Artificial intelligence subway construction carbon emission reduction management method and system

Abstract

The invention belongs to the technical field of subway construction carbon emission reduction management, and particularly relates to an artificial intelligence subway construction carbon emission reduction management method and system. Through combination of a clustering algorithm, an attention enhanced deep belief network model, a time sequence attention enhanced dynamic Bayesian network model, a multi-target deep reinforcement learning generation recessive carbon emission optimization strategy and a recessive carbon emission optimization closed loop iteration mechanism, temporary facility repeated construction loss, material inventory overstocked loss, and a recessive carbon emission optimization closed loop iteration mechanism are calculated. And cross operation waiting energy consumption, equipment no-load transfer energy consumption and recessive carbon emission of a non-standard process are calculated and analyzed, so that an optimal carbon emission reduction strategy is generated.

Description

Technical Field

[0001] This invention belongs to the field of carbon emission reduction management technology for subway construction, specifically relating to an artificial intelligence-based method and system for carbon emission reduction management in subway construction. Background Technology

[0002] In the process of subway construction, in addition to the visible carbon emissions from direct energy consumption of construction equipment and production of building materials, there are also a large number of hidden carbon emissions that are not easily detected. These carbon emissions originate from inefficient aspects of construction organization and management, and are characterized by their high degree of concealment, dispersion, and difficulty in quantification. However, they account for 30% to 40% of the total carbon emissions from subway construction, making them a key blind spot in carbon emission reduction management. Existing technologies only address the single factor of equipment idling for hidden carbon emissions, and there are no corresponding management methods for key hidden sources such as repeated construction of temporary facilities, material inventory backlog, and waiting time for cross-operations. To address this, an improved artificial intelligence-based carbon emission reduction management method and system for subway construction has been designed. Summary of the Invention

[0003] In view of the above-mentioned shortcomings in the existing technology, the present invention provides an artificial intelligence-based method and system for managing carbon emission reduction in subway construction to solve the problems mentioned in the background technology.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: An artificial intelligence-based method and system for carbon emission reduction management in subway construction includes the following steps: S1. Five core hidden carbon emission factors were identified through clustering algorithms: losses from repeated construction of temporary facilities, losses from stockpiled materials, energy consumption during cross-operations, energy consumption during idle equipment transport, and additional energy consumption from non-standard processes. A nonlinear mapping relationship between each factor and construction parameters was established. S2. Deploy a micro-feature perception network to collect associated data, and achieve dynamic frequency adjustment through an adaptive sampling mechanism; S3. Construct an attention-enhanced deep belief network model to quantify implicit carbon emissions. Implicit features are extracted through a 3-layer restricted Boltzmann machine. A scene-adaptive attention mechanism is introduced to assign differentiated weights to features of different construction scenarios, and output accurate values ​​of implicit carbon emissions. S4. A dynamic Bayesian network model with enhanced temporal attention is proposed. By using dynamic temporal attention weights, the dynamic correlation strength of latent factors in different construction stages is captured, the cascade impact paths and contributions of key latent emission sources are identified, and the final latent carbon emission measurement value is obtained. S5. Generate implicit carbon emission optimization strategies based on multi-objective deep reinforcement learning, with the implicit carbon emission reduction rate as the core optimization objective, coupled with construction constraints, and output targeted control strategies for different processes under different scenarios; S6. Establish a closed-loop iterative mechanism for optimizing implicit carbon emissions, and feed back the optimization effect data of implicit carbon emissions in actual construction to the attention-enhanced deep belief network model in step S3, so as to achieve rapid adaptation of model parameters under new construction scenarios through transfer learning.

[0005] Preferably, the associated data in S2 includes temporary facility turnover rate, material inventory turnover days, process connection waiting time, equipment no-load operation trajectory, and process parameter deviation value.

[0006] Preferably, in S3, the attention-enhanced deep belief network model assigns an attention weight of 0.6-0.7 to the material inventory features of the station construction scenario, an attention weight of 0.5-0.6 to the equipment transfer features of the open-cut section scenario, and an attention weight of 0.4-0.5 to the process deviation features of the auxiliary structure scenario. Furthermore, an adaptive module assigns differentiated weights to the features of different construction scenarios.

[0007] Preferably, the formula for calculating the differential weight is as follows: ,in For the first In the scenario, the first The similarity score for each feature is calculated using the following formula: , The total number of features.

[0008] Preferably, the process of extracting latent features using the 3-layer restricted Boltzmann machine in S3 is as follows: The first layer RBM receives 15-dimensional input features, and the number of hidden layer nodes is set to 30. The underlying correlation features between material inventory and secondary transfer, and between equipment idleness and energy consumption are captured by the contrastive divergence algorithm. The number of nodes in the second RBM hidden layer is set to 20, and the middle-layer features of process connection and equipment waiting time are further extracted. The number of nodes in the third RBM hidden layer is set to 10 to explore the high-level features of excavation process deviation and increased energy consumption. Stack the 3-layer RBM to form a complete deep belief network (DBN).

[0009] Preferably, the dynamic temporal attention weights in S4 are calculated as follows: the temporal attention weights are calculated using a stage adaptation function, and the formula is as follows: ,in For time t, the first The construction elements and the first Attention weighting of hidden carbon emission factors The fit score is composed of both the stage matching score and the time decay score. If the elements... If it is the core factor at time t, the score is 1; otherwise, it is 0.3~0.7. The time decay score adopts an exponential decay model. , The stage start time is then multiplied by the conditional probability to obtain the dynamic temporal attention weight.

[0010] Preferably, the formula for calculating the final implicit carbon emission measurement value in step S4 is as follows: ,in The activation function for the output layer is Sigmoid. The connection weights between the output layer and the feature extraction layer. The activation function for the feature extraction layer is Sigmoid. This is the input to hidden layer 3. For output layer bias terms, These serve as benchmark values ​​for carbon emissions in various scenarios. This is the final output value representing the implicit carbon emissions.

[0011] Preferably, the multi-objective deep reinforcement learning in step S5 adopts the DuelingDQN algorithm, which divides the state space into three subspaces: implicit carbon emission state, resource allocation state, and process progress state. The action space is used to identify all core implicit carbon emission sources in S1.

[0012] Preferably, the implicit carbon emission optimization closed-loop iterative mechanism in S6 is implemented as follows: weekly execution data of optimization strategies during actual construction is collected, and after compliance verification, it is fed back to the attention-enhanced deep belief network model in step S3. The parameters of the attention-enhanced deep belief network model are updated using incremental training to adapt to the actual working conditions. After the attention-enhanced deep belief network model is updated, a new round of strategy generation is automatically triggered. This new round of strategy adjusts the action space weights according to the implicit carbon emission contribution in the current scenario, thereby generating a new optimization strategy adapted to the current scenario.

[0013] An AI-powered carbon emission reduction management system for subway construction includes a hidden carbon emission feature identification module, a data acquisition module, a hidden carbon emission quantification module, a hidden carbon emission correlation analysis module, an optimization strategy generation module, and a closed-loop iteration module. The hidden carbon emission feature identification module is used to identify hidden carbon emission factors. The data acquisition module is used to collect related data. The hidden carbon emission quantification module is used to extract features from the collected data and output the hidden carbon emission measurement value. The hidden carbon emission correlation analysis module is used to perform correlation strength analysis on the collected data to obtain the final hidden carbon emission measurement value. The optimization strategy generation module: obtains the control strategy through constraints based on the final implicit carbon emission measurement value; The closed-loop iteration module collects the data after the strategy is executed and feeds it back to the implicit carbon emission quantification module. Compared with the prior art, the present invention has the following beneficial effects: 1. Clustering algorithms accurately identify five core latent factors: repeated construction of temporary facilities, material inventory backlog, waiting for cross-operations, idle equipment transport, and additional energy consumption from non-standard processes, achieving comprehensive coverage of latent carbon emission sources. Then, by establishing a non-linear mapping relationship between these five latent factors and construction parameters, they are quantified for computational analysis. Differential weights are assigned to core features under different construction scenarios, and a three-layer Restricted Boltzmann Machine (RBM) is used to extract features from the bottom, middle, and upper layers, further improving quantification accuracy. Finally, a dynamic temporal attention weight is used to capture latent factors at different stages. The correlation strength of elements is combined with the stage matching score and the exponential time decay score to dynamically adjust the contribution weight of each factor, and finally obtain the implicit carbon emission measurement value. Then, through the DuelingDQN multi-objective deep reinforcement learning algorithm, the state space is divided into three subspaces: "implicit carbon emission state", "resource allocation state", and "process progress state". A balance is achieved between emission reduction target and construction resource constraints and progress constraints, thereby generating the optimal strategy. This can analyze and calculate the implicit carbon emission sources of temporary facilities repeated construction, material inventory backlog, cross-operation waiting, equipment idle transfer and non-standard processes under different construction environments, and obtain the best strategy under this working condition. 2. By setting up a closed-loop management logic, optimization strategy execution data is collected weekly during actual construction. After compliance verification, the data is fed back to the attention-enhanced DBN. The model parameters are updated through incremental training algorithms. Without reconstructing the model, it can quickly adapt to changes in actual working conditions. After the model parameters are updated, a new round of strategy generation is automatically triggered. The new strategy adjusts the action space weights according to the real-time contribution of hidden carbon emissions in the current scenario, achieving continuous looping and ensuring that the emission reduction effect is continuously strengthened as construction progresses. Moreover, this method ensures scenario adaptability through multi-dimensional design, making it applicable to different environments and versatile. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating an embodiment of an artificial intelligence-based method and system for carbon emission reduction management in subway construction according to the present invention; Figure 2 This is a framework diagram of an artificial intelligence-based carbon emission reduction management system for subway construction according to the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0016] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0017] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present patent. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0018] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0019] Example 1: like Figure 1-2 The present invention discloses an artificial intelligence-based method and system for managing carbon emission reduction in subway construction. The present invention includes the following steps: S1. Five core hidden carbon emission factors were identified through clustering algorithms: losses from repeated construction of temporary facilities, losses from stockpiled materials, energy consumption during cross-operations, energy consumption during idle equipment transport, and additional energy consumption from non-standard processes. A nonlinear mapping relationship between each factor and construction parameters was established. S2. Deploy a micro-feature perception network to collect associated data, and achieve dynamic frequency adjustment through an adaptive sampling mechanism; S3. Construct an attention-enhanced deep belief network model to quantify implicit carbon emissions. Implicit features are extracted through a 3-layer restricted Boltzmann machine. A scene-adaptive attention mechanism is introduced to assign differentiated weights to features of different construction scenarios, and output accurate values ​​of implicit carbon emissions. S4. A dynamic Bayesian network model with enhanced temporal attention is proposed. By using dynamic temporal attention weights, the dynamic correlation strength of latent factors in different construction stages is captured, the cascade impact paths and contributions of key latent emission sources are identified, and the final latent carbon emission measurement value is obtained. S5. Generate implicit carbon emission optimization strategies based on multi-objective deep reinforcement learning, with the implicit carbon emission reduction rate as the core optimization objective, coupled with construction constraints, and output targeted control strategies for different processes under different scenarios; S6. Establish a closed-loop iterative mechanism for optimizing implicit carbon emissions, and feed back the optimization effect data of implicit carbon emissions in actual construction to the attention-enhanced deep belief network model in step S3, so as to achieve rapid adaptation of model parameters under new construction scenarios through transfer learning.

[0020] The associated data in S2 includes temporary facility turnover rate, material inventory turnover days, process connection waiting time, equipment no-load operation trajectory, and process parameter deviation value.

[0021] The attention-enhanced deep belief network model in S3 assigns an attention weight of 0.6-0.7 to the material inventory features of the station construction scenario, an attention weight of 0.5-0.6 to the equipment transfer features of the open-cut section scenario, and an attention weight of 0.4-0.5 to the process deviation features of the auxiliary structure scenario. It also assigns differentiated weights to the features of different construction scenarios through an adaptive module.

[0022] The formula for calculating the differential weight is as follows: ,in For the first In the scenario, the first The similarity score for each feature is calculated using the following formula: , The total number of features.

[0023] The process of extracting latent features using the 3-layer restricted Boltzmann machine in S3 is as follows: The first layer RBM receives 15-dimensional input features, and the number of hidden layer nodes is set to 30. The underlying correlation features between material inventory and secondary transfer, and between equipment idleness and energy consumption are captured by the contrastive divergence algorithm. The number of nodes in the second RBM hidden layer is set to 20, and the middle-layer features of process connection and equipment waiting time are further extracted. The number of nodes in the third RBM hidden layer is set to 10 to explore the high-level features of excavation process deviation and increased energy consumption. Stack the 3-layer RBM to form a complete deep belief network (DBN).

[0024] The dynamic temporal attention weights in S4 are calculated as follows: the temporal attention weights are calculated using a stage adaptation function, and the formula is as follows: ,in For time t, the first The construction elements and the first Attention weighting of hidden carbon emission factors The fit score is composed of both the stage matching score and the time decay score. If the elements... If it is the core factor at time t, the score is 1; otherwise, it is 0.3~0.7. The time decay score adopts an exponential decay model. , The stage start time is then multiplied by the conditional probability to obtain the dynamic temporal attention weight.

[0025] The formula for calculating the final hidden carbon emission measurement value in S4 is as follows: ,in The activation function for the output layer is Sigmoid. The connection weights between the output layer and the feature extraction layer. The activation function for the feature extraction layer is Sigmoid. This is the input to hidden layer 3. For output layer bias terms, These serve as benchmark values ​​for carbon emissions in various scenarios. This is the final output value representing the implicit carbon emissions.

[0026] The multi-objective deep reinforcement learning in step S5 uses the DuelingDQN algorithm, which divides the state space into three subspaces: implicit carbon emission state, resource allocation state, and process progress state. The action space is used to identify all the core implicit carbon emission sources in S1.

[0027] The implicit carbon emission optimization closed-loop iterative mechanism in S6 is implemented as follows: weekly execution data of optimization strategies during actual construction is collected, and after compliance verification, it is fed back to the attention-enhanced deep belief network model in step S3. The parameters of the attention-enhanced deep belief network model are updated using incremental training to adapt to the actual working conditions. After the attention-enhanced deep belief network model is updated, a new round of strategy generation is automatically triggered. This new round of strategy adjusts the action space weights according to the implicit carbon emission contribution in the current scenario, thereby generating a new optimization strategy adapted to the current scenario.

[0028] An AI-powered carbon emission reduction management system for subway construction includes a hidden carbon emission feature identification module, a data acquisition module, a hidden carbon emission quantification module, a hidden carbon emission correlation analysis module, an optimization strategy generation module, and a closed-loop iteration module. The hidden carbon emission feature identification module is used to identify hidden carbon emission factors. The data acquisition module is used to collect related data. The hidden carbon emission quantification module is used to extract features from the collected data and output the hidden carbon emission measurement value. The hidden carbon emission correlation analysis module is used to perform correlation strength analysis on the collected data to obtain the final hidden carbon emission measurement value. The optimization strategy generation module: obtains the control strategy through constraints based on the final implicit carbon emission measurement value; The closed-loop iteration module collects the data after the strategy is executed and feeds it back to the implicit carbon emission quantification module. Example 2: S1 constructs an improved density clustering algorithm to mine historical construction data, accurately identify hidden carbon emission factors, and build a feature system. The specific process is as follows: Data preprocessing: Construction logs, energy consumption records, material ledgers, and equipment operation data from multiple similar subway projects over the past 5 years were collected. Outliers were detected and removed using the 3x standard deviation principle and box plot method. Reasonable extreme values ​​caused by unexpected construction events were retained and their causes were marked. Missing values ​​were then filled in using Lagrange interpolation. Cross-source data comparison was then performed on the collected data, comparing the actual recorded values ​​of smart meters with the energy consumption estimates from equipment operating condition sensors on a time-period basis. When the relative error between the two exceeded a threshold, the actual measured value recorded by the meters was used. Real-time inventory records from warehouse smart inventory terminals were periodically compared with supplier delivery notes and on-site material requisition forms, satisfying the formula... After cross-source data comparison, the temporal logic is verified by comparing the completion time of the preceding process with the start time of the subsequent process step by step to prevent process errors. The actual sampling frequency is compared with the theoretical frequency set by the adaptive sampling mechanism over time; the actual frequency must match the theoretical frequency. After verification, the consistency of the data in terms of numerical accuracy and temporal logic is ensured. The verified data is then standardized using Z-score to eliminate the influence of different data dimensions. The mean of each feature dimension is calculated separately using the following formula: , Indicates the total amount of data. Indicates the first Data points, This represents the mean of all data points in the original dataset; the standard deviation is calculated using the following formula: , The standard deviation of all data points in the original dataset is represented by the standardization transformation formula: Each standardized data point Its local density Defined at the initial radius The number of neighbors within the range, where a neighbor is defined as a point. The distance is less than or equal to For other data points, the distance calculation formula is: ,in For feature dimension, For point In the Standardized values ​​on each feature, initial radius To take half of the median distance between all pairs of points, for the global average density The calculation is performed using the following formula: , This represents the average clustering degree of the entire dataset, and is then calculated based on the radius of the neighboring region. Adjusting the density characteristics of standardized data, the specific formula is as follows: For points in dense areas, the radius is reduced to avoid misclassifying too many points as neighbors; for points in sparse areas, the radius is increased to ensure that similar points in sparse areas are included as neighbors, thereby... The adjustments fully reflect the true density differences of the data, ultimately making the clustering results more consistent with the actual distribution of hidden carbon emission factors, and then constructing a random forest mapping model.

[0029] S2 deploys a micro-feature perception network to collect associated data. Distributed sensing devices will be deployed in the construction area to build a multi-dimensional data acquisition network, enabling real-time and accurate collection of data related to hidden carbon emissions. The specific plan is as follows: In the temporary facility area, RFID tags and infrared calculators are installed to record the frequency of equipment disassembly and assembly and the duration of each disassembly and assembly. Temperature and humidity sensors and smart inventory counting terminals are deployed in the material warehouse area to collect inventory turnover data once an hour. UWB positioning base stations and video analysis equipment are deployed on the work site to capture the waiting time between processes using the YOLOv8 target detection algorithm. GNSS positioning modules and operating condition sensors are equipped on construction vehicles to record their no-load operation trajectory and duration. Deploy parameter acquisition terminals on key equipment in the process execution stage to record the deviation values ​​between actual parameters and standard parameters; Then, the sampling frequency is dynamically adjusted based on the construction intensity. The construction intensity evaluation formula is: ,when When the value is greater than 1.2, it indicates the peak construction period. At this time, the sampling frequency is set to 3-5Hz. During the normal construction period, the sampling frequency was set to 1–3 Hz. During the off-season for construction, the sampling frequency was set to 0.5–1 Hz, and the collected data was uploaded to the cloud database in real time.

[0030] S3 constructs an attention-enhanced deep belief network to quantify implicit carbon emissions. The network structure includes an input layer, a feature extraction layer, an attention mechanism layer, and an output layer: The number of nodes in the input layer is equal to the total number of feature indicators collected in step S, which is used to receive multi-dimensional feature data. Feature Extraction Layer: The historical data of S2 is divided into a training set, and normalization is used to map the data to the range of 0 to 1. Pre-training is performed using a three-layer Restricted Boltzmann Machine (RBM). The first layer RBM receives 15-dimensional input features, and the number of hidden layer nodes is set to 30. The contrastive divergence algorithm is used to capture low-level correlation features such as material inventory and secondary transfer, and equipment idle time and energy consumption. The number of hidden layer nodes in the second layer RBM is set to 20 to further extract mid-level features such as process connection and equipment waiting time. The number of hidden layer nodes in the third layer RBM is set to 10 to mine high-level features such as process deviation and increased energy consumption. The three layers of RBM are stacked into a complete Deep Belief Network (DBN), and the weight parameters of each layer are obtained. The parameters of all layers are fine-tuned using the validation set labeled data to align the model output with the actual values. Finally, the weight parameters of each layer that can accurately measure hidden carbon emissions are obtained. The larger the absolute value of the weight, the stronger the influence of the corresponding feature on carbon emission measurement.

[0031] The attention mechanism layer employs a scene-adaptive attention module, assigning differentiated weights to features of different construction scenarios. Its attention weight calculation function is as follows: ,in For the first In the scenario, the first The similarity score for each feature is calculated using the following formula: , The total number of features is set, and a fixed weight ratio is set to determine the initial cognitive framework of the model. This allows the model to know the features that should be focused on in this scenario at the start-up stage, avoiding the model from getting into meaningless random exploration. The weight ratio is set according to different construction scenarios. In the station construction scenario, based on historical experience, the hidden carbon emissions caused by material inventory backlog account for 60% to 70% of the total hidden emissions. Therefore, the material inventory characteristic weight is assigned to 0.6 to 0.7, and the cross-operation weight is assigned to 0.2 to 0.3. In the open-cut tunnel scenario, due to the long construction range and frequent material transfer, the carbon emissions from equipment transfer without load are high. Therefore, the equipment transfer characteristic weight is assigned to 0.5 to 0.6, and the temporary facility characteristic weight is assigned to 0.3 to 0.4. In the auxiliary structure scenario, there are non-standard construction processes, such as the processing of irregularly shaped components, which require additional energy consumption. Therefore, the process deviation characteristic weight is assigned to 0.4 to 0.5, the cross-operation weight to 0.3 to 0.4, and other weights to 0.1 to 0.3.

[0032] Then, the obtained attention weights are fused with the features, and the calculation formula is as follows: ,in For the calculated attention weights, The original feature data, These are the weighted eigenvalues. Input to output layer; Output layer: Weighted eigenvalues After entering the output layer, the network parameters are initialized through training an attention-enhanced deep belief network using a 3-layer Restricted Block Model (RBM) to learn data features layer by layer. Then, the parameters of all layers are corrected using backpropagation based on the actual value of implicit carbon emissions. The fine-tuned parameters are then substituted into the formula, mapped through the output layer activation function, and multiplied by a scenario-based standard value to finally obtain the calculated value of implicit carbon emissions. The formula is as follows: ,in The activation function for the output layer is Sigmoid. The connection weights between the output layer and the feature extraction layer (hidden layer 3) are... The activation function for the feature extraction layer is Sigmoid. This is the input to hidden layer 3. For output layer bias terms, These serve as benchmark values ​​for carbon emissions in various scenarios. This is the final output value representing the implicit carbon emissions.

[0033] Given the phased nature of subway construction and the dynamic changes in the relationships between elements, the temporal attention weight is calculated using a phase adaptability function, the formula of which is: ,in For time t, the first The construction elements and the first Attention weighting of hidden carbon emission factors The fit score is composed of both the stage matching score and the time decay score. If the elements... If it is the core factor at time t, the score is 1; otherwise, it is 0.3~0.7. The time decay score adopts an exponential decay model. , The starting time of the phase is set, and then the temporal attention weight is multiplied by the conditional probability of the DBN. The conditional probability of the DBN is calculated based on historical data, extracting the joint occurrence frequency of each node state. For example, the probability of secondary transfer when inventory backlog occurs is calculated to obtain the dynamic conditional probability. Then, the network inference capability is extended through the joint tree algorithm, which can not only directly identify operational problems, but also trace multi-level indirect transmission paths, such as supplier delay → inventory backlog → secondary transfer → equipment idle. The path saliency is evaluated by the Bayesian information criterion BIC. The smaller the BIC value, the more critical the path. The calculation formula is as follows: ,in Let be the likelihood function. The number of path parameters. For the sample size, the contribution of each node is calculated using the reverse tracing method for the selected significant paths. The node characteristic parameters are dynamically updated based on the real-time data collected by S2, thereby uncovering the multi-level transmission relationship of hidden carbon emissions. The time-series attention weights enable the model to respond in real time to the transition of construction stages, thus ensuring that the correlation analysis is always synchronized with the actual working conditions.

[0034] S5. Generating Latent Carbon Emission Optimization Strategies Based on Multi-Objective Deep Reinforcement Learning By constructing a multi-objective optimization model, a hidden carbon emission optimization strategy adapted to dynamic operating conditions is generated. This is achieved through a dual-depth Q-network algorithm, where the state space is defined as a high-dimensional vector. It contains three key types of information: hidden carbon emission status. This corresponds to the real-time values ​​of the five types of hidden carbon emissions quantified in step S3; resource allocation status. This includes metrics reflecting the efficiency of construction resource utilization, such as personnel load rate, equipment load rate, and material inventory turnover rate; and the status of work process progress. Optimization strategies, including critical path buffer time, process completion rate, and overlap of cross-operations, ensure that the project schedule is not affected. The state space is automatically updated with S2 to ensure that the model reflects the latest state of the construction scenario. The action space is designed according to the classification of hidden carbon emission factors to ensure that it can cover all core hidden carbon emission sources identified in S1. A dynamically adjusted reward function is set to adapt to different project durations. The reward function calculation formula is as follows: ,in As a carbon emission incentive, For carbon emission weights, As a weight for resource utilization rate, Incentives for resource utilization As a weight for the smoothness of progress, To reward smooth progress, the weights of various components are adjusted in real time at different project milestones to ensure no delays. Strategy generation consists of two phases: model training and online strategy inference. An experience sample library is built using historical subway project data, and the DuelingDQN network is constructed to calculate the final output Q-value. The calculation formula is as follows: ,in and The network extracts evaluation values ​​for state value and action advantage, and introduces a priority experience replay mechanism to assign higher sampling weights to high-reward samples, accelerating model convergence. The network parameters are updated through the Adam optimizer. Training stops when the probability that the policy makes the reward function greater than 0 is stable above 85%. Then, the model calculates the Q-value of all actions in the current state according to the implicit carbon emission factors, filters the actions with the highest Q-value, and converts the selected actions into specific construction instructions and pushes them to the construction management system. One hour after execution, carbon emissions, resource utilization rate, and progress smoothness data in the new state are collected, the reward is calculated, and the data in this way is stored in the sample database.

[0035] S6 establishes a closed-loop iterative mechanism for optimizing hidden carbon emissions. Data on the implementation of optimization strategies during actual construction is collected weekly and then... After the principle is followed, the feedback is fed into the attention-enhanced DBN model of S3. The model parameters are updated using incremental training to ensure that the model continues to adapt to the actual working conditions. After the model is updated, a new round of policy generation is automatically triggered. This policy adjusts the weights of the action space according to the implicit carbon emission contribution in the scenario, thereby changing the contribution in the new scenario and generating a new policy.

[0036] In new scenarios, a cross-scenario general model is constructed using a model-independent meta-learning framework. Through preprocessing and fine-tuning, rapid adaptation to new scenarios is achieved. This cross-scenario general model first collects historical data from subway construction, including data on different geological, climatic, and structural types. Then, it employs an attention-enhanced DBN structure similar to the S3 system, but adds scenario adaptation parameters to its output layer. The meta-objective function maximizes the average fine-tuning accuracy across scenarios. ,in For meta-parameters, For the scene, To fine-tune the number of steps, The learning rate is used to obtain the meta-initialization parameters, and then the obtained meta-initialization parameters are used to achieve high-precision adaptation through multiple rounds of iteration.

[0037] Example 3: The first phase of a certain subway line 0 project encompasses 8 underground stations, 9.2km of tunnel, and 4 ancillary structures, with a construction period of 4 years. It involves various scenarios including open-cut station construction, tunnel boring, and fabrication of irregularly shaped components for ancillary structures. Preliminary project research revealed hidden carbon emission issues during construction, such as material inventory backlog, idle equipment transport, and waiting times between work processes, accounting for 32% of the project's total carbon emissions. The specific control and implementation process is as follows: S1: Collect construction logs, energy consumption records, material ledgers, and equipment operation data from similar projects of this subway project over the past 5 years. Use the 3x standard deviation principle combined with box plots to detect outliers. Identify abnormal data in the energy consumption records, such as daily power consumption exceeding 8000kWh for a single tunnel boring machine and a daily decrease of 50 tons in steel reinforcement inventory. These were verified to be sensor malfunctions or data entry errors and were removed. Meanwhile, reasonable extreme values ​​such as a 12-hour delay in concrete transportation due to heavy rain and a sudden increase in energy consumption due to equipment malfunctions were retained, and weather-related or equipment failure-related anomalies were marked as the cause. Also, consider the monthly cement loss in the material ledgers. Missing data such as consumption and equipment maintenance record intervals were supplemented using Lagrange interpolation to ensure data continuity. The real-time recorded values ​​of smart meters were compared hourly with the energy consumption estimates from equipment operating condition sensors, with a relative error threshold of 5%. When the error exceeded the limit, the actual measured value of the meters was used. The inventory records of the warehouse smart inventory terminal were compared weekly with the supplier's delivery note and the on-site material requisition form, which had to meet the formula "ending inventory = beginning inventory + current period inbound - current period outbound". Unregistered steel bar inbound and duplicate cement requisition records were corrected, and the carbon emission baseline value during the construction process was calculated to be 60tCO2e / week.

[0038] Define the local density of each data point after standardization. The number of neighbors within the initial radius r is defined as the number of data points with a distance d ≤ r. The median distance between all data points is calculated using a formula, which is 8.6. The initial radius r is set to 4.3, and the global average density is calculated. Adjustments are made based on the calculation results; for regions where the local density is greater than the global average density, the radius is reduced to [a smaller value]. For regions where the local density is less than the global average density, the radius is expanded to... The final feature system is as follows: Materials: Material inventory turnover rate, inventory backlog duration, and frequency of secondary transfers; Equipment: Equipment idle running time, equipment disassembly and assembly frequency, and equipment parameter deviation rate; Processes: Process connection waiting time, overlap of cross-operations, and critical path buffer time; Temporary facilities: Temporary building area, temporary power load, and facility disassembly and assembly time; Technology: Process parameter deviation value, number of times non-standard processes are executed, and processing volume of irregularly shaped components.

[0039] S2: Based on the feature system constructed in S1, distributed sensing devices will be deployed in the construction area of ​​Station 3, the tunnel section, and the ancillary structures of Metro Line 10. The scheme is as follows: Temporary facility area: Install RFID tags attached to temporary prefabricated houses, power distribution boxes and other facilities, and infrared timers to record the frequency of facility disassembly and assembly and the duration of each disassembly and assembly in real time; Materials warehouse area: Temperature and humidity sensors are deployed to collect temperature and humidity data once per hour to ensure that the cement storage environment meets the standards. The intelligent inventory terminal is based on RFID technology and automatically counts the inventory of major materials such as steel bars and cement every hour to generate inventory turnover data. Work area: UWB positioning base stations and video analysis equipment are deployed at the station foundation pit work area and the tunnel shield work area to capture the waiting time between processes through the YOLOv8 target detection algorithm; Construction vehicles: GNSS positioning modules and working condition sensors are equipped on 12 concrete mixer trucks and 8 cranes to record the no-load running trajectory and no-load duration; Key equipment: Deploy parameter acquisition terminals on key equipment such as tunnel boring machines, rebar cutters, and concrete pump trucks to collect the deviation values ​​between actual parameters and standard parameters in real time.

[0040] According to the construction intensity evaluation formula, I = (actual number of workers / planned number of workers) × (actual number of equipment in operation / planned number of equipment in operation).

[0041] S3: Quantify implicit carbon emissions through attention-enhanced deep belief networks. The attention-enhanced deep belief network (DBN) consists of an input layer, three feature extraction layers (Restricted Boltzmann Machine (RBM)), an attention mechanism layer, and an output layer.

[0042] The number of input layer nodes = the total number of features in the S1 feature system = 15, including 3 material categories, 3 equipment categories, 3 process categories, 3 temporary facility categories, and 3 technology categories. It receives multi-dimensional feature data collected by S2. Feature Extraction Layer: Data Partitioning and Preprocessing: The 12 months of historical data collected by S2 were divided into training and validation sets. Min-Max normalization was used to map the data to the [0,1] interval. The first RBM (input layer → hidden layer 1) received 15-dimensional input features and the number of hidden layer nodes was set to 30. It was trained by the contrastive divergence algorithm to capture low-level related features such as "material inventory and secondary transfer" and "equipment idle and energy consumption". The second RBM (hidden layer 1 → hidden layer 2) had 20 hidden layer nodes and further extracted mid-level features such as "process connection and equipment waiting". The third RBM (hidden layer 2 → hidden layer 3) had 10 hidden layer nodes and mined high-level features such as "process deviation and energy consumption increase". Then, the three RBMs were stacked into a complete deep belief network (DBN). The validation set labeled data (actual implicit carbon emissions, calculated by "material loss carbon emission coefficient + equipment energy consumption carbon emission coefficient") was used to fine-tune all layer parameters to align the model output with the actual values. Finally, the weight parameters of each layer were obtained. Among them, the absolute value of the weight corresponding to the feature "material inventory backlog duration" is 0.82, indicating that it has the strongest impact on carbon emission measurement. Then, the attention weight was calculated. In the station construction scenario, the material inventory backlog duration was calculated to be 3 days. It was then weighted with the initial weights in the station construction scenario: material inventory feature 0.65, cross-operation feature 0.25, and other features 0.1. The weighted feature value was 1.95. The weighted feature value was then input into the output layer. The results showed that in the station construction scenario, the final output implicit carbon emissions of materials were 68.5 tCO2e / week, equipment was 18.2 tCO2e / week, and process was 9.5 tCO2e / week.

[0043] S4 then uses a temporal attention-enhanced dynamic Bayesian network to mine associated paths and find the paths that cause hidden carbon emissions. Taking the material inventory in the construction of the station scene as the core factor, its stage matching score is 1, and the time decay score is calculated to be 0.37. Thus, the attention weight of the hidden carbon emissions of material inventory backlog under this construction element in the station scene can be obtained. Then, this attention weight is combined with the conditional probability of DBN to obtain dynamic conditional probability. Then, the joint tree algorithm is used to mine and identify the reasons for the material inventory backlog duration of this project as follows: supplier delay (t1) → inventory backlog (t2) → secondary transfer (t3) → equipment idle (t4) → increase in hidden carbon emissions (t5). The BIC is evaluated by Bayesian Information Criterion (BIC) and obtained as -128.6, which verifies that this path is a critical path.

[0044] S5 reads the current data state collected by S2 using a pre-trained multi-objective deep reinforcement learning model. Then, it calculates the Q-value of all actions in the action space using the DuelingDQN network and removes actions that violate basic constraints, such as shortening the steel bar procurement cycle to 2 days, but the supplier's minimum delivery cycle is 3 days. The remaining available actions are prioritized according to their Q-values ​​from high to low, with the first priority being the optimal strategy and the second priority being the alternative strategy. The output of this project is to sign a supply response agreement with the core material supplier, requiring the delivery cycle to be shortened from the original 7 days to 5 days. At the same time, the supplier is required to synchronize the daily inventory progress, prioritize the consumption of existing cement and steel bars in the inventory, suspend new purchases of the same type of materials, re-plan the material stacking area, and reduce the transfer distance.

[0045] The above are merely embodiments of the present invention. The circuits, electronic components, and modules involved are all prior art, fully achievable by those skilled in the art, and require no further explanation. The scope of protection in this application does not involve improvements to the software and methods. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all prior art in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. An artificial intelligence-based method and system for managing carbon emission reduction in subway construction, characterized in that: Includes the following steps: S1. Five core hidden carbon emission factors were identified through clustering algorithms: losses from repeated construction of temporary facilities, losses from stockpiled materials, energy consumption during cross-operations, energy consumption during idle equipment transport, and additional energy consumption from non-standard processes. A nonlinear mapping relationship between each factor and construction parameters was established. S2. Deploy a micro-feature perception network to collect associated data, and achieve dynamic frequency adjustment through an adaptive sampling mechanism; S3. Construct an attention-enhanced deep belief network model to quantify implicit carbon emissions from the collected associated data. Extract implicit features through a 3-layer restricted Boltzmann machine, introduce a scene-adaptive attention mechanism to assign differentiated weights to features of different construction scenarios, and output accurate measurement values ​​of implicit carbon emissions. S4. A dynamic Bayesian network model with enhanced temporal attention is proposed. By using dynamic temporal attention weights, the dynamic correlation strength of latent factors in different construction stages is captured, the cascade impact paths and contributions of key latent emission sources are identified, and the final latent carbon emission measurement value is obtained. S5. Generate implicit carbon emission optimization strategies based on multi-objective deep reinforcement learning, with the implicit carbon emission reduction rate as the core optimization objective, coupled with construction constraints, and output targeted control strategies for different processes under different scenarios; S6. Establish a closed-loop iterative mechanism for optimizing implicit carbon emissions, and feed back the optimization effect data of implicit carbon emissions in actual construction to the attention-enhanced deep belief network model in step S3, so as to achieve rapid adaptation of model parameters under new construction scenarios through transfer learning.

2. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 1, characterized in that: The associated data in S2 includes temporary facility turnover rate, material inventory turnover days, process connection waiting time, equipment no-load operation trajectory, and process parameter deviation value.

3. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 2, characterized in that: The attention-enhanced deep belief network model in S3 assigns an attention weight of 0.6-0.7 to the material inventory features of the station construction scenario, an attention weight of 0.5-0.6 to the equipment transfer features of the open-cut section scenario, and an attention weight of 0.4-0.5 to the process deviation features of the auxiliary structure scenario. It also assigns differentiated weights to the features of different construction scenarios through an adaptive module.

4. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 3, characterized in that: The formula for calculating the differential weight is as follows: ,in For the first In the scenario, the first The similarity score for each feature is calculated using the following formula: , The total number of features.

5. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 1, characterized in that: The process of extracting latent features using the 3-layer restricted Boltzmann machine in S3 is as follows: The first layer RBM receives 15-dimensional input features, and the number of hidden layer nodes is set to 30. The underlying correlation features between material inventory and secondary transfer, and between equipment idleness and energy consumption are captured by the contrastive divergence algorithm. The number of nodes in the second RBM hidden layer is set to 20, and the middle-layer features of process connection and equipment waiting time are further extracted. The number of nodes in the third RBM hidden layer is set to 10 to explore the high-level features of excavation process deviation and increased energy consumption. Stack the 3-layer RBM to form a complete deep belief network (DBN).

6. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 1, characterized in that: The dynamic temporal attention weights in S4 are calculated as follows: the temporal attention weights are calculated using a stage adaptation function, and the formula is as follows: ,in For time t, the first The construction elements and the first Attention weighting of hidden carbon emission factors The fit score is composed of both the stage matching score and the time decay score. If the elements... If it is the core factor at time t, the score is 1; otherwise, it is 0.3~0.

7. The time decay score adopts an exponential decay model. , The stage start time is then multiplied by the conditional probability to obtain the dynamic temporal attention weight.

7. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 1, characterized in that: The formula for calculating the final hidden carbon emission measurement value in S4 is as follows: ,in The activation function for the output layer is Sigmoid. The connection weights between the output layer and the feature extraction layer. The activation function for the feature extraction layer is Sigmoid. This is the input to hidden layer 3. For output layer bias terms, These serve as benchmark values ​​for carbon emissions in various scenarios. This is the final output value representing the implicit carbon emissions.

8. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 1, characterized in that: The multi-objective deep reinforcement learning in step S5 uses the DuelingDQN algorithm, which divides the state space into three subspaces: implicit carbon emission state, resource allocation state, and process progress state. The action space is used to identify all the core implicit carbon emission sources in S1.

9. The artificial intelligence-based carbon emission reduction management method and system for subway construction as described in claim 1, characterized in that: The implicit carbon emission optimization closed-loop iterative mechanism in S6 is implemented as follows: the execution data of the optimization strategy in actual construction is collected weekly, and after compliance verification, it is fed back to the attention-enhanced deep belief network model in step S3. The parameters of the attention-enhanced deep belief network model are updated by incremental training to adapt to the actual working conditions. After the attention-enhanced deep belief network model is updated, a new round of policy generation is automatically triggered. This new round of policy adjusts the action space weights according to the contribution of implicit carbon emissions in the current scenario, thereby generating a new optimized policy that is adapted to the current scenario.

10. An artificial intelligence-based carbon emission reduction management system for subway construction, characterized in that: It includes a hidden carbon emission feature identification module, a data acquisition module, a hidden carbon emission quantification module, a hidden carbon emission correlation analysis module, an optimization strategy generation module, and a closed-loop iteration module; The hidden carbon emission feature identification module is used to identify hidden carbon emission factors. The data acquisition module is used to collect related data. The hidden carbon emission quantification module is used to extract features from the collected data and output the hidden carbon emission measurement value. The hidden carbon emission correlation analysis module is used to perform correlation strength analysis on the collected data to obtain the final hidden carbon emission measurement value. The optimization strategy generation module: obtains the control strategy through constraints based on the final implicit carbon emission measurement value; The closed-loop iteration module collects the data after the strategy is executed and feeds it back to the implicit carbon emission quantification module.

Images