Intelligent elevator detecting and dispatching method for special crowds

Abstract

The invention discloses an intelligent elevator detecting and dispatching method for special crowds. According to the method, edge computing nodes are deployed to locally process video data, and an improved YOLOv7-tiny model integrated with an attention mechanism is used for recognizing the age, the luggage state and the use condition of auxiliary tools of passengers in real time; according to an identification result, dynamically distributing a priority weight, and constructing a multi-objective optimization function fusing time cost, energy consumption cost and a special service weight; the function is solved online by adopting a parameter dynamic self-adaptive whale optimization algorithm, and an optimal elevator dispatching instruction sequence is generated and issued to an elevator group control system for execution; and the system continuously monitors operation data, and performs feedback and closed-loop adaptive adjustment and optimization on the attribute identification model and optimization algorithm parameters based on historical data. According to the method, accurate perception and personalized scheduling under high real-time performance and high privacy protection are achieved, and the service fairness and the overall operation efficiency of the elevator system for special crowds are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of elevator intelligent control, and particularly relates to an elevator intelligent detection and scheduling method for special groups, which is suitable for providing personalized elevator scheduling services for the elderly, the disabled and passengers carrying large luggage. BACKGROUND

[0002] In modern high-rise buildings, the scheduling efficiency of the elevator group control system directly affects the travel experience of passengers and building energy consumption. Traditional elevator scheduling systems mostly use strategies based on fixed rules, such as "first-come, first-served" or "minimum average waiting time", which have a single optimization goal and are difficult to adapt to the individual needs of diverse passengers. Especially for special groups such as the elderly, wheelchair users or passengers carrying large luggage, their elevator behavior has obvious differences and priority needs, and the traditional system lacks the ability to perceive the identity of passengers, resulting in a lack of humanization in scheduling strategies.

[0003] To improve the system's perception ability, some existing technologies try to introduce computer vision technology for passenger detection. However, these solutions mostly rely on cloud processing of video data, which has problems such as large transmission delay, high bandwidth occupation, and privacy leakage risk, and cannot meet the real-time and reliability requirements of elevator scheduling. In terms of optimization algorithms, although some studies use meta-heuristic methods such as genetic algorithms and particle swarm algorithms for multi-objective scheduling, the optimization functions are still limited to traditional indicators such as time and energy consumption, and cannot integrate real-time perception of passenger attribute information, making it impossible to achieve truly personalized dynamic priority scheduling. In addition, existing systems generally lack self-learning and adaptive capabilities, and cannot optimize their performance based on long-term operation data.

[0004] Therefore, the existing technology has the following deficiencies: lack of fine-grained and real-time perception of passenger attributes; single scheduling goal, without considering the priority service needs of special groups; poor system adaptability, unable to dynamically adjust strategies based on operation data; and real-time deficiency and privacy risks caused by cloud processing mode. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies of the prior art and provide an elevator intelligent detection and scheduling method for special groups, which integrates edge computing, improved lightweight deep learning models and dynamic multi-objective optimization to build an intelligent scheduling system that is real-time, accurate, personalized and has closed-loop self-optimization capabilities.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] An elevator intelligent detection and scheduling method for special groups, comprising the following steps:

[0008] S1, through the edge computing nodes deployed in the elevator hall and the car, real-time collection and processing of multi-channel video sensor data, when detecting the appearance of a waiting passenger in the picture, starting the passenger identity attribute recognition process;

[0009] S2, based on the improved lightweight deep learning target detection model, real-time detection and fine-grained attribute recognition of passengers, the recognition attributes at least include age stage, whether carrying large luggage, and whether using auxiliary action tools; wherein the improvement is realized by introducing attention mechanism to improve the accuracy of fine-grained attribute recognition;

[0010] S3, according to the recognized passenger identity attribute, dynamically assigning individualized scheduling priority weight, which is based on the real-time change of passenger attribute and adaptive adjustment; and constructing a multi-objective optimization function including system total time cost, system total energy consumption cost and individualized service weight;

[0011] S4, based on the real-time updated call request and passenger attribute information, using meta-heuristic optimization algorithm to solve the multi-objective optimization function online, generating the current optimal elevator dispatching instruction sequence;

[0012] S5, the dispatching instruction sequence is sent to the corresponding elevator group control system to control the elevator operation;

[0013] S6, the system continuously monitors the elevator running state and passenger flow data, and regularly based on the accumulated historical scheduling data, feedback and adaptive tuning of the parameters of the attribute recognition model and optimization algorithm; the adaptive tuning includes online learning and optimization of model and algorithm parameters based on historical data to form a closed-loop control system.

[0014] Preferably, the improved lightweight deep learning target detection model used in step S2 is a model based on YOLOv7-tiny architecture and integrated with efficient channel attention (ECA) module, wherein the ECA module is embedded in the feature extraction layer of YOLOv7-tiny backbone network, used to enhance the weight of key feature channels, thereby improving the recognition accuracy of passenger age stage, luggage carrying condition and auxiliary action tool use condition.

[0015] Preferably, the implementation of the model comprises the following steps: S2.1, constructing a passenger image dataset containing images labeled with different age stages, luggage carrying conditions and auxiliary action tool usage conditions; the auxiliary action tools include wheelchairs, strollers, crutches, walking aids; S2.2, integrating an efficient channel attention (ECA) module in the feature extraction layer of the YOLOv7-tiny backbone network to improve the feature representation ability through cross-channel interaction; S2.3, using the dataset to perform multi-attribute recognition training on the YOLOv7-tiny network integrated with the ECA module, and using a multi-task learning loss function to simultaneously optimize the target detection and attribute classification tasks; S2.4, deploying the trained model on an edge computing node and adapting to the limitations of edge computing resources through model compression techniques.

[0016] Preferably, the multi-objective optimization function constructed in step S3 is: wherein, represents the total waiting and boarding time of the system, represents the total energy consumption of the system, represents the priority service score for special needs passengers; α, β, γ are weight coefficients and are dynamically adjusted according to the number and type of special passengers identified, with the adjustment rule being: when a passenger using an auxiliary action tool is identified, the weight of γ is increased, and when the system energy consumption exceeds a threshold, the weight of β is increased.

[0017] Preferably, the priority service score P special is calculated according to the formula: wherein w i is the priority weight of passenger i, with different weight values preset according to its attribute category, d i is a factor representing the degree of service delay, defined as the ratio of passenger waiting time to preset maximum tolerance time.

[0018] Preferably, the meta-heuristic optimization algorithm used in step S4 is the whale optimization algorithm, and the fitness function of the whale optimization algorithm is set as the multi-objective optimization function F, and dynamic adjustment of search parameters is introduced in the solving process, including adaptive adjustment of the convergence factor and spiral update weight of the whale optimization algorithm according to real-time passenger flow.

[0019] Preferably, the feedback and adaptive tuning in step S6 include: incrementally learning the attribute recognition model based on historical scheduling data to update model parameters to adapt to new scenarios; and dynamically adjusting the parameters of the whale optimization algorithm and the weight coefficients α, β, γ in the multi-objective optimization function based on scheduling effect evaluation, wherein the scheduling effect evaluation is performed by comparing the deviation between actual scheduling time and predicted time.

[0020] Preferably, in the step S5, the edge computing node issues the dispatching instruction to the elevator group control system core controller through an industrial bus, the industrial bus adopts a CAN bus or a Modbus protocol, and the real-time performance and reliability of the instruction transmission are ensured.

[0021] Compared with the prior art, the present application has the beneficial effects that: through the edge computing node and the lightweight model integrated with the ECA attention mechanism, high-precision and low-delay identification of passenger fine-grained attributes is realized locally, and privacy is effectively protected; innovatively, the identified attributes are converted into dynamic priority weights and integrated into a multi-objective optimization function, so that the scheduling can balance efficiency and economy while significantly improving the service fairness to special groups; the improved whale optimization algorithm with dynamic and adaptive parameters is adopted to quickly generate approximately optimal dispatching instructions in a high-dynamic environment, thereby ensuring real-time response; and through historical data feedback, the model and parameters are continuously optimized, forming a technology closed loop with self-learning ability, thereby improving the long-term robustness and comprehensive performance of the system. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The overall flowchart of the elevator intelligent dispatching optimization method for special groups according to the present application is shown in the figure.

[0023] Figure 2 The improved YOLOv7-tiny model structure integrated with ECA attention is shown in the figure.

[0024] Figure 3 The multi-objective dispatching decision flowchart based on the whale optimization algorithm is shown in the figure. DETAILED DESCRIPTION

[0025] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0026] As shown in the figure, the elevator intelligent detection and dispatching method for special groups provided by the present application has the core of constructing a closed-loop intelligent system for perception, decision-making, execution and optimization. The specific implementation is as follows: Figure 1

[0027] Step S1: The edge computing node collects and processes video data in real time

[0028] ​Deploy edge computing nodes based on ARM architecture (such as NVIDIA Jetson Nano, TX2 series) or lightweight GPU in the elevator waiting hall and the car. The node is connected to a high-definition wide-angle camera, runs an embedded Linux operating system and integrates computer vision and inference acceleration libraries such as OpenCV and TensorRT. The node has a built-in motion detection and background subtraction algorithm for real-time perception of passengers waiting in the video stream. Once the target is detected, the subsequent attribute recognition process is triggered. In order to fundamentally guarantee real-time performance and user privacy, all video stream analysis and processing are completed locally on the edge node, and the original video data is discarded immediately after processing. Only the structured recognition results (including passenger position bounding box, age stage, luggage status, auxiliary tool type, and other attribute labels and their confidence) are uploaded to the central dispatcher through the internal network. Each edge node has local cache and offline transmission functions to deal with temporary network interruptions and ensure that scheduling instructions are not lost.

[0029] Step S2: Passenger fine-grained attribute recognition based on improved lightweight deep learning model

[0030] The S2 step is the basis for realizing personalized scheduling, and the key is to use the improved YOLOv7-tiny model for high-precision and efficient attribute recognition.

[0031] S2.1 Dataset construction and annotation: Collect and build a passenger image dataset covering diverse scenarios (different lighting, angles, and occlusions) with a total of no less than 10,000 images. The dataset specifically includes images and video sequences of different age groups (children, young adults, middle-aged, and elderly), different luggage states (none, small, and large), and the use of various auxiliary mobility tools (wheelchairs, strollers, crutches, and walkers). Use tools such as LabelImg to finely annotate the dataset, including target bounding boxes and corresponding multi-attribute labels (such as "elderly", "carrying large luggage", "using a wheelchair", etc.)

[0032] S2.2 Model structure improvement: In the feature extraction layer of the standard YOLOv7-tiny backbone network, integrate an efficient channel attention (ECA) module after the CBS (Conv-BN-SiLU) module. This module performs global average pooling (GAP) on the feature map, then performs one-dimensional convolution (the convolution kernel size k is determined by the channel number C through the function where γ and b are hyperparameters), and Sigmoid activation to generate a weight vector for each channel, and multiply it with the original feature map channel by channel to achieve feature re-labeling. This allows the model to pay more attention to feature channels useful for fine-grained attribute discrimination while maintaining lightweight, thereby improving recognition accuracy.

[0033] S2.3 Model training and deployment: Using the annotated dataset, the improved model is trained end-to-end in PyTorch or similar frameworks. A composite loss function is adopted , where is the bounding box regression loss, is the objectness loss, is the cross-entropy loss for multi-attribute classification, , , is the balancing weight. After training, the model is converted to an efficient inference format such as ONNX or TensorRT, and deployed to edge computing nodes using tools such as OpenVINO or TensorRT. To further adapt to edge resources, INT8 quantization technology is used to compress the model, ensuring that the inference speed on the target hardware is stable at ≥15 fps, meeting the real-time requirements.

[0034] Step S3: Dynamic priority assignment and multi-objective optimization function construction

[0035] The central dispatcher dynamically assigns individualized scheduling priority weights w i based on the passenger attributes identified in S2. The preset rules for weights are as follows: the baseline weight for ordinary passengers is 1; the weight for elderly passengers or passengers carrying large luggage is 2; the weight for passengers using wheelchairs, crutches, or walkers is 3. This weight allocation rule can be configured according to the building type (e.g., hospital, office building).

[0036] The multi-objective optimization function for dispatching decisions is constructed as follows:

[0037] where, represents the total waiting and boarding time of all passengers in the system (which can be estimated based on the elevator operation curve and floor distance); represents the total energy consumption of all elevators in the system (which can be estimated based on the start-stop and running distance model); represents the total priority service score for special passengers. Among them, d i is the service delay factor of passenger i, defined as , is the current waiting time of the passenger, The preset maximum tolerance waiting time corresponding to its attribute (e.g. 90 seconds for wheelchair users) is set. The initial values of a, b, and g are set as [0.5, 0.2, 0.3]. The system dynamically adjusts according to the real-time scene: when a passenger using an auxiliary mobility tool is identified, g is increased by a certain step (e.g. 0.1); when the instantaneous power or total energy consumption of the system exceeds the preset threshold, b is increased; and during the flat peak period, a can be appropriately increased to improve efficiency.

[0038] Step S4: Online solution based on improved whale optimization algorithm (WOA)

[0039] The whale optimization algorithm (WOA) is used as a meta-heuristic optimization algorithm to quickly solve the above multi-objective function F online. Each possible dispatching instruction sequence (e.g. assigning a passenger to a certain elevator) is encoded as a position vector of a whale individual in the search space.

[0040] The fitness function of the algorithm is directly set as F, and the goal is to find the dispatching scheme that minimizes F. Based on the standard WOA algorithm, dynamic adjustment of search parameters is introduced to improve performance in high-dynamic scenarios of elevator dispatching: the convergence factor a is linearly decreased from 2 to 0, and is changed to adaptively adjust the decreasing rate according to the total number of passengers waiting for the elevator. When the number of passengers is large, the decreasing rate is slowed down to enhance global exploration, and when the number of passengers is small, the decreasing rate is accelerated to quickly converge. The spiral update weight b is also randomly adjusted within a certain range according to the dynamic changes of the system state. The algorithm iteratively updates the population position through three stages of simulation: surrounding the prey, bubble net attack, and random search. Within a certain number of iterations (e.g. 50-100 times) or within a convergence threshold, the approximate optimal solution is found. This solving process is triggered every time there is a new call request, a new passenger is identified, or the system state changes significantly, ensuring the real-time response capability of the dispatching.

[0041] Step S5: Dispatching instruction execution

[0042] The optimal dispatching instruction sequence obtained by solving is sent from the central dispatcher to the core controller of the elevator group control system through a high-reliability industrial bus (such as CAN bus or Modbus protocol), thereby controlling the corresponding elevator to execute the running task. This communication method ensures the real-time and reliability of the instruction transmission.

[0043] Step S6: System monitoring and feedback optimization (closed-loop control)

[0044] This step is the key to realizing long-term self-adaptation and continuous optimization of the system. The system continuously monitors and records historical data such as elevator running state (e.g. actual waiting time, boarding time, energy consumption), passenger flow data, and accuracy of attribute recognition, forming a dispatching log database.

[0045] Feedback and adaptive tuning mainly includes the following two aspects:

[0046] (1) Online incremental learning of the model: periodically (such as every week) or when the recognition accuracy continuously falls below a threshold (such as 92%), the system automatically extracts image data with correct attribute labels (a small part can be reviewed and confirmed by the administrator) from the historical data in a recent period (such as the past month) to perform incremental learning on the attribute recognition model in S2, fine-tune the parameters of the fully connected layer or part of the convolutional layer of the model, so that it can adapt to changes in the environment such as new lighting conditions, passenger clothing, and new auxiliary tools, and prevent model performance degradation.

[0047] (2) Dynamic tuning of algorithm and parameters: based on the evaluation of historical scheduling results, dynamically adjust the system parameters. For example, calculate the mean absolute error (MAE) of the predicted passenger waiting time and the actual measured value in the past period every week. If the MAE continues to increase, automatically adjust the parameters of the whale optimization algorithm in S4 (such as appropriately increasing the population size or the number of iterations). At the same time, based on the completion of the scheduling target (such as special passenger satisfaction survey, total energy consumption bill), use regression analysis or reinforcement learning approach to dynamically optimize the weight coefficients α, β, γ of the multi-objective optimization function in S3, so that the system strategy continuously approaches the optimal balance point of actual operation.

[0048] Through the continuous feedback and optimization of S6, the entire system constitutes a closed-loop control system with self-learning and self-adjusting capabilities, which can maintain efficient, reliable, and humanized operation for a long time.

[0049] Embodiment:

[0050] This embodiment deploys the present application in the group control system of 3 elevators in a 5-story office building. The system recognizes a passenger using a wheelchair through the edge node on the 3rd floor (hardware: NVIDIA Jetson Xavier NX) (attribute recognition confidence 96%, the system assigns priority weight w i= 3). When the central dispatcher receives this information, the total energy consumption of the system is at a normal level, so the weight coefficients are dynamically adjusted to [a = 0.4, b = 0.2, g = 0.4] to give priority to the special passenger. When the improved WOA algorithm (population size 30, maximum iteration 80 times) is used to solve the multi-objective function F, due to the high weight of this passenger and the increasing delay factor d_i, the generated elevator assignment scheme gives priority to assigning the elevator A that is about to arrive at the 3rd floor to serve, rather than following the traditional "first come, first served" principle. The actual running results show that the waiting time of the wheelchair passenger is significantly reduced from an average of 65 seconds before the system is deployed to 22 seconds. After three months of operation and closed-loop optimization of step S6, the attribute recognition model is incrementally learned based on the accumulated about 1000 pieces of labeled data, and the recognition accuracy of the model for "wheelchair" and "large luggage" is improved from the initial 92% and 88% to 96% and 93%, respectively. At the same time, by analyzing the dispatching deviation, the population size of the WOA is adaptively adjusted to 40. Through a questionnaire survey, the satisfaction of the special passenger group is improved from the original 70% to 92%. This embodiment fully proves the effectiveness, adaptability and superiority of the scheme of the present application.

[0051] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements shall be covered within the protection scope of the present application.

Claims

1. A method for intelligent elevator detection and scheduling for special populations, characterized in that, Includes the following steps: S1. By deploying edge computing nodes in the elevator lobby and car, data from multiple video sensors is collected and processed in real time. When passengers waiting for the elevator are detected in the video, the passenger identity attribute recognition process is initiated. S2. Based on an improved lightweight deep learning target detection model, passengers are detected in real time and their attributes are identified in fine detail. The identified attributes include at least age group, whether they are carrying large luggage, and whether they are using assistive devices. S3. Based on the identified passenger identity attributes, dynamically assign personalized scheduling priority weights to them, and construct a multi-objective optimization function that includes the total system time cost, the total system energy consumption cost, and the personalized service weights. S4. Based on real-time updated elevator call requests and passenger attribute information, a metaheuristic optimization algorithm is used to solve the multi-objective optimization function online, generating the current optimal elevator dispatch instruction sequence; S5. Send the elevator dispatch command sequence to the corresponding elevator group control system to control the elevator operation; S6. The system continuously monitors the elevator's operating status and passenger flow data, and periodically provides feedback and adaptive optimization of the parameters of the attribute recognition model and optimization algorithm based on the accumulated historical scheduling data. In step S2, the improved lightweight deep learning object detection model enhances the accuracy of fine-grained attribute recognition by introducing an attention mechanism. In step S3, the dynamic priority weight allocation is adaptively adjusted based on real-time changes in passenger attributes. In step S6, the adaptive tuning includes online learning and optimization of model and algorithm parameters based on historical data to form a closed-loop control system.

2. The method according to claim 1, characterized in that, The improved lightweight deep learning object detection model used in step S2 is a model based on the YOLOv7-tiny architecture and integrating an efficient channel attention (ECA) module. The ECA module is embedded in the feature extraction layer of the YOLOv7-tiny backbone network to enhance the weights of key feature channels, thereby improving the accuracy of identifying passenger age groups, luggage carrying conditions, and use of assistive devices.

3. The method according to claim 2, characterized in that, The implementation of the model includes the following steps: S2.1 Construct a passenger image dataset, which includes images labeled with different age groups, luggage carrying conditions, and use of assistive devices; the assistive devices include wheelchairs, strollers, canes, and walking aids; S2.2 Integrate an efficient channel attention (ECA) module into the feature extraction layer of the YOLOv7-tiny backbone network to improve feature representation capabilities through cross-channel interaction; S2.

3. Use the dataset to train the YOLOv7-tiny network with integrated ECA module for multi-attribute recognition, and use a multi-task learning loss function to simultaneously optimize object detection and attribute classification tasks; S2.4 Deploy the trained model on edge computing nodes and adapt it to the limitations of edge computing resources through model compression technology.

4. The method according to claim 1, characterized in that, The multi-objective optimization function constructed in step S3 is:

5. Among them, This indicates the total waiting time and travel time for the elevator. This represents the total energy consumption of the system. This represents the priority service score for passengers with special needs; α, β, and γ are weighting coefficients, which are dynamically adjusted according to the number and type of special passengers identified. The adjustment rules are as follows: when a passenger using assistive devices is identified, the weight of γ is increased; when the system energy consumption exceeds the threshold, the weight of β is increased.

6. The method according to claim 4, characterized in that, Priority service score The calculation formula is: , where w i For passenger i, different weight values ​​are preset according to its attribute category, d i The factor characterizing the degree of service delay is defined as the ratio of passenger waiting time to the preset maximum tolerable time.

7. The method according to claim 1, characterized in that, The metaheuristic optimization algorithm used in step S4 is the whale optimization algorithm. The fitness function of the whale optimization algorithm is set as the multi-objective optimization function F, and dynamically adjusted search parameters are introduced in the solution process, including adaptively adjusting the convergence factor of the whale optimization algorithm and spiral updating the weights according to the real-time passenger flow.

8. The method according to claim 6, characterized in that, The feedback and adaptive optimization in step S6 includes: incrementally learning the attribute recognition model based on historical scheduling data to update the model parameters to adapt to new scenarios; and dynamically adjusting the parameters of the whale optimization algorithm and the weight coefficients α, β, γ in the multi-objective optimization function based on scheduling effect evaluation, wherein the scheduling effect evaluation is performed by comparing the deviation between the actual scheduling time and the predicted time.

9. The method according to claim 1, characterized in that, In step S5, the edge computing node sends the elevator dispatch command to the core controller of the elevator group control system through the industrial bus. The industrial bus adopts the CAN bus or Modbus protocol to ensure the real-time performance and reliability of the command transmission.

Images