. The starting point of the traversal is selected as the node located at the boundary between the surface and the main mining area. During the traversal, the numbers of all nodes passed in the path are recorded, and the structural path set is established in the order of access.

[0066] For all nodes in each structural path, extract the structural parameters corresponding to the nodes in the path order. The structural parameters include coal pillar size, burial depth, roof lithology code, fault boundary distance, and settlement monitoring point number corresponding to the node.

[0067] Specifically, the roof lithology is mapped to a fixed-dimensional lithology category vector using a one-hot encoding method. The fault boundary distance is calculated numerically as the shortest spatial distance to the center point of the structural path and normalized to a floating-point value between 0 and 1. The one-hot encoding method maps roof lithology categories to a discrete vector representation with a fixed dimension. Assuming there are n possible lithology types, each lithology category is assigned a binary vector of length n, where only the position corresponding to the current lithology category is 1, and all other positions are 0. In this way, each lithology has a unique identifier in the vector space, avoiding interference from the size order of categories on the model learning process. The encoded lithology category vector can be directly spliced ​​or embedded with other continuous geological parameters in the same dimensional space for neural network feature fusion and state modulation operations.

[0068] The various structural parameters of a node are concatenated along the same node dimension to form a node feature vector. All node feature vectors are arranged sequentially along the structural path to form the input tensor of a structural path. After performing the above steps on all structural paths, the input tensors of all structural paths are summarized to form a set of structural path feature tensors.

[0069] During tensor construction, zero-padding is performed on structural paths whose length is less than the preset minimum number of nodes, and truncation is performed on structural paths whose length exceeds the upper limit. All structural path tensors have the same dimension. Specifically:

[0070] In the construction of the structural path feature tensor, a dynamic standardization strategy with classification judgment is adopted for structural paths of different lengths. Let the path tensor be... Among them, l i Let d be the number of nodes in the structural path, d be the feature dimension of a single node, and let L be the minimum node count threshold. min The maximum number of nodes is L. max When the path length satisfies l i <L min At that time, zero-vector padding is performed to construct a tensor:

[0071] Make

[0072] The padded zero vectors are masked in subsequent attention calculations and do not participate in weight calculations or state updates, when the path length satisfies l. i >L max At that time, perform a tail truncation operation, retaining only the first L. max Construct a tensor based on the features corresponding to each structural node:

[0073] Make

[0074] This classification strategy, while maintaining the uniformity of the input tensor dimensions, creatively combines the semantic structure of geological paths for truncation and completion, avoiding the introduction of false structures or the loss of key structural information due to uneven path lengths, thereby improving the robustness and semantic integrity of structural path inputs in the improved Tree-LSTM model.

[0075] The path construction method is based on structural topology logic and does not rely on absolute spatial grids, thereby ensuring that the structural information carried by each path has a clear "mining structure evolution chain" attribute. This makes it easier for the improved Tree-LSTM to perform targeted modeling of the path structure state.

[0076] In this embodiment, the process of inputting the set of structural path feature tensors into the improved Tree-LSTM network for structural state update calculation includes:

[0077] For each sub-node in each structural path, the rock layer category, porosity, and water content corresponding to the sub-node are extracted. The rock layer category is converted into a fixed-length category vector through one-hot encoding. The porosity and water content are converted into floating-point values ​​through maximum and minimum value normalization. The three are concatenated in sequence to form a lithology modulation vector. The lithology modulation vector is concatenated with the hidden state vector of the sub-node by dimension to form a new extended node vector input. The hidden state vector is the output state vector of the sub-node at the current moment in the structural path calculation process, which is generated by the structural path feature tensor and gating calculation.

[0078] When performing the forget gate calculation, the extended node vector is called as input. By weighting it with the current input vector of the parent node, the retention coefficient of the child node in the current propagation cycle is generated. The retention coefficient controls the strength of the influence of the child node on the memory state of the parent node.

[0079] The calculation process of the retention coefficient is directly affected by the values ​​of each element in the lithology modulation vector. Different rock strata types, porosity levels, and water content will change the value of the retention coefficient, thereby changing the amount of structural state information transmitted by the sub-node in the path; specifically, the impact is as follows:

[0080] First, after converting the rock strata category into a unique heat code, it is embedded into the gating computation vector. Different rock strata codes are linearly combined with the gating weight matrix to enhance the neuron response corresponding to hard rock strata, thereby obtaining a higher retention output in the forget gate computation. The child nodes of hard rock strata are more likely to retain their state, while soft rock strata are more likely to be forgotten or weakened.

[0081] Second, porosity, as a continuous value, is normalized and then used as a gating input. The higher the porosity, the looser the structure, and the lower the retention coefficient is assigned to it by the system. In actual calculation, the system forms a negative correlation mapping with the weight terms of the gating weight matrix, so that nodes with high porosity output a lower retention value, and the information weight is weakened in path aggregation.

[0082] Third, the water content, after being normalized, is used as the third dimension input, indicating that the rock mass at this node may be in a state of water softening or swelling. The system provides "inhibitory adjustment" to the high water content input according to preset parameters, that is, adding a penalty term to the gated neuron to further reduce its retention probability.

[0083] Subsequently, element-wise multiplication is performed using the retention coefficient and the memory state vector of the child nodes to obtain the weighted memory state result under lithological modulation. The weighted memory state results of all child nodes are summed element-wise and used as the update value of the memory state vector of the parent node, thus completing the progressive propagation of the structural state.

[0084] This step proposes constructing vectors from lithological parameters and directly injecting them into the neural network structure update calculation process. This allows lithological information to no longer be merely a model input feature, but rather an internal regulating factor influencing the memory propagation path. For the first time, this achieves a deep integration of coupled control of geological attributes and propagation mechanisms in structural modeling, thereby significantly improving the model's responsiveness and prediction accuracy to differences in structural states across different lithological regions. The differentiated modulation operation is implemented through "node feature vectors, input expansion, gated neuron transformation, retention coefficients, and memory state weighting".

[0085] In this embodiment, the process of generating the stability focus representation of the current path includes:

[0086] For each structural node in each structural path, extract the settlement monitoring time series corresponding to the node, calculate the difference between adjacent time steps to obtain the settlement rate, and calculate the settlement acceleration based on the rate change.

[0087] The settlement rate and settlement acceleration of each node are normalized to the interval between 0 and 1. A stability score is constructed using a weighted average function, in which the weight of settlement acceleration is greater than that of settlement rate to enhance the responsiveness to abrupt changes. Specifically, the normalized result of settlement rate is set as v, the normalized result of settlement acceleration is set as a, and the weighting coefficient w is selected. v with w a , making w v +w a =1 and w a >w v This leads to an evaluation strategy that is more sensitive to acceleration, using the following weighted average method to calculate the node stability score s:

[0088] s = w v ·v+w a ·a;

[0089] The weight can be set as w. v =0.4, w a =0.6 indicates that a higher weight is given to the sudden increase trend of acceleration change, so as to enhance the model's ability to pay attention to the abnormal regions of structural response. The score s is the stability score of this node, which participates in the path aggregation process as an attention weighting coefficient.

[0090] The stability score of each child node is used as its attention input weight, and element-wise multiplication is performed with the hidden state vector of that child node to generate a weighted state vector. The weighted state vectors of all child nodes are accumulated and fused according to their dimensions to generate an aggregated state representation. The aggregated state representation is used as the stability focus representation of the current path to express the dominant influence of unstable structures on the path state in the whole path. It participates in the propagation of the improved Tree-LSTM network as the hidden state of the parent node.

[0091] The operation process involves incorporating the stability change characteristics of nodes into the structural state propagation weight calculation, so that substructures with different response intensities contribute unequally to the overall path state aggregation, thereby enabling state modeling focused on nodes with high settlement risk and improving the sensitivity of path structure representation to the identification of unstable regions.

[0092] In this embodiment, the stability focus representation is input to the processing suggestion generation unit, multi-layer linear mapping and parameter regression operations are performed, and the foundation treatment thickness value and processing method number of the corresponding path are output.

[0093] The processing suggestion generation unit includes the following components:

[0094] The input interface layer is used to receive the stability focus representation corresponding to the structure path;

[0095] The feature transformation subunit includes at least two layers of linear mapping structure, which are used to perform dimensional transformation and nonlinear activation processing on the input features and output intermediate feature representations for prediction.

[0096] The thickness regression sub-unit is used to perform parametric regression operations based on intermediate feature representations to generate predicted values ​​for the foundation treatment thickness of the structure path.

[0097] The processing method classification subunit is used to perform multi-class classification operations based on intermediate feature representations and outputs the processing method type number.

[0098] The output combination layer is used to combine the predicted thickness value and the processing method type number to form a processing suggestion vector, which is then bound to the corresponding structure path label to output the structure processing suggestion record.

[0099] In this embodiment, the output process of the foundation treatment thickness value and treatment method number for the corresponding path includes:

[0100] The input interface layer receives the stability-focused representation corresponding to the structural path, which serves as the model input feature.

[0101] The feature transformation subunit first feeds the input vector into the first linear mapping layer, performs a linear affine transformation with a bias term, completes the activation through the ReLU activation function, and then feeds it into the second linear mapping layer to compress the dimension and generate an intermediate feature representation;

[0102] The thickness regression sub-unit receives intermediate feature representations, calls a linear regression function to generate floating-point predicted values, which are used as the foundation treatment thickness corresponding to the structural path.

[0103] The processing method classification subunit receives the same intermediate feature representation, performs a linear mapping and then a softmax function, outputs the probability distribution corresponding to each processing method, and selects the category index corresponding to the maximum probability as the processing method number of the structure path.

[0104] In this embodiment, the structural treatment suggestion data record includes spatial coordinate labels of the structural path, predicted values ​​of foundation treatment thickness, and treatment method number. The structural treatment suggestion data record set is used to complete the identification of site stability in the goaf and to recommend foundation treatment suggestions, specifically including:

[0105] The spatial coordinate labels of each structural path are bound to the corresponding predicted values ​​of foundation treatment thickness and treatment method number to form a structural treatment suggestion data record. All suggestions are recorded on the geological structure topology map for spatial mapping.

[0106] Based on the range of processed thickness values, the paths are divided into three categories of path records: those with thickness less than the set threshold, those in the middle range, and those exceeding the threshold.

[0107] Based on the treatment method number identifier, match the corresponding treatment process parameter table for each path, including grouting parameters, filling material type and interlayer spacing requirements;

[0108] For paths with a thickness less than the threshold, it is recommended to set up a single borehole grouting point in the corresponding area of ​​the path, using a low-pressure injection method, with the injection depth not exceeding the bottom interface of the path control layer; wherein, the low-pressure injection method: the grouting pressure is controlled at 0.2 to 0.5 MPa;

[0109] For paths with intermediate thickness, segmented grouting is recommended. Set up two or more grouting holes, control the grouting spacing to be no more than 3 meters, and use medium-pressure quantitative injection. The cumulative grouting volume is linearly distributed according to the treatment thickness. In the medium-pressure quantitative injection method, the grouting pressure is controlled at 0.5 to 0.8 MPa.

[0110] For paths where the thickness exceeds the upper limit threshold, it is recommended to perform layer-by-layer filling operations in the target area, using a layered compaction method of crushed stone and silty clay, controlling the thickness of each layer to not exceed 30 cm. After each layer is filled, a standard penetration test is performed to confirm the bearing capacity. After passing the test, the next layer of construction can continue. Specifically, the filling construction is carried out using a layered compaction method of crushed stone and silty clay, with each layer not exceeding 30 cm in thickness. A layer of crushed stone and a layer of silty clay are filled in sequence and compacted separately. The number of compaction passes is not less than three. After each layer is compacted, a compaction degree test is performed. The compaction degree is required to be not less than 95% of the standard compaction test value. The next layer of construction is not allowed for the section that does not meet the compaction degree requirements.

[0111] For records with path numbers indicating combined treatment methods, a combined grouting and filling process is performed, with grouting done first and then filling. Settlement monitoring is conducted after the first phase of construction is completed, and subsequent layer construction begins once the settlement stabilizes.

[0112] Finally, all specific operational parameters are bound to the corresponding path spatial coordinates to form a set of structural treatment suggestion data records, thus completing the stability identification of the goaf area and the recommendation of foundation treatment based on structural feature differences.

[0113] Example 1:

[0114] To verify the feasibility of this invention in practice, it was applied to a power transmission and transformation project foundation planning site with typical goaf geological features. The geological structure of this area was modeled, stability was assessed, and foundation treatment recommendations were made to evaluate the intelligent identification and differentiated decision-making capabilities of this invention under complex geological conditions.

[0115] The site is located in the shallow overburden zone of a historical coal seam mining area. The strata consist of top silty clay, lower gravel interbedded with siltstone transition layer and multiple fault zones. Some areas have subsidence, cracks and local slight inclination. It belongs to the type of goaf with discontinuous structure, complex lithology and obvious impact of historical mining. By deploying multi-point settlement monitoring instruments, shallow drilling lithology probes and three-dimensional laser scanners, a total of 472 structural path node data, including rock layer thickness, lithology code, porosity, water content, settlement rate and settlement acceleration, were collected.

[0116] In the data preprocessing stage, spatial coordinate calibration is first performed on the collected raw node data to construct a set of structural nodes and a set of spatial connection edges. Then, structural parameters are sequentially arranged within each structural path to construct a set of structural path feature tensors. Zero vector padding is performed on some paths with fewer than 10 nodes, and tail truncation is performed on paths with a length exceeding 24. Finally, a feature tensor with a uniform dimension of 24×6 is constructed and input into the improved Tree-LSTM network.

[0117] In the network model structure, a lithology modulation vector is introduced. The modulation factor is formed by the unique thermal encoding of rock strata and the normalization of porosity and water content. It is embedded into the forget gate calculation to participate in the adjustment of the differential retention coefficient. Taking path number P113 as an example, its upper lithology is medium-dense silty clay with a porosity of 0.34 and a water content of 0.28. In the forget gate update, the weight retention coefficient of the state of high water content structural sections is significantly improved, which effectively enhances the memory ability of easily subsidence path sections.

[0118] During the path state aggregation stage, stability scores are calculated by combining the normalized results of settlement rate and settlement acceleration. A weighted average strategy is used to calculate the score of each node, with node scores ranging from 0.21 to 0.92. Higher scores indicate greater instability. In path P113, the average score is 0.79, which is significantly higher than the sample average of 0.51, thus entering the model's attention focus region. Finally, after generating a stability focus representation vector, it is input into the processing suggestion generation unit, generating a processing thickness prediction value of 1.25m. The processing method is numbered 2, corresponding to dual-hole quantitative grouting.

[0119] In the output recommendation stage, based on the spatial coordinates of the structural path labels, a set of processing suggestion records is automatically generated, and classification suggestions are recommended within different path clustering areas. The model identified 43 high-risk paths and generated 43 processing suggestion records. Among them, 18 paths adopted the grouting and pressurization method, accounting for 41.9%; 12 paths adopted the layered filling method, accounting for 27.9%; and 13 paths adopted a combined process, accounting for 30.2%. In the later comparison with the results of artificial geotechnical expert annotation, the system-generated recommendation schemes and expert suggestions had an overlap rate of 86%, with the thickness prediction error for high consistency treatment being less than ±0.15m and the classification accuracy reaching 89.7%.

[0120] The table lists the input lithological characteristics, settlement scores, predicted treatment thickness, and recommended treatment methods for some typical structural paths, verifying the present invention's ability to respond to different path states, identify stability risks, and recommend foundation treatment schemes under complex goaf structural conditions. This demonstrates the integrated advantages of the present invention in intelligent identification, parameter modulation, and engineering decision-making.

[0121] Table 1: Comparison of Typical Structural Path Status Scores and Foundation Treatment Recommendations

[0122]

[0123] As can be seen from the table data, different structural paths exhibit significant structural differences and logical correspondences in lithological parameters, settlement response, and recommended treatment methods. This fully verifies the practicality and accuracy of the stability assessment and treatment suggestion generation method based on lithological modulation and improved Tree-LSTM structure proposed in this invention in engineering scenarios.

[0124] Taking path P113 as an example, its rock strata are medium-dense silty clay (unique thermal code: [0,1,0,0,0]), with a porosity of 0.34 and a water content of 0.28. After model evaluation, the stability score of this path reached 0.79, reflecting that it is in a high-risk area for settlement. Combined with the enhancement mechanism of state transfer by modulation vector, the model finally predicted that it needs to be treated with 1.25 meters of grouting. The recommended treatment method is dual-hole quantitative grouting (number 2). This path has similar characteristics to P122, with a score of 0.74. The recommended treatment method is consistent, showing that the model's treatment strategy is stable under similar structural conditions.

[0125] Paths P087 and P199 exhibit a medium risk level, with scores of 0.68 and 0.47 respectively. Both are recommended to use the single-hole low-pressure grouting method (number 1). Although the two have different lithological categories, their porosity and water content are at relatively low levels. The model assigns them a smaller retention coefficient when modulating the forget gate, thereby reducing the state intensity transmitted to the upper structure and ultimately corresponding to a lighter treatment requirement.

[0126] High-scoring paths such as P234 and P305 are characterized by lithological complexity and water content, with water contents of 0.36 and 0.42, respectively, and stability scores as high as 0.84 and 0.89, respectively. The model generated high-thickness treatment suggestions (1.62m and 1.85m) for these paths, and recommended more complex layered filling and composite treatment methods (method number 3), demonstrating the model's sensitivity and strategy optimization capabilities in identifying high-risk paths and composite structure areas.

[0127] The lowest-scoring path, P176, has a stability score of only 0.41, corresponding to a porosity of 0.27 and a water content of 0.19. It belongs to a path segment with stable lithology and low water sensitivity. The model judges that its structural stability is relatively high, requiring only 0.55-meter-thick grouting treatment. The recommended treatment method is single-hole low-pressure grouting to ensure a balance between cost control and structural safety.

[0128] In summary, this table reflects that the model generates stability scores for different structural paths based on lithological parameters and settlement characteristics, and maps these scores to differentiated processing thicknesses and suggested methods, forming a three-dimensional matching structure of structure-score-recommendation. In multiple path samples, the deviation between predicted thickness and actual processed thickness is controlled within ±0.15m, and the matching rate between the recommended method and the manual solution is close to 90%, which is better than the traditional experience-based judgment method. This demonstrates the accuracy, responsiveness, and engineering adaptability of this invention in intelligent structure identification and processing decision recommendation.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A ground-based intelligent recommendation method for infrastructure processing, characterized in that, Includes the following steps: Collect data from the goaf area and preprocess it to generate a geological structure topology map; A depth-first traversal is performed on the geological structure topology map to construct several structural paths. The structural parameters corresponding to each node in each path are arranged in the order of the path to form a set of structural path feature tensors. The set of structural path feature tensors is input into an improved Tree-LSTM network that incorporates lithology modulation vectors for structural state update calculation. Specifically, a lithology modulation vector consisting of rock type, porosity, and water content is added during the forget gate calculation process to differentially modulate the memory state of each sub-node. The settlement rate and settlement acceleration of each node in each structural path are standardized and calculated to construct a node stability score. During the path aggregation stage, attention weighting coefficients are calculated based on the node stability scores, and the hidden states of all child nodes are weighted and fused to generate a stability-focused representation of the current path. The stability focus representation is input into the treatment suggestion generation unit, multi-level linear mapping and parameter regression operations are performed, and the foundation treatment thickness value and treatment method number of the corresponding path are output. Bind the processing thickness value and the processing method number to the spatial coordinate label corresponding to the path to generate a structural processing suggestion data record; Output a set of data records suggesting structural processing, and complete the identification of site stability and recommendation of foundation treatment in the goaf area.

2. The intelligent recommendation method for ground-based processing based on neural networks according to claim 1, characterized in that, The collected data on the goaf area includes coal pillar size, coal seam depth, roof lithology, fault distribution, and settlement monitoring sequences. The preprocessing process includes performing spatial coordinate registration, numerical normalization, and time series completion operations on various structural parameters to generate a geological structure topology map.

3. The intelligent recommendation method for ground-based processing based on neural networks according to claim 2, characterized in that, The process of generating the structural path feature tensor set includes the following steps: A depth-first traversal operation is performed on the geological structure topology map. The starting point of the traversal is selected as the node located at the boundary between the surface and the main mining area. During the traversal, the numbers of all nodes passed in the path are recorded, and the structural path set is established in the order of access. For all nodes in each structural path, extract the structural parameters corresponding to the nodes in the path order. The structural parameters include coal pillar size, burial depth, roof lithology code, fault boundary distance, and settlement monitoring point number corresponding to the node. Among them, the lithology code of the roof is mapped to a fixed-dimensional lithology category vector using the unique thermal coding method, and the fault boundary distance is calculated numerically as the shortest spatial distance to the center point of the structural path and normalized to a floating-point value in the range of 0 to 1. The various structural parameters of a node are concatenated along the same node dimension to form a node feature vector. All node feature vectors are arranged sequentially along the structural path to form the input tensor of a structural path. After performing the above steps on all structural paths, the input tensors of all structural paths are summarized to form a set of structural path feature tensors. During tensor construction, zero-padding is performed on structural paths whose length is less than the preset minimum number of nodes, and truncation is performed on structural paths whose length exceeds the upper limit. All structural path tensors have the same dimension.

4. The intelligent recommendation method for ground-based processing depth based on neural networks according to claim 3, characterized in that, The process of inputting the set of structural path feature tensors into the improved Tree-LSTM network for structural state update calculation includes: For each sub-node in each structural path, extract the rock layer type, porosity, and water content corresponding to the sub-node. Convert the rock layer type into a fixed-length category vector using a unique thermal encoding method. Convert the porosity and water content into floating-point values ​​by normalizing the maximum and minimum values ​​respectively. Concatenate the three in order to form a lithology modulation vector. Concatenate the lithology modulation vector with the hidden state vector of the sub-node according to the dimension to form a new extended node vector input. When performing the forget gate calculation, the extended node vector is called as input, and the retention coefficient of the child node in the current propagation cycle is generated by weighting it with the current input vector of the parent node. The calculation process of the retention coefficient is directly affected by the values ​​of each element in the lithology modulation vector. Different rock strata, porosity levels and water content will change the value of the retention coefficient, thereby changing the amount of structural state information transmitted by the sub-node in the path. Subsequently, element-wise multiplication is performed using the retention coefficient and the memory state vector of the child nodes to obtain the weighted memory state result under lithological modulation. The weighted memory state results of all child nodes are summed element-wise and used as the update value of the memory state vector of the parent node, thus completing the progressive propagation of the structural state.

5. The intelligent recommendation method for ground-based treatment depth based on neural networks according to claim 4, characterized in that, The process of generating the stability focus representation of the current path includes: For each structural node in each structural path, extract the settlement monitoring time series corresponding to the node, calculate the difference between adjacent time steps to obtain the settlement rate, and calculate the settlement acceleration based on the rate change. The settling velocity and settling acceleration of each node are normalized to the interval between 0 and 1. A stability score is constructed using a weighted average function. The stability score of each child node is used as the attention input weight and multiplied element-wise with the hidden state vector of that child node to generate a weighted state vector. The weighted state vectors of all child nodes are accumulated and fused according to their dimensions to generate an aggregated state representation. The aggregated state representation is used as the stability focus representation of the current path to express the dominant influence of unstable structures on the path state. It participates in the propagation of the improved Tree-LSTM network as the hidden state of the parent node.

6. The intelligent recommendation method for ground-based processing based on neural networks according to claim 5, characterized in that, The processing suggestion generation unit includes the following components: The input interface layer is used to receive the stability focus representation corresponding to the structure path; The feature transformation subunit includes at least two layers of linear mapping structure, which are used to perform dimensional transformation and nonlinear activation processing on the input features and output intermediate feature representations for prediction. The thickness regression sub-unit is used to perform parametric regression operations based on intermediate feature representations to generate predicted values ​​for the foundation treatment thickness of the structure path. The processing method classification subunit is used to perform multi-class classification operations based on intermediate feature representations and outputs the processing method type number. The output combination layer is used to combine the predicted thickness value and the processing method type number to form a processing suggestion vector, which is then bound to the corresponding structure path label to output the structure processing suggestion record.

7. The intelligent recommendation method for ground-based processing depth based on neural networks according to claim 6, characterized in that, The process of outputting the foundation treatment thickness value and treatment method number for the corresponding path includes: The input interface layer receives the stability-focused representation corresponding to the structural path, which serves as the model input feature. The feature transformation subunit first feeds the input vector into the first linear mapping layer, performs a linear affine transformation with a bias term, completes the activation through the ReLU activation function, and then feeds it into the second linear mapping layer to compress the dimension and generate an intermediate feature representation; The thickness regression sub-unit receives intermediate feature representations, calls a linear regression function to generate floating-point predicted values, which are used as the foundation treatment thickness corresponding to the structural path. The processing method classification subunit receives the same intermediate feature representation, performs a linear mapping and then a softmax function, outputs the probability distribution corresponding to each processing method, and selects the category index corresponding to the maximum probability as the processing method number of the structure path.

8. The intelligent recommendation method for ground-based processing depth based on neural networks according to claim 7, characterized in that, The structural treatment suggestion data record includes spatial coordinate labels of the structural path, predicted values ​​of foundation treatment thickness, and treatment method number. This set of structural treatment suggestion data records is used to complete the identification of site stability in the goaf and to recommend foundation treatment suggestions, specifically including: The spatial coordinate labels of each structural path are bound to the corresponding predicted values ​​of foundation treatment thickness and treatment method number to form a structural treatment suggestion data record. All suggestions are recorded on the geological structure topology map for spatial mapping. Based on the range of processed thickness values, the paths are divided into three categories of path records: those with thickness less than the set threshold, those in the middle range, and those exceeding the threshold. Based on the treatment method number identifier, match the corresponding treatment process parameter table for each path, including grouting parameters, filling material type and interlayer spacing requirements; For paths with a thickness less than the threshold, it is recommended to set up a single borehole grouting point in the corresponding area of ​​the path, use a low-pressure injection method, and the injection depth should not exceed the bottom interface of the path control layer. For paths with thickness in the middle range, it is recommended to perform segmented grouting, set two or more grouting hole positions, control the grouting spacing to be no more than 3 meters, adopt medium-pressure quantitative injection method, and linearly distribute the cumulative grouting volume according to the treatment thickness. For paths where the thickness exceeds the upper limit threshold, it is recommended to perform layer-by-layer filling operations in the target area, using a layered compaction method of crushed stone and silty clay, controlling the thickness of each layer to not exceed 30 cm, and conducting a standard penetration test after each layer is filled to confirm the bearing capacity. After passing the test, the next layer of construction can continue. For records with path numbers indicating combined treatment methods, a combined grouting and filling process is performed, with grouting done first and then filling. Settlement monitoring is conducted after the first phase of construction is completed, and subsequent layer construction begins once the settlement stabilizes. Finally, all specific operational parameters are bound to the corresponding path spatial coordinates to form a set of structural treatment suggestion data records, thus completing the stability identification of the goaf area and the recommendation of foundation treatment based on structural feature differences.

Images