A method and device for coordinated control of multiple motor devices in property elevator management, and a storage medium

Abstract

The application relates to the technical field of property management, and provides a coordinated control method and device for multiple motor devices in property elevator management and a storage medium. The method comprises the following steps: S1, collecting real-time people flow data and motor operation state data; S2, generating fusion data with a structured metadata format; S3, constructing a distributed people flow prediction model in a building, and outputting a predicted passenger flow demand matrix of each area in a future period; S4, generating a cooperative scheduling instruction for an elevator traction motor cluster; and S5, controlling the elevator traction motor cluster according to the cooperative scheduling instruction. The device comprises a data collection module, a data fusion module, a passenger flow prediction module and a control execution module for executing the above method. The storage medium stores a computer program, and the computer program is executed by a processor to realize the method. In addition, the overall optimization of transportation efficiency, energy consumption management and user experience is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of property management, and in particular to a method and device for coordinated control of multiple motor devices in property elevator management, and a storage medium. BACKGROUND

[0002] Currently, elevator group control systems are widely used in large commercial buildings and high-rise buildings as the core solution for vertical transportation. The mainstream group control system is based on fixed rules to achieve scheduling, such as the shortest waiting time or the smallest call response strategy. Such systems essentially belong to the reactive scheduling mode, which can only respond passively after the passenger makes a call request, and have inherent defects such as response lag and insufficient global optimization.

[0003] Although the prior art attempts to improve by adding sensors to monitor car load and other methods, it still faces problems such as single sensing dimension, limited prediction ability, and isolated subsystems. These methods are difficult to accurately predict the distribution and trend of passenger flow in the building, and cannot support the collaborative and intelligent scheduling of the elevator system, resulting in high energy consumption, uneven equipment wear and tear, and congestion during peak hours.

[0004] Therefore, there is an urgent need for a coordinated control method that can deeply integrate multi-source sensing data, use artificial intelligence for accurate prediction, and collaboratively optimize control of motor clusters, to fundamentally improve the operating efficiency of the elevator group, reduce energy consumption, and improve user experience from the system level. SUMMARY

[0005] The purpose of the present application is to achieve active intelligent scheduling of elevator groups, improve the overall operating efficiency of elevator groups, reduce system energy consumption, and optimize the elevator experience.

[0006] To achieve the above purpose, in one embodiment of the present application, a method for coordinated control of multiple motor devices in property elevator management is proposed. The present application uses advanced technology centered on artificial intelligence, through an integrated architecture of multi-modal sensing, accurate prediction and collaborative control, to realize the transition of the elevator system from passive response to active scheduling. The method for coordinated control of multiple motor devices in property elevator management includes:

[0007] S1. Collect real-time passenger flow data and motor operating state data through a multi-modal sensor network deployed in the building;

[0008] S2. Perform multi-modal fusion processing on the real-time passenger flow data and the motor operating state data to generate fusion data in a structured metadata format;

[0009] S3. Based on the fusion data, build a distributed passenger flow prediction model in the building through an artificial intelligence model, and output a predicted passenger flow demand matrix for each region in the future period.

[0010] S4. inputting the predicted passenger flow demand matrix into a multi-agent collaborative control model to generate a collaborative scheduling instruction for a cluster of elevator traction motors;

[0011] S5. controlling the cluster of elevator traction motors according to the collaborative scheduling instruction to optimize overall operation efficiency.

[0012] Optionally, in some embodiments, the step S1 comprises:

[0013] S11. collecting video stream data through a high-definition network camera deployed at a key node of a building;

[0014] S12. performing real-time analysis on the video stream data to generate structured metadata including the number of people, moving direction and location information;

[0015] S13. collecting motor operating state data through a motor sensor deployed in an elevator car and a hoistway, the motor operating state data including current, voltage, temperature and vibration data;

[0016] S14. aligning and packaging the structured metadata and the motor operating state data to form the real-time passenger flow data and the motor operating state data.

[0017] Optionally, in some embodiments, the step S2 comprises:

[0018] S21. receiving the real-time passenger flow data and the motor operating state data from a multi-modal sensor network;

[0019] S22. performing timestamp alignment and data cleaning on the real-time passenger flow data and the motor operating state data to obtain standardized time series data;

[0020] S23. fusing the standardized time series data with external context data, the external context data including calendar event information and environmental information;

[0021] S24. constructing spatio-temporal semantic features based on the fused data to generate the fused data in the structured metadata format.

[0022] Optionally, in some embodiments, the artificial intelligence model is a spatio-temporal graph convolution network model, and the step S3 comprises:

[0023] S31. constructing a spatio-temporal graph structure representing the association relationship between regions based on the fused data in the structured metadata format;

[0024] S32. processing the spatio-temporal graph structure through the spatio-temporal graph convolution network model to extract regional association features in the spatial dimension.

[0025] S33. The region correlation features are processed using a multi-head self-attention mechanism to capture long-term dependencies in the time dimension;

[0026] S34. Based on the enhanced spatiotemporal feature sequence, the predicted passenger flow demand matrix of each region in the future period is output.

[0027] Steps S31 to S34 constitute a complete and internally causally linked passenger flow prediction technical solution. Through multi-step cooperative and progressive processing, the prediction accuracy and spatiotemporal perception ability are significantly improved, thereby providing a reliable basis for intelligent scheduling of the elevator cluster. Step S31 constructs a spatiotemporal graph structure representing the correlation between regions. This structure abstracts the floors, corridors, elevator halls, etc. in the physical space as nodes of the graph, and the possibility of people flow as edges, and embeds the state changes in the time dimension. This step changes the modeling method in the traditional method, which regards each region as an isolated unit, providing a data structure basis for capturing the cross-regional passenger flow migration pattern. Step S32 processes the spatiotemporal graph structure through a spatiotemporal graph convolution network. The graph convolution operation can effectively extract complex region correlation features in the spatial dimension, such as the interaction between adjacent floors or functionally complementary regions (such as conference rooms and elevator halls). This step helps to effectively express spatial dependency and achieve spatial semantic understanding of passenger flow distribution in the building. In step S33, the multi-head self-attention mechanism is used to process the extracted spatial feature sequence, which can capture long-term dependencies at different intervals in the time dimension in parallel, such as identifying periodic or event-based passenger flow patterns such as the morning peak, midday break, and large meeting dispersal. This overcomes the limitations of traditional recurrent neural networks in modeling long-term temporal correlations, ensuring the temporal coherence and accuracy of the model's prediction of future passenger flow trends.

[0028] Step S34 then outputs the predicted passenger flow demand matrix of each region in the future period based on the enhanced spatiotemporal feature sequence. This matrix not only contains the number of passengers predicted, but also implicitly contains flow direction and intensity information. The high-precision output directly benefits from the deep fusion and extraction of spatiotemporal features in the previous steps, providing a key and reliable input for subsequent scheduling decisions. This enables the system to transition from passive response to proactive scheduling, ultimately reducing passenger waiting time, optimizing energy consumption, and balancing equipment wear and tear.

[0029] Therefore, the four steps S31 to S34 are closely linked and gradually deepen the understanding of the spatiotemporal dynamics of passenger flow, collectively solving the core technical problem of lagging response and insufficient optimization in the prior art due to single perception dimension and lack of spatiotemporal globality in prediction.

[0030] In some embodiments, possible alternatives or variations are also proposed, and the step S3 further comprises:

[0031] S31’. Based on the fusion data with structured metadata format, a space-time graph neural network with building function areas as nodes is constructed;

[0032] S32’. Dynamic attention weights between adjacent nodes in the space-time graph neural network are calculated;

[0033] S33’. Node enhanced representations are generated by aggregating feature information of adjacent nodes according to the dynamic attention weights;

[0034] S34’. The node enhanced representations are input into a time convolution network to predict the predicted passenger flow demand matrix of each area in the future period.

[0035] The technical scheme constructs an advanced building passenger flow prediction method. In step S31’, a space-time graph neural network with building function areas as nodes is constructed, the physical space structure is converted into a calculable graph model, providing a basis for spatial relationship modeling, correlating discrete areas through a graph structure, providing topological support for subsequent feature propagation, and helping to express complex spatial dependencies between areas. Step S32’ calculates the dynamic attention weights between adjacent nodes in the graph, which realizes the adaptive quantization of the spatial correlation strength, and can dynamically adjust the influence degree between different areas according to the real-time passenger flow state and external context (such as activity events), thereby more finely capturing spatial heterogeneity and overcoming the limitations of fixed weight or simple adjacency relationship modeling. Step S33’ aggregates adjacent node features according to the weights to generate node enhanced representations, which realizes effective propagation and fusion of information in the spatial dimension, so that the feature representation of each node not only contains its own state, but also absorbs context information from related areas, significantly enhancing the richness and discriminability of feature expression. Step S34’ inputs the enhanced node representations into a time convolution network to complete the prediction of passenger flow demand in the future period. The time convolution network can efficiently capture medium and long term time dependencies, avoiding the gradient disappearance problem of recurrent neural networks, thereby realizing more stable and accurate prediction of passenger flow trends.

[0036] The above four steps are gradually promoted and closely coupled: step S31’ provides structural prior, step S32’ realizes dynamic correlation perception, step S33’ completes spatial information fusion, and step S34’ is responsible for time series prediction. This technical path effectively solves the prediction problem caused by the dynamic nature of passenger flow, spatial heterogeneity and event disturbance in the building environment through the combination of structured space-time feature learning and dynamic attention mechanism, and its prediction accuracy and robustness are significantly better than traditional time series prediction or simple graph model methods. The technical scheme effectively improves the perception and decision-making ability of the elevator dispatching system.

[0037] Optionally, in some embodiments, the motor operating state data is real-time state data of the elevator traction motor cluster, and the step S4 comprises:

[0038] S41. inputting the predicted passenger flow demand matrix and the real-time state data of the elevator traction motor cluster into a fuzzy logic inference engine to generate a preliminary scheduling preference vector for each elevator; S41. inputting the predicted passenger flow demand matrix and the real-time state data of the elevator traction motor cluster into a fuzzy logic inference engine to generate a preliminary scheduling preference vector for each elevator;

[0039] S42. constructing a multi-objective model predictive control optimization problem that integrates system total energy consumption, passenger total waiting time and equipment total wear cost, with the preliminary scheduling preference vector as a soft constraint; S43. solving the multi-objective model predictive control optimization problem by using an effective set algorithm to obtain a set of optimal control sequences within a limited prediction time domain;

[0040] S44. extracting the optimal instruction at the current time from the optimal control sequence to generate a coordinated scheduling instruction for the elevator traction motor cluster.

[0041] The four steps S41 to S44 together constitute a hierarchical coordinated decision-making and optimization control process. This process combines forward-looking passenger flow prediction with complex multi-objective system optimization problems, and through a sequential processing mechanism, it realizes efficient and flexible scheduling of the elevator traction motor cluster, effectively solving the core problems of insufficient global optimization and response lag in the prior art.

[0042] The scheme starts from step S41, the predicted passenger flow demand matrix and the real-time state data of the elevator traction motor cluster are jointly input into the fuzzy logic inference machine, the core function of which is to handle the preliminary trade-off between the system uncertainty and multiple objectives. Fuzzy logic is good at converting language variables such as "predicted high passenger flow at a certain floor" into preliminary scheduling preferences for each elevator based on expert experience or rule base, for example, "preferentially schedule nearby low-load elevators to assist". This step outputs a preliminary scheduling preference vector, which provides an important initial decision direction and soft constraint for subsequent precise optimization, avoiding blind search of the optimization algorithm in a huge solution space, thereby improving the calculation efficiency. Step S42 formally constructs a multi-objective model predictive control optimization problem that integrates the total energy consumption of the system, the total waiting time of the passengers, and the total wear cost of the equipment, with the above-mentioned preliminary scheduling preference vector as a soft constraint. This step integrates multiple conflicting optimization objectives (for example: shortening the waiting time may require the elevator to start and stop more frequently, thereby increasing the energy consumption and wear) into a unified mathematical framework. The introduction of model predictive control enables the optimization to be based not only on the current state but also to predict the system behavior in the future finite time domain and make decisions based on it, thereby realizing truly forward-looking scheduling.

[0043] Step S43 uses the efficient set algorithm to rollingly solve the above-mentioned multi-objective optimization problem in a limited prediction time domain. This algorithm is particularly suitable for solving quadratic programming problems with constraints, and has high calculation efficiency and good numerical stability, which can meet the stringent real-time requirements of elevator scheduling. The "rolling solution" mechanism means that the system re-solves the optimization problem at each control period and only executes the optimal control quantity at the current time, and repeats the process based on the new system state at the next moment. This mechanism enables the scheduling strategy to continuously adapt to the dynamic changes in the system state, such as the update of passenger flow prediction or the sudden state of a certain elevator, thereby ensuring the robustness and adaptability of the scheduling scheme. Step S44 extracts the optimal instruction at the current time from the optimal control sequence obtained by rolling solution, and distributes and executes it. At this point, a closed loop that deeply integrates AI prediction, multi-objective optimization theory, and real-time control technology is completed. Through the close cooperation and interaction of these steps, the system can ultimately optimize the overall operating efficiency, which is specifically reflected in multiple aspects, such as significantly reducing the total passenger waiting time, reducing the total system energy consumption, and balancing the equipment wear.

[0044] Optionally, in some embodiments, the step S5 comprises:

[0045] S51. Distribute the collaborative scheduling instruction to the main controller of the elevator group control system;

[0046] S52. Analyze and convert the instruction into specific control parameters for the corresponding traction motor through the main controller;

[0047] S53. Drive the elevator hoisting motor cluster to operate according to the specific control parameters, and collect operation feedback data in real time.

[0048] Optionally, in some embodiments, the step S12 further comprises:

[0049] S121. Perform real-time analysis on the video stream data by running a lightweight deep learning model on the edge computing node;

[0050] S122. Extract personnel positions and motion trajectories in the video frames by target detection and multi-target tracking algorithms;

[0051] S123. Generate the structured metadata containing timestamp, camera ID, number of people, moving direction, and position information based on the motion trajectories.

[0052] Optionally, in some embodiments, the step S23 further comprises:

[0053] S231. Receive the standardized time series data and the external context data, the external context data including calendar event information and environmental information;

[0054] S232. Perform time dimension alignment and feature encoding on the standardized time series data and the external context data;

[0055] S233. Fuse the encoded standardized time series data and external context data through an attention weighting mechanism;

[0056] S234. Output the fused multi-source data feature representation.

[0057] Optionally, in some embodiments, the step S24 further comprises:

[0058] S241. Extract a feature vector representing the spatio-temporal correlation between regions based on the fused multi-source data feature representation;

[0059] S242. Associate map the feature vector with predefined functional semantic labels;

[0060] S243. Generate the fused data in the structured metadata format containing spatio-temporal attributes and functional semantics.

[0061] Optionally, in some embodiments, the multi-agent collaborative control model generates the collaborative scheduling instructions by maximizing the following collaborative utility function:

[0062]

[0063] wherein,

[0064] A t =( a t 1 , a t 2 ,…, a t N ) represents the joint action vector of all agents at time t ;

[0065] O t represents the joint observation vector, which is composed of the predicted passenger flow demand matrix and the motor operating state data;

[0066] f k (•) is the utility function of the k th optimization objective, including the total energy consumption of the system, the total waiting time of passengers, and the total wear cost of equipment;

[0067] w k is the adaptive weight coefficient corresponding to each objective;

[0068] P(A t ) represents the action distribution output by the multi-agent policy;

[0069] Q(A t ) represents the action prior distribution defined by the preliminary scheduling preference vector generated by the fuzzy logic inference;

[0070] D KL (•││•) represents the KL divergence, which is used to measure the deviation between the action distribution and the prior distribution;

[0071] λ is the regularization coefficient, which is used to balance the optimization objective and the degree of policy deviation.

[0072] Optionally, in some embodiments, a coordination control device for implementing the above method is also provided, comprising:

[0073] a data acquisition module configured to acquire real-time passenger flow data and motor operating state data through a multi-modal sensor network deployed in the building;

[0074] a data fusion module configured to perform multi-modal fusion processing on the real-time passenger flow data and the motor operating state data to generate fusion data with a structured metadata format;

[0075] A passenger flow prediction module is configured to construct a distributed passenger flow prediction model in a building based on the fusion data by an artificial intelligence model, and output a predicted passenger flow demand matrix of each area in a future period;

[0076] A collaborative decision-making module is configured to input the predicted passenger flow demand matrix into a multi-agent collaborative control model to generate a collaborative scheduling instruction for the elevator traction motor cluster;

[0077] A control execution module is configured to control the elevator traction motor cluster according to the collaborative scheduling instruction to optimize the overall operation efficiency.

[0078] In some other embodiments of the present application, a computer readable storage medium is provided, which stores a computer program. The program is executed by a processor to implement the steps of the method of any other embodiment.

[0079]

Technical effects

[0080] The present scheme constructs a complete coordinated control method for multiple motor devices through steps S1 to S5, effectively solving the technical problems of elevator cluster response lag and low energy efficiency in property management.

[0081] Step S1 synchronously collects real-time passenger flow and motor operation state data through a multi-modal sensor network, providing comprehensive and real-time sensing input for the system, overcoming the scheduling deviation caused by single data source or update delay in traditional systems. Step S2 performs fusion processing on multi-source heterogeneous data to generate fusion data with a structured metadata format, providing high-quality and standardized input for subsequent artificial intelligence models. Step S3 constructs a distributed passenger flow prediction model based on the fusion data, accurately outputting a predicted passenger flow demand matrix of each area in a future period, realizing accurate prediction from multi-modal sensing. Step S4, as the intelligent decision-making core of the system, inputs the predicted passenger flow demand matrix into a multi-agent collaborative control model to generate a globally optimal collaborative scheduling instruction through a game optimization mechanism; this step fully utilizes the output results of the previous steps, realizing efficient conversion of prediction information to control strategy. Step S5 finally executes the coordinated control instruction to realize precise control of the elevator traction motor cluster, forming a complete closed loop from sensing, prediction, decision-making to execution. There is a strict technical correlation between each step: the multi-modal data output by step S1 is the basis for fusion processing in step S2; the fusion data generated by step S2 is the prerequisite for constructing the prediction model in step S3; the predicted passenger flow demand matrix output by step S3 is the input of the multi-agent collaborative model in step S4; and the scheduling instruction generated by step S4 is the basis for control execution in step S5.

[0082] To sum up, the cascading technology chain ensures the improvement of the overall operation efficiency of the system, solves the problems of response delay and resource waste in elevator cluster scheduling. In addition, the device and storage medium provided by the present application provide specific hardware and software carriers for the implementation of the above-mentioned method, facilitate the popularization and application of the technology, and constitute a complete technical solution. BRIEF DESCRIPTION OF DRAWINGS

[0083] The accompanying drawings, which are included to provide a further understanding of the application and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0084] Figure 1 A flowchart of a coordination control method for multiple motor devices in property elevator management according to an embodiment of the present application;

[0085] Figure 2 A flowchart of the main steps of a coordination control method for multiple motor devices in property elevator management according to another embodiment of the present application;

[0086] Figure 3 A block diagram of a device for a coordination control method for multiple motor devices in property elevator management according to an exemplary embodiment. DETAILED DESCRIPTION

[0087] The inventors, through a large number of researches, adopt advanced technologies with artificial intelligence as the core, through an integrated architecture of multi-modal perception, accurate prediction and collaborative control, propose a coordination control method for multiple motor devices in property elevator management, and realize the leap of elevator system from passive response to active scheduling.

[0088] Several embodiments are presented below to specifically implement the corresponding technical solutions of the present application.

[0089]

Embodiment 1

[0090] As shown in Figure 1 , a flowchart of a coordination control method for multiple motor devices in property elevator management is shown, and the specific implementation process of the present application will be described in detail below with an elevator intelligent scheduling system deployed in a super high-rise office building as an example. The office building has 50 floors above ground and is equipped with 16 elevators. During the morning peak period on weekdays, it faces serious passenger flow pressure and scheduling challenges.

[0091] In step S1 of this embodiment, multi-spectral high-definition network cameras are deployed in the elevator halls and key passage nodes of each floor of the super-high-rise office building. The cameras have infrared night vision and wide dynamic range functions, collect 1080P video stream data at a rate of 30 frames per second, and perform real-time H.265 encoding and local caching through the edge computing nodes deployed in the floor's weak current room, while marking the time stamp and device ID identifier.

[0092] The load changes are monitored in real time by a piezoelectric load sensor installed at the bottom of the elevator car, and the load data is recorded in kilograms, associated with the elevator ID and running direction state, while collecting current, voltage, temperature, and vibration data from the traction motor driver at a sampling frequency of 100 Hz, and transmitting them to the edge gateway through the CAN bus.

[0093] The existing access control system and WiFi probe of the building are used to capture personnel card swiping records and mobile phone MAC addresses, generate personnel trajectory sequences, and infer real-time positions combined with Bluetooth beacon data, to build a personnel distribution heat map with floors as units, with a data update interval of 5 seconds.

[0094] Then, access external event data interfaces to obtain weather forecasts, holiday arrangements, and conference schedule information, convert the external event data into structured JSON format, and align the time stamps based on NTP protocol with the sensor data to form a multi-source heterogeneous data stream.

[0095] In this embodiment, all collected data is transmitted to the central data collection server through industrial Ethernet, and data integrity verification and format standardization are performed to remove duplicate or abnormal records, and finally generate a set of real-time people flow data and motor running state data with space-time labels.

[0096] Step S2 receives real-time people flow data and motor running state data from the multi-modal sensor network, aligns the time stamps of the real-time people flow data and removes outliers, detects and repairs data missing points using a sliding window algorithm with a window size of 10 seconds, and generates cleaned standardized time series data.

[0097] The standardized time series data is fused with external context data at a feature level in step S2, the external context data including calendar event information and environment information, a weighting mechanism is used to calculate the weight of each data source, the weight is dynamically adjusted based on historical correlation, and a multi-source feature tensor is generated after fusion. Based on the multi-source feature tensor, spatio-temporal semantic features are extracted, a convolutional neural network is used to extract spatial local features, and a long short-term memory network is used to capture time dependence, and a feature vector representing the flow density, moving speed and residence time of each floor area is output. Then, the feature vector is mapped to a predefined functional semantic label, the functional semantic label includes "office area", "dining area" and "transportation hub", and a fusion data with a structured metadata format is generated, which is stored in Avro format to support serialization and fast query. The fusion data is subjected to Z-score standardization processing to eliminate dimensional differences and compressed into a low-dimensional embedding space to form a unified data representation for downstream models.

[0098] In step S3, a spatio-temporal graph structure representing the correlation of each floor area is constructed based on the fusion data with a structured metadata format, the node represents the floor number, the edge represents the elevator accessible path and the walking distance, and the edge weight is dynamically calculated based on the historical flow interaction frequency.

[0099] In step S3, the spatio-temporal graph structure is processed by a spatio-temporal graph convolutional network model, a graph convolutional layer is used to extract the correlation features of the areas in the spatial dimension, and a gated recurrent unit is used to capture the time dynamics, and a multi-scale spatio-temporal feature sequence is output.

[0100] At the same time, the multi-head self-attention mechanism is used to process the multi-scale spatio-temporal feature sequence, the attention weight of different time steps is calculated, the long-term dependence is captured, and the external event attention module is fused to dynamically adjust the influence factor of events such as conferences and weather.

[0101] Then, based on the enhanced spatio-temporal feature sequence, the elevator demand intensity and flow density of each floor in the next 15-60 minutes are predicted by a fully connected layer, and a predicted passenger flow demand matrix is output, the matrix dimension is the number of floors x the number of time intervals, and the numerical value is normalized to the range of 0-1.

[0102] Then, an online learning mechanism is introduced to adjust the model parameters based on real-time feedback data, and an incremental gradient descent algorithm is used to update the weight to ensure that the prediction accuracy is adaptively improved.

[0103] In step S4, the predicted passenger flow demand matrix and the real-time state data of the elevator traction motor cluster are jointly input into a fuzzy logic reasoner, the input variables are defined as demand intensity, load rate and motor temperature, and the output is a scheduling preference vector of each elevator, the vector dimension is consistent with the number of elevators.

[0104] With the scheduling preference vector as a soft constraint, a multi-objective model predictive control optimization problem is constructed, a target function including total energy consumption of the system, total waiting time of passengers and total wear cost of equipment, and a weighted summation method is used to integrate the multi-objective, and weights are dynamically adjusted based on time period.

[0105] Then, the multi-objective model predictive control optimization problem is solved by using an efficient set algorithm within a limited prediction time domain, the prediction time domain is set to 5 minutes, the control time domain is 1 minute, and a group of optimal control sequences is obtained, the sequences including speed and acceleration instructions of each elevator.

[0106] Through step S4, the optimal instruction at the current time is extracted from the optimal control sequence to generate the coordinated scheduling instruction of the elevator traction motor cluster, the instruction format is a CANopen protocol frame, including the target floor, the running curve and the priority identifier. A virtual call injection mechanism is introduced, virtual calls are injected into the group control system before physical calls occur according to the predicted passenger flow demand matrix, triggering elevator pre-scheduling and reducing response delay.

[0107] In step S5 of the embodiment, the coordinated scheduling instruction is distributed to the main controller of the elevator group control system, the main controller parses the instruction and converts it into specific control parameters of the corresponding traction motor, including target speed, acceleration curve and start-stop timing, and the parameters are issued through the PROFIBUS-DP bus.

[0108] According to the specific control parameters, the elevator traction motor cluster is driven to run, realizing dynamic speed curve optimization, calculating the optimal acceleration curve based on load and stroke in real time to reduce mechanical impact and energy consumption, while coordinating the energy feedback timing of multiple elevators to avoid power grid impact. Real-time collection of motor running feedback data, including current, voltage, temperature and vibration data, with a sampling frequency of 1 kHz, and calculation of key performance indicators such as energy efficiency ratio and comfort index after denoising by a digital filter.

[0109] Then, based on the running feedback data, the control parameters are dynamically adjusted, the PID controller is used to fine-tune the speed curve, forming a closed-loop optimization to ensure continuous and efficient operation of the system, and operation logs are recorded for subsequent analysis. At the same time, when abnormal vibration or temperature exceeds the threshold value is detected, the safety protection mechanism is automatically triggered to reduce the running speed or switch to the standby elevator, ensuring system reliability. Those skilled in the art should understand that some terms in the embodiment can be understood as follows:

[0110] spatiotemporal semantic feature

[0111] Spatiotemporal semantic feature refers to a high-level feature representation extracted from raw data that can express spatial position relationship, time evolution law and functional semantics (such as "aggregation", "diffusion", "go to exit") at the same time.

[0112] Predicted passenger flow demand matrix

[0113] The predicted passenger flow demand matrix is a mathematical tool quantitatively representing the scale and direction of personnel flow between functional areas in a building in a specific future period of time, and is a core basis for forward scheduling.

[0114] Example 2

[0115] As shown in the following, a large commercial complex is taken as an example to illustrate the implementation process of the technical solution of the present application. Figure 2

[0116] Specific application scenario: Consider a typical modern large commercial complex, which has 2 underground floors for parking and 8 above-ground floors for retail, catering and entertainment. During the lunch break period (11:30-13:30) on weekdays and all day on weekends, there will be a high peak of instantaneous, high-density and complex-direction people flow in the building. Traditional elevator group control systems (such as rule-based systems based on the nearest floor or up-down peak detection) perform poorly in such scenarios, often resulting in problems such as long waiting time for passengers on some floors, high elevator empty running rate, and serious energy waste. The technical solution of the present application aims to fundamentally solve these problems.

[0117] Step S11 is implemented in detail as follows:

[0118] High-definition network cameras deployed at key nodes in the building collect video stream data. These "key nodes" include but are not limited to elevator halls on each floor, main entrance gates, escalator intersections, and key passenger flow collection and distribution points such as the entrances of various brand stores. The camera network constitutes the "visual nerve" of the system, continuously capturing raw and unstructured video information. In the complex, a hierarchical zoning strategy is adopted for camera deployment to achieve a combination of no dead angle coverage and key monitoring. For example, one camera is deployed on each side of the elevator hall on each floor to ensure coverage of the entire waiting area and reduce blind spots; high-definition cameras with wider dynamic range are deployed at the main entrance and key passages leading to the parking lot to cope with complex lighting conditions such as backlight; panoramic cameras are deployed in the atrium area to macroscopically grasp the flow trend of people flow on each floor. All cameras transmit the collected high-definition video stream data to the next processing node in real time through the gigabit Ethernet or optical fiber network pre-laid in the building.

[0119] ​Step S121 implementation details: the system needs to process massive video streams in real time near the data source to reduce the pressure on network bandwidth and central servers. The edge computing gateway (equipped with a dedicated AI acceleration chip) deployed in each layer of the weak electricity well undertakes this responsibility. Lightweight deep learning models (such as pruned and quantized YOLOv5 or MobileNet-SSD) are preloaded into the edge gateway. Their function is not to perform complex scene understanding, but to efficiently and accurately perform two core tasks: identify each "person" in the video frame and assign a temporary ID to it. The model runs on the edge side, greatly reducing the huge delay and bandwidth consumption caused by transmitting raw video streams to the central server, so that the time from data collection to preliminary analysis results is controlled within milliseconds, providing a solid foundation for subsequent real-time scheduling.

[0120] Step S122 implementation details: after identifying the personnel, the system needs to continuously track their movements to understand their intentions. The detection model in step S121 provides the bounding box coordinates of all personnel in each frame for step S122, which uses a multi-object tracking (MOT) algorithm (such as DeepSORT) to associate these cross-frame detection boxes to form continuous and smooth motion trajectories. The algorithm combines appearance features (extracted by a lightweight ReID model) and motion prediction (such as Kalman filtering) to maintain the continuity of the trajectory, even if there is temporary occlusion, it can be re-associated. For elevator scheduling, a valuable motion trajectory not only contains the real-time pixel coordinates of the personnel, but more importantly, it can infer the moving direction (towards or away from the elevator hall), speed (running or walking), and dwell time at the elevator hall door. All this information is a strong signal for judging whether the passenger has the intention to take the elevator and the urgency of the intention.

[0121] It can be understood that the multi-object tracking algorithm (MOT) is an important field in computer vision, which aims to track the motion of multiple targets (such as pedestrians, vehicles) in a video sequence. The core challenge is to correctly associate the detected targets with existing tracking trajectories in each frame, and handle problems such as the appearance of new targets, the disappearance of old targets, and the cross-occlusion between targets. A high-performance MOT algorithm can output stable and identity-consistent trajectory sequences, providing a high-quality data foundation for subsequent behavior analysis.

[0122] In step S22: analysis results from each edge node are aggregated, but these data can be inconsistent in time stamp and format, and need to be regularized. Step S22 timestamps aligns and data cleanses the real-time people flow data and the motor operation state data, to obtain standardized time series data. The central server receives metadata such as time stamp, personnel coordinates, trajectory ID from all camera edge nodes, and motor state data from the elevator control cabinet. A high-precision network time protocol (NTP) server ensures that the time synchronization error of all devices is within milliseconds.

[0123] The data cleaning process includes: rejecting obvious outliers (such as "people" appearing on the wall due to lens distortion or transient false detection); interpolating and repairing trajectory breaks caused by temporary occlusion; and unifying all data to the same space-time coordinate system (for example, mapping pixel coordinates to real floor plan coordinates through perspective transformation). Finally, the state of each person at each time can be represented by a set of standardized data, such as (timestamp, tracker_id, floor, x, y, velocity_x, velocity_y). Motor state data is also processed into a standard format of (timestamp, elevator_id, current, voltage, temperature, vibration).

[0124] Step S23 implements detailed: in order to make more accurate predictions, the system cannot rely only on the current snapshot of people flow, but also needs to combine external factors that can affect people flow patterns. Step S23 fuses the standardized time series data with external context data, which is achieved by including calendar event information and environmental information. External context data constitutes the system's knowledge base for understanding "why people flow is like this". "Calendar event information" can be automatically obtained from the API interface of the building management system, including but not limited to: the start and end times of each session in the cinema, the scheduled peak period of popular restaurants, the schedule of promotional activities held by shopping malls, the marking of weekdays / weekends / holidays. "Environmental information" includes real-time weather data (rainy and snowy weather will significantly increase people's demand for elevators), current time (lunch time, evening closing time). These context data are also encoded into time series features. For example, 10 minutes before a movie ends, the system will generate a high-weight "event signal", indicating that a large number of people will soon flow from the cinema to the elevator hall.

[0125] In step S31: After acquiring high-quality, multi-source fused data, the system needs to extract deeper, predictive regularities. This is achieved by constructing a spatiotemporal graph representing the relationships between different areas based on the previously fused data with structured metadata in step S31. This step is a crucial leap in transforming data into knowledge, mathematically modeling building space and human behavior. The system models the entire commercial complex as a graph G = (V, E, A). Here, the node set V can represent different functional areas (e.g., the east hall on the first floor, the west hall on the third floor, the cinema on the fifth floor, etc.). The edge set E represents the connectivity between areas (connected by stairs, escalators, and elevators). The adjacency matrix A not only represents physical connections, but its weights A{ij} can also dynamically represent the historical or real-time probability of pedestrian flow from area i to area j. Each node v... i Each node is accompanied by a feature vector, which is composed of standardized time-series data and external context data that have been processed in steps S22 and S23 and converged into that region. This graph structure is dynamic, and its node features and edge weights are continuously updated over time.

[0126] It should be understood that a spatiotemporal graph structure is a special type of graph data structure that simultaneously captures both spatial relationships (topological connections between graph nodes) and temporal dynamics (features on each node changing over time). In elevator scheduling scenarios, spatial relationships reflect the physical and functional connections between different areas of a building, while temporal dynamics reflect the periodic, trend-based, and event-driven changes in the number of people flowing into each area. Spatiotemporal graphs are ideal inputs for subsequent accurate spatiotemporal predictions.

[0127] Step S33 Implementation Details: After constructing the graph structure, the system needs a powerful model to simultaneously learn complex spatial and temporal dependencies. Although the graph convolution operation in step S31 effectively extracts spatial features, traditional recurrent neural networks (RNNs) suffer from vanishing or exploding gradients when handling long-term temporal dependencies. Self-attention mechanisms perfectly compensate for this deficiency. When processing the temporal features of each node, the model uses a multi-head self-attention mechanism. This mechanism allows the model to focus its attention on different points in time simultaneously. For example, one "head" might focus on the peak pattern at "12 noon every Friday" (periodic dependency), another "head" might focus on "the start time of a large event 90 minutes before the current time" (event dependency), and a third "head" might focus on "the sharp increase in passenger flow in the last 5 minutes" (short-term dependency). Through this mechanism, the model can flexibly and explicitly model the dependencies between any two points in time, regardless of their distance, thus greatly improving the accuracy of predicting future passenger flow. Its robustness in dealing with sudden surges in passenger flow is unmatched by traditional methods.

[0128] In step S43, how to convert it into the optimal elevator control instruction is a complex multi-objective optimization problem. Step S43 solves the multi-objective optimization problem using the efficient set algorithm in a limited prediction horizon to obtain a set of optimal control sequences. This step is the "brain" of the scheduling system, responsible for making the final decision. The objective function of the optimization problem is to minimize the weighted sum of the total energy consumption, total waiting time of passengers, and total wear cost of equipment. This is a quadratic programming (QP) or nonlinear programming problem with constraints such as elevator capacity, maximum acceleration of the motor, safety distance, etc. The efficient set algorithm is an iterative algorithm very suitable for solving such problems: by actively identifying and processing the active constraint boundary, it efficiently finds the optimal solution. The model predictive control (MPC) framework enables it to "roll optimization": that is, at each control period, the system re-solves the optimization problem in a limited time domain (for example, the next 3 minutes) based on the latest prediction data and system state, but only implements the optimal control instruction at the current time (such as assigning which elevator to respond to which floor call). Repeat this process at the next time. This rolling feedback mechanism enables the system to continuously adapt to real-time changes, overcoming the inevitable errors in the prediction model, and showing strong robustness.

[0129] This multi-objective optimization problem can be formalized as solving the following problem at each decision time t:

[0130]

[0131] where H is the prediction horizon, x is the system state (including elevator position, speed, and floor waiting number), u is the control input (acceleration, target floor assignment), f is the model describing the dynamics of the elevator system, U and X are the constraint sets of control input and system state, respectively. J energy ,J wait ,J wear are energy consumption, waiting time, and wear cost functions, respectively. ω1, ω2, ω3 are adaptive weight coefficients.

[0132] Step S53 is implemented in detail:

[0133] The optimized decision needs to be precisely executed, and the system needs to perceive the execution effect to form a closed loop. This is completed through step S53, driving the elevator traction motor cluster to operate according to the specific control parameters, and collecting operation feedback data in real time. The main controller converts the obtained optimal instruction (such as "Elevator 3 runs to Floor 5 with acceleration a") into a control signal (such as voltage, frequency setting value) that the frequency converter can receive. The control signal is issued to the special frequency converter driver of each elevator through an industrial bus (such as CAN or Profinet). The frequency converter precisely controls the torque and speed of the traction motor, thereby achieving smooth and efficient start-stop and operation. At the same time, high-precision sensors deployed on the motor and the car continuously collect the latest operation state data (current, voltage, actual position, speed, vibration, etc.). These operation feedback data are transmitted back to the central server in real time. This part of data is crucial, which is used to update the motor operation state data in step S22, providing the real system state for the next round of optimization solving; on the other hand, it can also be used for online calibration and adaptive learning of the prediction model and control model, so that the whole system can become "smarter" and continuously improve the scheduling performance.

[0134] By implementing the patent solution in the above commercial complex scenario, it realizes the transition from "passive response" to "active scheduling", and achieves the following technical effects:

[0135] 1) Overall operation efficiency is significantly improved: the average passenger waiting time for the traditional system during the noon peak is as long as 120 seconds, and some passengers even wait for more than 3 minutes. After using the solution, through accurate prediction of passenger flow trend and advanced optimization scheduling, the average waiting time is greatly reduced. More importantly, the standard deviation of the waiting time is greatly reduced, eliminating the situation of extremely long waiting time, and optimizing the average travel time of the elevator group (the average time from call generation to arrival at the destination).

[0136] 2) System energy consumption is greatly reduced: under the traditional scheduling, the phenomena of elevator empty running, frequent start-stop, and invalid round trip are serious. The solution takes energy consumption as one of the core optimization targets through multi-objective optimization. The optimization algorithm will tend to combine adjacent floor calls, reduce unnecessary stops, and plan the most energy-saving speed curve. Actual operation data shows that under the condition of equivalent passenger flow, the total energy consumption of the elevator group system is reduced by 15%~20%. This is mainly due to the "pre-scheduling" based on prediction, which reduces the blind movement of the elevator, making its operation more "targeted".

[0137] In addition, the elevator experience is also optimized: the solution brings not only "faster", but also "smarter" and "more comfortable". The system can predict the large passenger flow after the movie, and dispatch standby elevators from other floors in advance (active scheduling). The multi-head self-attention mechanism ensures that the system can understand and respond to this short-term, high-intensity demand pulse. For passengers, the most intuitive feeling is that "the elevator seems to know I'm coming". At the same time, the optimization algorithm smooths the acceleration curve, reduces sudden stops and starts, and effective load balancing reduces the burden on the motor and vibration, improving the stability and quietness of the car operation.

[0138] Furthermore, equipment life is extended and maintenance costs are reduced: including total wear and tear costs in the optimization goal makes the scheduling strategy not only pursue efficiency, but also consider the load of mechanical components such as motors and brake pads. The algorithm will avoid running a certain elevator at high load for a long time, thereby achieving load balancing within the cluster. Feedback from vibration sensor data allows the system to sense changes in mechanical state, and even before potential failures occur, it can avoid risks by adjusting the scheduling strategy. According to actual measurement statistics, it is expected that the maintenance period of the main transmission components can be extended by more than 10%, significantly reducing the operating costs throughout the life cycle.

[0139] In summary, the patent technical solution builds a digital twin base through multi-modal perception, understands passenger flow patterns through data fusion and accurate prediction, and finally realizes decision value closed loop through multi-agent collaborative optimization control. The three are closely linked to form an integrated active scheduling agent. Its successful application in the commercial complex scene proves that the solution produces synergistic gains in efficiency, energy consumption, and experience optimization, which are superior to traditional passive scheduling methods.

[0140]

Example 3

[0141] Here, the technical solution consisting of steps S31-S34 described above can be applied to an application scenario (elevator intelligent scheduling during lunchtime passenger flow peak in a large commercial complex) to detail the specific implementation of the four steps.

[0142] This embodiment takes a 50-story large commercial complex as the application scenario, and details how the technical solution of the invention optimizes elevator scheduling during the lunchtime (11:30-13:30) passenger flow peak through accurate prediction and collaborative control. The complex has 2 underground floors for parking, 1-3 floors above ground for retail stores, 4-5 floors for dining areas, 6-8 floors for cinemas and entertainment facilities, and 9 floors and above for office areas. During the lunchtime peak, a large number of people flow from the office area and the parking lot, with a clear target pointing to the dining floors, followed by the subsequent diversion demand of returning to the office area or going to the cinema after dining, and the flow presents a high degree of dynamics and concentration in space and time, posing a great challenge to the elevator scheduling system.

[0143] In this embodiment, in step S31, the spatio-temporal features of each floor, including the average crowd density, the variance of moving speed, and the event impact factor, are extracted from the fusion data with the structured metadata format as the node feature vector of the spatio-temporal graph structure, and the feature dimension is set to 64 dimensions, and the numerical values are normalized to the range of zero to one using the Min-Max normalization.

[0144] Then, based on the building layout of the super-high office building, a node set is defined, each node corresponding to a floor number from 1 to 50, and an edge set is defined to represent the elevator accessible path, including the standard connection between adjacent floors and the cross-layer express elevator path, and the edge direction is set according to the main flow direction of the crowd.

[0145] Further, the weight value of each edge is calculated, based on the interaction frequency in the historical crowd data, the Jaccard similarity coefficient algorithm is used to measure the crowd correlation strength between floors, and the calculated weight value is normalized.

[0146] Then, an adjacency matrix is constructed to represent the spatio-temporal graph structure, the matrix size is 50 by 50, and a self-loop edge is added to ensure the connection of the node itself, and is stored in the CSR sparse matrix format to optimize memory usage, while recording the node coordinate mapping relationship.

[0147] In this embodiment, the connectivity of the spatio-temporal graph structure is verified, the depth-first search algorithm is used to detect whether there is an isolated node, and based on the principle of geographical proximity, the missing edges are automatically supplemented to ensure the completeness of the graph structure for model input.

[0148] In step S32, the spatio-temporal graph structure is input into the spatio-temporal graph convolution network model, first the spatial dimension is processed by the graph convolution layer, the Chebyshev polynomial is used for graph filter approximation, the polynomial order is set to 3, to extract the local neighborhood features of the node, and output the spatial enhanced feature tensor.

[0149] In step S32, the spatial enhanced feature tensor is applied to the time convolution layer, a one-dimensional convolution kernel is used to slide along the time axis, the convolution kernel size is set to 5, and the step size is set to 1, to capture the short-term time dependence, and the residual connection is fused to prevent gradient disappearance.

[0150] Further, the gated recurrent unit network is introduced to process the time series, the number of network units is set to 128, to learn the long-term time pattern, and a multi-scale spatio-temporal feature sequence is output, the sequence length is aligned with the prediction period.

[0151] Then, the multi-scale spatio-temporal feature sequence is subjected to a pooling operation, the maximum pooling method is used to reduce the data dimension while retaining significant features, generating a compressed feature representation for subsequent modules.

[0152] At the same time, the model training process is monitored, and when the loss function converges slowly, the learning rate is dynamically adjusted and an exponential decay strategy is used to ensure the stability and efficiency of the training process.

[0153] Further, in step S33, the region correlation features are processed using a multi-head self-attention mechanism, the input feature sequence is divided into multiple subsequences, the number of attention heads is set to 8, and each subsequence independently generates query, key and value vectors through linear transformation.

[0154] In step S33, the scaled dot-product attention weight is calculated for each attention head, the similarity between the query vector and the key vector is compared and the softmax function is applied, and then the value vector is weighted and summed to capture the long-term dependencies in the time dimension, generating an attention-weighted feature representation.

[0155] Then, the outputs of different attention heads are fused, a full connection layer is applied after the splicing operation to integrate information, the output dimension is kept consistent with the input, and layer normalization operation is added to smooth the data distribution.

[0156] Then, a feedforward neural network is introduced for nonlinear transformation, ReLU is used as the activation function to enhance the expression ability of the model, and finally the enhanced spatio-temporal feature sequence is output. The calculated attention weight is visualized and analyzed to identify the time steps that have a key impact on the output. When the weight distribution is abnormal, the model calibration mechanism is triggered to improve the robustness of the system.

[0157] Step S34 is based on the enhanced spatio-temporal feature sequence, which is mapped to the output space through a two-layer fully connected layer, the number of hidden layer units is set to 256, and the Dropout method is used for regularization to prevent overfitting, with a dropout rate of zero point two.

[0158] In step S34, the passenger flow demand matrix of each region in the future period is predicted, the matrix dimension is defined as 50 floors multiplied by 12 time intervals, each interval is 5 minutes, the output value represents the normalized demand intensity, and the Sigmoid function is used to constrain the value between zero and one.

[0159] Then, the prediction results are post-processed, a sliding average filter with a window size of 3 is applied to smooth the time series, reduce random fluctuations, and verify the internal consistency of the matrix to ensure that the total demand of each floor is logical.

[0160] At the same time, an online learning mechanism is introduced, the prediction error is calculated based on the real-time collected passenger flow data, and when the average absolute error exceeds the preset threshold of zero point one, the incremental update process is triggered, and the model parameters are fine-tuned using the stochastic gradient descent method.

[0161] In this embodiment, the final predicted passenger flow demand matrix is stored in a time series database and labeled with a confidence score for evaluation by downstream modules, and a prediction report is automatically generated for monitoring by operation and maintenance personnel.

[0162] The present embodiment realizes accurate prediction of distributed human flow in the building through the cooperative execution of the four sub-steps S31-S34. Compared with traditional prediction methods based on historical averages or single sensors, the present application integrates multi-modal data and spatio-temporal graph structures, significantly improving prediction accuracy and robustness, especially in handling sudden changes in passenger flow and cross-regional human flow migration scenarios. In addition, the combination of the spatio-temporal graph convolution network and the multi-head attention mechanism effectively captures the complex interaction of spatial correlation and temporal dependence, providing a more reliable data foundation for the generation of subsequent elevator dispatching instructions.

[0163] In this embodiment, the spatio-temporal graph is a mathematical model used to represent the spatial relationships between entities and their dynamic changes over time. This structure converts abstract building space into a mathematical object that can be processed and learned by computers.

[0164]

Embodiment 4

[0165] In the following, the technical solution composed of steps S41-S44 described above is applied to an application scenario (multi-target collaborative scheduling of elevator clusters in a large office building during the evening peak period) to explain the specific implementation of the four steps in detail.

[0166] The embodiment takes a smart office building with 30 floors of office area and equipped with 8 elevators as the application scenario. From 17:00 to 18:30 on weekdays, it is the peak of off-work passenger flow, a large number of people almost simultaneously flow from each floor to the elevator hall, and the target is the 1st floor lobby and B1, B2 parking lot. In this scenario, the traditional group control system is prone to problems such as excessive concentration of some elevators, frequent floor stops, long waiting time of passengers, and sharp increase of elevator energy consumption. The invention realizes intelligent collaborative scheduling through the following steps. The camera and edge computing node deployed in each floor elevator hall generate structured metadata including the number of passengers waiting on each floor and the moving direction in real time by running target detection and tracking algorithms. At the same time, the sensors installed on the elevator traction system continuously collect three-phase current, winding temperature, instantaneous power and vertical vibration acceleration data of each motor. Then, the real-time passenger flow data and the motor operating state data are processed by multi-modal fusion to generate fusion data with structured metadata format. This process timestamps the multi-source data, cleanses it, and fuses it with the external context information "today is a weekday" to finally generate a standardized and time-aligned fusion data frame. Further, the artificial intelligence model uses the spatio-temporal graph convolution network (STGCN) combined with the multi-head self-attention mechanism to predict the waiting passenger number change matrix of each floor hall in the next 5 minutes at 30-second intervals. The prediction result shows that the 20th and 25th floors will have a peak of waiting passengers in the next 90 seconds.

[0167] Based on the above prediction information, the multi-agent collaborative control module of the system is activated to start executing the core scheduling instruction generation step.

[0168] In this embodiment, step S41 inputs the predicted passenger flow demand matrix and the real-time state data of the elevator traction motor cluster into the fuzzy logic reasoner, defines the input variables including the demand intensity of each floor, the current load rate of the elevator, the motor operating temperature and the vibration amplitude, and sets the fuzzy interval for them.

[0169] Then, a fuzzy set is defined for each input variable, including "low", "medium" and "high" three levels, and a triangular membership function is used for fuzzy processing. The fuzzy rule base is initialized according to expert experience, and the number of rules is 27. Further, the Mamdani reasoning method is applied in the fuzzy logic reasoner to reason the fuzzy input according to the fuzzy rules to obtain a fuzzy output set representing the scheduling priority of each elevator. Then, the fuzzy output obtained by reasoning is de-fuzzied, and the gravity method is used to calculate the accurate scheduling preference value to generate a preliminary scheduling preference vector with a dimension equal to the number of elevators.

[0170] In this embodiment, the preliminary scheduling preference vector is normalized to ensure that the sum of the elements is 1, and a timestamp identifier is added for use by the subsequent optimization module.

[0171] In step S42, a multi-objective model predictive control optimization problem is constructed with the preliminary dispatch preference vector as a soft constraint, and the optimization objectives include the total energy consumption of the system, the total waiting time of passengers, and the total wear cost of equipment, and a normalization coefficient is set for each objective function.

[0172] In step S42, the preliminary dispatch preference vector is converted into a constraint condition, which limits the deviation of the dispatch decision variable of each elevator within a certain threshold, and the threshold is dynamically adjusted according to the real-time load of the system. The decision variable is defined as the speed instruction sequence of each elevator in the future prediction time domain, and the hard constraint conditions include the maximum acceleration of the motor, the maximum speed, and the safe operation range. Then, the multi-objective is converted into a single-objective optimization problem by using the weighted summation method, and the weight coefficient is dynamically configured according to the operation period, with the waiting time being emphasized during peak hours and the energy consumption being emphasized during off-peak hours. At the same time, the optimization problem is expressed in the form of a standard quadratic programming, and the convexity of the problem is verified to ensure the existence of a global optimal solution.

[0173] Further, in step S43, the multi-objective model predictive control optimization problem is solved by using the efficient set algorithm in a limited prediction time domain, the prediction time domain is set to 5 minutes, the control time domain is set to 1 minute, and the discrete time interval is 10 seconds.

[0174] In step S43, the efficient set is initialized, a feasible point that satisfies all the constraint conditions is selected as the initial iteration point, and the gradient vector and the Lagrange multiplier of the point are calculated.

[0175] Then, in each rolling solution, the active constraint of the current efficient set is judged, the simplified equality constraint problem is solved, the decision variable is updated, and the objective function value is calculated.

[0176] Then, the optimality condition is checked, if the KKT condition is satisfied, the current solution is output as a local optimum, otherwise the efficient set is updated and the next iteration is entered, until convergence or the maximum number of iterations is reached.

[0177] Preferably, the solving process is monitored, and when the number of iterations exceeds the preset upper limit or the solving time is exceeded, a backup heuristic rule is enabled to generate a feasible dispatch instruction, ensuring the real-time performance of the system.

[0178] Further, in step S44, the optimal instruction at the current time is extracted from the optimal control sequence, which includes the target speed, the target floor, and the acceleration curve of each elevator.

[0179] In step S44, the extracted instructions are checked for feasibility, checking whether they violate the physical limits of the motor or safety regulations, and are smoothed to reduce mechanical impact. Then, the instructions that pass the check are encoded into a protocol format recognizable by the elevator group control system, including CANopen object dictionary entries and transmission PDO mapping relationships. Then, the coordinated scheduling instructions for the elevator traction motor cluster are generated, including execution timestamps, instruction priorities, and checksums to ensure the integrity and timeliness of the instructions.

[0180] At the same time, a log of the instructions is recorded, including the solving time, the target function value, and the constraint violation situation, for subsequent performance analysis and model improvement.

[0181] [Example 5]

[0182] Here, the technical solution of steps S31'-S34' described above can be applied to an application scenario (precise prediction of midday and event-driven passenger flow in a large commercial complex) to describe the specific implementation of the four steps in detail.

[0183] This embodiment takes a 50-story commercial complex with retail, catering, and entertainment functions as an application scenario. The building has a high degree of dynamicity and complexity in the spatial and temporal distribution of passenger flow during the lunch hour (11:30-13:30) and at specific events (such as movie theater exits). Traditional time series prediction methods are difficult to effectively capture the spatial correlation between different functional areas and the long-term temporal dependence of event influences, resulting in limited passenger flow prediction accuracy and inability to provide high-quality input for elevator active scheduling. The invention achieves precise prediction of passenger flow demand in such complex scenarios through the steps defined in steps S31'-S34' described above. Real-time passenger flow data and motor operating state data are collected by a multi-modal sensor network deployed in the building, and the real-time passenger flow data and motor operating state data are processed through multi-modal fusion to generate fusion data in a structured metadata format. The processed fusion data includes timestamp-aligned regional passenger statistics, movement vectors, and calendar event labels (such as "theater on the fifth floor at 12:30 p.m. exits") from the building management system.

[0184] Based on this data, the prediction model performs the following core steps:

[0185] In this embodiment, in step S31', feature data for each functional area of the building is extracted from the fusion data with structured metadata format, including the average number of passenger flow in each area in the last 15 minutes, the maximum and minimum values of passenger flow movement speed, which are combined into an initial feature vector for each area.

[0186] Then, 50 floors of the building are divided into different area nodes according to functions, including office floor nodes, elevator hall nodes and corridor area nodes, each node containing a floor number and a function area type identifier.

[0187] Further, the connection relationship between the nodes is determined according to the floor plan, and a connection edge is established between adjacent floor nodes and between different function area nodes of the same floor.

[0188] Then, a weight value is assigned to each connection edge, and the weight is calculated according to the personnel flow frequency between the two areas within the past one hour. The higher the flow frequency, the greater the weight value.

[0189] In this embodiment, an adjacency matrix is used to store the information of all nodes and edges. The matrix rows and columns correspond to node numbers, and the matrix element values represent the connection weights between the corresponding nodes.

[0190] Alternatively, in step S32', the importance weight of each node with its adjacent nodes is calculated, and a historical human flow correlation-based calculation method is adopted to measure the similarity of the human flow trend of two nodes using the Pearson correlation coefficient.

[0191] Through step S32', the calculated correlation coefficient is normalized to convert it into a weight value between 0 and 1. The greater the correlation coefficient, the higher the weight value.

[0192] Further, the actual distance factor between nodes is considered, and higher weight adjustment coefficients are given to node pairs with closer distances, and lower adjustment coefficients are given to node pairs with farther distances.

[0193] Then, the weight is fine-tuned in combination with the current time period characteristics, and the weight of the elevator hall node is increased during the rush hour, and the weight of the dining area node is increased during the lunch break.

[0194] At the same time, the maximum weight is standardized to ensure that the sum of all adjacent node weights of each node is equal to 1.

[0195] Further, in step S33', according to the calculated dynamic attention weight, the feature vectors of all adjacent nodes of each node are weighted and averaged to obtain a comprehensive representation of the neighbor features.

[0196] In step S33', the feature vector of each node itself is linearly combined with the comprehensive representation of the neighbor features, and a fixed proportion coefficient is used to balance the importance of the self-feature and the neighbor feature.

[0197] Then, the combined feature vector is standardized to make the numerical values of each feature dimension in the same order of magnitude range.

[0198] Then, the features are reduced in dimension through a simple linear transformation layer, retaining the main feature information and reducing data redundancy.

[0199] Preferably, an outlier detection is performed on the feature aggregation process, and when the feature value of a certain node deviates significantly from the normal range, the feature value of the adjacent node is used for replacement.

[0200] Further, in step S34', the enhanced feature representation of each node is arranged in time sequence to construct time series data, and the time granularity is 5 minutes per data point.

[0201] In step S34', a moving average algorithm is used to smooth the time series data, and a simple moving average with a window size of 6 is used to eliminate random fluctuations.

[0202] Then, based on the smoothed time series data, a linear regression model is used to predict the passenger flow trend in each area in the next 60 minutes, and a prediction value is output every 5 minutes.

[0203] Then, the prediction results of all areas are combined into a passenger flow demand matrix, and the rows correspond to different areas and the columns correspond to different time points.

[0204] At the same time, the prediction results are corrected according to the historical prediction error, and an error feedback mechanism is used to gradually improve the prediction accuracy.

[0205] The spatio-temporal graph neural network in this embodiment can be understood as follows:

[0206] The spatio-temporal graph neural network is a deep learning architecture specially designed for processing graph-structured spatio-temporal data. It has the ability to capture both the spatial dependency between nodes in the graph and the temporal dynamic changes. Its "spatio-temporal" characteristics are reflected in: on the spatial level, the graph convolution operation is used to aggregate the neighbor node information; on the time level, the sequence model (such as recurrent neural network, temporal convolution network or self-attention mechanism) is used to learn the time evolution pattern of node features.

[0207] Figure 3 is a block diagram of a coordination control device 300 for a plurality of motor devices in property elevator management according to an example embodiment. For example, the coordination control device 300 can be a mobile phone, a computer, a digital broadcast terminal, a messaging equipment, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, and the like.

[0208] Referring to Figure 3 , the coordination control device 300 can include one or more of the following components: a processing component 302, a memory 304, a power supply component 306, a multimedia component 308, an audio component 310, an input / output (I / O) interface 312, a sensor component 314, and a communication component 316.

[0209] Figure 3 The coordination control device 300, the processing component 302, and the memory 304 are each configured to perform a number of operations associated with the coordination control device 300 as described herein. In this regard, the processing component 302 can be used to implement many of the steps associated with the data fusion, passenger flow prediction, and collaborative decision making described herein. For example, the processing component 302 can be used to implement the data fusion, passenger flow prediction, and collaborative decision making described herein. The memory 304 can be used to store the fused data, the prediction model, the optimization algorithm, and the computer program. The communication component 316 can be used to communicate with the multi-modal sensor network and the elevator group control system.

[0210] The processing component 302 typically controls the overall operations of the coordination control device 300, such as operations associated with display, telephony calls, data communications, camera operations, and recording operations. The processing component 302 can include one or more processors 320 to execute instructions to complete all or part of the steps associated with the methods described above. In addition, the processing component 302 can include one or more modules to facilitate the interaction between the processing component 302 and other components. For example, the processing component 302 can include a multimedia module to facilitate the interaction between the multimedia component 308 and the processing component 302.

[0211] The memory 304 is configured to store various types of data to support operations of the coordination control device 300. Examples of such data include instructions for any application or method operating on the coordination control device 300, contact data, phonebook data, messages, pictures, videos, and the like. The memory 304 can be implemented by any type of volatile or non-volatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable programmable read only memory (EPROM), programmable read only memory (PROM), read only memory (ROM), magnetic storage, flash memory, magnetic or optical disks.

[0212] The power component 306 provides power to the various components of the coordination control device 300. The power component 306 can include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the coordination control device 300.

[0213] The multimedia component 308 includes a screen providing an output interface between the coordinating control device 300 and a user. In some embodiments, the screen includes a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touch, swiping, and gestures on the touch panel. The touch sensors can not only sense a boundary of a touching or swiping action, but also detect duration and pressure related to the touching or swiping action. In some embodiments, the multimedia component 308 includes a front camera and / or a back camera. The front and / or back camera can receive external multimedia data when the coordinating control device 300 is in an operation mode, such as a shooting mode or a video mode. Each of the front and back camera can be a fixed optical lens system or have a focal length and optical zoom capability.

[0214] The audio component 310 is configured to output and / or input audio signals. For example, the audio component 310 includes a microphone (MIC) that is configured to receive an external audio signal when the coordinating control device 300 is in an operation mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signal can be further stored in the memory 304 or transmitted via the communication component 316. In some embodiments, the audio component 310 further includes a speaker for outputting audio signals.

[0215] The I / O interface 312 provides an interface between the processing component 302 and peripheral interface modules, which can be a keypad, a click wheel, buttons, and the like. The buttons can include, but are not limited to, a home button, a volume button, a start button, and a lock button.

[0216] The sensor component 314 includes one or more sensors to provide various state assessments for the coordinating control device 300. For example, the sensor component 314 can detect an open / closed state of the coordinating control device 300, relative positioning of components, such as a display and a keypad of the coordinating control device 300, a change in position of the coordinating control device 300 or a component of the coordinating control device 300, presence or absence of user contact with the coordinating control device 300, an orientation or acceleration / deceleration of the coordinating control device 300, and a temperature change of the coordinating control device 300. The sensor component 314 can include a proximity sensor configured to detect presence of a nearby object without any physical touch. The sensor component 314 can further include a light sensor, such as a CMOS or CCD image sensor, for use in an imaging application. In some embodiments, the sensor component 314 can further include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0217] The communication component 316 is configured to facilitate wired or wireless communication between the coordination control device 300 and other devices. The coordination control device 300 can access a wireless network based on a communication standard, such as WiFi, 2G, or 3G, or a combination thereof. In an example embodiment, the communication component 316 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel. In an example embodiment, the communication component 316 further includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0218] In an example embodiment, the coordination control device 300 can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, micro-controllers, microprocessors, or other electronic elements for performing the above-described methods.

[0219] In an example embodiment, a non-transitory computer readable storage medium including instructions, such as the memory 304 including instructions, is also provided, which can be executed by the processor 320 of the coordination control device 300 to complete the above-described methods. For example, the non-transitory computer readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0220] In some embodiments of the present application, a non-transitory computer readable storage medium is also provided, which, when instructions in the storage medium are executed by a processor of a mobile terminal, enables the mobile terminal to perform a method for coordinating control of multiple motor devices in property elevator management.

Claims

1. A method of coordinated control of a plurality of motor devices in a property elevator management, characterized by, Comprising: S1. Collecting real-time passenger flow data and motor operating state data through a multi-modal sensor network deployed in the building; S2. Performing multi-modal fusion processing on the real-time passenger flow data and the motor operating state data to generate fusion data in a structured metadata format; S31. Based on the fusion data in the structured metadata format, constructing a spatio-temporal graph structure representing the correlation between regions; S32. Processing the spatio-temporal graph structure through a spatio-temporal graph convolution network model to extract regional correlation features in the spatial dimension; S33. Processing the regional correlation features using a multi-head self-attention mechanism to capture long-term dependencies in the temporal dimension; S34. Based on the enhanced spatio-temporal feature sequence, predicting a predicted passenger flow demand matrix for each region in the future time period; S4. Inputting the predicted passenger flow demand matrix into a multi-agent collaborative control model to generate a collaborative scheduling instruction for the elevator traction motor cluster; S5. Controlling the elevator traction motor cluster according to the collaborative scheduling instruction to optimize overall operating efficiency; Wherein the motor operating state data is real-time state data of the elevator traction motor cluster, and the S4 comprises: Inputting the predicted passenger flow demand matrix and the real-time state data of the elevator traction motor cluster into a fuzzy logic reasoner to generate a preliminary scheduling preference vector for each elevator; Using the preliminary scheduling preference vector as a soft constraint, constructing a multi-objective model predictive control optimization problem that integrates system total energy consumption, passenger total waiting time, and equipment total wear cost; Rolling solving the multi-objective model predictive control optimization problem within a limited prediction time domain using an efficient set algorithm to obtain a set of optimal control sequences; Extracting the optimal instruction at the current time from the optimal control sequence to generate the collaborative scheduling instruction for the elevator traction motor cluster.

2. The method of claim 1, wherein the method further comprises: The S1 comprises: S11. Collecting video stream data through high-definition network cameras deployed at key nodes in the building; S12. Performing real-time analysis on the video stream data to generate structured metadata including the number of people, moving direction, and location information; S13. Collecting motor operating state data through motor sensors deployed in the elevator car and the shaft, the motor operating state data including current, voltage, temperature, and vibration data; S14. Aligning and packaging the structured metadata and the motor operating state data to form the real-time passenger flow data and the motor operating state data.

3. The method of claim 1, wherein the method further comprises: The S2 comprises: S21. Receiving the real-time passenger flow data and the motor operating state data from the multi-modal sensor network; S22. Time stamp alignment and data cleaning of the real-time passenger flow data and the motor operating state data to obtain standardized time series data; S23. Fusing the standardized time series data with external context data, including calendar event information and environmental information; S24. Based on the fused data, constructing spatio-temporal semantic features to generate the fusion data in the structured metadata format.

4. The method of claim 1, wherein the method further comprises: The S5 comprises: S51. Distributing the collaborative scheduling instruction to the main controller of the elevator group control system; S52. Resolving and converting the instructions into specific control parameters of the traction motor through the main controller; S53. Driving the elevator traction motor cluster to operate according to the specific control parameters and collecting operation feedback data in real time.

5. The method of claim 2, wherein the method further comprises: The S12 comprises: S121. Analyzing the video stream data in real time by running a lightweight deep learning model on the edge computing node; S122. Extracting the personnel position and motion trajectory in the video frame through the target detection and multi-target tracking algorithm; S123. Generating the structured metadata containing the timestamp, camera ID, number of people, moving direction and position information based on the motion trajectory.

6. The method of claim 3, wherein the method further comprises: The S23 comprises: S231. Receiving the standardized time series data and the external context data, wherein the external context data includes calendar event information and environmental information; S232. Aligning the standardized time series data and the external context data in the time dimension and encoding the features; S233. Fusing the encoded standardized time series data and external context data through an attention weighting mechanism; S234. Outputting the fused multi-source data feature representation.

7. The method of claim 6, wherein the method further comprises: The S24 comprises: S241. Extracting a feature vector representing the spatio-temporal correlation between regions based on the fused multi-source data feature representation; S242. Associating and mapping the feature vector with a predefined functional semantic label; S243. Generating the fused data in the structured metadata format containing spatio-temporal attributes and functional semantics.

8. The method of claim 1, wherein the method further comprises: The multi-agent collaborative control model generates the collaborative scheduling instructions by maximizing the following collaborative utility function: wherein, A t =( a t 1 , a t 2 ,…, a t N ) represents the joint action vector of all agents at time t t O t represents the joint observation vector, which is composed of the predicted passenger flow demand matrix and the motor operation state data f k ( ) is the first k The utility function of the optimization objective includes total system energy consumption, total passenger waiting time, and total equipment wear and tear cost; w k adaptive weight coefficients corresponding to each target; P(A t ) denotes the action distribution output by the multi-agent policy; Q(A t ) denotes the action prior distribution defined by the preliminary schedule preference vector generated by the fuzzy logic reasoner; D KL ( ││ ) denotes the KL divergence, which measures the deviation between the action distribution and the prior distribution; λ is a regularization coefficient, used to balance the degree of deviation of the optimization objective and the strategy.

9. A coordinated control device for a plurality of motor devices in a property elevator management, characterized by, comprises: a data acquisition module configured to acquire real-time passenger flow data and motor operating state data through a multi-modal sensor network deployed in a building; a data fusion module configured to perform multi-modal fusion processing on the real-time passenger flow data and the motor operating state data to generate fused data in a structured metadata format; a passenger flow prediction module configured to: based on the fused data in the structured metadata format, construct a spatio-temporal graph structure representing the correlation between regions; process the spatio-temporal graph structure through a spatio-temporal graph convolution network model to extract regional correlation features in the spatial dimension; process the regional correlation features through a multi-head self-attention mechanism to capture long-term dependencies in the time dimension; based on the enhanced spatio-temporal feature sequence, predict a predicted passenger flow demand matrix for each region in a future time period; and the motor operating state data is real-time state data of an elevator traction motor cluster; The cooperative decision module is configured to: input the predicted passenger flow demand matrix and real-time state data of the elevator traction motor cluster into a fuzzy logic inference engine together to generate a preliminary scheduling preference vector for each elevator; construct a multi-objective model predictive control optimization problem that integrates system total energy consumption, passenger total waiting time and equipment total wear cost, taking the preliminary scheduling preference vector as a soft constraint; solve the multi-objective model predictive control optimization problem by using an effective set algorithm to obtain a group of optimal control sequences within a limited prediction time domain; extract an optimal instruction at the current time from the optimal control sequences to generate a cooperative scheduling instruction for the elevator traction motor cluster; The control execution module is configured to control the elevator traction motor cluster according to the cooperative scheduling instruction to optimize overall operation efficiency.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps of the method of any one of claims 1 to 8.

Images