Elevator edge computing device and elevator event recognition method

Abstract

This invention discloses an elevator edge computing device, including a communication interface, a control and acquisition interface, and an AI processing unit. The communication interface includes dual Ethernet ports and onboard WiFi and Bluetooth modules. The control and acquisition interface includes a USB interface, a CAN bus, an RS485 or RS232 interface, digital input, and digital output interfaces. The CAN bus interfaces with the elevator control system, and the RS485 and RS232 interfaces interface with various sensors or peripherals. The AI ​​processing unit is equipped with a trained AI model that features adaptive parameter optimization. The elevator edge computing device is installed on the elevator car in a locally deployed manner, enabling data communication and cloud-edge collaboration with a cloud platform. This invention has significant value in processing dense sensor signal data in the elevator car and optimizing local performance and cost.

Description

Technical Field

[0001] This invention relates to the field of elevator technology, specifically to an elevator edge computing device and an elevator event recognition method. Background Technology

[0002] In the practical implementation of elevator edge computing devices, their physical deployment location is one of the key factors determining system performance, cost, and reliability. The elevator machine room, shaft, and car (car top) are the main candidate locations for installing edge computing devices. Considering that the car is a core data source related to safety, passengers, and the environment (such as video analytics, behavior recognition, and environmental monitoring), and that sensor information is dense and processing requirements are complex, current technologies have not optimized the placement of edge computing devices, which is one of the existing problems.

[0003] Existing elevator edge computing solutions, exemplified by Chinese Patent Document 1 (CN120308780A), typically deploy gateway devices in the machine room and rely primarily on reading information from the elevator control system's internal data bus (such as CAN or RS485) for analysis. This architecture suffers from inherent flaws that prevent it from meeting the core requirements of efficient, reliable, and low-cost operation of modern elevator IoT systems: On the one hand, the limitations of the system architecture lead to a fundamental contradiction between the "flood of raw data" and the "harsh communication environment": the data source of this solution highly depends on the internal state parameters (such as floor, speed, and fault codes) already processed and encapsulated by the elevator control system. However, to achieve accurate fault warning, safety monitoring, and performance optimization, a large number of sensors outside the control system must be introduced (such as high-definition cameras for behavior recognition, high-sampling-rate vibration sensors for abnormal noise analysis, and audio sensors for abnormal sound detection). Under the existing architecture, the sensor data located at remote locations such as the car and hoistway needs to be transmitted back to the edge gateway on the machine room side via long-distance cables, or directly uploaded to the cloud. This firstly significantly increases the complexity and cost of dedicated cabling. More importantly, the massive volume of these multimodal raw data (especially video and audio streams) poses an extreme challenge to the bandwidth and stability of the communication link when it must be transmitted through the heavily shielded and interference-prone interior environment of the building (such as the hoistway). The existing solution essentially fails to solve the fundamental problem of "transporting massive amounts of raw data from the end of the harsh communication environment," forcing reliance on high-bandwidth or high-cost communication solutions with poor stability.

[0004] On the other hand, insufficient local processing capabilities lead to "traffic dependence" and a lack of real-time performance: Since existing technologies can only process local data, their self-iteration potential and convenience are insufficient. Especially when the analysis results of some raw signals still have uncertainties, local processing solutions lack long-term improvement mechanisms. For scenarios requiring complex analysis of raw high-dimensional data (such as images, vibrations, and noise), general-purpose computing devices often suffer from unstable network access, difficulty in transmitting large amounts of raw data, or even inability to transmit to the cloud due to their location (such as in elevator shafts or interference-prone computer rooms). However, the raw data still requires other methods for analysis. This situation may lead to high costs in human resources or latency caused by poor network conditions, resulting in high traffic costs. This violates the core principles of edge computing: "real-time and proximity."

[0005] As a common specialized piece of equipment in buildings, elevators, when integrated into building systems, often rely on their own sensors or controllers for basic data. Directly interfacing with this low-information-density data significantly increases communication costs and algorithmic complexity. Building systems require access to general-purpose data and unified processing of data from various heterogeneous devices such as HVAC, lighting, security, and elevators. However, this "unified" design philosophy leads to the following shortcomings when dealing with the highly complex, specialized, and spatially universal vertical scenario of elevators: 1. Generalized scenarios lead to a lack of depth, failing to unlock the "black box" of elevator data and control: Building systems treat elevators as ordinary terminals providing limited data points (such as operating status and fault codes) through standard protocols (e.g., BACnet / IP, CAN). However, the elevator system is a deeply integrated mechatronic closed loop, with extremely rich and specialized real-time information flows within its control system (MCU / CPU / PLC) (such as vector control curves, door operator torque feedback, safety loop status, load spectrum, etc.). General-purpose building edge gateways lack sufficient domain knowledge and processing permissions to analyze this deep-level information, let alone perform high-real-time, high-reliability secure interaction with the elevator control system. This results in the building management system's "mastery" of elevators remaining superficial, unable to achieve truly in-depth predictive maintenance and performance optimization.

[0006] 2. Building systems' training data is constrained by building boundaries, preventing the formation of cross-geographical "network intelligence." The cloud-based large-scale model of a building system is strictly limited to the learning and optimization scope of a single building or a local building cluster. Its training data comes from the equipment within that building, and the generated model primarily serves the building's energy conservation and maintenance. However, the operating logic, failure modes, and performance degradation patterns of elevator equipment have strong cross-regional and cross-building type universality. Buildings of the same type may have similar data scenarios within their elevators. Fully utilizing data collected by in-car sensors across regions and performing sample backflow on edge computing devices can significantly improve the performance of edge computing devices globally. However, this is difficult to directly implement in elevators within a building system. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an elevator edge computing device, comprising a communication interface, a control and acquisition interface, and an AI processing unit; the communication interface includes dual Ethernet ports and onboard WiFi and Bluetooth modules; the control and acquisition interface includes a USB interface, a CAN bus, an RS485 interface, an RS232 interface, digital input interfaces, and digital output interfaces. The CAN bus connects to the elevator control system, and the RS485 and RS232 interfaces connect to various sensors or peripherals; the AI ​​processing unit is equipped with a trained AI model with adaptive parameter optimization; the elevator edge computing device is deployed on the elevator car to communicate with the cloud platform and perform cloud-edge collaboration.

[0008] Because the elevator edge computing device is deployed close to the elevator car, the physical distance between it and the sensor is shortened, resulting in advantages such as reduced bandwidth pressure and improved real-time performance.

[0009] Preferably, the elevator edge computing device includes a power outage recovery module; when the power outage recovery module detects an abnormal disconnection of the external power supply, it supports the elevator edge computing device to continue operating for at least a first preset time. During this period, the elevator edge computing device captures the sensor information of the last moment in the car and triggers alarm information to be uploaded to the cloud or to send emergency information to the cloud platform through the elevator's internal communication network data channel.

[0010] Preferably, the elevator edge computing device includes a hardware watchdog; when the main program of the elevator edge computing device crashes, the hardware watchdog will forcibly reset the system within a preset time and record the fault log.

[0011] Preferably, the elevator edge computing device reads and parses the elevator system's private protocol data via the CAN bus and stores the operating condition information; the elevator edge computing device performs rolling storage of sensor data; the operating condition information includes faults and operating scenarios.

[0012] Preferably, the elevator edge computing device performs local calculations on the collected data to obtain data features and processes them into simplified labels or result data.

[0013] This invention deploys an elevator edge computing device with abundant expansion interfaces on the elevator car (e.g., on the car top), which can bring the following advantages: 1. Significantly shorten sensor wiring distance: Reduce attenuation, interference risks, and cable costs associated with long-distance analog signal transmission.

[0014] 2. Improve signal quality and simplify hardware design: Short-distance transmission helps maintain a higher signal-to-noise ratio, which can reduce the requirements for sensor output drive capability or front-end anti-interference circuits, providing greater flexibility for sensor selection.

[0015] 3. Enables local real-time decision-making: Reduces the overall communication bandwidth requirements from the elevator car to the machine room. Enables edge computing devices to perform low-latency processing of car scene events (various data collected by sensors) locally, and quickly feeds the results back to the elevator control system as high-value input, enhancing the overall system's intelligent response capability.

[0016] Therefore, deploying edge computing devices close to the car roof is of great value in processing dense sensor signal data in the car, optimizing local performance and cost. At this time, the distance of the sensor access cable can be significantly controlled / reduced, the risk of interference is reduced, and the cost is lowered.

[0017] The present invention also provides an elevator event recognition method, which uses the aforementioned elevator edge computing device to recognize elevator events; the elevator edge computing device acquires real-time video streams through a camera mounted in the car and uses the AI ​​processing unit to recognize elevator events.

[0018] Preferably, the elevator edge computing device collects audio inside the elevator car through a microphone array and uses the AI ​​processing unit to identify elevator events.

[0019] Preferably, when the confidence level of elevator event identification is lower than the threshold and within a preset range, the relevant data is used as a difficult case and uploaded to the cloud platform in real time; the cloud platform uses the relevant data to train the AI ​​model and then sends the trained new AI model to the elevator edge computing device.

[0020] Preferably, the elevator edge computing device uses the new AI model as a backup model, and the actual output is determined by the original old AI model; it also inputs the newly acquired sensor data into both the new AI model and the old AI model simultaneously, and records the output differences between the two; within an observation period of a given time parameter length, it statistically analyzes the consistency between the inference results of the new AI model and the old AI model; when the output of the new AI model meets the preset indicators, the elevator edge computing device activates the new AI model and retains the old AI model as a backup. Attached Figure Description

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Figure 1 This is a schematic diagram of the hardware interface of the elevator edge computing device in Example 1; Figure 2 This is a schematic diagram of the connection between the elevator edge computing device and external devices in Example 1; Figure 3 This is a schematic diagram of how a suspected elevator entrapment incident is processed by the elevator edge computing device in Example 4. Detailed Implementation

[0022] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can fully understand other advantages and technical effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through different specific embodiments, and the details in this specification can also be applied based on different viewpoints, with various modifications or changes made without departing from the overall design concept of the invention. It should be noted that, unless otherwise specified, the following embodiments and features can be combined with each other. The following exemplary embodiments of the present invention can be implemented in many different forms and should not be construed as being limited to the specific embodiments set forth herein. It should be understood that these embodiments are provided to make the disclosure of the present invention thorough and complete, and to fully convey the technical solutions of these exemplary embodiments to those skilled in the art. Example 1

[0023] like Figure 1 As shown, this embodiment provides an elevator edge computing device, including a communication interface, a control and acquisition interface, and an AI processing unit.

[0024] The communication interface includes dual Ethernet ports (WAN / LAN) and onboard WiFi and Bluetooth modules. The dual Ethernet ports are used to connect to building networks, 4G / 5G routers, and network cameras; the onboard WiFi and Bluetooth modules are used for near-field debugging.

[0025] The control and data acquisition interfaces include a USB interface (for connecting external storage or other general-purpose devices), a CAN bus, RS485 and RS232 interfaces, as well as several sets of digital input (DI) and digital output (DO) interfaces. The CAN bus is used to interface with the elevator control system, while the RS485 and RS232 interfaces are used to connect various sensors or peripherals.

[0026] The AI ​​processing unit is equipped with a trained AI model capable of adaptive parameter optimization. This elevator edge computing device is installed in the elevator car and communicates with the cloud platform for data exchange and cloud-edge collaboration.

[0027] like Figure 2 As shown, the device in this embodiment serves as the core node of the elevator Internet of Things (IoT) and adopts an industrial-grade embedded architecture design. It connects to the elevator control system, elevator-related controllable devices (such as air conditioners), and common sensors in the elevator environment (such as temperature and humidity sensors, vibration sensors, cameras, microphone arrays, etc.), and communicates with the cloud platform to achieve cloud-edge collaboration.

[0028] The innovation of this embodiment lies in its departure from the traditional approach of centrally deploying computing nodes in a server room. Instead, it deploys high-performance edge intelligent devices directly at the core source of data generation—the top or interior of the elevator car. This device is physically close to the data source, connecting to and managing multi-source sensors (such as cameras, microphones, vibration sensors, etc.) in and around the car and key areas. The elevator itself can also function as a sensor, transmitting data to the edge computing device via a bus for collaborative fusion processing with data from other sensors. Its core improvement lies in utilizing a built-in dedicated AI processing unit to run a lightweight yet high-performance intelligent model at the very first moment and location of data generation, transforming the continuous, high-dimensional raw data stream into low-volume, highly structured semantic information in real time.

[0029] This device not only achieves static, localized data processing but also constructs a self-reinforcing performance optimization loop. By deploying AI models with adaptive parameter optimization, it achieves efficient information condensation. Through continuous technological iteration, it constantly improves the analytical accuracy, feature extraction capabilities, and scene coverage of the edge-side model. More powerful models can extract more accurate, abstract, and information-density feature descriptions from the same raw data and upload data to the cloud via the elevator's own IoT data upload channel, ensuring network communication quality.

[0030] For example, early models might need to upload a compressed audio clip (tens of KB) for secondary verification in the cloud, while the improved model can directly determine "timestamp, abnormal voiceprint matching 'insufficient guide shoe lubrication', confidence level 98%" locally, requiring only a conclusion of less than 100 bytes to be uploaded. Only cases that are difficult to determine are uploaded to the cloud server for subsequent model training.

[0031] Event assessment is achieved through cloud-edge collaboration: the edge performs local computations on the data, extracting features and processing them into concise labels or result data, rather than uploading the full data, thereby reducing bandwidth consumption and cloud computing pressure. The cloud then performs event analysis and further processing based on the uploaded label data, and intervenes as necessary on the edge based on the analysis results. Example 2

[0032] Based on Example 1, and considering the special characteristics of the shaft operating environment, the elevator edge computing device in this example also includes the following key modules: Power Outage Resumption Module (Backup Battery): This module incorporates a built-in power outage detection circuit and a short-term backup module. When an abnormal disconnection of external power supply is detected (such as a power outage accompanying elevator entrapment), this module can support continued operation for at least 30 seconds. During this period, the device performs an "emergency snapshot" task, capturing the last sensor information inside the elevator car and triggering alarm information to be uploaded to the cloud or the elevator's internal communication network data channel. It also sends emergency information to the cloud platform, ensuring "no loss of connection during power outage."

[0033] Hardware Watchdog: A built-in hardware watchdog circuit independent of the main processor. When the main program crashes due to electromagnetic interference or software logic errors (heartbeat loss), the watchdog will force a system reset within 10 seconds and record the fault log. This hardware bypass design ensures that faults in peripheral devices do not affect the operation of the elevator's original safety circuit. Example 3

[0034] Building upon Example 1, the elevator edge computing device in this example supplements the elevator control system. It reads and parses the elevator system's proprietary protocol data (including operating status, load, door status, fault status, etc.) via the CAN / 485 interface and stores the corresponding operating condition information. Simultaneously, when sensors are connected, sensor data can be stored in a rolling manner; the operating condition information includes faults and operating scenarios.

[0035] This storage mode can save scene data when communication is interrupted, which facilitates subsequent information review (such as building scene analysis) and provides data basis for the improvement of equipment and models. Example 4

[0036] During operation, elevator systems identify potential entrapment risks based on their response to commands and by combining existing signals and combinations within the control system. However, due to the limitations of this method, human intervention is necessary to identify or filter suspected entrapment incidents.

[0037] In the process of recognizing and filtering elevator entrapment alarms, it is generally done by uploading several real-time camera images, which are then used by cloud services for entrapment identification. However, this process is limited by network signal strength and communication overhead, resulting in a certain lag in event recognition. By using edge computing devices, images from cameras within the local area network are acquired at the edge, and the edge model directly assesses the situation inside the elevator, processing the data to determine whether an entrapment is real or a false alarm. This allows for intervention in subsequent processes, significantly improving the algorithm's real-time performance while reducing communication overhead.

[0038] This embodiment provides an elevator event recognition method that uses an elevator edge computing device to identify elevator events. The device acquires real-time video streams through a camera mounted in the elevator car and uses an AI processing unit for event recognition. Taking a suspected elevator entrapment incident as an example, such as... Figure 3 As shown, with the addition of more sensors, the image re-judgment service will form a multimodal fusion judgment, achieving high-precision identification and handling of trapped people.

[0039] In this embodiment, when the edge computing device is simultaneously connected to a noise sensor (such as a microphone), an "audiovisual fusion" strategy is adopted, as follows: - Visual Channel: Real-time video stream is acquired through a camera on the car roof, and a lightweight target detection model (such as an improved version of YOLO-Tiny) is used to identify whether there are people in the car and whether there are violent physical movements (such as waving for help or falling to the ground).

[0040] - Auditory channel: Audio inside the elevator car is collected through a microphone array, and after noise reduction processing, a voiceprint recognition model is used to detect specific abnormal sounds (such as shouts for help, violent banging on the elevator walls, and abnormal noises during elevator operation).

[0041] - Decision logic: Edge devices perform weighted fusion of visual confidence and auditory confidence. When image information and voiceprint information are present at the same time, the system adds a "confirmed to be a trapped person (high confidence)" label to the original reported information.

[0042] Furthermore, the edge computing device sends the image re-judgment results to the elevator control system, which then provides localized audio-visual reminders (such as playing soothing voice prompts).

[0043] In this embodiment, the elevator's associated event information is jointly analyzed and calculated by the elevator and edge computing devices. The results serve as the elevator data access object for the building system to access. The elevator's analysis mode adaptively iterates based on scene data, and the building system is treated as an accessible "sensor." Example 5

[0044] Building upon Example 4, when the edge-side model deployed on the edge computing device has doubts about the elevator entrapment identification (e.g., the confidence level is near a threshold of 50%, and the identification result is 45%), the relevant data will be uploaded to the cloud platform in real time as a "difficult case." The cloud platform will then call a higher-precision model (which is usually not suitable for direct deployment on the edge) and combine it with more multi-dimensional data (such as abnormal events and fault records within the elevator cycle) to comprehensively judge the result of the difficult case and send it back to the edge side for auxiliary judgment.

[0045] Similarly, when deploying electric vehicle recognition algorithms, certain ambiguous objects (such as electric vehicles with raincoats, tricycles, and electric wheelchairs) may lead to misidentification. To improve recognition performance, images with a confidence level of 25% to 50% are classified as difficult examples and uploaded to the cloud periodically. The cloud uses a large-scale model algorithm to further classify the specific categories of difficult examples and uses this as feedback data to optimize subsequent models. The optimized model is then distributed in batches to each edge computing device via OTA (Over-The-Air).

[0046] After the new model is downloaded, it does not immediately replace the old model. Instead, it is started in the background and runs in "shadow mode." During this time, the actual control output is still determined by the old model. Edge devices simultaneously input sensor data from the same time period into both the old and new models, recording the differences in their outputs. During a 72-hour observation period, the consistency of the inference results between the new and old models is statistically analyzed, as well as whether the new model generates false alarms in the "safe scenarios" defined by the old model. Only when the new model's performance in shadow mode meets preset indicators (such as the number of serious logic errors and average / maximum inference time) will the system perform a master-slave switchover, officially activating the new model, while the old model is retained as a backup for rapid rollback in case of anomalies.

[0047] This embodiment fully leverages the performance of edge computing devices, using the devices themselves as the object to iterate and form a cloud-edge-device intelligent architecture of "deep integration of vertical scenarios and collaborative evolution of horizontal networks".

[0048] By fully utilizing elevator scenario data, sensor data is analyzed by edge computing devices to collect challenging scenarios, and the data is uploaded to a cloud server using the elevator's built-in IoT access channel. The cloud server analyzes these challenging scenarios as a supplement to edge computing. This challenging data comes from a wide range of sources, covering different building types, elevator cars, and surrounding environments. This data accumulation across time, region, and environment supports further improvements in model performance.

[0049] By collecting data across time, region, and environment, the optimized model can be distributed to various edge computing devices via cloud servers. This enhances their ability to analyze sensor data in elevator scenarios, enabling cross-spatial collaborative learning, ultimately improving model efficiency, reducing the likelihood of backtracking to known difficult examples, and minimizing communication traffic in known scenarios. Communication transmission focuses on the event itself and device performance improvements (such as new scenario analysis and mining).

[0050] By collecting data on the edge, computing on the edge, analyzing difficult cases in the cloud, and optimizing and deploying models, cloud-edge-device collaboration is achieved, ultimately forming a highly efficient analysis that spans time, region, and environment.

[0051] The present invention has been described in detail above through specific embodiments and examples, but these are not intended to limit the invention. Many modifications and improvements can be made by those skilled in the art without departing from the principles of the invention, and these should also be considered within the scope of protection of the present invention.

Claims

1. An elevator edge computing device, characterized in that, It includes a communication interface, a control and acquisition interface, and an AI processing unit; the communication interface includes dual Ethernet ports and onboard WiFi and Bluetooth modules; the control and acquisition interface includes a USB interface, a CAN bus, an RS485 interface, an RS232 interface, digital input, and digital output interfaces; The CAN bus interfaces with the elevator control system, and the RS485 and RS232 interfaces interfaces with various sensors or peripherals. The AI ​​processing unit is equipped with a trained AI model that features adaptive parameter optimization. The elevator edge computing device is deployed on the elevator car and communicates with the cloud platform for data exchange and cloud-edge collaboration.

2. The elevator edge computing device according to claim 1, characterized in that, Including a power-off battery recovery module; When the power failure recovery module detects an abnormal disconnection of the external power supply, it supports the elevator edge computing device to continue operating for at least the first preset time. During this period, the elevator edge computing device captures the sensor information inside the car at the last moment and triggers alarm information to be uploaded to the cloud or to send emergency information to the cloud platform through the elevator's internal communication network data channel.

3. The elevator edge computing device according to claim 1, characterized in that, This includes a hardware watchdog; when the main program of the elevator edge computing device crashes, the hardware watchdog will force a system reset within a preset time and record a fault log.

4. The elevator edge computing device according to claim 1, characterized in that, The elevator edge computing device reads and parses the elevator system's private protocol data via the CAN bus and stores the operating condition information; the elevator edge computing device performs rolling storage of sensor data; the operating condition information includes faults and operating scenarios.

5. The elevator edge computing device according to claim 1, characterized in that, The elevator edge computing device performs local calculations on the collected data to obtain data features and processes them into simplified labels or result data.

6. A method for elevator event recognition, characterized in that, Elevator events are identified using an elevator edge computing device, wherein the elevator edge computing device is any one of claims 1 to 5; The elevator edge computing device acquires real-time video streams through cameras mounted in the car and uses the AI ​​processing unit to identify elevator events.

7. The elevator event recognition method according to claim 6, characterized in that, The elevator edge computing device collects audio from inside the elevator car through a microphone array and uses the AI ​​processing unit to identify elevator events.

8. The elevator event recognition method according to claim 6 or 7, characterized in that, When the confidence level of elevator event identification is lower than the threshold and within the preset range, the relevant data is used as a difficult case and uploaded to the cloud platform in real time. The cloud platform uses the relevant data to train the AI ​​model and then sends the trained new AI model to the elevator edge computing device.

9. The elevator event recognition method according to claim 8, characterized in that, The elevator edge computing device uses the new AI model as a backup model, while the actual output is determined by the original old AI model. Newly acquired sensor data is simultaneously input into both the new and old AI models, and the output differences between the two are recorded. During an observation period of a given time parameter, the consistency between the inference results of the new AI model and the old AI model is statistically analyzed. When the output of the new AI model meets the preset indicators, the elevator edge computing device activates the new AI model and retains the old AI model as a backup.

Images