A power distribution room inspection method based on a wheeled robot

Abstract

The application discloses a power distribution room inspection method based on a wheeled robot, and comprises the following steps: step S1, inspection point calibration and task configuration; step S2, multi-source data acquisition; step S3, multi-dimensional state identification; step S4, abnormality checking and grading; and step S5, result output and alarm. The application overcomes the problems of low frequency, limited detection items, high failure omission rate, great safety hazards and non-standard inspection records of traditional manual inspection.

Description

Technical Field

[0001] This invention relates to the operation and maintenance inspection technology of sewage treatment power distribution rooms, and more specifically, to a power distribution room inspection method based on a wheeled robot. Background Technology

[0002] The wastewater treatment plant's power distribution room (or distribution center) is the core power supply hub of the plant area, responsible for receiving, distributing, and protecting electrical energy to ensure the continuous and stable operation of the wastewater treatment process. It is typically designed as a secondary load, requiring dual-circuit 10kV power supply and an emergency diesel generator. It houses high-voltage switchgear, transformers, low-voltage distribution panels, and protection and control devices, supplying power to critical equipment such as booster pumps, blowers, and sludge dewatering machines. Power outages could lead to the collapse of the biological treatment system, excessive emissions, or even safety accidents.

[0003] In the existing component selection method, the inspection of the power distribution room of the wastewater treatment plant is usually carried out on a fixed schedule (such as daily or per shift) by staff entering the power distribution room to check the meter readings, switch status, indicator lights, temperature, and for any abnormal noises or odors, and manually record the data. Overall, this method suffers from low inspection efficiency and is prone to data recording errors.

[0004] Therefore, the traditional manual inspection of power distribution rooms has the following shortcomings: numerous inspection points require manual inspection and recording of each unit with tools; overall inspection efficiency is low, making it difficult to respond quickly to emergencies; and data recording is prone to subjective errors or omissions. Furthermore, existing power distribution room inspections mainly rely on regular manual inspections, which suffer from low inspection frequency, limited inspection items, high requirements for personnel professional skills, and non-standard inspection records. The routine frequency of manual inspections is only 1-2 times per week. Some latent defects, such as early partial discharge and minor overheating, are difficult to detect through manual inspection, resulting in a fault omission rate of over 30%, posing serious safety hazards. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a power distribution room inspection method based on a wheeled robot, overcoming the problems of low frequency of traditional manual inspections, limited inspection items, high failure rate, significant safety hazards, and non-standard inspection records.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for inspecting a power distribution room based on a wheeled robot includes the following steps:

[0008] Step S1: Inspection point location marking and task configuration;

[0009] Step S2: Multi-source data acquisition;

[0010] Step S3, multi-dimensional state recognition;

[0011] Step S4, anomaly verification and classification;

[0012] Step S5: Output the results and issue alarms.

[0013] Preferably, step S1 specifically includes:

[0014] All the aforementioned inspection points within the power distribution room are pre-marked;

[0015] An improved A* algorithm is used to plan the inspection path;

[0016] Configure the inspection frequency and inspection items according to the inspection requirements.

[0017] Preferably, the A* algorithm is as follows:

[0018] In standard A*, the evaluation function for each node is:

[0019]

[0020] in, From the starting point to the node The actual cost of movement; For the node Heuristic estimation of the target point;

[0021] Frequent turns during the inspection path reduce movement efficiency and increase energy consumption, so a turning penalty is introduced:

[0022]

[0023]

[0024] in, The turning penalty coefficient is set to 0.1~0.5. The steering radius coefficient is 0.2 for a 45° turn, 0.5 for a 90° turn, 0.8 for a 135° turn, and 1.0 for a 180° turn. This represents the step size from the parent node to the current node.

[0025] Preferably, the calibration of all the inspection points includes recording the device type, normal status parameters, and thresholds;

[0026] The calibration time for a single inspection point is ≤30 seconds;

[0027] The time required to plan the inspection path is ≤10 seconds, and the repetition rate of the inspection path is ≤1%.

[0028] The inspection frequency is from 15 minutes to 30 days.

[0029] Preferably, step S2 specifically includes:

[0030] The wheeled robot moves along the inspection path and automatically adjusts its body posture and sensor angle upon reaching each inspection point. It then sequentially collects visible light images, infrared thermal images, and ultrasonic partial discharge signals from the equipment, and simultaneously records the coordinates and collection time of each inspection point.

[0031] Preferably, the wheeled robot uses laser SLAM + QR code for assisted positioning, with a positioning accuracy of ±1cm;

[0032] The pitch angle adjustment range of the wheeled robot's gimbal is -90° to +90°, and the horizontal rotation range is 0° to 360°.

[0033] The data time synchronization error of each sensor on the wheeled robot is ≤50ms.

[0034] Preferably, step S3 specifically includes:

[0035] The visible light image is input into the multi-task recognition model, which outputs the indicator light status, switch position, and knob gear information, and compares them with the preset normal state.

[0036] Preferably, the multi-task recognition model is designed to recognize the color of indicator lights, the on / off position of switches, and the position of knobs in visible light images. It employs a deep learning-based multi-task target detection and classification framework; the specific process is as follows:

[0037] (1) Data collection and labeling

[0038] Acquire visible light images of the device under different lighting conditions and angles;

[0039] Label each target area (indicator light, switch, knob) with its bounding box and category:

[0040] Indicator lights: Red, Green, Yellow, Off (optional)

[0041] Switch: Opening, closing

[0042] Knob: Gear 1, Gear 2, ... (depending on the actual number of gears)

[0043] (2) Model design

[0044] Backbone network: Lightweight CNN is used to extract image features.

[0045] Region proposal: If a two-stage detector (Faster R-CNN) is used, candidate regions are first generated by the RPN; if a single-stage detector (YOLOv8) is used, bounding boxes and categories are predicted directly through a grid.

[0046] Multi-task branching: Three independent sub-networks are connected in parallel on a shared feature map or RoI feature:

[0047] Indicator light branch: Fully connected layer + Softmax (outputs red / green / yellow / other)

[0048] Switching branch: Fully connected layer + Sigmoid (Binary classification: Opening / Closing)

[0049] Knob branch: Fully connected layer + Softmax (outputs the probability of each gear position)

[0050] Loss function: joint optimization ,in, For bounding box regression loss and object confidence loss (such as CIoU); This is either cross-entropy loss or binary classification loss;

[0051] (3) Training and Reasoning

[0052] During training, transfer learning (pre-trained on ImageNet) was used, and data augmentation (rotation, brightness adjustment, affine transformation) was employed to improve generalization ability;

[0053] During inference, a single image is input, and the model outputs for each detected target: bounding box position, object category (indicator light / switch / knob), corresponding state, and confidence level (maximum probability output by Softmax or Sigmoid output value).

[0054] (4) Post-processing

[0055] Non-maximum suppression (NMS) removes overlapping bounding boxes;

[0056] For each target, the state with the highest confidence level is taken as the final output; if the confidence level is lower than the preset threshold, it is marked as "uncertain" and manual review or re-sampling is triggered.

[0057] Preferably, step S4 specifically includes:

[0058] After an anomaly is detected, the system automatically retrieves and compares the historical inspection data of the last three inspections for that inspection point. Once the anomaly is confirmed to be persistent, it is classified according to the type of anomaly.

[0059] Preferably, if the anomaly type involves a change in the state of a switch or knob, it needs to be compared with the SCADA system operation record simultaneously to exclude state changes caused by legitimate operations. For anomalies that cannot be confirmed, a manual review process should be automatically triggered.

[0060] Preferably, step S5 specifically includes:

[0061] Generate structured inspection reports, including inspection data and anomalies for all equipment;

[0062] The alarm is pushed to the corresponding maintenance personnel according to the alarm level, and an emergency alarm will trigger an audio and visual prompt simultaneously.

[0063] Inspection data is automatically stored in the management platform, forming a health record for the entire life cycle of the equipment.

[0064] This invention provides a power distribution room inspection method based on a wheeled robot, overcoming the problems of low frequency, limited inspection items, high failure rate, significant safety hazards, and non-standard inspection records associated with traditional manual inspections. It offers a fully automated inspection method based on a wheeled robot, achieving high-frequency, all-weather, multi-dimensional (visible light, infrared) equipment status identification and hierarchical alarms. This shortens fault detection time, reduces the risk of electric shock to personnel and maintenance costs, and supports integration with SCADA systems for error prevention verification and standardized report generation, thereby improving the safety and reliability of the power distribution system. The robot can replace manual entry into high and low voltage areas. Inspections are recorded in real time through video, images, temperature measurements, and alarm data, forming electronic records for traceability. The robot can operate continuously, unaffected by fatigue or line of sight; it can be on standby 24 / 7, and can quickly respond and confirm abnormal situations (such as smoke or fire), significantly shortening anomaly reporting time. This reduces equipment failure losses caused by inadequate inspections. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the power distribution room inspection method of the present invention. Detailed Implementation

[0066] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0067] Combination Figure 1 As shown, this invention provides a method for inspecting power distribution rooms based on a wheeled robot, comprising the following steps:

[0068] Step S1, Inspection Point Marking and Task Configuration, specifically includes:

[0069] All inspection points in the power distribution room are pre-calibrated, and the equipment type, normal status parameters, and thresholds are entered; the calibration time for a single inspection point is ≤30 seconds.

[0070] An improved A* algorithm is used to plan inspection paths; the planning time for inspection paths is ≤10 seconds, and the repetition rate of inspection paths is ≤1%.

[0071] The A* algorithm is as follows:

[0072] In standard A*, the evaluation function for each node is:

[0073]

[0074] in, From the starting point to the node The actual cost of movement (usually Euclidean distance or Manhattan distance); For the node Heuristic estimation of the target point (e.g., Euclidean distance);

[0075] Frequent turns during the inspection path reduce movement efficiency and increase energy consumption, so a turning penalty is introduced:

[0076]

[0077]

[0078] in, The turning penalty coefficient is set to 0.1~0.5. The steering radius coefficient is 0.2 for a 45° turn, 0.5 for a 90° turn, 0.8 for a 135° turn, and 1.0 for a 180° turn. This represents the step size from the parent node to the current node.

[0079] Configure inspection frequency and inspection items according to inspection needs, and support custom task types such as full inspection, special inspection, and special inspection; the inspection frequency is from 15 minutes to 30 days.

[0080] Step S2, multi-source data acquisition, specifically includes:

[0081] The wheeled robot moves along the inspection path at an adjustable speed of 0.3m / s to 1m / s. It uses laser SLAM and QR code for assisted positioning with an accuracy of ±1cm. Upon reaching each inspection point, the robot automatically adjusts its posture and sensor angles. The tilt angle of the robot's gimbal is adjustable from -90° to +90°, and the horizontal rotation range is 0° to 360°. The robot sequentially collects visible light images, infrared thermal images, and ultrasonic partial discharge signals from equipment (distribution cabinets, indicator lights, knobs, switches, busbars, cable connectors, etc.), covering all types of inspection needs for power distribution rooms with voltage levels of 10kV and below. It also synchronously records the coordinates of the inspection points and the acquisition time. The data synchronization error of each sensor on the wheeled robot is ≤50ms.

[0082] The wheeled robot is equipped with a 4K high-definition visible light camera, an infrared thermal imager with a temperature measurement range of -20℃ to 300℃ (temperature measurement accuracy of ±2℃ or ±2%, whichever is greater), and an ultrasonic partial discharge detector with a detection frequency of 20kHz to 100kHz.

[0083] Step S3, multi-dimensional state recognition, specifically includes:

[0084] The visible light image is input into the multi-task recognition model, and the target detection algorithm outputs the indicator light status (red / green / yellow color), switch position (open / closed), and knob position information. The recognition accuracy is ≥90%. The model is compared with the preset normal status to determine whether the equipment is operating normally; the confidence level of the status judgment is ≥0.98.

[0085] To achieve multi-task recognition of indicator light colors, switch open / closed positions, and knob positions in visible light images, the multi-task recognition model employs a deep learning-based multi-task object detection and classification framework; the specific process is as follows:

[0086] (1) Data collection and labeling

[0087] Acquire visible light images of the device under different lighting conditions and angles;

[0088] Label each target area (indicator light, switch, knob) with its bounding box and category:

[0089] Indicator lights: Red, Green, Yellow, Off (optional)

[0090] Switch: Opening, closing

[0091] Knob: Gear 1, Gear 2, ... (depending on the actual number of gears)

[0092] (2) Model design

[0093] Backbone network: Lightweight CNN is used to extract image features.

[0094] Region proposal: If a two-stage detector (Faster R-CNN) is used, candidate regions are first generated by the RPN; if a single-stage detector (YOLOv8) is used, bounding boxes and categories are predicted directly through a grid.

[0095] Multi-task branching: Three independent sub-networks are connected in parallel on a shared feature map or RoI feature:

[0096] Indicator light branch: Fully connected layer + Softmax (outputs red / green / yellow / other)

[0097] Switching branch: Fully connected layer + Sigmoid (Binary classification: Opening / Closing)

[0098] Knob branch: Fully connected layer + Softmax (outputs the probability of each gear position)

[0099] Loss function: joint optimization ,in, For bounding box regression loss and object confidence loss (such as CIoU); This is either cross-entropy loss or binary classification loss;

[0100] (3) Training and Reasoning

[0101] During training, transfer learning (pre-trained on ImageNet) was used, and data augmentation (rotation, brightness adjustment, affine transformation) was employed to improve generalization ability;

[0102] During inference, a single image is input, and the model outputs for each detected target: bounding box position, object category (indicator light / switch / knob), corresponding state, and confidence level (maximum probability output by Softmax or Sigmoid output value).

[0103] (4) Post-processing

[0104] Non-maximum suppression (NMS) removes overlapping bounding boxes;

[0105] For each target, the state with the highest confidence level is taken as the final output; if the confidence level is lower than the preset threshold (such as 0.9), it is marked as "uncertain" and manual review or resampling is triggered.

[0106] A database of associated power distribution equipment ledgers is built, linking each inspection point with equipment number, rated parameters, and historical operating data. It supports batch import of equipment ledger information with an import efficiency of ≥1000 entries / minute. The identification results are automatically compared with preset normal states to automatically determine whether there are any abnormalities, with an abnormality identification response time of ≤2 seconds.

[0107] Step S4, anomaly verification and classification, specifically includes:

[0108] After an anomaly is detected, the system automatically retrieves and compares the historical inspection data of the last three inspections for that inspection point. Once the anomaly is confirmed to be persistent, it is classified into three levels according to the type of anomaly: warning, general alarm, and emergency alarm. An abnormal indicator light is considered a warning; an abnormal switch position is considered a general alarm. Different alarm levels trigger different push rules: emergency alarm push response time ≤ 10 seconds, general alarm ≤ 1 minute, and warning ≤ 5 minutes.

[0109] If the anomaly involves changes in the status of switches or knobs, it must be compared with the SCADA system operation records simultaneously to exclude status changes caused by legitimate operations. For anomalies that cannot be confirmed, a manual review process will be automatically triggered.

[0110] It supports anti-misoperation verification, automatically comparing the detected switch / knob status changes with the SCADA system's operation records. The default query time window is 10 minutes; if no corresponding operation record is found, it is considered a misoperation, triggering an alarm immediately. The misoperation identification accuracy is ≥99%, effectively preventing safety risks from unauthorized operations. For temporary construction and equipment maintenance in power distribution rooms, it supports one-click activation of the inspection shielding mode, automatically ignoring status change alarms in preset maintenance areas to avoid false alarms. The shielding mode operation log is fully traceable, achieving 100% traceability.

[0111] Step S5, result output and alarm, specifically includes:

[0112] Generate structured inspection reports, including inspection data and anomalies for all equipment;

[0113] The alarm is pushed to the corresponding maintenance personnel according to the alarm level, and an emergency alarm will trigger an audio and visual prompt simultaneously.

[0114] Inspection data is automatically stored in the management platform, forming a health record for the entire life cycle of the equipment.

[0115] Output structured inspection reports and automatically generate inspection logs, including all equipment testing data, abnormal status, and handling suggestions. The reports can be directly exported and archived, with a generation time of ≤1 minute per report. It supports seamless integration with the power operation and maintenance management system to achieve automatic archiving of inspection data and meet the archiving requirements of power safety regulations.

[0116] In this invention, the wheeled robot supports automatic return-to-base charging. It automatically returns to the charging dock when the inspection task is completed or the battery level drops below 20%. A full charge takes ≤2 hours, and a fully charged robot can continuously inspect for over 8 hours, ensuring 24 / 7 uninterrupted inspection capability. It also supports multi-robot collaborative inspection; for large power distribution rooms, multiple robots can be configured to divide the inspection area and collaboratively complete the inspection task, increasing inspection efficiency by N times (N being the number of robots). For power distribution rooms with high humidity and high dust conditions, an IP65 protection rating is available, allowing for long-term stable operation in environments with relative humidity ≤95% and dust concentration ≤10mg / m³, with a mean time between failures (MTBF) ≥5000 hours.

[0117] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A method for inspecting power distribution rooms based on a wheeled robot, characterized in that, Includes the following steps: Step S1: Inspection point location marking and task configuration; Step S2: Multi-source data acquisition; Step S3, multi-dimensional state recognition; Step S4, anomaly verification and classification; Step S5: Output the results and issue alarms.

2. The method for inspecting a power distribution room based on a wheeled robot according to claim 1, characterized in that, Step S1 specifically includes: All the aforementioned inspection points within the power distribution room are pre-marked; An improved A* algorithm is used to plan the inspection path; Configure the inspection frequency and inspection items according to the inspection requirements.

3. The method for inspecting a power distribution room based on a wheeled robot according to claim 2, characterized in that, The A* algorithm is described in detail below: In standard A*, the evaluation function for each node is: ； in, From the starting point to the node The actual cost of movement; For the node Heuristic estimation of the target point; Frequent turns during the inspection path reduce movement efficiency and increase energy consumption, so a turning penalty is introduced: ； ； in, The turning penalty coefficient is set to 0.1~0.

5. The steering radius coefficient is 0.2 for a 45° turn, 0.5 for a 90° turn, 0.8 for a 135° turn, and 1.0 for a 180° turn. This represents the step size from the parent node to the current node.

4. The power distribution room inspection method based on a wheeled robot according to claim 2, characterized in that: The calibration of all the aforementioned inspection points includes recording the device type, normal status parameters, and thresholds; The calibration time for a single inspection point is ≤30 seconds; The time required to plan the inspection path is ≤10 seconds, and the repetition rate of the inspection path is ≤1%. The inspection frequency is from 15 minutes to 30 days.

5. The method for inspecting a power distribution room based on a wheeled robot according to claim 2, characterized in that, Step S2 specifically includes: The wheeled robot moves along the inspection path and automatically adjusts its body posture and sensor angle upon reaching each inspection point. It then sequentially collects visible light images, infrared thermal images, and ultrasonic partial discharge signals from the equipment, and simultaneously records the coordinates and collection time of each inspection point.

6. The power distribution room inspection method based on a wheeled robot according to claim 5, characterized in that: The wheeled robot uses laser SLAM + QR code for assisted positioning, with a positioning accuracy of ±1cm. The pitch angle adjustment range of the wheeled robot's gimbal is -90° to +90°, and the horizontal rotation range is 0° to 360°. The data time synchronization error of each sensor on the wheeled robot is ≤50ms.

7. The method for inspecting a power distribution room based on a wheeled robot according to claim 5, characterized in that, Step S3 specifically includes: The visible light image is input into the multi-task recognition model, which outputs the indicator light status, switch position, and knob gear information, and compares them with the preset normal state.

8. The method for inspecting a power distribution room based on a wheeled robot according to claim 7, characterized in that, The multi-task recognition model is designed to recognize indicator light colors, switch open / closed positions, and knob positions in visible light images. It employs a deep learning-based multi-task object detection and classification framework; the specific process is as follows: (1) Data collection and labeling Acquire visible light images of the device under different lighting conditions and angles; Label the bounding box and category for each target region; (2) Model design Backbone network: Lightweight CNN is used to extract image features; Region proposal: If a two-stage detector is used, candidate regions are first generated by the RPN; if a single-stage detector is used, bounding boxes and categories are predicted directly through a grid. Multi-task branching: Three independent sub-networks are connected in parallel on a shared feature map or RoI feature: Indicator light branch: Fully connected layer + Softmax Switching Branch: Fully Connected Layer + Sigmoid Knob Branch: Fully Connected Layer + Softmax Loss function: joint optimization ,in, The bounding box regression loss and object confidence loss are used. This is either cross-entropy loss or binary classification loss; (3) Training and Reasoning Transfer learning is used during training, and data augmentation is employed to improve generalization ability; During inference, a single image is input, and the model outputs for each detected target: bounding box location, object category, corresponding state, and confidence level. (4) Post-processing Non-maximum suppression removes overlapping bounding boxes; For each target, the state with the highest confidence level is taken as the final output; if the confidence level is lower than the preset threshold, it is marked as "uncertain" and manual review or re-sampling is triggered.

9. The method for inspecting a power distribution room based on a wheeled robot according to claim 7, characterized in that, Step S4 specifically includes: After an anomaly is detected, the system automatically retrieves and compares the historical inspection data of the last three inspections for that inspection point. Once the anomaly is confirmed to be persistent, it is classified according to the type of anomaly.

10. The method for inspecting a power distribution room based on a wheeled robot according to claim 9, characterized in that: If the anomaly type involves a change in the state of a switch or knob, it must be compared with the SCADA system operation record simultaneously to exclude state changes caused by legitimate operations. For anomalies that cannot be confirmed, a manual review process will be automatically triggered.

11. The method for inspecting a power distribution room based on a wheeled robot according to claim 9, characterized in that, Step S5 specifically includes: Generate structured inspection reports, including inspection data and anomalies for all equipment; The alarm is pushed to the corresponding maintenance personnel according to the alarm level, and an emergency alarm will trigger an audio and visual prompt simultaneously. Inspection data is automatically stored in the management platform, forming a health record for the entire life cycle of the equipment.

Images