Project cost prediction method and system based on deep learning

Abstract

The invention provides a project cost prediction method and system based on deep learning, and the method comprises the steps: extracting a geological anomaly type and a support adjustment amplitude, and determining an initial anomaly feature set through combining the real-time monitoring of data triggering an anomaly condition; quantitatively classifying geological anomaly types and identifying anomaly distribution density through the initial anomaly feature set, separating abnormal point locations in the foundation pit area according to the anomaly distribution density to form an anomaly distribution data set, and identifying the geological anomaly types and the support adjustment amplitude in the anomaly distribution data set; constructing a knowledge graph through the geological anomaly type and the support adjustment amplitude, updating and mapping the knowledge graph into a graph structure, identifying an association adjustment rule and node weight distribution in the graph structure, and determining a local support adjustment knowledge graph; and extracting cost splitting details from the cost budget and the settlement file, counting historical support parameter adjustment data to obtain a parameter adjustment frequency, and determining a unified cost feature set in combination with the initial cost fluctuation interval.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and particularly to a project cost prediction method and system based on deep learning. Background Art

[0002] When studying the construction project cost knowledge graph construction and prediction method based on multi-source heterogeneous data fusion, for the deep foundation pit support project of high-rise buildings, it is necessary to deeply analyze the complex relationship between geological conditions, support structures, construction processes and project costs. During the construction process, due to sudden geological anomalies, extreme weather or surrounding environmental restrictions, the original design plan is forced to be adjusted. For example, the number of temporary support piles is increased or the support material is replaced, which will affect the final project cost prediction. At this time, different data sources, including geological exploration reports, construction logs, meteorological records, and design change documents, will have semantic conflicts and inconsistencies when describing the same event due to differences in collection methods, recording entities or time. For example, the geological exploration report marks a certain area as a stable sand layer, while the construction log records that quicksand appears in this area; the design change document may update the support structure parameters but not synchronize them to the cost estimation system in time. How to identify and process the subtle deviations in attribute description, timestamp alignment and causal association of cross-source data caused by sudden events when fusing these data is a technical difficulty that needs to be solved urgently. Especially when some data sources lack context annotations or have subjective tendencies due to manual input, how to quantify the impact of these conflicts on the accuracy of entity relationship reasoning in the knowledge graph without introducing additional deviations. Further, whether this deviation will be amplified in project cost prediction, for example, due to incomplete tracing of support structure adjustments, resulting in missing key variables in cost estimation, is the key technical problem of this solution. Summary of the Invention

[0003] The present invention provides a project cost prediction method based on deep learning, mainly including:

[0004] Extract the types of geological anomalies and the amplitude of support adjustment, and combine the data triggering the abnormal conditions through real-time monitoring to determine the initial abnormal feature set;

[0005] Through the initial abnormal feature set, quantitatively classify the types of geological anomalies and identify the abnormal distribution density, separate the abnormal points within the foundation pit area according to the abnormal distribution density to form an abnormal distribution data set, and identify the types of geological anomalies and the amplitude of support adjustment therein;

[0006] Construct a knowledge graph through the types of geological anomalies and the amplitude of support adjustment, update and map the knowledge graph into a graph structure, identify the associated adjustment rules and node weight distributions in the graph structure, and determine the local support adjustment knowledge graph;

[0007] Analyze the local support adjustment knowledge graph, obtain the correlation relationship between various geological anomaly types and the support adjustment amplitude, obtain the inference path length and the priority of inference rules according to the correlation relationship, identify the cost impact factors of geological anomaly types on the support adjustment amplitude, and obtain the preliminary cost fluctuation range;

[0008] Extract the cost breakdown details from the cost budget and settlement documents, count the historical support parameter adjustment data, obtain the parameter adjustment frequency, and combine it with the preliminary cost fluctuation range to determine the unified cost feature set;

[0009] Analyze and quantify the degree of confusion of semantic differences in describing cost features in the unified cost feature set on the association rules, adjust the node weight distribution in the knowledge graph according to the degree of confusion, process the unified cost feature set, and obtain the consistency verification value between the support adjustment amplitude and the cost fluctuation range;

[0010] If the consistency verification value does not meet the requirements, update the local support adjustment knowledge graph, recalculate the cost impact factors to determine the correction value of the cost prediction, compare the correction value with the actual data, if the deviation is greater than the preset threshold, adjust the node weight distribution, update and train the unified cost feature set, and determine the final cost prediction value.

[0011] Furthermore, extract the geological anomaly types and the support adjustment ranges, and combine the data of the real-time monitoring triggering the anomaly conditions to determine the initial anomaly feature set, including: collect the real-time displacement data of each measuring point of the surrounding rock by the rock stratum displacement monitoring and collection device at a preset sampling interval, eliminate the noise and remove the outliers from the original data through the data preprocessing module, and calculate the displacement data after preprocessing. For the displacement data after preprocessing, use the data analysis processor to calculate the deformation rate and the cumulative displacement of the surrounding rock. When the continuous duration of the deformation rate of a single measuring point exceeding the preset threshold reaches the anomaly triggering condition, generate a measuring point anomaly record. For the anomaly record, calculate the deformation acceleration, deformation trend coefficient and deformation cumulative amount of the surrounding rock through the feature extraction module, and construct an anomaly feature vector in combination with the surrounding rock level parameters. Scan the rock mass structure image around the anomaly record point by the geological structure detector, and use the image preprocessing module for noise reduction and enhancement to obtain the standardized rock mass structure image. Use the random forest algorithm to compare the rock mass structure feature library to identify geological structure categories such as faults and joint zones, and calculate the rock mass integrity score based on the joint development degree and lithology characteristics. The rock mass structure feature library is generated by training with the labeled data of historical engineering cases. According to the surrounding rock stress state data collected by the in-situ stress sensors at the monitoring layout points and the groundwater seepage flow monitoring data, and in combination with the anomaly feature vector, use the support vector machine classifier to judge the support form selection scheme. The support vector machine classifier is trained based on the historical data set including stress data, seepage flow data, anomaly feature data and corresponding support schemes. For the monitoring data, rock mass integrity score and support parameters of the anomaly record point, use the data analysis processor to establish a support parameter adjustment function based on the deformation rate of the surrounding rock, the rock mass integrity score and the in-situ stress state, and calculate the support parameter adjustment amount. According to the calculated support parameter adjustment amount, dynamically adjust the support structure parameters according to the preset adjustment rules to generate a new support parameter set.

[0012] Furthermore, through the initial abnormal feature set, quantitatively classify geological anomaly types and identify the abnormal distribution density, separate abnormal points within the foundation pit area according to the abnormal distribution density, form an abnormal distribution data set, and identify the geological anomaly types and support adjustment amplitudes therein, including: obtaining geological structure distribution data according to the foundation pit area range map, calculating the abnormal point density distribution map in the initial abnormal feature set using a pre-established point density calculation module, comparing the relationship between the point density value and the geological structure boundary line position through a density clustering algorithm to obtain the abnormal point clustering result. For the area where the density value in the abnormal point clustering result exceeds the preset clustering threshold, obtain the geological structure feature indicators within the area using the geological structure distribution map, establish a geological anomaly quantification function based on the feature indicators, and calculate the abnormal degree quantification values of each point in the abnormal point density distribution map. Divide the foundation pit sub-areas according to the abnormal point clustering result, establish area division boundary values through the boundary points of the sub-areas, and use a decision tree classifier to divide the abnormal point coordinate sets within multiple groups of foundation pit areas. The decision tree classifier is trained based on the historical engineering abnormal point distribution data. For the divided abnormal point coordinate sets, combined with the geological structure feature indicators of the sub-areas, establish an abnormal feature classification function through a geological anomaly identifier, and compare with the pre-stored abnormal type database to obtain the abnormal feature type identifiers of each point. Construct a point abnormal feature vector using the abnormal feature type identifier and the abnormal degree quantification value, generate an abnormal distribution data set through a feature vector clustering algorithm, and calculate the abnormal density distribution values of each sub-area based on this data set. For the abnormal density distribution values of each sub-area in the abnormal distribution data set, combined with the geological structure feature indicators, identify key abnormal points through a key abnormal position discriminator, and calculate the support parameter adjustment value using a neural network trained based on historical support adjustment records. According to the abnormal feature vector and the support parameter adjustment value of the key abnormal points, establish a support adjustment amplitude calculation function to obtain the support structure adjustment amplitudes of each key point.

[0013] Furthermore, a knowledge graph is constructed based on the geological anomaly types and the support adjustment amplitudes, and the updated knowledge graph is mapped into a graph structure. The associated adjustment rules and the node weight distribution in the graph structure are identified to determine the local support adjustment knowledge graph, including: constructing a knowledge graph through geological anomaly and anomaly type data, representing the corresponding relationship between geological anomalies and support adjustments using a graph structure to obtain an initial graph structure; using a mapping update method to refresh the graph structure in combination with the adjustment amplitude data to generate an updated knowledge graph; analyzing the node weights and weight distribution through the graph structure, calculating the quantization values of each node to obtain the node weight distribution result; extracting association rules for the node weight distribution result, and using a rule analysis method to identify the logical connection between support adjustments and anomaly types to generate an association rule set; judging, according to the association rule set and the local support data, that if the node weight in the local support area exceeds a preset threshold, then a local support adjustment plan is generated in combination with the adjustment amplitude to obtain a set of adjustment plans; and updating the weight distribution and association rules through the set of adjustment plans and the graph structure of the knowledge graph to obtain an optimized local support knowledge graph.

[0014] Furthermore, collect actual engineering cases and extract the corresponding relationship between geological anomaly types and support adjustment measures, convert the corresponding relationship into a rule form to obtain associated adjustment rules, statistically analyze the anomaly distribution data set to obtain the frequencies of different geological anomaly types and support adjustment measures, and numerically simulate the effects of different support plans under different geological anomaly conditions. Assign initial weights to each node in the graph structure based on the effects of different support plans to obtain the node weight distribution, including: according to the engineering case library identifier, using an association rule mining module to extract the corresponding relationship between geological anomaly types and support adjustment measures from the actual engineering data, and generating an associated adjustment rule set through rule strength quantization calculation; for the anomaly distribution data set, calculating the occurrence frequency values of different geological anomaly types and the distribution proportion of support adjustment measures through mathematical statistics methods to generate an anomaly-adjustment association frequency distribution diagram; constructing a support plan combination library based on the anomaly-adjustment association frequency distribution diagram, calculating the support structure force distribution function using preset support structure parameters to obtain the initial support effect index; establishing a geomechanics numerical calculation model based on the initial support effect index, obtaining the surrounding rock stress field distribution under different support plans through hierarchical progressive calculation, and calculating the support adaptability evaluation value; establishing the node connection relationship of the graph structure for the support adaptability evaluation value, quantitatively calculating the association strength between nodes using a preset connection degree threshold to generate the graph structure connection degree distribution diagram; based on the graph structure connection degree distribution diagram, traversing the support effect index space using a depth-first search algorithm, and calculating the weights of the nodes on the search path; and calculating the normalized weights of each node in the graph structure according to the node search path and the support adaptability evaluation value through a weight distribution function to obtain the node weight distribution diagram.

[0015] Furthermore, analyze the local support adjustment knowledge graph to obtain the correlation relationship between various geological anomaly types and the support adjustment amplitude. Based on the correlation relationship, obtain the inference path length and the priority of inference rules, identify the cost impact factors of geological anomaly types on the support adjustment amplitude, and obtain the preliminary cost fluctuation range, including: obtaining the correlation data between geological anomaly types and the support adjustment amplitude from the local support adjustment knowledge graph, calculating the shortest path length between nodes through the graph searcher, traversing all paths between the anomaly type nodes and the support adjustment nodes using the breadth-first search algorithm to obtain the initial knowledge path set. For the initial knowledge path set, calculate the frequency of each node in the path through the node correlation degree calculation module, and combine the connection strength between nodes to calculate the support node impact factor to obtain the node weight distribution map. Based on the node weight distribution map, use the path optimization function to score and sort all paths, and eliminate redundant paths through the path length threshold filter to generate the optimized knowledge path set. According to the optimized knowledge path set, obtain the unit cost parameters from the preset cost benchmark library, establish a support adjustment amplitude quantization function, and calculate the cost change value under different adjustment amplitudes. For the cost change value, use the interval division function to establish the cost fluctuation range, and train the random forest regressor in combination with historical engineering data to generate the cost impact prediction model. Based on the cost impact prediction model, calculate the cost impact coefficient corresponding to different anomaly types through the anomaly type feature vector, and establish the cost fluctuation mapping function. According to the cost fluctuation mapping function and the preset fluctuation threshold, divide the cost range of the support adjustment plan under different anomaly types to generate the cost fluctuation range distribution map. The local support adjustment knowledge graph contains various correlation relationships between geological anomaly type nodes and support adjustment nodes. For example, there is a direct connection between the fault fracture zone node and the lengthened bolt node, and the joint development zone node is indirectly connected to the specific adjustment measure node through the support reinforcement node.

[0016] Furthermore, through the analysis of the inference path length, the distribution characteristics of geological anomalies are judged. Combining with the priority of inference rules according to the distribution characteristics, the classification result of anomaly types is determined. The preliminary requirements for support adjustment are obtained from the classification result, and the initial range of the adjustment amplitude is obtained. The influence factors are calculated for the initial range, and the specific value of the amplitude change is judged. The support vector machine algorithm is used to determine the prediction model of cost impact, and the cost impact factor of the impact of geological anomalies on cost is output, including: obtaining the depth data of the geological anomaly inference path through the inference path length calculator, calculating the frequency of path node occurrences through the node traversal statistic, and using the depth-first search algorithm to perform spatial clustering on the anomaly distribution characteristics to obtain the initial anomaly distribution map. For the anomaly nodes in the initial anomaly distribution map, the priority evaluation function is used to calculate the inference rule weight value, and the rule priority ranking table is established by combining the anomaly feature indicators. According to the rule priority ranking table, the decision tree classifier is used to classify the characteristics of anomaly types, and the anomaly classification result is generated based on the pre-stored anomaly feature standard library. For the anomaly classification result, the support adjustment parameter library is constructed by using the adjustment records in the historical support dataset, and the change range of support parameters is quantified through the adjustment amplitude calculation function. Based on the change range of support parameters, the support vector machine classifier is trained by combining with the engineering practice data, and the classifier parameters are optimized by using the cross-validation method to establish the cost impact prediction function. According to the cost impact prediction function, the cost change amplitude under different anomaly types is calculated through the influence factor quantification module, and the cost impact mapping relationship is established. For the cost impact mapping relationship, the interval division function is used to classify the cost impact factors to generate the cost impact factor distribution map.

[0017] Furthermore, the cost split details are extracted from the cost budget and settlement documents, the historical support parameter adjustment data are counted, the parameter adjustment frequency is obtained, and the unified cost feature set is determined in combination with the preliminary cost fluctuation range, including: establishing a cost project database according to the cost budget document, extracting the support component number and construction process number through the cost project identification module, and using the project classification statistics to classify and summarize the material consumption value and the labor time consumption to generate a cost split detailed table. For the cost split detailed table, the pre-established support cost benchmark library is used to compare the project costs, and the support cost proportion of each project is obtained through the comparison calculation function, and a cost composition distribution diagram is established. Support adjustment records are extracted from the historical engineering database, and the adjustment parameter statistics are used to calculate the frequency of occurrence of different parameter adjustment types, and the parameter adjustment frequency distribution diagram is generated through the frequency sorting function. For the parameter adjustment frequency distribution diagram, combined with the support cost proportion data, the association rule mining algorithm is used to identify the association between high-frequency adjustment parameters and cost items, and a parameter cost association table is generated. Based on the parameter cost association table, the mapping relationship between the support parameter adjustment amount and the cost fluctuation amplitude is established through the random forest regressor, and the historical engineering cost data is used to train the regressor to obtain the cost fluctuation interval table. According to the cost fluctuation interval table, the interval feature extractor is used to calculate the fluctuation interval feature parameters, and the fluctuation interval is classified and labeled in combination with the preset feature threshold to establish a fluctuation feature library. According to the fluctuation feature library and the cost splitting details table, the support vector machine classifier is used to classify the support cost features, and a unified cost feature set is generated based on the preset classification standards. The cost budget file contains multiple support projects, such as reinforced concrete support walls, anchor support, shotcrete, etc. Each project has a corresponding component number and construction process number.

[0018] Furthermore, analyze and quantify the degree of confusion of the semantic differences in describing cost features in the unified cost feature set for the association rules, adjust the node weight distribution in the knowledge graph according to the degree of confusion, process the unified cost feature set to obtain the consistency verification value of the support adjustment range and the cost fluctuation range, including: According to the semantic word vector set in the unified cost feature set, use a word vector similarity calculator to calculate the semantic similarity of different feature descriptions, and identify semantic confusion feature pairs through a similarity threshold comparator to generate a feature confusion comparison table. For the feature confusion comparison table, use a feature vector calculation module to construct cost feature vectors, and calculate the degree of confusion between different feature vectors through a deep neural network, where the deep neural network is trained based on an annotated semantic data set. According to the degree of confusion of the feature vectors, use a confusion weight calculation function to generate a rule confusion matrix, obtain the confusion feature vectors through matrix eigenvalue decomposition, and establish a confusion degree quantification index. For the confusion degree quantification index, use a node weight update function to dynamically adjust the node weights in the knowledge graph, and generate a new weight distribution map based on a preset weight threshold. According to the new weight distribution map, standardize the adjustment parameters through a support adjustment range normalization function, and construct a feature mapping relationship using a standardized feature set. For the feature mapping relationship, use a random forest regressor to establish a cost fluctuation prediction function, where the random forest regressor is trained based on historical engineering feature data. According to the cost fluctuation prediction results, generate consistency verification parameters through a verification threshold calculator, calculate the matching degree between the support adjustment range and the cost fluctuation range, and obtain a verification value distribution curve. The unified cost feature set contains multiple semantic description methods, such as "the support wall is thickened" and "the support structure is thickened" expressing the same meaning.

[0019] Further, if the consistency check value does not meet the requirements, update the local support adjustment knowledge graph, recalculate the cost impact factor to determine the correction value of the cost prediction, compare the correction value with the actual data. If the deviation is greater than the preset threshold, adjust the node weight distribution, update and train the unified cost feature set to determine the final cost prediction value, including: if the check failure flag is triggered, obtain the adjustment record set according to the local support adjustment knowledge graph, construct the adjustment feature vector through the graph feature extractor, and retrain the cost impact prediction function based on the historical adjustment data using a deep neural network to obtain the updated cost impact factor. For the updated cost impact factor, verify the reliability of the prediction result through the feature verification function, generate the prediction deviation distribution map using the cost deviation calculator, and perform deviation classification based on the preset deviation threshold. According to the deviation classification result, optimize the structure of the local support adjustment knowledge graph using the graph update function, and generate the node weight adjustment parameter through the node importance calculator. For the node weight adjustment parameter, reassign the weights of the knowledge graph nodes using the weight distribution optimizer, and generate a new weight distribution table based on the preset weight adjustment rule. According to the new weight distribution table, restructure the unified cost feature set through the feature reconstruction function, and construct an extended feature set based on the updated features using the feature extender. Based on the extended feature set, perform feature training using a random forest classifier, and optimize and adjust the classifier parameters through the cross-validation optimizer to generate a new version of the feature model. For the new version of the feature model, calculate the final cost prediction value through the cost prediction function, and perform accuracy verification using the prediction result validator.

[0020] The present invention provides a deep learning-based engineering cost prediction system, mainly including:

[0021] An initial abnormal feature set extraction module, which is used to extract the geological anomaly type and the support adjustment range, and determine the initial abnormal feature set by combining the data that triggers the abnormal condition in real-time monitoring;

[0022] An abnormal distribution data set generation module, which is used to quantify and classify the geological anomaly type through the initial abnormal feature set, identify the abnormal distribution density, separate the abnormal points in the foundation pit area according to the abnormal distribution density to form an abnormal distribution data set, and identify the geological anomaly type and the support adjustment range therein;

[0023] A local support adjustment knowledge graph construction module, which is used to construct a knowledge graph through the geological anomaly type and the support adjustment range, update and map the knowledge graph into a graph structure, identify the associated adjustment rules and node weight distribution in the graph structure, and determine the local support adjustment knowledge graph;

[0024] The preliminary construction cost fluctuation range determination module is used to analyze the local support adjustment knowledge graph, obtain the correlation relationship between each geological anomaly type and the support adjustment range, obtain the inference path length and the priority of the inference rule according to the correlation relationship, identify the cost impact factor of the geological anomaly type on the support adjustment range, and obtain the preliminary construction cost fluctuation range;

[0025] The unified construction cost feature set determination module is used to extract the cost breakdown details from the cost budget and settlement documents, count the historical support parameter adjustment data, obtain the parameter adjustment frequency, and combine the preliminary construction cost fluctuation range to determine the unified construction cost feature set;

[0026] The consistency check value calculation module is used to analyze and quantify the confusion degree of the semantic differences in describing the cost features in the unified construction cost feature set on the association rules, adjust the node weight distribution in the knowledge graph according to the confusion degree, process the unified construction cost feature set, and obtain the consistency check value between the support adjustment range and the construction cost fluctuation range;

[0027] The final construction cost prediction value determination module is used to, if the consistency check value does not meet the requirements, update the local support adjustment knowledge graph, recalculate the cost impact factor to determine the correction value of the construction cost prediction, compare the correction value with the actual data, and if the deviation is greater than the preset threshold, adjust the node weight distribution, update and train the unified construction cost feature set, and determine the final construction cost prediction value.

[0028] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:

[0029] The present invention discloses a construction cost prediction method based on deep learning. This method extracts the geological anomaly type and the support adjustment range, constructs an initial anomaly feature set in combination with real-time monitoring data, and then quantifies and classifies geological anomalies and identifies the anomaly distribution density. Based on this, the present invention constructs and updates the local support adjustment knowledge graph, analyzes the correlation relationship between the geological anomaly type and the support adjustment range, and identifies the cost impact factor. Through inference path analysis and the support vector machine algorithm, the present invention determines the cost impact prediction model, and establishes a unified construction cost feature set in combination with historical data and cost breakdown details. Finally, through consistency check and iterative optimization, the present invention continuously adjusts the knowledge graph and the feature set, and finally obtains an accurate construction cost prediction value. This method can effectively cope with the construction cost fluctuation problem caused by geological anomalies in foundation pit engineering, and improve the accuracy and reliability of construction cost prediction. Description of the Drawings

[0030] Figure 1 It is a flowchart of a construction cost prediction method based on deep learning of the present invention.

[0031] Figure 2 It is a schematic diagram of a construction cost prediction method based on deep learning of the present invention. Specific Embodiments

[0032] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0033] As Figure 1-2 , a method and system for predicting project cost based on deep learning in this embodiment may specifically include:

[0034] S101. If it is necessary to predict the cost of deep foundation pit support engineering, first extract the geological anomaly type and support adjustment amplitude, generate an initial anomaly feature set in combination with real-time monitoring data, and analyze the anomaly distribution characteristics according to the feature set.

[0035] S1011. In practical applications, use the rock formation displacement monitoring equipment to collect the real-time displacement data of the surrounding rock measurement points at a preset sampling interval. The data preprocessing module filters out noise and removes outliers from the collected raw data to ensure data quality. Then, the feature extraction module calculates the deformation acceleration and deformation trend coefficient of the surrounding rock, generates an anomaly feature vector in combination with the surrounding rock level parameter. At the same time, scan the rock mass structure image of the anomaly area with a geological structure detector. After noise reduction and enhancement by the image preprocessing module, identify the geological structure category based on the rock mass structure feature library and calculate the rock mass integrity score.

[0036] S1012. According to the real-time displacement data of the surrounding rock collected by the rock formation displacement monitoring equipment, the data analysis processor calculates the deformation rate and cumulative displacement. When the deformation rate of the measurement point continuously exceeds the preset threshold, an anomaly record is generated. Then, the feature extraction module calculates the deformation acceleration and cumulative deformation in combination with the surrounding rock level parameter to construct an anomaly feature vector. In addition, for the rock mass structure image obtained by the geological structure detector, use the random forest algorithm to identify the type of fault or joint zone, calculate the rock mass integrity score based on the joint development degree and lithology characteristics. At the same time, in combination with the stress sensor and seepage flow monitoring data, use the support vector machine classifier to analyze the selection of support schemes, and establish a support parameter adjustment function according to the surrounding rock deformation rate and stress state, calculate the adjustment amount and generate a new set of support parameters.

[0037] In some optional embodiments, the rock displacement monitoring equipment collects displacement, inclination and stress data at a sampling interval of 300 seconds, filters out noise caused by equipment jitter or electromagnetic interference through wavelet transform, sets a 3-times standard deviation threshold to eliminate abnormal values, and obtains pre-processed monitoring data. Based on this, the deformation rate of the surrounding rock is calculated to be 0.5 mm / hour, and the cumulative displacement is 35 mm. If the deformation rate exceeds 0.3 mm / hour for 8 consecutive hours, an abnormal record is triggered. For the abnormal area, the deformation acceleration is calculated to be 0.08 mm / square hour, and the deformation coefficient is 1.2. Combined with the level III surrounding rock parameters, an abnormal feature vector is formed.

[0038] Furthermore, the geological structure detector was used to scan the rock structure at a depth of 30 meters with high-frequency electromagnetic waves. The images were processed by median filtering and histogram equalization to identify a fault fracture zone with an inclination of 75 degrees, with moderate joint development, sandstone lithology, and a rock integrity score of 68 points. Combined with stress monitoring data, vertical stress of 12 MPa, horizontal stress of 8 MPa, and seepage of 25 liters / hour, the support vector machine classifier trained based on 2000 sets of historical data was input to determine the anchor rod plus steel mesh support scheme, and a new parameter set was obtained by adjusting the function calculation, with the anchor rod length increased by 0.5 meters, the spacing reduced by 0.3 meters, and the steel mesh specification increased to 8 mm.

[0039] In the embodiment of the present invention, after applying the new support parameters, the deformation rate of the surrounding rock dropped to below 0.2 mm / h, the cumulative displacement tended to be stable, and the stress distribution was uniform, indicating that the adjustment scheme was effective. This dynamic analysis method based on multi-source data can capture geological anomalies in a timely manner and optimize support parameters, laying a solid foundation for cost prediction. It is understandable that the specific implementation method of this step can be adjusted according to the actual engineering scenario. For example, the sampling interval or threshold setting can be optimized by the technician according to the needs without too much limitation.

[0040] S102. Analyze the characteristics of geological anomalies through the initial anomaly feature set, quantify and classify the types of geological anomalies and identify the anomaly distribution density, and then separate the abnormal points in the foundation pit area according to the density distribution to form an anomaly distribution data set including the geological anomaly type and the support adjustment range.

[0041] In some optional embodiments, a pre-built point density calculation module is used to process the foundation pit area range map to generate an abnormal point density distribution map, which is based on the multi-dimensional data in the initial abnormal feature set, reflecting the spatial distribution law of abnormal points, and then the relative position relationship between the point density value and the geological structure boundary line is analyzed by the density clustering algorithm to obtain the clustering grouping results of the abnormal points, and the foundation pit sub-areas are divided based on this for refined analysis. Based on the clustering results, geological structural characteristic indicators such as joint development degree and rock mass integrity are used to establish a geological anomaly quantification function through mathematical modeling to calculate the quantitative value of the abnormal degree of each point.

[0042] S1021. In the embodiment of the present invention, geological structure distribution data is obtained based on the foundation pit area range map, and a density distribution map of the initial abnormal feature set is generated by a point density calculation module. The density clustering algorithm such as DBSCAN is used to process this distribution map, where the clustering radius and the minimum number of points are set as parameters. By comparing the position relationship between the point density value and the geological structure demarcation line, the clustering result of abnormal points is generated. At the same time, for the area where the density exceeds the preset threshold, geological structure characteristic indexes such as rock layer dip angle and joint density are extracted from the geological structure distribution map, and the abnormal degree quantization value is calculated through a geological anomaly quantization function. This function takes the characteristic indexes as inputs, and the output range is between 0 and 1, reflecting the severity of the anomaly.

[0043] S1022. In practical applications, foundation pit sub-regions are divided according to the clustering result of abnormal points. A division threshold for the boundary points of the sub-regions is set, and the coordinate set of abnormal points is classified by a decision tree classifier. This classifier is trained based on historical engineering data and includes feature inputs such as coordinates and geological parameters, and can effectively separate the coordinate groups of different abnormal types of points. Subsequently, combined with the geological structure characteristic indexes in the sub-region, a geological anomaly recognition device is used to construct an abnormal feature classification function. This function identifies specific abnormal types such as faults and joint zones by comparing with a pre-stored abnormal type database, and generates corresponding feature type identifiers. Further, the identifier and the abnormal degree quantization value are combined into a point abnormal feature vector, and the K-means clustering algorithm is used to group the feature vectors to generate an abnormal distribution data set, which includes the abnormal density distribution values of each sub-region.

[0044] For the high-density sub-regions in the abnormal distribution data set, a key abnormal position discriminator is used to identify the key abnormal points. This discriminator combines geological structure characteristic indexes and determines the core influencing points in the abnormal distribution through a logistic regression method. Then, a multi-layer perceptron neural network trained based on historical support adjustment records is used to calculate the support parameter adjustment value of the key points. The neural network takes the abnormal feature vector as the input and outputs adjustment parameters such as pile spacing or stiffness change value. Based on this, through a support adjustment amplitude calculation function, the abnormal feature vector and the adjustment value of the key points are mapped to the specific support structure adjustment amplitude, such as increasing the anchor rod length or adjusting the steel bar specification.

[0045] In a specific scenario, assume that the area of the foundation pit construction area is 8,000 square meters, and the initial abnormal feature set contains 127 abnormal points. Through the analysis of the point density calculation module, there are 1.6 abnormal points evenly distributed per 100 square meters on average, and the high-density area reaches 5.8 points per 100 square meters. Using the density clustering algorithm, with a clustering radius of 20 meters and a minimum number of points of 5, 3 abnormal dense areas are identified. For these areas, geological structure characteristic indicators are extracted, such as the joint development degree of 0.72, the rock mass integrity of 62 points, and the rock layer dip angle of 75 degrees. A quantization function is established to calculate the degree of abnormality, and it is found that the quantization value of some points reaches 0.82, indicating the existence of serious abnormalities.

[0046] Furthermore, the foundation pit is divided into 12 sub-areas, each area is about 670 square meters, and the spacing between boundary points is 5 meters. The decision tree classifier is used to separate the coordinates of abnormal points. This classifier is trained with 2,000 sets of historical data and can accurately group according to the coordinates and geological features. Combining with the abnormal type database, it is identified that the dense areas are mainly fault fracture zones and joint zones. After constructing the feature vector, an abnormal distribution data set is generated through the K-means algorithm. The abnormal densities of sub-areas 3, 7, and 9 are calculated to be 0.82, 0.76, and 0.68 respectively, indicating that the geological conditions in these areas are poor.

[0047] For the key points in the high-density area, using the multi-layer perceptron neural network, trained based on 500 sets of historical support records, it is calculated that the spacing of the support piles needs to be reduced by 0.3 meters, the pile length needs to be increased by 0.8 meters, and the stiffness needs to be increased by 25%. Then, through the adjustment function, it is determined that the length of the anchor rod in sub-area 3 is increased by 0.5 meters and the spacing is reduced by 0.2 meters, and the anchor rod specifications in sub-areas 7 and 9 are increased to 32 millimeters. This adjustment method based on density analysis and neural network can accurately reflect the dynamic impact of geological abnormalities on the support requirements, ensuring the adaptability and reliability of the cost prediction. It can be understood that the specific parameter settings of the density clustering algorithm and the neural network in the embodiments of the present invention can be optimized by technicians according to the actual engineering needs, such as adjusting the clustering radius or the number of network layers, to improve the accuracy of abnormal recognition and adjustment calculation, and no excessive limitations are made.

[0048] S103. By analyzing the types of geological abnormalities and the support adjustment amplitude, a knowledge graph is constructed using multi-source data and mapped into a graph structure. The associated adjustment rules and node weight distributions are identified, and finally, the local support adjustment knowledge graph is determined to provide a structured basis for the cost prediction of deep foundation pits.

[0049] In some alternative embodiments, based on the geological structure distribution data in the foundation pit area map, a point density distribution map is generated by the point density calculation module. The density clustering analysis method is used to compare the spatial relationship between the point density and the geological structure boundary line, identify the abnormal point aggregation area, and then a geological anomaly quantification function is constructed using indicators such as joint development degree, rock mass integrity score, and rock layer dip angle. The quantitative parameter of the anomaly degree is calculated, and sub-regions are divided based on this. A classification function is established by combining geological parameters and anomaly characteristics to identify geological anomaly types such as faults, joint zones, and weak interlayers. The optimization parameter set of the support structure is generated by feature vector clustering to improve the construction process of the knowledge graph.

[0050] For the abnormal point aggregation area, rock layer displacement, dip angle, and stress data are collected using real-time monitoring equipment. The number of abnormal points per unit area is counted through grid division to form a density distribution map. When the point density in a certain area exceeds the preset threshold, the DBSCAN density clustering algorithm is used to identify the aggregation area. This algorithm takes the clustering radius and the minimum number of points as core parameters, and determines the spatial distribution characteristics of abnormal points by iteratively calculating the distance and density connectivity between points. For example, the area of the foundation pit is 8000 square meters, and 127 abnormal points are monitored. The density distribution map shows that the number of points per 100 square meters in some areas is as high as 5.8, far exceeding the average value of 1.6, indicating a significant abnormal aggregation.

[0051] S1031. In the embodiment of the present invention, geological structure feature data is extracted for the abnormal point aggregation area, including indicators such as joint development degree, rock mass integrity score, and rock layer dip angle. A geological anomaly quantification function is established through mathematical modeling to calculate the quantitative parameter of the anomaly degree of each point. Among them, the joint development degree is calculated based on the number and length of joints per unit volume, the rock mass integrity score is generated from rock mechanics test data, and the rock layer dip angle is obtained by sensor measurement. Then, according to the anomaly degree parameter and the foundation pit engineering area division rule, the aggregation area is divided into multiple sub-regions. A decision tree classifier is used to classify the abnormal points in the sub-regions. This classifier is trained with historical engineering data, and the input features include point coordinates, geological parameters, and anomaly degree. The output is the identification result of geological anomaly types such as faults and joint zones. Further, a feature vector is constructed by combining the identification result and the quantitative parameter, and an abnormal distribution data set is generated by the K-means clustering algorithm in the feature space to calculate the abnormal density distribution parameter of the sub-region.

[0052] In the high abnormal density sub-region, a neural network calculation model is used to adjust the support parameters. This model takes 500 sets of historical support engineering records as training data, inputs the abnormal feature vectors, and outputs the adjusted reference values such as the spacing, length, and stiffness of the support piles. For example, in the area with a joint development degree of 0.72, a rock mass integrity score of 62 points, and an inclination angle of 75 degrees, the neural network calculates that the pile spacing is reduced by 0.3 meters and the pile length is increased by 0.8 meters. Through the support adjustment amplitude calculation function, the feature vectors and the adjusted reference values are mapped into an optimized parameter set, such as the anchor rod length is increased to 4.5 meters and the spacing is adjusted to 0.8 meters. This method ensures a high degree of matching between the support parameters and the geological abnormal characteristics through quantitative analysis and intelligent calculation.

[0053] S1032. In practical applications, based on the abnormal distribution data set and historical engineering cases, the association rule mining module is used to extract the corresponding relationship between the geological abnormal types and the support adjustment measures, forming an association adjustment rule set. For example, the fault fracture zone corresponds to the lengthening of the anchor rod and the densification of the spacing, and the joint zone corresponds to the improvement of the concrete strength. Then, through mathematical statistics, the occurrence frequency of the abnormal types and the distribution proportion of the support measures are analyzed to generate an abnormal-adjustment association frequency distribution diagram. Based on this, a support plan combination library is constructed, and the initial effect indexes of different support plans, such as the stress concentration coefficient and the deformation amount, are calculated using the stress distribution function. Subsequently, a geomechanics numerical model is established, and the stress field distribution of the surrounding rock is calculated layer by layer to evaluate the support adaptability, and the evaluation value is mapped into the connection relationship between the nodes of the graph structure. The depth-first search algorithm is used to traverse the effect index space, calculate the normalized weights of each node, generate a node weight distribution diagram, and improve the local support adjustment knowledge graph.

[0054] In a specific scenario, the engineering case library shows that the frequency proportion of the fault fracture zone is 35%, the soft interlayer is 28%, and the joint zone is 22%. In the support adjustment, the adjustment proportion of the anchor rod parameters is 42%, and the adjustment of the concrete strength is 31%. The association rule mining takes the minimum support of 0.3 and the confidence of 0.8 as thresholds to extract high-strength rules. The support plan combination library includes various forms, such as bolt-shotcrete support, and the initial parameters are set as a thickness of 300 mm and an anchor rod length of 4 m. The numerical calculation shows that the stress concentration coefficient in the fault area is 2.3 and the deformation amount is 15 mm, and the stress coefficient in the joint zone is 1.8 and the deformation amount is 8 mm. In the graph structure, the connection degree between the lengthening of the anchor rod and the improvement of the concrete strength reaches 0.8, and the weights are 0.35 and 0.28 respectively, reflecting their importance in the optimization.

[0055] It can be understood that through multi-source data fusion and intelligent algorithms, the embodiments of the present invention realize the full-chain analysis from geological anomaly identification to support parameter optimization, significantly improving the practicality and adaptability of the knowledge graph. The specific algorithm parameters can be adjusted according to actual needs and are not overly limited.

[0056] S104. By parsing the local support adjustment knowledge graph, the correlation between geological anomaly types and the support adjustment range is mined. Based on the inference path length and rule priority, the cost impact factors are identified, and finally the preliminary cost fluctuation range is generated.

[0057] Based on the local support adjustment knowledge graph, the shortest path length between the geological anomaly type node and the support adjustment node is calculated using the graph searcher. The breadth-first search algorithm is used to traverse all possible paths to generate the initial knowledge path set. Then, the node association degree calculation module counts the frequency of node occurrences in the path, combines the connection strength to generate the node weight distribution map, and then uses the path optimization function to score and sort the paths. The path set is optimized by filtering through the length threshold. Based on this set and combined with the cost benchmark library, a support adjustment range quantization function is established to calculate the cost change value. Finally, a cost impact prediction model is trained using a random forest regressor, and the cost fluctuation range distribution map is output.

[0058] S1041. In practical applications, for the associated data in the knowledge graph, the graph searcher calculates the shortest path from the anomaly type node to the support adjustment node using the breadth-first search algorithm. For example, the path length from the fault fracture zone to the lengthened bolt is 2, and the path length from the joint zone to the concrete strength improvement is 3. This algorithm is based on the direct or indirect connection between nodes and expands the search layer by layer until all reachable paths are covered, generating an initial knowledge path set containing dozens of paths. Subsequently, the node association degree calculation module analyzes the frequency of each node occurrence in the path. For example, the support reinforcement node appears in more than 40% of the paths. Combining the connection strength between nodes, such as the strength of 0.9 between the lengthened bolt and the reinforcement, a node weight distribution map is generated. Then, the path optimization function comprehensively scores the paths based on the path length and node weight, and filters out the optimized path set with a length not exceeding 4 and a score higher than 0.8. For example, the path "fault fracture zone - support reinforcement - lengthened bolt" has a score of 0.85.

[0059] For the optimized knowledge path set, the unit cost parameters are extracted based on the preset cost benchmark library, such as 300 yuan per meter for bolts and 750 yuan per cubic meter for concrete. The adjusted cost change is calculated through the support adjustment range quantization function. For example, when the bolt length increases by 0.5 meters and the spacing decreases by 0.2 meters, the cost per square meter increases by approximately 180 yuan. Then, the interval division function is used to map the change value to the fluctuation range. Combining historical project data, a random forest regressor is trained. This regressor takes the anomaly type feature vector as the input and predicts the cost impact coefficient through the integration of multiple decision trees. For example, the cost fluctuation range of the fault fracture zone is 1.25 times to 1.35 times the original cost, and that of the soft interlayer is 1.15 times to 1.25 times, generating an intuitive cost fluctuation range distribution map.

[0060] S1042. In the embodiment of the present invention, the distribution characteristics of geological anomalies are analyzed through the inference path length. The depth-first search algorithm is used to perform spatial clustering on anomaly nodes to generate an initial anomaly distribution map. Then, a priority evaluation function is adopted to calculate the weight values of inference rules and generate a rule priority ranking table in combination with anomaly feature indicators. The decision tree classifier is used to classify the anomaly types. Based on the classification results and historical support data, a support vector machine classifier is trained. The cost impact factor is calculated and a prediction model is established. Finally, the specific impact range of geological anomalies on the cost is output.

[0061] In a specific scenario, the inference path length calculator analyzes the path depth and finds that the average length of the fault fracture zone path is 4 and that of the joint zone is 5. Starting from the high-frequency nodes, the depth-first search algorithm is used to cluster the anomaly distribution along the paths with high connection strength. For example, the frequency of the support reinforcement node reaches 65%, generating an initial anomaly distribution map. Subsequently, the priority evaluation function calculates the rule weights according to the node frequency and association strength. For example, the rule weight of the fault fracture zone is 0.85 and that of the weak interlayer is 0.78. Then, in combination with indicators such as rock mass integrity and joint development degree, a priority ranking table is generated. The decision tree classifier identifies the type based on the anomaly feature standard library. For example, when the rock mass integrity is lower than 0.6 and the joint development degree is greater than 0.75, it is determined as the fault fracture zone.

[0062] For the classification results, an adjustment parameter library is constructed using the historical support data set. For example, the bolt length of the fault fracture zone is increased by 0.5 meters to 1 meter, and the support thickness of the weak interlayer is increased by 100 millimeters to 200 millimeters. The change range is quantified through the adjustment amplitude calculation function. Then, in combination with 500 sets of engineering practice data, a support vector machine classifier is trained. The radial basis kernel function and cross-validation are used to optimize the parameters. The predicted cost impact factor of the fault fracture zone is 1.3 and that of the weak interlayer is 1.2. Finally, a cost impact factor distribution map is generated through the interval division function, showing that the high-impact area is concentrated in the fault fracture zone with a fluctuation range of up to 1.35 times, and the medium-impact area is the weak interlayer with a fluctuation range of 1.15 times to 1.25 times.

[0063] It can be understood that in the embodiment of the present invention, through the combination of map analysis and intelligent algorithms, an accurate mapping from geological anomalies to cost impacts is realized. The algorithm parameters and thresholds can be flexibly adjusted according to the actual engineering requirements to improve the adaptability and accuracy of the prediction.

[0064] S105. Extract the cost breakdown details by parsing the cost budget and settlement documents, adjust the statistical adjustment frequency in combination with the historical support parameter adjustment data, and fuse the preliminary cost fluctuation range to construct a unified cost feature set.

[0065] Based on the cost budget file, the cost item identification module is used to extract the support component number and construction process number, and a cost item database is established. The project cost comparison is then performed through the preset support cost benchmark library, and the support cost proportion distribution data is generated using the comparison calculation function. The historical support adjustment records are then analyzed using the adjustment parameter counter, and the frequency of parameter adjustment types is calculated. The parameter adjustment frequency distribution diagram is generated through the frequency sorting function. The mapping relationship between the adjustment amount and the cost fluctuation is established in combination with the random forest regressor. Finally, a unified cost feature set is generated based on the support vector machine classifier.

[0066] S1051. In actual application, support component information is extracted from the cost budget file, such as the component number and corresponding construction process number of the reinforced concrete support wall. The material consumption and labor time consumption are classified and counted through the cost item identification module. For example, each cubic meter of concrete consumes 80 kg of steel bars and the labor time required for steel bar workers is 2.5 hours. A detailed cost breakdown list is generated, and then the support cost benchmark library is used to compare the standard cost of each project. It is calculated that the reinforced concrete support wall accounts for 45% of the total cost and the anchor support accounts for 28%. The cost composition distribution diagram is generated by comparing the calculation function, which clearly reflects the cost weight of each support form.

[0067] For the support adjustment records in the historical engineering database, the adjustment parameter counter is used to calculate the frequency of occurrence of various adjustment types. For example, the adjustment of support wall thickness accounts for 40%, and the adjustment of anchor rod length accounts for 35%. The parameter adjustment frequency distribution diagram is generated by the frequency sorting function. It is found that the adjustments of increasing thickness by 100 mm and anchor rod length by 500 mm are the most common. The association rule mining algorithm is then used to analyze the correlation between high-frequency adjustment parameters and cost items. For example, the correlation coefficient between the increase in support wall thickness and the increase in material cost is 0.85. A parameter cost correlation table is generated to further clarify the impact of adjustments on construction costs.

[0068] S1052. In an embodiment of the present invention, based on the parameter adjustment frequency distribution diagram and the support cost ratio data, a mapping relationship between the support parameter adjustment amount and the cost fluctuation is established through a random forest regressor. The regressor takes the adjustment amount, geological conditions and original cost as input features, and predicts cost changes by integrating multiple decision trees. For example, an increase in the thickness of the support wall leads to a 20% to 30% increase in cost, and a lengthening of the anchor rod leads to an increase of 15% to 25%. The model is trained using 1,000 sets of historical engineering data to obtain a cost fluctuation range table, and then the characteristic parameters of the fluctuation range are analyzed through an interval feature extractor, and the fluctuations are divided into three levels: large, medium and slight. For example, a large fluctuation corresponds to a cost change of more than 25%, and the characteristics include an increase in the thickness of the support wall and an increase in the density of the anchor rod, thereby generating a fluctuation feature library.

[0069] According to the fluctuation feature library and cost split details table, the support vector machine classifier is used to classify the support cost features. The classifier takes the fluctuation feature parameters as input and divides the projects into high, medium and low cost feature categories based on the preset classification standards. For example, when the thickness of the support wall is increased and the anchor rod is encrypted, it is classified as a high cost category, and when only the anchor rod length is adjusted, it is classified as a medium cost category. Through training and optimization of the classifier parameters, the classification accuracy rate reaches 88%, and a unified cost feature set is generated, which realizes the accurate mapping from parameter adjustment to cost features.

[0070] In a specific scenario, the cost budget document shows that the cost of anchor support materials is about 300 yuan per meter, and the cost of concrete support is about 750 yuan per cubic meter. When the thickness of the support wall increases by 100 mm, the material cost increases by about 25% and the labor cost increases by 15%. The random forest regressor predicts that in areas with complex geological conditions, the cost fluctuation may reach 1.3 times the original budget. This analysis method reveals the specific impact of adjusting parameters on cost through data-driven.

[0071] It can be understood that the embodiment of the present invention ensures the comprehensiveness and accuracy of the cost feature set through the combination of multi-level data analysis and intelligent algorithms. The specific thresholds and classification standards can be adjusted according to actual engineering needs to adapt to the prediction requirements in different scenarios.

[0072] S106. By analyzing the semantic differences in the unified cost feature set, quantifying its impact on the confusion degree of association rules, and dynamically adjusting the node weight distribution of the knowledge graph according to the confusion degree, the feature set is finally optimized to generate a consistency check value between the support adjustment range and the cost fluctuation range.

[0073] Based on the feature description information in the unified cost feature set, the word vector similarity calculator is used to analyze the semantic similarity. The confused feature pairs are identified by setting the threshold, and a feature confusion comparison table is generated. The deep neural network is then used to calculate the degree of confusion between the feature vectors, and a rule confusion matrix is constructed. The knowledge graph weights are then adjusted through the node weight update function to form a new weight distribution graph. Subsequently, the random forest regressor is combined to establish the cost fluctuation prediction function, and the consistency check value is calculated through the verification threshold.

[0074] In some optional embodiments, semantic descriptions are extracted from the unified cost feature set, such as "support wall thickening" and "support structure thickening", and the similarity between the two is calculated by a word vector similarity calculator based on a pre-trained word vector model. For example, if the similarity reaches 0.92, which is higher than the threshold value of 0.85, it is determined to be a confusing pair. Similarly, "anchor rod lengthening" and "anchor extension" have a similarity of 0.88. Subsequently, a feature confusion comparison table is generated, listing all potentially confusing feature pairs. This method can effectively identify semantic repetition or ambiguity, thereby improving the standardization of the feature set.

[0075] S1061. In an embodiment of the present invention, for the feature confusion comparison table, the cost feature is converted into a multi-dimensional vector by a feature vector calculation module, including information such as support type, adjustment method, and magnitude. Then, the confusion degree is calculated through a deep neural network, which is trained with 3000 groups of labeled semantic data and adopts a three-layer hidden layer structure. The number of nodes in each layer is 256, 128, and 64 respectively. The semantic associations between features are extracted layer by layer through an activation function, and a confusion degree score from 0 to 1 is output. For example, the confusion degree between "thickening of the retaining wall" and "thickening of the support structure" is 0.72. Subsequently, a regular confusion matrix is generated based on the confusion degree, and the source of confusion is analyzed through eigenvalue decomposition. It is found that the proportion of the adjustment method description is as high as 65%, which is the main factor of confusion. Furthermore, a confusion degree quantification index is established.

[0076] For the confusion degree quantification index, a node weight update function is used to dynamically adjust the nodes in the knowledge graph. For example, the weight of the high-confusion node "thickening of the retaining wall" is reduced from 0.8 to 0.65, while the weight of "thickening of the support structure" remains unchanged at 0.8. A new weight distribution map is generated by screening through a preset weight threshold of 0.3 to make the weight difference more significant. Subsequently, the support adjustment range is normalized. For example, an increase in thickness of 100 mm is standardized to 0.4, and an increase in anchor rod length of 500 mm is standardized to 0.6, ensuring the comparability of parameters with different dimensions and optimizing the construction process of the feature mapping relationship.

[0077] S1062. In practical applications, based on the new weight distribution map, a cost fluctuation prediction function is established using a random forest regressor, which takes 1000 groups of historical project data as the training set. The input features include standardized adjustment parameters, original cost, and geological conditions. The cost change is predicted through the integration of multiple decision trees. For example, a standard value of 0.4 for the retaining wall thickness corresponds to a 25% cost increase, and a standard value of 0.6 for the anchor rod length corresponds to a 30% increase, generating a calibration value distribution curve. Then, a deviation threshold of 0.15 is set through a calibration threshold calculator to calculate the consistency between the support adjustment range and the cost fluctuation range. For example, the calibration value distribution of the retaining wall adjustment is between 0.08 and 0.12, and the anchor adjustment is between 0.1 and 0.14, both meeting the requirements, indicating that the prediction result highly coincides with the actual fluctuation.

[0078] In a specific scenario, the unified cost feature set may cause semantic confusion due to different description habits. After analyzing the confusion degree and adjusting the weights through a deep neural network, the knowledge graph can more accurately reflect the true impact of support adjustment on cost. For example, the cost prediction deviation of the anchor rod lengthening plan is reduced from 20% to within 10%. This method significantly improves the prediction accuracy by eliminating semantic ambiguity.

[0079] It can be understood that in the embodiments of the present invention, through semantic analysis and intelligent algorithm optimization, the consistency between the cost feature set and the actual engineering requirements is ensured. The specific thresholds and network parameters can be flexibly adjusted according to the actual situation to meet the prediction requirements of deep foundation pit projects of different scales.

[0080] S107. If the consistency verification value fails to meet the expected requirements, update the local support adjustment knowledge graph and recalculate the cost impact factor to determine the correction value of the cost prediction. Then compare the correction value with the actual data. If the deviation exceeds the preset threshold, optimize the node weight distribution and train the unified cost feature set to generate the cost prediction value.

[0081] In some alternative embodiments, when the verification result shows inconsistency, extract the adjustment record set from the local support adjustment knowledge graph, use the graph feature extractor to generate the adjustment feature vector, train the historical data based on the deep neural network to update the cost impact factor. Subsequently, analyze the prediction deviation distribution through the cost deviation calculator. If it exceeds the threshold, perform classification optimization, then adjust the node weights of the knowledge graph and reconstruct the feature set, and optimize the training through the random forest classifier to output the cost prediction value that conforms to the actual situation.

[0082] S1071. In practical applications, if the consistency verification fails, extract the adjustment record set based on the local support adjustment knowledge graph, such as data on the adjustment of the retaining wall thickness and the change in the anchor rod length. Convert the records into multi-dimensional feature vectors including support type, adjustment range, and construction technology through the graph feature extractor. Then use the deep neural network to train 2000 groups of historical adjustment data. This network adopts a multi-layer perceptron structure, receives the feature vector through the input layer, and outputs the updated cost impact factor after being processed by three hidden layers. For example, the impact factor of the retaining wall thickness adjustment is increased from 0.7 to 0.85. Subsequently, compare the predicted value with the actual value through the feature verification function, generate the deviation distribution diagram using the cost deviation calculator, and find that the deviation of the retaining wall adjustment is concentrated between 15% and 30%, indicating that the model needs to be further optimized.

[0083] For the deviation distribution diagram, when the prediction deviation exceeds the preset threshold, such as 15%, classify the deviation into high, medium, and low categories through the deviation classifier, optimize the structure of the knowledge graph in combination with the graph update function, evaluate the influence of each node using the node importance calculator. For example, the importance of the retaining wall thickness node reaches 0.85, and the importance of the anchor rod parameter node is 0.78. Then redistribute the weights according to the importance through the weight distribution optimizer, adjust the weight of the retaining wall thickness node from 0.6 to 0.75, generate a new weight distribution table to ensure that the graph more accurately reflects the impact of the adjustment on the cost. Subsequently, refine the feature classification through the feature reconstruction function, such as classifying the retaining wall thickness adjustment into small and large adjustments, corresponding to cost fluctuations of 10% to 20% and 20% to 35% respectively.

[0084] During the optimization process, the construction difficulty index is introduced as a new feature through the feature expander. This index is calculated based on geological conditions and construction environment and is positively correlated with cost fluctuations. For example, for every 0.1 increase in complexity, the cost deviation increases by approximately 5%. Based on the expanded feature set, a random forest classifier is used for training. This classifier takes adjusted parameters and construction difficulty as inputs and optimizes the parameters through 5-fold cross-validation. For example, the learning rate and tree depth are adjusted. After training, the accuracy of predicting minor adjustment working conditions reaches 94% and that of major adjustments reaches 88%, generating a new version of the feature model, significantly enhancing the prediction ability under complex working conditions.

[0085] S1072. In the embodiment of the present invention, based on the new weight distribution table and the expanded feature set, the final cost prediction value is calculated through the cost prediction function. For example, when the thickness of the retaining wall increases by 200 mm, considering the construction difficulty, the predicted cost increase is 32%, with a deviation of only 1.6% from the actual value of 31.5%. When the length of the anchor rod increases by 1 m, the predicted increase is 25%, with a deviation of 0.8% from the actual value of 24.8%. Then, the prediction result validator is used to evaluate the accuracy of the model to ensure that the deviation is controlled within the preset range. This iterative optimization method enhances the adaptability of the cost prediction model to complex working conditions through multi-dimensional feature analysis and intelligent algorithm training.

[0086] In a specific scenario, if the initial prediction shows that the adjustment of the retaining wall thickness results in a cost deviation of 28%, after updating the atlas and feature set, the deviation is reduced to less than 2%, indicating that the model can dynamically adapt to the changes in the retaining wall adjustment, ensuring the coincidence degree between the predicted value and the actual engineering data, providing a more accurate decision-making basis for the budget management of deep foundation pit projects, and the threshold and algorithm parameters can be adjusted according to actual needs to further optimize the effect.

[0087] The present invention provides a deep learning-based engineering cost prediction system, mainly including:

[0088] An initial abnormal feature set extraction module, used to extract geological abnormal types and the amplitude of retaining wall adjustment, and determine the initial abnormal feature set by combining the data triggering abnormal conditions in real-time monitoring;

[0089] An abnormal distribution data set generation module, used to quantify and classify geological abnormal types through the initial abnormal feature set, identify the abnormal distribution density, separate abnormal points within the foundation pit area according to the abnormal distribution density to form an abnormal distribution data set, and identify the geological abnormal types and the amplitude of retaining wall adjustment therein;

[0090] A local retaining wall adjustment knowledge graph construction module, used to construct a knowledge graph through geological abnormal types and the amplitude of retaining wall adjustment, update and map the knowledge graph into a graph structure, identify the associated adjustment rules and node weight distributions in the graph structure, and determine the local retaining wall adjustment knowledge graph;

[0091] The preliminary cost fluctuation range determination module is used to analyze the local support adjustment knowledge graph, obtain the correlation relationship between various geological anomaly types and the support adjustment amplitude, obtain the inference path length and the priority of inference rules according to the correlation relationship, identify the cost impact factor of the geological anomaly type on the support adjustment amplitude, and obtain the preliminary cost fluctuation range;

[0092] The unified cost feature set determination module is used to extract the cost breakdown details from the cost budget and settlement documents, count the historical support parameter adjustment data, obtain the parameter adjustment frequency, and combine the preliminary cost fluctuation range to determine the unified cost feature set;

[0093] The consistency check value calculation module is used to analyze and quantify the degree of confusion of the semantic differences in describing cost features in the unified cost feature set on the association rules, adjust the node weight distribution in the knowledge graph according to the degree of confusion, process the unified cost feature set, and obtain the consistency check value between the support adjustment amplitude and the cost fluctuation range;

[0094] The final cost prediction value determination module is used to, if the consistency check value does not meet the requirements, update the local support adjustment knowledge graph and recalculate the cost impact factor to determine the correction value of the cost prediction, compare the correction value with the actual data, and if the deviation is greater than the preset threshold, adjust the node weight distribution, update and train the unified cost feature set, and determine the final cost prediction value.

[0095] The above embodiments are only one of the preferred embodiments of the present invention and should not be used to limit the protection scope of the present invention. Any meaningless changes or polishing made on the main design idea and spirit of the present invention, as long as the technical problems solved are still the same as those of the present invention, should be included in the protection scope of the present invention.

Claims

1. A deep learning-based engineering cost prediction method, characterized in that, The method includes: Extracting the geological anomaly types and the support adjustment amplitudes, and combining the data of the real-time monitoring triggering anomaly conditions to determine the initial anomaly feature set; Through the initial anomaly feature set, quantitatively classifying the geological anomaly types and identifying the anomaly distribution density, separating the anomaly points within the foundation pit area according to the anomaly distribution density to form an anomaly distribution data set, and identifying the geological anomaly types and the support adjustment amplitudes therein; Through the geological anomaly types and the support adjustment amplitudes, constructing a knowledge graph, updating and mapping the knowledge graph into a graph structure, identifying the associated adjustment rules and the node weight distributions in the graph structure, and determining the local support adjustment knowledge graph; Analyzing the local support adjustment knowledge graph, obtaining the association relationships between the various geological anomaly types and the support adjustment amplitudes, obtaining the inference path lengths and the inference rule priorities according to the association relationships, identifying the cost impact factors of the geological anomaly types on the support adjustment amplitudes, and obtaining the preliminary cost fluctuation range; Extracting the cost breakdown details from the cost budget and settlement documents, counting the historical support parameter adjustment data, obtaining the parameter adjustment frequency, and combining with the preliminary cost fluctuation range to determine the unified cost feature set; Analyzing and quantifying the confusion degree of the semantic differences in describing the cost features in the unified cost feature set on the association rules, adjusting the node weight distributions in the knowledge graph according to the confusion degree, processing the unified cost feature set, and obtaining the consistency verification value of the support adjustment amplitude and the cost fluctuation range; If the consistency verification value does not meet the requirements, update the local support adjustment knowledge graph, recalculate the cost impact factors to determine the correction value of the cost prediction, compare the correction value with the actual data, if the deviation is greater than the preset threshold, adjust the node weight distributions, update and train the unified cost feature set, and determine the final cost prediction value.

2. The method according to claim 1, wherein The extraction of the geological anomaly types and the support adjustment amplitudes, and the combination of the data of the real-time monitoring triggering anomaly conditions to determine the initial anomaly feature set includes: According to the real-time displacement data of the surrounding rock measurement points collected by the rock layer displacement monitoring device at the preset sampling interval, performing noise elimination and outlier removal on the original data to obtain the preprocessed displacement data; For the preprocessed displacement data, calculating the deformation acceleration and the deformation trend coefficient of the surrounding rock, and combining with the surrounding rock level parameters to obtain the anomaly feature vector; Scanning and obtaining the image of the rock mass structure around the anomaly recording point through a geological structure detector, performing noise reduction and enhancement, and identifying the fault joint zone category according to the rock mass structure feature library to obtain the rock mass integrity score; Using a support vector machine classifier to analyze the anomaly feature vector and the stress sensor data to determine the initial anomaly feature set.

3. The method according to claim 1, wherein The quantification and classification of the geological anomaly types through the initial anomaly feature set and the identification of the anomaly distribution density, separating the anomaly points within the foundation pit area according to the anomaly distribution density to form an anomaly distribution data set, and identifying the geological anomaly types and the support adjustment amplitudes therein includes: Obtaining the anomaly point density distribution map in the foundation pit area range map, and the anomaly point density distribution map contains the density distribution data of the initial anomaly feature set; Based on the abnormal point density distribution map, compare the relationship between the point density value and the geological structure boundary line through the density clustering algorithm to obtain the clustering result of abnormal points; For the clustering result of abnormal points, establish a geological anomaly quantification function based on geological structure characteristic indicators to obtain the quantification value of the anomaly degree; Use the quantification value of the anomaly degree to construct an abnormal point feature vector, and generate an abnormal distribution data set through the feature vector clustering algorithm; Identify the abnormal feature type according to the abnormal distribution data set; Combine the abnormal density distribution values of each sub-region included in the abnormal distribution data set to identify key abnormal points, and determine the support adjustment range based on the historical support adjustment records.

4. The method according to claim 1, wherein Based on the geological anomaly type and the support adjustment range, construct a knowledge graph, update and map the knowledge graph into a graph structure, identify the associated adjustment rules and node weight distributions in the graph structure, and determine the local support adjustment knowledge graph, including: Construct a knowledge graph of geological anomaly and support adjustment, use a graph structure to represent the corresponding relationship between geological anomaly and support adjustment, and obtain the initial graph structure; Use the mapping update method to refresh the initial graph structure in combination with the adjustment range data to generate an updated knowledge graph; Analyze the weights and weight distributions of the nodes in the updated knowledge graph, calculate the quantification values of each node, and obtain the node weight distribution result; For the node weight distribution result, identify the logical connection between support adjustment and anomaly type, and generate an associated rule set; According to the associated rule set, if the node weight of the local support area exceeds the preset threshold, generate a local support adjustment plan in combination with the adjustment range to obtain a set of adjustment plans; Through the graph structure of the set of adjustment plans and the updated knowledge graph, update the weight distribution and associated rules to obtain the local support adjustment knowledge graph.

5. The method according to claim 4, wherein It also includes: Collect actual engineering cases and extract the corresponding relationship between geological anomaly types and support adjustment measures, convert the corresponding relationship into a rule form to obtain associated adjustment rules, statistically analyze the abnormal distribution data set to obtain the frequencies of different geological anomaly types and support adjustment measures, and numerically simulate the effects of different support plans under different geological anomaly conditions. Assign initial weights to each node in the graph structure through the effects of different support plans to obtain the node weight distribution, specifically including: Extract the corresponding relationship between geological anomaly types and support adjustment measures from actual engineering data to obtain a set of associated adjustment rules; Calculate the occurrence frequency values of different geological anomaly types and the distribution proportion of support adjustment measures according to the set of associated adjustment rules, and generate an abnormal and adjustment associated frequency distribution map; Construct a support plan combination library for the abnormal and adjustment associated frequency distribution map, and calculate the support structure force distribution function through preset support structure parameters to obtain the initial support effect index; Establish a geomechanical numerical calculation model according to the initial support effect index, obtain the surrounding rock stress field distribution under different support plans through hierarchical progressive calculation, and calculate the support adaptability evaluation value; Establish a connection relationship between the nodes of the graph structure for the support adaptability evaluation value, and use a preset connection degree threshold to quantitatively calculate the association strength between nodes to generate a graph structure connection degree distribution map; Based on the graph structure connectivity distribution diagram, the depth-first search algorithm is used to traverse the support effect index space and calculate the weights of the nodes on the search path; According to the node search path and support adaptability evaluation value, the normalized weight of each node in the graph structure is calculated through the weight distribution function to obtain the node weight distribution.

6. The method according to claim 1, wherein The local support adjustment knowledge graph is analyzed to obtain the correlation between the geological anomaly types and the support adjustment range, and the reasoning path length and the reasoning rule priority are obtained according to the correlation relationship, and the cost impact factor of the geological anomaly type on the support adjustment range is identified to obtain the preliminary cost fluctuation range, including: The shortest path length between the abnormal type node and the support adjustment node is calculated by the graph searcher, and the initial knowledge path set is obtained by using the breadth-first search algorithm; According to the initial knowledge path set, the frequency of occurrence of nodes in the path is counted, and a node weight distribution diagram is obtained by combining the connection strength between nodes; For the node weight distribution graph, a path optimization function is used to score and sort the paths, and an optimized knowledge path set is obtained through a path length threshold filter; According to the optimized knowledge path set, an interval partitioning function is used to establish a cost fluctuation interval, the cost impact prediction model is obtained by training a random forest regressor, and the cost impact coefficients corresponding to different anomaly types are calculated through anomaly type feature vectors to generate a preliminary cost fluctuation interval; it also includes: judging the distribution characteristics of geological anomalies through reasoning path length analysis, determining the classification results of the anomaly types based on the distribution characteristics and the priority of the reasoning rules, obtaining the preliminary demand for support adjustment from the classification results, obtaining the initial range of the adjustment amplitude, calculating the impact factor for the initial range, judging the specific value of the amplitude change, using the support vector machine algorithm to determine the cost impact prediction model, and outputting the cost impact factor of the geological anomaly on the cost.

7. The method according to claim 1, characterized in that, The cost split details are extracted from the cost budget and settlement documents, historical support parameter adjustment data are counted, parameter adjustment frequency is obtained, and a unified cost feature set is determined in combination with the preliminary cost fluctuation range, including: Obtain the support component number and construction process number according to the cost budget document to form a cost project database; For the cost project database, the support cost benchmark database is used to perform project cost comparison, and the support cost proportion distribution data is obtained through a preset comparison calculation function; According to the support cost proportion distribution data, the parameter adjustment type frequency is calculated using an adjustment parameter counter, and a parameter adjustment frequency distribution diagram is obtained through a frequency sorting function; According to the parameter adjustment frequency distribution diagram, a random forest regressor is used to establish a mapping relationship between parameter adjustment amount and cost fluctuation, and a cost fluctuation range table is obtained through training with historical engineering cost data; According to the cost fluctuation interval table, the interval feature extractor is used to calculate the fluctuation interval feature parameters, and the fluctuation intervals are classified and labeled in combination with the preset feature threshold to establish a fluctuation feature library; For the fluctuation feature library and cost split details table, a unified cost feature set is generated based on preset classification standards.

8. The method according to claim 1, characterized in that In the analysis and quantification of the unified cost feature set, the degree of confusion of the semantic differences in describing cost features for association rules is described, and the node weight distribution in the knowledge graph is adjusted according to the degree of confusion, and the unified cost feature set is processed to obtain the consistency verification value of the support adjustment range and the cost fluctuation range, including: Receiving the feature description information in the unified cost feature set, where the feature description information is generated from the cost feature set; Calculating the semantic similarity between features using a word vector similarity calculator according to the feature description information, and obtaining a feature confusion comparison table through comparison with a similarity threshold; Calculating the degree of confusion between feature vectors for the feature confusion comparison table using a deep neural network, and obtaining a rule confusion matrix through training the deep neural network based on the labeled semantic data set; Processing the rule confusion matrix using a node weight update function, adjusting the node weights in the knowledge graph, and obtaining a new weight distribution map through a preset weight threshold; For the new weight distribution map, using a random forest regressor to establish a cost fluctuation prediction function, and generating a consistency verification value of the support adjustment range and the cost fluctuation range through a verification threshold calculator according to the cost fluctuation prediction result.

9. The method according to claim 1, characterized in that, If the consistency verification value does not meet the requirements, then update the local support adjustment knowledge graph, recalculate the cost impact factor to determine the correction value of the cost prediction, compare the correction value with the actual data, and if the deviation is greater than the preset threshold, then adjust the node weight distribution, update and train the unified cost feature set, and determine the final cost prediction value, including: Obtaining an adjustment record set according to the local support adjustment knowledge graph, and processing the adjustment record set to obtain an adjustment feature vector; For the adjustment feature vector, using a deep neural network to train the historical adjustment data to obtain an updated cost impact factor; Processing the updated cost impact factor through a cost deviation calculator to obtain a prediction deviation distribution map, and if the prediction deviation distribution map exceeds the preset deviation threshold, then generating a deviation classification result; For the deviation classification result, using a weight distribution optimizer to allocate weights to the nodes in the knowledge graph, and generating a new weight distribution table through a preset weight adjustment rule; According to the new weight distribution table, using a random forest classifier for feature training and calculating the final cost prediction value.

10. A project cost prediction system based on deep learning, characterized in that, The system includes: An initial abnormal feature set extraction module, which is used to extract geological anomaly types and support adjustment ranges, and determine an initial abnormal feature set by combining data that triggers abnormal conditions in real-time monitoring; An abnormal distribution data set generation module, which is used to quantify and classify geological anomaly types through the initial abnormal feature set, identify the abnormal distribution density, separate abnormal points in the foundation pit area according to the abnormal distribution density to form an abnormal distribution data set, and identify the geological anomaly types and support adjustment ranges therein; A local support adjustment knowledge graph construction module, which is used to construct a knowledge graph through geological anomaly types and support adjustment ranges, update and map the knowledge graph into a graph structure, identify the associated adjustment rules and node weight distribution in the graph structure, and determine the local support adjustment knowledge graph; The preliminary cost fluctuation range determination module is used to analyze the local support adjustment knowledge graph, obtain the correlation relationship between various geological anomaly types and the support adjustment amplitude, obtain the inference path length and the priority of inference rules according to the correlation relationship, identify the cost impact factor of the geological anomaly type on the support adjustment amplitude, and obtain the preliminary cost fluctuation range; The unified cost feature set determination module is used to extract the cost breakdown details from the cost budget and settlement documents, count the historical support parameter adjustment data, obtain the parameter adjustment frequency, and combine the preliminary cost fluctuation range to determine the unified cost feature set; The consistency check value calculation module is used to analyze and quantify the degree of confusion of the semantic differences in describing cost features in the unified cost feature set on the association rules, adjust the node weight distribution in the knowledge graph according to the degree of confusion, process the unified cost feature set, and obtain the consistency check value between the support adjustment amplitude and the cost fluctuation range; The final cost prediction value determination module is used to, if the consistency check value does not meet the requirements, update the local support adjustment knowledge graph and recalculate the cost impact factor to determine the correction value of the cost prediction, compare the correction value with the actual data, and if the deviation is greater than the preset threshold, adjust the node weight distribution, update and train the unified cost feature set, and determine the final cost prediction value.

Images