CN-122009929-A - Elevator toughness operation and maintenance decision method and system

CN122009929ACN 122009929 ACN122009929 ACN 122009929ACN-122009929-A

Abstract

The application provides a method and a system for deciding toughness operation and maintenance of an elevator, which relate to the technical field of intelligent operation and maintenance of elevators and are used for realizing the method. The method comprises the steps of generating a standardized lightweight digital twin data packet by an edge node, enabling a cloud to migrate knowledge of a general teacher model to an edge student model through federal knowledge distillation, performing local fine tuning to generate a personalized health evolution index, and executing hierarchical triggering type cluster collaborative decision-making, wherein the hierarchical triggering type cluster collaborative decision-making comprises three layers of edge autonomous emergency, cloud rapid reasoning and periodic optimization. The application solves the problem of multi-source data fusion, realizes low-cost and high-efficiency deployment of the model, ensures the real-time performance of system response and the global performance of decision, and endows the system with strong toughness.

Inventors

ZHENG DUAN
LI JIAYU
HU JIANXIONG
WANG LI
ZHANG WEI
HU WEI

Assignees

日立电梯（成都）有限公司

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. The method for deciding the toughness operation and maintenance of the elevator is characterized by comprising the following steps of: S1, generating an edge side lightweight digital twin segment, namely monitoring a complete elevator operation segment at each edge node, and after the segment is finished, fusing multi-source sensing data in the segment with a local analysis result by the edge node to generate and upload a standardized lightweight digital twin data packet, wherein the data packet at least comprises a segment identifier, a start-stop time stamp, an operation mode label, a fused health feature vector with fixed dimension and a local early warning mark; S2, cloud federal knowledge distillation and model updating, wherein a cloud decision platform maintains a general health prediction teacher model based on historical data training; for a target elevator, migrating predicted knowledge of a teacher model to a lightweight student model through a federal knowledge distillation frame, issuing the student model to a corresponding edge node, and performing local fine tuning by utilizing a twin data packet generated by the node later so as to generate a personalized health evolution index of the elevator; S3, hierarchical triggering type cluster collaborative decision-making, which comprises the following steps: s31, an edge autonomous emergency layer, wherein an edge node immediately executes a preset emergency control instruction for a predefined highest-level local early warning; S32, a cloud rapid reasoning layer, wherein when the twin data packet uploaded by the edge node comprises an early warning mark, the cloud decision platform inputs the health evolution index, early warning information and current global resource state of the elevator into a pre-trained deep reinforcement learning decision network, and outputs preliminary treatment priority or dispatch advice in millisecond time; And S33, a cloud period optimization layer, namely starting a background global optimization algorithm periodically or when the accumulated tasks reach a threshold value, rescheduling and distributing resources for all tasks to be processed in a period of time, and updating the strategy of the deep reinforcement learning decision network by utilizing an optimization result.
2. The method according to claim 1, wherein in step S3, the hierarchical triggered cluster collaborative decision-making comprises: S31, an edge autonomous emergency layer immediately executes a preset emergency control instruction for pre-defined highest-level early warning identified by an edge node local algorithm, so that highest safety is ensured; S32, a cloud rapid reasoning layer, wherein when a data packet uploaded by an edge node contains a non-highest-level early warning mark, the cloud decision platform immediately triggers rapid decision, and inputs the health evolution index, specific early warning information and current global resource state of an elevator triggering early warning into a pre-trained deep reinforcement learning decision network; And S33, a cloud period optimization layer, wherein the cloud decision platform periodically or when the accumulated tasks to be processed reach a certain threshold value, a background global optimization algorithm is started, rescheduling and resource allocation are carried out by the algorithm based on all tasks in a longer time window, resource constraint and more complex optimization targets, and the new strategy obtained by optimization can be used as training data to update the strategy of the deep reinforcement learning decision network in S32 so that the strategy can make more optimal rapid decisions in the future.
3. The method according to claim 1, wherein in step S1, the generating process of the lightweight digital twin packet specifically includes: s11, the edge node takes the elevator door closing start as a starting point and takes the next door opening flat layer as an ending point to define an operation segment; S12, locally processing multisource data in real time, wherein the multisource data comprises the steps of carrying out feature extraction on acceleration data acquired by a vibration sensor to obtain vibration root mean square, vibration peak value, specific frequency band energy and vibration short-time energy features, carrying out feature extraction on current data acquired by a current sensor to obtain current effective values, current harmonic distortion rate and current imbalance degree features, carrying out feature extraction on rotating speed data acquired by an encoder to obtain real-time rotating speed and rotating speed fluctuation rate features, carrying out feature extraction on data acquired by a temperature sensor to obtain real-time temperature values and temperature rise rate features, carrying out local analysis on video streams acquired by a camera to only output anonymous passenger counting and behavior statistical features, carrying out feature extraction on audio streams acquired by a microphone to obtain mel frequency cepstrum coefficient features, carrying out feature extraction on data acquired by a photoelectric sensor and a temperature and humidity sensor to obtain transmission light intensity, smoke absorbance, temperature and humidity features, carrying out feature extraction on data acquired by a flat layer sensor to obtain flat layer error values, carrying out feature extraction on data read by an elevator main controller to obtain instruction response delay time and running state features, carrying out feature extraction on data acquired by a symmetrical heavy load carrier controller, carrying out feature extraction on current door opening and closing time and motion state feature, and carrying out feature extraction on motion door opening and closing state feature in a hoistway door, and controlling the motion state feature extraction process, so as to obtain real-time response time and real-time feature; S13, after the segment is finished, inputting all the extracted features, together with metadata comprising operation start-stop floors, time and energy consumption logs, into a feature fusion coding network which is pre-deployed at the edge, and outputting a low-dimension dense fusion health feature vector; And S14, packaging fragment metadata, fusing health feature vectors and local early warning marks to form a standardized twin data packet and uploading the standardized twin data packet to a cloud decision platform.
4. The elevator toughness operation and maintenance decision method according to claim 3, wherein the feature fusion encoding network is a small fully-connected neural network or a convolutional neural network, and is cured and deployed to edge nodes after pre-training on a cloud decision platform.
5. The method for deciding toughness operation and maintenance of an elevator according to claim 3, wherein in step S14, the local early warning mark is generated by an edge node in a segment processing process, and the generating manner includes comparing a characteristic value extracted in real time with an early warning rule threshold value preset in the node, and if the characteristic value exceeds the threshold value, generating a local early warning mark of a corresponding grade and type.
6. The method for determining toughness operation and maintenance of an elevator according to claim 1, wherein in step S2, the specific flow of federal knowledge distillation is as follows: S21, the cloud decision platform utilizes a large number of accumulated historical twin data packets and corresponding actual maintenance records to train a health prediction teacher model offline; S22, for a target elevator, the cloud decision platform starts a federal knowledge distillation process: The cloud decision platform initializes a lightweight student model; The cloud decision platform performs reasoning on a group of generalized twin data by using the healthy prediction teacher model, and the obtained output probability distribution is used as a soft label; the cloud decision platform sends the initialized student model and the soft label to an edge node of a target elevator; the edge node of the target elevator performs preliminary training on the student model by utilizing the received soft label and a small amount of initial twin data possibly existing locally, so that the student model learns the generalization prediction capability of the teacher model; S23, issuing the preliminarily trained student model to a target edge node, wherein after the node generates a new twin data packet, the node takes the fused health feature vector as input, and performs incremental fine adjustment on the student model to enable the student model to gradually adapt to the specific running mode and degradation rule of the elevator; And S24, periodically uploading the parameters subjected to fine adjustment by each edge node to a cloud decision-making platform, and aggregating the parameters by the cloud decision-making platform for updating the teacher model to realize the continuous evolution of knowledge in the group.
7. The elevator toughness operation and maintenance decision method according to claim 1, wherein the health evolution index calculation process is as follows: ① Real-time calculation, namely, when each operation segment is finished, the edge node fuses health feature vectors generated by the segment Inputting a local student model Obtain the original output , Is a constant; ② The index is generated by a preset linear mapping function Converting the original output into visual index value, said health evolution index The calculation formula of (2) is as follows: , Wherein, the For the preset scaling factor to be used, The offset is preset, and the offset and the model output are used for normalizing the model output to be in the interval of [0, 100 ]; the piecewise variation of (a) constitutes the healthy evolution track of the elevator.
8. The elevator flexible operation and maintenance decision method according to any one of claims 1 to 7, further comprising step S4: The edge node extracts localized characteristics of the original audio and video data, only uses the abstract statistical characteristics which are irreversible and can not identify personal identities to generate the operation mode label, and the original audio and video data is destroyed immediately after the local characteristics are extracted and is not stored and uploaded.
9. The method for determining toughness operation and maintenance of an elevator according to claim 8, wherein the localization feature extraction is specifically that a video analysis module operates in an edge device memory, converts an image frame into mathematical features of an optical flow field, edge density and contour number in real time, and discards an original image frame immediately.
10. An elevator flexible operation and maintenance decision system for implementing the method of any of claims 1-9, the system comprising: A plurality of edge nodes, each node configured to perform step S1 of claim 1; The cloud decision platform is configured to perform steps S2, S32 and S33 in claim 1, and is communicatively connected to the edge node.

Description

Elevator toughness operation and maintenance decision method and system Technical Field The invention relates to the technical field of intelligent operation and maintenance of elevators, in particular to a method and a system for deciding toughness operation and maintenance of an elevator. Background The existing intelligent elevator operation and maintenance system generally adopts an architecture of edge collection and cloud decision platform analysis, and aims to realize fault early warning and preventive maintenance (as shown in a patent document with the application number of CN 202511015676.2). However, in evolving towards more accurate and efficient clustered intelligent operations and maintenance, several core technical hurdles are faced: 1. The multi-source heterogeneous data fusion is difficult, and mechanical sensing data, video behavior data and operation logs of the elevator are required to be associated for realizing accurate maintenance. The data has huge difference in frequency, transmission delay and format, and the realization of high-precision space-time alignment in a distributed system is extremely complex, which is very easy to cause the distortion of subsequent feature analysis and risk assessment. 2. The intelligent model deployment and updating cost is high, and establishing a personalized health prediction model based on machine learning for each elevator is a key of accurate operation and maintenance. But new elevators or elevators after overhaul lack historical data and face the difficult problem of cold start, and the required data, calculation force and engineering cost are prohibitive for training and deploying high-quality models for thousands of types of elevators with different working conditions respectively. 3. When the number of elevators is increased suddenly, the cloud decision platform center needs to process mass data in real time and solve complex multi-objective optimization problems (such as maintenance task scheduling), the traditional optimization algorithm is time-consuming to calculate, the minute-level response requirement of an emergency fault is difficult to meet, and the system expansibility is limited. 4. Data privacy and security compliance risk are prominent-schemes that rely on video analysis of passenger behavior to evaluate usage strength face significant compliance pressure and user trust challenges under increasingly stringent personal data protection regulations. Therefore, a new generation of operation and maintenance architecture capable of realizing accurate health prediction and real-time toughness decision of elevator clusters with low cost and high efficiency on the premise of protecting privacy is needed. Disclosure of Invention In view of the above, the present invention provides a method for determining toughness operation and maintenance of an elevator, which aims to overcome at least one of the drawbacks of the prior art. In order to achieve the above purpose, the present invention provides the following technical solutions: in a first aspect, the invention provides an elevator toughness operation and maintenance decision method, comprising the following steps: S1, generating an edge side lightweight digital twin segment, namely monitoring a complete elevator operation segment at each edge node, and after the segment is finished, fusing multi-source sensing data in the segment with a local analysis result by the edge node to generate and upload a standardized lightweight digital twin data packet, wherein the data packet at least comprises a segment identifier, a start-stop time stamp, an operation mode label, a fused health feature vector with fixed dimension and a local early warning mark; S2, cloud federal knowledge distillation and model updating, wherein a cloud decision platform maintains a general health prediction teacher model based on historical data training; for a target elevator, migrating predicted knowledge of a teacher model to a lightweight student model through a federal knowledge distillation frame, issuing the student model to a corresponding edge node, and performing local fine tuning by utilizing a twin data packet generated by the node later so as to generate a personalized health evolution index of the elevator; S3, hierarchical triggering type cluster collaborative decision-making, which comprises the following steps: s31, an edge autonomous emergency layer, wherein an edge node immediately executes a preset emergency control instruction for a predefined highest-level local early warning; S32, a cloud rapid reasoning layer, wherein when the twin data packet uploaded by the edge node comprises an early warning mark, the cloud decision platform inputs the health evolution index, early warning information and current global resource state of the elevator into a pre-trained deep reinforcement learning decision network, and outputs preliminary treatment priority or dispatch advice in millisecond time; And S33, a clou