Elevator taking optimization system and method based on artificial intelligence

Abstract

The invention discloses an elevator taking optimization system and method based on artificial intelligence, and relates to an intelligent elevator control technology combining multi-mode perception, prediction model dynamic adjustment and reinforcement learning scheduling strategies. The method comprises the steps that firstly, elevator running states, environment information and passenger behavior data are collected through multiple types of sensors, and multi-modal scene perception vectors are generated; and secondly, elevator loads and floor requirements are predicted in real time through a dynamically-adjusted prediction model, the reward function weight in reinforcement learning is adjusted in a self-adaptive mode on the basis, and accurate response and intelligent scheduling of different scenes are achieved. The system forms a perception-prediction-scheduling closed loop, significantly reduces the waiting time of passengers through multi-objective optimization, reduces the energy consumption, and improves the safety and emergency processing capability. The method is suitable for various high-rise building elevator group control systems, and has high intelligence, flexibility and wide application value.

Description

Technical Field

[0001] The present invention relates to the field of elevator control, and in particular to an elevator riding optimization system and method based on artificial intelligence. Background Art

[0002] With the rapid development of high-rise and intelligent buildings in cities, elevators, as core vertical transportation equipment, have a significant impact on the smooth flow of people and the overall operational efficiency of buildings. Traditional elevator control systems are mostly rule-driven, employing simple first-come, first-served (FCFS) strategies or strategies based on the closest response between the elevator call signal and the elevator's current location. While these scheduling methods were able to meet basic elevator needs in their early days, they have shown significant limitations in modern buildings.

[0003] First, existing elevator systems lack the ability to dynamically perceive environmental conditions and passenger behavior. They are unable to comprehensively consider multiple factors, such as real-time floor density, elevator load, and historical usage patterns, to make scheduling decisions. During peak hours, when elevators frequently start and stop, and when loads fluctuate, scheduling delays, uneven responses, and irrational resource allocation are common. These issues lead to long passenger wait times, high elevator idle rates, and decreased operational efficiency and user satisfaction.

[0004] In existing technologies, the demand forecasting module and scheduling optimization module (such as those based on reinforcement learning or fixed-priority algorithms) of an elevator system are typically independent and statically invoked. The forecasting module is used only to estimate future demand and does not dynamically adjust based on new environmental data perceived during operation. The scheduling module operates based on a fixed or offline trained reward function, making it difficult to respond to emergencies (such as crowds and abnormal events) in a timely manner. This disconnect between forecast results and actual scheduling makes it difficult for the system to respond efficiently, personalized, and securely. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the deficiencies of the prior art, the present invention aims to provide an elevator ride optimization system and method based on artificial intelligence, which solves the problems existing in the prior art. In response to the problem that the prediction module and the scheduling module in the existing system are independent and separated from each other, the system innovatively deeply couples multimodal real-time perception data (such as thermal imaging, sound, vibration, floor call, load, etc.) with the demand prediction model and reinforcement learning scheduling strategy to form a closed-loop adaptive mechanism. Specifically, the system dynamically adjusts the parameters of the demand prediction model and the weight of the reinforcement learning reward function through the perceived real-time environment and passenger status, realizes the synchronous linkage optimization of the prediction and scheduling strategies, can respond to emergencies and crowd anomalies in real time, accurately schedule elevator operation paths, significantly reduce passenger waiting time, reduce energy consumption, and improve safety and service personalization in emergency scenarios, overcoming the defects of delayed scheduling response and overall low efficiency in the prior art.

[0007] (2) Technical solution

[0008] To achieve the above object, the present invention provides the following technical solution: an elevator riding optimization method based on artificial intelligence, comprising the following steps:

[0009] An artificial intelligence-based elevator ride optimization system and method, comprising the following steps:

[0010] S1 Data Collection: Collect multi-modal data through multiple types of sensors to generate raw data stream X t ;

[0011] S2 Data preprocessing: Preprocess the multimodal data and fuse them into a time series feature matrix F t ;

[0012] S3 Passenger behavior and environment analysis: Based on the time series feature matrix F t , analyze the elevator riding mode, crowd density and congestion mode, behavior pattern recognition and environmental anomalies, and generate a multimodal scene perception vector M t ;

[0013] S4 Elevator Demand Forecast: Forecast elevator load and floor demand distribution, detect abnormal demand, and calculate the abnormal adjustment factor ω based on the actual deviation degree to obtain the revised predicted demand

[0014] S5 intelligent scheduling optimization: based on multimodal scene perception vector M t Dynamically adjust the prediction model parameters and adaptively adjust the weight parameters in the reinforcement learning reward function to output the optimal scheduling strategy;

[0015] S6 cloud-edge collaborative management: The cloud analyzes global data, and edge computing enables local rapid response.

[0016] Preferably, in the S4 elevator demand prediction step, the prediction of the elevator load includes: based on historical load sensor data, floor call records and passenger entry and exit statistics, using the long short-term memory network LSTM to build a time series prediction model, using the load characteristics of the past n moments as input, predicting the elevator load change trend in the next t seconds or t moments, and generating the future elevator load prediction value.

[0017] Preferably, in the S4 elevator demand prediction step, the prediction of floor demand includes: based on historical elevator records and floor call data, combined with periodic and sudden demand characteristics, through the autoregressive integrated moving average ARIMA model or the Prophet time series model, predicting the demand intensity of each floor at several future moments, and outputting the floor demand distribution prediction results.

[0018] Preferably, in the step S4 of predicting elevator demand, abnormal demand detection determines the floor demand abnormality flag by the following formula:

[0019]

[0020] Where, d t，i represents the actual demand value of the i-th floor at the current time t, μ t，i represents the historical average demand value,

[0021] σ t，i represents the historical standard deviation, and k is the preset threshold coefficient, which is used to control the sensitivity of anomaly detection.

[0022] Preferably, in the step S4 of predicting the demand for elevators, the calculation formula of the abnormal adjustment factor ω is: Where λ is a weight hyperparameter and ∈ is a small constant that prevents the denominator from being zero.

[0023] Preferably, in the step S4 of predicting the demand for taking the elevator, the revised predicted demand The expression is:

[0024]

[0025] Where D t+H,i is the original forecast value generated based on historical statistical data, and H is the forecast time window.

[0026] Preferably, the S5 intelligent scheduling decision step includes:

[0027] S51 Scheduling Optimization Strategy: Based on multimodal scenario perception results and prediction model output, it designs peak-period optimization strategies, off-peak energy-saving strategies, and special event response strategies to achieve dynamic scheduling optimization for different scenarios.

[0028] Implementation of the S52 scheduling algorithm: Use the deep Q network DQN to optimize elevator scheduling, build a Markov decision process model, define the state space, action space, reward function, state transition probability and discount factor, and combine the experience replay mechanism with the ε-greedy strategy to realize the training and update of the adaptive scheduling strategy.

[0029] Preferably, in the implementation of the S52 scheduling algorithm, the reward function is a reward function including adaptive weights, and the expression is:

[0030] R t =-(α×W t +β×E t -γ×O t )

[0031] Where W t is the average waiting time, E t is the energy consumption index of the elevator, O t is the special optimization score, and α, β, γ are the weight factors of the reward function.

[0032] An elevator riding optimization system based on artificial intelligence, comprising:

[0033] 1. AI Elevator Request Processing Unit: Responsible for receiving and interpreting user elevator requests (including voice, app, touchscreen input, etc.). It intelligently allocates elevator resources based on request priority, user history, real-time elevator status, and floor-level passenger flow forecasts.

[0034] 2. Elevator Status Monitoring Unit: This unit monitors the elevator's operating status in real time using various sensors, including: Position sensor: Monitors the elevator's current floor and direction of travel. Load sensor: Detects the elevator's load to prevent overloading. Speed ​​and acceleration sensor: Determines whether the elevator is experiencing abnormal vibrations or operational issues. Door status sensor: Detects the opening and closing of the elevator doors to determine if there are any door malfunctions.

[0035] 3. AI security monitoring unit: Uses cameras and sensors to conduct real-time security monitoring inside and outside the elevator, automatically detecting abnormal behavior such as fighting and falls in the elevator, and analyzing human posture and behavior trajectory through computer vision.

[0036] Illegal occupation: Detects if the door is blocked while no one is there for a long time.

[0037] Smoke and temperature anomalies: Detect dangerous situations such as fire and smoke.

[0038] 4. User Interaction Interface: Provides multiple ways for users to submit elevator requests, check elevator status, or receive safety alerts, including: voice interaction (voice assistant wakes up the elevator), touch screen input (app or screens inside and outside the elevator to select floors), and contactless interaction (such as facial recognition or gesture control)

[0039] 5. Cloud-edge collaborative computing system: This system is used to run complex AI training models and perform deep learning optimization on elevator usage patterns. Edge computing devices: This system deploys AI models in local elevator controllers to achieve fast response and reduce latency.

[0040] (3) Beneficial effects

[0041] This application aims to provide an artificial intelligence-based elevator ride optimization system and method. This system significantly improves the operational efficiency and intelligence of elevator systems through a multimodal data-driven dynamic prediction-scheduling coupling mechanism. First, the system uses multiple sensors (including thermal imaging, sound, vibration, smoke, floor calls, and load) to perceive the building environment and passenger status in real time. This allows for accurate identification of peak passenger flows, unusual crowds, emergencies, and diverse passenger demands. Compared to traditional prediction models that rely solely on fixed historical data or a single sensor signal, the present invention dynamically adjusts demand prediction model parameters and the reward function weights of reinforcement learning scheduling strategies during operation, enabling a closed-loop real-time optimization of prediction and scheduling, rapidly responding to dynamic changes within the building. By real-time updating of floor demand heat maps and adaptively adjusting scheduling strategies, elevators can pre-assemble and divert passengers to high-demand areas, effectively reducing passenger wait times during peak periods and improving transport capacity. During off-peak or nighttime hours, requests are automatically consolidated to reduce energy consumption and extend the life of elevator equipment. Furthermore, for special events (such as fires, emergency evacuations, and conferences), the system automatically prioritizes capacity allocation, enhancing safety and emergency response capabilities. Overall, the present invention has brought significant technical effects in terms of improving operational efficiency, reducing energy consumption, improving elevator riding experience, enhancing safety and providing intelligent services, and has broad promotion and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0043] Figure 2 Schematic diagram of the flow of the scheduling optimization feature extraction step of the present invention.

[0044] Figure 3 It is a flow chart of the scheduling optimization strategy of the present invention. DETAILED DESCRIPTION

[0045] The following is a combination of the examples of the present invention Figure 1-Figure 3 A clear and complete description of the technical solutions in the embodiments of the present invention is given. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0046] The present invention provides a technical solution:

[0047] An elevator riding optimization system based on artificial intelligence, comprising:

[0048] 1. AI Elevator Request Processing Unit: Responsible for receiving and interpreting user elevator requests (including voice, app, touchscreen input, etc.). It intelligently allocates elevator resources based on request priority, user history, real-time elevator status, and floor-level passenger flow forecasts.

[0049] 2. Elevator Status Monitoring Unit: This unit monitors the elevator's operating status in real time using various sensors, including: Position sensor: Monitors the elevator's current floor and direction of travel. Load sensor: Detects the elevator's load to prevent overloading. Speed ​​and acceleration sensor: Determines whether the elevator is experiencing abnormal vibrations or operational issues. Door status sensor: Detects the opening and closing of the elevator doors to determine if there are any door malfunctions.

[0050] 3. AI security monitoring unit: Uses cameras and sensors to conduct real-time security monitoring inside and outside the elevator, automatically detecting abnormal behavior such as fighting and falls in the elevator, and analyzing human posture and behavior trajectory through computer vision.

[0051] Illegal occupation: Detects if the door is blocked while no one is there for a long time.

[0052] Smoke and temperature anomalies: Detect dangerous situations such as fire and smoke.

[0053] 4. User Interaction Interface: Provides multiple ways for users to submit elevator requests, check elevator status, or receive safety alerts, including: voice interaction (voice assistant wakes up the elevator), touch screen input (app or screens inside and outside the elevator to select floors), and contactless interaction (such as facial recognition or gesture control)

[0054] 5. Cloud-edge collaborative computing system: This system is used to run complex AI training models and perform deep learning optimization on elevator usage patterns. Edge computing devices: This system deploys AI models in local elevator controllers to achieve fast response and reduce latency.

[0055] An elevator riding optimization method based on artificial intelligence includes the following steps:

[0056] S1 Data Collection: Data collection is the foundation of intelligent elevator dispatching and safety monitoring systems. Its purpose is to collect information about elevator operating status, passenger behavior, and the environment. This system uses a multimodal data collection solution. Through various sensors installed inside and outside the elevator, combined with historical data and data input from intelligent devices, it forms a complete real-time monitoring network.

[0057] 1. Elevator status data collection: Through position encoders, speed sensors, load sensors, etc. installed in the elevator car and hoistway, the current elevator position (floor number and precise height), movement speed (up, down, stop) and car load weight (percentage or absolute value) are obtained in real time.

[0058] 2. Floor call and button data collection: Monitor the operation records of the external call panels (up and down buttons) on each floor and the target floor buttons inside the car, and obtain the floor call request status in real time.

[0059] 3. Passenger distribution and crowd sensing: Thermal imaging sensors and infrared array detectors are used to detect the number and distribution of passengers in the elevator car and in the waiting areas on each floor, generating a thermal distribution map of the number of passengers.

[0060] 4. Environmental status data collection: Ambient light intensity sensors, smoke concentration sensors, and temperature and humidity sensors installed inside and outside the car are used to sense changes in ambient light, air quality, temperature and humidity, and to comprehensively judge safety or comfort indicators.

[0061] 5. Sound and vibration perception: A microphone array is set on the top or inside the elevator car to collect the sound inside the elevator and form a continuous audio stream signal x audio (t).

[0062] Using the accelerometer (IMU), real-time detection of three-axis acceleration [a x (t), a y (t), a z (t)], used to monitor vibration patterns, sudden stops, or abnormal mechanical jitter.

[0063] The above multimodal sensor data is uploaded in real time to generate the original data stream X t :

[0064] {x pos (t),x vel (t),x load (t),x call (t),x thermal (t),x light (t),x smoke (t),x audio (t),x acc (t)}

[0065] S2 data preprocessing

[0066] Clean, normalize, align and extract features of the multimodal data collected by S1 to construct the time series feature matrix F t , providing high-quality input for subsequent behavior analysis, prediction, and scheduling. It includes the following sub-steps:

[0067] S21 Data cleaning: Data cleaning aims to remove noise and outliers from the collected data and standardize it to ensure data consistency and reliability.

[0068] S211 Outlier Detection and Removal: The purpose is to eliminate abnormal data caused by sensor failure or external interference to ensure data stability and reliability.

[0069] (1) For outliers of a single variable, the Z-score method is used, which includes the following steps:

[0070] a. Calculate the mean and standard deviation:

[0071]

[0072] Among them, x i represents the i-th data point (e.g., the load sensor reading at a certain moment). μ is the mean of all data points, i.e., the central tendency of the data. σ is the standard deviation of the data, which measures the degree of dispersion of the data.

[0073] b. Calculate the Z-score and filter outliers:

[0074]

[0075] Among them, Z i Represents data point x i The normalized distance relative to the mean μ. If |Z i |>3 means that the data point is far away from the normal distribution range and is an outlier and is removed.

[0076] (2) For anomaly detection of multidimensional data, the local outlier factor (LOF) is used to remove the anomaly, which includes the following steps:

[0077] a. Calculate the average reachability density of the kth nearest neighbor:

[0078]

[0079] Among them, lrd k (A) represents the local density of point A. d(A, N i ) is the distance from point A to its k nearest neighbor points N i The Euclidean distance.

[0080] If the density of data points around A is low, it indicates that it may be an outlier.

[0081] b. Calculate the LOF value:

[0082]

[0083] If LOF k(A)>τ, then the point is considered abnormal.

[0084] The LOF value is used to measure the degree of deviation of the density of point A relative to its neighbors.

[0085] If LOF k (A) is much greater than 1, then A may be an outlier.

[0086] S212 missing value filling: Use linear interpolation to fill missing values. The expression is:

[0087]

[0088] Among them, x t Indicates the currently missing value. t-1 and x t+1 are the observations preceding and following the missing value. The missing value is estimated using the mean of the two points.

[0089] S22 Data Fusion: The purpose of data fusion is to synchronize and correlate multiple sensor data to form complete input data. This includes time alignment and multimodal data association:

[0090] (1) Time alignment: Align different sensor data to a unified time base to ensure that there is no time deviation during data fusion. Based on hierarchical time window matching, it includes the following steps:

[0091] a. Timestamp normalization: All sensor times are converted to millisecond-level UTC time.

[0092] b. Set the time window size W:

[0093] W=max(Δt sensor1 , Δt sensor2 , …)

[0094] Where W represents the maximum time interval between different sensors, which is used to ensure that all data points are matched within the same time window.

[0095] c. Calculate the optimal alignment path P:

[0096]

[0097] P represents the optimal solution for data point alignment. i and Y j They are the time series data of different sensors. Dynamic programming is used to find the minimum time difference matching solution.

[0098] (2) Multimodal data association: Correlate different data sources to build a complete elevator behavior model:

[0099] a. Passenger residence time = time entering the elevator - time exiting the elevator

[0100] Count the number of people entering and leaving through thermal imaging sensors and match them with timestamps.

[0101] b. Elevator load = number of passengers × average weight (dynamic estimation)

[0102] Combine load sensors with passenger statistics to optimize load forecasting.

[0103] c. Floor demand distribution = call button pressing frequency

[0104] Combined with historical call data, analyze high-demand floors.

[0105] S23 Feature Extraction: Feature extraction is used to extract key variables from raw data for subsequent modeling and optimization. The extracted features include sound spectrum features and vibration frequency features:

[0106] (1) Sound spectrum characteristics: used for abnormal sound detection (cries for help, mechanical noises). The spectrum of the signal changes over time by short-time Fourier transform (STFT) is expressed as:

[0107]

[0108] Here, x(n) is the input signal, w(n) is the window function (often a Hanning window), and R is the frame shift, which controls the temporal resolution. The STFT results are then classified using a CNN+LSTM model to identify abnormal sounds.

[0109] (2) Vibration frequency characteristics: used to detect sudden stops and shaking of elevators. Perform frequency analysis and calculate the frequency characteristics of the acceleration sensor signal.

[0110] First, obtain the dominant frequency of the signal through Fast Fourier Transform (FFT):

[0111]

[0112] If the high-frequency components increase, it may indicate a mechanical failure or emergency stop.

[0113] The vibration energy distribution is then calculated by the power spectral density (PSD):

[0114]

[0115] It is used to detect abnormal vibration patterns.

[0116] S24 timing feature matrix generation

[0117] Finally, all the denoised, normalized and extracted multimodal features are combined to generate a multimodal time series feature matrix:

[0118] Used for subsequent behavior and environment analysis, predictive model input and reinforcement learning decision-making.

[0119] S3 Passenger Behavior and Environment Analysis: Based on Time Series Feature Matrix F t , real-time analysis of multimodal sensor data (floor call information, load, thermal imaging, sound, vibration, smoke, etc.) extracts elevator behavior characteristics and environmental status characteristics, providing an accurate basis for subsequent dynamic prediction and scheduling decisions.

[0120] S31 Elevator Riding Mode Analysis

[0121] The system conducts statistical analysis on the elevator demand for each floor and time period based on historical data and real-time perception data. Specifically, it includes:

[0122] Divide into different time periods: such as morning peak (7:00–9:30), midday flat peak, evening peak (17:30–19:30), and late night low peak.

[0123] Statistics on the frequency of elevator calls, average waiting time, average elevator dwell time and other indicators for each floor in each time period to obtain a floor demand heat map

[0124]

[0125] Where N call (f, t) represents the number of elevator calls from floor f within time t, T interval Indicates the statistical time window.

[0126] Analysis of S32 Crowd Density and Congestion Pattern

[0127] Analyze crowd density and distribution patterns through thermal imaging sensors, load sensors, ambient light and entrance flow sensors. Specifically, it includes:

[0128] Calculate the current thermal distribution map of the area inside and outside the elevator

[0129] Count the instantaneous number of people in the waiting area on each floor

[0130] According to the distribution, evaluate whether each area is congested. If It is judged as congested and the scene label is updated.

[0131] S33 Behavioral Pattern Recognition and Anomaly Detection

[0132] Identify behaviors inside and outside the elevator to detect safety risks or abnormal events. Including:

[0133] ① Sound anomaly detection: Use short-time Fourier transform (STFT) to extract the spectral characteristics of the sound signal in the elevator s :

[0134] F s =STFT(x(t))

[0135] Where x(t) is the time domain sound signal.

[0136] The CNN+LSTM combined network model detects abnormal audio events such as cries for help and crashes. CNN extracts time-frequency graph features, while LSTM identifies the temporal dependency patterns of the sounds and outputs classification results.

[0137] ② Elevator vibration abnormality detection

[0138] Based on the acceleration sensor, the elevator acceleration signal a(t) is extracted and the frequency distribution is calculated: S a =FFT(a(t)).

[0139] If the frequency is abnormally high or fluctuates violently, it is judged as an emergency stop or abnormal shaking, and an alarm is triggered.

[0140] S34 Environmental Abnormal Perception

[0141] According to the smoke concentration S collected by the environmental sensor smoke , ambient temperature T, light intensity L and other characteristics to determine whether there is a fire, power outage or other emergency.

[0142] If S smoke >S threshold or T>T threshold , then in M t Add high-risk environment labels and prioritize the adjustment of emergency evacuation strategies.

[0143] S35 multimodal scene perception vector M t Build, contains the following sub-items:

[0144]

[0145] in, This is the floor demand heat map. Distribute thermal information to the population; F is the number of people queuing on the floor; s is the sound spectrum feature; S a is the acceleration signal characteristic; S smoke , T, and L are environmental states.

[0146] Finally, the system integrates multimodal information such as floor demand matrix, crowd density distribution, sound and vibration abnormality characteristics, and environmental abnormality signs to generate a multimodal scene perception vector Mt The vector M t It will be used to dynamically adjust the prediction model and reward function to achieve real-time closed-loop optimization of elevator scheduling.

[0147] S4 scheduling optimization feature extraction:

[0148] After completing the cleaning, fusion, and behavioral pattern analysis of multimodal data, the scheduling optimization feature extraction phase begins. This phase provides high-quality, quantifiable input features for subsequent prediction models and reinforcement learning scheduling strategies. This phase primarily includes: S41 Elevator Load Forecasting: analyzing elevator load fluctuations based on time series to achieve short-term and medium- to long-term forecasts; and S42 Floor Demand Distribution Forecasting: analyzing floor demand trends using historical data to optimize elevator scheduling strategies.

[0149] S41 Elevator Load Forecast

[0150] (1) Data source: To ensure the accuracy of the forecast, the input data for load forecasting includes:

[0151] a. Load sensor data (real-time detection of elevator weight);

[0152] b. Floor call data (recording elevator button call status);

[0153] c. Passenger entry and exit statistics (thermal imaging / door control sensors detect changes in the number of people);

[0154] d. Historical load data (load trends over the past few days / weeks);

[0155] These data constitute a time series and are input into the LSTM (Long Short-Term Memory Network) model for prediction.

[0156] (2) LSTM time series prediction model: Since elevator load is a dynamic process that changes over time, an LSTM neural network is used for time series prediction. This includes the following steps:

[0157] ① Data preprocessing

[0158] a. Data normalization: Processed by Min-Max normalization:

[0159]

[0160] The normalized data is easier to process by LSTM and improves the training convergence speed.

[0161] b. Constructing a time series window

[0162] Input: Load data for the past n moments

[0163] Output: Predict the elevator load in the next t seconds

[0164] ②LSTM structure: LSTM prediction process:

[0165] Input layer: time series window data (load values ​​at the past n moments);

[0166] LSTM layer: learns time-dependent features and extracts elevator load patterns;

[0167] Fully connected layer (Dense): outputs the predicted future load value;

[0168] Loss function: mean square error (MSE), expressed as:

[0169]

[0170] ③ Prediction results and application strategies: Predict loads at different time scales to optimize elevator scheduling:

[0171]

[0172] High load forecast (>80%): Dispatch additional elevators in advance to reduce waiting time.

[0173] Low load prediction (<20%): Consolidated scheduling to avoid no-load operation and improve energy efficiency.

[0174] S42 Floor Demand Distribution Forecast

[0175] (1) Data statistics: To analyze changes in floor demand, we collect elevator records from the past 7-30 days, including the number of calls per floor, the number of passengers entering and exiting, and the demand distribution in different time periods. This forms a time series dataset that can be used to predict future floor demand distribution.

[0176] (2) Prediction model

[0177] The changes in floor demand have certain periodicity and randomness, and ARIMA and Prophet are used for prediction respectively.

[0178] ① For office building demand during working hours, which has relatively stable demand changes, the ARIMA model is used for forecasting. The modeling includes the following steps:

[0179] a. Check for stationarity: Calculate the autocorrelation function (ACF) and partial autocorrelation function (PACF) to determine whether the data is stationary.

[0180] b. Differential operation (if not stable): Y′ t =Y t —Y t-1

[0181] c. Fitting the ARIMA model:

[0182] Y t =φ1Yt -1 +φ2Y t-2 +...+θ1e t-1 +θ2e t-2 +e t

[0183] d. Predict future floor demand and calculate the error.

[0184] ② For environments with obvious demand peaks such as shopping malls and elevator buildings, use the Prophet time series model for prediction. The Prophet prediction process is as follows:

[0185] y(t)=g(t)+s(t)+h(t)+∈ t

[0186] Among them, g(t) is the trend model, s(t) is the periodic component (peak / trough), h(t) is the emergency event (special case), ∈ t is the noise term.

[0187] (3) Optimization strategy: Based on the floor demand forecast results, optimize elevator scheduling, including:

[0188] a. Predict peak floors and dispatch in advance: Increase elevator dwell time on high-demand floors during morning and evening peaks. Reduce unnecessary stops on low-demand floors.

[0189] b. Avoid ineffective stops: Reduce the frequency of stops at low-demand floors to improve operational efficiency. Intelligently optimize stop strategies based on passenger entry and exit history.

[0190] S43 Abnormal demand feature extraction: Combined with the abnormal peak demand and abnormal behavior patterns detected in step S3, the prediction model is dynamically adjusted to enable the elevator system to instantly perceive and adaptively respond to emergencies or abnormal crowd flows.

[0191] Specifically, based on multimodal environmental perception and passenger behavior analysis, the system dynamically monitors real-time floor call patterns and crowd distribution to determine if there are any abnormalities in demand. If the call frequency on a particular floor is significantly higher than the historical average within a short period of time, this is considered abnormal demand.

[0192] To achieve the above anomaly detection, the system determines the floor demand anomaly flag using the following formula:

[0193]

[0194] Among them, d t，i represents the actual demand value of the i-th floor at the current time t, μ t，irepresents the historical average demand value, σ t，i represents the historical standard deviation, and k is the preset threshold coefficient, which is used to control the sensitivity of anomaly detection.

[0195] When abnormal demand is detected, the system calculates the abnormal adjustment factor ω based on the actual deviation degree to dynamically correct the subsequent floor demand forecast results. The specific calculation formula is:

[0196]

[0197] Here, λ is a weight hyperparameter, and ∈ is a small constant that prevents the denominator from being zero.

[0198] The system applies the abnormal adjustment factor ω to the future floor demand forecast value to obtain the revised forecast demand

[0199]

[0200]

[0201] Among them, D t+H，i is the original forecast value generated based on historical statistical data (ARIMA model or Prophet model), and H is the forecast time window.

[0202] The system also uses environmental perception data (such as smoke concentration detection, thermal imaging analysis, and sound anomaly detection) as a basis for dynamic adjustments. When it detects unexpected crowd gatherings, unusual sounds, or emergency evacuation orders, it can simultaneously adjust the anomaly adjustment factor ω and the prediction window H, further modifying the scheduling model parameters and making real-time adjustments to elevator parking strategies, priority settings, and the weights in the reinforcement learning reward function.

[0203] Through the above design, the present invention achieves a close coupling between abnormal demand detection results and prediction models and scheduling decisions, which can significantly improve the adaptability of the elevator system in highly dynamic environments and emergency scenarios, reduce passenger waiting time, and improve transportation efficiency and safety levels, thereby achieving the technical effect of optimizing overall scheduling performance.

[0204] S5 intelligent scheduling decision:

[0205] Goal: Improve elevator response efficiency and reduce passenger wait times. Optimize elevator routes, reduce ineffective stops, and lower energy consumption. Intelligent scheduling optimization is implemented for peak and off-peak periods, as well as special events. Reinforcement learning (DQN) is used to optimize elevator scheduling strategies and achieve adaptive scheduling. This system consists of two modules: S51 Scheduling Optimization Strategy (peak-period optimization, off-peak energy conservation, and special event response) and S52 Scheduling Algorithm Implementation (DQN scheduling strategy based on reinforcement learning).

[0206] S51 scheduling optimization strategy

[0207] Based on the multimodal perception results M t Based on the prediction model output, a multi-level, dynamically adjustable scheduling optimization strategy is designed for different scenarios to achieve high efficiency, adaptability and safety of elevator system operation, specifically including the following aspects:

[0208] First, in the peak-period optimization strategy, concentrated demand areas on each floor are predicted in advance through a comprehensive analysis of elevator demand heat maps, real-time passenger density distribution, and call request distribution. Some elevators are pre-dispatched to high-demand floors, creating a preemptive assembly mechanism to reduce passenger wait times and congestion during peak periods. For example, during the morning rush hour in an office building, the system automatically configures some elevators to go directly to higher floors or limit service to high-rise areas based on data analysis, effectively alleviating delays caused by frequent stops at lower floors.

[0209] Secondly, in its off-peak energy-saving strategy, the system dynamically adjusts elevator operating frequency and stopping strategies during low-traffic periods, such as off-peak and late at night, to reduce empty or low-load operations and lower overall energy consumption. The system can consolidate call requests from nearby floors, centralizing stops and implementing a consolidated service logic. Furthermore, during periods of predicted extremely low demand, it automatically dispatches some elevators into a dormant or standby state, significantly improving energy efficiency.

[0210] Thirdly, for special event response strategies, the present invention can be based on the multimodal perception results M t Automatically identify and quickly respond to emergencies (such as the end of a large conference, peak-hour collective evacuation, fire alarms, and other special circumstances). The system can preemptively dispatch some elevators to specific floors in the early stages of an incident, creating an emergency priority protection mechanism. After the incident is over, it automatically guides elevators to evacuate passengers in an orderly manner, reducing chaos and congestion.

[0211] S52 scheduling algorithm implementation: Using the Deep Q Network (DQN) framework, we implement the comprehensive optimization scheduling of multiple objectives of the elevator. Specifically, we include the following steps:

[0212] First, a Markov decision process model (MDP) for elevator scheduling is constructed, and the state space S, action space A, reward function R, state transition probability P and discount factor γ are defined. t It includes the current elevator location (floor number), the current elevator load rate, the floor call request distribution, the predicted floor demand heat map and the multimodal scene perception features, which are specifically expressed as follows:

[0213] S t =[L t , V t , D t ,H t , Mt ]

[0214] Among them, L t Indicates the current elevator position, V t Indicates load conditions, D t Indicates floor call status, H t is the floor demand heat map output by the prediction model, M t Represents scene perception features (such as crowd density, environmental anomalies, special group requests, etc.).

[0215] Action space A t This includes selecting the next target stop floor, deciding the direction of the elevator (up, down, or stay still), whether to merge requests for floors in the same direction, and special priority adjustments.

[0216] The present invention specifically designs a reward function containing adaptive weights, which is defined as follows:

[0217] R t =-(α×W t +β×E t -γ×O t )

[0218] Among them, W t is the average waiting time, E t is the energy consumption index of the elevator, O t is the special optimization score, α, β, γ are the weight factors of the reward function. The above weight factors are based on the multimodal scene perception result M t Real-time adjustment, that is: [α, β, γ] = f(M t ).

[0219] For example, when peak passenger flow or emergency evacuation scenarios are detected, the system automatically increases α and decreases β to shorten passenger waiting times. At night or during off-peak hours, β is automatically increased to enhance energy conservation and reduce frequent stops. This weight adjustment function f(·) can be automatically fitted using a policy network or rule-based approach, without the need for human intervention.

[0220] The elevator dispatching strategy updates the Q value through deep Q learning. The update formula is:

[0221]

[0222] Combined with the experience replay mechanism, the experience pool D = {(S t ,A t ,R t ,S t+1 )}, by randomly sampling batch updates, we can avoid sample sequence correlation and improve training stability. The loss function is defined as:

[0223] Among them, the target value y t express:

[0224]

[0225] θ represents the current network parameters, θ - is the target network parameter, and is updated synchronously with a fixed step size C to ensure training convergence.

[0226] In addition, to balance exploration and utilization, the system adopts an ε-greedy strategy:

[0227]

[0228] Among them, ∈ is gradually reduced in an exponential decay manner, and the formula is:

[0229] ∈=∈0×exp(-k×t)

[0230] Ultimately, by introducing a closed-loop adaptive optimization chain of multimodal perception → dynamic prediction model → weight adaptive adjustment → reinforcement learning scheduling strategy → action execution → state feedback, this system achieves real-time scheduling optimization and intelligent self-learning of the elevator system for different scenarios, significantly reducing passenger waiting time, improving energy utilization efficiency, and providing priority response guarantees when dealing with emergencies or special events.

[0231] S6 Cloud-Edge Collaboration and System Management: To ensure the efficient operation and long-term stability of the elevator system, this solution combines cloud computing with edge computing to achieve real-time scheduling, remote management, and security monitoring.

[0232] S61 cloud-based big data analysis and model update: The system uploads all elevator operation data to the cloud computing platform, optimizes cross-building scheduling strategies through big data analysis and deep learning technology, and regularly updates AI prediction models.

[0233] (1) Cloud Data Analysis

[0234] Operation data: collect the usage frequency, average waiting time, operating energy consumption, etc. of each elevator.

[0235] Peak pattern recognition: Use time series analysis to identify elevator demand during rush hour and large-scale events, and optimize scheduling in advance.

[0236] Energy-saving analysis: Evaluate the elevator's invalid parking rate and no-load operation rate, and dynamically adjust the energy-saving strategy.

[0237] (2) AI model training and optimization

[0238] Reinforcement learning (DQN) optimized scheduling: Through long-term data accumulation, the scheduling strategy is continuously optimized to improve operational efficiency.

[0239] Forecast model update: LSTM / Prophet is used to regularly update the load forecast model to improve accuracy.

[0240] (3) Cross-system intelligent scheduling

[0241] Multi-building complex management: The cloud platform can manage elevators in multiple buildings and optimize cross-building scheduling (such as sharing scheduling strategies among multiple office buildings).

[0242] Elevator and smart building linkage: Combined with the building automation system (BAS), elevators can be linked with access control, parking systems, and intelligent security to improve management efficiency.

[0243] S62 edge computing real-time response: Deploy edge computing devices (Edge Computing) at key nodes to process data locally, improve real-time performance, and reduce the burden of cloud computing.

[0244] (1) Edge computing application scenarios

[0245] Local decision-making during peak hours: During rush hours or emergencies, edge computing devices can directly optimize scheduling, avoiding waiting for cloud responses.

[0246] Safety monitoring: Real-time analysis of surveillance videos inside and outside the elevator to detect abnormal behavior (such as people being stranded, elevator failure, etc.).

[0247] (2) Low-latency scheduling decisions

[0248] Local data storage + computing: Reduces cloud communication, and scheduling algorithms can be quickly calculated and executed locally.

[0249] Combined with 5G network: further reduce scheduling delay and improve elevator response speed.

[0250] S73 Remote Monitoring and Maintenance: The elevator system is equipped with a remote monitoring platform, allowing managers to check the system status at any time and perform remote maintenance.

[0251] (1) Remote real-time monitoring

[0252] Monitor elevator operation status: display elevator position, load, fault conditions, etc.

[0253] Record historical data: Provide operation logs and fault logs to facilitate analysis and maintenance.

[0254] Safety warning: The system automatically detects anomalies such as overload, elevator jams, abnormal noises, etc., and sends alerts.

[0255] (2) Remote maintenance & OTA updates

[0256] Online diagnosis: Technicians can remotely connect to the elevator system to perform software debugging and troubleshooting.

[0257] OTA upgrade: push software updates through the cloud, optimize scheduling algorithms, and fix vulnerabilities.

[0258] (3) Fault prediction and automatic dispatch

[0259] AI predictive maintenance: Based on equipment health monitoring and machine learning, it detects elevator anomalies in advance and reduces the rate of sudden failures.

[0260] Automatic dispatch: When the system detects a possible elevator failure, it automatically notifies the operation and maintenance personnel and dispatches an order for maintenance.

[0261] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An elevator riding optimization method based on artificial intelligence, characterized in that: The following steps are involved: S1 Data Collection: Collect multi-modal data through multiple types of sensors to generate raw data stream X t ; S2 Data preprocessing: Preprocess the multimodal data and fuse them into a time series feature matrix F t ; S3 Passenger behavior and environment analysis: Based on the time series feature matrix F t , analyze the elevator riding mode, crowd density and congestion mode, behavior pattern recognition and environmental anomalies, and generate a multimodal scene perception vector M t ; S4 Elevator Demand Forecast: Forecast elevator load and floor demand distribution, detect abnormal demand, and calculate the abnormal adjustment factor ω based on the actual deviation degree to obtain the revised predicted demand S5 intelligent scheduling optimization: based on multimodal scene perception vector M t Dynamically adjust the prediction model parameters and adaptively adjust the weight parameters in the reinforcement learning reward function to output the optimal scheduling strategy; S6 cloud-edge collaborative management: The cloud analyzes global data, and edge computing enables local rapid response.

2. The method for optimizing elevator riding based on artificial intelligence according to claim 1, characterized in that: In the S4 elevator demand prediction step, the prediction of elevator load includes: based on historical load sensor data, floor call records and passenger entry and exit statistics, using the long short-term memory network LSTM to build a time series prediction model, using the load characteristics of the past n moments as input, predicting the elevator load change trend in the next t seconds or t moments, and generating the future elevator load prediction value.

3. The method for optimizing elevator riding based on artificial intelligence according to claim 1, characterized in that: In the S4 elevator demand prediction step, the prediction of floor demand includes: based on historical elevator records and floor call data, combined with periodic and sudden demand characteristics, through the autoregressive integrated moving average ARIMA model or the Prophet time series model, predicting the demand intensity of each floor at several future moments, and outputting the floor demand distribution prediction results.

4. The method for optimizing elevator riding based on artificial intelligence according to claim 1, characterized in that: In the step S4 of elevator demand prediction, abnormal demand detection determines the abnormal floor demand flag by the following formula: Where, d t，i represents the actual demand value of the i-th floor at the current time t, μ t，i represents the historical average demand value, σ t，i represents the historical standard deviation, and k is the preset threshold coefficient, which is used to control the sensitivity of anomaly detection.

5. The method for optimizing elevator riding based on artificial intelligence according to claim 1, characterized in that: In the S4 elevator demand forecasting step, the calculation formula of the abnormal adjustment factor ω is: Where λ is a weight hyperparameter and ∈ is a small constant that prevents the denominator from being zero.

6. The method for optimizing elevator riding based on artificial intelligence according to claim 1, characterized in that: In the S4 elevator demand forecasting step, the revised forecast demand The expression is: Where D t+H，i is the original forecast value generated based on historical statistical data, and H is the forecast time window.

7. The method for optimizing elevator riding based on artificial intelligence according to claim 1, characterized in that: The S5 intelligent scheduling decision-making step includes: S51 Scheduling Optimization Strategy: Based on multimodal scenario perception results and prediction model output, it designs peak-period optimization strategies, off-peak energy-saving strategies, and special event response strategies to achieve dynamic scheduling optimization for different scenarios. Implementation of the S52 scheduling algorithm: Use the deep Q network DQN to optimize elevator scheduling, build a Markov decision process model, define the state space, action space, reward function, state transition probability and discount factor, and combine the experience replay mechanism with the ε-greedy strategy to realize the training and update of the adaptive scheduling strategy.

8. The method for optimizing elevator riding based on artificial intelligence according to claim 1, characterized in that: In the implementation of the S52 scheduling algorithm, the reward function is a reward function including adaptive weights, and the expression is: R t =-(α×W t +β×E t -γ×O t ) Where W t is the average waiting time, E t is the energy consumption index of the elevator, O t is the special optimization score, and α, β, γ are the weight factors of the reward function.

9. A system for executing the elevator riding optimization method based on artificial intelligence according to any one of claims 1 to 8, characterized in that: include: AI Elevator Request Processing Unit: Configured to analyze user requests and intelligently allocate elevator resources based on priority, historical habits, and real-time data; Elevator status monitoring unit: It is used to integrate position, load, thermal imaging, sound and other sensors to collect elevator operation and environmental data in real time; AI security monitoring unit: Enabled to detect abnormal behavior, smoke / temperature anomalies, and mechanical failures through cameras and sensors; User interaction interface: It is used to support voice, touch screen and contactless interaction; Cloud-edge collaborative computing system: It is used to run AI models in the cloud to optimize global scheduling, and edge devices to process real-time decisions locally.

Images