A multi-axis robot method and device for power distribution room safety inspection AI monitoring

Abstract

The application discloses a kind of multi-axis robot method and device for power distribution room safety inspection AI monitoring, to solve the pain points such as supervision fault, illegal discovery difficult in traditional power distribution room inspection.The system integrates track walking, visual recognition, multi-sensor fusion and other core modules, relies on 5G private network and industrial Ethernet to realize high-reliability data transmission.Through UWB positioning and digital twin technology, personnel dynamic tracking is realized, and a "small model real-time identification + large model dynamic evaluation" collaborative mechanism is adopted to accurately detect illegal behavior and equipment anomalies.The robot can move autonomously along the I-shaped track, and through the improved YOLOv7-tiny model and multi-sensor fusion technology, it can realize instrument identification, abnormal heating detection and other functions.The system realizes automatic inspection of the whole process, strengthens the implementation of safety management system, has strong expansibility and is suitable for various inspection scenarios in power distribution rooms.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and robotics technology for power distribution rooms, and in particular to a method and device for AI monitoring of multi-axis robots for safety inspection of power distribution rooms. Background Technology

[0002] As the power industry's operational scenarios become increasingly complex, with more high-risk scenarios such as cross-regional maintenance, high-altitude operations, and nighttime emergency repairs, the traditional "person-to-person" supervision model is no longer sufficient to meet safety requirements. Issues such as personnel violations, unqualified personnel, and unauthorized absence from duty have become major causes of safety accidents. There is an urgent need to build a comprehensive, all-encompassing personnel safety management system using intelligent methods to shift from "post-accident remediation" to "prevention."

[0003] Breakthroughs in artificial intelligence technology have provided core support for power safety management and control. "AI visual recognition + big data analysis" has moved from concept to large-scale application. Power plants can accurately capture violations and issue warnings within seconds through scenario-optimized AI models. Combined with technologies such as UWB positioning, this enables dynamic tracking of personnel, breaking the limitations of traditional "passive recording" monitoring. The collaborative application of large and small models has shifted safety management from "experience-driven" to "data-driven decision-making," promoting the upgrading of personnel management from a fragmented and reactive model to a full-process, intelligent, proactive defense system.

[0004] Regulatory gaps and difficulty in detecting violations: In nighttime, remote areas, or high-risk scenarios, manual supervision is insufficient to cover all areas, making it difficult to detect and intervene in violations in a timely manner; Passive and inefficient personnel status and qualification management: Personnel qualification verification relies on manual processes, which are cumbersome, inefficient, and make it difficult to achieve real-time dynamic matching and auditing; Incomplete closed-loop for handling violations: The handling of violations lacks a systematic tracking mechanism, resulting in repeated violations and delayed rectification, failing to form a complete management closed loop; Lack of accuracy in risk prediction: Operational risk assessment relies heavily on experience-based judgment and lacks intelligent analysis and hierarchical early warning capabilities based on multi-source data; Difficulty in implementing accountability and assessment: Data such as work records and surveillance videos are scattered, making it difficult to quickly link "personnel-behavior-responsibility," and assessments lack quantitative basis; Low efficiency in emergency response: In case of emergencies, the lack of intelligent guidance and collaborative command on-site leads to slow emergency response and is prone to secondary accidents. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] Therefore, the first objective of this invention is to propose a multi-axis robot method for AI monitoring of power distribution room safety inspection.

[0007] Another objective of this invention is to propose an AI-controlled multi-axis robot device for safety inspection of power distribution rooms.

[0008] The third objective of this invention is to provide a computer device.

[0009] The fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0010] To achieve the above objectives, a first aspect of the present invention proposes a method for AI-controlled multi-axis robot monitoring for safety inspection of power distribution rooms, comprising:

[0011] S1: Construct a high-precision three-dimensional spatial map of the power distribution room and plan the inspection route. The track-mounted mobile platform carries multiple sensors and moves autonomously along the track to generate multi-source heterogeneous sensor data. S2, based on a lightweight AI model deployed on an edge computing unit, analyzes the video stream of the work area in real time, and combines UWB positioning data with a digital twin system to dynamically identify personnel identity, behavior and location, and generate real-time monitoring status information. S3 integrates real-time monitoring status information with multi-source heterogeneous sensor data, and adopts a non-same-source dual-confirmation mechanism to cross-verify the device status and environmental parameters, triggering abnormal warnings and generating structured alarm information. S4, through the background intelligent management and control platform, links the edge model and the cloud big model to execute anomaly handling instructions and record the handling process.

[0012] In one embodiment of the present invention, S1 includes: Track and power supply unit: The I-shaped track laid on the top of the power distribution room, and the seamless sliding contact line laid along the track, are used to provide continuous power supply and data communication channel for the robot; The walking drive unit includes a first drive motor, a first planetary reducer, a transmission system consisting of five meshing first gears, and two walking wheels that engage with the side of the track. The walking speed satisfies the formula:

[0013] in, This is the reduction ratio of the first planetary gearbox. , , The number of teeth for the driving, intermediate, and driven gears are respectively. The radius of the traveling wheel, This is the speed of the first drive motor.

[0014] In one embodiment of the present invention, S2 includes: The face feature extraction algorithm based on ArcFace is used, and its loss function is expressed as:

[0015] Where (s) is the feature scaling factor, and (m) is the angle margin. The angle between the current sample and its corresponding class center; Based on the improved YOLOv7-tiny model, the ECA-Net attention mechanism is introduced to enhance small object detection capabilities, and the loss function adopted is CIoULoss.

[0016] in, (c) is the distance between the center points of the predicted bounding box and the ground truth bounding box, (v) is the length of the diagonal of the minimum bounding rectangle, and (v) is the aspect ratio consistency measure.

[0017] In one embodiment of the present invention, S3 includes: The instrument and status recognition unit employs a CRNN+Attention structure for instrument digit recognition and enhances the model's ability to interpret readings from pointer-type instruments. Convolutional layers extract image features, recurrent layers model sequence dependencies, and the attention mechanism focuses on the pointer or digit region, outputting the recognition result. Its forward propagation process is represented as follows:

[0018] It supports real-time identification of various targets, including ammeters, voltmeters, and indicator lights.

[0019] In one embodiment of the present invention, S4 includes: The physical marking mechanism is controlled to mark the faulty equipment. The physical marking mechanism is a coding machine, a stamping device or a label attaching mechanism. The marking content includes the type of abnormality, timestamp and equipment number. The alarm information is encapsulated in JSON format and pushed to the backend platform and mobile terminal via the MQTT protocol. The JSON structure includes the event type, occurrence time, camera location, associated work ticket ID, and information of the personnel involved.

[0020] In one embodiment of the present invention, the method further includes: It adopts a dual-redundant design of sliding contact line power supply and lithium battery. The lithium battery capacity is 48V / 50Ah, which supports continuous operation for no less than 2 hours after the robot reports a power failure. Power Management Unit: Enables real-time monitoring of voltage and current and overload protection, and supports remote power status query and fault diagnosis.

[0021] In one embodiment of the present invention, the method further includes: The environmental safety verification process involves using commands to drive a robot to scan sensor data in the target area to determine if the SF6 gas concentration is below a certain level. Is the oxygen content higher than 19.5%? Is the ambient temperature lower than 19.5%? If any indicator is abnormal, the work ticket will be frozen and an early warning work order will be generated.

[0022] To achieve the above objectives, a second aspect of the present invention provides a multi-axis robot device for AI monitoring of power distribution room safety inspection, comprising: The 3D spatial map construction and path planning module is used to build a high-precision 3D spatial map of the power distribution room and plan the inspection path. It uses a track-type mobile platform to carry multiple sensors and move autonomously along the track to generate multi-source heterogeneous sensor data. The edge AI video analysis and dynamic recognition module is used to analyze video streams in the work area in real time based on lightweight AI models deployed on edge computing units. It combines UWB positioning data and digital twin systems to dynamically identify personnel identity, behavior and location, and generate real-time monitoring status information. The multi-source data fusion and anomaly warning module is used to fuse real-time monitoring status information and multi-source heterogeneous sensor data. It adopts a non-same-source dual-confirmation mechanism to cross-verify the device status and environmental parameters, trigger anomaly warnings, and generate structured alarm information. The intelligent management and control module is used to link the edge model and the cloud-based big model through the background intelligent management and control platform, execute anomaly handling instructions and record the handling process.

[0023] This invention discloses a method and device for AI-controlled multi-axis robot monitoring of power distribution room safety inspection, which realizes all-round real-time monitoring and multi-source data fusion analysis of the 6kV power distribution room operation area. Through the collaboration of large and small models and physical marking mechanism, it effectively forms a safety closed loop of "discovery-early warning-disposal-tracing", significantly improves the accuracy of violation identification and emergency response efficiency, and reduces the risk of human operation.

[0024] To achieve the above objectives, a third aspect of this application provides a computer device comprising a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, for implementing a multi-axis AI monitoring robot for power distribution room safety inspection as described in the first aspect embodiment.

[0025] To achieve the above objectives, the fourth aspect of this application proposes a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a multi-axis robot for AI monitoring of power distribution room safety inspection as described in the first aspect embodiment.

[0026] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] Figure 1 This is a flowchart of a multi-axis robot method for AI monitoring of power distribution room safety inspection according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall architecture of an AI monitoring robot system for safety inspection of power distribution rooms according to an embodiment of the present invention; Figure 3 This is a cross-sectional schematic diagram of the mechanical structure of the track walking module according to an embodiment of the present invention; Figure 4 This is a block diagram of the improved YOLOv7 visual recognition model network structure according to an embodiment of the present invention; Figure 5 This is a logic flowchart of multi-sensor fusion detection and emergency response according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the "no operation without monitoring" business process of the robot system and the back-end intelligent management and control platform according to an embodiment of the present invention. Figure 7 This is a structural diagram of a multi-axis robot device for AI monitoring of power distribution room safety inspection according to an embodiment of the present invention; Figure 8 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0030] The following description, with reference to the accompanying drawings, describes a method and apparatus for AI monitoring of multi-axis robots for safety inspection of power distribution rooms, based on embodiments of the present invention.

[0031] Figure 1 This is a flowchart of a multi-axis robot method for AI monitoring of power distribution room safety inspection according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1: Construct a high-precision three-dimensional spatial map of the power distribution room and plan the inspection route. The track-mounted mobile platform carries multiple sensors and moves autonomously along the track to generate multi-source heterogeneous sensor data. S2, based on a lightweight AI model deployed on an edge computing unit, analyzes the video stream of the work area in real time, and combines UWB positioning data with a digital twin system to dynamically identify personnel identity, behavior and location, and generate real-time monitoring status information. S3 integrates real-time monitoring status information with multi-source heterogeneous sensor data, and adopts a non-same-source dual-confirmation mechanism to cross-verify the device status and environmental parameters, triggering abnormal warnings and generating structured alarm information. S4, through the background intelligent management and control platform, links the edge model and the cloud big model to execute anomaly handling instructions and record the handling process.

[0032] Furthermore, this invention proposes an AI monitoring multi-axis robot system for safety inspection of 6kV substations in smart power plants. The system includes modules such as a track walking module, a vision recognition module, a multi-sensor fusion module, a communication control module, a power supply module, an emergency power supply response module, and a back-end intelligent management and control platform.

[0033] The track-walking module is used to achieve precise positioning and movement of the robot within the power distribution room. It includes: a track and power supply unit: an I-shaped track laid on the ceiling of the power distribution room, and a seamless sliding contact line laid along the track, providing continuous power supply and data communication channels for the robot; and a walking drive unit: including a first drive motor, a first planetary reducer, a transmission system consisting of five meshing first gears, and two walking wheels that engage with the sides of the track. The walking speed satisfies the formula:

[0034] in, This is the reduction ratio of the first planetary gearbox. , , The number of teeth for the driving, intermediate, and driven gears are respectively. The radius of the traveling wheel, The speed is the speed of the first drive motor. The lifting drive unit includes a second drive motor, a second planetary reducer, a transmission system consisting of five meshing second gears, and two reels. Vertical lifting is achieved by connecting the monitoring pod below via a steel wire rope. The lifting speed satisfies the formula:

[0035] in, The radius of the winding reel. Clamping and guiding unit: including the suspension wheel assembly and a clamping assembly consisting of springs and adjusting screws, ensuring stable robot suspension and automatically compensating for track clearance and wear.

[0036] Furthermore, the visual recognition module is integrated into the monitoring pod for strong binding of tickets and monitoring data, full-process control of operations, and handling of violations, including: Work ticket creation and associated monitoring unit: When a work ticket is initiated, the system forcibly selects the camera associated with the work area. A ticket can only be successfully created if the area is effectively monitored. The system automatically binds the camera number to the ticket ID, creating a "one ticket, one monitoring unit" file. Face recognition unit: Employs an ArcFace-based face feature extraction algorithm, whose loss function is expressed as:

[0037] Where (s) is the feature scaling factor, and (m) is the angle margin. This refers to the perspective of the current sample and its corresponding class center. Real-time face detection is performed using the MobileNetV3 lightweight network, achieving face comparison at over 30 frames per second on embedded devices. Real-time matching is then performed with a unified access control platform to achieve "card, ID, and face consistency" verification. When personnel enter the work area, cameras capture their faces, and a personnel-access control matching model is used to compare these images with data from the unified platform. Once "card, ID, and face consistency + monitoring coverage" are met, entry is permitted.

[0038] Location Monitoring Unit: Pre-work monitoring and verification: The person in charge must use a mobile device to confirm the start-up conditions, including whether the monitoring angle and safety measures meet the requirements. Work access can only be unlocked after all requirements are met; In-work monitoring + real-time model monitoring: The AI ​​vision model analyzes violations frame by frame and adjusts the recognition sensitivity according to the work risk level, promptly freezing violations and triggering alarms; Work is suspended upon monitoring interruption: Once monitoring fails, the system immediately locks the work ticket and notifies relevant personnel. Work can only be restarted after monitoring is restored and verified by the large model; Behavior analysis unit: Based on the improved YOLOv7-tiny model, the ECA-Net attention mechanism is introduced to enhance small target detection capabilities. The loss function adopts CIoULoss.

[0039] in, (c) represents the distance between the center point of the predicted bounding box and the ground truth bounding box, (v) represents the length of the diagonal of the minimum bounding rectangle, and (c) represents the aspect ratio consistency metric. The small target detection capability has been optimized for power distribution room scenarios. This model supports real-time detection of safety helmet wearing, workwear recognition, absence from duty / sleeping on duty recognition, and violations (such as unlicensed work and work outside designated areas), with an accuracy rate ≥95%. Instrument and Status Recognition Unit: This unit employs a CRNN+Attention structure for instrument digit recognition and enhances the model's ability to interpret readings from pointer-type instruments. Convolutional layers extract image features, recurrent layers (LSTM) model sequence dependencies, and the attention mechanism focuses on the pointer or digit region, outputting the recognition result. Its forward propagation process can be represented as:

[0040] It supports real-time identification of various targets such as ammeters, voltmeters, and indicator light status.

[0041] Operation Retention Unit: When the operation is completed, the system automatically generates a dual archive including full video recording. Failure to archive as required will prevent the completion of the ticket archiving process. Responsibility Tracing Unit: Inadequate personal protection: Camera capture → Small model recognition → Large model association with ticket freezing permissions → Push to responsible person for rectification → Take a photo and upload after rectification → Small model verification → Unlock ticket Working without a ticket in the monitoring area: Camera recognizes face → Small model matches no ticket → Large model triggers serious alarm → Push to safety supervision and industry safety department → On-site removal → System records violation → Included in contractor / individual assessment. Work exceeding the permitted scope: Camera locates personnel position → Small model compares with permitted area on ticket → Large model generates handling suggestions → Ticket suspension → On-site confirmation by person in charge → Scope adjustment → Large model reviews and restarts. Absenteeism / Sleeping on Duty under Monitoring: Camera behavior analysis → Small model identifies abnormal status → Large model associates with core position information in ticket → Pushes to team leader to arrange replacement → Verifies cause → Ticket archive records rectification process. Outsourced Personnel and Access Management Unit: Outsourced personnel must be approved and formally authorized. Authorization permissions must clearly define the scope and time limit of work; exceeding the scope of inspection work is strictly prohibited. The system supports temporary authorization management, with an authorization validity period not exceeding one day. Daily reassessment and approval are required, and all entry and exit records are kept for at least one year. Electronic access control is linked to permissions; unauthorized or expired personnel cannot enter the work area. Operation Ticket Review and Shift Handover Mechanism Unit: Operation tickets must be jointly reviewed and confirmed by the supervisor and the shift leader to ensure the accuracy of the steps; the incoming shift personnel must re-examine the operation ticket content to ensure complete information connection and prevent handover omissions; the system supports multi-role collaborative confirmation and electronic signatures to form a traceable review record; Manual Supervision and Robot Collaboration Unit: When the robot performs key operations such as switching, a designated person must supervise the entire process with the operation ticket; key operation nodes require manual confirmation of the consistency between equipment status and instructions; in case of abnormalities, the supervisor should intervene in a timely manner to ensure safe and controllable operation; Multi-Source Equipment Status Confirmation Unit: The "non-same-source dual confirmation" principle is adopted to prevent misjudgment caused by a single signal source failure; cross-verification is performed by combining multi-source information such as mechanical position indication, electrical signals, monitoring screens, and sensor data; Furthermore, the multi-sensor fusion module is integrated into the robot body and the monitoring pod, used to collect multi-dimensional information about the environment and equipment, including: an infrared thermal imager: employing an uncooled focal plane array with a resolution of 640×512 and a temperature measurement range of -20℃ to 550℃, used to detect abnormal heating in electrical equipment joints, bushings, and other parts; a gas sensor: using an electrochemical SF6 sensor with a detection range of 0~2000ppm and a response time of <30 seconds; O2 and CO2 sensors using NDIR (non-dispersive infrared) technology to achieve continuous monitoring of ambient gas concentrations; and a humidity sensor: employing a high-precision capacitive sensing element with a measurement range of 0~100%. RH, with an accuracy of ±2%RH, is deployed inside the pod to monitor ambient humidity in real time and participate in insulation status assessment; Smoke detection unit: includes a laser scattering smoke sensor, which, together with a stepper motor-driven worm gear mechanism, adjusts the fan angle to achieve active aspiration smoke detection with a sensitivity of up to 0.1%obs / m; Partial discharge detector: uses an ultra-high frequency (UHF) sensor with a frequency band of 300MHz~1.5GHz, and uses wavelet noise reduction algorithm to extract and identify partial discharge signals; Collision avoidance sensor: uses multiple sets of infrared and ultrasonic sensors, covering the four directions of the pod (front, back, left, and right), with a detection distance of 0.1~3m, to achieve dynamic obstacle avoidance and collision protection.

[0042] Furthermore, the communication control module, the "brain" of the robot, includes: a main controller: an embedded industrial computer using the ARM Cortex-A72 architecture, running a Linux system, responsible for task scheduling, motion control, data fusion, and real-time decision-making; a communication unit: supporting wired communication via a seamless sliding contact line using the Modbus TCP / IP protocol, and also supporting 5G SA networking and Wi-Fi 6 wireless communication, achieving low-latency, high-bandwidth data backhaul and command issuance; and an edge computing unit: equipped with NVIDIA Jetson Xavier NX, with a built-in TensorRT acceleration engine, capable of locally running lightweight AI models such as YOLOv7-tiny and MobileNetV3, enabling real-time analysis of video streams and violation recognition.

[0043] Furthermore, the power module includes a dual-redundant design employing a sliding contact line power supply and a lithium battery. The lithium battery has a capacity of 48V / 50Ah, supporting continuous operation for at least 2 hours after the robot reports a power outage. The power management unit (PMU) enables real-time monitoring of voltage and current, overload protection, and supports remote power status query and fault diagnosis.

[0044] Furthermore, the emergency power supply response module includes: a real-time monitoring and analysis unit: the system accesses a lightweight behavior recognition model based on YOLOv5 to continuously analyze monitoring video streams and automatically identify high-risk abnormal behaviors such as open flames / smoke, prolonged stillness of personnel (suspected fainting), and falls from heights; an emergency alarm unit: once an emergency is confirmed, the system generates structured alarm information in JSON format within 10 seconds, including: event type, occurrence time, camera location, associated work ticket ID, and information of the personnel involved, and pushes it to the backend platform and mobile terminals via the MQTT protocol; an alarm information push unit: alarm information is simultaneously pushed to the mobile terminals of the on-site supervisor, safety supervisor, and regional duty personnel, supporting one-click confirmation and rapid response, and supporting alarm push through multiple channels such as DingTalk, WeChat Work, and SMS to ensure that relevant personnel receive and respond in a timely manner; and an alarm information archiving unit: all alarm records are automatically archived to the corresponding ticket file as a basis for subsequent review and improvement, strengthening the role of "no operation without monitoring" as a safety net in extreme scenarios. All alarm records are automatically associated with work ticket IDs and stored in a time-series database, supporting subsequent query and statistical analysis.

[0045] Furthermore, the aforementioned intelligent management and control platform is deployed in a remote monitoring center, working in conjunction with the robot system to achieve intelligent management and control of personnel, equipment, processes, and data. Specifically, it includes: a personnel-permission dynamic matching and auditing subsystem: using the graph database Neo4j to store the complex relationships between personnel, equipment, work areas, and permissions. Compared to traditional relational databases, graph databases can more efficiently traverse and query multi-layered network relationships, thereby achieving real-time, accurate matching and dynamic auditing of personnel qualifications and work permissions, providing core technical support for the automated verification of consistency between "person-position-certificate-ticket". A panoramic digital management and control subsystem for work processes: the platform as a whole is built using a microservice architecture, decomposing business functions such as work invoicing, monitoring binding, process review, and archiving auditing into a set of independent, loosely coupled services. This architecture allows each part of the system to be developed, deployed, and expanded independently, greatly improving the platform's reliability, maintainability, and elastic scalability when facing multi-task concurrency and high-frequency data interaction. Big Data Analytics and Intelligent Decision Support Engine: Based on the Apache Spark distributed computing framework, this engine performs high-speed processing and analysis of massive historical inspection data (such as instrument reading time series, temperature change trends, and alarm event records). Leveraging Spark's powerful in-memory computing and machine learning libraries, the platform can perform complex clustering analysis, trend prediction, and pattern recognition, thereby enabling in-depth assessment of equipment health status and automatic generation of predictive maintenance strategies, driving the evolution of safety management from "post-event response" to "pre-event warning."

[0046] Furthermore, Figure 2This is a schematic diagram of the overall architecture of the AI ​​monitoring robot system for safety inspection of power distribution rooms provided in this embodiment of the invention. As shown in the figure, the system is built on a four-layer architecture of "perception-transmission-platform-application". The perception layer integrates multiple sensors such as high-definition vision, infrared thermal imaging, gas, and UWB positioning through the robot to achieve comprehensive and high-precision perception of the power distribution room environment and equipment status. Its design capacity supports more than 500 monitoring points and can switch to backup power within 0.1 seconds after the main power supply fails, ensuring at least 15 minutes of emergency operation capability in critical scenarios. The transmission layer adopts 5G private network and industrial Ethernet to ensure real-time and reliable data synchronization to the platform and supports real-time calibration of UWB positioning data. The platform layer, as a digital twin and intelligent decision engine, uses a large and small model collaboration mechanism (real-time detection of small edge models and online learning of large cloud models) for data processing and risk assessment, and uses the ICP algorithm to map physical world information to virtual models, achieving high-precision digital mapping with a positioning error of less than 10 centimeters. The application layer ultimately forms a complete safety closed loop of "discovery → alarm → location → handling → traceability".

[0047] Figure 3This is a cross-sectional schematic diagram of the mechanical structure of the track walking module in this embodiment of the invention. The diagram shows in detail the robot's suspension walking and lifting mechanism. The robot body (3) is suspended on the I-shaped track (1) by the cantilever assembly (4). The walking drive unit (5) is driven by a servo motor through a planetary reducer (reduction ratio 1:50) to drive the gear system, which in turn drives the walking wheels to move on the track. The lifting drive unit (6) controls the winding wheel through a similar transmission system to realize the vertical lifting of the detector pod (7). The key design is the use of a non-contact sliding contact line (2) for power supply and data transmission. Its conductivity is ≥98% IACS, width is 120mm, height is 80mm, and conductor diameter is ≥8mm, which can adapt to the high dust environment of the power distribution room. The walking wheels and the track slot form a three-point contact, which, together with the spring preload (1.5N±0.1N), ensures stable conductivity and mechanical locking. In terms of safety mechanism, an electromagnetic brake is installed at the output end of the planetary reducer. When communication is interrupted for more than 1 second or the robot tilt angle is greater than 3°, a braking torque of ≥5Nm can be applied within 0.5 seconds to achieve emergency stop. The robot body (3) integrates a main controller (9) which is responsible for coordinating the actions of the walking drive unit (5) and the lifting drive unit (6) and monitoring the robot tilt angle, communication status and other operating parameters in real time. The data analysis center (11) is connected to the detection pod (7) to process and analyze the collected images and environmental parameters (such as temperature and partial discharge signals) in real time. The processing results are transmitted remotely to the monitoring station through the communication unit (10) using an industrial-grade wireless module to achieve two-way data interaction. In terms of safety mechanism, an electromagnetic brake (8) is installed at the output end of the planetary reducer of the walking drive unit (5). It adopts a normally closed design. When communication is interrupted for more than 1 second or the robot tilt angle is greater than 3°, the power is immediately cut off to trigger braking and a braking torque of ≥5Nm is applied within 0.5 seconds to ensure emergency stop.

[0048] Figure 4 This is a block diagram of the improved YOLOv7 visual recognition model network structure in this embodiment of the invention. The diagram illustrates a lightweight model architecture optimized for small target recognition in power distribution rooms (such as indicator lights and meter cracks). In the backbone network, CSPDarknet combined with the CBAM attention mechanism is used to enhance feature extraction capabilities. The neck network uses BiFPN and PANet structures for multi-scale feature fusion, significantly improving the recall rate of small targets. The head network outputs predictions at three scales: 13×13, 26×26, and 52×52, with the 52×52 high-resolution feature layer specifically used to detect tiny 2×2 pixel-level indicator lights. After optimization, the model is compressed to 15MB on embedded devices (such as Jetson Nano), achieving an inference speed of 28 FPS, and improving the average precision (mAP) for small target detection from 72.1% to 85.6%, effectively solving the problem of missed detection of small targets in complex backgrounds.

[0049] Figure 5 This is a logic flowchart of multi-sensor fusion detection and emergency response in this embodiment of the invention. The diagram clearly depicts the decision-making process from multi-source data acquisition to closed-loop handling. The system continuously collects temperature (threshold <40℃), humidity (threshold 30%~70%), and SF6 gas pressure. The rated value is 0.6 MPa, the alarm value is 0.55 MPa, and the lockout value is 0.5 MPa. During normal operation, the gas pressure should be maintained at the rated value of 0.6 MPa; when the pressure is below 0.6 MPa, the system is in a non-ideal state, but an alarm has not yet been triggered. When the pressure drops to 0.55 MPa (alarm value), the system issues an alarm signal, prompting maintenance personnel to handle the situation promptly. If the pressure continues to drop to 0.5 MPa (lockout value) or below, the system will automatically lock the switch operation, prohibiting opening and closing actions, and the equipment will lose its protection capability. Data such as smoke and visual recognition results (e.g., insulator damage probability >0.9) are also collected. The core principle is "non-same-source dual confirmation," meaning that an alarm is triggered only when both independent signal sources (such as mechanical and electrical signals, or visual recognition and sensor data) confirm an anomaly, effectively preventing misjudgments caused by a single signal source failure. For example, when gas leakage and temperature anomalies are detected simultaneously, the system determines it as a risk of equipment overheating or insulation failure. For smoke detection, a "fan enhancement" strategy is adopted: when the initial concentration reaches the threshold, the fan is activated for active sampling. If the secondary detection still exceeds the fire point threshold, it is confirmed as a real fire; otherwise, a 30-second delay is used to clear the false alarm. After confirming a real anomaly, the process enters a closed-loop handling phase: on-site audible and visual alarms, platform pop-ups, UWB coordinates overlaid on the digital twin model for precise positioning, robot-triggered physical markers, platform-pushed intelligent guidance, and finally, all data is recorded and archived, with difficult cases used for AI model iteration and updates.

[0050] Figure 6This is a schematic diagram of the "no operation without monitoring" business process of the robot system and the back-end intelligent management and control platform in this embodiment of the invention. The diagram shows the complete technical closed loop of operation safety management and control. The process begins on the platform side: the work ticket is initiated through the two-ticket system interface (17), and the back-end intelligent management and control platform (16) performs three mandatory checks in sequence: 1) whether the work area is within the robot's monitoring range; 2) whether the person in charge's qualifications are valid (interfacing with the HPC system); 3) whether the operation ticket has been electronically reviewed by the supervisor and the duty supervisor. If any check fails, the system will freeze the ticket and notify the safety officer. Only after all access conditions are met will the task be officially issued to the robot. On the robot side, the robot receives the task and walks to the target location, performs face and permission verification on the operator through the visual recognition unit, and performs anomaly detection in real time during the operation (such as personnel exceeding authority, violation of operation). Once the monitoring is interrupted or the robot leaves the monitoring area, the system immediately suspends the operation and notifies the supervisor. When the operation ends, the system automatically archives the entire video and operation log (including digital signature) to the data storage center. All records are kept for no less than 365 days to form a strictly traceable safety file.

[0051] In one embodiment of the present invention, a smart power plant plans to conduct regular inspections and temperature measurements on the "No. 3 main transformer switchgear" in its 6kV power distribution room. After the system is started, the robot autonomously maps along track 1 (S1), using its onboard 2D LiDAR and visual SLAM algorithm to construct a high-precision two-dimensional grid map and a three-dimensional point cloud map of the power distribution room, and automatically marks the location of all cabinets.

[0052] The inspection task is generated by the intelligent management platform 16. The platform obtains the operation information through the two-ticket system interface 17 and automatically performs four access verifications (S2): Monitoring coverage verification: automatically verifies whether the operation area is within the robot's monitoring range; Personnel qualification verification: automatically verifies the qualifications of the person in charge of this operation and related personnel; Environmental safety verification: the platform drives the robot to scan or retrieve the latest sensor data of the target area through instructions to verify whether the SF6 gas concentration is within the safe range (<0.55MPa), whether the oxygen content is normal (>19.5%), whether there is smoke or open flame warning, and whether the ambient temperature exceeds the warning threshold (e.g., >40℃). If any environmental safety indicator is abnormal, the system will determine that the environment is "unsafe" and freeze the operation ticket, while generating an early warning work order and pushing it to the maintenance team; Process compliance verification: confirms that the operation ticket has been electronically reviewed by the supervisor and the duty supervisor. After all four verifications are passed, the platform officially issues the inspection task to the robot main controller 14.

[0053] Furthermore, after receiving the instruction, the robot's walking drive unit 5 is activated, moving to the target cabinet according to the planned path. The specific mechanical transmission structure and safety braking mechanism of the walking drive unit are as follows: Figure 3As shown, its highly reliable contactless power supply and three-point contact design ensure continuous and stable operation in complex environments. During movement, the communication unit 15 maintains real-time data synchronization with the platform via the sliding contact line 2. Upon arrival at the inspection point, the lifting drive unit 6 is activated, controlling the monitoring pod 7 to descend to the instrument's visible height.

[0054] The visual recognition module begins operation: High-definition camera 8 captures images of the switch cabinet panel, and these images are analyzed in real-time by the edge computing unit. The improved lightweight YOLOv7-tiny model used has network structure and performance optimization details as follows: Figure 4 As shown, this model has been specifically enhanced for small targets such as indicator lights and small instruments in the power distribution room, achieving a balance between high accuracy and high frame rate. The improved YOLOv7-tiny model identifies targets such as ammeters, voltmeters, and circuit breaker indicator lights, and the CRNN model reads the ammeter pointer reading as "315A". Simultaneously, the infrared thermal imager 9 acquires a temperature distribution map of the cabinet and detects that the temperature of a certain connector is 85℃ (exceeding the threshold of 75℃). The gas sensor 12 detects that the SF6 concentration is normal, and the smoke sensor 13 does not trigger an alarm.

[0055] The multi-sensor data is fused and analyzed by the main controller 14 to determine that the joint has an overheating defect. The core logic of multi-sensor data fusion and emergency response is as follows: Figure 5 As shown, the system uses the "non-same-source double confirmation" principle for anomaly detection, which greatly reduces the false alarm rate. The robot marks the anomaly as a "general defect" and packages the data into JSON format via the communication unit 15, then uploads it to the platform 16 via the 5G network. At the same time, the main controller 14 sends a command to the robotic arm 10 to control the physical marking mechanism 11 to affix an "overheated and pending inspection" mark next to the corresponding cabinet door number.

[0056] After receiving the data, platform 16 automatically associates the defect with the equipment ledger, generates a maintenance work order, and pushes it to the maintenance team. The entire process, including video and data, is archived to the data storage and analysis center 18.

[0057] If the smoke sensor 13 detects an anomaly during the inspection, the fan assembly activates enhanced detection, and upon confirmation of smoke, immediately triggers the highest-level alarm. This process follows... Figure 5 The detailed process of "enhanced smoke detection and false alarm handling for wind turbines" is described in the document. The robot triggers an on-site audible and visual alarm, and the platform sends alarm information including video clips to multiple parties, such as safety officers and the fire control room, via the MQTT protocol. It also activates the emergency plan, and the robot can switch to voice guidance mode to assist on-site personnel evacuation.

[0058] Throughout the entire process, from work ticket issuance, authorization verification, process monitoring to archiving and traceability, the complete business process strictly adheres to... Figure 6The defined "no operation without monitoring" control loop allows each module to be uniformly scheduled and coordinated through the main controller 14: the hardware communicates internally with each other via CAN bus and RS485, and the software system uses the ROS (Robot Operating System) framework for task orchestration and message passing, realizing a fully closed-loop automated inspection of perception, decision-making and execution.

[0059] In one embodiment of the present invention, a scenario for outsourced personnel inspection and access management is described: Personnel from an outsourced unit need to enter a power distribution room to perform equipment inspection tasks. Their supervisor submits a work application via a mobile device. The system automatically verifies whether the personnel have been approved and authorized, and checks the scope and timeliness of their work. Simultaneously, the system detects the real-time environmental safety status of the target work area (e.g., gas, temperature, smoke). If the environment is unsafe, even if the personnel's permissions are valid, the system will refuse to generate electronic access control permissions and display a message stating "Safety risks exist in the work area; entry is prohibited." If the authorization is valid, the system automatically generates temporary electronic access control permissions and binds them to the personnel's identity information. When personnel enter, a dual verification process is performed using facial recognition and a permission matching model. Entry is only permitted after confirming that the "person-identity-ticket-permission" matches. During the work, a robot monitors their behavior in real time. If the work exceeds the authorized scope, the system immediately alarms and freezes their ticket permissions, while simultaneously notifying the on-site supervisor and safety management personnel to intervene. After the work is completed, the system automatically archives the inspection records and videos. All operation logs are stored for at least one year, supporting post-event auditing and accountability.

[0060] In one embodiment of the present invention, a human-machine collaborative monitoring scenario is described: During a switching operation, the robot moves to the corresponding switchgear as instructed, preparing to perform the opening operation. A supervisor, holding an operation ticket, monitors the entire process on-site and confirms the robot's positioning via a mobile device. Before opening the switch, the robot confirms the mechanical position of the switch and the status of electrical indicator lights using a visual recognition unit, and performs "non-same-source dual confirmation" based on sensor data (such as current and voltage). After confirmation, the supervisor clicks "Confirm Execution" on the mobile device, allowing the robot to perform the opening action. If communication interruption, abnormal sensor data, or inconsistent visual recognition occurs during the process, the system automatically pauses the operation and prompts the supervisor to intervene for inspection and handling. All operation steps, confirmation records, timestamps, and operator and supervisor information are automatically linked to the electronic operation ticket, forming a complete and traceable operation file.

[0061] The embodiments of this invention also have the following technical effects: Full automation and intelligence of the inspection process: From path planning, data collection, status recognition to report generation, the entire process requires no manual intervention, greatly improving efficiency and reducing labor costs. High-precision, lightweight visual recognition: Adopting an improved YOLOv7 model, it has been specifically optimized to address the difficulty of identifying small targets in power distribution rooms, achieving high-precision real-time detection on embedded devices and solving the problems of missed and false detections in complex scenarios. Multi-dimensional, proactive safety perception: Integrating multiple sensors such as temperature, gas, and smoke, especially the enhanced smoke detection design for fans, it achieves early, proactive, and sensitive detection of risks such as electrical fires and insulation faults. Deep integration of safety management systems: It creatively realizes intelligent linkage between the inspection robot and the "two-ticket" system, forcibly implementing "no operation without monitoring" through technical means, elevating safety procedures from "human prevention" to a new level of "technical prevention," eliminating management blind spots. Forming a complete emergency closed loop: Not only does it "discover" anomalies, but it also forms a complete safety closed loop of "discovery-alarm-location-response-tracing" through on-site alarms, physical marking, intelligent guidance, and platform linkage, significantly improving emergency response speed and handling effectiveness. High system scalability: Adopting a modular design, sensors and functional modules can be flexibly selected according to the needs of different power distribution rooms. The AI ​​model can also continuously evolve through online learning, adapting to new equipment types and defect patterns.

[0062] To achieve the above embodiments, such as Figure 7 As shown, this embodiment also provides a multi-axis robot device 10 for AI monitoring of power distribution room safety inspection, including: The 3D spatial map construction and path planning module 100 is used to construct a high-precision 3D spatial map of the power distribution room and plan the inspection path. It can autonomously move along the track with multiple sensors on a track-type mobile platform to generate multi-source heterogeneous sensor data. The edge AI video analysis and dynamic recognition module 200 is used to analyze the video stream of the work area in real time based on the lightweight AI model deployed on the edge computing unit, and to perform dynamic recognition of personnel identity, behavior and location by combining UWB positioning data and digital twin system, and to generate real-time monitoring status information. The multi-source data fusion and anomaly warning module 300 is used to fuse real-time monitoring status information and multi-source heterogeneous sensor data, and adopts a non-same-source dual confirmation mechanism to cross-verify the device status and environmental parameters, trigger anomaly warnings and generate structured alarm information. The intelligent control module 400 is used to link the edge model and the cloud-based big model through the background intelligent control platform, execute anomaly handling instructions and record the handling process.

[0063] This invention discloses an AI-controlled multi-axis robot device for safety inspection of power distribution rooms. It enables comprehensive real-time monitoring and multi-source data fusion analysis of the 6kV power distribution room work area. Through the collaboration of large and small models and physical marking mechanisms, it effectively forms a safety closed loop of "discovery-early warning-handling-tracing", significantly improving the accuracy of violation identification and emergency response efficiency, and reducing the risk of human operation.

[0064] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 8 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the above-described method for AI monitoring multi-axis robot for safety inspection of power distribution rooms.

[0065] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a multi-axis robot method for AI monitoring of power distribution room safety inspection as described in the foregoing embodiments.

[0066] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0067] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for AI-controlled multi-axis robot monitoring of power distribution room safety inspection, characterized in that, include: S1: Construct a high-precision three-dimensional spatial map of the power distribution room and plan the inspection route. The track-mounted mobile platform carries multiple sensors and moves autonomously along the track to generate multi-source heterogeneous sensor data. S2, based on a lightweight AI model deployed on an edge computing unit, analyzes the video stream of the work area in real time, and combines UWB positioning data with a digital twin system to dynamically identify personnel identity, behavior and location, and generate real-time monitoring status information. S3 integrates real-time monitoring status information with multi-source heterogeneous sensor data, and adopts a non-same-source dual-confirmation mechanism to cross-verify the device status and environmental parameters, triggering abnormal warnings and generating structured alarm information. S4, through the background intelligent management and control platform, links the edge model and the cloud big model to execute anomaly handling instructions and record the handling process.

2. The method as described in claim 1, characterized in that, S1 includes: Track and power supply unit: The I-shaped track laid on the top of the power distribution room, and the seamless sliding contact line laid along the track, are used to provide continuous power supply and data communication channel for the robot; The walking drive unit includes a first drive motor, a first planetary reducer, a transmission system consisting of five meshing first gears, and two walking wheels that engage with the side of the track. The walking speed satisfies the formula: in, This is the reduction ratio of the first planetary gearbox. , , The number of teeth for the driving, intermediate, and driven gears are respectively. The radius of the traveling wheel, This is the speed of the first drive motor.

3. The method as described in claim 1, characterized in that, The S2 includes: The face feature extraction algorithm based on ArcFace is used, and its loss function is expressed as: Where (s) is the feature scaling factor, and (m) is the angle margin. The angle between the current sample and its corresponding class center; Based on the improved YOLOv7-tiny model, the ECA-Net attention mechanism is introduced to enhance small object detection capabilities, and the loss function adopted is CIoULoss. in, (c) is the distance between the center points of the predicted bounding box and the ground truth bounding box, (v) is the length of the diagonal of the minimum bounding rectangle, and (v) is the aspect ratio consistency measure.

4. The method as described in claim 1, characterized in that, The S3 includes: The instrument and status recognition unit employs a CRNN+Attention structure for instrument digit recognition and enhances the model's ability to interpret readings from pointer-type instruments. Convolutional layers extract image features, recurrent layers model sequence dependencies, and the attention mechanism focuses on the pointer or digit region, outputting the recognition result. Its forward propagation process is represented as follows: It supports real-time identification of various targets, including ammeters, voltmeters, and indicator lights.

5. The method as described in claim 1, characterized in that, The S4 includes: The physical marking mechanism is controlled to mark the faulty equipment. The physical marking mechanism is a coding machine, a stamping device or a label attaching mechanism. The marking content includes the type of abnormality, timestamp and equipment number. The alarm information is encapsulated in JSON format and pushed to the backend platform and mobile terminal via the MQTT protocol. The JSON structure includes the event type, occurrence time, camera location, associated work ticket ID, and information of the personnel involved.

6. The method as described in claim 1, characterized in that, The method further includes: It adopts a dual-redundant design of sliding contact line power supply and lithium battery. The lithium battery capacity is 48V / 50Ah, which supports continuous operation for no less than 2 hours after the robot reports a power failure. Power Management Unit: Enables real-time monitoring of voltage and current and overload protection, and supports remote power status query and fault diagnosis.

7. The method as described in claim 1, characterized in that, The method further includes: The environmental safety verification process involves using commands to drive a robot to scan sensor data in the target area to determine if the SF6 gas concentration is below a certain level. Is the oxygen content higher than 19.5%? Is the ambient temperature lower than 19.5%? If any indicator is abnormal, the work ticket will be frozen and an early warning work order will be generated.

8. A multi-axis robot device for AI monitoring of safety inspection in power distribution rooms, characterized in that, include: The 3D spatial map construction and path planning module is used to build a high-precision 3D spatial map of the power distribution room and plan the inspection path. It uses a track-type mobile platform to carry multiple sensors and move autonomously along the track to generate multi-source heterogeneous sensor data. The edge AI video analysis and dynamic recognition module is used to analyze video streams in the work area in real time based on lightweight AI models deployed on edge computing units. It combines UWB positioning data and digital twin systems to dynamically identify personnel identity, behavior and location, and generate real-time monitoring status information. The multi-source data fusion and anomaly warning module is used to fuse real-time monitoring status information and multi-source heterogeneous sensor data. It adopts a non-same-source dual-confirmation mechanism to cross-verify the device status and environmental parameters, trigger anomaly warnings, and generate structured alarm information. The intelligent management and control module is used to link the edge model and the cloud-based big model through the background intelligent management and control platform, execute anomaly handling instructions and record the handling process.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the AI ​​monitoring multi-axis robot method for safety inspection of power distribution rooms as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for AI monitoring multi-axis robot for safety inspection of power distribution rooms as described in any one of claims 1-7.

Images