A CBR-based green construction decision method for mountain railway tunnels

Abstract

The application discloses a mountainous railway tunnel green construction decision-making method based on CBR, and relates to the field of construction decision-making, and comprises the following steps: step 1: integrating the scene attribute system of the construction site of the hole mouth and the hole body, obtaining the case representation including the construction scene attribute, the attribute value and the green construction measure, and generating a three-dimensional scene through parameterization of Rhino software; step 2: for the case determined as data fuzzy or missing due to special terrain or weather, based on the standard completeness case index, the fuzzy degree coefficients of the geological conditions, the environmental sensitive area, the construction machinery type, the energy consumption index and the material transportation path parameters of the case in the scene attribute system are evaluated; a fuzzy data virtual filling mechanism is set, the most similar data in the complete case is found based on the characteristic label matching to carry out virtual replacement, the data missing and fuzzy in the real scene are effectively dealt with, and the similarity between the non-complete characteristic cases is accurately calculated instead of simply discarding the cases with missing characteristics.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of construction decision-making, in particular to a mountain railway tunnel green construction decision-making method based on CBR. BACKGROUND

[0002] Mountain railway tunnel construction has long been faced with severe challenges such as high energy consumption and high pollution, and the implementation of green construction has become an inevitable choice for the sustainable development of the industry. However, the realization of green construction of tunnel engineering highly depends on the green construction measures selected. Due to the complex and changeable construction scene of mountain railway tunnel engineering, such as sudden gushing water, high altitude, high cold and oxygen deficiency, the geological conditions and environmental requirements are different, and the traditional green construction measure decision-making is made by the on-site technical personnel according to the current tunnel engineering construction scene through qualitative analysis to select the green construction measure, which not only requires high knowledge reserve and practical experience of the construction personnel, but also aggravates the problems of blind and lack of targeted green construction measure decision-making.

[0003] However, the complex and changeable construction scene and the green construction measure decision-making present a certain suitability relationship, for example: the side slope of the mountain railway tunnel portal is high and steep, which is easy to cause soil erosion, and in such a construction scene, the green construction measure of preventing and controlling soil erosion needs to be taken at the tunnel portal; in the high-altitude and high-cold area tunnel construction, the vegetation is dense in the necessary construction area, and in such a scene, the green construction measure of protecting the vegetation needs to be taken. At the same time, there are a large number of similar construction scenes in the green construction of mountain railway tunnel engineering, but the green construction cases of completed projects are not used as data support for auxiliary decision-making in the corresponding green construction measure decision-making, resulting in a blank in providing green construction measures according to the similar construction scenes in the past.

[0004] The existing mountain railway tunnel engineering green construction measure decision-making still has the following defects:

[0005] The related case data of green construction measures is scattered, and the case system is not combed, managed and utilized; in the face of complex and changeable construction scenes, the green construction measures are mainly selected by the subjective experience of construction personnel, and lack of applicability and pertinence; the construction environment in the mountain area is poor, and the terrain and weather often lead to incomplete or fuzzy data records, so the traditional method either directly discards the case or simply fills it in, which seriously restricts the usability of the system in the real scene. SUMMARY

[0006] (I) Technical problems solved

[0007] In view of the above-mentioned defects existing in the prior art, the application provides a mountain railway tunnel green construction decision-making method based on CBR, which can effectively solve the problems of the prior art.

[0008] (II) Technical scheme

[0009] To achieve the above object, the present application is realized by the following technical solutions:

[0010] The application discloses a mountainous railway tunnel green construction decision-making method based on CBR, and comprises the following steps:

[0011] Step 1: integrate the scene attribute system of the construction site of the hole mouth and the hole body, obtain the case representation including the construction scene attribute, the attribute value and the green construction measure, and generate a three-dimensional scene through parameterization of Rhino software;

[0012] Step 2: for the case determined to be fuzzy or missing due to special terrain or weather, based on the standard completeness case index, the fuzzy degree coefficients of the geological conditions, the environmental sensitive area, the construction machinery type, the energy consumption index and the material transportation path parameter of the case in the scene attribute system are evaluated, and the indexes with the comprehensive fuzzy degree coefficient higher than the preset threshold value are selected;

[0013] Step 3: the characteristic label of the missing data under the selected index is extracted, matched with the complete characteristic label of the same index in the case library, virtual supplementary data is generated through simulation replacement, temporarily stored in the case library and added with a fuzzy mark;

[0014] Step 4: mixed coding is performed on the numerical, category and Boolean heterogeneous characteristics in the case library, the category characteristics are subjected to One-Hot coding containing a missing mark, and a binary mask vector is constructed to identify the effectiveness of the characteristics;

[0015] Step 5: a twin neural network structure is constructed, the preprocessed feature vector is input, processed through 3 layers of full connection layer, GELU activation function, layer normalization and Dropout layer, and an L2 normalized embedding vector is output;

[0016] Step 6: based on the mask representation of the query case, the candidate cases without exclusive characteristics in the mask vector of the training data are screened, and the intersection of the effective characteristics of the query case and the candidate cases is calculated;

[0017] Step 7: the Euclidean distance of the embedding vectors of the query case and the candidate cases is calculated, the measure sorting output with the distance less than the dynamic threshold value is extracted, the new case verified by construction is fed back to the case library, and the virtual supplementary data is replaced to trigger model retraining.

[0018] Further, the evaluation process of the fuzzy degree coefficient in step 2 is as follows:

[0019] For the target index, the total number of data subitems that must be included in the standard complete case is counted as the reference quantity, the percentage of the number of data subitems actually missing in the target index in the current case in the reference quantity is calculated as the basic missing rate;

[0020] According to the index type, the geological conditions, the environmental sensitive area, the energy consumption index, and the construction machinery and transportation path weight are preset.

[0021] The base missing rate is multiplied by the weight factor of the corresponding index to generate a weighted missing rate. Based on historical case library statistics, when the target index is missing, the data anomaly probability of the associated index is used as the fluctuation coefficient. The final ambiguity coefficient = weighted missing rate x (1 + fluctuation coefficient).

[0022] Further, the logic of the characteristic label matching in step 3 is:

[0023] The geological conditions, environmental sensitive area, construction machinery type, energy consumption index, and material transportation path index of each complete case in the case library are pre-defined with characteristic labels.

[0024] According to the selected high ambiguity index type, the characteristic label requirements of the missing data under the target index are extracted, and complete cases with the same construction site and index type are selected from the case library. The matching degree of the characteristic label and the missing label is calculated. The matching degree is calculated based on the semantic similarity of the label description text and the numerical interval overlap degree comprehensive evaluation;

[0025] The characteristic label with the highest matching degree is selected for data simulation replacement. The source label case number and replacement timestamp are retained when generating virtual supplementary data, and the corresponding case in the case library is marked as virtual supplementary state.

[0026] Further, the geological condition index corresponding characteristic label includes rock layer permeability grade and fault fracture zone range grade, the environmental sensitive area index corresponding characteristic label includes water source straight line distance and core vegetation coverage rate, the construction machinery type index corresponding characteristic label includes equipment rated power range and fuel type category, the energy consumption index corresponding characteristic label includes power supply source and unit engineering quantity oil consumption interval, and the material transportation path parameter index corresponding characteristic label includes maximum slope angle and minimum turning radius.

[0027] Further, the virtual supplementary data is marked with a semi-transparent identifier in the case library. When searching, real cases are called first. When real data is obtained, virtual data is automatically replaced and fuzzy marks are cleared, triggering incremental training of the twin neural network.

[0028] Further, the execution process of the mixed encoding in step 4 includes: normalizing and scaling all numerical features in the case library to convert the value range to the interval of 0 to 1; performing one-hot encoding on the category features to generate an independent binary vector for each category value, and creating a special identification bit for the missing value; converting the Boolean feature directly into a numerical representation of 1 or 0 to construct an encoding vector; generating a binary identification vector with the same feature dimension as each case, in which the effective feature bit is marked as 1 and the missing feature bit is marked as 0 to construct a mask vector; and splicing the resulting encoding vector and the mask vector in order according to the feature dimension to form a joint feature vector for the final input neural network.

[0029] Further, the twin neural network structure in step 5 is composed of two artificial neural networks with the same structure and shared weights, and the loss function selects a triplet loss, which learns a discriminative embedding space by minimizing the distance between the anchor and the positive sample while maximizing the distance between the anchor and the negative sample, and uses squared Euclidean distance to enhance gradient stability, and its expression is:

[0030] ;

[0031] In the formula, represents the loss function, a, p and n represent the embedding vectors of the anchor, positive sample and negative sample respectively, is the interval parameter, and N represents the total number of samples.

[0032] Further, the training process of the twin neural network model in step 5 adopts hierarchical cross-validation and full-volume training, and uses cosine annealing to schedule the learning rate to avoid local optimum, and its expression is:

[0033] ;

[0034] In the formula, represents the actual learning rate of the neural network at the t-th iteration, represents the current iteration step.

[0035] Further, in the process of calculating the intersection of the effective features of the query case and the candidate case in step 6, the features that exist in both directions are indicated by dynamic masks to ensure that only the features that exist in common are compared, and its expression is:

[0036] ;

[0037] In the formula, represents the intersection of the effective features, represents the mask vector of the query case, represents the mask vector of the candidate case.

[0038] Further, the dynamic threshold in step 7 is set according to the P90 quantile of the similarity distribution of the training set, and the output measure list is arranged in ascending order of Euclidean distance, if the distance exceeds the specified threshold, it is skipped, otherwise it is added to the measure list of the valid case.

[0039] (Three) beneficial effects

[0040] Compared with the known prior art, the technical scheme provided by the application has the following beneficial effects:

[0041] By setting the fuzzy data virtual filling mechanism, the completion priority is determined by the blur coefficient evaluation, the most similar data is found from the complete case based on the characteristic label matching to perform virtual replacement, and the case library is temporarily supplemented, so that the incomplete case is effectively utilized, the decision-making process is ensured not to be interrupted, and the data loss and blur in the real scene are effectively dealt with.

[0042] Through the cooperation of the mask mechanism, the twin network can accurately calculate the similarity between non-complete feature cases, instead of simply discarding the cases with missing features. By using the mask vector, the similarity of the effective features shared by the query case and the candidate case is calculated, which perfectly adapts to the query scene with any missing features, so that the decision-making suggestion is highly matched with the current complex actual construction scene, and the accuracy of the measure recommendation is improved.

[0043] By recommending green measures highly matched with the scene, the energy consumption and carbon emissions in the construction process are directly reduced from the source, scientific decision-making avoids rework, resource waste or environmental accidents caused by improper measures, saves construction cost, and achieves direct economic benefits. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical scheme in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0045] Figure 1 The flowchart of the application;

[0046] Figure 2 The three-dimensional model diagram of the green construction measure case library construction process in the application;

[0047] Figure 3 The training loss curve and accuracy curve diagram of the embodiment of the application;

[0048] Figure 4 The cross-validation average training loss, accuracy and validation accuracy curve diagram of the embodiment of the application. DETAILED DESCRIPTION

[0049] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will be combined with the accompanying drawings of the embodiments of the present application to make a clear and complete description of the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work belong to the scope of protection of the present application.

[0050] The present application will be further described below in combination with the embodiments.

[0051] A CBR-based green construction decision method for mountain railway tunnels in the present embodiment, as shown in FIG. 1, includes the following steps: Figures 1-4

[0052] Step 1: Integrate the scene attribute system of the construction site of the tunnel portal and the tunnel body, obtain the case representation including the construction scene attribute, the attribute value and the green construction measure, and generate a three-dimensional scene through the parameterization of Rhino software.

[0053] Step 2: For the cases determined to be fuzzy or missing due to special terrain or weather, based on the standard completeness case index, evaluate the fuzzy degree coefficient of the geological condition, the environmental sensitive area, the construction machinery type, the energy consumption index and the material transportation path parameter of the case in the scene attribute system, and select the index with the comprehensive fuzzy degree coefficient higher than the preset threshold; the evaluation process of the fuzzy degree coefficient is as follows:

[0054] For the target index, count the total number of data sub-items that must be included in the standard complete case as the reference quantity, calculate the percentage of the number of data sub-items actually missing in the current case in the reference quantity as the basic missing rate;

[0055] According to the index type, preset the weights of the geological condition, the environmental sensitive area, the energy consumption index, the construction machinery and the transportation path;

[0056] Multiply the basic missing rate by the weight factor of the corresponding index to generate a weighted missing rate, and based on the historical case library statistics, when the target index is missing, the data anomaly probability of the associated index is taken as the fluctuation coefficient, and finally the fuzzy degree coefficient = weighted missing rate x (1+ fluctuation coefficient).

[0057] Step 3: Extract the characteristic label of the missing data under the selected index, match it with the complete characteristic label of the same index in the case library, generate virtual supplementary data through simulation replacement, and temporarily store it in the case library and add a fuzzy mark.

[0058] ​Step 4: Perform hybrid encoding on the numerical, categorical and Boolean heterogeneous features in the case base. One-Hot encoding with missing value markers is used for categorical features, and a binary mask vector is constructed to identify feature validity. The hybrid encoding process includes: normalizing and scaling all numerical features in the case base to convert their value range to the 0 to 1 interval; performing One-Hot encoding on categorical features to generate independent binary vectors for each category value and creating a dedicated identification bit for missing values; converting Boolean features directly into numerical representations of 1 or 0 to construct an encoding vector; generating a binary identification vector for each case with the same feature dimension as the mask vector, where valid feature bits are marked as 1 and missing feature bits are marked as 0 to construct a mask vector; concatenating the resulting encoding vector and mask vector in the order of feature dimensions to form a joint feature vector as the final input to the neural network.

[0059] Step 5: Construct a Siamese neural network structure, input the pre-processed feature vector, pass it through 3 fully connected layers, GELU activation function, layer normalization and Dropout layer, and output an L2 normalized embedding vector.

[0060] The Siamese neural network structure is composed of two artificial neural networks with the same structure and shared weights, and its loss function selects a triplet loss. By minimizing the distance between the anchor and the positive sample, while maximizing the distance between the anchor and the negative sample, a discriminative embedding space is learned, and the squared Euclidean distance is used to enhance gradient stability. Its expression is:

[0061] ;

[0062] In the formula, represents the loss function, a, p and n represent the embedding vectors of the anchor, positive sample and negative sample, is the margin parameter, and N represents the total number of samples;

[0063] The training process of the Siamese neural network model adopts hierarchical cross-validation and full-volume training, and uses cosine annealing to schedule the learning rate to avoid local optimum. Its expression is:

[0064] ;

[0065] In the formula, represents the actual learning rate of the neural network at the t-th iteration, represents the current iteration step.

[0066] Step 6: Based on the mask representation of the query case, filter out candidate cases that do not have exclusive features in the mask vector of the training data, and calculate the intersection of valid features between the query case and the candidate cases.

[0067] In the process of calculating the intersection of the effective features of the query case and the candidate case, the features existing in both directions are indicated by dynamic masks to ensure that only the features existing in common are compared, and the expression is:

[0068] ;

[0069] In the formula, represents the intersection of the effective features, represents the query case mask vector, represents the candidate case mask vector.

[0070] Step 7: Calculate the Euclidean distance of the embedding vectors of the query case and the candidate case, extract the measure ranking output with a distance less than the dynamic threshold, and feed back the applicable new case to the case library after construction verification. The virtual supplementary data triggers model retraining after being replaced; the dynamic threshold is set according to the P90 quantile of the similarity distribution of the training set, and the output measure list is arranged in ascending order of Euclidean distance. If the distance exceeds the specified threshold, skip, otherwise add to the measure list of the effective case.

[0071] Compared with the prior art, the present application can intelligently process data missing, fuse multi-source heterogeneous information and perform deep similarity matching, ensure information integrity by quantifying the ambiguity of evaluation indicators and generating virtual data, uniformly process complex features using hybrid encoding and mask mechanism, and perform accurate case retrieval in embedding space with the help of twin neural networks. Finally, through dynamic threshold screening and case library continuous feedback optimization, the accuracy and self-adaptive ability of green construction decision of mountainous railway tunnel in uncertain environment are improved.

[0072] On the other hand, the embodiment provides a logic for characteristic label matching, specifically:

[0073] The geological conditions, environmental sensitive areas, construction machinery types, energy consumption indicators and material transportation path indicators of each complete case in the case library are pre-defined with characteristic labels; the characteristic labels corresponding to the geological condition indicators include rock layer permeability grade and fault fracture zone range grade, the characteristic labels corresponding to the environmental sensitive area indicators include water source straight line distance and core vegetation coverage, the characteristic labels corresponding to the construction machinery type indicators include equipment rated power range and fuel type category, the characteristic labels corresponding to the energy consumption indicators include power supply source and unit engineering quantity oil consumption interval, and the characteristic labels corresponding to the material transportation path parameter indicators include maximum slope angle and minimum turning radius;

[0074] According to the selected high-ambiguity indicator type, the characteristic label requirements of the missing data under the target indicator are extracted, complete cases with the same construction site and the same indicator type are screened in the case library, and the matching degree of their characteristic labels and missing labels is calculated. The matching degree is calculated based on the semantic similarity of the label description text and the numerical interval overlap degree.

[0075] The highest matching characteristic label is selected for data simulation replacement, the source label case number and replacement timestamp are retained when generating virtual supplementary data, and the corresponding case in the case library is marked as virtual supplementary state; the virtual supplementary data is marked with a semi-transparent identifier in the case library, and the real case is called first when searching, and when the real data is obtained, the virtual data is automatically replaced and the blurred mark is cleared, triggering the incremental training of the twin neural network.

[0076] Compared with the prior art, by constructing a characteristic label system and based on semantic and numerical interval double matching degree evaluation, the traceable generation of virtual supplementary data in the case of data missing is realized, and at the same time, through the state marking and incremental updating mechanism, the dynamic optimization of the case library data and the continuous improvement of the reliability of the decision are ensured.

[0077] Embodiment 2 provides a green construction measure hierarchical retrieval system for mountainous railway tunnel engineering, comprising:

[0078] Collect cases that meet the conditions:

[0079] Select the query case, adopt the same data preprocessing mode of the green construction measure case library scene attribute characteristics, and output the query case mask. Then, all cases in the case library are traversed, and cases in the case library that contain certain scene attribute characteristics but the query case does not contain are filtered out.

[0080] The selection of the query case of the present application is a railway tunnel construction engineering project. The project is located in a high-altitude area with low oxygen content and high coldness. The average altitude of the work site is high, and it is located in a high-altitude area with low oxygen content and high coldness. Under the conditions of high coldness and oxygen deficiency, the physical function of the construction personnel decreases, the work efficiency decreases, the power of the mechanical equipment decreases, and the requirements for personnel health protection and mechanical equipment performance are higher. The surface water system along the line is developed, and the vegetation is mainly shrubs and weeds, which are sensitive and fragile, have weak anti-interference ability and self-repairing ability, and have long natural recovery time. The proportion of adverse geology is large, the surrounding rock grade is poor, the construction risk is high, it is close to residential areas and the line along the line is surrounded by many mountains, and there are many wild animals, so the noise control requirements are high. Therefore, the main points of green construction of the project include: energy saving and efficiency improvement of mechanical equipment, protection of human resources, soil and water conservation, vegetation protection, noise pollution prevention and control, etc. According to the hole and body scene attribute system formed in step 1, the scene attribute characteristics of the query case are extracted, as shown in the following table.

[0081] Query case scene attribute characteristics:

[0082] Scenario attribute Attribute value Elevation 3000m Burial depth 380m Self-stability of surrounding rock Poor stability Tunnel length 2275.55m Air temperature 3.7℃ Natural slope of portal 35° Vegetation coverage 0.78 Construction area Seismic zone Groundwater outflow state 100 min·10 m Construction wastewater generation 3809 m³ / d Near residential area 160m Near surface water body 280m Near wildlife habitat 553m Dust concentration 35 mg / m³ Slope rock-soil type Type II rock-soil Slope stability Stable Slope gradient 33.69° Slope height 6.84m

[0083] Green construction measure generation and evaluation

[0084] According to the effective feature intersection of the calculation query case and the candidate case, the feature vectors of different scene attribute combinations are converted based on the common existing features, the embedded vectors are obtained by inputting the trained twin neural network in step 5, the Euclidean distance is calculated, and the green construction measure list suitable for the query case is output, and the smaller the distance is, the more similar it is, as shown in the following table.

[0085] Green construction measure list

[0086] Similar case Euclidean distance Green construction measures adopted Portal scenario 4942 0.00 1. Excavate shock-absorbing ditch before blasting operation 2. Cover the surface of the blasting area and sprinkle water 3. Pre-splitting blasting technology Portal scenario 4522 0.00 Diesel generator sound insulation treatment Air compressor sound insulation treatment Crusher sound insulation treatment Portal scenario 4129 0.01 Releasing (scientifically selecting releasing species and ensuring releasing quality) Portal scenario 4353 0.01 Select low-noise earthmoving machinery, transportation equipment, and electric equipment Portal scenario 710 0.01 Zero-excavation into the hole Portal scenario 1625 0.03 Vegetation reinforced slope (soil improvement, selection of plants with developed root systems and strong adaptability, use of grass mats, straw covers, and reasonable selection of plant species and planting density) Portal scenario 330 0.05 1. Artificial stripping of surface soil in necessary construction areas 2. Set up retaining walls Portal scenario 6200 0.10 Set up thermal curtains (color striped cloth, cotton felt) Portal scenario 3788 0.11 Build a roof gutter Portal scenario 5618 0.12 Double-layer cover plate water channel Portal scenario 164 0.27 Mechanical stripping of surface soil in necessary construction areas Tunnel body scenario 6078 0.00 Movable inverted arch trestle Tunnel body scenario 4086 0.00 Air supply ventilation Tunnel body scenario 2919 0.00 1. Pre-grouting water plugging 2. Post-excavation grouting (viscosity time-dependent material) Tunnel body scenario 5469 0.02 Select small wet dust collectors Tunnel body scenario 3564 0.06 Diffusion oxygen supply at the working face Tunnel body scenario 2744 0.09 Post-excavation grouting (viscosity time-dependent material) Tunnel body scenario 3742 0.12 Personal oxygen supply Tunnel body scenario 1426 0.13 Construction equipment equipped with polymer pumps (concrete conveying, drainage system) Tunnel body scenario 1069 0.13 Thermal blanket wrapped around the tank body of the concrete transport vehicle Tunnel body scenario 4288 0.19 Combined ventilation Tunnel body scenario 976 0.20 Excavation by cantilevered tunneling machine

[0087] After analysis, the output green construction measures meet the various green construction points required by the project, vegetation protection, soil and water conservation, material consumption reduction, noise pollution prevention and control, air pollution prevention and control, highland construction human resource protection and the like, and cover green construction measures in all aspects of tunnel construction, such as blasting, mechanized excavation, ventilation, waterproofing and drainage, etc., which shows that the green construction measures output by the model are practical and feasible. The field construction personnel can further judge whether the similar green construction measure case output is applicable in combination with the specific construction actual situation, and complete the final green construction measure decision of the project

[0088] In summary, the present application can accurately identify and quantify the missing influence of key indicators by using the comprehensive fuzziness coefficient evaluation method combined with weight and data anomaly probability, effectively alleviates the problem of incomplete data caused by special conditions such as terrain and weather through the characteristic label matching and virtual data supplement mechanism, and ensures the continuity of the decision-making process;

[0089] The mixed coding scheme is adopted, and the binary mask vector is combined to uniformly process the multi-source heterogeneous features, clearly identify the data validity, and provide high-quality and complete information input for the subsequent deep learning model; the triple loss based twin neural network is constructed, the deep and discriminative features between cases are accurately captured in the embedding space through metric learning, and the dynamic mask intersection calculation is used to ensure that only the common effective features are compared, so that the similarity measurement is still fair and accurate under the condition of data missing;

[0090] The dynamic threshold is introduced to adaptively select the most similar cases, and the measure list sorted by similarity is output, the new cases verified to be effective in the field are fed back to the case library, and the model retraining can be triggered after the virtual data is replaced by real data, forming a continuous learning decision system with high intelligence.

[0091] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some of the technical features can be replaced by equivalents; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A CBR-based mountain railway tunnel green construction decision-making method, characterized in that, The method comprises the following steps: Step 1: integrate the scene attribute system of the construction site of the hole opening and hole body, obtain the case representation including the construction scene attribute, attribute value and green construction measure, and generate a three-dimensional scene through parameterization of Rhino software; Step 2: for the case determined to be data fuzzy or missing due to special terrain or weather, based on the standard completeness case index, evaluate the fuzzy degree coefficient of the case in the scene attribute system in terms of geological conditions, environmental sensitive area, construction machinery type, energy consumption index and material transportation path parameter, and select the index with a comprehensive fuzzy degree coefficient higher than a preset threshold; Step 3: extract the characteristic label of the missing data under the selected index, match it with the complete characteristic label of the same index in the case library, generate virtual supplementary data through simulation replacement, temporarily store it in the case library and add a fuzzy mark; Step 4: perform hybrid coding on the numerical, categorical and Boolean heterogeneous features in the case library, use One-Hot coding with missing marks for the categorical features, and construct a binary mask vector to identify feature validity; Step 5: construct a twin neural network structure, input the preprocessed feature vector, process it through 3 layers of full connection, GELU activation function, layer normalization and Dropout layer, and output an L2 normalized embedding vector; Step 6: based on the mask representation of the query case, filter the candidate cases that do not have exclusive features in the mask vector of the training data, and calculate the intersection of the valid features of the query case and the candidate cases; Step 7: calculate the Euclidean distance between the embedding vectors of the query case and the candidate cases, extract the measure ranking output with a distance less than a dynamic threshold, feed the new case verified by construction back to the case library, and trigger model retraining after the virtual supplementary data is replaced; The evaluation process of the fuzzy degree coefficient in step 2 is as follows: For the target index, count the total number of data sub-items that must be included in the standard complete case as the baseline quantity, calculate the percentage of the number of data sub-items actually missing in the target index in the current case in the baseline quantity as the basic missing rate; According to the index type, preset the weights of geological conditions, environmental sensitive area, energy consumption index, construction machinery and transportation path; Multiply the basic missing rate by the weight factor of the corresponding index to generate a weighted missing rate. Based on the historical case library statistics, when the target index is missing, the data anomaly probability of the associated index is taken as the fluctuation coefficient. Finally, the fuzzy degree coefficient = weighted missing rate x (1 + fluctuation coefficient).

2. The CBR-based green construction decision method for mountain railway tunnels according to claim 1, characterized in that, The logic of characteristic label matching in step 3 is as follows: Predefine the characteristic labels of the geological conditions, environmental sensitive area, construction machinery type, energy consumption index and material transportation path index of each complete case in the case library; According to the selected high-fuzzy index type, extract the characteristic label requirement of the missing data under the target index, filter the complete cases of the same construction site and index type in the case library, calculate the matching degree of the characteristic labels and the missing labels, and comprehensively evaluate the matching degree according to the semantic similarity of the label description text and the numerical interval overlap degree; Select the characteristic label with the highest matching degree to perform data simulation replacement, generate virtual supplementary data while retaining the source label case number and replacement timestamp, and mark the corresponding case in the case library as a virtual supplementary state.

3. The CBR-based green construction decision method for mountain railway tunnels according to claim 2, characterized in that, The geological condition index corresponds to a characteristic label including a rock layer permeability level and a fault fracture zone range, the environmental sensitive area index corresponds to a characteristic label including a water source straight line distance and a core vegetation coverage, the construction machinery type index corresponds to a characteristic label including an equipment rated power range and a fuel type category, the energy consumption index corresponds to a characteristic label including a power supply source and a unit engineering quantity oil consumption interval, and the material transportation path parameter index corresponds to a characteristic label including a maximum slope angle and a minimum turning radius.

4. The CBR-based green construction decision method for mountain railway tunnels according to claim 2, characterized in that, The virtual supplementary data is marked with a semi-transparent label in the case library, and when the real case is obtained, the virtual data is automatically replaced and the fuzzy mark is cleared, and the twin neural network incremental training is triggered.

5. The CBR-based green construction decision method for mountain railway tunnels according to claim 1, characterized in that, The execution process of the mixed coding in step 4 includes: normalizing and scaling all numerical features in the case library to convert the value range to the interval of 0 to 1; performing one-hot encoding on the category features to generate an independent binary vector for each category value and create a special identification bit for the missing value; converting the Boolean features directly into numerical representation of 1 or 0 to build the encoding vector; generating a binary identification vector with the same feature dimension as each case, where the effective feature bit is marked as 1 and the missing feature bit is marked as 0 to build the mask vector; and splicing the obtained encoding vector and mask vector in order according to the feature dimension to form a joint feature vector for the final input neural network.

6. The CBR-based green construction decision method for mountain railway tunnels according to claim 1, characterized in that, The twin neural network structure in step 5 is formed by splicing two artificial neural networks with the same structure and shared weights, and the loss function selects a ternary loss to minimize the distance between the anchor point and the positive sample while maximizing the distance between the anchor point and the negative sample, learn a discriminative embedding space, and use the squared Euclidean distance to enhance the gradient stability, and its expression is: ； In the formula, represents a loss function, a, p and n represent embedding vectors of anchor points, positive samples and negative samples respectively, is an interval parameter, and N represents the total number of samples.

7. The CBR-based green construction decision method for mountain railway tunnels according to claim 1, characterized in that, The twin neural network model training process in step 5 adopts hierarchical cross-validation and full-volume training, uses cosine annealing to schedule the learning rate to avoid local optimum, and its expression is: ； In the formula, represents the actual learning rate of the neural network at the tth iteration, represents the current iteration step number.

8. The CBR-based green construction decision method for mountain railway tunnels according to claim 1, characterized in that, In step 6, the process of calculating the intersection of the effective features of the query case and the candidate case, the features existing in both directions are indicated by dynamic masks to ensure that only the common features are compared, and its expression is: ； wherein represents the effective feature intersection, represents the query case mask vector, represents the candidate case mask vector.

9. The CBR-based green construction decision method for mountain railway tunnels according to claim 1, characterized in that, The dynamic threshold in step 7 is set according to the P90 quantile of the similarity distribution of the training set, the output measure list is arranged in ascending order of Euclidean distance, if the distance exceeds the specified threshold, it is skipped, otherwise it is added to the measure list of the effective case.

Images