Elevator energy-saving operation control system based on large model

Abstract

The invention provides an elevator energy-saving operation control system based on a large model, and relates to the technical field of elevator control, and the system comprises a data collection module which 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 relation; the structure construction module is used for performing coordinate positioning in a predefined state reference system based on the fused feature set, and selecting key feature points of the current operation situation core representation to construct a closed multi-dimensional feature structure; and the link identification module is used for calculating mutual information and a dynamic correlation degree between vertexes connected with each edge in the multi-dimensional feature structure, and weighting each edge according to the correlation degree to obtain a key link. According to the method, multi-module cooperation is used for capturing multi-element association of elevator group operation, analyzing the situation and adapting to double models to achieve self-adaptive scheduling, and collaborative optimization of energy conservation and service quality is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of elevator control, in particular to an elevator energy-saving operation control system based on a large model. BACKGROUND

[0002] The energy-saving optimization of an elevator group control system is an important part of building energy management. Current common energy-saving control schemes mostly use scheduling strategies based on statistical laws or preset rules, such as elevator scheduling according to time-of-use electricity prices, historical average passenger flow data, or simple real-time signal thresholds. Such methods can achieve certain energy-saving effects in scenarios where passenger flow is relatively stable and predictable. However, in actual operating environments with complex dynamic changes in passenger flow, such as a large general hospital during the daily peak visiting period, the elevator call requests from different floors, such as the outpatient department, the laboratory department, and the inpatient department, exhibit sudden, interwoven, and continuously evolving characteristics within a short period of time. Meanwhile, the load, position, running direction, and energy consumption of multiple elevators interact in real time. Existing methods may face some deficiencies when analyzing this high-dimensional, strongly coupled real-time operating situation. Their state analysis mostly focuses on monitoring and responding to a single or a small number of isolated parameters, such as waiting time and car load. There may be limitations in capturing and quantifying the complex, nonlinear internal relationships between multiple factors in the system, such as the spatial and temporal distribution of elevator signals, the coordination state of multiple cars, and the instantaneous energy consumption. This limitation may lead to an inaccurate model of the overall operating situation, which in turn may affect the adaptability and optimization potential of subsequent energy-saving scheduling strategies. There is room for further improvement in the balance between energy efficiency and service quality when dealing with complex and dynamic passenger flow patterns. SUMMARY

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

[0004] To solve the above technical problems, the technical solutions of the present application are as follows:

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

[0006] A data acquisition module for acquiring real-time operating parameters of an elevator group to obtain a multi-dimensional state vector, and generating a fusion feature set according to a pre-set space-time mapping relationship;

[0007] A structure construction module for coordinate positioning in a pre-defined state reference system based on the fusion feature set, selecting key feature points that represent the core characteristics of the current operating situation, and constructing a closed multi-dimensional feature structure.

[0008] a link identification module, configured to calculate mutual information and dynamic correlation degree between vertices connected by edges in the multi-dimensional feature structure, weight the edges according to the correlation degree, and obtain key links;

[0009] a manifold reconstruction module, configured to divide the multi-dimensional feature structure into a plurality of sub-regions according to the key links, and perform manifold reconstruction of a state vector field for each sub-region to obtain a reconstructed manifold of each sub-region;

[0010] a coupling analysis module, configured to generate a corresponding local covariance matrix according to the reconstructed manifold of each sub-region, and calculate a state coupling degree sequence;

[0011] a factor generation module, configured to extract a state correction factor representing an overall operation situation according to the state coupling degree sequence, and fuse the state correction factor with operation parameters to determine an elevator operation state judgment result;

[0012] a scheduling decision module, configured to obtain a first elevator scheduling instruction or a second elevator scheduling instruction by the first scheduling model or the second scheduling model in the pre-trained double model library based on the elevator operation state judgment result, and control the elevator group by executing the first elevator scheduling instruction or the second elevator scheduling instruction.

[0013] The above scheme of the application at least has the following beneficial effects:

[0014] The data acquisition module is used to collect elevator call signals, car load and position and other multi-source parameters in real time, and generate a fusion feature set through spatiotemporal mapping relationship processing, breaking the limitation of single parameter analysis; the feature coordinate positioning and key feature point screening are used to construct a closed multi-dimensional feature structure, so that the dispersed operation data is converted into structured information focusing on the core situation, and the system can comprehensively capture the multi-element correlation in the elevator group operation; the mutual information and dynamic correlation degree of each node in the feature structure are calculated to quantitatively evaluate the strength of the multi-element correlation, and the key links affecting the operation situation are screened out; the complex feature structure is divided into sub-regions and manifold reconstruction is completed, so that the internal law of the local operation state is clearly represented, and the problem of insufficient capture of complex correlation is solved; the state coupling degree sequence is generated by the coupling analysis module, the state correction factor reflecting the overall situation is extracted, and the state judgment is performed after the state correction factor is fused with the original operation parameters; this judgment method based on multi-dimensional feature fusion can not only identify normal operation state, but also effectively capture special modes such as emergency evacuation, and improve the intelligent level and reliability of state judgment; based on the accurate state judgment result, the scheduling strategy is adaptively switched through the double model library, and this differentiated scheduling method enables the system to match the optimal scheduling strategy according to different operation states, realizing the collaborative optimization of energy efficiency and service quality. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a schematic diagram of an elevator energy-saving operation control system based on a large model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0016] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0017] As Figure 1 shown, an embodiment of the present application proposes an elevator energy-saving operation control system based on a large model, comprising:

[0018] A data acquisition module is configured to acquire operation parameters of an elevator group in real time, obtain a multi-dimensional state vector, and generate a fusion feature set according to a preset space-time mapping relationship;

[0019] A structure construction module is configured to perform coordinate positioning in a predefined state reference system based on the fusion feature set, select key feature points representing a current operation situation core, and construct a closed multi-dimensional feature structure;

[0020] A link identification module is configured to calculate mutual information and dynamic correlation between vertices connected by each edge in the multi-dimensional feature structure, weight each edge according to the correlation, and obtain key links;

[0021] A manifold reconstruction module is configured to divide the multi-dimensional feature structure into a plurality of sub-regions according to the key links, and perform manifold reconstruction of a state vector field for each sub-region to obtain a manifold of each sub-region after reconstruction;

[0022] A coupling analysis module is configured to generate a corresponding local covariance matrix according to the manifold of each sub-region after reconstruction, and calculate a state coupling degree sequence;

[0023] A factor generation module is configured to extract a state correction factor representing an overall operation situation according to the state coupling degree sequence, and fuse the state correction factor with the operation parameters to determine an elevator operation state judgment result;

[0024] A scheduling decision module is configured to obtain a first elevator scheduling instruction or a second elevator scheduling instruction based on the elevator operation state judgment result through a first scheduling model or a second scheduling model in a pre-trained double model library; and execute the first elevator scheduling instruction or the second elevator scheduling instruction to control the elevator group.

[0025] In the embodiment of the present application, the spatio-temporal characteristics of the elevator call signals, the car load and the position and other multi-source parameters are collected in real time by the data collection module, and a fusion feature set is generated through spatio-temporal mapping relationship processing, breaking the limitation of single parameter analysis; through feature coordinate positioning and key feature point screening, a closed multi-dimensional feature structure is constructed, the dispersed operation data is converted into structured information of focused core situation, so that the system can fully capture the multi-element correlation in the operation of the elevator group; through the calculation of mutual information and dynamic correlation degree of each node in the feature structure, the quantitative evaluation of the correlation strength of multi-element is realized, and the key link affecting the operation situation is screened out; the complex feature structure is divided into sub-regions and the manifold reconstruction is completed, so that the internal law of the local operation state is clearly represented, and the problem of insufficient capture of complex correlation relationship is solved; through the generation of state coupling degree sequence, the state correction factor reflecting the overall situation is extracted, which is fused with the original operation parameters for state judgment, improving the intelligent level and reliability of state judgment; based on the state judgment result, the adaptive switching of the dispatching strategy is realized through the double model library, so that the system can match the optimal dispatching strategy according to different operation states, and the collaborative optimization of energy efficiency and service quality is realized.

[0026] In a preferred embodiment of the present application, the operation 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 the preset spatio-temporal mapping relationship, including:

[0027] The time stamp of the elevator call signal of each floor, the floor number, the real-time load value of each car, and the real-time position of each car in the shaft are collected in real time to obtain an original operating parameter set, specifically including: integrating a signal collection module in the elevator call panel of each floor, which is linked with the trigger circuit of the elevator call button. When a passenger presses the elevator call button, the module immediately records the time stamp of the trigger action and the corresponding floor number. The time stamp uses a high-precision clock signal unified by the system; 4 groups of pressure strain gauge load sensing devices are symmetrically installed on the load-bearing beams at the bottom of each elevator car. When the car is carrying a load, the load-bearing beams will produce a slight deformation, which will drive the strain gauges to deform synchronously, causing the resistance value of the strain gauges to change; the resistance change is converted into a voltage signal (i.e. a pressure electric signal) through a Wheatstone bridge circuit. After the voltage signal is processed by a signal amplification module, it is transmitted to a data processing unit for load calculation. The real-time load value is first established by offline calibration to establish the corresponding relationship between the pressure electric signal and the load, i.e. placing standard weights of known weights in the car in turn, recording the average voltage value output by the 4 groups of sensing devices under the weight of each weight, and forming a calibration curve of voltage value-load value. A linear regression equation is fitted, i.e. load value = (current average voltage value-no load voltage value) x calibration coefficient + no load compensation value, wherein the no load voltage value is the output voltage of the sensing device when the car is not loaded, the calibration coefficient is the slope of the regression equation, and the no load compensation value is used to correct the system error; after the data processing unit receives the real-time voltage signal, it first eliminates the abnormal values in the 4 groups of data, such as data deviating from the average value by more than 5%, then calculates the average voltage value of the remaining data, and substitutes it into the above equation to obtain the real-time load value; position sensors (such as photoelectric sensors) are installed on the inner wall of the elevator shaft at equal intervals along the vertical direction. The interval between adjacent sensors matches the distance between 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 turn. Combined with the pulse counting principle of the elevator operation (counting one pulse for each sensor trigger, and the running distance is the product of the pulse number and the sensor interval), the absolute position coordinates of the car in the shaft are obtained in real time; each collection module is connected with the elevator group control main controller through industrial Ethernet, and data is transmitted in a combination of polling and interruption. For the elevator call signal, the interruption trigger mode is adopted to ensure that the signal is uploaded immediately after it is generated; for the load and position data, the polling mode with a 100 millisecond cycle is adopted to avoid data redundancy; after the main controller receives the data uploaded by each module, the data is stored in categories according to the data type, elevator number / floor number, and time stamp format, forming an original operating parameter set containing elevator call signal data (time stamp, floor number), car load data (elevator number, time stamp, real-time load value), and car position data (elevator number, time stamp, real-time position coordinates). At the same time, missing or abnormal data, such as load values exceeding the equipment range and position coordinates suddenly changing, are marked.

[0028] The data collected in the same time slice in the original running parameter set is aligned according to the three dimensions of the elevator signal, the real-time load, and the real-time position, and is respectively vectorized to form a multi-dimensional state vector, which specifically includes: dividing the time into continuous time slices of fixed length, and setting the time slice length to be consistent with the polling period of the load and position data, that is, 100 milliseconds, while covering all elevator signals generated in the time slice; for each time slice, all data with time stamps falling within the time slice are screened from the original running parameter set, and then classified and arranged according to the three dimensions of the elevator signal, the real-time load, and the real-time position, that is, the elevator signal dimension summarizes the elevator call situation of each floor in the time slice to form a corresponding relationship of floor number-whether there is an elevator call; the real-time load dimension summarizes the load values of each car according to the elevator number; the real-time position dimension also summarizes the position coordinates of each car according to the elevator number, realizing data alignment; the aligned data of each dimension is respectively vectorized, the elevator signal dimension takes the total number of building floors as the vector length, each element corresponds to a floor, if the floor has an elevator signal in the time slice, the element value is 1, and if there is no elevator signal, the element value is 0, forming an elevator signal vector; the real-time load dimension takes the total number of elevators in the elevator group as the vector length, each element corresponds to the real-time load value of one elevator, forming a load vector; the real-time position dimension also takes the total number of elevators as the vector length, each element corresponds to the real-time position coordinates of one elevator, forming a position vector, finally the elevator signal vector, the load vector, and the position vector are sequentially spliced to form a multi-dimensional state vector containing all dimension information, each time slice corresponds to an independent multi-dimensional state vector.

[0029] The multi-dimensional state vector is input into a preset space-time mapping relationship. According to the corresponding rules of floors and time, the space-time mapping relationship extracts trajectory data describing the displacement process of the car, intensity data describing the cumulative situation of the call signal of each floor, and coordination data describing the relative position relationship of multiple cars from the multi-dimensional state vector. Specifically, the three core rules of the space-time mapping relationship are defined, the trajectory extraction rule defines the time window length of the position data, the speed calculation method and other parameters, wherein the time window length is set according to the time consumption of the average elevator running one floor (about 2 seconds); the intensity extraction rule defines the time range of the call signal accumulation and the intensity value statistical method, and the cumulative time range is determined with reference to the time interval (about 1 second) of the passenger concentrated call; the coordination extraction rule specifies the combination method of the elevator pair, the calculation logic of the relative position and direction coordination, and ensures that the interaction state of the multiple cars can be accurately reflected; for each elevator, the position vector element in the multi-dimensional state vector of the current time slice and the previous five consecutive time slices (600 milliseconds, covering the time consumption of the elevator running about 1 / 3 floor) is called to obtain the position sequence of the elevator; the position difference value of adjacent time slices is calculated, and each difference value is divided by the time slice length (100 milliseconds, i.e. 0.1 second) to obtain the instantaneous running speed; the position sequence and the speed sequence are combined in time sequence to form the trajectory data describing the displacement process of the elevator, and the trajectory data of the elevator group is the trajectory information set of all elevators; for each floor, the call signal vector element in the multi-dimensional state vector of the current time slice and the previous nine consecutive time slices (10 time slices, 1 second) is called, each element value is 1 indicating that there is a call in the time slice, and 0 indicating no call, the number of elements with value 1 in the 10 elements is counted, and the number is directly taken as the call signal intensity value of the floor, if the counted number is 0, the intensity value is 0, and if the number is 3, the intensity value is 3, the maximum intensity value is 10, i.e. there is a call in each time slice within 1 second, and the intensity values of all floors are arranged in order from low to high according to the floor number to form the intensity data with the same dimension as the total number of floors, reflecting the cumulative activity degree of the call signal of each floor; a combination algorithm is used to generate all non-repeating elevator pairs in the elevator group, such as elevator 1 and elevator 2, elevator 1 and elevator 3, etc., for each pair of elevators, set as elevator A and elevator B, the position vector element in the multi-dimensional state vector of the current time slice is called to obtain the position PA of A and the position PB of B, the relative position value of the two is calculated =PA-PB, if is positive, A is above B; if it is negative, A is below B; the running direction (V>0 for upward, V<0 for downward, V=0 for stationary) is judged in combination with the current speed (VA, VB) of A and B in the trajectory data to generate the direction coordination value (same direction is 1, opposite direction is -1, at least one is stationary is 0); the relative position value-direction coordination value of each pair of elevators is taken as the coordination information, and the information of all elevator pairs is summarized to form the coordination data, reflecting the interaction state of multiple cars.

[0030] The trajectory data, intensity data and coordination data are normalized and spliced to form a fusion feature set, specifically including: eliminating dimensional differences by using min-max normalization to make various types of data in the same numerical interval, in the trajectory data, the position value is bounded by the top layer (Hmax) and the bottom layer (Hmin) of the shaft, and is mapped to the 0-1 interval by normalizing position=(original position-Hmin) / (Hmax-Hmin); the speed value is bounded by the rated maximum speed (Vmax) of the elevator, and is mapped to the 0-1 interval (covering the up and down two-way speed) by normalizing speed=(original speed+Vmax) / (2*Vmax); the intensity data is bounded by the maximum number of calls (10 times) in 1 second, and is mapped to the 0-1 interval by normalizing intensity=original intensity / 10; in the coordination data, the relative position value is bounded by the maximum height (Hmax) of the shaft, and is mapped to the 0-1 interval by normalizing relative position=(ΔP+Hmax) / (2*Hmax); the direction coordination value retains the original value of-1, 0 and 1 (which is the standardized result); the normalized three types of data are spliced horizontally according to the fixed order of the trajectory data, the intensity data and the coordination data to form a high-dimensional feature vector, and identification information (time slice number+elevator group number) is added to each feature vector, the feature vectors of all time slices are arranged in time sequence, and finally a fusion feature set is formed, which completely contains the three types of core information of elevator operation, i.e., trajectory, call intensity and car coordination.

[0031] In this embodiment, the data classification storage and abnormal marking mechanism solve the situation awareness deviation problem caused by one-sided data collection; the time slice division realizes data alignment, so that the elevator group operation state of each time slice is structurally represented; based on the space-time mapping relationship, three types of core data, i.e., trajectory, intensity and coordination, are extracted, the key features of elevator operation are focused, and the interference of redundant data is avoided; the normalization processing eliminates the dimensional differences, improves the usability of data and the accuracy of state representation; through the space-time mapping relationship, the internal relationship between multiple factors of elevator operation is captured; the fusion feature set integrates these related information, so that the system can perceive the complex situation such as elevator signal burst and frequent car interaction under the scene of hospital peak, and solves the problem of insufficient capture of complex correlation.

[0032] In a preferred embodiment of the present application, based on the fusion feature set, coordinate positioning is performed in a pre-defined state reference system, and key feature points of the core representation of the current operation situation are selected to construct a closed multi-dimensional feature structure, including:

[0033] According to the type of each feature item in the fusion feature set, the feature item is mapped to the corresponding coordinate axis in the predefined state reference system, specifically including: explicitly defining the structure of the state reference system, which is a multidimensional Euclidean space, and the coordinate axes are divided according to the feature types of the fusion feature set, and three major coordinate axis clusters are set, wherein the first type is a trajectory feature coordinate axis cluster, including the position coordinate axis of each elevator and the speed coordinate axis (2 coordinate axes corresponding to a single elevator, and 2U coordinate axes for U elevators); the second type is a strength feature coordinate axis cluster, and each floor corresponds to one call strength coordinate axis, and M floors contain M coordinate axes; the third type is a coordination feature coordinate axis cluster, and each pair of elevators corresponds to two coordinate axes (relative position coordinate axis and direction coordination coordinate axis), and 2K coordinate axes for K pairs of elevators form a standardized state reference system; each feature item in the fusion feature set has a type identifier (trajectory class / strength class / coordination class) and a subdivision attribute (such as elevator 1-position, 3rd floor-call strength, and elevator 1 and 2-relative position), and the mapping rule is that the corresponding coordinate axis cluster is matched according to the type identifier of the feature item, and then the specific coordinate axis in the cluster is matched through the subdivision attribute; for example, the feature item of elevator 3-speed in the fusion feature set is matched to the trajectory feature coordinate axis cluster through the trajectory class identifier, and then the speed coordinate axis of elevator 3 in the cluster is matched through the subdivision attribute of elevator 3-speed; the 5th floor-call strength feature item is matched to the strength feature coordinate axis cluster through the strength class identifier, and then the call strength coordinate axis of the 5th floor is matched; all feature items in the fusion feature set are traversed through the program, and each feature item is assigned a unique corresponding coordinate axis in the state reference system according to the above mapping rule, and a feature item-coordinate axis mapping relationship table is recorded, and if the feature item attribute is ambiguous, such as no elevator number is labeled, an abnormal processing mechanism is triggered, the feature item is marked as to be confirmed, and the mapping is suspended, and after manual verification, it is executed again.

[0034] According to the numerical size of each feature item, the position of the feature item on the mapping coordinate axis is determined, and a coordinate set of all feature items is obtained, specifically including: each coordinate axis in the state reference system is pre-set with a fixed numerical range, in the trajectory feature, the position coordinate axis range is the distance from the bottom layer to the top layer of the elevator shaft, such as 0 meters to 50 meters, the speed coordinate axis range is the rated maximum downward speed to the rated maximum upward speed of the elevator, such as -2 meters / second to 2 meters / second; in the intensity feature, the call intensity coordinate axis range is 0 to 10; in the coordination feature, the relative position coordinate axis range is -50 meters to 50 meters (matching the maximum height of the shaft), the direction coordination coordinate axis range is -1 to 1 (consistent with the value range of the direction coordination value), each coordinate axis is divided into scales according to the interval of the range / 1000; the numerical value of the feature item in the fusion feature set has completed normalization processing, and it is necessary to convert the normalized numerical value into the actual coordinate value on the coordinate axis, that is, coordinate value=(normalized numerical value×coordinate axis numerical range span)+coordinate axis starting value, wherein the coordinate axis numerical range span=coordinate axis maximum-value-coordinate axis minimum value; after the coordinate value calculation is performed on all feature items in the fusion feature set one by one, the feature identification, corresponding coordinate axis and coordinate value information of each feature item are associated and stored, and a coordinate set containing the spatial position information of all feature items is obtained, which is presented in the form of a data table, and each row corresponds to the coordinate data of a feature item.

[0035] The coordinate set is analyzed by using a density-based clustering algorithm to identify areas with dense distribution of coordinate points, specifically including: using a density-based spatial clustering application algorithm to analyze the coordinate set, first determining the core parameters of the algorithm (neighborhood radius and minimum point number), and adapting the elevator operation data characteristics through offline training; first, prepare the training data, select four typical passenger flow scenarios: morning peak (7:30-9:00), noon flat peak (12:00-13:30), hospital treatment peak (10:00-11:30), and night low peak (21:00-22:30), collect one hour of operation data for each scenario, generate four sets of coordinate sets, and comprehensively cover different passenger flow densities and operation states; the search range of the neighborhood radius is determined as 0.1 to 1.0 coordinate axis scale units, divided into 10 candidate values at a step of 0.1; the search range of the minimum point number is determined as 3 to 10, divided into 8 candidate values at a step of 1, generate several groups of independent parameter combinations through full combination to meet the parameter requirements in different density scenarios; then, apply all parameter combinations to the coordinate sets of the four typical passenger flow scenarios, each parameter independently completes clustering operation on each coordinate set, and synchronously records four core quantitative indicators, namely, 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 intersection area ratio of the spatial boundaries of different clusters, a value of 0 indicates no overlap), core feature proportion (the proportion of trajectory and intensity feature items in the total feature items in the cluster), and noise point proportion (the proportion of noise points in the total number of coordinate points); based on the quantitative indicators, a double screening mechanism is constructed, the first level is the hard indicator threshold, requiring the inter-cluster boundary overlap to be 0 (to ensure clear dense area contours), the core feature proportion to be greater than or equal to 90%, and the noise point proportion to be less than or equal to 5% (to ensure high core feature aggregation); the second level is the comprehensive quality indicator screening, introducing the contour coefficient to quantify the clustering effect from a global dimension, which evaluates the clustering rationality by measuring the intra-cluster tightness and inter-cluster separation, for each feature point, calculate a value (the average Euclidean distance of the point and all other points in the cluster, reflecting the intra-cluster tightness, the smaller the value, the better) and b value (the average Euclidean distance of the point and all points in the nearest non-belonging cluster, reflecting the inter-cluster separation, the larger the value, the better), the contour coefficient of a single point is calculated according to the formula (b-a) / max(a,b), the value range is [-1, 1] (when the value is close to 1, the point is tightly clustered in the cluster and the boundary with other clusters is clear; when the value is close to -1, the point clustering attribution is wrong; when the value is close to 0, the point is at the inter-cluster boundary); take the average of the contour coefficients of all feature points to get the overall contour coefficient of the parameter combination in the current scenario, set the qualified threshold to be greater than or equal to 0.7 (verified by historical data, corresponding to the adaptive effect of closely related intra-cluster feature and significantly different inter-cluster sub-state); for the six parameter combinations screened in the first level, calculate their overall contour coefficients on the four coordinate sets and sort them from high to low, finally select the neighborhood radius5 coordinate axis scale units, the minimum point number 5, the optimal parameter combination is excellent in all scenarios, that is, the cluster boundary overlap degree is 0, the core feature proportion is greater than or equal to 92%, the noise point proportion is less than or equal to 3%, and the overall contour coefficient is greater than or equal to 0.75, which is much higher than the qualified threshold, and has the full working condition adaptation ability; when a certain feature point contains at least 5 other feature points in the neighborhood with itself as the center and 0.5 coordinate axis scale as the radius, it is determined to be a core point (corresponding to the stable state of multi-feature cooperation in elevator operation, such as the region of continuous high call intensity and the uniform speed running track feature of the elevator); all points in the coordinate set are traversed, and the point type is determined according to the parameter standard, the core point (the number of points in the neighborhood is greater than or equal to 5), the density reachable point (the number of points in the neighborhood is less than 5 but falls in the neighborhood of the core point, corresponding to the subordinate feature related to the stable feature, such as the elevator speed change around the high call area), and the noise point (neither core point nor density reachable point, corresponding to accidental fluctuation, such as false call signal and instantaneous vibration offset data of the elevator); starting from the first unmarked core point, the core point and all density reachable points are included in the initial cluster, and then the other core points in the cluster are taken as the starting point for recursive expansion until no new density reachable point can be added, forming a complete clustering cluster; after marking the points belonging to the cluster, the next unmarked core point is repeated, and finally a plurality of mutually non-overlapping clustering clusters are obtained, and the noise points are separately marked and do not participate in subsequent analysis; for each clustering cluster, the maximum and minimum values of all points in the cluster on each coordinate axis are extracted to form a rectangular boundary in a multi-dimensional space (such as elevator 1-position axis 10 to 15 meters, 3-floor-call intensity axis 2 to 5); small accidental clusters containing less than 10 points are removed, and large clusters containing more than 10 points are retained, and the corresponding space range is the core dense area (reflecting multi-feature cooperation, such as the traction association of a certain position elevator and the corresponding floor call).

[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 variance contribution degree core is to see the fluctuation of the efficiency index. How much is caused by this feature. The higher the explanation ratio, the greater the value. Taking the load of elevator 1 and the energy consumption per unit time as an example, in addition to the load of elevator 1, other features that may affect energy consumption are selected, such as the speed of elevator 1, the running floor height, a total of 3 features. Before calculating the variance contribution degree, the calculation logic of energy consumption fluctuation needs to be determined. It is essentially to measure the difference in elevator energy consumption in different time periods. First, select 10 consecutive 1-minute time periods (such as 9:00 to 9:01 every day) in the past 3 months, and record the energy consumption per unit time of the elevator. Add the 10 data and divide by 10 to get the average energy consumption value for this period. Subtract the average energy consumption from the actual energy consumption of each time period to get the difference value (positive or negative represents higher or lower than the benchmark). Square all difference values (eliminate positive and negative effects) and add them together. This total is the total energy consumption fluctuation. The larger the value, the more obvious the energy consumption difference between different time periods. After determining the energy consumption fluctuation calculation, the variance contribution degree of the load of elevator 1 is calculated. First, assume that the energy consumption fluctuation is only related to the 3 features (load of elevator 1, speed of elevator 1, running floor height), and calculate how much the 3 features can explain the energy consumption fluctuation together (for example, a total of 70%). Then remove the load of elevator 1 feature and see how much the remaining 2 features can explain the fluctuation (for example, it becomes 40%). The explanation ratio with load minus the explanation ratio without load is the variance contribution degree of the load of elevator 1. For example, 30%, the corresponding variance contribution degree is 0.3, indicating that load changes can explain 30% of the energy consumption fluctuation. The variance contribution degree of each feature item is calculated with the two efficiency indicators respectively, and the average value is taken as the final result. Weight value = (mutual information value x 0.6) + (variance contribution degree x 0.4), where the mutual information value can capture non-linear correlation (most of the elevator operation is non-linear relationship), which is more critical to the state characterization, so it is given a higher weight (0.6). The variance contribution degree focuses on linear explanation ability and is used as a supplement (0.4). Calculate the weight value of all feature items, and take the 75th percentile as the preset importance weight threshold (finally determined as 0.6), that is, 75% of the feature item weight is less than the value, to ensure that the threshold can effectively filter out a few high importance features. In each dense area, calculate the Euclidean distance between all points in the cluster and the regional center coordinates, and select the first 10 points with the smallest distance as candidate center feature points to ensure that the candidate points have spatial centrality. Query the weight value of the feature item corresponding to the 10 candidate points, and retain the candidate points with a weight value greater than 0.6. These points meet both spatial centrality and high importance, and are the key feature points of the region. If there are no candidate points with a weight value that meets the requirements in a certain region, such as the weak feature correlation in the low peak period, select the candidate point with the largest weight value as the key feature point to ensure that each dense area has a corresponding key feature point, covering the core operation state comprehensively.

[0038] The two key feature points are connected to form an edge, and the Pearson correlation coefficient of the original running parameter set corresponding to the two key feature points connected by each edge in the set time length is calculated, specifically including: all key feature points are connected by a full connection strategy, that is, a non-directional edge (no direction distinction, only represents the association relationship) is constructed between any two different key feature points; each edge is assigned a unique identifier, and the information of the two key feature points connected by the edge is associated and stored, such as point A corresponding to elevator 1-speed, and point B corresponding to 3rd floor-intensity, to form an association table of edges and points; the set time length is 10 seconds, corresponding to 100 time slices, the original running parameter set of the current time slice and the previous 99 time slices is selected, the time length can cover the typical period of 1 to 2 times of starting and stopping of the elevator, reflect the short-term dynamic association, and avoid the distortion of the association relationship caused by too long time; for the two key feature points (corresponding to feature items X and Y) connected by each edge, the original parameter sequence (sequence X and sequence Y, both with a length of 100) of the two key feature points in 10 seconds is extracted, and the Pearson correlation coefficient is calculated, with a value range of -1 to 1, a positive value indicating that the feature item change trend is consistent (such as the elevator speed increasing when the call intensity increases), a negative value indicating that the trend is opposite, and the greater the absolute value, the stronger the association degree.

[0039] The edges with a Pearson correlation coefficient higher than a preset connection threshold are retained, and the closed multi-dimensional feature structure is formed by all the retained edges and the key feature points connected by the retained edges, and specifically includes: determining a preset connection threshold, which is used to distinguish significant correlation edges and weak correlation edges, and the determination process is as follows: collecting key feature point pair correlation data in 8 typical passenger flow scenes in the past 1 month, a total of 1000 groups of effective feature point pairs are obtained, and the Pearson correlation coefficient of each group is calculated; 1000 correlation coefficients are counted, and it is found that the absolute value of the coefficient is concentrated between 0 and 0.8, and the coefficients with an absolute value greater than or equal to 0.3 account for about 30%, and the feature point pairs corresponding to these coefficients, such as the 5th floor call intensity-elevator 2 position and the elevator 1 load-elevator 1 speed, have clear logical correlation in actual operation; and the coefficients with an absolute value less than 0.3 are mostly accidental correlations, such as the 10th floor call intensity-elevator 3 speed weak correlation in the low peak period, that is, 0.3 is selected as the preset connection threshold, and the edges with a correlation coefficient absolute value greater than or equal to 0.3 are determined as effective edges with significant correlation; all the edges are traversed, the Pearson correlation coefficient absolute value of each edge is calculated, and the threshold is compared for screening, the edges with a coefficient absolute value greater than 0.3 are recorded, and the two key feature points connected by the edges, the correlation coefficient value and the corresponding feature item information are recorded, such as edge-1: point A (elevator 1 speed)-point B (3rd floor intensity), coefficient 0.52; the edges with a coefficient absolute value less than 0.3 are marked as redundant edges and deleted; based on the retained effective edges and the corresponding key feature points, a multi-dimensional feature graph model is constructed through a graph structure modeling tool (such as NetworkX), wherein the key feature points are used as the vertices of the graph, each vertex is labeled with the corresponding feature item information, such as elevator 2-position; the effective edges are used as the edges connecting the vertices, and the weight of the edge is set as the corresponding Pearson correlation coefficient absolute value (reflecting the correlation strength); since the effective edges are all core correlation edges, the vertices naturally form multiple connected subgraphs through the edges, such as a subgraph composed of call intensity-elevator position-elevator speed; within each subgraph, the vertices form a ring through the mutual connection of the edges, such as point A-point B-point C-point A or a mesh structure, such as point A connecting point B, point C and point D, and point B connecting point C and point E. These structures form a closed area in space and can completely reflect a certain running sub-situation, such as the elevator response situation of the high call area; all the connected subgraphs are combined to form a complete closed multi-dimensional feature structure, which presents the correlation network between the core features of the elevator operation in the form of a visual graph, and each closed substructure corresponds to a running sub-situation.

[0040] The embodiment realizes the structured and visualized representation of the multi-dimensional features of the elevator operation by establishing a standardized state reference system, accurately mapping each feature item in the fused feature set to the corresponding coordinate axis, determining the spatial coordinate position of each feature item, and solving the problem of feature management confusion; the dense area of feature points is automatically identified by the density clustering algorithm, and the core area of feature coordination aggregation in the elevator operation is focused; the key feature points are screened in combination with the spatial centrality and feature importance double standards, so that the selected feature points can reflect the core characteristics of the operation situation, and the pertinence and accuracy of the situation representation are improved; the linear correlation strength between features is quantified by calculating the Pearson correlation coefficient of the corresponding parameters of the key feature points; the core correlation edge is screened based on the correlation strength threshold, the closed multi-dimensional feature structure is constructed, and the problem of fuzzy correlation analysis is solved.

[0041] In a preferred embodiment of the present application, mutual information and dynamic correlation degree between the vertices connected by each edge in the multi-dimensional feature structure are calculated, each edge is weighted according to the correlation degree, and a key link is obtained, including:

[0042] For each edge in the closed multi-dimensional feature structure, the original operation parameter set corresponding to the two key feature points connected by each edge is obtained, the mutual information value of the original operation parameter set corresponding to the two key feature points within a set time span is calculated, and the specific steps include: first, each edge in the closed multi-dimensional feature structure is traversed, the two key feature points connected by the edge are queried through the identification information of the edge, and then the original operation parameter set corresponding to the two feature points is extracted according to the feature identification-time slice range association table of the feature points; the set time span is 30 seconds (matching the typical operation period of the elevator), corresponding to 300 time slice data (100 milliseconds / time slice); for example, the edge connects the two key feature points of elevator 1-speed and 3rd floor-call intensity, then 300 original parameter data of the two features within the same 30 seconds are extracted, forming two parameter sequences (sequence A: elevator 1 speed data, sequence B: 3rd floor call intensity data); second, the mutual information value is calculated, which reflects the nonlinear correlation by measuring the information sharing degree of the two parameter sequences, and the calculation steps are as follows: the two parameter sequences are discretized according to the numerical interval, the elevator speed sequence is divided into four intervals according to [-2,-1), [-1,0), [0,1), [1,2] m / s, and the call intensity sequence is divided into four intervals according to [0,2), [2,4), [4,6), [6,10] times / s, and the continuous data is converted into discrete categories; the occurrence frequency of sequence A in each interval is counted, and the marginal probability P(A) is calculated, such as the proportion of the number of times of A in the [1,2] interval to the total number of times; the marginal probability P(B) of sequence B is calculated in the same way; the occurrence frequency of the combination of A and B in the interval is counted, and the joint probability P(A,B) is calculated, such as the proportion of the number of times of A in the [1,2] and B in the [6,10]; the information entropy and mutual information are calculated, first the information entropy of a single sequence is calculated, the formula is i.e. the sum of the probabilities of each interval multiplied by their logarithm and then negated, resulting in the information entropy and H(B); and then the joint information entropy ; the final mutual information value is , the value range is 0 to 1, the greater the value, the stronger the nonlinear correlation between the two parameters; in a set time span, divide the time into multiple continuous sliding windows, and in each sliding window, calculate the Pearson correlation coefficient of the original running parameter set corresponding to the two key feature points, specifically including: based on a 30-second time span, set the window size to 5 seconds (including 50 time slice data, which can reflect the short-term correlation characteristics of the parameters), the sliding step is 2 seconds (including 20 time slices, to avoid too much window overlap causing data redundancy), the number of windows that can be divided in a 30-second time span is calculated as follows: window number = (total number of time slices - window time slice number) ÷ step time slice number + 1, to ensure complete coverage of the entire time span; taking a window as an example, extract the parameter sequence of the two key feature points in the window (sequence X and sequence Y, both with a length of 50), calculate the Pearson correlation coefficient, the value range is -1 to 1, positive value indicates positive correlation, negative value indicates negative correlation, and the greater the absolute value, the stronger the linear correlation; perform the above calculation on the 13 windows one by one to obtain 13 Pearson correlation coefficient values.

[0043] Calculate the average value of the Pearson correlation coefficients in all sliding windows, and calculate the variance of the Pearson correlation coefficients in all sliding windows, take the average value and the variance as the measurement value of the dynamic correlation degree, specifically including: sum the 13 correlation coefficient values, and then divide by the number of windows to obtain the average value of the Pearson correlation coefficient, the average value reflects the average linear correlation strength of the two parameters in the entire time span, the closer the value is to 1 or -1, the more stable the overall correlation is; calculate the Pearson correlation coefficient variance, the variance reflects the fluctuation degree of the correlation coefficient, reflects the dynamic characteristics of the correlation, the smaller the variance, the more stable the linear correlation between the two parameters is; the greater the variance, the more obvious the correlation strength changes with time; take the calculated average value of the Pearson correlation coefficient and the Pearson correlation coefficient variance as the two core measurement values of the dynamic correlation degree, which jointly represent the strength level and dynamic stability of the correlation.

[0044] The mutual information value and the dynamic correlation degree measurement value are normalized respectively to obtain normalized mutual information value and normalized dynamic correlation degree measurement value; the normalized mutual information value and the normalized dynamic correlation degree measurement value are weighted and summed according to a preset weight to obtain a comprehensive correlation strength value corresponding to an edge, which specifically includes: using the min-max normalization method, the mutual information value and the two measurement values (average value, variance) of the dynamic correlation degree are respectively mapped to the interval of 0 to 1; the mutual information value normalization is to collect the mutual information values of all edges in the closed multi-dimensional feature structure, and the global minimum value and the global maximum value of the index are obtained by traversal statistics; then the original mutual information value of each edge is substituted into the formula normalized value=(original value-global minimum value) ÷ (global maximum value-global minimum value) to calculate the normalized mutual information value of each edge; when the dynamic correlation degree measurement value is normalized, the dynamic correlation degree includes two measurement values of the average value of the Pearson correlation coefficient and the variance, which need to be processed differently according to the characteristics of the index, wherein the average value normalization, the average value is a positive index, the larger the value, the stronger the overall linear strength of the feature correlation, the min-max normalization formula is directly used, the global minimum value and the maximum value of the average value of the Pearson correlation coefficient of all edges are first counted, and then the original average value of each edge is substituted into the calculation to obtain the normalized average value; when the variance is normalized, the variance is a reverse index, the smaller the value, the more stable the linear strength of the feature correlation, if the conventional formula is directly used, the result will be contrary to the actual meaning, therefore, the formula is adjusted to normalized value=(global maximum value-original value) ÷ (global maximum value-global minimum value), which maps the variance to a positive index, ensuring that the larger the normalized value, the stronger the correlation stability; after the above operations, each edge will correspond to three normalized indexes, namely the normalized mutual information value, the normalized average value and the normalized variance; the normalized average value and the normalized variance are fused into a single dynamic correlation degree comprehensive value to realize the collaborative evaluation of the correlation strength level and the dynamic stability; considering that the average value reflects the correlation strength and the variance reflects the correlation stability, which are equally important for the dynamic correlation degree, the weights of both are set to 0.5, and the weighted sum method is used for calculation, the formula is dynamic correlation degree comprehensive value=(normalized average value×0.5)+(normalized variance×0.5); the comprehensive correlation strength value is synthesized combined with the elevator running characteristics, the correlation between features in the elevator running is mostly nonlinear, such as the correlation between the call strength and the elevator response speed, which is affected by many factors such as elevator position and passenger capacity, therefore, the mutual information value that can capture the nonlinear correlation is more valuable than the dynamic correlation degree comprehensive value that focuses on linear dynamic correlation; based on this, the preset weight is set, the normalized mutual information value accounts for 0.6, and the dynamic correlation degree comprehensive value accounts for 0.4, and the final comprehensive correlation strength value is obtained by weighted sum, the formula is comprehensive correlation strength value=(normalized mutual information value×0.6)+(dynamic correlation degree comprehensive value×0.4), the value of which is fixed in the range of 0 to 1, and the closer the value is to 1, the more stable the non-linear correlation and reliable linear dynamic correlation between the two feature points connected by the corresponding edge, which is a key edge reflecting the core correlation of elevator operation.

[0045] According to the comprehensive correlation strength value, each edge in the closed multi-dimensional feature structure is weighted and assigned, each edge is assigned a weighted value, all edges in the closed multi-dimensional feature structure are sorted in descending order of the weighted value of the edges, and the top N% of the sorted edges are selected as the key links, where N is a preset integer proportion threshold, which specifically includes: each edge in the closed multi-dimensional feature structure is assigned a unique weighted value to form an association table of edge identifier-weighted value; N is a preset integer proportion threshold, which is determined by historical data verification, that is, 8 typical passenger flow scenarios (weekday morning peak, hospital visit peak, shopping mall weekend peak, noon flat peak, night low peak, holiday flat peak, and sudden passenger flow peak) are selected, 30 days of continuous operation data (12 hours per day) are collected for each scenario, 100 sets of closed multi-dimensional feature structure-edge weighted value data sets can be generated for each scenario, a total of 800 verification samples, and the edges with the top 5% of the comprehensive correlation strength value in each scenario are defined as the benchmark core correlation edges, which correspond to the core collaborative relationship of elevator operation, such as the correlation between call intensity and elevator response speed, and the correlation between elevator load and energy consumption; the statistical index is set as the proportion of the top N% edges covering the benchmark core correlation edges, the higher the proportion, the better the screening effect of N value, and the calculation formula is Coverage Proportion = (number of benchmark core correlation edges contained in the top N% edges ÷ total number of benchmark core correlation edges in the scenario) x 100%; four candidate values of N=15, 20, 25, and 30 are tested respectively, and the average coverage proportion of each N value in the 800 samples is calculated, when N=15, the average coverage proportion is 68%, and in some scenarios, such as sudden passenger flow peak, the coverage proportion is only 52%, and there is a core correlation omission; when N=20, the coverage proportion of the 8 scenarios is all ≥80%, the average coverage proportion is 86%, and there is no core correlation omission; when N=25, the average coverage proportion is increased to 92%, but the number of links is increased by 42%, which will lead to redundancy in subsequent analysis; when N=30, the number of links is increased by 88%, and the redundancy problem is more prominent; considering the core coverage integrity and link simplicity, N=20 is determined as the preset integer proportion threshold, which can be adjusted according to the particularity of the elevator use scenario, such as super high-rise elevator and medical elevator, and after adjustment, the coverage proportion ≥80% needs to be verified by small sample data again; all edges are sorted in descending order of the weighted value, the total number of sorted edges is calculated, and the screening number is determined according to the total number x N%, for example, there are 100 edges, the top 20 edges are selected, and the edges ranked in the top N% are the key links, which represent the correlation between the core features in the elevator operation.

[0046] In this embodiment, the key links are screened by weighted ranking to filter the top N% key links, focus on the core correlation of elevator operation, eliminate the redundant edges of weak correlation, and reduce the amount of decision data and the complexity of calculation; through the sliding window mechanism, the correlation fluctuation is quantified by variance, so that the correlation strength evaluation can reflect the short-term dynamic changes in real time, for example, the correlation fluctuation between the call intensity and the elevator speed is larger during the morning peak, and this mechanism can accurately capture this feature to provide a basis for dynamically adjusting the scheduling strategy; a standardized process of parameter extraction, index calculation, normalization, and weighted screening is established, all calculation processes are based on quantitative data, and the weights and thresholds are determined through historical data verification to avoid the subjectivity of manual intervention and improve the reliability of key link identification.

[0047] In a preferred embodiment of the present application, according to the key links, the multi-dimensional feature structure is divided into several sub-regions, and the state vector field manifold reconstruction is performed on each sub-region to obtain the reconstructed manifold of each sub-region, including:

[0048] According to the key links and the weighted values assigned to each edge in the key links, the key feature points connected by the key links in the closed multi-dimensional feature structure are clustered and divided to form several topologically connected sub-regions, specifically including: taking the determined key link as the core edge and the key feature points connected by the key link as the vertices, a key correlation graph is constructed through a graph structure modeling tool (such as NetworkX), the key correlation graph clearly defines two rules, namely: the vertices only contain key feature points connected by key links (eliminate isolated feature points not associated with key links); the edges only retain key links, and the weights of the edges follow the comprehensive correlation strength values; the topological connectivity determination standard is that if there is one or more continuous key links between two key feature points, i.e., indirect connection through other key feature points and key links, then the two feature points are determined to be topologically connected; for example, feature point A is connected to feature point B through a key link, and feature point B is connected to feature point C through a key link, then A, B, and C all belong to a topologically connected set; the connected component extraction algorithm (such as the depth-first search algorithm) of the key correlation graph is called to traverse all vertices in the key correlation graph, i.e., starting from the first unmarked key feature point, recursively finding all topologically connected feature points with the point, to form an initial sub-region; mark the ownership of all feature points in the sub-region, and repeat the operation from the next unmarked feature point until all key feature points are assigned to a sub-region; each connected component corresponds to a topologically connected sub-region, which is named sub-region 1, sub-region 2, …, sub-region Q (Q is the total number of sub-regions) in order of clustering, and a sub-region-key feature point-key link association table is generated to record the feature point type (such as elevator speed, call intensity), quantity, and associated key link information included in each sub-region.

[0049] For the first sub-region, extract the multi-dimensional state vector corresponding to all key feature points contained in the first sub-region, and perform dimension reduction processing on the multi-dimensional state vector corresponding to all key feature points contained in the first sub-region to obtain a low-dimensional data point set corresponding to the first sub-region, which specifically includes: first, according to 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 running parameter sequence within a set time span, the time span is 30 seconds, corresponding to 300 time slices (100 milliseconds / time slice), forming a parameter sequence (length 300) of the feature point; align the parameter sequences of all key feature points in the sub-region by time slice, and combine them into a multi-dimensional state vector, the dimension of the vector = the number of key feature points in the sub-region × 300; for example, sub-region 1 contains elevator 1 speed and 3 floor call intensity 2 feature points, and the parameter sequence length of each feature point is 300, so the dimension of the multi-dimensional state vector is 600, the first to 300th positions in the vector are elevator 1 speed data, and the 301st to 600th positions are 3 floor call intensity data; second, to eliminate the dimensional differences of different feature parameters, the multi-dimensional state vector is standardized, and the processing formula is standardized value = (original parameter value - average value of parameter sequence of the feature point) ÷ standard deviation of the parameter sequence of the feature point, to ensure that the value range of each feature parameter after standardization is unified as [-1, 1]; third, t-SNE (t-distributed Stochastic Neighbor Embedding) algorithm is used for dimension reduction, the core of which is to preserve the local association structure of high-dimensional data, and the multi-dimensional state vector is mapped to a 2-dimensional space, and the specific steps are as follows: for the standardized multi-dimensional state vector, first calculate the high-dimensional similarity of any two time slice data in the vector, which is used to quantify the association closeness of the two time slice data in the high-dimensional space, and the value closer to 1 indicates that the feature trends of the two are more consistent (such as elevator speed and call intensity of the two time slices are changed synchronously), and the value closer to 0 indicates that the features are more different; the high-dimensional similarity is measured by a Gaussian kernel function, and the closer the distance, the higher the similarity, and the specific formula is similarity , wherein the Euclidean distance of two data is calculated according to high-dimensional coordinates, for example, the high-dimensional coordinates of time slice A are ( ), and the high-dimensional coordinates of time slice B are ( ) (v represents elevator speed and c represents call intensity), and the Euclidean distance calculation formula is Euclidean distance , the smaller the distance, the closer the molecular part (the square of the Euclidean distance of the two data) in the above Gaussian kernel function to 0, the closer the exponential operation result to 1, and the higher the similarity; the bandwidth parameter, the core role is to adjust the sensitive range of the Gaussian kernel function, that is, to control how large the Euclidean distance can still be determined as similar, and its value needs to be adaptively determined according to the local density of the feature point data to avoid local correlation distortion caused by fixed parameters. The specific determination method is: for each time slice data, statistics of its nearest 10 adjacent data in high-dimensional space (adjacent data = 10, adapt to the time continuity of the elevator data), calculate the average Euclidean distance of the 10 adjacent data and the data, and take this average distance as the local density index of the data; then take the median of the local density index of all time slices, and take 0.8 times of the median as the final bandwidth parameter; for example, the local density of a certain sub-area data is high (such as during the morning peak period, the feature data is dense), the average distance of the adjacent data is small, and the bandwidth parameter decreases accordingly, and the sensitive range of the Gaussian kernel function becomes narrow, and only the data with extremely close distance is determined as similar; if the local density is low (such as at night during the low peak), the bandwidth parameter increases, and the sensitive range is widened, avoiding missing potential associated data.

[0050] After completing the high-dimensional similarity calculation, a low-dimensional coordinate is randomly initialized for each time slice data in the 2-dimensional target space (the coordinate value range is set to [-5, 5] to ensure uniform initial distribution), and then the low-dimensional similarity of any two data in the low-dimensional space is calculated. The similarity is a mirror mapping target of the high-dimensional similarity in the low-dimensional space, which is used to measure whether the low-dimensional coordinate preserves the high-dimensional association characteristics. The low-dimensional similarity is calculated using the t-distribution kernel function, which has a wider tail and can provide more sufficient low-dimensional distribution space for different data. The specific calculation logic is similar to that of high-dimensional, first calculate the Euclidean distance of two data in 2-dimensional space, and then substitute it into the t-distribution kernel function formula. The formula also follows the rule that the closer the distance, the higher the similarity, but compared with the Gaussian kernel function, the decay speed of the low-dimensional similarity with the increase of the distance is more gentle, which can effectively avoid data congestion. Finally, the KL divergence (relative entropy) between the high-dimensional similarity distribution and the low-dimensional similarity distribution is minimized by the gradient descent method. The KL divergence is used to quantify the difference between the two distributions, and the value closer to 0 indicates that the low-dimensional distribution is more consistent with the high-dimensional distribution. The calculation formula is, In the iterative optimization process, the low-dimensional coordinates of each data are adjusted according to the change of the KL divergence in each iteration. For example, if the low-dimensional similarity of a certain data to the adjacent data is much lower than the high-dimensional similarity, the low-dimensional coordinates of the data are moved in the direction of the adjacent data. Here, the low-dimensional similarity is much lower than the high-dimensional similarity in combination with the correlation characteristics of the elevator feature data, and is specifically defined as that the ratio of the low-dimensional similarity to the high-dimensional similarity is less than 0.3, that is, the correlation degree of the two data in the low-dimensional space is far from reaching the correlation level in the high-dimensional space. At this time, the low-dimensional coordinates of the data need to be moved in the direction of the adjacent data, and the moving amplitude is positively correlated with the similarity difference. The greater the similarity difference, the farther the distance of single movement (the moving step is set to 0.01, which can be adjusted according to the convergence efficiency). The above adjustment process continues until the KL divergence converges (the number of iterations is set to 1000, and if the KL divergence change amount of continuous 50 iterations is less than the convergence threshold 0.001, it is also determined to be converged), and finally the 2-dimensional coordinates corresponding to each time slice data are obtained. The 2-dimensional coordinates of all time slice data are summarized to form a low-dimensional data point set corresponding to the first sub-area. Each point in the point set corresponds to a time slice state in the high-dimensional space, and the local correlation characteristics of the sub-area characteristics are retained.

[0051] Based on the low-dimensional data point set corresponding to the first sub-region, a triangular patch is formed by selecting every three adjacent low-dimensional data points, and the normal vector direction of each triangular patch is calculated; a plurality of triangular patches are spliced into a continuous surface to obtain a low-dimensional manifold corresponding to the first sub-region, representing the local data structure characteristics of the first sub-region, denoted as the manifold of the first sub-region after reconstruction, which specifically comprises the following steps: first, for each point in the low-dimensional data point set, the K nearest neighbor algorithm (K value is set to 5) is used to find the nearest 5 adjacent points, and the distance is calculated by using the Euclidean distance in 2D space; second, for each point and its adjacent points, a Delaunay triangulation algorithm is used to construct a triangular patch, which ensures that the internal angles of the generated triangle are greater than 30 degrees, avoiding the generation of narrow triangles, and the specific rules are as follows: three adjacent points that are not collinear are selected to form a triangle, and it is ensured that the circumcircle of the triangle does not contain other low-dimensional data points, and all points are iterated to generate an initial triangular patch set; third, the normal vector of the triangular patch is calculated, which is used to represent the spatial orientation of the patch and reflect the trend of the change of the sub-region characteristics, and the calculation steps are as follows: taking the coordinates of the three vertices of the triangular patch A (x1, y1), B (x2, y2), and C (x3, y3), vectors AB and AC are constructed, the coordinates of vector AB are (x2-x1, y2-y1), and the coordinates of vector AC are (x3-x1, y3-y1); since it is a 2D coordinate, it is extended to a 3D space (z coordinate is set to 0), and the z component of the cross product result = (x component of AB * y component of AC) - (y component of AB * x component of AC), the positive and negative of the z component represent the direction of the normal vector (positive for upward, negative for downward); the cross product result is divided by its module length to obtain the unit normal vector, ensuring that the magnitudes of the normal vectors of all patches are uniform; fourth, the generated triangular patches are spliced, and the included angle between the normal vectors of adjacent patches is calculated before splicing, which is used to judge whether the spatial orientations of the two patches are consistent, and the core calculation principle is to convert it into an angle value through the dot product operation of the normal vector, and the specific calculation steps and judgment logic are as follows: the normal vectors of adjacent patches are unit normal vectors (module length is 1, direction represents the orientation of the patch), and the unit normal vector of patch A is vector n1 (x1, y1, z1), and the unit normal vector of adjacent patch B is vector n2 (x2, y2, z2); since the patches are in the same plane in the 2D manifold reconstruction, the z component of the normal vector is the core of the orientation judgment (positive z represents upward, negative z represents downward), first, the dot product of the two unit normal vectors is calculated, and the dot product result directly reflects the coincidence degree of the vectors; second, the dot product is converted into an included angle (unit: radian) through the inverse cosine function; third, the radian is converted into an angle, i.e. the included angle (angle) = included angle (radian) * (180 / π); based on the included angle result, the splicing operation is performed, if the included angle is less than 15 degrees (judged as consistent in orientation, belonging to continuous structure), then the two patches are spliced along the common edge, ensuring smooth transition of the splicing surface;If the included angle is greater than 15 degrees, it indicates that there is an obvious turning of the two pieces of the surface, at this time, it is necessary to backtrack to the low-dimensional data point set, check whether there is a missing point near the common edge of the two pieces of the surface which is not included in the calculation, if there is a missing point, take the missing point as the vertex, and construct a new transition triangular surface piece with the end points of the common edge of the two pieces of the surface respectively, realize the smooth connection of the direction through the transition surface piece, and eliminate the gap; if there is no missing point, adjust the selection range of the adjacent point of one of the surface pieces (temporarily increase the K value of K-neighborhood to 6), and generate a new surface piece matching the direction of the other surface piece; finally, all the surface pieces are spliced into a continuous surface without overlapping and gap, and the surface is the low-dimensional manifold corresponding to the first subzone, which is recorded as the manifold of the first subzone after reconstruction, and the geometric shape directly reflects the local data structure of the core characteristics of the subzone, for example, the protrusion of the surface corresponds to the peak value of the characteristic parameter (such as the period of sudden increase of the call intensity), the depression of the surface corresponds to the trough of the characteristic parameter (such as the period of empty load of the elevator), and the inclination direction of the surface corresponds to the trend of the characteristic parameter (such as the process of increasing the speed of the elevator with the increase of the call intensity).

[0052] For the second divided subzone, the multi-dimensional state vectors corresponding to all the key feature points contained in the second subzone are extracted, and the multi-dimensional state vectors corresponding to all the key feature points contained in the second subzone are processed by dimension reduction to obtain a low-dimensional data point set corresponding to the second subzone, which specifically includes: second subzone state vector extraction and dimension reduction First, according to the subzone-key feature point association table, the original running parameter sequence of all key feature points in the second subzone is extracted, if the second subzone contains 3 characteristic points of elevator 2 load, 5th floor call intensity and elevator 2 running floor, the time span is still 30 seconds (300 time slices), the dimension of the multi-dimensional state vector is 3*300=900, and the vector is combined in the order of elevator 2 load (1 to 300 bits), 5th floor call intensity (301 to 600 bits) and elevator 2 running floor (601 to 900 bits); then the standardization processing is performed, the average value and the standard deviation of each characteristic point parameter sequence are calculated, and the formula standardized value=(original parameter value-average value)÷standard deviation is used to complete the processing; the t-SNE algorithm is still used in the dimension reduction stage, if the second subzone is in the morning peak scene (the local density of data is high), the bandwidth parameter is determined according to the adaptive rule, that is, the average Euclidean distance of the nearest 10 adjacent points of each time slice data is calculated, the median of all average distances is taken, and 0.8 times of the median is taken as the final bandwidth parameter (for example, if the median is 0.6, the bandwidth parameter is set to 0.48); after calculating the high-dimensional similarity by the Gaussian kernel function and the low-dimensional similarity by the t-distribution kernel function, the low-dimensional coordinates are optimized with KL divergence convergence (1000 iterations or continuous 50 times change amount <0.001) as the target, and finally the low-dimensional data point set of the second subzone is obtained.

[0053] Based on the low-dimensional data point set corresponding to the second sub-region, a triangular patch is formed by selecting every three adjacent low-dimensional data points, and the normal vector direction of each triangular patch is calculated; a plurality of triangular patches are spliced into a continuous surface to obtain a low-dimensional manifold corresponding to the second sub-region, representing the local data structure characteristics of the second sub-region, denoted as the manifold of the reconstructed second sub-region, which specifically includes: the low-dimensional manifold reconstruction of the second sub-region is performed on the low-dimensional data point set of the second sub-region, and the K nearest neighbor algorithm is used for adjacent point selection, and the value of K remains unchanged at 5 because the number of characteristic points of the sub-region is 3 (medium-sized sub-region); after the triangular patches are generated by Delaunay triangulation, the unit normal vector of each patch is calculated according to the vertex vector construction, cross multiplication to obtain the z component, and standardization, for example, the patch vertices A (1.2, 3.1), B (1.5, 3.3), and C (1.3, 3.5), the 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 normalized normal vector is (0, 0, 1); when splicing the patches, the dot product of the normal vectors of adjacent patches is calculated and converted into an angle, and if the included angle is less than 15 degrees, the patches are directly spliced, and if the included angle is greater than 15 degrees, a transition patch is supplemented, and finally a continuous surface is formed, which is the low-dimensional manifold of the second sub-region, and the protrusion thereof may correspond to a scene in which the load of the elevator 2 and the call intensity are simultaneously increased.

[0054] For the third sub-region to the last sub-region, the above process is repeated to extract the multi-dimensional state vectors corresponding to all key feature points in each sub-region and perform dimension reduction processing to obtain a low-dimensional data point set corresponding to each sub-region, and based on the low-dimensional data point set corresponding to each sub-region, a triangular patch is formed by selecting every three adjacent low-dimensional data points, and the normal vector direction of each triangular patch is calculated; a plurality of triangular patches are spliced into a continuous surface to obtain a low-dimensional manifold corresponding to each sub-region, until the manifolds of all reconstructed sub-regions are obtained, which specifically includes: the manifold reconstruction of the third sub-region to the last sub-region is repeated for subsequent sub-regions according to the above process, and the reconstruction effect is guaranteed by three main adaptation points, one is the dimension adaptation of the state vector, the dimension is dynamically adjusted according to the number of characteristic points of the sub-region x 300, and the statistical value of the characteristic data of the sub-region itself is used for standardization; two is the dimension reduction parameter fine tuning, the bandwidth parameter of the low-peak sub-region (sparse data) is enlarged to 1.0 to retain the correlation by widening the sensitive range of the Gaussian kernel; three is the K value adjustment of the triangulation, the K value of the small sub-region (characteristic points ≤2) is reduced to 3 to avoid insufficient adjacent points, and the K value of the large sub-region (characteristic points ≥5) is maintained at 5 to ensure patch coverage; after the processing of all sub-regions is completed, a corresponding table of sub-region number-low-dimensional manifold-core feature association is formed, and the geometric form of each manifold corresponds to the elevator operation sub-state of the sub-region one by one.

[0055] In the embodiment, each sub-area is divided based on the topology connectivity of the key link, each sub-area is formed by core feature association, a core association clustering and sub-potential isolation analysis mode is realized, and interference of irrelevant features is avoided; local association of high-dimensional data is optimized and reserved through similarity distribution, low-dimensional data point sets and original features are ensured; low-dimensional manifolds are constructed through triangular patch splicing, dynamic association of high-dimensional features is converted into a visual continuous surface, a bulge of the surface corresponds to a peak value of a feature parameter, a flat area of the surface corresponds to a stable state of the feature, a change in a normal vector direction corresponds to a turning point of the feature trend, and the understanding cost of an operation and maintenance personnel is reduced; different sub-areas are processed differently, and the manifold reconstruction of each sub-area is adapted to the feature density and association characteristics of the sub-area.

[0056] In a preferred embodiment of the application, according to the manifold of each sub-area after reconstruction, 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 sub-region after reconstruction, the local covariance matrix corresponding to the first sub-region is calculated; based on the coordinate distribution of all low-dimensional data points on the manifold of the second sub-region after reconstruction, the local covariance matrix corresponding to the second sub-region is calculated, which specifically includes: the local covariance matrix is used to measure the correlation degree and discrete characteristics of the low-dimensional data points of the sub-region in the x and y dimensions of the 2-dimensional space, the larger the matrix element value is, the more significant the correlation or fluctuation of the corresponding dimension is, and the calculation is based on the low-dimensional data point coordinates of the sub-region manifold as the core, the process is unified and needs to be adapted to the data characteristics of the sub-region, and the first sub-region and the second sub-region are taken as examples for detailed description; first, each data point of the low-dimensional manifold is a 2-dimensional coordinate (x, y), and the total number of low-dimensional data points of a sub-region is E (E is consistent with the number of time slices, i.e. 300), all data points are arranged into two sequences according to the coordinate dimension, i.e. x dimension sequence and y dimension sequence, and the value of each sequence corresponds to the low-dimensional feature projection result of a time slice; second, the average values of the x dimension and the y dimension are calculated respectively as the center reference of data distribution, third, the x and y coordinates of each data point are subtracted by the average value of the corresponding dimension respectively to obtain the x deviation and the y deviation, all data points are converted into deviation data with the average value as the origin through this step, and the interference of absolute value on correlation calculation is eliminated; fourth, the local covariance matrix is constructed, i.e. the covariance matrix of 2-dimensional data is a 2×2 matrix, the matrix elements reflect the covariant relationship between dimensions, the calculation logic is the average value of the corresponding deviation product sum, the formula is as follows, the first row and the first column element (x-x covariance) of the matrix = the sum of squares of x deviations of all data points ÷ (the total number of data points -1), which reflects the fluctuation degree of x dimension itself; the first row and the second column element (x-y covariance) of the matrix = the sum of x deviation × y deviation of all data points ÷ (the total number of data points -1), which reflects the correlation strength of x and y dimensions; the second row and the first column element (y-x covariance) of the matrix is the same as the x-y covariance value, because the covariant relationship has symmetry; the second row and the second column element (y-y covariance) of the matrix = the sum of squares of y deviations of all data points ÷ (the total number of data points -1), which reflects the fluctuation degree of y dimension itself; the calculation process of the second sub-region is completely consistent with that of the first sub-region, and only the low-dimensional data point coordinate sequence of the second sub-region needs to be replaced to obtain the exclusive local covariance matrix based on the data distribution characteristics thereof; for example, the second sub-region is in the morning peak scene, the data points are more densely distributed, and the sum of squares of x and y dimension deviations is smaller, so the element value of the final covariance matrix will be closer to 0 than that of the sub-region in the low peak scene, reflecting that the characteristic fluctuation is more gentle.

[0058] Based on the coordinate distribution of all low-dimensional data points on the manifold of the third sub-region after reconstruction, the local covariance matrix corresponding to the third sub-region is calculated; for each of the fourth sub-region to the last sub-region after reconstruction, based on the coordinate distribution of all low-dimensional data points on the manifold of the corresponding sub-region, the local covariance matrix of the corresponding sub-region is calculated, which specifically includes: for the third sub-region to the last sub-region, the standardized process of data arrangement, dimension mean calculation, deviation value calculation and covariance matrix construction is sequentially performed, and two adaptation points need to be noted in the operation, one is data quantity adaptation, if the total number of data points of a sub-region is less than 300 (such as some low peak sub-regions excluding abnormal data points) due to weak feature correlation, the actual remaining data point number is used as the criterion for calculating the mean and covariance, and the total number of data points in the formula is replaced by the actual effective data point number, so as to ensure that the calculation result is not affected by abnormal values; the second is the dimension meaning adaptation, the x and y dimensions of the low-dimensional manifold of different sub-regions correspond to different high-dimensional feature correlations (such as some sub-regions x corresponding to elevator speed and y corresponding to call intensity, and some sub-regions x corresponding to load and y corresponding to energy consumption), but the calculation logic of the covariance matrix is irrelevant to the physical meaning represented by the dimension, and only needs to be uniformly processed according to the coordinate sequence; after sequentially completing the calculation of all sub-regions, a one-to-one correspondence relationship between the sub-region number and the local covariance matrix is formed.

[0059] Eigenvalue decomposition operation is performed on the local covariance matrix corresponding to the first sub-region, the local covariance matrix corresponding to the second sub-region, and the local covariance matrix corresponding to the last sub-region, and the maximum eigenvalue is extracted from each local covariance matrix as the principal eigenvalue, which specifically includes: first, for the local covariance matrix (set as matrix C) of each sub-region, the eigenvalue λ and the eigenvector v that satisfy C×v=λ×v are solved by linear algebra decomposition algorithm (such as QR decomposition method), wherein λ is the eigenvalue and v is the corresponding eigenvector; since the covariance matrix is a symmetric matrix, the same number of eigenvalues (2) as the matrix dimension will be obtained after decomposition, and the eigenvalues are all non-negative; second, the two eigenvalues obtained by decomposition are sorted according to the numerical value, 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 core feature correlation of the sub-region, and the larger the value is, the more concentrated the data points on the low-dimensional manifold are in the direction corresponding to the eigenvalue, that is, the coupling relationship of the features in the sub-region is more prominent, such as the correlation between elevator speed and call intensity plays a leading role in the sub-region.

[0060] The main eigenvalues of all main features are arranged in the order of the first sub-region, the second sub-region, and the last sub-region to form a state coupling degree sequence, which specifically includes: arranging in the order of sub-region 1, sub-region 2, sub-region 3, and the last sub-region to ensure that the sequence order is consistent with the logical order of the physical running scenarios corresponding to the sub-regions, such as the call response sub-region and the energy consumption control sub-region; the main eigenvalues of each sub-region are arranged in the above order to form a one-dimensional numerical sequence, which is named as the state coupling degree sequence. The core value of the sequence is to transform the high-dimensional and complex sub-region feature correlation into an ordered quantitative index. Each value in the sequence represents the feature coupling strength of the corresponding sub-region. The subsequent analysis of the sequence characteristics such as fluctuation and peak value can realize the analysis of the global running situation of the elevator (for example, if the value at a certain position in the sequence suddenly increases, it means that the feature correlation of the corresponding sub-region enters a strong coupling state, which may exist a running efficiency bottleneck).

[0061] In this embodiment, the covariance matrix is used to transform the correlation of the sub-region features into specific numerical values, and the extracted main eigenvalues further focus on the core correlation strength, so that the feature coupling characteristics of each sub-region are changed from qualitative description to quantitative index, avoiding the deviation of subjective judgment. The main eigenvalues are extracted by eigenvalue decomposition, which eliminates the interference of secondary variation directions, integrates the main eigenvalues into a sequence in order, simplifies the global analysis from multiple matrix parallel processing to single sequence feature analysis, and improves the analysis efficiency. The arrangement order of the state coupling degree sequence is consistent with the logical order of the physical running scenarios corresponding to the sub-regions. The time sequence characteristics of the sequence are combined with the spatial characteristics, which can directly reflect the dynamic changes of the elevator running state.

[0062] In a preferred embodiment of the present application, according to the state coupling degree sequence, a state correction factor representing the overall running situation is extracted, and the state correction factor is fused with the running parameters to determine the elevator running state judgment result, including:

[0063] The state coupling degree sequence is statistically analyzed to extract the mean of the state coupling degree sequence, the variance of the state coupling degree sequence, and the skewness feature of the state coupling degree sequence, and the mean of the state coupling degree sequence, the variance of the state coupling degree sequence, and the skewness feature of the state coupling degree sequence are combined to form a state correction factor representing the overall operation situation, specifically including: the state correction factor is a core composite index representing the global operation situation of the elevator, which is composed of the mean, variance, and skewness of the state coupling degree sequence, wherein the mean reflects the overall level of coupling strength, the variance reflects the coupling fluctuation characteristics, and the skewness reflects the asymmetry of the coupling distribution, and the combination of the three realizes comprehensive quantification of the global situation, and the specific process is as follows: first, the state coupling degree sequence is a one-dimensional numerical sequence, which is denoted as S=[s1, s2, s3, …, sL], wherein L represents the total number of sub-regions, and each s represents the characteristic coupling strength of the corresponding sub-region; second, the mean of the sequence is calculated, which is the average level of all values in the sequence and reflects the overall strength of the characteristic coupling of each sub-region of the elevator; third, the variance of the sequence is calculated, which reflects the fluctuation degree of the main characteristic values in the sequence and embodies the difference between the coupling strengths of the sub-regions, and the greater the fluctuation, the more prominent the imbalance of the sub-region situation, and the calculation steps are as follows: first, calculate the deviation, subtract the mean of the sequence from each main characteristic value to obtain a deviation sequence, then calculate the sum of the squares of the deviations, and finally the variance = sum of squares of deviations ÷ (total number of sub-regions -1); fourth, the skewness of the sequence is calculated, which reflects the asymmetric direction of the sequence distribution, and a positive value indicates that the larger main characteristic values in the sequence are more concentrated (some sub-regions have extremely strong coupling), and a negative value indicates that the smaller main characteristic values are more concentrated (some sub-regions have extremely weak coupling), and the skewness = (total number of sub-regions ÷ [(total number of sub-regions -1) × (total number of sub-regions -2)]) × (sum of cubes of deviations ÷ cube of standard deviation); fifth, the mean, variance, and skewness calculated are combined in the order of mean, variance, and skewness to form a three-dimensional feature vector, which is the state correction factor, and the vector contains three types of information of the strength level, fluctuation characteristics, and distribution form of the global coupling.

[0064] The state correction factor is fused with the elevator signal, real-time load and real-time position data in the original operating parameter set to generate fused operating feature information, specifically including: first, three types of core real-time data are extracted from the original operating parameter set, i.e. elevator signal (intensity of elevator request at each floor, unit: times / sec), real-time load (passenger weight in the current elevator, unit: kg), and real-time position (current floor and accurate position of the elevator, unit: m, with the ground floor as the 0 point); standardization processing is performed on the three types of data to eliminate dimensional differences, the standardized value of the elevator signal = (current elevator intensity - minimum value of the historical elevator intensity of the elevator) ÷ (maximum value of the historical elevator intensity of the elevator - minimum value of the historical elevator intensity of the elevator), which is mapped to the [0, 1] interval; the standardized value of the real-time load = (current load - minimum value of the rated load of the elevator 0) ÷ (rated load of the elevator - 0), which is mapped to the [0, 1] interval (for example, the rated load is 1000 kg, the current load is 500 kg, and the standardized value is 0.5); the standardized value of the real-time position = (current position - minimum position 0 of the elevator operation) ÷ (maximum position of the elevator operation - 0), which is mapped to the [0, 1] interval (for example, the elevator runs up to the 10th floor, the floor height is 3 m, the maximum position is 30 m, and the current position is at the 5th floor 15 m, and the standardized value is 0.5); second, in order to unify the value range of the correction factor and the real-time parameter, standardization processing is performed on the three-dimensional state correction factor, and the processing manner is consistent with that of the real-time parameter, the historical value range of the mean value, variance, and skewness is counted (through the past 30 days of elevator operation data), and each dimension is mapped to the [0, 1] interval according to the formula (current value - minimum value) ÷ (maximum value - minimum value); third, the fusion weight is set in combination with the elevator operation characteristics, the state correction factor reflects the global situation, and the weight accounts for 40%; the three types of real-time parameters directly reflect the current state of the elevator, and the total weight accounts for 60%, among which the elevator signal is most closely related to the operation state judgment, and the weight is set to 30%, the real-time load and the real-time position are each set to 15%, and the fusion formula is: fused feature value = (standardized mean value of the correction factor × 0.4 × (1 / 3)) + (standardized variance of the correction factor × 0.4 × (1 / 3)) + (standardized skewness of the correction factor × 0.4 × (1 / 3)) + (standardized elevator signal × 0.3) + (standardized real-time load × 0.15) + (standardized real-time position × 0.15), wherein the three dimensions of the correction factor each account for 1 / 3 of the total weight (i.e. 0.4 ÷ 3 ≈ 0.133), ensuring that the three-dimensional information is balanced in the fusion; the standardized data is substituted into the formula to obtain the fused feature value corresponding to each time slice, and the fused feature values of all time slices are summarized to obtain the fused operating feature information.

[0065] analyzing whether the fused running feature information contains a preset emergency evacuation feature mode, the emergency evacuation feature mode being defined as all floors generating a continuous down call signal in a continuous time window and the signal strength exceeding a preset threshold; if the fused running feature information contains the emergency evacuation feature mode, determining that the elevator is running in an emergency evacuation state; if the fused running feature information does not contain the emergency evacuation feature mode, determining that the elevator is running in a normal running state, to obtain an elevator running state judgment result, specifically including: first, combining building safety specifications and elevator running characteristics, defining the emergency evacuation feature mode as all floors generating a continuous down call signal in a continuous time window and the signal strength exceeding a preset threshold, wherein the key parameters are quantified as follows: the continuous time window is set to 10 seconds, corresponding to 100 time slices, all floors refer to all floors served by the elevator, for example, if a building has 10 floors, floors 1 to 10 all need to meet the conditions, excluding unopened or malfunctioning disabled floors; the continuous down call signal exists in each time slice of the 10-second time window for each floor, with no interruption; the signal strength preset threshold is determined by historical emergency evacuation drill data statistics, and is set to 3 times the average call intensity of the building during the daily peak period, to ensure the distinction between normal down calls and emergency evacuation calls; second, separating the down call signal intensity sequence of each floor (which has been standardized and needs to be restored to the actual intensity value through inverse standardization, the inverse standardization formula being actual intensity value = standardized value x (historical maximum value - historical minimum value) + historical minimum value) from the fused running feature information; third, verifying in the order of time window continuity, floor coverage integrity, and signal strength compliance, wherein when verifying the time window, a 10-second continuous time window is slid and intercepted, and it is checked whether there is a continuous down call signal in each time window; when verifying the floor coverage, it is checked whether there is a down call signal for all served floors in the window that meets the time continuity; when verifying the strength compliance, it is checked whether the actual intensity of the call signal of each floor exceeds the preset threshold for the window that meets the previous two conditions, the preset threshold being 3 times the average call intensity of the building during the daily peak period; fourth, outputting the state judgment result, if there is a time window that meets all three conditions, determining that the elevator is running in an emergency evacuation state; if no window meets all conditions after traversing all time windows, determining that the elevator is running in a normal running state; the judgment result needs to be updated in real time, and the matching verification is re-executed every 2 seconds to ensure a quick response to state changes.

[0066] The embodiment ensures that the fusion process is quantifiable and reproducible by standardizing and unifying the parameter value range and setting scientific weights in combination with the elevator operation characteristics; the quantified definition of the emergency evacuation mode avoids the subjectivity of experience judgment and improves the reliability of the judgment; the threefold verification of the time window, the floor coverage and the intensity threshold value captures the typical call characteristics of the emergency evacuation, can quickly distinguish the normal downward call from the emergency evacuation call, and strengthens the safety guarantee capability of the elevator operation; the fusion of the global situation reflected by the state correction factor and the real-time parameters can quickly capture the change of the operation state, and the judgment result can be directly used as the basis for adjusting the scheduling strategy of the elevator group control system, so that the scheduling scheme is more suitable for the real-time operation demand, and the elevator operation efficiency and service quality are improved.

[0067] In a preferred embodiment of the present application, based on the elevator operation state 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 double model library; the first elevator scheduling instruction or the second elevator scheduling instruction is executed to control the elevator group, including:

[0068] If it is judged that the elevator operation is in a normal operation state, the fused operation feature information is input to a first scheduling model in a pre-trained double model library to generate a first elevator scheduling instruction, specifically including: the double model library is a special model set trained and verified through historical data, wherein 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 association between features and scheduling indicators, such as the relationship between call intensity and elevator response time, and the neural network is responsible for mining nonlinear associations, such as the coupling relationship between elevator position, load and energy consumption. The model input is the fused operation feature information, including the standardized state correction factor and real-time parameters, and the output is a scheduling instruction set including elevator number, target floor, operation speed and other parameters; the construction and training process of the first scheduling model needs to follow the logic of data preparation, structure design and iterative optimization; in the model construction link, first, collect the normal operation data of the elevator in the past 12 months, covering the fused feature information from 7:00 to 23:00 every day, including 5 types of feature groups such as global situation, call, load, etc. These data are used as input data; at the same time, the scheduling effect indicators corresponding to the scheduling results are labeled as label data, which specifically include three core indicators of elevator response time, passenger average waiting time and unit time energy consumption; after completing data preparation, enter the model structure design stage, the gradient boosting tree part is composed of 100 CART decision trees, each tree is limited to 6 layers to avoid overfitting, and the input layer receives a 32-dimensional feature vector of 5 types of feature groups; the neural network part adopts a 3-layer fully connected structure, the number of input layer neurons is consistent with the output dimension of the gradient boosting tree, which is set to 16 dimensions, the hidden layer is divided into two layers, the first layer is configured with 32 neurons, the second layer is configured with 16 neurons, and the output layer is set with 8 neurons, corresponding to the predicted values of 8 types of basic scheduling parameters; the last step of structure design is the fusion layer design, which weights and sums the outputs of the gradient boosting tree and the neural network according to the weight of 4:6, integrates the linear and nonlinear feature mapping results in this way, and obtains the final feature mapping output.

[0069] After the model is built, it enters the training stage. First, the 120,000 collected data is divided into training set, validation set and test set according to the proportion of 7:2:1. Standardization processing is performed on the input features, and the standardization formula is standardized value=(original value-feature mean) / feature standard deviation. The label is normalized to map it to the [0, 1] interval. After data preprocessing, a multi-objective weighted loss function is used to integrate the prediction errors of the three scheduling effect indicators. The total loss calculation method is 0.4×|response time predicted value-real value|+0.4×|waiting time predicted value-real value|+0.2×|energy consumption predicted value-real value|, 0.4 is the weight of response time and waiting time. Because in the normal operating state, the core goal of elevator scheduling is to improve operating efficiency and reduce passenger waiting, these two indicators directly reflect the efficiency level, so they are given a higher weight. 0.2 is the weight of the energy consumption indicator. Energy optimization is a secondary goal of normal scheduling. Therefore, the weight is relatively low. The initial learning rate is set to 0.001, and the learning rate is attenuated to 0.9; training rounds are set to 500 rounds, and an early stopping mechanism is set, when the total loss of the validation set does not decrease for 20 consecutive rounds, the training is stopped to prevent model overfitting; after the training is completed, the model must go through the model verification and optimization link, and the performance of the model is verified with the test set, if it is found that the error of a certain dispatching scenario such as the early morning peak is higher, then 30,000 special data for this scene are supplemented for fine tuning; in the second step, the fused running feature information is structured and split according to the input dimension requirements of the first dispatching model, and finally five functional features are formed; specifically, the global situation group directly corresponds to the mean, variance and skewness of the state correction factor; the call feature group specially stores the call signal strength of each floor and the duration data of each signal; the load feature group covers the real-time load data of all elevators and the load rate calculated from the real-time load; the position feature group is responsible for recording the real-time position of each elevator and the running direction of the elevator; the historical dispatching group is used to save the execution feedback results of all dispatching instructions in the last 5 minutes, including instruction completion, execution deviation and other information; each feature group needs to complete normalization and recalibration according to the weight proportion set during pre-training, the specific formula for recalibration is, the calibrated feature value = the original fused feature value x the training weight proportion of the feature group, the weight proportions of the five feature groups have been clearly divided according to the normal dispatching requirements, the global situation group is 20%, the call feature group is 30%, the load feature group is 15%, the position feature group is 20%, and the historical dispatching group is 15%, the total weight is 100%, this weight distribution is just to highlight the core role of call features (directly related to passenger demand) and global situation (reflecting the overall running state) in dispatching decision-making; in the third step, the inference calculation link of the first dispatching model is entered, the whole inference process is divided into linear calculation and nonlinear calculation two closely connected stages, the linear calculation stage is responsible by gradient boosting tree, the core is to weight the sum of the calibrated features through 100 basic CART decision trees; because the accuracy of each decision tree on the validation set is different, the contribution of each tree to the final result also exists difference, therefore, it is necessary to allocate accuracy weight for each tree, the calculation method of the weight is, the accuracy of a single decision tree on the validation set, divided by the sum of the accuracy of 100 decision trees, the final linear calculation result is, multiply the output value of each decision tree by the corresponding accuracy weight, and then add all the products to get the total sum; the nonlinear calculation stage is completed by a 3-layer fully connected neural network, first, the linear calculation result is input into the network, and the feature mapping is performed through the ReLU activation function, the mapping formula is, the activated value = max(0, linear calculation result x weight matrix + bias term), wherein, the weight matrix is determined by repeated iteration and optimization during the training of the model, that is, combined with the structure of the 3-layer fully connected model (input layer 16 dimensions, first hidden layer 32 dimensions, second hidden layer 16 dimensions, output layer 8 dimensions), assuming that the weight matrix value range from input layer to first hidden layer is [-0.12, 0.12], the weight matrix from the first hidden layer to the second hidden layer ranges from -0.15 to 0.15, and the weight matrix from the second hidden layer to the output layer ranges from -0.2 to 0.2; the bias term is set to 0.01, which aims to prevent the gradient vanishing problem in the feature mapping process. After the nonlinear calculation, the neural network outputs a set of numerical results, which need to be converted into specific scheduling parameters executable by the elevator. For example, when the output value corresponding to the running priority is 0.8, it will be converted into the scheduling instruction of the highest priority. When the output value corresponding to the target floor is a certain value, the mapping rule based on the physical characteristics of the elevator floor is used to complete the matching. Specifically, the integer part of the neural network output value 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, and if the decimal part is less than 0.5, it is rounded down, while ensuring that the final floor number is within the range of the elevator service floors (excluding unopened or malfunctioning disabled floors). For example, the decimal part of the value 5.2 is 0.2, which is less than 0.5, corresponding to the actual 5th floor. The decimal part of the value 5.6 is 0.6, which is greater than or equal to 0.5, corresponding to the actual 6th floor. The decimal part of the value 4.1 is 0.1, which is less than 0.5, corresponding to the actual 4th floor. Through this explicit numerical discretization rule, the output value and the specific floor number are accurately matched. The fourth step is to integrate the scheduling parameters output by the model and generate the first elevator scheduling instruction in the fixed format of elevator number, scheduling type, core parameter, and execution time limit. All elevator scheduling instructions need to meet the non-conflict constraint, that is, there will be no situation where two elevators respond to the same floor call at the same time. If there is a conflict in the model output, the call intensity priority principle is used to adjust, and the scheduling instruction corresponding to the floor with higher call signal intensity is retained to ensure reasonable resource allocation.

[0070] If it is judged that the elevator operation is in an emergency evacuation state, the fused operation feature information is input to a second dispatching model in a pre-trained double model library to generate a second elevator dispatching instruction, specifically including: first, the core characteristics of the second dispatching model are determined, the model is constructed based on a deep Q network, the training data covers simulated evacuation data and historical exercise data of multiple emergency scenarios such as building fires and earthquakes, and the input features are based on the fused operation feature information, with the addition of three types of emergency scenario-specific features: safe passage location, floor personnel density estimation value, and emergency power status, and the output instruction prioritizes the three core needs of downlink evacuation, empty response, and nearby landing; the construction and training process of the second dispatching model is centered on the interaction between the reinforcement learning agent and the environment, and the decision-making strategy is optimized through continuous trial and error; in the model construction link, environment modeling is performed first, and a simulation environment for elevator emergency evacuation is built, which needs to include elevator operation physical rules such as speed limit and load limit, building structure information such as floor distribution and safe passage location, and personnel flow models such as evacuation speed and call behavior in emergency situations, and the environment state will be updated in real time with dispatching actions; the agent design takes the deep Q network as the core, the input layer receives a 42-dimensional feature vector composed of the original fused 32-dimensional features and 10-dimensional emergency-specific features; the hidden layer adopts the structure of two convolution layers and two fully connected layers, the convolution layers are configured with 64 and 32 convolution 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 dispatching actions; the definition of state and action is a key link in the construction, the state space is the real-time feature set in the emergency scenario, including elevator state, personnel distribution, environmental risk, etc.; the action space covers 12 types of specific dispatching actions such as immediately descending to the first floor and responding to calls after emptying the car, ensuring coverage of various emergency disposal needs.

[0071] The model training link first prepares the data, generates 50,000 multi-scenario emergency evacuation simulation data, including fire, earthquake and other scenarios, and collects 10,000 real data of elevator emergency drills to form a training data set to improve the generalization ability of the model; After completing the data preparation, an experience replay mechanism needs to be built. Specifically, an experience pool with a capacity of 100,000 is built. After the agent (deep Q network) executes a certain scheduling action in the emergency evacuation simulation environment, the current state of the environment is obtained, such as the personnel density of each floor, the position / weight of the elevator, the occupancy of the safety passage, the scheduling action performed, such as the nearest stop of the elevator, the priority downward, the immediate reward corresponding to the action, and the next environment state updated after the action is triggered (such as the change of floor personnel density, the adjustment of elevator running track), and the complete four-tuple composed of the four elements is continuously stored in the experience pool; 256 data are randomly extracted from the experience pool as batch samples each time the model is trained; The core form of the reward function is the total reward calculation method, and its detailed design is the key to guide the model to optimize towards the efficient evacuation and safe bottom line goal. After supplementing the intermediate reward node on the basis of the original, the total reward (i.e. the core formula of the reward function) is determined as 0.5 x (evacuated number / total number of the floor) + 0.3 x (1 - personnel gathering density / safety density threshold) - 0.2 x overload risk, wherein 0.5 is the weight of the evacuation efficiency index. In any emergency scenario, the first scheduling goal is to evacuate personnel to a safe area as soon as possible, and the index of the number of evacuated personnel / the total number of the floor can directly quantify the progress of the evacuation task of a single floor, so it is given the highest weight. 0.3 is the weight of the personnel gathering risk index, which is the core safety target next to the evacuation efficiency; 0.2 is the weight of the overload risk index. The overload of the elevator will directly lead to equipment failure, and even cause fatal accidents such as falling elevator, but compared with the group safety risk caused by the gathering of non-evacuated personnel, the overload risk of a single elevator has a smaller range of influence, so the weight is relatively low; The safety density threshold needs to be set in combination with the characteristics of the elevator service building. For places such as office buildings and shopping malls where personnel flow is frequent and instantaneous gathering probability is high, the safety density threshold is set to 0.8 person / square meter, and the gathering risk is controlled by strictly controlling the threshold; For residential buildings and other places where personnel distribution is fixed and peak gathering is low, the safety density threshold is relaxed to 0.6 person / square meter, taking into account safety and scheduling flexibility; When the actual personnel gathering density (calculation method: floor personnel number / floor area) of a certain floor exceeds the safety density threshold of the corresponding building type, the calculation result of 1 - personnel gathering density / safety density threshold will decrease simultaneously, directly leading to the decrease of the total reward value of the model. This quantitative feedback will force the model to actively adjust the scheduling strategy, such as preferentially sending empty elevators to high-density floors, extending the elevator door opening time to speed up personnel diversion, etc. Through the linkage of weight allocation and quantitative index, the dynamic balance of evacuation efficiency and safety risk is realized.

[0072] The training process needs to consider exploration and utilization. In the initial stage, an epsilon-greedy strategy is used with an initial exploration probability epsilon of 0.9, which is reduced by 0.05 every 100 iterations, with a minimum of 0.1. The RMSprop optimizer is used with a learning rate of 0.002 and 1000 training rounds, each containing 1000 environment interactions. The model is verified every 200 rounds during training using real emergency exercise data. If the evacuation time error exceeds 10%, the environment model parameters are optimized for that scenario and retrained. In the second step, the newly added emergency-specific features are quantified and weighted. The estimated personnel density is calculated by multiplying the elevator signal strength, the floor area, and the personnel density coefficient. The personnel density coefficient is set according to building type, with 0.15 people per square meter for office buildings and 0.1 people per square meter for residential buildings. The distance between the safe passage and the elevator is calculated as a straight line distance, which is the core basis for the priority stop weight. The closer the distance, the higher the weight. The priority stop weight = 1 ÷ (distance between safe passage and elevator + 0.1). Adding 0.1 avoids an infinite weight when the distance is 0. In the feature fusion stage, the emergency-specific features are fused with the original fusion features at a ratio of 60% for emergency features and 40% for original features. By increasing the weight of emergency features, the model focuses on key information in emergency scenarios and ensures that the model focuses on emergency core needs. In the third step, the model determines the optimal instruction by calculating the Q value (action value) of each potential dispatch action. Q value = immediate reward + discount factor × future reward expectation. The future reward expectation is a specific value calculated by the model based on the current environment state, such as floor personnel distribution, elevator operation status, trained parameters, and typical emergency evacuation procedures (up to 20 core evacuation actions). The calculation logic is as follows: First, according to the Q value table trained by the model, query the estimated reward corresponding to the environment state that may be entered after executing the current action (the calculation logic of this estimated reward is consistent with the total reward function defined earlier, and each step of the estimated reward is based on the Q value of the optimal action in that state). Then, according to the formula , where is the estimated reward of the first step of the subsequent action, is the estimated reward of the second step of the subsequent action, is the discount factor ( = 0.9), i.e., the estimated reward of the first step of the subsequent action multiplied by , the second step multiplied by , and so on until the 20th step; finally, the specific value of the future reward expectation is obtained through the summation formula; immediate reward = (number of evacuated people ÷ total number of people on the floor) × 0.7 - (elevator overload risk × 0.3), elevator overload risk = (real-time load - rated load × 0.8) ÷ rated load, and if the result is negative, take 0; the discount factor is set to 0.9, ensuring that the model takes into account both current and subsequent evacuation effects; during reasoning, the action with the maximum Q value is selected as the dispatching instruction, typical instructions such as immediately emptying the elevator car and stopping near the 1st floor safety passage; in the fourth step, the second elevator dispatching instruction is generated, the instruction format adds a safety prompt information and an emergency linkage signal field based on the first dispatching instruction, and the core parameters need to meet the emergency specifications, for example, the upper limit of the running speed is increased to 2 meters per second but not exceeding the rated speed of the elevator, the door opening delay is extended to 5 seconds, and the stopping floor is preferentially selected to be the 1st floor, the 3rd floor, and other designated floors close to the safety passage, and at the same time, the instruction needs to include a failure emergency backup plan.

[0073] The first elevator dispatching instruction or the second elevator dispatching instruction is issued to an elevator group control execution unit, and the elevator group is controlled according to the first elevator dispatching instruction or the second elevator dispatching instruction by the elevator group control execution unit, specifically including: converting the first or second elevator dispatching instruction from a model output format into a control signal format recognizable by the elevator group control execution unit, and after conversion, performing double-checking to ensure the validity of the instruction, and syntax checking focuses on checking whether the instruction format conforms to the communication protocol of the group control unit to avoid failure to parse due to format errors; logic checking checks whether the instruction exceeds the physical operation limit of the elevator, such as speed not exceeding the rated value and floor within the service range; if the checking fails, the model generates an instruction again, forming the first fault tolerance barrier; the instruction is issued in a dual-link mode of industrial Ethernet and wireless backup communication, the main link selects industrial Ethernet to ensure that the transmission delay is less than or equal to 100 milliseconds; the backup link uses 4G industrial wireless communication, and when the main link is interrupted, it automatically switches to the backup link within 50 milliseconds, ensuring uninterrupted transmission of the instruction through dual-link redundancy, and the instruction is sent one by one in the order of elevator number, and each elevator's instruction is attached with a unique identity, effectively avoiding confusion; after the group control execution unit receives the instruction, it first parses the control parameters of each elevator, and then sends control signals to the drive system, door system and display system of the elevator through 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 sheave diameter), wherein the reduction ratio uses the standard value 30 commonly used in the elevator industry (adapted to the transmission characteristics of the mainstream elevator traction system); the door system accurately controls the door machine opening and closing timing and duration according to the door opening delay parameter; the display system synchronously updates the elevator running state and safety prompt information; during execution, the group control execution unit collects the running data of the elevator in real time, including position, speed, load and other key parameters, and feeds back the execution status to the dispatching system once every 500 milliseconds; if there is an instruction execution deviation, the dispatching system generates a correction instruction and issues it, forming a closed-loop control of issuance, execution, feedback and correction, to ensure that the dispatching effect meets the expectations.

[0074] In this embodiment, the dispatching scheme is highly matched with the elevator running state through the design of dual model libraries, solving the scene adaptation limitation problem of single model; the global situation features, real-time running features and historical dispatching features are fused into the model, and the deep correlation between the features is mined through linear and nonlinear calculation, avoiding the one-sidedness of single parameter-based dispatching; dual communication links are used to avoid interruption of instruction transmission, double-checking is used to ensure the correctness of the instruction, real-time feedback and correction mechanisms are used to solve the execution deviation problem, reducing the risks of instruction transmission failure and execution deviation; the second dispatching model generates targeted instructions through feature enhancement and reward mechanism design, and quickly issues and executes them, improving the emergency response capability and safety guarantee level of the elevator.

[0075] The above is the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the principles described in the present application, can also be made several improvements and refinements, these improvements and refinements should also be considered the scope of protection of the present application.

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 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. 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. 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; 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, The system collects real-time operating parameters of the elevator group to obtain a multi-dimensional state vector, and generates a fused feature set based on a preset spatiotemporal mapping relationship, including: The system collects the timestamps and floor numbers of call signals from each floor in real time, the real-time load values ​​of each car, and the real-time position of each car in the hoistway to obtain the original set of operating parameters. The data collected in the same time slice from the original set of operating parameters are aligned according to three dimensions: call signal, real-time load, and real-time location, and then vectorized and combined to form a multi-dimensional state vector. The multidimensional state vector is input into a preset spatiotemporal mapping relationship for processing. 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 relationship of multiple cars from the multidimensional state vector according to the correspondence rules between floors and time. Trajectory data, intensity data, and collaborative data are normalized and spliced ​​together to form a fusion feature set.

3. The elevator energy-saving operation control system based on a large model according to claim 2, characterized in that, Based on the fused feature set, coordinate positioning is performed in a predefined state reference frame, and key feature points representing the core characteristics of the current operational state are selected to construct a closed multi-dimensional feature structure, including: Based on the type of each feature item in the fused feature set, the feature items are mapped to the corresponding coordinate axes in the predefined state reference system; Based on the numerical value of each feature item, determine the position of the feature item on the mapped coordinate axis, and obtain the coordinate set of all feature items; A density-based clustering algorithm is used to analyze the coordinate set to identify regions with densely distributed coordinate points; In each densely distributed region, the coordinate points located at the center of the region and whose corresponding feature items have an importance weight higher than a preset importance weight threshold are selected as key feature points. Connect any two key feature points to form an edge, and calculate the Pearson correlation coefficient of the original set of running parameters corresponding to the two key feature points connected by each edge within a set time length. Edges with a Pearson correlation coefficient higher than a preset connection threshold are retained, and all retained edges and the key feature points connected by the retained edges together form a closed multidimensional feature structure.

4. The elevator energy-saving operation control system based on a large model according to claim 3, characterized in that, Calculate the mutual information and dynamic correlation degree between the vertices connected by each edge in the multidimensional feature structure, and weight each edge according to the correlation degree to obtain the critical links, including: For each edge in the closed multidimensional feature structure, obtain the original set of running parameters corresponding to the two key feature points connected by each edge, and calculate the mutual information value of the original set of running parameters corresponding to the two key feature points within a set time span. Within a set time span, the time is divided into multiple consecutive sliding windows, and within each sliding window, the Pearson correlation coefficient of the original set of operating parameters corresponding to two key feature points is calculated. Calculate the average of the Pearson correlation coefficients within all sliding windows, and calculate the variance of the Pearson correlation coefficients within all sliding windows. Use the average and variance as a measure of dynamic correlation. The mutual information value and the dynamic correlation degree metric are normalized respectively 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. 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.

5. The elevator energy-saving operation control system based on a large model according to claim 4, 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.

6. The elevator energy-saving operation control system based on a large model according to claim 5, 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.

7. The elevator energy-saving operation control system based on a large model according to claim 6, 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.

8. The elevator energy-saving operation control system based on a large model according to claim 7, 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