Decision-making method and system for flexible operation and maintenance of elevator

Abstract

The invention provides an elevator flexible operation and maintenance decision-making method and system, relates to the technical field of elevator intelligent operation and maintenance, and is used for realizing the method. The method comprises the following steps: an edge node generates a standardized lightweight digital twin data packet; the cloud end migrates knowledge of the general teacher model to the edge student model through federal knowledge distillation, and local fine tuning is carried out to generate a personalized health evolution index; and executing layered trigger type cluster collaborative decision-making which comprises three layers of edge autonomous emergency, cloud fast reasoning and periodic optimization. According to the method, the problem of multi-source data fusion is solved, low-cost and efficient deployment of the model is realized, the real-time performance of system response and the globality of decision making are guaranteed, and the system is endowed with strong toughness.

Description

Technical Field

[0001] This invention relates to the field of intelligent elevator operation and maintenance technology, and in particular to a method and system for elevator resilience operation and maintenance decision-making. Background Technology

[0002] Existing intelligent elevator operation and maintenance systems typically adopt an architecture of "edge data acquisition + cloud-based decision-making platform analysis," aiming to achieve fault early warning and preventative maintenance (as shown in the patent document with application number CN202511015676.2). However, in the process of evolving towards more precise and efficient clustered intelligent operation and maintenance, several core technical obstacles are encountered:

[0003] 1. Difficulty in fusing multi-source heterogeneous data: To achieve accurate maintenance, it is necessary to link the elevator's mechanical sensor data, video behavior data, and operation logs. These data vary greatly in terms of generation frequency, transmission latency, and format, making high-precision spatiotemporal alignment in a distributed system extremely complex and prone to distortion in subsequent feature analysis and risk assessment.

[0004] 2. High Cost of Deploying and Updating Intelligent Models: Establishing a personalized health prediction model based on machine learning for each elevator is key to accurate operation and maintenance. However, new elevators or elevators that have undergone major repairs lack historical data and face the challenge of a "cold start." Training and deploying high-quality models for thousands of elevator models with different operating conditions requires prohibitive data, computing power, and engineering costs.

[0005] 3. The contradiction between real-time performance and scalability in cluster decision-making: When the number of elevators surges, the cloud-based decision-making platform center needs to process massive amounts of data in real time and solve complex multi-objective optimization problems (such as maintenance task scheduling). Traditional optimization algorithms are computationally time-consuming and cannot meet the minute-level response requirements for emergency faults, thus limiting the system's scalability.

[0006] 4. Data privacy and security compliance risks become prominent: Solutions that rely on video analysis of passenger behavior to assess usage intensity face enormous compliance pressure and user trust challenges under increasingly stringent personal data protection regulations.

[0007] Therefore, there is an urgent need for a new generation of operation and maintenance architecture that can achieve accurate health prediction and real-time resilience decision-making for elevator clusters at low cost and high efficiency while protecting privacy. Summary of the Invention

[0008] In view of the above, the present invention provides an elevator resilience operation and maintenance decision-making method, which aims to overcome at least one of the defects of the prior art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] In a first aspect, the present invention provides an elevator resilience operation and maintenance decision-making method, comprising:

[0011] S1. Lightweight Digital Twin Segment Generation at the Edge: At each edge node, a complete elevator operation segment is monitored; after the segment ends, the edge node fuses the multi-source sensor data and local analysis results within the segment to generate and upload a standardized lightweight digital twin data package; the data package includes at least the segment identifier, start and end timestamps, operation mode label, fixed-dimensional fused health feature vector, and local early warning marker;

[0012] S2, Cloud-based Federated Knowledge Distillation and Model Update: The cloud-based decision platform maintains a general health prediction teacher model trained on historical data; for the target elevator, the predicted knowledge of the teacher model is transferred to a lightweight student model through a federated knowledge distillation framework; the student model is distributed to the corresponding edge node and fine-tuned locally using the twin data package subsequently generated by the node to generate the personalized health evolution index of the elevator.

[0013] S3, hierarchical trigger-based cluster collaborative decision-making, including:

[0014] S31, Edge Autonomous Emergency Layer: When an edge node receives a predefined highest-level local early warning, it immediately executes pre-set emergency control commands.

[0015] S32, Cloud-based Fast Inference Layer: When the twin data packet uploaded by the edge node contains a warning marker, the cloud decision platform inputs the elevator's health evolution index, warning information, and current global resource status into the pre-trained deep reinforcement learning decision network, and outputs preliminary handling priorities or dispatch suggestions within milliseconds.

[0016] S33, Cloud-based Periodic Optimization Layer: The cloud-based decision platform periodically or when the accumulated tasks reach a threshold, initiates a background global optimization algorithm to reschedule and allocate resources for all pending tasks within a certain period, and uses the optimization results to update the strategy of the deep reinforcement learning decision network.

[0017] In some embodiments of the present invention, step S3, the hierarchical triggered cluster collaborative decision-making, includes:

[0018] S31, Edge Autonomous Emergency Layer: For the highest level of warning identified by the local algorithm of the edge node, the preset emergency control instructions are immediately executed to ensure the highest level of security;

[0019] S32, Cloud-based Fast Inference Layer: When the data packet uploaded by the edge node contains a warning marker that is not the highest level, the cloud decision platform immediately triggers a fast decision. It inputs the health evolution index of the elevator that triggered the warning, the specific warning information, and the current global resource status into a pre-trained deep reinforcement learning decision network. The network outputs a preliminary handling suggestion within milliseconds, such as assigning a priority to the event or recommending which maintenance personnel to assign the task to.

[0020] S33, Cloud-based Periodic Optimization Layer: The cloud-based decision platform periodically or when the accumulated pending tasks reach a certain threshold, initiates a background global optimization algorithm. This algorithm reschedules and allocates resources based on all tasks, resource constraints, and more complex optimization objectives within a longer time window. The optimized new strategy can be used as training data to update the strategy of the deep reinforcement learning decision network in S32, enabling it to make better and faster decisions in the future.

[0021] In some embodiments of the present invention, the generation process of the lightweight digital twin data package specifically includes:

[0022] S11: An edge node defines a running segment, starting from the elevator's first door closing and starting, and ending with the next door opening and leveling.

[0023] S12: Within the segment, local real-time processing of multi-source data: feature extraction is performed on acceleration data collected by vibration sensors to obtain vibration root mean square, vibration peak value, specific frequency band energy, and short-time vibration energy characteristics; feature extraction is performed on current data collected by current sensors to obtain current RMS value, current harmonic distortion rate, and current imbalance characteristics; feature extraction is performed on rotational speed data collected by encoders to obtain real-time rotational speed and rotational speed fluctuation rate characteristics; feature extraction is performed on data collected by temperature sensors to obtain real-time temperature value and temperature rise rate characteristics; local analysis is performed on video streams collected by cameras, outputting only anonymized passenger counts and behavioral statistics. Features were extracted from the audio stream collected by the microphone to obtain the Mel-frequency cepstral coefficient features; features were extracted from the data collected by the photoelectric sensor and the temperature and humidity sensor to obtain the transmitted light intensity, smoke absorbance, shaft temperature and humidity features; features were extracted from the data collected by the leveling sensor to obtain the leveling error value; features were extracted from the data read by the elevator main controller to obtain the command response delay time and operating status features; features were extracted from the data collected by the load cell to obtain the real-time load and load change rate features; and features were extracted from the data read by the door operator controller to obtain the door opening and closing time and the resistance current during the door movement process.

[0024] S13: After the segment ends, all extracted features, along with metadata including the start and end floors, time, and energy consumption logs, are input into a feature fusion encoding network pre-deployed at the edge, and output a low-dimensional, dense fused health feature vector.

[0025] S14: Package fragment metadata, integrate health feature vectors and local early warning markers to form a standardized twin data package and upload it to the cloud decision-making platform.

[0026] In some embodiments of the present invention, the feature fusion encoding network is a small fully connected neural network or convolutional neural network, which is pre-trained on a cloud decision platform and then deployed to edge nodes.

[0027] In some embodiments of the present invention, in step S14, the local warning marker is generated by the edge node during the segment processing. The generation method includes: comparing the real-time extracted feature values ​​with a preset warning rule threshold within the node; if the threshold is exceeded, a local warning marker of the corresponding level and type is generated.

[0028] In some embodiments of the present invention, the specific process of federated knowledge distillation is as follows:

[0029] S21: Cloud-based decision-making platform. The cloud-based decision-making platform utilizes a large amount of accumulated historical twin data packages and corresponding actual maintenance records to train a health prediction teacher model offline.

[0030] S22: For the target elevator, the cloud-based decision-making platform initiates the federated knowledge distillation process:

[0031] The cloud-based decision-making platform initializes a lightweight student model;

[0032] The cloud-based decision-making platform infers the health prediction teacher model on a set of generalized twin data, and the resulting output probability distribution is used as a soft label.

[0033] The cloud-based decision-making platform sends the initialized student model and the soft tags to the edge node of the target elevator;

[0034] The edge nodes of the target elevator use the received soft tags and a small amount of initial twin data that may exist locally to perform preliminary training on the student model, enabling it to learn the generalization prediction ability of the teacher model.

[0035] S23: The initially trained student model is sent to the target edge node; thereafter, each time the node generates a new twin data package, it uses the fused health feature vector in it as input to incrementally fine-tune the student model, so that it gradually adapts to the unique operating mode and degradation pattern of this elevator; the fine-tuned student model can output the health evolution index of the elevator in real time.

[0036] S24: Each edge node can periodically upload the fine-tuned parameters of the student model to the cloud decision platform; the cloud decision platform aggregates these parameters to update the teacher model, enabling the continuous evolution of knowledge in the group.

[0037] In some embodiments of the present invention, the calculation process of the health evolution index is as follows:

[0038] ① Real-time computation: At the end of each running segment, the edge node will generate a fused health feature vector for that segment. Input local student model , to obtain the original output , It is a constant;

[0039] ② Exponential generation: Subsequently, through a pre-defined linear mapping function... The raw output is converted into intuitive index values; the health evolution index The calculation formula is:

[0040] ,

[0041] in, The preset scaling factor. These are preset offsets used to normalize the model output to the range [0, 100]. The gradual changes in each segment constitute the health evolution trajectory of the elevator.

[0042] In some embodiments of the present invention, step S4 is also included:

[0043] Privacy protection and compliance processing: Edge nodes perform localized feature extraction on the original audio and video data, and only use irreversible and unidentifiable abstract statistical features to generate the operating mode label. The original audio and video data is destroyed immediately after local feature extraction and is not stored or uploaded.

[0044] In some embodiments of the present invention, the localized feature extraction specifically involves: the video analysis module running in the memory of the edge device converting image frames into mathematical features such as optical flow field, edge density, and contour quantity in real time, and immediately discarding the original image frames.

[0045] Secondly, the present invention provides an elevator resilience operation and maintenance decision-making system for implementing the above-mentioned method, the system comprising:

[0046] Multiple edge nodes, each configured to execute step S1;

[0047] The cloud-based decision-making platform is configured to execute steps S2, S32, and S33 and communicate with the edge nodes.

[0048] The embodiments of the present invention have at least the following advantages or beneficial effects:

[0049] 1. Solves the fundamental problem of multi-source data fusion and alignment: By defining a standardized “lightweight digital twin data package”, the complex spatiotemporal alignment and feature fusion process is pushed down to the edge with a precise local clock. The cloud decision platform only needs to process high-value data with unified structure and abstract information, which greatly improves system reliability and data processing efficiency.

[0050] 2. Enabled low-cost, high-efficiency, and universally accessible deployment of intelligent models: Through the "Federated Knowledge Distillation" architecture, new elevators or elevators lacking data can quickly obtain a lightweight model with prior knowledge, solving the cold start problem; at the same time, it protects the data privacy of each elevator and enables the sharing of group experience and continuous model evolution.

[0051] 3. Ensures real-time system response and global optimality of decision-making: The "layered trigger-based decision-making mechanism" combines millisecond-level edge autonomous response, second-level cloud-based intelligent recommendation, and hourly-level global optimization planning. This not only meets the immediate handling needs of emergency faults but also achieves near-optimal real-time scheduling through deep reinforcement learning, resulting in strong system scalability.

[0052] 4. It endows the system with strong "resilience": Even if the cloud decision-making platform network is temporarily interrupted, the edge nodes can still maintain a certain degree of independent assessment and emergency response capabilities based on the local student model and twin data. After the cloud decision-making platform is restored, it can be seamlessly synchronized, and the overall system's resilience and availability are significantly enhanced.

[0053] Other features and advantages of the present invention will be set forth in the following description. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating the elevator resilience operation and maintenance decision-making method in Example 1. Detailed Implementation

[0056] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention.

[0057] The embodiments of the present invention will be described in detail below.

[0058] Example 1

[0059] Firstly, see [the following] Figure 1 This embodiment provides an elevator resilience operation and maintenance decision-making method, applied to a system containing multiple edge nodes and a cloud decision-making platform. The method includes:

[0060] S1: Lightweight Digital Twin Segment Generation at the Edge: At each edge node, a complete elevator operation segment is monitored; after the segment ends, the edge node fuses the multi-source sensor data and local analysis results within the segment to generate and upload a standardized lightweight digital twin data package; the data package includes at least a unique segment identifier, start and end timestamps, operation mode label, a fixed-dimensional fused health feature vector, and a local early warning marker;

[0061] S2: Cloud-based Federated Knowledge Distillation and Model Update: The cloud-based decision platform maintains a general health prediction teacher model trained on a large number of historical twin data packages; for the target elevator (new elevator / specific elevator with scarce data), the cloud-based decision platform transfers the predictive knowledge of the health prediction teacher model to a lightweight student model through a federated knowledge distillation framework; the student model is distributed to the corresponding edge node and fine-tuned locally using the twin data packages subsequently generated by the node to generate the personalized health evolution index of the elevator;

[0062] S3: Layered trigger-based cluster collaborative decision-making, including:

[0063] S31: Edge Autonomous Emergency Layer: Edge nodes immediately execute preset emergency control commands upon receiving a predefined highest-level local early warning;

[0064] S32: Cloud-based fast inference layer: When the twin data packet uploaded by the edge node contains a warning marker, the cloud decision platform inputs the elevator's health evolution index, warning information, and current global resource status into a pre-trained deep reinforcement learning decision network, and outputs preliminary handling priorities or dispatch suggestions within milliseconds.

[0065] S33: Cloud-based periodic optimization layer: The cloud-based decision platform periodically or when the accumulated tasks reach a threshold, activates the background global optimization algorithm to reschedule and allocate resources for all pending tasks within a certain period of time, and uses the optimization results to update the strategy of the deep reinforcement learning decision network.

[0066] S4: Privacy protection and compliance processing; edge nodes perform localized feature extraction on the original audio and video data, and only use irreversible and unidentifiable abstract statistical features to generate the operating mode label. The original audio and video data is destroyed immediately after local feature extraction and is not stored or uploaded.

[0067] Secondly, this embodiment provides an elevator resilience operation and maintenance decision-making system for implementing the following method, the system comprising:

[0068] Multiple edge nodes, each configured to execute step S1;

[0069] The cloud-based decision-making platform is configured to execute steps S2, S32, and S33 of claim 1 and to communicate with the edge node.

[0070] More specifically, edge nodes are deployed at the site of each elevator and are typically embedded devices or industrial gateways with certain computing, storage, and communication capabilities. They are responsible for connecting various elevator sensors (such as vibration, current, temperature, cameras, microphones, etc.) and controllers, performing data acquisition, local feature extraction, lightweight digital twin fragment generation, local model inference, and autonomous emergency control with the highest level of early warning.

[0071] The cloud-based decision-making platform is deployed on a remote server cluster. It is responsible for receiving and storing twin data packets from each edge node, maintaining and updating the global model (the aforementioned health prediction teacher model), performing federated knowledge distillation to generate and distribute student models, running a deep reinforcement learning decision network for rapid inference, and performing background global optimization tasks.

[0072] Edge nodes communicate with the cloud decision-making platform via wired or wireless networks (such as 4G / 5G, Ethernet) to transmit standardized twin data packets, model parameters, and control commands.

[0073] The above methods and systems have at least the following beneficial effects:

[0074] 1. Solves the fundamental problem of multi-source data fusion and alignment: By defining a standardized “lightweight digital twin data package”, the complex spatiotemporal alignment and feature fusion process is pushed down to the edge with a precise local clock. The cloud decision platform only needs to process high-value data with unified structure and abstract information, which greatly improves system reliability and data processing efficiency.

[0075] 2. Enabled low-cost, high-efficiency, and universally accessible deployment of intelligent models: Through the "Federated Knowledge Distillation" architecture, new elevators or elevators lacking data can quickly obtain a lightweight model with prior knowledge, solving the cold start problem; at the same time, it protects the data privacy of each elevator and enables the sharing of group experience and continuous model evolution.

[0076] 3. Ensures real-time system response and global optimality of decision-making: The "layered trigger-based decision-making mechanism" combines millisecond-level edge autonomous response, second-level cloud-based intelligent recommendation, and hourly-level global optimization planning. This not only meets the immediate handling needs of emergency faults but also achieves near-optimal real-time scheduling through deep reinforcement learning, resulting in strong system scalability.

[0077] 4. Meets the highest standards of privacy compliance: Through the design principle of "edge characterization and destruction of raw data", the system achieves usage pattern analysis without touching or storing raw audio and video data, fundamentally eliminating the risk of privacy leakage and clearing obstacles for deployment in areas with strict regulations.

[0078] 5. It endows the system with strong "resilience": Even if the cloud decision-making platform network is temporarily interrupted, the edge nodes can still maintain a certain degree of independent assessment and emergency response capabilities based on local student models and twin data. After the cloud decision-making platform is restored, it can be seamlessly synchronized, and the overall system's resilience and availability are significantly enhanced.

[0079] Example 2

[0080] This embodiment is a specific implementation of each step in Embodiment 1.

[0081] In step S1, the generation process of the lightweight digital twin data package specifically includes:

[0082] S11: An edge node defines a running segment, starting from the elevator's first door closing and starting, and ending with the next door opening and leveling.

[0083] S12: Within the segment, local real-time processing of multi-source data, specifically collecting and processing the following three types of data:

[0084] Mechanical operating data, including:

[0085] Vibration data: collected by triaxial vibration acceleration sensors installed on the top of the car, the traction machine housing and the guide rail bracket. The sampling frequency is adjustable from 100-500Hz. The root mean square vibration, peak vibration, specific frequency band (such as the characteristic frequency band of the traction machine) energy and short-time vibration energy characteristics are extracted.

[0086] Current data: The instantaneous values ​​of three-phase current are collected by Hall current sensors installed on the power input line of the traction machine at a sampling frequency of 1kHz. The effective value of current, current harmonic distortion rate and current imbalance characteristics are extracted.

[0087] Speed ​​data: Acquired through an incremental encoder installed on the output shaft of the traction machine, synchronized with the motor, and real-time speed and speed fluctuation rate (standard deviation) characteristics are extracted;

[0088] Temperature data: collected by PT100 thermal resistors installed on the bearing housing, brake surface and control cabinet of the traction machine, with a sampling frequency of 1Hz, and real-time temperature value and temperature rise rate (differential) characteristics are extracted.

[0089] Behavioral and environmental data include:

[0090] Video data: Video streams are captured by a wide-angle camera installed on the top of the car at 15fps. A lightweight object detection model (such as MobileNet-SSD) is run locally, and only the number of passengers, distribution area statistics, and overall motion energy index are output. No images are stored or uploaded.

[0091] Audio data: Audio streams were acquired using a microphone array installed on the ceiling of the car, with a sampling frequency of 16kHz. Mel-frequency cepstral coefficient features were extracted.

[0092] Environmental data: Data was collected using photoelectric sensors and temperature and humidity sensors installed in the middle of the shaft, with a sampling frequency of 10Hz. Transmitted light intensity, smoke absorbance, shaft temperature and humidity characteristics were extracted.

[0093] Basic operational data includes:

[0094] Leveling data: acquired through magnetic induction or optical leveling devices, triggered by events to obtain leveling error values;

[0095] Control data: Read operating commands and feedback signals from the elevator main controller, with a sampling frequency of 10Hz, and extract command response delay time and operating status characteristics;

[0096] Load data: collected by the weighing sensor under the car, with a sampling frequency of 1Hz, and real-time load and load change rate features are extracted;

[0097] Door operator data: Read from the car door controller, sampling frequency 10Hz, extracting the resistance current characteristics during door opening and closing time and door movement;

[0098] S13: After the segment ends, all extracted features, along with metadata including the start and end floors, time, and energy consumption logs, are input into a feature fusion encoding network pre-deployed at the edge, and output a low-dimensional (128-dimensional), dense fused health feature vector.

[0099] S14: Package fragment metadata, integrate health feature vectors and local early warning markers (if any), form a standardized twin data package and upload it to the cloud decision-making platform.

[0100] In step S14, the local warning marker is generated by the edge node during the segment processing. The generation method includes: comparing the real-time extracted feature values ​​(such as vibration peak value, leveling error) with the warning rule threshold preset in the node. If the threshold is exceeded, a local warning marker of the corresponding level and type is generated.

[0101] In step S13, the feature fusion encoding network is a small fully connected neural network (such as a 3-layer fully connected network) or a convolutional neural network, which is pre-trained on the cloud decision platform and then deployed to edge nodes. The function of the feature fusion encoding network is to map high-dimensional heterogeneous features into a low-dimensional, dense fused health feature vector with a fixed dimension (such as 128 dimensions). This vector comprehensively represents the overall health status of the elevator within the operating segment.

[0102] In step S2, the specific process of federated knowledge distillation is as follows:

[0103] S21: Teacher Model Training: Cloud Decision Platform The cloud decision platform utilizes a large amount of accumulated historical twin data packages and corresponding actual maintenance records (as labels) to train a powerful and general health prediction teacher model offline. This model is based on an LSTM encoder-decoder architecture and can predict future health trends based on historical twin sequences.

[0104] S22: Knowledge Transfer (Distillation): For the target elevator (new elevator or data-scarce elevator), the cloud-based decision platform initiates a federated knowledge distillation process:

[0105] The cloud-based decision-making platform initializes a lightweight student model (with a simpler structure than the teacher model).

[0106] The cloud-based decision-making platform infers the health prediction teacher model on a set of generalized twin data (e.g., typical data fragments extracted from other elevators and anonymized), and the resulting output probability distribution serves as a soft label.

[0107] The cloud-based decision-making platform sends the initialized student model and the soft tags to the edge node of the target elevator;

[0108] The edge nodes of the target elevator use the received soft tags and a small amount of initial twin data that may exist locally to perform preliminary training on the student model, enabling it to learn the generalization prediction ability of the teacher model.

[0109] S23: Local Personalized Fine-Tuning: The initially trained student model is sent to the target edge node; thereafter, each time the node generates a new twin data package, it uses the fused health feature vector in it as input to incrementally fine-tune the student model, so that it gradually adapts to the unique operating mode and degradation pattern of this elevator; the fine-tuned student model can output the health evolution index of the elevator in real time.

[0110] S24: Model Evolution: Each edge node can periodically upload the fine-tuned parameters (after anonymization) of the student model to the cloud decision platform; the cloud decision platform aggregates these parameters to update the teacher model (global model), realizing the continuous evolution of knowledge in the group.

[0111] In step S23, the health evolution index is a normalized numerical indicator used to quantify the current health status of the elevator and its changing trend. It is calculated by a personalized student model deployed on the edge node based on real-time twin data (the lower the value, the worse the health status or the higher the risk of degradation). The specific process is as follows:

[0112] ① Real-time computation: At the end of each running segment, the edge node will generate a fused health feature vector for that segment. Input local student model , to obtain the original output , It is a constant;

[0113] ② Exponential generation: Subsequently, through a pre-defined linear mapping function... The raw output is converted into intuitive index values; the health evolution index The calculation formula is:

[0114] ,

[0115] in, The preset scaling factor. These are preset offsets used to normalize the model output to the range [0, 100]. The gradual changes in each segment constitute the health evolution trajectory of the elevator.

[0116] In step S3, the hierarchical triggered cluster collaborative decision-making includes:

[0117] S31, Edge Autonomous Emergency Layer: For the highest level of predefined warnings (such as "safety loop abnormality", "overspeed" etc.) identified by the local algorithm of the edge node, immediately execute (without relying on the cloud) the pre-set emergency control commands (such as leveling the nearest floor, opening the door, stopping operation) to ensure the highest safety.

[0118] S32, Cloud-based Fast Inference Layer: When the data packet uploaded by the edge node contains a warning marker that is not the highest level, the cloud-based decision platform immediately triggers a fast decision. It inputs the health evolution index of the elevator that triggered the warning, the specific warning information, and the current global resource status (such as the location / idleness of each maintenance personnel, spare parts inventory) into a pre-trained deep reinforcement learning (DRL) decision network. The network outputs a preliminary handling suggestion within milliseconds, such as assigning a priority to the event or recommending which maintenance personnel to assign the task to.

[0119] S33, Cloud-based Periodic Optimization Layer: The cloud-based decision-making platform periodically (e.g., every 6 hours) or when the accumulated pending tasks reach a certain threshold, initiates a background global optimization algorithm (e.g., genetic algorithm or mixed integer programming). This algorithm reschedules and allocates resources based on all tasks within a longer time window, resource constraints, and more complex optimization objectives (e.g., shortest total completion time, lowest total cost). The optimized new strategy can be used as training data to update the strategy of the deep reinforcement learning decision network in S32, enabling it to make better and faster decisions in the future.

[0120] In step S32, the deep reinforcement learning decision network adopts an actor-critic architecture, and its state space includes the health evolution index of each elevator in the cluster, the current warning level, the location and status of maintenance personnel, and the spare parts inventory; the action space is to allocate different priorities or maintenance resources to different elevators.

[0121] In step S4, the localized feature extraction specifically involves the video analysis module running in the memory of the edge node device, converting the image frames into mathematical features such as optical flow field, edge density, and contour quantity in real time, and immediately discarding the original image frames.

[0122] In step S4, the operating mode label is a classification label used to summarize the operating conditions of a single elevator operation segment, such as "empty up", "heavy down", "frequent stops during peak hours", etc., which is generated by the edge node based on information such as anonymized passenger statistics and starting and ending floors.

[0123] Finally, it should be noted that the above are merely preferred embodiments of this application and are not intended to limit this application. For those skilled in the art, this application can have various modifications and variations. Without conflict, the embodiments and features described in the embodiments of this application can be arbitrarily combined with each other. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for elevator resilience operation and maintenance decision-making, characterized in that, include: S1. Lightweight Digital Twin Segment Generation at the Edge: At each edge node, a complete elevator operation segment is monitored; after the segment ends, the edge node fuses the multi-source sensor data and local analysis results within the segment to generate and upload a standardized lightweight digital twin data package; the data package includes at least the segment identifier, start and end timestamps, operation mode label, fixed-dimensional fused health feature vector, and local early warning marker; S2, Cloud-based Federated Knowledge Distillation and Model Update: The cloud-based decision platform maintains a general health prediction teacher model trained on historical data; for the target elevator, the predicted knowledge of the teacher model is transferred to a lightweight student model through a federated knowledge distillation framework; the student model is distributed to the corresponding edge node and fine-tuned locally using the twin data package subsequently generated by the node to generate the personalized health evolution index of the elevator. S3, hierarchical trigger-based cluster collaborative decision-making, including: S31, Edge Autonomous Emergency Layer: When an edge node receives a predefined highest-level local early warning, it immediately executes pre-set emergency control commands. S32, Cloud-based Fast Inference Layer: When the twin data packet uploaded by the edge node contains a warning marker, the cloud decision platform inputs the elevator's health evolution index, warning information, and current global resource status into the pre-trained deep reinforcement learning decision network, and outputs preliminary handling priorities or dispatch suggestions within milliseconds. S33, Cloud-based Periodic Optimization Layer: The cloud-based decision platform periodically or when the accumulated tasks reach a threshold, initiates a background global optimization algorithm to reschedule and allocate resources for all pending tasks within a certain period, and uses the optimization results to update the strategy of the deep reinforcement learning decision network.

2. The elevator resilience operation and maintenance decision-making method according to claim 1, characterized in that, In step S3, the hierarchical triggered cluster collaborative decision-making includes: S31, Edge Autonomous Emergency Layer: For the highest level of warning identified by the local algorithm of the edge node, the preset emergency control instructions are immediately executed to ensure the highest level of security; S32, Cloud-based Fast Inference Layer: When the data packet uploaded by the edge node contains a warning marker that is not the highest level, the cloud decision platform immediately triggers a fast decision. It inputs the health evolution index of the elevator that triggered the warning, the specific warning information, and the current global resource status into a pre-trained deep reinforcement learning decision network. The network outputs a preliminary handling suggestion within milliseconds, such as assigning a priority to the event or recommending which maintenance personnel to assign the task to. S33, Cloud-based Periodic Optimization Layer: The cloud-based decision platform periodically or when the accumulated pending tasks reach a certain threshold, initiates a background global optimization algorithm. This algorithm reschedules and allocates resources based on all tasks, resource constraints, and more complex optimization objectives within a longer time window. The optimized new strategy can be used as training data to update the strategy of the deep reinforcement learning decision network in S32, enabling it to make better and faster decisions in the future.

3. The elevator resilience operation and maintenance decision-making method according to claim 1, characterized in that, In step S1, the generation process of the lightweight digital twin data package specifically includes: S11: An edge node defines a running segment, starting from the elevator's first door closing and starting, and ending with the next door opening and leveling. S12: Within the segment, local real-time processing of multi-source data: feature extraction is performed on acceleration data collected by vibration sensors to obtain vibration root mean square, vibration peak value, specific frequency band energy, and short-time vibration energy characteristics; feature extraction is performed on current data collected by current sensors to obtain current RMS value, current harmonic distortion rate, and current imbalance characteristics; feature extraction is performed on rotational speed data collected by encoders to obtain real-time rotational speed and rotational speed fluctuation rate characteristics; feature extraction is performed on data collected by temperature sensors to obtain real-time temperature value and temperature rise rate characteristics; local analysis is performed on video streams collected by cameras, outputting only anonymized passenger counts and behavioral statistics. Features were extracted from the audio stream collected by the microphone to obtain the Mel-frequency cepstral coefficient features; features were extracted from the data collected by the photoelectric sensor and the temperature and humidity sensor to obtain the transmitted light intensity, smoke absorbance, shaft temperature and humidity features; features were extracted from the data collected by the leveling sensor to obtain the leveling error value; features were extracted from the data read by the elevator main controller to obtain the command response delay time and operating status features; features were extracted from the data collected by the load cell to obtain the real-time load and load change rate features; and features were extracted from the data read by the door operator controller to obtain the door opening and closing time and the resistance current during the door movement process. S13: After the segment ends, all extracted features, along with metadata including the start and end floors, time, and energy consumption logs, are input into a feature fusion encoding network pre-deployed at the edge, and output a low-dimensional, dense fused health feature vector. S14: Package fragment metadata, integrate health feature vectors and local early warning markers to form a standardized twin data package and upload it to the cloud decision-making platform.

4. The elevator resilience operation and maintenance decision-making method according to claim 3, characterized in that, The feature fusion encoding network is a small, fully connected neural network or convolutional neural network, which is pre-trained on a cloud decision-making platform and then deployed to edge nodes.

5. The elevator resilience operation and maintenance decision-making method according to claim 3, characterized in that, In step S14, the local warning marker is generated by the edge node during the fragment processing. The generation method includes: comparing the feature value extracted in real time with the warning rule threshold preset in the node. If the threshold is exceeded, a local warning marker of the corresponding level and type is generated.

6. The elevator resilience operation and maintenance decision-making method according to claim 1, characterized in that, In step S2, the specific process of federated knowledge distillation is as follows: S21: Cloud-based decision-making platform. The cloud-based decision-making platform utilizes a large amount of accumulated historical twin data packages and corresponding actual maintenance records to train a health prediction teacher model offline. S22: For the target elevator, the cloud-based decision-making platform initiates the federated knowledge distillation process: The cloud-based decision-making platform initializes a lightweight student model; The cloud-based decision-making platform infers the health prediction teacher model on a set of generalized twin data, and the resulting output probability distribution is used as a soft label. The cloud-based decision-making platform sends the initialized student model and the soft tags to the edge node of the target elevator; The edge nodes of the target elevator use the received soft tags and a small amount of initial twin data that may exist locally to perform preliminary training on the student model, enabling it to learn the generalization prediction ability of the teacher model. S23: The initially trained student model is sent to the target edge node; thereafter, each time the node generates a new twin data package, it uses the fused health feature vector in it as input to incrementally fine-tune the student model, so that it gradually adapts to the unique operating mode and degradation pattern of this elevator; the fine-tuned student model can output the health evolution index of the elevator in real time. S24: Each edge node can periodically upload the fine-tuned parameters of the student model to the cloud decision-making platform; The cloud-based decision-making platform aggregates these parameters to update the teacher model, enabling the continuous evolution of knowledge within the group.

7. The elevator resilience operation and maintenance decision-making method according to claim 1, characterized in that, The calculation process for the health evolution index is as follows: ① Real-time computation: At the end of each running segment, the edge node will generate a fused health feature vector for that segment. Input local student model , to obtain the original output , It is a constant; ② Exponential generation: Subsequently, through a pre-defined linear mapping function... The raw output is converted into intuitive index values; the health evolution index The calculation formula is: ， in, The preset scaling factor. These are preset offsets, used to normalize the model output to the range [0, 100]. The gradual changes in each segment constitute the health evolution trajectory of the elevator.

8. The elevator resilience operation and maintenance decision-making method according to any one of claims 1-7, characterized in that, It also includes step S4: Privacy protection and compliance processing: Edge nodes perform localized feature extraction on the original audio and video data, and only use irreversible and unidentifiable abstract statistical features to generate the operating mode label. The original audio and video data is destroyed immediately after local feature extraction and is not stored or uploaded.

9. The elevator resilience operation and maintenance decision-making method according to claim 8, characterized in that, The localized feature extraction specifically involves the video analysis module running in the memory of the edge device, converting image frames into mathematical features such as optical flow field, edge density, and contour quantity in real time, and immediately discarding the original image frames.

10. An elevator resilience operation and maintenance decision-making system, used to implement the method of any one of claims 1-9, the system comprising: Multiple edge nodes, each configured to perform step S1 of claim 1; The cloud-based decision-making platform is configured to execute steps S2, S32, and S33 of claim 1 and to communicate with the edge node.

Images