Elevator energy-saving operation control system based on large model

Abstract

The application provides an elevator energy-saving operation control system based on a large model, relates to the technical field of elevator control, and comprises a data acquisition module, a structure construction module and a link identification module.The data acquisition module is used for collecting operation parameters of an elevator group in real time, obtaining a multi-dimensional state vector, and generating a fusion feature set according to a preset space-time mapping relationship.The structure construction module is used for performing coordinate positioning in a predefined state reference system based on the fusion feature set, selecting key feature points of a current operation situation core representation, and constructing a closed multi-dimensional feature structure.The link identification module is used for calculating mutual information and dynamic correlation degrees between vertices connected by each edge in the multi-dimensional feature structure, weighting each edge according to the correlation degrees, and obtaining key links.The application cooperates with multiple modules to capture multi-element correlation of elevator group operation, analyze the situation and adapt to double models to realize adaptive scheduling, and achieves the collaborative optimization of energy saving and service quality.

Description

Technical Field

[0001] This invention relates to the field of elevator control technology, and in particular to an elevator energy-saving operation control system based on a large model. Background Technology

[0002] Energy optimization of elevator group control systems is a crucial aspect of building energy management. Current common energy-saving control schemes often employ scheduling strategies based on statistical patterns or preset rules, such as time-of-use pricing, historical average passenger flow data, or simple real-time signal thresholds. These methods can achieve certain energy-saving effects in scenarios with relatively stable and predictable passenger flow patterns. However, in real-world operating environments with high passenger flow and complex dynamic changes, such as a large general hospital during peak hours, elevator calls from different floors (outpatient, laboratory, inpatient departments) exhibit sudden, interwoven, and continuously evolving characteristics within a short period. Simultaneously, the load, location, direction of travel, and energy consumption of multiple elevators are monitored in real-time. Interactive influences present limitations in existing methods for analyzing such high-dimensional, strongly coupled real-time operational situations. Their state analysis often focuses on monitoring and responding to single or a few isolated parameters, such as waiting time and car load. This approach may be limited in capturing and quantifying the complex, nonlinear relationships between multiple system elements, such as spatiotemporally distributed call signals, multi-car collaborative states, and real-time energy consumption. These limitations could lead to inaccurate modeling of the overall operational situation, potentially affecting the adaptability and optimization potential of subsequent energy-saving scheduling strategies. Furthermore, there is room for improvement in balancing energy efficiency and service quality when dealing with complex and dynamic passenger flow patterns. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide an elevator energy-saving operation control system based on a large model, which can identify normal operating conditions and effectively capture special modes such as emergency evacuation, thereby improving the intelligence level and reliability of state judgment.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] An elevator energy-saving operation control system based on a large model includes:

[0006] The data acquisition module is used to collect the operating parameters of the elevator group in real time, obtain a multi-dimensional state vector, and generate a fusion feature set according to the preset spatiotemporal mapping relationship.

[0007] The structure construction module is used to perform coordinate positioning in a predefined state reference system based on the fused feature set, select key feature points representing the core of the current operating state, and construct a closed multi-dimensional feature structure.

[0008] The link identification module is used to calculate the mutual information and dynamic correlation between the vertices connected by each edge in the multidimensional feature structure, and to weight each edge according to the correlation to obtain the key link;

[0009] The manifold reconstruction module is used to divide the multidimensional feature structure into several sub-regions according to the critical link, and to reconstruct the manifold of the state vector field for each sub-region to obtain the reconstructed manifold of each sub-region.

[0010] The coupling analysis module is used to generate the corresponding local covariance matrix based on the manifold of each sub-region after reconstruction, and to calculate the state coupling degree sequence.

[0011] The factor generation module is used to extract state correction factors that characterize the overall operating status based on the state coupling degree sequence, and to fuse the state correction factors with the operating parameters to determine the elevator operating status judgment result.

[0012] The scheduling decision module is used to determine the elevator's operating status and obtain a first elevator scheduling instruction or a second elevator scheduling instruction through a pre-trained dual-model library using either the first or second scheduling model. It then executes the first or second elevator scheduling instruction to control the elevator group.

[0013] The above-described solution of the present invention has at least the following beneficial effects:

[0014] The system aggregates multi-source parameters in real time, including the spatiotemporal characteristics of elevator call signals, car load, and position, through a data acquisition module. These parameters are then processed using spatiotemporal mapping to generate a fused feature set, overcoming the limitations of single-parameter analysis. By using feature coordinate positioning and key feature point filtering, a closed multi-dimensional feature structure is constructed, transforming scattered operational data into structured information focusing on the core situation. This allows the system to comprehensively capture the multi-element relationships in elevator group operation. By calculating the mutual information and dynamic correlation degree of each node in the feature structure, a quantitative assessment of the strength of multi-element relationships is achieved, identifying key links affecting the operational status. Finally, the complex feature structure is divided into sub-regions and its manifold is reconstructed, revealing the inherent rules of local operational states. The laws are clearly represented, solving the problem of insufficient capture of complex relationships. The state coupling degree sequence is generated by the coupling analysis module, and the state correction factor reflecting the overall situation is extracted. After being fused with the original operating parameters, the state is judged. This judgment method based on multi-dimensional feature fusion can not only identify normal operating states, but also effectively capture special modes such as emergency evacuation, improving the intelligence level and reliability of state judgment. Based on the accurate state judgment results, the scheduling strategy is adaptively switched through dual model libraries. This differentiated scheduling method enables the system to match the optimal scheduling strategy according to different operating states, realizing the coordinated optimization of energy efficiency and service quality. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of an elevator energy-saving operation control system based on a large model, provided by an embodiment of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0017] like Figure 1 As shown, embodiments of the present invention propose an elevator energy-saving operation control system based on a large model, comprising:

[0018] The data acquisition module is used to collect the operating parameters of the elevator group in real time, obtain a multi-dimensional state vector, and generate a fusion feature set according to the preset spatiotemporal mapping relationship.

[0019] The structure construction module is used to perform coordinate positioning in a predefined state reference system based on the fused feature set, select key feature points representing the core of the current operating state, and construct a closed multi-dimensional feature structure.

[0020] The link identification module is used to calculate the mutual information and dynamic correlation between the vertices connected by each edge in the multidimensional feature structure, and to weight each edge according to the correlation to obtain the key link;

[0021] The manifold reconstruction module is used to divide the multidimensional feature structure into several sub-regions according to the critical link, and to reconstruct the manifold of the state vector field for each sub-region to obtain the reconstructed manifold of each sub-region.

[0022] The coupling analysis module is used to generate the corresponding local covariance matrix based on the manifold of each sub-region after reconstruction, and to calculate the state coupling degree sequence.

[0023] The factor generation module is used to extract state correction factors that characterize the overall operating status based on the state coupling degree sequence, and to fuse the state correction factors with the operating parameters to determine the elevator operating status judgment result.

[0024] The scheduling decision module is used to determine the elevator's operating status and obtain a first elevator scheduling instruction or a second elevator scheduling instruction through a pre-trained dual-model library using either the first or second scheduling model. It then executes the first or second elevator scheduling instruction to control the elevator group.

[0025] In this embodiment of the invention, a data acquisition module gathers multi-source parameters such as the spatiotemporal characteristics of elevator call signals, car load, and position in real time. These parameters are then processed through spatiotemporal mapping to generate a fused feature set, overcoming the limitations of single-parameter analysis. By using feature coordinate positioning and key feature point filtering, a closed multi-dimensional feature structure is constructed, transforming scattered operational data into structured information focusing on the core situation. This enables the system to comprehensively capture the multi-element relationships in elevator group operation. By calculating the mutual information and dynamic correlation degree of each node in the feature structure, a quantitative assessment of the strength of multi-element relationships is achieved, identifying key links affecting the operational situation. Dividing the complex feature structure into sub-regions and completing manifold reconstruction clearly represents the inherent laws of local operational states, solving the problem of insufficient capture of complex relationships. By generating a state coupling degree sequence, a state correction factor reflecting the overall situation is extracted and fused with the original operational parameters for state judgment, improving the intelligence and reliability of state judgment. Based on the state judgment results, adaptive switching of scheduling strategies is achieved through a dual model library, enabling the system to match the optimal scheduling strategy according to different operational states, realizing the synergistic optimization of energy efficiency and service quality.

[0026] In a preferred embodiment of the present invention, the operating parameters of the elevator group are collected in real time to obtain a multi-dimensional state vector, and a fusion feature set is generated according to a preset spatiotemporal mapping relationship, including:

[0027] The system collects real-time timestamps and floor numbers of call signals from each floor, real-time load values ​​of each elevator car, and real-time position of each car in the hoistway to obtain a set of raw operating parameters. Specifically, this includes: integrating a signal acquisition module inside the call panel on each floor. This module is linked to the trigger circuit of the call button. When a passenger presses the call button, the module immediately records the timestamp of the trigger action and the corresponding floor number. The timestamp uses a system-wide unified high-precision clock signal. Four sets of pressure strain gauge load sensors are symmetrically installed on the load-bearing beam at the bottom of each elevator car. When the car bears weight, the load-bearing beam undergoes slight deformation, causing the strain gauges to deform synchronously, resulting in a change in the resistance value of the strain gauges. This change in resistance is converted into a Wheatstone bridge circuit. The voltage signal (i.e., the pressure electrical signal) is processed by the signal amplification module and then transmitted to the data processing unit for load calculation. The real-time load value is first established offline through calibration to establish the correspondence between the pressure electrical signal and the load. This involves placing standard weights of known weight sequentially inside the car and recording the average voltage values ​​output by four sets of sensors for each weight, forming a voltage-load value calibration curve. A univariate linear regression equation is then fitted: Load value = (Current average voltage value - No-load voltage value) × Calibration coefficient + No-load compensation value. Here, the no-load voltage value is the output voltage of the sensor when the car is unloaded, the calibration coefficient is the slope of the regression equation, and the no-load compensation value is used to correct system errors. The data processing unit receives the real-time voltage signal. Next, outliers in the four sets of data are removed, such as data deviating from the average by more than 5%. Then, the average voltage value of the remaining data is calculated and substituted into the above equation to obtain the real-time load value. Positioning sensors (such as photoelectric sensors) are installed at equal intervals along the vertical direction on the inner wall of the elevator shaft. The spacing between adjacent sensors matches the spacing of the guide shoes of the elevator car. At the same time, a signal receiving module is installed on the top of the car. When the car is running, the receiving module triggers the sensors in the shaft in sequence. Combining the pulse counting principle of elevator operation (each triggering a sensor counts as one pulse, and the number of pulses multiplied by the sensor spacing is the running distance), the absolute position coordinates of the car in the shaft are obtained in real time. Each acquisition module establishes a communication connection with the elevator group control main controller through an industrial Ethernet. Next, data is transmitted using a combination of polling and interrupts. For elevator call signals, an interrupt-triggered method is used to ensure that the data is uploaded immediately after the signal is generated. For load and position data, a polling method with a period of 100 milliseconds is used to upload the data to avoid data redundancy. After receiving the data uploaded by each module, the main controller classifies and stores the data according to the data type, elevator number / floor number, and timestamp format, forming a set of raw operating parameters containing elevator call signal data (timestamp, floor number), car load data (elevator number, timestamp, real-time load value), and car position data (elevator number, timestamp, real-time position coordinates). At the same time, missing or abnormal data, such as load values ​​exceeding the equipment range or sudden changes in position coordinates, are marked.

[0028] Data collected within the same time slice from the original operating parameter set is aligned according to three dimensions: call signal, real-time load, and real-time location. These dimensions are then vectorized and combined to form a multi-dimensional state vector. Specifically, time is divided into continuous time slices of fixed length, with the slice length set to match the polling period for load and location data (100 milliseconds), covering all call signals generated within that time slice. For each time slice, all data with timestamps falling within that time slice are selected from the original operating parameter set and then categorized according to three dimensions: call signal, real-time load, and real-time location. Specifically, the call signal dimension summarizes the call status of each floor within that time slice, forming a correspondence between floor number and whether a call has occurred; the real-time load dimension summarizes the load values ​​of each elevator car according to the elevator number; and the real-time location dimension… Similarly, the position coordinates of each car are summarized according to the elevator number to achieve data alignment. The aligned data for each dimension is vectorized. The call signal dimension uses the total number of floors in the building as the vector length, with each element corresponding to a floor. If there is a call signal for that floor in the time slice, the element value is 1, and if there is no call signal, it is 0, forming a call signal vector. The real-time load dimension uses the total number of elevators in the elevator group as the vector length, with each element corresponding to the real-time load value of one elevator, forming a load vector. The real-time position dimension also uses the total number of elevators as the vector length, with each element corresponding to the real-time position coordinates of one elevator, forming a position vector. Finally, the call signal vector, load vector, and position vector are concatenated in sequence to form a multi-dimensional state vector containing information from all dimensions. Each time slice corresponds to an independent multi-dimensional state vector.

[0029] The multidimensional state vector is input into a preset spatiotemporal mapping relationship for processing. This spatiotemporal mapping relationship, based on the correspondence rules between floors and time, extracts trajectory data describing the car's displacement process, intensity data describing the accumulated call signals at each floor, and collaborative data describing the relative positional relationships of multiple cars from the multidimensional state vector. Specifically, this includes: defining three core rules for the spatiotemporal mapping relationship; the trajectory extraction rules define parameters such as the time window length for position data and the speed calculation method, where the time window length is set based on the average time it takes for the elevator to travel one floor (approximately 2 seconds); the intensity extraction rules define the call signal... The signal accumulation time range and intensity value statistics method are defined, with the accumulation time range determined by referencing the time interval (approximately 1 second) of concentrated passenger elevator calls. The collaborative extraction rules specify the elevator pair combination method and the calculation logic for relative position and direction coordination, ensuring accurate reflection of the interaction status of multiple elevator cars. For each elevator, the position vector elements in the multi-dimensional state vector of the current time slice and the preceding five consecutive time slices (a total of 600 milliseconds, covering approximately 1 / 3 of the elevator's floor travel time) are retrieved to obtain the elevator's position sequence. The position difference between adjacent time slices is calculated, and each difference is divided by the time slice length (100 milliseconds). The instantaneous operating speed is obtained by combining the position sequence and speed sequence in chronological order to form trajectory data describing the elevator's displacement process. The trajectory data of the elevator group is a collection of trajectory information for all elevators. For each floor, the call signal vector elements in the multidimensional state vector of the current time slice and the previous 9 consecutive time slices (a total of 10 time slices, 1 second) are retrieved. Each element with a value of 1 indicates that there is a call in that time slice, and 0 indicates that there is no call. The number of times the value of 1 is counted among these 10 elements, and this number is directly used as the call signal strength value for that floor. If the count is 0, the strength value is... 0 indicates an intensity value of 3, and the maximum intensity value is 10, meaning there is an elevator call in every time slot within 1 second. The intensity values ​​of all floors are arranged in ascending order of floor number, forming intensity data with the same dimension as the total number of floors, reflecting the cumulative activity level of elevator call signals on each floor. A combination algorithm is used to generate all unique elevator pairs in the elevator group, such as elevator 1 and elevator 2, elevator 1 and elevator 3, etc. For each elevator pair, designated as elevator A and elevator B, the position vector elements in the multidimensional state vector of the current time slot are called to obtain the position PA of A and the position PB of B, and the relative position value between them is calculated. =PA-PB, if A positive value indicates that A is above B; a negative value indicates that A is below B. The direction of travel is determined by combining the current speeds (VA, VB) of A and B in the trajectory data (V>0 is upward, V<0 is downward, V=0 is stationary), and a direction coordination value is generated (1 for the same direction, -1 for the opposite direction, and 0 for at least one elevator stationary). The relative position value of each pair of elevators minus the direction coordination value is used as coordination information. The information of all elevator pairs is summarized to form coordination data, which reflects the interactive state of multiple elevator cars.

[0030] Trajectory data, intensity data, and coordination data are normalized and concatenated to form a fusion feature set. Specifically, this includes: using min-max normalization to eliminate dimensional differences, ensuring all data types fall within the same numerical range; in trajectory data, position values ​​are mapped to the 0-1 range using the top and bottom floors of the shaft (Hmax and Hmin) as boundaries, through normalized position = (original position - Hmin) / (Hmax - Hmin); speed values ​​are mapped to the 0-1 range using the elevator's rated maximum speed (Vmax) as boundaries, through normalized speed = (original speed + Vmax) / (2 × Vmax) (covering both upward and downward speeds); intensity data is mapped to the maximum number of calls per second (10 times) as boundaries, through normalized intensity... =Original intensity / 10 is mapped to the 0 to 1 interval; In the collaborative data, the relative position value is mapped to the 0 to 1 interval by normalizing the relative position = (ΔP+Hmax) / (2×Hmax) with the maximum height of the shaft (Hmax) as the boundary; The directional collaborative value retains the original values ​​of -1, 0, and 1 (which are already standardized results); According to the fixed order of trajectory data, intensity data, and collaborative data, the three types of normalized data are horizontally spliced ​​to form a high-dimensional feature vector. Identification information (time slice number + elevator group number) is added to each feature vector. The feature vectors of all time slices are arranged in chronological order to finally form a fused feature set. This feature set completely contains the three core information of elevator operation trajectory, call intensity, and car collaboration.

[0031] This embodiment addresses the situational awareness bias caused by incomplete data collection through data classification and storage and anomaly marking mechanisms. Time-slice division enables data alignment, providing a structured representation of the elevator group's operational status for each time slice. Based on spatiotemporal mapping, it extracts three core data categories: trajectory, intensity, and coordination, focusing on key elevator operation characteristics and avoiding interference from redundant data. Normalization eliminates dimensional differences, improving data usability and the accuracy of status representation. Spatiotemporal mapping captures the inherent relationships between multiple elevator operation elements. A feature set integration consolidates this related information, enabling the system to perceive complex situations such as sudden elevator call signals and frequent car interactions during peak hospital visit periods, thus resolving the problem of insufficient capture of complex relationships.

[0032] In a preferred embodiment of the present invention, based on a fused feature set, coordinate positioning is performed in a predefined state reference system, and key feature points representing the core characteristics of the current operating state are selected to construct a closed multidimensional feature structure, including:

[0033] Based on the type of each feature item in the fused feature set, the feature items are mapped to the corresponding coordinate axes in a predefined state reference system. Specifically, this includes: defining the structure of the predefined state reference system, which is a multi-dimensional Euclidean space. Its coordinate axes are divided according to the feature type of the fused feature set, with three main categories of coordinate axis clusters. The first category is the trajectory feature coordinate axis cluster, containing the position and velocity coordinate axes of each elevator (two coordinate axes for a single elevator, and 2U coordinate axes for U elevators); the second category is the intensity feature coordinate axis cluster, with one call intensity coordinate axis for each floor, and M axes for M floors; the third category is the coordination feature coordinate axis cluster, with two coordinate axes for each pair of elevators (relative position coordinate axis and directional coordination coordinate axis), and K pairs of elevators containing 2K axes, forming a standardized state reference system. Each feature item in the fused feature set carries a type identifier (trajectory type / intensity type / coordination type) and subdivision attributes (such as...). The mapping rules for elevators (elevator 1 - location, floor 3 - call intensity, elevators 1 and 2 - relative positions) are as follows: The corresponding coordinate axis cluster is matched based on the type identifier of the feature item, and then the specific coordinate axis within the cluster is matched through subdivision attributes. For example, for the feature item elevator 3 - speed in the fusion feature set, the trajectory feature coordinate axis cluster is first matched through the trajectory class identifier, and then the speed coordinate axis of elevator 3 within the cluster is matched through the subdivision attributes of elevator 3 - speed. Similarly, the feature item floor 5 - call intensity is matched through the intensity class identifier to the intensity feature coordinate axis cluster, and then the call intensity coordinate axis of floor 5 is matched. The program iterates through all feature items in the fusion feature set, assigns a unique corresponding coordinate axis in the state reference system to each feature item according to the above mapping rules, and records the feature item-coordinate axis mapping relationship table. If a feature item attribute is ambiguous, such as without an elevator number, an exception handling mechanism is triggered, marking the feature item as pending confirmation and pausing the mapping until manual verification before re-execution.

[0034] Based on the numerical value of each feature, the position of the feature on the mapped coordinate axis is determined, resulting in the coordinate set of all feature items. Specifically, each coordinate axis in the state reference system has a preset fixed numerical range. In the trajectory feature, the position coordinate axis ranges from the bottom to the top of the elevator shaft, such as 0 meters to 50 meters; the speed coordinate axis ranges from the maximum rated downward speed to the maximum rated upward speed of the elevator, such as -2 meters / second to 2 meters / second; in the intensity feature, the call intensity coordinate axis ranges from 0 to 10; in the coordination feature, the relative position coordinate axis ranges from -50 meters to 50 meters (matching the maximum height of the shaft), and the directional coordination coordinate axis ranges from -1 to 1 (corresponding to the directional coordination value). (The value range is consistent), each coordinate axis is divided into scales at intervals of range / 1000; the feature item values ​​in the fused feature set have been normalized, and the normalized values ​​need to be converted into actual coordinate values ​​on the coordinate axes, that is, coordinate value = (normalized value × coordinate axis value range span) + coordinate axis starting value, where the coordinate axis value range span = coordinate axis maximum value - coordinate axis minimum value; after performing coordinate value calculation on all feature items in the fused feature set one by one, the feature identifier, corresponding coordinate axis, and coordinate value information of each feature item are associated and stored to obtain a coordinate set containing the spatial location information of all feature items. This set is presented in the form of a data table, with each row corresponding to the coordinate data of one feature item.

[0035] A density-based clustering algorithm was used to analyze the coordinate set to identify areas with dense coordinate point distribution. Specifically, this involved: using a density-based spatial clustering algorithm to analyze the coordinate set, first determining the algorithm's core parameters (neighborhood radius and minimum number of points), and then adapting it to the characteristics of elevator operation data through offline training; training data preparation was conducted by selecting four typical passenger flow scenarios: weekday morning peak (7:30-9:00), midday off-peak (12:00-13:30), hospital peak (10:00-11:30), and nighttime off-peak (21:00-22:30). One hour of operation data was collected for each scenario, generating four sets of coordinates to comprehensively cover different passenger flow densities and operating states; the search range for the neighborhood radius was determined as follows. The coordinate axis scale is divided into 0.1 to 1.0 units, resulting in 10 candidate values ​​with a step size of 0.1. The search range for the minimum number of points is determined to be 3 to 10, with a step size of 1, resulting in 8 candidate values. Several independent parameter combinations are generated through a full combination method to meet the parameter requirements under different density scenarios. Then, all parameter combinations are applied to the coordinate sets of 4 typical passenger flow scenarios. Each parameter set independently performs clustering operations on each coordinate set, and 4 core quantitative indicators are recorded simultaneously: intra-cluster point spatial concentration (measured by the average Euclidean distance of all points in the cluster, the smaller the value, the higher the concentration), inter-cluster boundary overlap (measured by the proportion of the intersection area of ​​the spatial boundaries of different clusters, a value of 0 indicates no overlap), core feature proportion (intra-cluster trajectory). The criteria for clustering are: the proportion of class and intensity-based feature items to the total number of feature items, and the proportion of noise points (the proportion of noise points to the total number of points in the coordinate set). A dual screening mechanism is constructed based on quantitative indicators. The first level is a hard threshold, requiring inter-cluster boundary overlap = 0 (ensuring clear outlines of dense regions), core feature proportion ≥ 90%, and noise point proportion ≤ 5% (ensuring high core feature aggregation). The second level is a comprehensive quality index screening, introducing a silhouette coefficient to quantify the clustering effect from a global perspective. This coefficient assesses the rationality of clustering by measuring intra-cluster compactness and inter-cluster separation. For each feature point, the a-value (the average Euclidean distance between this point and all other points in the cluster, reflecting intra-cluster compactness; a smaller value is better) and the b-value (the distance between this point and all points not belonging to its own cluster) are calculated. The average Euclidean distance reflects the inter-cluster separation; a larger value is better. The silhouette coefficient of a single point is calculated using the formula (b-a) / max(a, b), with a value range of [-1, 1]. (When the value is close to 1, the point is tightly clustered within the cluster and has clear boundaries with other clusters; when the value is close to -1, the point is incorrectly assigned to a cluster; when the value is close to 0, the point is at the boundary between clusters.) The average silhouette coefficient of all feature points is taken to obtain the overall silhouette coefficient of this parameter combination in the current scene. The qualified threshold is set to ≥0.7 (verified by historical data, corresponding to a close correlation of features within the cluster and significant differences in sub-situations between clusters). For the 6 sets of parameters after the first screening, their overall silhouette coefficients on 4 sets of coordinates are calculated and sorted from high to low. Finally, the neighborhood radius of 0 is selected.The optimal parameter combination with 5 coordinate axis scale units and a minimum of 5 points performs excellently in all scenarios, namely, the overlap between cluster boundaries is 0, the proportion of core features is ≥92%, the proportion of noise points is ≤3%, and the overall contour coefficient is ≥0.75, far exceeding the qualified threshold, demonstrating full-condition adaptability. When a feature point contains at least 5 other feature points within a neighborhood centered on itself with a radius of 0.5 coordinate axis scale units, it is determined to be a core point (corresponding to the stable state of multi-feature collaboration in elevator operation, such as the area of ​​continuous high-volume calls and the feature of the elevator's uniform speed trajectory). Traversing all points in the coordinate set, the point type is determined according to the parameter standard: core point (number of points in the neighborhood ≥5), density-reachable point (number of points in the neighborhood <5 but falls in the neighborhood of the core point, corresponding to subordinate features associated with stable features, such as the elevator speed change around the high-volume call area), and noise point (neither a core point nor a density-reachable point). For occasional fluctuations, such as accidental call signals or instantaneous elevator vibration offset data; starting from the first unlabeled core point, it and all density-reachable points are included in the initial cluster. This process is then recursively expanded from other core points within the cluster until no new density-reachable points can be added, forming a complete cluster. After labeling the points within a cluster, the operation is repeated from the next unlabeled core point, ultimately resulting in multiple non-overlapping clusters. Noise points are labeled separately and not included in subsequent analysis. For each cluster, the maximum and minimum values ​​of all points within the cluster on each coordinate axis are extracted to form rectangular boundaries in multidimensional space (e.g., elevator 1 - position axis 10 to 15 meters, 3rd floor - call intensity axis 2 to 5 meters). Small, occasional clusters containing less than 10 points are removed, while large clusters containing ≥10 points are retained. The corresponding spatial range is the core dense area (reflecting the synergistic effect of multiple features, such as the traction correlation between an elevator at a certain location and the call on the corresponding floor).

[0036] In each densely distributed area, the coordinate points located at the center of the area and whose corresponding feature items have importance weights higher than the preset importance weight threshold are selected as key feature points. Specifically, this includes: for each identified dense area, determining the center coordinates of the area by calculating the arithmetic mean of the coordinate values ​​of all coordinate points within it on each coordinate axis. This center coordinate represents the central tendency of the feature distribution within the area; calculating the importance weights of feature items, which quantify the degree of influence of feature items on elevator operation status. The calculation logic is as follows: selecting average waiting time and energy consumption per unit time as core indicators of elevator operation efficiency. These two indicators directly reflect the scheduling effect and energy-saving goals, and are key bases for status evaluation; collecting elevator operation data from the past 3 months, and for each feature item, such as the speed of elevator 1 and the call intensity on the 5th floor, calculating its mutual information value and variance contribution with the two efficiency indicators, with both values ​​ranging from 0 to 1, and then obtaining the final weight through weighted summation; the core of the mutual information value is to see whether the efficiency indicators change when the feature item changes, and the more synchronous the change, the higher the value; taking the call intensity on the 5th floor and the average waiting time as examples, the past 3 months of elevator operation data are used to calculate the mutual information value and variance contribution with the two efficiency indicators. For three months, the call intensity (recorded once every 100 milliseconds) and average waiting time (recorded once every minute) on the 5th floor were mapped by time. For example, the average call intensity within one minute was paired with the waiting time within that minute to form a data set, resulting in over 10,000 paired data sets (call intensity value, waiting time value). The call intensity was divided into three groups (low: 0-2 calls / minute, medium: 3-6 calls / minute, high: 7-10 calls / minute), and the waiting time was also divided into three groups (short: 0-10 seconds, medium: 11-20 seconds, long: over 21 seconds). Then, statistics were compiled on the call intensity and waiting time. How many pairs of synchronously changing combinations are there, such as long elevator call times, low call intensity, and short waiting times? And how many pairs of asynchronous combinations are there, such as high call intensity but short waiting times? The similarity ratio is obtained by dividing the number of synchronously changing combinations by the total number of data sets. This ratio is the simplified mutual information value. For example, if synchronous combinations account for 80%, the mutual information value is 0.8, indicating that the higher the call intensity, the longer the waiting time, and the strong non-linear correlation between the two. The mutual information value of each feature item with the average waiting time and energy consumption per unit time is calculated separately. Finally, the average of the two values ​​is taken as the final mutual information value of the feature item.

[0037] The core of variance contribution is to examine the fluctuations in efficiency indicators and determine how much of the fluctuation is caused by this feature. The higher the explained proportion, the larger the value. Taking elevator 1's load and energy consumption per unit time as an example, besides elevator 1's load, other features that may affect energy consumption are selected, such as elevator 1's speed and the floor height it travels to, for a total of three features. Before calculating variance contribution, it is necessary to clarify the calculation logic of energy consumption fluctuation. Essentially, it measures the magnitude of the difference in elevator energy consumption over different time periods. First, select 10 consecutive 1-minute time periods over the past 3 months (e.g., 9:00 to 9:01 every day) and record the elevator's energy consumption per unit time. Add the 10 data points together and divide by 10; this is the baseline energy consumption value for this period. Subtract the actual energy consumption of each time period from the actual energy consumption of the previous period. Average energy consumption is calculated to obtain the variance value (positive and negative numbers represent values ​​higher or lower than the benchmark). All variance values ​​are squared (to eliminate positive and negative effects) and then summed. This sum represents the total energy consumption fluctuation; a larger value indicates a more significant difference in energy consumption across different time periods. After clarifying the calculation of energy consumption fluctuation, the variance contribution of elevator 1's load is calculated. First, it is assumed that energy consumption fluctuation is only related to three characteristics (elevator 1 load, elevator 1 speed, and operating floor height). The combined variance of these three characteristics is calculated to determine how much energy consumption fluctuation can be explained (e.g., 70%). Then, the elevator 1 load characteristic is removed, and only the remaining two characteristics are considered to explain how much fluctuation (e.g., 40%). The difference between the explained ratio when the elevator is loaded and the explained ratio when it is unloaded is the total energy consumption fluctuation. 1. The variance contribution of load; for example, 30%, the corresponding variance contribution is 0.3, indicating that the load change can explain 30% of the energy consumption fluctuation alone; each feature item should also have its variance contribution calculated separately for the two efficiency indicators, and the average value is taken as the final result; weight value = (mutual information value × 0.6) + (variance contribution × 0.4), where the mutual information value can capture nonlinear correlations (most of the relationships in elevator operation are nonlinear), which is more critical for situational characterization, so it is given a higher weight (0.6); the variance contribution focuses on linear explanatory power and serves as a supplement (0.4); the weight values ​​of all feature items are statistically analyzed, and the 75th quantile is taken as the preset importance weight threshold (finally determined to be 0.6), that is, 75% of the features If the feature weight is lower than this value, the threshold can effectively filter out a few highly important features. Within each dense region, the Euclidean distance between all points in the cluster and the coordinates of the region center is calculated, and the 10 points with the smallest distance are selected as candidate central feature points to ensure that the candidate points have spatial centrality. The weight values ​​of the feature items corresponding to the 10 candidate points are queried, and candidate points with weight values ​​> 0.6 are retained. These points satisfy both spatial centrality and high importance, and are the key feature points of the region. If there are no candidate points with sufficient weight in a certain region, such as when the feature correlation is weak in some regions during off-peak periods, the candidate point with the largest weight value is selected as the key feature point to ensure that each dense region has a corresponding key feature point and to fully cover the core operational status.

[0038] Connect any two key feature points to form an edge, and calculate the Pearson correlation coefficient of the original set of operating parameters corresponding to the two key feature points connected by each edge within a set time length. Specifically, this includes: using a fully connected strategy to connect all key feature points, that is, constructing an undirected edge (without direction distinction, only representing the association) between any two different key feature points; assigning a unique identifier to each edge and storing the information of the two key feature points it connects, such as point A corresponding to elevator 1-speed, point B corresponding to the 3rd floor-intensity), forming an association table of edges and points; setting the time length to 10 seconds, corresponding to 100 time slices, and selecting... The original operating parameters set for the current time slice and the preceding 99 time slices are used. This time length can cover the typical cycle of the elevator completing 1 to 2 starts and stops, reflecting short-term dynamic correlation, while avoiding the distortion of the correlation caused by excessively long time. For the two key feature points connected by each edge (corresponding to feature terms X and Y), their original parameter sequences within 10 seconds are extracted (sequence X and sequence Y, both with a length of 100). The Pearson correlation coefficient is calculated, with a value range of -1 to 1. A positive value indicates that the trend of feature terms changes is consistent (such as the elevator speed increases when the call intensity increases), and a negative value indicates that the trend is opposite. The larger the absolute value, the stronger the correlation.

[0039] Edges with Pearson correlation coefficients higher than a preset connection threshold are retained. A closed multidimensional feature structure is formed by all retained edges and the key feature points connected to them. Specifically, this involves determining the preset connection threshold, which distinguishes between significantly correlated and weakly correlated edges. The determination process is as follows: Key feature point pair correlation data for eight typical passenger flow scenarios over the past month are collected, resulting in 1000 valid feature point pairs. The Pearson correlation coefficient for each pair is calculated. Statistical analysis of the 1000 correlation coefficients reveals that the absolute values ​​of the coefficients are concentrated between 0 and 0.8, with coefficients having an absolute value ≥ 0.3 accounting for approximately 30%. These coefficients correspond to feature point pairs, such as the call intensity on the 5th floor - elevator 2 location, and elevator 1 load - elevator 1 speed, which have a clear logical relationship in actual operation. Coefficients with an absolute value < 0.3 are mostly accidental relationships, such as the weak relationship between the call intensity on the 10th floor and elevator 3 speed during off-peak hours. Therefore, 0.3 is selected as the preset connection threshold, and edges with an absolute correlation coefficient ≥ 0.3 are judged as valid edges with significant correlation. All constructed edges are traversed, and the absolute value of the Pearson correlation coefficient for each edge is calculated. This is then compared to the threshold for filtering. Edges with an absolute coefficient ≥ 0.3 are recorded, along with their two connected key feature points, correlation coefficient values, and corresponding... Feature information, such as edge-1: point A (elevator 1 speed) - point B (3rd floor strength), coefficient 0.52; edges with an absolute coefficient value < 0.3 are marked as redundant edges and deleted; based on the retained valid edges and corresponding key feature points, a multi-dimensional feature graph model is constructed using graph structure modeling tools (such as NetworkX), where key feature points are the vertices of the graph, and each vertex is labeled with corresponding feature information, such as elevator 2 - location; valid edges are the edges connecting vertices, and the weight of the edges is set to the absolute value of the corresponding Pearson correlation coefficient (reflecting the correlation strength); since valid edges are all core correlation edges, vertices are connected by edges Multiple connected subgraphs are naturally formed, such as a subgraph consisting of call intensity, elevator location, and elevator speed. Within each subgraph, vertices are interconnected by edges to form loops, such as point A-point B-point C-point A, or a network structure, such as point A connecting points B, C, and D, and point B connecting points C and E. These structures form closed regions in space, which can completely reflect a certain operational sub-state, such as the elevator response state in the high call area. All connected subgraphs combine to form a complete closed multi-dimensional feature structure. This structure presents the network of connections between the core features of elevator operation in the form of a visual graph, with each closed substructure corresponding to an operational sub-state.

[0040] This embodiment establishes a standardized state reference system, accurately mapping various feature items in the fused feature set to their corresponding coordinate axes, and determining the spatial coordinate position of each feature item. This achieves a structured and visual representation of the multi-dimensional features of elevator operation, solving the problem of chaotic feature management. It automatically identifies dense regions of feature points using a density clustering algorithm, focusing on the core areas where features co-aggregate during elevator operation. By combining spatial centrality and feature importance as dual criteria to select key feature points, it ensures that the selected feature points reflect the core characteristics of the operational status, improving the relevance and accuracy of the status representation. By calculating the Pearson correlation coefficient of the parameters corresponding to the key feature points, it quantifies the linear correlation strength between features. Based on the correlation strength threshold, it selects core correlation edges and constructs a closed multi-dimensional feature structure, solving the problem of fuzzy correlation analysis.

[0041] In a preferred embodiment of the present invention, the mutual information and dynamic correlation degree between the vertices connected by each edge in the multidimensional feature structure are calculated, and each edge is weighted according to the correlation degree to obtain the critical link, including:

[0042] For each edge in the closed multidimensional feature structure, the original operating parameter set corresponding to the two key feature points connected by each edge is obtained. The mutual information value of the original operating parameter set corresponding to the two key feature points within a set time span is calculated. Specifically, this includes: First, traversing each edge in the closed multidimensional feature structure, querying the two key feature points connected by the edge through the edge's identifier information, and then extracting the original operating parameter set corresponding to the two feature points based on the feature identifier-time slice range association table. The time span is set to 30 seconds (matching the typical elevator operating cycle), corresponding to 300 time slice data (100 milliseconds / time slice). For example, if an edge connects two key feature points, elevator 1 - speed and 3rd floor - call intensity, then 300 original parameter data of these two features within the same 30 seconds are extracted to form two parameter sequences (Sequence A: elevator 1 speed data, Sequence B: 3rd floor call intensity data). The second step is to calculate the mutual information value. The mutual information value reflects the nonlinear correlation by measuring the degree of information sharing between two parameter sequences. The calculation steps are as follows: Discretize the two parameter sequences according to numerical intervals. The elevator speed sequence is divided into 4 intervals according to [-2, -1), [-1, 0), [0, 1), and [1, 2] m / s, and the call intensity sequence is divided into 4 intervals according to [0, 2), [2, 4), [4, 6), and [6, 10] times / s, converting continuous data into discrete categories; count the number of occurrences of sequence A in each interval and calculate the marginal probability P(A), such as the proportion of the number of times A occurs in the interval [1, 2]; similarly, calculate the marginal probability P(B) of sequence B; then count the number of occurrences of A and B in the interval combination and calculate the joint probability P(A, B), such as the proportion of the number of times A occurs in [1, 2] and B occurs in [6, 10]; calculate the information entropy and mutual information. First, calculate the information entropy of a single sequence, the formula is: The information entropy is obtained by multiplying the probability of each interval by its logarithm, summing the results, and then taking the negative. And H(B); then calculate the joint information entropy. Final mutual information value = The value ranges from 0 to 1, with a larger value indicating a stronger non-linear correlation between the two parameters. Within a set time span, the time is divided into multiple consecutive sliding windows. Within each sliding window, the Pearson correlation coefficient of the original set of operating parameters corresponding to the two key feature points is calculated. Specifically, based on a 30-second time span, the window size is set to 5 seconds (containing 50 time slices of data, reflecting the short-term correlation characteristics of the parameters), and the sliding step size is 2 seconds (containing 20 time slices, avoiding excessive window overlap and data redundancy). At 30 seconds... The formula for calculating the number of windows that can be divided within a time span is: Number of windows = (Total number of time slices - Number of time slices in a window) ÷ Number of time slices in a step + 1, ensuring complete coverage of the entire time span. Taking a certain window as an example, extract the parameter sequences of two key feature points within the window (sequence X and sequence Y, both with a length of 50), calculate the Pearson correlation coefficient, and the value range is from -1 to 1. Positive values ​​indicate positive correlation, negative values ​​indicate negative correlation, and the larger the absolute value, the stronger the linear correlation. Perform the above calculation on each of the 13 windows to obtain 13 Pearson correlation coefficient values.

[0043] The average and variance of the Pearson correlation coefficients within all sliding windows are calculated. These averages and variances are used as measures of dynamic correlation strength. Specifically, this involves: summing the 13 correlation coefficient values ​​and dividing by the number of windows to obtain the average Pearson correlation coefficient. The average value reflects the average linear correlation strength between the two parameters over the entire time span; the closer the value is to 1 or -1, the more stable the overall correlation. The variance of the Pearson correlation coefficient is calculated; this reflects the degree of fluctuation in the correlation coefficient, embodying the dynamic characteristics of the correlation. The smaller the variance, the more stable the linear correlation between the two parameters; the larger the variance, the more significant the change in correlation strength over time. The calculated average and variance of the Pearson correlation coefficients are used as two core measures of dynamic correlation strength, jointly characterizing the strength and dynamic stability of the correlation.

[0044] The mutual information value and the dynamic correlation degree metric are normalized separately to obtain the normalized mutual information value and the normalized dynamic correlation degree metric. The normalized mutual information value and the normalized dynamic correlation degree metric are then weighted and summed according to preset weights to obtain the comprehensive correlation strength value of the corresponding edge. Specifically, this includes: using the min-max normalization method to map the two metrics (mean and variance) of mutual information value and dynamic correlation degree to the interval between 0 and 1 respectively; mutual information value normalization first collects the mutual information values ​​of all edges in the closed multidimensional feature structure, and obtains the global minimum and global maximum values ​​of the index through traversal statistics; then, the original mutual information value of each edge is substituted into the formula normalized value = ( The normalized mutual information value for each edge is calculated by dividing the original value by the global minimum value by the global maximum value. When normalizing the dynamic correlation metric, the dynamic correlation includes two measures: the average and variance of the Pearson correlation coefficient. Differential processing methods are needed based on the characteristics of the indicators. For average normalization, the average value is a positive indicator; a larger value indicates a stronger overall linear correlation. The min-max normalization formula is directly used. First, the global minimum and maximum values ​​of the average Pearson correlation coefficients for all edges are calculated. Then, the original average value of each edge is substituted into the calculation to obtain the normalized average value. For variance normalization, the variance is a negative indicator; a smaller value indicates a more stable linear correlation. If directly... Using the conventional formula leads to results that contradict practical implications. Therefore, the formula is adjusted to Normalized Value = (Global Maximum Value - Original Value) ÷ (Global Maximum Value - Global Minimum Value). This formula maps variance to a positive indicator, ensuring that a larger normalized value represents stronger association stability. After this operation, each edge will correspond to three normalized indicators: normalized mutual information value, normalized mean, and normalized variance. The normalized mean and variance are then combined into a single dynamic association degree comprehensive value, achieving a synergistic assessment of association strength and dynamic stability. Considering that the mean reflects association strength and the variance reflects association stability, both are equally important for dynamic association degree, and both are weighted at 0.5, using a weighted summation method. The formula for calculating the dynamic correlation strength is: Dynamic Correlation Comprehensive Value = (Normalized Mean × 0.5) + (Normalized Variance × 0.5). Combining elevator operating characteristics with the synthesized comprehensive correlation strength value, the correlations between features during elevator operation are often non-linear. For example, the correlation between call intensity and elevator response speed is affected by multiple factors such as elevator location and passenger load. Therefore, it can capture the mutual information value of non-linear correlations, which has more situational representation value than focusing on the dynamic correlation comprehensive value of linear dynamic correlations. Based on this, a preset weight is set: normalized mutual information value accounts for 0.6, and dynamic correlation comprehensive value accounts for 0.4. The final comprehensive correlation strength value is obtained by weighted summation, and the formula is: Comprehensive Correlation Strength Value = (Normalized Mutual Information Value × 0.6) + (Dynamic Correlation Comprehensive Value × 0.5).4) This value is fixed between 0 and 1. The closer the value is to 1, the more stable the non-linear relationship between the two feature points connected by the corresponding edge, and the more reliable the linear dynamic relationship. It is a key edge reflecting the core relationship of elevator operation.

[0045] Based on the comprehensive association strength value, each edge in the closed multidimensional feature structure is weighted and assigned a weight value. All edges in the closed multidimensional feature structure are sorted from largest to smallest according to their weight values. The top N percentile edges are selected as critical links, where N is a preset integer percentage threshold. Specifically, this involves: assigning a unique weight value to each edge in the closed multidimensional feature structure, forming an edge identifier-weight value association table; and determining the preset integer percentage threshold through historical data verification, specifically selecting eight typical passenger flow scenarios (weekday morning peak, hospital visit peak, shopping mall weekend peak, midday peak, etc.). The system collects continuous operational data for 30 days (12 hours per day) for each scenario (off-peak, nighttime low-peak, holiday off-peak, and sudden peak passenger flow). Each scenario generates 100 sets of closed multi-dimensional feature structures with edge weights, totaling 800 validation samples. The top 5% of edges with the highest overall correlation strength in each scenario are defined as the baseline core correlation edges. These edges correspond to the core collaborative relationships in elevator operation, such as the correlation between call intensity and elevator response speed, and the correlation between elevator load and energy consumption. The statistical indicator is set as the proportion of the top N% of edges covering the baseline core correlation edges. The higher this proportion, the better the selection effect of the N value. The calculation formula is: Coverage Ratio. Example = (Number of baseline core related edges included in the first N% of edges ÷ Total number of baseline core related edges in this scenario) × 100%; Test four candidate values ​​of N: 15, 20, 25, and 30, and calculate the average coverage ratio of each N value in 800 samples. When N=15, the average coverage ratio is 68%, but in some scenarios, such as sudden peak passenger flow, the coverage ratio is only 52%, indicating missing core related edges; when N=20, the coverage ratio of all 8 scenarios is ≥80%, the average coverage ratio reaches 86%, and there are no missing core related edges; when N=25, the average coverage ratio increases to 92%, but the number of links increases by 42%, which will lead to... This leads to redundancy in subsequent analysis; when N=30, the number of links increases by 88%, making the redundancy problem more prominent; considering both the integrity of core coverage and the simplification of links, N=20 is determined as the preset integer proportion threshold. This threshold can be fine-tuned according to the special characteristics of elevator usage scenarios, such as super high-rise elevators and medical elevators. After adjustment, the coverage ratio needs to be re-verified through small sample data to be ≥80%; sort all edges from largest to smallest according to their weight values, calculate the total number of edges after sorting, and determine the number of filters by multiplying the total number by N%. For example, if there are 100 edges in total, filter the top 20 edges. The edges ranked in the top N% are the key links. These edges represent the associations between the core features in elevator operation.

[0046] This embodiment uses a weighted sorting method to filter the top N% of critical links, focusing on the core correlations in elevator operation and eliminating redundant edges with weak correlations, thus reducing the amount of decision-making data and computational complexity. A sliding window mechanism quantifies correlation fluctuations through variance, enabling the correlation strength assessment to reflect short-term dynamic changes in real time. For example, the correlation between call intensity and elevator speed fluctuates significantly during morning rush hour; this mechanism can accurately capture this characteristic, providing a basis for dynamically adjusting scheduling strategies. A standardized process for parameter extraction, index calculation, normalization, and weighted filtering is established. All calculations are based on quantitative data, and weights and thresholds are determined through historical data verification, avoiding the subjectivity of manual intervention and improving the reliability of critical link identification.

[0047] In a preferred embodiment of the present invention, the multidimensional feature structure is divided into several sub-regions according to the critical link, and the manifold reconstruction of the state vector field is performed on each sub-region to obtain the manifold of each sub-region after reconstruction, including:

[0048] Based on the critical links and the weighted values ​​assigned to each edge within them, the key feature points connected by critical links in the closed multidimensional feature structure are clustered to form several topologically connected sub-regions. Specifically, this involves: using the identified critical links as core edges and the key feature points connected by these critical links as vertices, constructing a critical association graph using graph structure modeling tools (such as NetworkX). Two rules are clearly defined in the critical association graph: vertices only contain key feature points connected by critical links (isolated feature points not connected by critical links are removed); edges only retain key links, and the edge weights use the comprehensive association strength value. The criterion for topological connectivity is that if there is one or more continuous critical links between two key feature points, i.e., they are indirectly connected through other key feature points and critical links, then these two feature points are considered topologically connected. For example, feature point A is connected through a critical link... If feature point B is connected to feature point C via a critical link, then A, B, and C all belong to the set of topologically connected features. The connected component extraction algorithm of the critical association graph (such as a depth-first search algorithm) is invoked to traverse all vertices in the critical association graph. Starting from the first unlabeled critical feature point, all feature points topologically connected to that point are recursively searched to form an initial sub-region. The affiliation of all feature points within this sub-region is marked, and the operation is repeated starting from the next unlabeled feature point until all critical feature points are assigned to sub-regions. Each connected component corresponds to a topologically connected sub-region, named sequentially according to clustering order: sub-region 1, sub-region 2, ..., sub-region Q (Q being the total number of sub-regions). Simultaneously, a sub-region-critical feature point-critical link association table is generated, recording the type (e.g., elevator speed, call intensity), quantity, and associated critical link information of the feature points contained in each sub-region.

[0049] For the first sub-region, extract the multi-dimensional state vectors corresponding to all key feature points contained in the first sub-region, and perform dimensionality reduction processing on the multi-dimensional state vectors corresponding to all key feature points contained in the first sub-region to obtain the low-dimensional data point set corresponding to the first sub-region. Specifically, this includes: First, based on the sub-region-key feature point association table, obtain all key feature points contained in the first sub-region; for each key feature point, extract its original operating parameter sequence within a set time span, with the time span following 30 seconds, corresponding to 300 time slices (100 milliseconds / time slice), forming the parameter sequence (length 300) of that feature point; align the parameter sequences of all key feature points in the sub-region according to the time slices and combine them into a multi-dimensional state vector, the dimension of the vector = number of key feature points in the sub-region × 300; for example, if sub-region 1 contains two feature points, elevator 1 speed and 3rd floor call intensity, and the parameter sequence length of each feature point is 300, then the dimension of the multi-dimensional state vector is 600, the first to 300th positions in the vector are elevator 1 speed data, the 301st position is... The first step involves standardizing the multidimensional state vector to eliminate dimensional differences between different feature parameters. The standardization formula is: Standardized value = (Original parameter value - Average value of the parameter sequence of the feature point) ÷ Standard deviation of the parameter sequence of the feature point, ensuring that the value range of each feature parameter is uniformly [-1, 1]. The second step uses the t-SNE (t-distributed random neighborhood embedding) algorithm for dimensionality reduction. The core is to preserve the local correlation structure of high-dimensional data and map the multidimensional state vector to a 2D space. Specifically, for the standardized multidimensional state vector, the high-dimensional similarity between any two time slices within the vector is calculated. This similarity is used to quantify the degree of correlation between two time slices in the high-dimensional space. The closer the value is to 1, the more consistent the feature trends are (e.g., the elevator speed and call intensity of the two time slices change synchronously). The closer the value is to 0, the greater the feature difference. The high-dimensional similarity is measured using a Gaussian kernel function. The calculation logic is that the closer the distance, the higher the similarity. The specific formula is: Similarity = (Gaussian kernel function) / ... The Euclidean distance between the two data points is calculated using high-dimensional coordinates. For example, the high-dimensional coordinates of time slice A are (…). Time slice B is ( (v represents elevator speed, c represents call intensity), the Euclidean distance calculation formula is: Euclidean distance = The smaller the distance, the closer the numerator (the square of the Euclidean distance between the two data points) in the Gaussian kernel function is to 0, and the closer the exponential result is to 1, thus increasing the similarity. The bandwidth parameter's core function is to adjust the sensitivity range of the Gaussian kernel function, i.e., to control the Euclidean distance at which similarity can still be determined. Its value needs to be adaptively determined based on the local density of the feature point data to avoid local correlation distortion caused by fixed parameters. Specifically, for each time slice of data, its 10 nearest neighbors in the high-dimensional space are counted (neighboring data = 10, to adapt to the temporal continuity of the elevator data). The average Euclidean distance between the 10 nearest neighboring data points and the current data is calculated, and this average distance is used as the local density index of the data. Then, the median of the local density indices for all time slices is taken, and 0.8 times the median is used as the final bandwidth parameter. For example, if the local density of a certain sub-region is high (such as during the morning rush hour, when feature data is dense), the average distance between the neighboring data is small, and the bandwidth parameter is reduced accordingly. The sensitivity range of the Gaussian kernel function becomes narrower, and only data that are very close to each other are judged as similar. If the local density is low (such as during the nighttime low peak), the bandwidth parameter is increased, the sensitivity range is widened, and potential related data is avoided from being missed.

[0050] After calculating the high-dimensional similarity, a low-dimensional coordinate is randomly initialized for each time slice of data in the 2D target space (coordinate values ​​are set to [-5, 5] to ensure uniform initial distribution). Then, the low-dimensional similarity between any two data points in the low-dimensional space is calculated. This similarity is a mirror image of the high-dimensional similarity in the low-dimensional space, used to measure whether the low-dimensional coordinates retain the high-dimensional correlation characteristics. The low-dimensional similarity is calculated using the t-distribution kernel function, which is characterized by a wider tail, providing more sufficient low-dimensional distribution space for differing data. The specific calculation logic is similar to that of the high-dimensional similarity: first, the Euclidean distance between the two data points in the 2D space is calculated, and then substituted into the t-distribution kernel function formula. This formula also follows the rule that the closer the distance, the higher the similarity. However, compared to the Gaussian kernel function, the decay rate of low-dimensional similarity with increasing distance is more gradual, effectively avoiding data crowding. Finally, the KL divergence (relative entropy) between the high-dimensional and low-dimensional similarity distributions is minimized using the gradient descent method. The KL divergence is used to quantify the difference between the two distributions; the closer the value is to 0, the closer the low-dimensional distribution is to the high-dimensional distribution. The calculation formula is... During the iterative optimization process, each iteration adjusts the low-dimensional coordinates of each data point based on changes in the KL divergence. If the low-dimensional similarity of a data point with its neighboring data is significantly lower than its high-dimensional similarity, it is moved towards the neighboring data. This "significantly lower low-dimensional similarity" is defined in conjunction with the association characteristics of elevator feature data; specifically, if the ratio of low-dimensional similarity to high-dimensional similarity is less than 0.3, it indicates that the association between the two data points in the low-dimensional space is far from reaching the association level in the high-dimensional space. In this case, the low-dimensional coordinates of the data point need to be moved towards its neighboring data. The magnitude of the movement is positively correlated with the difference in similarity. The larger the degree difference, the farther the distance of a single movement (the movement step size is set to 0.01, which can be fine-tuned according to the convergence efficiency); the above adjustment process continues until the KL divergence converges (the number of iterations is set to 1000; if the change in KL divergence in 50 consecutive iterations is less than the convergence threshold of 0.001, it is also considered converged). Finally, the 2D coordinates corresponding to each time slice data are obtained. The 2D coordinates of all time slice data are summarized to form the low-dimensional data point set corresponding to the first sub-region. Each point in the point set corresponds to a time slice state in the high-dimensional space, and the local correlation features of the sub-region features are preserved.

[0051] Based on the low-dimensional data point set corresponding to the first sub-region, triangular patches are formed by selecting every three neighboring low-dimensional data points, and the normal vector direction of each triangular patch is calculated. Multiple triangular patches are then stitched together to form a continuous surface, resulting in a low-dimensional manifold representing the local data structure characteristics of the first sub-region. This manifold is denoted as the reconstructed manifold of the first sub-region. Specifically, this involves: First, for each point in the low-dimensional data point set, the K-nearest neighbor algorithm (K value set to 5) is used to find the five nearest neighbors, with the distance calculated using Euclidean distance in 2D space. Second, for each point and its neighbors, the Delaunay triangulation algorithm is used to construct triangular patches. This algorithm ensures that the interior angles of the generated triangles are all greater than 30 degrees, avoiding narrow triangles. For the long triangle, the specific rules are as follows: First, select three non-collinear neighboring points and connect them to form a triangle, ensuring that the circumcircle of this triangle does not contain other low-dimensional data points. Then, iterate through all points to generate an initial set of triangular facets. The third step is to calculate the normal vector of each triangle facet. The normal vector represents the spatial orientation of the facet and reflects the changing trend of the sub-region features. The calculation steps are as follows: Take the coordinates of the three vertices of the triangle facet A (x1, y1), B (x2, y2), and C (x3, y3), and construct vectors AB and AC. The coordinates of vector AB are (x2-x1, y2-y1), and the coordinates of vector AC are (x3-x1, y3-y1). Since these are 2D coordinates, extend them to 3D space (set the z-coordinate to 0). The cross product result is... The z-component = (x-component of AB × y-component of AC) - (y-component of AB × x-component of AC), where the sign of the z-component represents the direction of the normal vector (positive for upward, negative for downward). Dividing the cross product by its magnitude yields the unit normal vector, ensuring that the normal vector magnitudes of all facets are uniform. The fourth step involves splicing the generated triangular facets. Before splicing, the angle between the normal vectors of adjacent facets must be calculated. This angle is used to determine whether the spatial orientation of two facets is consistent. The core calculation principle is to convert the dot product of the normal vectors into an angle value. The specific calculation steps and judgment logic are as follows: the normal vectors of adjacent facets are all standardized unit normal vectors (magnitude 1, direction representing the facet's orientation). Let the unit normal vector of facet A be vector n1(x1, ... Given y1, z1), the unit normal vector of adjacent face B is vector n2(x2, y2, z2). Since the faces are in the extended space of the same plane in the 2D manifold reconstruction, the z component of the normal vector is the core of orientation determination (positive z represents upward, negative z represents downward). First, calculate the dot product of the two unit normal vectors. The dot product result directly reflects the degree of overlap of the vectors. Second, convert the dot product into an angle (in radians) using the inverse cosine function. Third, convert the radians into angles, i.e., angle (angle) = angle (radians) × (180 / π). Perform a splicing operation based on the angle result. If the angle is less than 15 degrees (determined to be consistent orientation, belonging to continuous structure), then directly splice the two faces along the common edge to ensure a smooth transition of the surface at the splicing point.If the included angle is greater than 15 degrees, it indicates a significant change in the orientation of the two faces. In this case, it is necessary to backtrack to the low-dimensional data point set and check whether there are any unincluded neighboring points near the common edge of the two faces. If there are missing points, new transition triangle faces are constructed with the missing points as vertices and the endpoints of the common edge of the two faces respectively. The orientation is smoothly connected through the transition faces, eliminating the gap. If there are no missing points, the selection range of neighboring points of one face is adjusted (the K value of K nearest neighbors is temporarily increased to 6), and a new face matching the orientation of the other face is regenerated. Finally, all faces are spliced ​​into a continuous surface without overlap or gaps. This surface is the low-dimensional manifold corresponding to the first sub-region, denoted as the manifold of the first sub-region after reconstruction. Its geometric shape directly reflects the local data structure of the core features of the sub-region. For example, the bulge of the surface corresponds to the peak of the feature parameters (such as the period when the call intensity increases sharply), the depression of the surface corresponds to the trough of the feature parameters (such as the period when the elevator is empty), and the tilt direction of the surface corresponds to the changing trend of the feature parameters (such as the process of the elevator speed increasing with the increase of the call intensity). ;

[0052] For the second sub-region, the multidimensional state vectors corresponding to all key feature points contained within the second sub-region are extracted. Then, the dimensionality of these multidimensional state vectors is reduced to obtain the low-dimensional data point set corresponding to the second sub-region. Specifically, this includes: Extraction and dimensionality reduction of the state vectors for the second sub-region. First, based on the sub-region-key feature point association table, the original operating parameter sequences of all key feature points within the second sub-region are extracted. If the second sub-region contains three feature points: elevator 2 load, 5th floor call intensity, and elevator 2 operating floors, and the time span is still 30 seconds (300 time slices), then the dimension of the multidimensional state vector is 3 × 300 = 900. The vectors are grouped in the order of elevator 2 load (positions 1 to 300), 5th floor call intensity (positions 301 to 600), and elevator 2 operating floors (positions 601 to 900). The process is as follows: Then, standardization is performed, calculating the mean and standard deviation of the parameter sequence for each feature point. The standardized value is then calculated as (original parameter value - mean) ÷ standard deviation. The dimensionality reduction stage still uses the t-SNE algorithm. If the second sub-region is a morning peak scenario (high local data density), the bandwidth parameter is determined according to adaptive rules. This involves calculating the average Euclidean distance between the 10 nearest neighbors for each time slice, taking the median of all average distances, and using 0.8 times the median as the final bandwidth parameter (e.g., if the median is 0.6, the bandwidth parameter is set to 0.48). After calculating high-dimensional similarity using the Gaussian kernel function and low-dimensional similarity using the t-distribution kernel function, the low-dimensional coordinates are optimized with the goal of KL divergence convergence (1000 iterations or 50 consecutive changes < 0.001). Finally, the low-dimensional data point set for the second sub-region is obtained.

[0053] Based on the low-dimensional data point set corresponding to the second sub-region, triangular patches are formed by selecting every three neighboring low-dimensional data points, and the normal vector direction of each triangular patch is calculated. Multiple triangular patches are then stitched together to form a continuous surface, resulting in a low-dimensional manifold representing the local data structure characteristics of the second sub-region, denoted as the reconstructed manifold of the second sub-region. Specifically, the low-dimensional manifold reconstruction of the second sub-region involves using the K-nearest neighbor algorithm for selecting neighboring points from the low-dimensional data point set of the second sub-region. Since the number of feature points in this sub-region is 3 (medium-sized sub-region), the value of K remains constant at 5. After generating triangular patches through Delaunay triangulation, the manifold is constructed based on vertex vectors, and the z-value is calculated using the cross product. The component- and standardized process calculates the unit normal vector of each facet. For example, facet vertices A (1.2, 3.1), B (1.5, 3.3), and C (1.3, 3.5), vectors AB (0.3, 0.2) and AC (0.1, 0.4), the cross product z-component = 0.3 × 0.4 - 0.2 × 0.1 = 0.1, and the standardized normal vector is (0, 0, 1). When facets are spliced, the dot product of the normal vectors of adjacent facets is calculated and converted into an angle. If the included angle is less than 15 degrees, they are spliced ​​directly. If the included angle is greater than 15 degrees, a transition facet is added. The resulting continuous surface is the low-dimensional manifold of the second sub-region, and its bulge may correspond to the scenario where the load and call intensity of elevator 2 increase simultaneously.

[0054] For the third to the last sub-region, the multi-dimensional state vectors corresponding to all key feature points within each sub-region are extracted and dimensionality reduced to obtain a low-dimensional data point set for each sub-region. Based on this low-dimensional data point set, triangular patches are formed by selecting every three neighboring low-dimensional data points, and the normal vector direction of each triangular patch is calculated. Multiple triangular patches are then stitched together to form a continuous surface, resulting in the low-dimensional manifold for each sub-region. This process is repeated until the manifolds of all sub-regions are reconstructed. Specifically, the manifold reconstruction for the third to the last sub-region is performed, and the above process is repeated for subsequent sub-regions. The key is to ensure the manifold reconstruction through three major adaptation points. The structure achieves the following effects: First, the state vector dimension is adapted by dynamically adjusting the dimension by 300 times the number of feature points in the sub-region, and using the statistical values ​​of the sub-region's own feature data during standardization. Second, the dimensionality reduction parameter is fine-tuned by increasing the bandwidth parameter to 1.0 for low-peak sub-regions (sparse data) and preserving correlations by widening the sensitive range of the Gaussian kernel. Third, the triangulation K value is adjusted by reducing the K value to 3 for small sub-regions (feature points ≤ 2) to avoid insufficient neighboring points, and maintaining K=5 for large sub-regions (feature points ≥ 5) to ensure patch coverage. After all sub-regions are processed, a correspondence table is formed between the sub-region number, the low-dimensional manifold, and the core feature association. The geometric shape of each manifold corresponds one-to-one with the elevator operation sub-state of that sub-region.

[0055] This embodiment divides sub-regions based on the topological connectivity of key links. Each sub-region is composed of core feature associations, realizing an analysis mode of core association clustering and sub-situation isolation, avoiding interference from irrelevant features. Local associations of high-dimensional data are preserved through similarity distribution optimization, ensuring the low-dimensional data point set and original features remain intact. A low-dimensional manifold is constructed by piecing together triangular facets, transforming the dynamic associations of high-dimensional features into a visualized continuous surface. Surface bulges correspond to peak values ​​of feature parameters, smooth areas correspond to stable feature states, and changes in the normal vector direction correspond to turning points in feature trends, reducing the understanding cost for maintenance personnel. Differentiated processing is applied to different sub-regions, ensuring that the manifold reconstruction of each sub-region is adapted to its own feature density and association characteristics.

[0056] In a preferred embodiment of the present invention, based on the manifold of each reconstructed sub-region, a corresponding local covariance matrix is ​​generated, and a state coupling degree sequence is calculated, including:

[0057] Based on the coordinate distribution of all low-dimensional data points on the manifold of the first reconstructed subregion, calculate the local covariance matrix corresponding to the first subregion; based on the coordinate distribution of all low-dimensional data points on the manifold of the second reconstructed subregion, calculate the local covariance matrix corresponding to the second subregion. Specifically, the local covariance matrix measures the correlation and discreteness of the low-dimensional data points in the subregion in the x and y dimensions of 2D space. The larger the matrix element value, the more significant the correlation or fluctuation in the corresponding dimension. The specific calculation is based on the coordinates of the low-dimensional data points of the subregion manifold. The process is unified and needs to be adapted to the data characteristics of the subregion. Taking the first and second subregions as examples... Detailed Explanation: Step 1: Each data point in the low-dimensional manifold is a 2D coordinate (x, y). Let the total number of low-dimensional data points in a sub-region be E (E is the same as the number of time slices, i.e., 300). Organize all data points into two sequences according to their coordinate dimensions: an x-dimensional sequence and a y-dimensional sequence. The values ​​in each sequence correspond to the low-dimensional feature projection result of a time slice. Step 2: Calculate the average values ​​of the x-dimensional and y-dimensional coordinates respectively, using them as the central reference for the data distribution. Step 3: For each data point, subtract the mean of the corresponding dimension from the x and y coordinates respectively to obtain the x-bias and y-bias. This step transforms all data points into biases with the mean as the origin. The data is processed to eliminate the interference of absolute values ​​on the correlation calculation; the fourth step is to construct a local covariance matrix, that is, the covariance matrix of 2D data is a 2×2 matrix, the matrix elements reflect the covariance relationship between dimensions, and the calculation logic is the average of the sum of the products of corresponding deviations, the formula is as follows: the element in the first row and first column of the matrix (xx covariance) = the sum of the squares of the x deviations of all data points ÷ (total number of data points - 1), reflecting the degree of fluctuation of the x dimension itself; the element in the first row and second column of the matrix (xy covariance) = the sum of the x deviations × y deviations of all data points ÷ (total number of data points - 1), reflecting the correlation strength between the x and y dimensions; the element in the second row and first column of the matrix (yx covariance) = the sum of the x deviations × y deviations of all data points ÷ (total number of data points - 1), reflecting the correlation strength between the x and y dimensions; the element in the second row and first column of the matrix (yx covariance) = the sum of the squares ... The variance (x) and the xy covariance have the same value because the covariance relationship is symmetrical; the element in the second row and second column of the matrix (yy covariance) = the sum of the squares of the y deviations of all data points ÷ (total number of data points - 1), reflecting the degree of fluctuation in the y dimension itself; the calculation process of the second sub-region is completely the same as that of the first sub-region, only needing to be replaced with its own low-dimensional data point coordinate sequence, and a unique local covariance matrix is ​​obtained based on its data distribution characteristics; for example, the second sub-region is the morning peak scene, the data points are more densely distributed, the sum of the squares of the x and y dimensions is smaller, and the final element values ​​of the covariance matrix will be closer to 0 than the sub-region of the low peak scene, reflecting a smoother feature fluctuation.

[0058] Based on the coordinate distribution of all low-dimensional data points on the manifold of the reconstructed third sub-region, calculate the local covariance matrix corresponding to the third sub-region. For each sub-region from the fourth to the last sub-region after reconstruction, calculate the local covariance matrix of the corresponding sub-region based on the coordinate distribution of all low-dimensional data points on the manifold of the corresponding sub-region. Specifically, for the third to the last sub-region, perform the standardized process of data preparation, dimension mean calculation, deviation calculation, and covariance matrix construction in sequence. Two adaptation points need to be noted during the operation: first, data volume adaptation. If the total number of data points after screening in a certain sub-region due to weak feature correlation is less than 300 (such as some sub-regions), the data volume should be adapted accordingly. (1) Outlier data points were removed from the low-peak sub-regions. When calculating the mean and covariance, the actual number of remaining data points should be used as the standard. The total number of data points in the formula should be replaced with the actual number of valid data points to ensure that the calculation results are not affected by outliers. (2) The meaning of dimensions should be adapted. The high-dimensional features corresponding to the low-dimensional manifold x and y dimensions of different sub-regions are different (e.g., in some sub-regions, x corresponds to elevator speed and y corresponds to call intensity, while in others, x corresponds to load and y corresponds to energy consumption). However, the calculation logic of the covariance matrix is ​​not related to the physical meaning represented by the dimensions. It only needs to be processed uniformly according to the coordinate sequence. After the calculation of all sub-regions is completed in sequence, a one-to-one correspondence between the sub-region number and the local covariance matrix is ​​formed.

[0059] Eigenvalue decomposition is performed on the local covariance matrix corresponding to the first sub-region, the local covariance matrix corresponding to the second sub-region, and so on up to the local covariance matrix corresponding to the last sub-region. The largest eigenvalue is extracted from each local covariance matrix as the principal eigenvalue. Specifically, the process includes: First, for the local covariance matrix (denoted as matrix C) of each sub-region, the eigenvalues ​​λ and eigenvectors v satisfying C×v=λ×v are solved using a linear algebraic decomposition algorithm (such as QR decomposition). Here, λ is the eigenvalue and v is the corresponding eigenvector. Since the covariance matrix is ​​a symmetric matrix, the decomposition will yield the same number of eigenvalues ​​as the matrix dimension (2), and all eigenvalues ​​are non-negative. Second, the two eigenvalues ​​obtained from the decomposition are sorted by their numerical values, and the maximum value is taken as the principal eigenvalue of the sub-region. The physical meaning of the principal eigenvalue is the dominant strength of the association between the core features of the sub-region. The larger the value, the more concentrated the data points on the low-dimensional manifold are in the direction corresponding to the eigenvalue, that is, the more prominent the coupling relationship of the features within the sub-region. For example, the association between elevator speed and call intensity plays a dominant role in this sub-region.

[0060] All principal feature values ​​are arranged in the order of the first sub-region, the second sub-region, and so on up to the last sub-region, forming a state coupling degree sequence. Specifically, this includes arranging them in the order of sub-region 1, sub-region 2, sub-region 3, and so on up to the last sub-region, ensuring that the sequence order is consistent with the logical order of the physical operation scenarios corresponding to the sub-regions, such as the call response sub-region and the energy consumption control sub-region. The principal feature values ​​of each sub-region are then arranged in the above order to form a one-dimensional numerical sequence, named the state coupling degree sequence. The core value of this sequence is that it transforms the high-dimensional and complex sub-region feature associations into ordered quantitative indicators. Each value in the sequence represents the feature coupling strength of the corresponding sub-region. Subsequently, by analyzing the fluctuations, peaks, and other characteristics of the sequence, the overall operation status of the elevator can be analyzed (for example, a sudden increase in the value at a certain position in the sequence indicates that the feature association of the corresponding sub-region has entered a strong coupling state, which may indicate a bottleneck in operating efficiency).

[0061] In this embodiment, the correlation of sub-region features is transformed into specific numerical values ​​through the covariance matrix. The extracted principal eigenvalues ​​further focus on the core correlation strength, transforming the feature coupling characteristics of each sub-region from qualitative descriptions to quantitative indicators, thus avoiding the bias of subjective judgment. By extracting principal eigenvalues ​​through eigenvalue decomposition, interference from secondary variation directions is eliminated, and the principal eigenvalues ​​are orderly integrated into a sequence, simplifying the global analysis from multi-matrix parallel processing to single-sequence feature analysis, thereby improving analysis efficiency. The arrangement order of the state coupling degree sequence is consistent with the logic of the physical operation scenario corresponding to the sub-region. The temporal and spatial characteristics of the sequence are combined to intuitively reflect the dynamic changes in the elevator's operating status.

[0062] In a preferred embodiment of the present invention, a state correction factor characterizing the overall operating status is extracted based on the state coupling degree sequence, and the state correction factor is fused with operating parameters to determine the elevator operating status judgment result, including:

[0063] Statistical analysis is performed on the state coupling degree sequence to extract its mean, variance, and skewness. These features are then combined to form a state correction factor characterizing the overall operational status. Specifically, the state correction factor is a core composite indicator characterizing the elevator's overall operational status, composed of the mean, variance, and skewness of the state coupling degree sequence. The mean reflects the overall coupling strength, the variance reflects coupling fluctuations, and the skewness reflects the asymmetry of the coupling distribution. The combination of these three features provides a comprehensive quantification of the overall situation. The specific process is as follows: First, the state coupling degree sequence is a constructed one-dimensional numerical sequence, containing L principal feature values ​​(L equals the total number of sub-regions), denoted as sequence S = [s1, s2, s3, ..., sL], where each s represents the feature coupling strength of the corresponding sub-region. Second, the mean of the sequence is calculated. The mean is the average level of all values ​​in the sequence, reflecting... The first step is to calculate the overall strength of the coupling characteristics of each sub-zone of the elevator. The second step is to calculate the variance of the sequence. The variance reflects the degree of fluctuation of the principal characteristic values ​​in the sequence and reflects the difference in the coupling strength of each sub-zone. The greater the fluctuation, the more prominent the imbalance of the sub-zone. The calculation steps are as follows: first, calculate the deviation, subtract the mean of the sequence from each principal characteristic value to obtain the deviation sequence, and then calculate the sum of squared deviations. The final variance = sum of squared deviations ÷ (total number of sub-zones - 1). The third step is to calculate the skewness of the sequence. Skewness reflects the asymmetric direction of the sequence distribution. A positive value indicates that the larger principal characteristic values ​​in the sequence are more concentrated (some sub-zones are extremely strongly coupled), and a negative value indicates that the smaller principal characteristic values ​​are more concentrated (some sub-zones are extremely weakly coupled). Skewness = (total number of sub-zones ÷ [(total number of sub-zones - 1) × (total number of sub-zones - 2)]) × (sum of cubed deviations ÷ cubed standard deviation). The fifth step is to combine the calculated mean, variance, and skewness in the order of mean, variance, and skewness to form a three-dimensional feature vector, which is the state correction factor. This vector contains three types of information: the strength level of global coupling, fluctuation characteristics, and distribution pattern.

[0064] The state correction factor is fused with the call signal, real-time load, and real-time location data from the original operating parameter set to generate fused operating characteristic information. Specifically, this involves: First, extracting three types of core real-time data from the original operating parameter set: call signal (intensity of call requests at each floor, in times / second), real-time load (current passenger weight in the elevator, in kilograms), and real-time location (the elevator's current floor and precise location, in meters, with the ground floor as 0). Standardization is then applied to these three types of data to eliminate dimensional differences. The standardized call signal value is calculated as: (Current call intensity - Minimum historical call intensity of this elevator) ÷ (Most historical call intensity of this elevator) The maximum value of the degree - the minimum value of the elevator's historical call intensity), mapped to the interval [0, 1]; the real-time load standardized value = (current load - minimum value of elevator rated load 0) ÷ (elevator rated load - 0), mapped to the interval [0, 1] (e.g., rated load 1000 kg, current load 500 kg, standardized value is 0.5); the real-time position standardized value = (current position - lowest position of elevator operation 0) ÷ (highest position of elevator operation - 0), mapped to the interval [0, 1] (e.g., elevator runs up to the 10th floor, building height 3 meters, highest position 30 meters, currently at the 5th floor 15 meters, standardized value is 0.5); the second step is to make the correction factor consistent with the real-time parameters. The numerical range of the numbers is unified, and the three-dimensional state correction factor is standardized. The processing method is consistent with the real-time parameters. The historical value ranges of the mean, variance, and skewness are statistically analyzed (based on the elevator's operating data over the past 30 days). Each dimension is mapped to the [0, 1] interval using the formula (current value - minimum value) ÷ (maximum value - minimum value). In the third step, the fusion weights are set based on the elevator's operating characteristics. The state correction factor reflects the global situation and has a weight of 40%. The three types of real-time parameters directly reflect the current state of the elevator and have a total weight of 60%. Among them, the call signal is most closely related to the operating status judgment and has a weight of 30%. The real-time load and real-time position are each set to 15%. The fusion formula is... The fusion feature value is calculated as follows: (mean after standardization of correction factor × 0.4 × (1 / 3)) + (variance after standardization of correction factor × 0.4 × (1 / 3)) + (skewness after standardization of correction factor × 0.4 × (1 / 3)) + (standardized call signal × 0.3) + (standardized real-time load × 0.15) + (standardized real-time position × 0.15). Each of the three dimensions of the correction factor accounts for 1 / 3 of the total weight (i.e., 0.4 ÷ 3 ≈ 0.133), ensuring that the three-dimensional information participates in the fusion in a balanced manner. Substituting the standardized data into the formula, the fusion feature value corresponding to each time slice is obtained. The sum of the fusion feature values ​​of all time slices is the fused operational feature information.

[0065] The analysis checks whether the fused operational feature information contains a preset emergency evacuation feature pattern. The emergency evacuation feature pattern is defined as all floors generating continuous downward call signals within a continuous time window, with signal strength exceeding a preset threshold. If the fused operational feature information contains an emergency evacuation feature pattern, the elevator is determined to be in an emergency evacuation state; if the fused operational feature information does not contain an emergency evacuation feature pattern, the elevator is determined to be in a normal operating state. The elevator operating state determination result is obtained, specifically including: First, combining building safety regulations and elevator operating characteristics, defining the emergency evacuation feature pattern as all floors... The elevator continuously sends a downward call signal within a 10-second time window, with the signal strength exceeding a preset threshold. Key parameters are quantified as follows: the continuous time window is set to 10 seconds, corresponding to 100 time slices; "all floors" refers to all floors served by the elevator. For example, if the building has 10 floors, floors 1 to 10 must meet the condition, excluding floors that are not open or disabled due to malfunction; the continuous downward call signal means that each floor has a downward call request in every time slice within the 10-second time window, without interruption; the preset signal strength threshold is determined through historical emergency evacuation drill data and is set to three times the average call strength during off-peak hours for the building, ensuring differentiation from normal downward calls. The first step involves identifying both standardization and emergency evacuation call signals. The second step involves separating the down-going call signal strength sequence for each floor from the fused operational characteristic information (already standardized, but requiring inverse standardization to restore the actual strength value; the inverse standardization formula is: actual strength value = standardized value × (historical maximum value - historical minimum value) + historical minimum value). The third step involves verifying the signal strength in the order of time window continuity, floor coverage integrity, and signal strength compliance. During time window verification, a 10-second continuous time window is slid-triggered to check for a continuous down-going call signal within each time window. During floor coverage verification, for windows that meet the time continuity requirement, the signal strength is checked... Check if all service floors have downlink call signals; during the intensity verification, for windows that meet the first two conditions, check if the actual intensity of the call signals on each floor exceeds a preset threshold, which is three times the average call intensity during the off-peak hours of the building; fourth, output the status judgment result. If there is a time window that meets all three conditions simultaneously, the elevator is determined to be in an emergency evacuation state; if after traversing all time windows, no window meets all the conditions, the elevator is determined to be in a normal operation state; the judgment result needs to be updated in real time, and the matching verification is re-executed every 2 seconds to ensure a rapid response to status changes.

[0066] This embodiment standardizes and unifies the value ranges of various parameters, and sets scientific weights based on elevator operating characteristics to ensure that the fusion process is quantifiable and reproducible. The quantitative definition of the emergency evacuation mode avoids the subjectivity of experience-based judgment and improves the reliability of the judgment. Through triple verification of time window, floor coverage, and intensity threshold, it captures typical elevator call characteristics of emergency evacuation, and can quickly distinguish between normal downward elevator calls and emergency evacuation elevator calls, thus strengthening the safety assurance capability of elevator operation. The fusion of the global situation reflected by the state correction factor and real-time parameters can quickly capture changes in the operating state. The judgment results can be directly used as the basis for adjusting the scheduling strategy of the elevator group control system, making the scheduling scheme more in line with real-time operating needs and improving elevator operating efficiency and service quality.

[0067] In a preferred embodiment of the present invention, based on the elevator operating status judgment result, a first elevator scheduling instruction or a second elevator scheduling instruction is obtained through a first scheduling model or a second scheduling model in a pre-trained dual-model library; executing the first elevator scheduling instruction or the second elevator scheduling instruction to control the elevator group includes:

[0068] If it is determined that the elevator is in normal operation, the fused operation feature information is input into the first scheduling model in the pre-trained dual-model library to generate the first elevator scheduling instruction. Specifically, the dual-model library is a set of dedicated models trained and validated using historical data. The first scheduling model adopts a composite structure of gradient boosting tree and neural network. The gradient boosting tree is responsible for capturing the linear relationship between features and scheduling indicators, such as the relationship between call intensity and elevator response time. The neural network is responsible for mining nonlinear relationships, such as the coupling relationship between elevator location, load and energy consumption. The model input consists of fused operational feature information, including standardized state correction factors and real-time parameters. The output is a set of scheduling instructions containing parameters such as elevator number, target floor, and operating speed. The construction and training process of the first scheduling model must follow the logic of data preparation, structural design, and iterative optimization. In the model construction stage, the elevator's normal operation data over the past 12 months is collected, covering fused feature information from 7:00 to 23:00 daily, including five feature groups such as global status, call status, and load. This data is used as input data. Simultaneously, the scheduling effect indicators corresponding to the scheduling results are labeled as tag data, specifically including three core indicators: elevator response time, average passenger waiting time, and energy consumption per unit time. After completing data preparation, the process continues... In the model structure design phase, the gradient boosting tree consists of 100 CART decision trees, each with a depth limited to 6 layers to avoid overfitting. The input layer receives 32-dimensional feature vectors from 5 feature groups. The neural network adopts a 3-layer fully connected structure. The number of neurons in the input layer is consistent with the output dimension of the gradient boosting tree, set to 16 dimensions. The hidden layer is divided into two layers, with the first layer configured with 32 neurons and the second layer configured with 16 neurons. The output layer has 8 neurons, corresponding to the predicted values ​​of 8 basic scheduling parameters. The final step in the structure design is the fusion layer design, which weights and sums the output of the gradient boosting tree and the output of the neural network with a weight of 4:6. In this way, the linear and non-linear feature mapping results are integrated to obtain the final feature mapping output.

[0069] After the model is built, the training phase begins. First, the collected 120,000 data points are divided into training, validation, and test sets in a 7:2:1 ratio. Input features are standardized using the formula: Standardized value = (Original value - Feature mean) / Feature standard deviation. Labels are normalized to the [0, 1] interval. After data preprocessing, a multi-objective weighted loss function is used to synthesize the prediction errors of three scheduling performance indicators. The total loss is calculated as: 0.4 × |Predicted response time - Actual response time| + 0.4 × |Predicted waiting time|. The formula is: Measured Value - Actual Value | + 0.2 × |Energy Consumption Prediction - Actual Value|. 0.4 represents the weight of response time and waiting time. Under normal operating conditions, the core objective of elevator scheduling is to improve operational efficiency and reduce passenger waiting time; these two indicators directly reflect efficiency levels, hence they are given a higher weight. 0.2 represents the weight of energy consumption. Energy optimization is a secondary objective of normal scheduling, taking into account energy-saving needs while ensuring efficiency; therefore, its weight is relatively low. The Adam optimizer is used, with an initial learning rate of 0.001, which decays to 0 every 100 iterations.9. The training rounds are set to 500 rounds, with an early stop mechanism. Training stops when the total loss of the validation set no longer decreases for 20 consecutive rounds to prevent overfitting. After training, a model validation and optimization phase is required. The model performance is validated using a test set. If a certain type of scheduling scenario, such as the morning rush hour, is found to have a high error, 30,000 additional specific data points are added for fine-tuning. The second step is to structurally decompose the fused operational feature information according to the input dimension requirements of the first scheduling model, ultimately forming five feature groups with clearly defined functions. Specifically, the global situation group directly corresponds to the mean, variance, and skewness of the state correction factor. The elevator call feature group specifically collects the elevator call signal strength of each floor and the duration data of each signal. The load feature group... The feature group covers real-time load data for all elevators, as well as the load rate calculated from the real-time load. The location feature group records the current real-time position of each elevator and the direction of travel it is currently executing. The historical dispatch group stores the execution feedback results of all dispatch instructions within the last 5 minutes, including instruction completion status and execution deviation. Each feature group needs to undergo normalization recalibration according to the weight ratio set during pre-training. The specific formula for recalibration is: calibrated feature value = original fused feature value × training weight ratio of the feature group. The weight ratios of the five feature groups have been clearly defined according to the needs of normal dispatching: global situation group 20%, call feature group 30%, load feature group 15%, location feature group 20%, and historical dispatch group 10%. 5%, with a total weight of 100%, this weight allocation method is precisely to highlight the core role of call characteristics (directly related to passenger demand) and the overall situation (reflecting the overall operational status) in dispatching decisions; the third step is to enter the inference and calculation stage of the first dispatching model. The entire inference process is divided into two closely linked stages: linear calculation and nonlinear calculation. The linear calculation stage is handled by gradient boosting trees, and the core is to perform a weighted summation of the calibrated features through 100 basic CART decision trees; since the accuracy of each decision tree on the validation set is different, its contribution to the final result also varies. Therefore, it is necessary to assign an accuracy weight to each tree. The calculation method for this weight is: the accuracy of a single decision tree on the validation set divided by the accuracy of 100 decision trees. The sum of the accuracies, the final linear calculation result, is the sum of the product of the output value of each decision tree and its corresponding accuracy weight, and then the sum of all the product results. The nonlinear calculation stage is completed by a 3-layer fully connected neural network. First, the result of the linear calculation is input into the network, and feature mapping is performed through the ReLU activation function. The mapping formula is: activated value = max(0, linear calculation result × weight matrix + bias term). The weight matrix is ​​determined by iterative optimization during model training. That is, combined with the 3-layer fully connected structure of this model (16-dimensional input layer, 32-dimensional first hidden layer, 16-dimensional second hidden layer, and 8-dimensional output layer), it is assumed that the weight matrix from the input layer to the first hidden layer takes values ​​in the range of [-0.12, 0].12] The weight matrix from the first hidden layer to the second hidden layer has a value range of [-0.15, 0.15], and the weight matrix from the second hidden layer to the output layer has a value range of [-0.2, 0.2]. The bias term is set to 0.01 to prevent the gradient vanishing problem during feature mapping. After nonlinear calculation, the neural network will output a set of numerical results. Finally, these values ​​need to be converted into specific scheduling parameters that the elevator can execute. For example, when the output value corresponding to the running priority is 0.8, it will be converted into the highest priority scheduling instruction. When the target floor outputs a certain value, the matching is completed through the preset mapping rule. This mapping rule is designed based on the physical characteristic that the elevator floors are integers. Specifically, the integer part of the output value of the neural network is taken as the target floor number. If the decimal part of the value is greater than or equal to 0.5, it is rounded up. If the decimal part is less than 0.5, it is rounded down. At the same time, it ensures that the minimum value is 0.5. The final floor number is within the floor range served by the elevator (excluding floors that are not open or disabled due to malfunction). For example, the decimal part of a value of 5.2 is 0.2, which is less than 0.5, corresponding to the actual 5th floor; the decimal part of a value of 5.6 is 0.6, which is greater than or equal to 0.5, corresponding to the actual 6th floor; and the decimal part of a value of 4.1 is 0.1, which is less than 0.5, corresponding to the actual 4th floor. This explicit numerical discretization rule ensures precise matching between the output value and the specific floor number. The fourth step involves integrating the scheduling parameters output by the model and generating the first elevator scheduling instruction in a fixed format: elevator number, scheduling type, core parameters, and execution time limit. All elevator scheduling instructions must meet the conflict-free constraint, meaning that no two elevators will respond to the same floor call simultaneously within the same time period. If there is a conflict in the model output, it is adjusted according to the call strength priority principle, prioritizing the scheduling instruction corresponding to the floor with the higher call signal strength to ensure reasonable resource allocation.

[0070] If the elevator is determined to be in an emergency evacuation state, the fused operational feature information is input into the second scheduling model in the pre-trained dual-model library to generate a second elevator scheduling instruction. Specifically, this includes: First, defining the core characteristics of the second scheduling model. The model is built based on a deep Q-network, and the training data covers simulated evacuation data and historical drill data for various emergency scenarios such as building fires and earthquakes. The input features, based on the fused operational feature information, include three emergency scenario-specific features: the location of safe passages, estimated floor personnel density, and emergency power status. The output instruction prioritizes three core requirements: downward evacuation, empty-load response, and nearest stop. The construction and training process of the second scheduling model is centered on reinforcement learning agent-environment interaction, continuously optimizing the decision-making strategy through trial and error. In the model construction phase, environmental modeling is first performed to build an elevator emergency evacuation simulation environment. This environment must include elevator operation physical rules such as speed limits, load limits, etc. The system is designed with strict constraints, including building structure information such as floor distribution and safety passage locations, as well as personnel flow models such as evacuation speed and elevator call behavior in emergencies. The environmental status is updated in real time with dispatch actions. The intelligent agent design uses a deep Q-network as its core. The input layer receives a 42-dimensional feature vector, which is composed of 32-dimensional original fused features and 10-dimensional emergency-specific features. The hidden layer adopts a structure of two convolutional layers and two fully connected layers. The convolutional layers are configured with 64 and 32 convolutional kernels, respectively, and the fully connected layers are configured with 64 and 32 neurons, respectively. The number of neurons in the output layer is consistent with the dimension of the action space, which is set to 12 dimensions, corresponding to 12 types of emergency dispatch actions. The definition of state and action is a key part of the construction. The state space is a set of real-time features in emergency scenarios, including elevator status, personnel distribution, and environmental risks. The action space covers 12 specific dispatch actions, such as immediately descending to the first floor and responding to elevator calls after emptying the car, to ensure coverage of various emergency response needs.

[0071] The model training phase begins with data preparation, generating 50,000 simulated emergency evacuation data points across multiple scenarios, including fire and earthquake scenarios. Simultaneously, 10,000 real-world data points from elevator emergency drills are collected to form the training dataset, enhancing the model's generalization ability. After data preparation, an experience playback mechanism needs to be constructed. Specifically, a pool of 100,000 experience points is built. When the agent (deep Q-network) executes a scheduling action in the emergency evacuation simulation environment, it simultaneously acquires the current environmental state, such as the density of people on each floor, elevator location / load, safety passage occupancy, executed scheduling actions (e.g., elevator stopping at the nearest stop, priority downward movement), the corresponding immediate reward, and the updated environmental state after the action is triggered. (Such as changes in floor occupancy density and adjustments to elevator trajectories), these four elements are continuously stored in the experience pool as a complete quadruple. During each model training iteration, 256 data points are randomly selected from the experience pool as batch samples. The core representation of the reward function is the total reward calculation method, and its detailed design is crucial for guiding the model towards efficient evacuation and a safety net. After supplementing intermediate reward nodes, the total reward (i.e., the core formula of the reward function) is determined as: 0.5 × (number of evacuees / total number of people on the floor) + 0.3 × (1 - occupancy density / safety density threshold) - 0.2 × overload risk, where 0.5 is the weight of the evacuation efficiency index, ensuring that in any emergency scenario, people are evacuated to a safe location as quickly as possible. The entire area is the primary dispatch target, and the ratio of evacuated persons to the total number of people on that floor directly quantifies the progress of evacuation on a single floor, thus giving it the highest weight. 0.3 is the weight for the personnel gathering risk indicator, a core safety objective second only to evacuation efficiency. 0.2 is the weight for the overload risk indicator; elevator overload can directly lead to equipment failure and even fatal accidents such as falls, but compared to the group safety risks posed by unevacuated personnel gathering, the impact of a single elevator overload risk is smaller, therefore its weight is relatively low. The safety density threshold needs to be set based on the building characteristics served by the elevator. For places with frequent personnel flow and high probability of instantaneous gathering, such as office buildings and shopping malls, the safety density threshold is set at 0.8 people / square meter. Strict thresholds are used to control the risk of gathering. For places with fixed population distribution and low peak gathering times, such as residential buildings, the safe density threshold is relaxed to 0.6 people / square meter, balancing safety and scheduling flexibility. When the actual population density on a certain floor (calculated as the number of people on the floor / floor area) exceeds the safe density threshold for the corresponding building type, the calculation result of 1 - population density / safe density threshold will decrease simultaneously, directly leading to a decrease in the model's total reward value. This quantitative feedback will force the model to actively adjust its scheduling strategy, prioritizing actions such as dispatching empty elevators to high-density floors and extending elevator door opening time to accelerate the diversion of people. Through the linkage of weight allocation and quantitative indicators, a dynamic balance between evacuation efficiency and safety risk is achieved.

[0072] The training process needs to balance exploration and utilization. In the initial stage, an ε-greedy strategy is adopted, with an initial exploration probability ε of 0.9, decreasing by 0.05 every 100 iterations, down to a minimum of 0.1. An RMSprop optimizer is used with a learning rate of 0.002, and 1000 training rounds are conducted, each containing 1000 environmental interactions. Every 200 rounds during training, real emergency drill data is used for verification. If the evacuation time error exceeds 10%, the environmental model parameters are optimized for that scenario, and the training is repeated. In the second step, newly added emergency-specific features are quantified and weighted. The estimated personnel density is calculated by multiplying the elevator call signal strength by the floor area by the personnel density coefficient. The personnel density coefficient is set according to building type: 0.15 people / square meter for office buildings and 0.1 people / square meter for residential buildings. The distance between the safety passage and the elevator is calculated as a straight-line distance and serves as the core basis for priority stopping weights; the closer the distance, the higher the weight. Priority stopping weight = 1 ÷ (distance between safety passage and elevator + 0.1). By adding 0.1, the weight becomes infinite when the distance is 0. In the feature fusion stage, emergency-specific features are fused with the original fusion features at a ratio of 60% for emergency features and 40% for the original features. By increasing the weight of emergency features, the influence of key information in emergency scenarios is strengthened, ensuring that the model focuses on the core needs of emergency response. In the third step, the model determines the optimal instruction by calculating the Q-value (action value) of each potential dispatch action. The Q-value = immediate reward + discount factor × expected future reward. The expected future reward is a specific value calculated by the model based on the current environmental state, such as the distribution of people on each floor, the elevator operation status, post-training parameters, and the typical emergency evacuation process (up to 20 core evacuation actions). The calculation logic is as follows: first, according to the Q-value table trained by the model, the estimated reward corresponding to each possible environmental state after executing the current action is queried (the calculation logic of this estimated reward is consistent with the total reward function defined above, and the estimated reward for each step is derived based on the Q-value of the optimal action in that state); then, according to the formula... Calculation, where For the first The estimated reward for subsequent actions. Discount factor ( =0.9), which is the estimated reward of the subsequent action in step 1 multiplied by Step 2 multiply by This process continues until step 20. Finally, the specific value of the expected future reward is obtained through the summation formula. The immediate reward is calculated as (number of evacuees ÷ total number of people on the floor) × 0.7 - (elevator overload risk × 0.3). The elevator overload risk is calculated as (real-time load - rated load × 0.8) ÷ rated load. If the result is negative, it is set to 0. The discount factor is set to 0.9 to ensure that the model takes into account both the current and subsequent evacuation effects. During inference, the action with the largest Q value is selected as the scheduling instruction. A typical instruction is to immediately empty the elevator car and descend to the vicinity of the safety passage on the 1st floor. In the fourth step, the second elevator scheduling instruction is generated. The instruction format adds safety prompt information and emergency linkage signal fields to the first scheduling instruction. The core parameters must meet the emergency specifications. For example, the upper limit of the running speed is increased to 2 meters / second but does not exceed the rated speed of the elevator. The door opening delay is extended to 5 seconds. The preferred stopping floors are the 1st floor, 3rd floor, etc., which are close to the safety passage. At the same time, the instruction must include a backup plan for emergency response.

[0073] The first or second elevator scheduling command is sent to the elevator group control execution unit, which then controls the elevator group according to the command. Specifically, this involves: converting the first or second elevator scheduling command from the model output format into a control signal format recognizable by the elevator group control execution unit; performing dual checks after conversion to ensure command validity; syntax checking to verify if the command format conforms to the group control unit's communication protocol to avoid parsing errors; and logic checking to ensure the command does not exceed the elevator's physical operating limits, such as speed not exceeding the rated value or floors within the service range. If the check fails, the command is regenerated from the model, forming the first fault tolerance barrier. Commands are sent using a dual-link approach: industrial Ethernet and wireless backup communication. The primary link uses industrial Ethernet to ensure a transmission delay of ≤100 milliseconds; the backup link uses 4G industrial-grade wireless communication. If the primary link is interrupted, it automatically switches to the backup link within 50 milliseconds. This dual-link redundancy ensures uninterrupted command transmission. Commands are sent according to elevator numbering. The commands are sent sequentially, with each elevator's command accompanied by a unique identifier to effectively avoid command confusion. After receiving the commands, the elevator group control execution unit first parses the control parameters of each elevator, and then sends control signals to the elevator's drive system, door system, and display system via the PLC (Programmable Logic Controller). The drive system adjusts the motor speed according to the running speed and target floor parameters, i.e., motor speed = running speed × reduction ratio ÷ (π × elevator traction sheave diameter), where the reduction ratio adopts the commonly used standard value of 30 in the elevator industry (adapting to the transmission characteristics of mainstream elevator traction systems). The door system precisely controls the timing and duration of door opening and closing based on the door opening delay parameters. The display system synchronously updates the elevator's operating status and safety prompts. During execution, the group control execution unit collects elevator operating data in real time, including key parameters such as position, speed, and load, and feeds back the execution status to the dispatching system every 500 milliseconds. If a command execution deviation occurs, the dispatching system regenerates and issues a correction command, forming a closed-loop control of issuance, execution, feedback, and correction to ensure that the dispatching effect meets expectations.

[0074] This embodiment, through a dual-model library design, ensures that the scheduling scheme is highly matched with the elevator's operating status, solving the limitation of single-model scenario adaptation. It integrates global situational features, real-time operational features, and historical scheduling features into the model, and uses linear and nonlinear calculations to uncover deep correlations between features, avoiding the one-sidedness of scheduling based on a single parameter. Dual communication links prevent interruptions in command transmission, double verification ensures command accuracy, and real-time feedback and correction mechanisms address execution deviations, reducing risks such as command transmission failures and execution errors. The second scheduling model generates targeted commands through feature enhancement and reward mechanisms, enabling rapid issuance and execution, thus improving the elevator's emergency response capabilities and safety assurance level.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An elevator energy-saving operation control system based on a large model, characterized in that, include: The data acquisition module is used to collect the timestamps and floor numbers of call signals on each floor, the real-time load values ​​of each car, and the real-time position of each car in the hoistway in real time, to obtain the original operating parameter set. The data collected within the same time slice in the original operating parameter set are aligned according to three dimensions: call signal, real-time load, and real-time position, and then vectorized and combined to form a multi-dimensional state vector. The multi-dimensional state vector is then processed by a preset spatiotemporal mapping relationship. Based on the correspondence rules between floors and time, the spatiotemporal mapping relationship extracts trajectory data describing the car displacement process, intensity data describing the cumulative status of call signals on each floor, and collaborative data describing the relative positional relationships of multiple cars from the multi-dimensional state vector. The trajectory data, intensity data, and collaborative data are then normalized and concatenated to form a fused feature set. The structure construction module is used to map each feature item in the fused feature set to the corresponding coordinate axis in a predefined state reference system according to the type of each feature item; determine the position of each feature item on the mapped coordinate axis according to the value of each feature item, and obtain the coordinate set of all feature items; analyze the coordinate set using a density-based clustering algorithm to identify regions with dense distribution of coordinate points; in each densely distributed region, select the coordinate points located at the center of the region and whose corresponding feature item importance weight is higher than a preset importance weight threshold as key feature points; connect any two key feature points to form an edge, and calculate the Pearson correlation coefficient of the original running parameter set corresponding to the two key feature points connected by each edge within a set time length; retain the edges with Pearson correlation coefficients higher than a preset connection threshold, and the retained edges and the key feature points connected by the retained edges together constitute a closed multidimensional feature structure; The link identification module is used to obtain the original set of operating parameters corresponding to the two key feature points connected by each edge in a closed multidimensional feature structure, calculate the mutual information value of the original set of operating parameters corresponding to the two key feature points within a set time span; divide the time span into multiple consecutive sliding windows, and calculate the Pearson correlation coefficient of the original set of operating parameters corresponding to the two key feature points within each sliding window; calculate the average value of the Pearson correlation coefficients within all sliding windows, and calculate the variance of the Pearson correlation coefficients within all sliding windows, using the average value and variance as the measure of dynamic correlation; normalize the mutual information value and the measure of dynamic correlation respectively to obtain the normalized mutual information value and the normalized measure of dynamic correlation; and weight the normalized mutual information value and the normalized measure of dynamic correlation according to preset weights to obtain the comprehensive correlation strength value of the corresponding edge. Based on the comprehensive correlation strength value, each edge in the closed multidimensional feature structure is weighted and assigned a weight value. All edges in the closed multidimensional feature structure are sorted from largest to smallest according to their weight values. The edges ranked in the top N percent after sorting are selected as critical links, where N is a preset integer percentage threshold. The manifold reconstruction module is used to divide the multidimensional feature structure into several sub-regions according to the critical link, and to reconstruct the manifold of the state vector field for each sub-region to obtain the reconstructed manifold of each sub-region. The coupling analysis module is used to generate the corresponding local covariance matrix based on the manifold of each sub-region after reconstruction, and to calculate the state coupling degree sequence. The factor generation module is used to extract state correction factors that characterize the overall operating status based on the state coupling degree sequence, and to fuse the state correction factors with the operating parameters to determine the elevator operating status judgment result. The scheduling decision module is used to determine the elevator's operating status and obtain a first elevator scheduling instruction or a second elevator scheduling instruction through a first scheduling model or a second scheduling model in a pre-trained dual-model library; and execute the first elevator scheduling instruction or the second elevator scheduling instruction to control the elevator group.

2. The elevator energy-saving operation control system based on a large model according to claim 1, characterized in that, Based on the critical link, the multidimensional feature structure is divided into several sub-regions, and the manifold of the state vector field is reconstructed for each sub-region to obtain the reconstructed manifold of each sub-region, including: Based on the critical links and the weighted values ​​assigned to each edge in the critical links, the key feature points connected by the critical links in the closed multidimensional feature structure are clustered and divided into several topologically connected sub-regions. For the first sub-region, extract the multi-dimensional state vectors corresponding to all key feature points contained in the first sub-region, and perform dimensionality reduction processing on the multi-dimensional state vectors corresponding to all key feature points contained in the first sub-region to obtain the low-dimensional data point set corresponding to the first sub-region. Based on the low-dimensional data point set corresponding to the first sub-region, a triangular patch is formed by selecting every three neighboring low-dimensional data points, and the normal vector direction of each triangular patch is calculated; multiple triangular patches are spliced ​​into a continuous surface to obtain the low-dimensional manifold corresponding to the first sub-region, which represents the local data structure characteristics of the first sub-region, and is denoted as the manifold of the reconstructed first sub-region. For the second sub-region, extract the multi-dimensional state vectors corresponding to all key feature points contained in the second sub-region, and perform dimensionality reduction processing on the multi-dimensional state vectors corresponding to all key feature points contained in the second sub-region to obtain the low-dimensional data point set corresponding to the second sub-region. Based on the low-dimensional data point set corresponding to the second sub-region, a triangular patch is formed by selecting every three neighboring low-dimensional data points, and the normal vector direction of each triangular patch is calculated; multiple triangular patches are spliced ​​into a continuous surface to obtain the low-dimensional manifold corresponding to the second sub-region, which represents the local data structure characteristics of the second sub-region, and is denoted as the reconstructed manifold of the second sub-region. For the third to the last sub-region, the multidimensional state vectors corresponding to all key feature points contained in each sub-region are extracted and dimensionality reduction is performed to obtain the low-dimensional data point set corresponding to each sub-region. Based on the low-dimensional data point set corresponding to each sub-region, a triangular patch is formed by selecting every three neighboring low-dimensional data points, and the normal vector direction of each triangular patch is calculated. Multiple triangular patches are spliced ​​into a continuous surface to obtain the low-dimensional manifold corresponding to each sub-region, until the manifold of all sub-regions after reconstruction is obtained.

3. The elevator energy-saving operation control system based on a large model according to claim 2, characterized in that, Based on the reconstructed manifold of each sub-region, the corresponding local covariance matrix is ​​generated, and the state coupling degree sequence is calculated, including: Based on the coordinate distribution of all low-dimensional data points on the manifold of the first sub-region after reconstruction, calculate the local covariance matrix corresponding to the first sub-region; based on the coordinate distribution of all low-dimensional data points on the manifold of the second sub-region after reconstruction, calculate the local covariance matrix corresponding to the second sub-region. Based on the coordinate distribution of all low-dimensional data points on the manifold of the third sub-region after reconstruction, calculate the local covariance matrix corresponding to the third sub-region; for each sub-region from the fourth to the last sub-region after reconstruction, calculate the local covariance matrix of the corresponding sub-region based on the coordinate distribution of all low-dimensional data points on the manifold of the corresponding sub-region. Eigenvalue decomposition is performed on the local covariance matrix corresponding to the first sub-region, the local covariance matrix corresponding to the second sub-region, and so on up to the local covariance matrix corresponding to the last sub-region. The largest eigenvalue is extracted from each local covariance matrix as the principal eigenvalue. All principal feature values ​​are arranged in the order of the first sub-region, the second sub-region, and so on up to the last sub-region, forming a state coupling degree sequence.

4. The elevator energy-saving operation control system based on a large model according to claim 3, characterized in that, Based on the state coupling degree sequence, a state correction factor characterizing the overall operating status is extracted, and this state correction factor is fused with operating parameters to determine the elevator operating status judgment result, including: Statistical analysis is performed on the state coupling degree sequence to extract the mean, variance, and skewness characteristics of the state coupling degree sequence. The mean, variance, and skewness characteristics of the state coupling degree sequence are then combined to form a state correction factor that characterizes the overall operating status. The status correction factor is fused with the call signal, real-time load and real-time location data in the original set of operating parameters to generate fused operating characteristic information; The system analyzes whether the fused operational feature information contains a preset emergency evacuation feature pattern. The emergency evacuation feature pattern is defined as all floors generating a continuous downward call signal within a continuous time window, and the signal strength exceeding a preset threshold. If the fused operational feature information contains an emergency evacuation feature pattern, the elevator is determined to be in an emergency evacuation state. If the fused operational feature information does not contain an emergency evacuation feature pattern, the elevator is determined to be in a normal operating state, thus obtaining the elevator operation status judgment result.

5. The elevator energy-saving operation control system based on a large model according to claim 4, characterized in that, Based on the elevator operation status judgment result, the first elevator scheduling instruction or the second elevator scheduling instruction is obtained through the first scheduling model or the second scheduling model in the pre-trained dual model library. Execute the first elevator dispatch command or the second elevator dispatch command to control the elevator group, including: If it is determined that the elevator is in normal operation, the fused operation feature information is input into the first scheduling model in the pre-trained dual-model library to generate the first elevator scheduling instruction. If it is determined that the elevator is in an emergency evacuation state, the fused operating feature information is input into the second scheduling model in the pre-trained dual-model library to generate a second elevator scheduling instruction. The first elevator dispatch instruction or the second elevator dispatch instruction is sent to the elevator group control execution unit, which then controls the elevator group according to the first elevator dispatch instruction or the second elevator dispatch instruction.

Images