Construction elevator intelligent dynamic scheduling system and method based on multi-modal deep learning

Abstract

The invention discloses a construction elevator intelligent dynamic scheduling system and method based on multi-modal deep learning, and relates to the technical field of elevator scheduling management, the system comprises a multi-modal data acquisition module, a multi-modal deep learning processing module, a hierarchical decision module, an adaptive learning module, a hybrid reinforcement learning scheduling module, and a conflict prediction and solution module; time sequence data, image data and text requirements are processed at the same time through a multi-mode deep learning architecture, a three-layer neural network decision architecture is adopted to realize full-process intelligent decision from strategy to execution, and a learning strategy and decision weight are automatically adjusted according to a project progress stage in combination with an adaptive learning system. The optimal scheduling scheme can be automatically generated according to information such as material types, quantity and latest completion time submitted by a team, the use efficiency of the construction elevator is effectively improved, resource waste and scheduling conflicts are reduced, and the method is suitable for scenes such as building construction elevator scheduling and a multi-elevator cooperative working environment.

Description

Technical Field

[0001] The present invention relates to the technical field of elevator dispatching and management, and in particular to an intelligent dynamic dispatching system and method for construction elevators based on multimodal deep learning. Background Art

[0002] With the continuous expansion of construction projects and the increasing number of high-rise buildings, construction elevators, as key equipment for vertical transportation on construction sites, have a direct impact on project progress and cost control in terms of scheduling and management efficiency. Currently, there are the following main problems with construction elevator scheduling: 1. Traditional scheduling methods rely on manual experience and lack scientific and forward-looking thinking. Construction site elevator scheduling is usually arranged by on-site managers based on their experience. This is easily influenced by subjective factors and is difficult to cope with complex and changing construction needs.

[0003] 2. Unbalanced resource allocation leads to elevator congestion during certain periods and idleness during other periods. Due to a lack of overall understanding of the construction progress and the needs of each team, multiple teams often compete for elevator resources during certain periods, while elevator utilization is low during other periods.

[0004] 3. Delayed scheduling decisions and inability to respond to emergencies. Traditional scheduling methods are slow to react. When urgent tasks or equipment failures occur, it is difficult to quickly adjust the original plan, affecting construction efficiency.

[0005] 4. Low efficiency of multi-elevator coordination and lack of a unified dispatching mechanism. In a multi-elevator working environment, each elevator often operates independently and lacks a coordination mechanism, resulting in low overall transportation efficiency.

[0006] 5. Insufficient data utilization and inability to learn from historical experience for optimization. Traditional scheduling methods lack systematic analysis and utilization of historical data, making continuous optimization impossible.

[0007] Several existing intelligent dispatching systems have been applied to ordinary passenger elevators, such as elevator group control systems based on fuzzy logic and elevator dispatch optimization methods based on genetic algorithms. However, these technologies primarily target passenger elevators in commercial buildings, with the dispatching objective of minimizing passenger wait times and energy consumption. Construction elevators differ significantly from ordinary passenger elevators in their usage scenarios, dispatching objectives, and constraints. Construction elevators primarily transport construction materials, not personnel, and multiple factors such as material weight, volume, and urgency must be considered. Construction elevator usage patterns exhibit distinct phases and fluctuations, closely tied to project progress. Construction elevator dispatching requires coordinating the needs of multiple construction teams and managing resource competition and conflicts.

[0008] In recent years, artificial intelligence technologies, particularly deep learning and reinforcement learning, have made significant progress in the field of scheduling optimization. Examples include traffic signal control based on deep reinforcement learning and logistics scheduling based on multi-agent systems. These technologies offer new approaches to solving scheduling problems in complex environments. However, applying these advanced technologies to construction elevator scheduling still faces numerous challenges, such as multimodal data processing, dynamic environment adaptation, and multi-objective optimization.

[0009] Therefore, there is an urgent need to develop an intelligent dynamic scheduling system specifically tailored to the characteristics of construction elevators, which can comprehensively process multi-source heterogeneous data, adapt to the needs of different stages of the project, realize the intelligent allocation and dynamic adjustment of construction elevator resources, and improve the utilization efficiency and management level of vertical transportation equipment during the construction process. Summary of the Invention

[0010] In order to solve the above technical problems, the present invention provides a construction elevator intelligent dynamic scheduling system and method based on multimodal deep learning. The following technical solutions are adopted: An intelligent dynamic scheduling system for construction elevators based on multimodal deep learning, characterized by: including a multimodal data acquisition module, a multimodal deep learning processing module, a hierarchical decision module, an adaptive learning module, a hybrid reinforcement learning scheduling module, and a conflict prediction and resolution module; The multimodal data acquisition module is used to collect time series data, image data and text demand data of the construction elevator; The multimodal deep learning processing module is used to process image data, time series data and text demand data; The hierarchical decision-making module is used to realize intelligent decision-making in the whole process from global resource allocation to real-time adjustment; The adaptive learning module is used to automatically adjust the learning strategy and decision weight according to the project progress stage; The hybrid reinforcement learning scheduling module is used to generate a scheduling plan for construction elevators; The conflict prediction and resolution module is used to predict potential scheduling conflicts and automatically generate solutions.

[0011] Optionally, the multimodal data acquisition module collects data based on communication sampling with the elevator's sensors, and the acquisition frequency is dynamically adjusted according to actual needs; The multimodal deep learning processing module includes an LSTM module, a YOLO module, and a Transformer module. The LSTM module is used to process time series data; the YOLO module is used to process image data; and the Transformer module is used to process text demand data. The hierarchical decision-making module includes a strategic layer, a tactical layer, and an execution layer. The strategic layer is based on a global resource allocation network of a graph neural network; the tactical layer is based on a time window optimization network of a variational autoencoder; and the execution layer is based on a real-time adjustment network of deep reinforcement learning. The adaptive learning module includes a meta-learning framework, a multi-task learning component, and a hyperparameter automatic adjustment component; the meta-learning framework is used to dynamically adjust the learning rate; the multi-task learning component is used to optimize the objective function at different stages; the hyperparameter automatic adjustment component is based on Bayesian optimization; The hybrid reinforcement learning scheduling module includes a multi-agent reinforcement learning collaborative decision-making component, a hierarchical reinforcement learning component for processing long-term and short-term decisions, a model-based reinforcement learning prediction component, a model-free reinforcement learning real-time adjustment component, and a reward shaping technology component. The conflict prediction and resolution module includes a spatiotemporal graph convolutional network, a conflict severity assessment component, a multi-objective optimization algorithm, and a decision tree selection component; The spatiotemporal graph convolutional network is used to predict potential conflicts; the conflict severity assessment component is implemented based on the attention mechanism; the multi-objective optimization algorithm is used to generate the Pareto optimal solution set; and the decision tree selection component is used to select the final solution.

[0012] Optionally, the multimodal deep learning processing module also includes a multi-head attention mechanism and a custom loss function. The multi-head attention mechanism is used to fuse information from different modalities; the custom loss function is used to optimize the multi-objective decision-making process. The hierarchical decision-making module also includes an inter-layer information transmission mechanism for realizing information interaction and feedback adjustment between layers; The adaptive learning module also includes an incremental learning mechanism for adapting to engineering changes.

[0013] Optionally, the intelligent dynamic scheduling system for construction elevators based on multimodal deep learning as described in claim 3 is used to implement intelligent dynamic scheduling of construction elevators, comprising the following steps: Step 1: The multimodal data acquisition module collects multimodal data of the construction elevator, which includes time series data, image data, and text demand data; Step 2: Process the multimodal data separately through the multimodal deep learning processing module and fuse the different modal information through the multi-head attention mechanism; Step 3: Generate a scheduling plan through a hierarchical decision-making module. The scheduling plan includes global resource allocation at the strategic level, time window optimization at the tactical level, and real-time adjustment at the execution level. Step 4: Automatically adjust the learning strategy and decision weights according to the project progress stage through the adaptive learning module; Step 5: Optimize the scheduling scheme through the hybrid reinforcement learning scheduling module; Step 6: Predict potential conflicts and automatically generate solutions through the conflict prediction and resolution module; Step 7: Execute the final scheduling plan.

[0014] Optionally, step 8 is also included to collect feedback data for continuous optimization of the intelligent dynamic scheduling system of construction elevators.

[0015] Optionally, the multimodal deep learning process in step 2 includes the following sub-steps: Step 21: Use the LSTM module to process the historical usage time series data; Step 22: Use the YOLO module to process the elevator load image recognition; Step 23: Use the Transformer module to process the team demand text analysis; Step 24: Fusion of different modal information through multi-head attention mechanism; Step 25: Use the custom loss function to optimize the multi-objective decision.

[0016] Optionally, the hierarchical decision-making process in step 3 includes the following sub-steps: Step 31, the strategic layer performs global resource allocation based on GraphNeuralNetwork; Step 32: The tactical layer performs time window optimization based on the variational autoencoder. Step 33: The execution layer makes real-time adjustments based on deep reinforcement learning. Step 34: information exchange and feedback adjustment between layers are realized through the inter-layer information transmission mechanism.

[0017] Optionally, the adaptive learning process in step 4 includes: Step 41, dynamically adjust the learning rate through the meta-learning framework; Step 42: Optimize the objective functions of different stages through multi-task learning; Step 43, automatically adjusting hyperparameters through Bayesian optimization; Step 44: Adapt to engineering changes through incremental learning mechanism.

[0018] Optionally, the hybrid reinforcement learning scheduling process in step 5 includes the following sub-steps: Step 51: Implement collaborative decision-making through multi-agent reinforcement learning; Step 52, processing long-term and short-term decisions through hierarchical reinforcement learning; Step 53, predicting future states through model-based reinforcement learning; Step 54, real-time adjustment through model-free reinforcement learning; Step 55: Optimize learning efficiency through reward shaping technology.

[0019] Optionally, the conflict prediction and resolution process in step 6 includes the following sub-steps: Step 61, predicting potential conflicts through a spatiotemporal graph convolutional network; Step 62, evaluating the severity of the conflict by using an attention-based mechanism; Step 63, generating a Pareto optimal solution set through a multi-objective optimization algorithm; Step 64: Select the final solution through the decision tree.

[0020] In summary, the present invention includes at least one of the following beneficial technical effects: 1. Improve scheduling efficiency: Through a multimodal deep learning architecture, time series data, image data, and text requirements are simultaneously processed to achieve intelligent allocation and dynamic adjustment of construction elevator resources, improving scheduling efficiency by more than 50%.

[0021] 2. Optimize resource utilization: A three-layer neural network decision-making architecture is used to achieve intelligent decision-making throughout the entire process from strategy to execution, increasing resource utilization by 40% and reducing elevator idle time.

[0022] 3. Reduce scheduling conflicts: Integrate a conflict prediction and automatic resolution mechanism based on graph neural networks, with an automatic conflict resolution rate of 95%, significantly reducing the need for manual intervention.

[0023] 4. Rapid response to changes: Combined with an adaptive learning system, the system automatically adjusts learning strategies and decision weights according to the project progress stage. The system response time is less than 3 seconds, enabling rapid response to emergencies.

[0024] 5. Improved prediction accuracy: Through a hybrid reinforcement learning scheduling algorithm, the prediction accuracy exceeds 92%, providing reliable decision support for construction management.

[0025] 6. Achieve multi-elevator collaboration: Through multi-agent reinforcement learning collaborative decision-making components, multiple elevators can work together to improve overall transportation efficiency.

[0026] 7. Continuous self-optimization: Through incremental learning mechanisms and feedback adjustment mechanisms, the system can learn from historical experience and continuously optimize to adapt to engineering changes.

[0027] 8. Reduce management costs: reduce the need for manual scheduling, reduce the workload of managers, and save labor costs.

[0028] 9. Accelerate construction progress: By optimizing vertical transportation efficiency and reducing material waiting time, the construction progress and project quality can be indirectly accelerated.

[0029] 10. Strong adaptability: The system is suitable for a variety of scenarios such as elevator dispatching in high-rise building construction, multi-elevator collaborative working environment, emergency task insertion processing, and emergency adjustment of equipment failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a system overall architecture diagram of a construction elevator intelligent dynamic scheduling system and method based on multimodal deep learning according to an embodiment of the present invention; Figure 2 is a flow chart of multimodal data collection and processing according to an embodiment of the present invention; Figure 3 This is a diagram of the multimodal deep learning architecture of an embodiment of the present invention; Figure 4 is a hierarchical decision framework diagram of an embodiment of the present invention; Figure 5 is a flowchart of the adaptive learning system according to an embodiment of the present invention; Figure 6 is a flow chart of the conflict prediction and resolution mechanism according to an embodiment of the present invention; Figure 7 Schematic diagram of the system interface of an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below with reference to the accompanying drawings.

[0032] The embodiments of the present invention disclose a construction elevator intelligent dynamic scheduling system and method based on multimodal deep learning.

[0033] Reference Figure 1-Figure 7 ,Example 1: Overall system architecture; like Figure 1 As shown, the intelligent dynamic scheduling system for construction elevators based on multimodal deep learning provided by the present invention includes six core modules: a multimodal data acquisition module 10, a multimodal deep learning processing module 20, a hierarchical decision-making module 30, an adaptive learning module 40, a hybrid reinforcement learning scheduling module 50, and a conflict prediction and resolution module 60. The modules exchange information via a data bus 70, and unified scheduling is performed by a system control center 80.

[0034] Multimodal data acquisition module; like Figure 2 As shown, the multimodal data acquisition module 10 includes: The time series data acquisition unit 11 is used to collect time series data such as elevator running speed, elevator load, and elevator floor; Image acquisition units 12, located at key locations inside and outside the construction elevator, are used to capture the type and quantity of materials in the elevator; The text demand collection unit 13 is used to receive and process the elevator demand text information submitted by each construction team.

[0035] The time-series data acquisition unit 11 includes a variety of sensors, such as an inertial measurement unit, a weighing sensor, and a rotary encoder. The sampling frequency can be dynamically adjusted based on actual needs and is generally set to 10 times per second. The image acquisition unit 12 uses a high-definition camera array with a resolution of at least 1080P and a frame rate of at least 30fps to ensure clear images even under complex lighting conditions. The text demand acquisition unit 13 receives demand information submitted by work teams through a mobile terminal application, including key parameters such as material type, quantity, and latest completion time.

[0036] Multimodal deep learning processing module; like Figure 3 As shown, the multimodal deep learning processing module 20 includes: LSTM module 21, which uses a bidirectional LSTM network to process historical usage time series data; YOLO module 22, using the YOLO11 architecture, is used to process elevator load image recognition; Transformer module 23 uses the BERT pre-trained model to process team demand text analysis; Multi-head attention mechanism24, used to fuse information from different modalities; Custom loss function 25 is used to optimize the multi-objective decision-making process.

[0037] The LSTM module 21 employs a bidirectional structure with 128 hidden units, effectively capturing long-term and short-term dependencies in elevator usage. The YOLO module 22, specifically trained for construction materials, can identify 20 common building material types with over 95% accuracy. The Transformer module 23, based on the pre-trained BERT model, uses transfer learning to adapt to specialized construction terminology, achieving text comprehension accuracy exceeding 90%.

[0038] The multi-head attention mechanism24 employs an eight-head attention structure, with each attention head having a dimension of 64. This mechanism effectively integrates information by calculating the correlation between features from different modalities. A custom loss function25 comprehensively considers scheduling efficiency, resource utilization, and conflict rate, balancing the multi-objective optimization problem through a weighted approach.

[0039] Hierarchical decision-making module; like Figure 4 As shown, the hierarchical decision module 30 adopts a three-layer architecture: Strategic layer 31, global resource allocation network based on graph neural network; Tactical layer 32, time window optimization network based on variational autoencoder; Execution layer 33, real-time adjustment network based on deep reinforcement learning; The inter-layer information transmission mechanism 34 is used to realize information interaction and feedback adjustment between layers.

[0040] The strategic layer 31 models construction elevators and teams as nodes in a graph network, with edges representing resource allocation relationships. The GraphSAGE algorithm is used to learn node representations and output a global resource allocation plan. The tactical layer 32 uses a β-VAE architecture with a latent variable dimension of 64. By encoding and decoding time windows, it optimizes the elevator schedule for each team. The execution layer 33 uses a Double DQN architecture with a state space dimension of 128 and an action space consisting of 11 basic scheduling operations, allowing for real-time adjustments to the scheduling plan.

[0041] The inter-layer information transmission mechanism 34 adopts residual connections and gated update units to ensure that the upper-layer decisions can guide the lower-layer execution. At the same time, the lower-layer feedback can also affect the upper-layer strategy adjustment, forming a closed-loop optimization system.

[0042] Adaptive learning modules; like Figure 5 As shown, the adaptive learning module 40 includes: A meta-learning framework41 for dynamically adjusting learning rates; A multi-task learning component 42 for optimizing objective functions at different stages; Hyperparameter automatic tuning component based on Bayesian optimization43; Incremental learning mechanism 44 is used to adapt to engineering changes.

[0043] The meta-learning framework 41 uses the MAML (Model-Agnostic Meta-Learning) algorithm, with an initial inner loop learning rate of 0.01 and an outer loop learning rate of 0.001. This allows for rapid adaptation to new learning tasks based on the project's progress. The multi-task learning component 42 employs hard parameter sharing, with the underlying network sharing four convolutional layers and two task-specific layers each, balancing general feature extraction with task-specific optimization.

[0044] The Bayesian optimization-based hyperparameter automatic tuning component43 uses a Gaussian process regression model with 50 sampling points and an exploration-exploitation balance parameter κ = 2.5. It automatically optimizes 10 key hyperparameters, including the learning rate, batch size, and network depth. The incremental learning mechanism44 employs knowledge distillation to transfer knowledge from the old model to the new model, while retaining 20% ​​of historical data for mixed training with new data, ensuring that the model can adapt to engineering changes without forgetting historical experience.

[0045] Hybrid reinforcement learning scheduling module; The hybrid reinforcement learning scheduling module 50 includes: Multi-agent reinforcement learning collaborative decision-making component51; Hierarchical reinforcement learning handles long-term and short-term decision components 52; Model-based reinforcement learning prediction component 53; Model-free reinforcement learning real-time adjustment component 54; Reward Shaping Technology Component 55.

[0046] The multi-agent reinforcement learning collaborative decision-making component 51 uses the MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm. Each elevator acts as an agent, sharing a critic network but maintaining its own independent actor network, enabling multi-elevator collaboration. The hierarchical reinforcement learning component 52 for long- and short-term decision-making uses a FeudalNetwork architecture, with high-level policies responsible for setting long-term goals and low-level policies for executing short-term actions. The timescale ratio is 1:8.

[0047] The model-based reinforcement learning prediction component 53 uses a WorldModel structure, comprising a VAE encoder, an MDN-RNN dynamic model, and a Controller policy network. It can simulate and predict system states up to 10 time steps into the future. The model-free reinforcement learning real-time adjustment component 54 uses the SAC (Soft Actor-Critic) algorithm with an entropy regularization coefficient α = 0.2 to achieve robust decision-making in uncertain environments. The reward shaping technology component 55 designs a hierarchical reward function with six sub-goals, accelerating the learning process by decomposing potential rewards.

[0048] Conflict prediction and resolution module; like Figure 6 As shown, the conflict prediction and resolution module 60 includes: Spatiotemporal graph convolutional networks61 for predicting potential conflicts; Conflict severity assessment component based on attention mechanism62; Multi-objective optimization algorithm 63 for generating Pareto optimal solution sets; The decision tree selection component 64 is used to select the final solution.

[0049] The spatio-temporal graph convolutional network 61 uses the ST-GCN (Spatio-Temporal Graph Convolutional Network) architecture, consisting of three spatio-temporal convolutional layers. This architecture captures the spatiotemporal dependencies in elevator scheduling and predicts potential scheduling conflicts within the next 30 minutes. The attention-based conflict severity assessment component 62 uses a Transformer encoder structure, taking in a conflict feature vector and outputting a conflict severity score (0-10).

[0050] The multi-objective optimization algorithm 63 uses NSGA-II (Non-dominated Sorting Genetic Algorithm II), with a population size of 100, 50 generations, a crossover probability of 0.8, and a mutation probability of 0.1. This algorithm generates a Pareto-optimal solution set that considers scheduling efficiency, resource utilization, and fairness. The decision tree selection component 64, built on the C4.5 algorithm, takes the current system state and user preferences as input and selects the most appropriate solution from the Pareto solution set.

[0051] Example 2: System workflow; The intelligent dynamic scheduling method for construction elevators based on multimodal deep learning of the present invention comprises the following steps: Step 1: Multimodal data acquisition; The system collects multimodal data of the construction elevator through the multimodal data acquisition module, including: Various sensors are used to collect time series data such as elevator speed, elevator load, and elevator floor, with a frequency of 10 times per second; The image data inside the elevator is collected through a high-definition camera at a frequency of once every 5 seconds; like Figure 7 As shown, the mobile terminal application receives the text information of the elevator demand submitted by the team, including material type, quantity, latest completion time, etc.

[0052] After preprocessing, the collected data is stored in a distributed database, providing a data foundation for subsequent processing. Preprocessing includes outlier detection and interpolation of time series data, image denoising, scaling, and normalization, and word segmentation and vectorization of text data.

[0053] Step 2: Multimodal deep learning processing; The system processes various types of data separately through a multimodal deep learning processing module and fuses different modal information through a multi-head attention mechanism, specifically including: Step 21: Time series data processing; Use the LSTM module to process historical time series data and analyze elevator usage patterns and regularities. The LSTM module uses a bidirectional structure, a 24-hour time window, and a 5-minute step size to output the temporal distribution characteristics of elevator usage.

[0054] Step 22: image data processing; The YOLO module is used to process elevator load image recognition and identify the material type, quantity, and space occupancy within the elevator. The YOLO module uses the YOLO11 architecture, takes a 640×640 input image, and outputs the detected object category, location, and confidence score.

[0055] Step 23: Text data processing; The Transformer module processes text analysis of team requirements and extracts key information such as material type, quantity, and urgency. Based on the BERT pre-trained model, the Transformer module is fine-tuned to accommodate specialized construction terminology and outputs structured requirements information.

[0056] Step 24: Multimodal information fusion; The multi-head attention mechanism fuses information from different modalities, calculates the correlation weights between the modalities, and generates a unified feature representation. The multi-head attention mechanism uses an 8-head structure, with each attention head having a dimension of 64, and outputs a fused 512-dimensional feature vector.

[0057] Step 25: Multi-objective decision optimization; Use a custom loss function to optimize multi-objective decision making and balance multiple objectives such as scheduling efficiency, resource utilization, and fairness. The loss function is in the form of ,in and They represent efficiency loss, resource waste loss and unfair loss respectively. and is the corresponding weight.

[0058] Step 3: Hierarchical decision generation; The system generates scheduling plans through hierarchical decision modules, including global resource allocation at the strategic level, time window optimization at the tactical level, and real-time adjustments at the execution level. Specifically, it includes: Step 31: Strategic level decision making; The strategic layer performs global resource allocation based on a graph neural network. Elevators and teams are modeled as graph network nodes, with edges representing resource allocation relationships. The GraphSAGE algorithm is used to learn node representations, aggregate neighbor information to update node features, and ultimately output a global resource allocation plan.

[0059] Step 32: Tactical level decision making; The tactical layer uses a variational autoencoder to optimize time windows, dividing a 24-hour day into 48 time windows of 30 minutes each. The variational autoencoder encodes the time window requirements into a latent space and generates a smooth and reasonable time window allocation solution by optimizing the latent variable distribution.

[0060] Step 33: Execution layer decision; The execution layer makes real-time adjustments based on deep reinforcement learning. The state space includes information such as the current elevator position, load, and waiting queue, while the action space includes operations such as going up, going down, stopping, and adjusting priority. The DoubleDQN algorithm is used to generate real-time scheduling decisions through experience replay and target network stabilization training.

[0061] Step 34: inter-layer information transmission; Information interaction and feedback adjustment between layers are achieved through the inter-layer information transmission mechanism. The upper-layer decision-making results serve as the constraints of the lower layer, and the lower-layer execution feedback serves as the basis for upper-layer strategy adjustment, forming a closed-loop optimization system.

[0062] Step 4: Adaptive learning adjustment; The system automatically adjusts learning strategies and decision weights based on the project progress stage through the adaptive learning module, including: Step 41: Dynamic adjustment of learning rate; The learning rate is dynamically adjusted through a meta-learning framework, adaptively adjusting the learning rate parameters of each module based on the data characteristics and learning difficulty at different stages of the project. The MAML algorithm is used to quickly adapt to new learning tasks through two-level optimization of the inner and outer loops.

[0063] Step 42: Optimize the objective function; Through multi-task learning, we optimize the objective functions at different stages. The foundational phase focuses on resource allocation fairness, the mid-term on scheduling efficiency, and the final phase on emergency response capabilities. Using hard parameter sharing, the underlying network extracts common features, while the task-specific layers optimize the objectives at each stage.

[0064] Step 43: Automatically adjust hyperparameters; Bayesian optimization is used to automatically adjust hyperparameters, including network structure parameters, optimizer parameters, and reinforcement learning parameters. A Gaussian process regression model is used to find the optimal hyperparameter combination through iterative sampling and evaluation.

[0065] Step 44: Incremental learning adaptation; Adapting to engineering changes through incremental learning, the system retains historical knowledge while learning new patterns and rules. Using knowledge distillation technology, it transfers knowledge from old models to new models while retaining some historical data for mixed training with new data.

[0066] Step 5: Hybrid reinforcement learning scheduling optimization; The system optimizes the scheduling plan through a hybrid reinforcement learning scheduling module, including: Step 51: Multi-agent collaborative decision-making; Collaborative decision-making is achieved through multi-agent reinforcement learning. Each elevator acts as an agent, sharing environmental information but making independent decisions. Using the MADDPG algorithm, each agent has its own actor network but shares a critic network, enabling multi-elevator collaboration.

[0067] Step 52: long-term and short-term decision processing; Hierarchical reinforcement learning is used to handle long-term and short-term decision-making. High-level policies are responsible for setting long-term goals, such as balanced resource allocation, while low-level policies are responsible for executing short-term actions, such as specific scheduling operations. Using a FeudalNetwork architecture with a time scale ratio of 1:8, it achieves coordinated optimization of long-term and short-term goals.

[0068] Step 53: Future state prediction; This system uses model-based reinforcement learning to predict future states, build a dynamic model of the environment, and simulate and predict system evolution. It uses a WorldModel architecture, including a VAE encoder, an MDN-RNN dynamic model, and a Controller policy network, capable of predicting system states up to 10 time steps into the future.

[0069] Step 54: Real-time adjustment processing; Model-free reinforcement learning handles real-time adjustments, learning optimal strategies directly from experience to cope with unexpected situations. Using the SAC algorithm, it achieves robust decision-making in uncertain environments by maximizing the weighted sum of expected return and policy entropy.

[0070] Step 55: Optimize learning efficiency; We optimize learning efficiency through reward shaping techniques, decomposing complex goals into multiple simple sub-goals and designing hierarchical reward functions. We also employ a latent reward decomposition method to accelerate the learning process through intermediate reward signals, improving training efficiency.

[0071] Step 6: Conflict prediction and resolution; The system uses the Conflict Prediction and Resolution module to predict potential conflicts and automatically generate solutions, including: Step 61: Conflict prediction; Predicting potential conflicts through a spatiotemporal graph convolutional network (STGCN) modeled elevator scheduling as a spatiotemporal graph, with nodes representing elevators and shifts and edges representing scheduling relationships. Using the ST-GCN architecture, STGCN captures spatiotemporal dependencies in scheduling through spatiotemporal convolution operations, predicting potential future scheduling conflicts.

[0072] Step 62: Conflict severity assessment; By assessing conflict severity through an attention-based mechanism, factors such as the scope, duration, and criticality of the conflict are considered. Using a Transformer encoder structure, it inputs a conflict feature vector and outputs a conflict severity score, providing a basis for prioritizing solutions.

[0073] Step 63: Generate the optimal solution set; A multi-objective optimization algorithm is used to generate a Pareto optimal solution set, taking into account multiple objectives such as scheduling efficiency, resource utilization, and fairness. The NSGA-II algorithm is used to generate a diverse set of Pareto optimal solutions through non-dominated sorting and congestion distance calculation.

[0074] Step 64: Final solution selection; The final solution is selected through a decision tree. Based on the current system state and user preferences, the most suitable solution is selected from the Pareto solution set. The C4.5 algorithm is used to construct the decision tree. The input includes system state characteristics and user preference parameters, and the output is the final solution.

[0075] Step 7: The system executes the final dispatch plan and sends it to each elevator control system and team mobile terminal; Step 8: Collect feedback data for continuous system optimization, including: Monitor the execution of the plan in real time and collect status data and result data during the execution process; Calculate key performance indicators such as scheduling efficiency, resource utilization, conflict resolution rate, etc.; The feedback data is stored in the database for the next round of learning and optimization.

[0076] Example 3: Application scenario example: For example, a 30-story high-rise building project was equipped with three construction elevators and had 12 teams working simultaneously. Under traditional scheduling methods, each team manually negotiated elevator schedules, which frequently resulted in time conflicts and wasted resources. On average, 15 scheduling conflicts occurred daily, and the average elevator utilization rate was only 60%.

[0077] After implementing the intelligent dynamic scheduling system of the present invention, the system first collects elevator demand and operation data from each team through a multimodal data acquisition module. The multimodal deep learning processing module analyzes time-series data to discover that elevator usage peaks between 9:00 AM and 11:00 AM and 2:00 PM and 4:00 PM. It then analyzes image data to identify material type and load, and analyzes text demand data to extract key information such as material type, quantity, and latest completion time for each team.

[0078] The hierarchical decision-making module generates scheduling plans based on the collected data: the strategic layer assigns 12 teams to three elevators, rationally grouping them according to material type and work floor; the tactical layer allocates the optimal time window for each team, avoiding peak hours and taking into account the latest completion time of material delivery; the execution layer dynamically adjusts the elevator operation strategy according to real-time conditions, optimizing the docking sequence and waiting time.

[0079] The adaptive learning module automatically adjusts system parameters based on the project's progress: focusing on fair resource allocation during the foundation phase, improving transportation efficiency during the structural phase, and refined scheduling and emergency response during the renovation phase. The hybrid reinforcement learning scheduling module enables the coordinated operation of three elevators through multi-agent collaborative decision-making. It handles daily planning and real-time adjustments through hierarchical reinforcement learning, and accelerates the learning process through reward shaping technology.

[0080] The conflict prediction and resolution module predicts potential scheduling conflicts 30 minutes in advance, assesses their severity, generates multiple solutions, and selects the optimal one. For example, if it detects a potential conflict involving three shifts using the elevator at 10:00 AM, the system automatically adjusts one shift's elevator time to 9:30 AM and another's to 10:30 AM, while also optimizing elevator routes to maximize transportation efficiency.

[0081] After implementing this system, scheduling conflicts on this project were reduced to fewer than three per day, average elevator utilization increased to 85%, waiting time for material deliveries decreased by 40%, and the average work efficiency of each team increased by 25%. The system's response time is kept under 2 seconds, and its prediction accuracy reaches 93%, meeting the requirements for efficient, intelligent, and dynamic scheduling.

[0082] Example 4: System deployment method: The system of the present invention can be deployed in a variety of ways to adapt to construction projects of different scales and requirements: Cloud deployment method: The system's core algorithms and data processing modules are deployed on a cloud server, communicating with field equipment over the network. This approach offers advantages such as ample computing resources and strong scalability, making it suitable for large-scale projects. However, it also offers disadvantages such as high network dependency, which can affect real-time performance.

[0083] Edge computing deployment methods: Lightweight models are deployed on edge servers at the construction site, while complex computing tasks are still handled by the cloud. The advantages are reduced network latency, improved real-time performance, and suitable for medium-sized projects. However, the disadvantage is that the computing power of edge devices is limited, making it difficult to deploy complex models.

[0084] Hybrid deployment mode: Using an edge-cloud collaborative architecture, time-critical modules, such as executive-level decision-making, are deployed on edge devices, while computationally intensive modules, such as deep learning training, are deployed in the cloud. This approach offers both real-time performance and computational power, resulting in strong adaptability. However, it also increases system complexity and maintenance costs.

[0085] Different deployment methods can be selected for projects of different sizes: small projects (1-2 elevators) can adopt lightweight edge deployment; medium-sized projects (3-5 elevators) can adopt hybrid deployment; large projects (more than 6 elevators) are recommended to adopt cloud deployment or distributed hybrid deployment.

[0086] Example 5: System interface design: The system of the present invention provides multiple interfaces to facilitate use by different user roles: Device management interface: The Equipment Management interface, primarily for project managers, provides functions for monitoring elevator equipment status, viewing scheduling plans, and setting system parameters. The left side of the interface features a navigation menu, the center displays real-time elevator status, and the right side features an operator control panel. The status display visually displays information such as each elevator's location, load, and task queue, supporting multi-dimensional data filtering and time range selection.

[0087] Team reservation interface: The team reservation interface, designed for construction team leaders, allows users to schedule and query elevator usage times. The interface displays available time slots in a calendar view, allowing users to drag and drop to select a time slot. The system automatically checks for conflicts and suggests alternatives. Once a reservation is successfully made, users can view details, modify, or cancel it, and the system will automatically send a notification.

[0088] Team demand submission interface: The team request submission interface is used for teams to submit detailed elevator request information. The interface includes a material information input area, time requirement settings, a priority selector, and a text box for describing special requests. Material information supports type selection, quantity input, and image upload. Time requirements include setting expected and latest completion times. Priorities are categorized as normal, important, and urgent. The text box for special requests supports natural language input. The right side of the interface displays the system's real-time optimal time recommendations based on the current scheduling situation, helping users make informed choices.

[0089] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent dynamic dispatching system for construction elevators based on multimodal deep learning, characterized by: It includes multimodal data acquisition module, multimodal deep learning processing module, hierarchical decision-making module, adaptive learning module, hybrid reinforcement learning scheduling module, and conflict prediction and resolution module; The multimodal data acquisition module is used to collect time series data, image data and text demand data of the construction elevator; The multimodal deep learning processing module is used to process image data, time series data and text demand data; The hierarchical decision-making module is used to realize intelligent decision-making in the whole process from global resource allocation to real-time adjustment; The adaptive learning module is used to automatically adjust the learning strategy and decision weight according to the project progress stage; The hybrid reinforcement learning scheduling module is used to generate a scheduling plan for construction elevators; The conflict prediction and resolution module is used to predict potential scheduling conflicts and automatically generate solutions.

2. The intelligent dynamic scheduling system for construction elevators based on multimodal deep learning according to claim 1, characterized in that: The multimodal data acquisition module collects data based on communication sampling with the elevator's sensors, and the acquisition frequency is dynamically adjusted according to actual needs; The multimodal deep learning processing module includes an LSTM module, a YOLO module, and a Transformer module. The LSTM module is used to process time series data; the YOLO module is used to process image data; and the Transformer module is used to process text demand data. The hierarchical decision-making module includes a strategic layer, a tactical layer, and an execution layer. The strategic layer is based on a global resource allocation network of a graph neural network; the tactical layer is based on a time window optimization network of a variational autoencoder; and the execution layer is based on a real-time adjustment network of deep reinforcement learning. The adaptive learning module includes a meta-learning framework, a multi-task learning component, and a hyperparameter automatic adjustment component; the meta-learning framework is used to dynamically adjust the learning rate; the multi-task learning component is used to optimize the objective function at different stages; the hyperparameter automatic adjustment component is based on Bayesian optimization; The hybrid reinforcement learning scheduling module includes a multi-agent reinforcement learning collaborative decision-making component, a hierarchical reinforcement learning component for processing long-term and short-term decisions, a model-based reinforcement learning prediction component, a model-free reinforcement learning real-time adjustment component, and a reward shaping technology component. The conflict prediction and resolution module includes a spatiotemporal graph convolutional network, a conflict severity assessment component, a multi-objective optimization algorithm, and a decision tree selection component; The spatiotemporal graph convolutional network is used to predict potential conflicts; the conflict severity assessment component is implemented based on the attention mechanism; the multi-objective optimization algorithm is used to generate the Pareto optimal solution set; and the decision tree selection component is used to select the final solution.

3. The intelligent dynamic scheduling system for construction elevators based on multimodal deep learning according to claim 2 is characterized in that: The multimodal deep learning processing module also includes a multi-head attention mechanism and a custom loss function. The multi-head attention mechanism is used to fuse information from different modalities; the custom loss function is used to optimize the multi-objective decision-making process. The hierarchical decision-making module also includes an inter-layer information transmission mechanism for realizing information interaction and feedback adjustment between layers; The adaptive learning module also includes an incremental learning mechanism for adapting to engineering changes.

4. A construction elevator intelligent dynamic scheduling method based on multimodal deep learning, characterized by: The intelligent dynamic scheduling system for construction elevators based on multimodal deep learning as described in claim 3 is used to implement intelligent dynamic scheduling of construction elevators, comprising the following steps: Step 1: The multimodal data acquisition module collects multimodal data of the construction elevator, which includes time series data, image data, and text demand data; Step 2: Process the multimodal data separately through the multimodal deep learning processing module and fuse the different modal information through the multi-head attention mechanism; Step 3: Generate a scheduling plan through a hierarchical decision-making module. The scheduling plan includes global resource allocation at the strategic level, time window optimization at the tactical level, and real-time adjustment at the execution level. Step 4: Automatically adjust the learning strategy and decision weights according to the project progress stage through the adaptive learning module; Step 5: Optimize the scheduling scheme through the hybrid reinforcement learning scheduling module; Step 6: Predict potential conflicts and automatically generate solutions through the conflict prediction and resolution module; Step 7: Execute the final scheduling plan.

5. The method for intelligent dynamic scheduling of construction elevators based on multimodal deep learning according to claim 4 is characterized in that: It also includes step 8, collecting feedback data for continuous optimization of the construction elevator intelligent dynamic scheduling system.

6. The intelligent dynamic scheduling method for construction elevators based on multimodal deep learning according to claim 4 is characterized in that: The multimodal deep learning process in step 2 includes the following sub-steps: Step 21: Use the LSTM module to process the historical usage time series data; Step 22: Use the YOLO module to process the elevator load image recognition; Step 23: Use the Transformer module to process the team demand text analysis; Step 24: Fusion of different modal information through multi-head attention mechanism; Step 25: Use the custom loss function to optimize the multi-objective decision.

7. The method for intelligent dynamic scheduling of construction elevators based on multimodal deep learning according to claim 4 is characterized in that: The hierarchical decision-making process in step 3 includes the following sub-steps: Step 31, the strategic layer performs global resource allocation based on GraphNeuralNetwork; Step 32: The tactical layer performs time window optimization based on the variational autoencoder. Step 33: The execution layer makes real-time adjustments based on deep reinforcement learning. Step 34: information exchange and feedback adjustment between layers are realized through the inter-layer information transmission mechanism.

8. The method for intelligent dynamic scheduling of construction elevators based on multimodal deep learning according to claim 4 is characterized in that: The adaptive learning process in step 4 includes: Step 41, dynamically adjust the learning rate through the meta-learning framework; Step 42: Optimize the objective functions of different stages through multi-task learning; Step 43, automatically adjusting hyperparameters through Bayesian optimization; Step 44: Adapt to engineering changes through incremental learning mechanism.

9. The method for intelligent dynamic scheduling of construction elevators based on multimodal deep learning according to claim 4, characterized in that: The hybrid reinforcement learning scheduling process in step 5 includes the following sub-steps: Step 51: Implement collaborative decision-making through multi-agent reinforcement learning; Step 52, processing long-term and short-term decisions through hierarchical reinforcement learning; Step 53, predicting future states through model-based reinforcement learning; Step 54, real-time adjustment through model-free reinforcement learning; Step 55: Optimize learning efficiency through reward shaping technology.

10. The method for intelligent dynamic scheduling of construction elevators based on multimodal deep learning according to claim 4, characterized in that: The conflict prediction and resolution process in step 6 includes the following sub-steps: Step 61, predicting potential conflicts through a spatiotemporal graph convolutional network; Step 62, evaluating the severity of the conflict by using an attention-based mechanism; Step 63, generating a Pareto optimal solution set through a multi-objective optimization algorithm; Step 64: Select the final solution through the decision tree.

Images