Electric energy meter appearance defect detection method and system based on AI automatic identification

Abstract

The invention relates to the technical field of electric energy meter appearance detection, and discloses an AI automatic identification-based electric energy meter appearance defect detection method and system. The method comprises the steps that electric energy meters are grouped to form detection units, and a detection device is started after an initial detection instruction is received; when initial detection is executed, basic appearance information of the electric energy meter of each detection unit is collected, and corresponding historical defect records are called; the defect probability grade and the criticality grade of each detection unit are evaluated by combining the two types of information, and a personalized detection sequence of the detection device is planned accordingly. The detection device captures real-time image information of the electric energy meter when executing tasks according to the sequence, obtains field interference factor data to judge whether the electric energy meter is interfered or not, and corrects the electric energy meter when the electric energy meter is interfered and a detection track deviates; the real-time image information is analyzed through an artificial intelligence algorithm to find appearance flaws, an alarm is generated after the appearance flaws are found, and flaw sites and related image information are transmitted to a user display end.

Description

Technical Field

[0001] This invention relates to the field of electricity meter appearance inspection technology, specifically to an AI-based method and system for detecting electricity meter appearance defects. Background Technology

[0002] As a core device for measuring electricity consumption, the appearance quality of electricity meters directly affects the stability of equipment installation, safety of use, and long-term operational reliability. During the production, transportation, and maintenance of electricity meters, defects such as scratches, deformation, damage, and blurred markings are prone to occur on the exterior. If such defects are not detected and addressed in a timely manner, it may lead to difficulties in subsequent installation, or even expose the internal circuitry due to damage to the casing, increasing the safety risks of electric shock and short circuits. Therefore, the detection of appearance defects is a crucial aspect of electricity meter quality control.

[0003] Currently, the detection of visual defects in electricity meters mainly relies on two methods: manual inspection and traditional machine inspection. Manual inspection requires inspectors to visually inspect each meter individually, identifying defects by sight. However, this method is heavily influenced by the inspector's subjective experience, visual fatigue, and concentration. In batch inspection scenarios, it is not only inefficient and unsuitable for the high-speed flow of production lines, but also prone to missed or false detections due to human error, especially for minor scratches, slight deformations, and other inconspicuous defects, where accuracy is difficult to guarantee. Furthermore, prolonged and repetitive visual inspections significantly increase the workload of inspectors, leading to high inspection costs. Moreover, the lack of standardized manual inspection criteria, with different inspectors using varying judgment standards, can result in inconsistent inspection results for the same batch of meters, affecting the stability of quality control.

[0004] Traditional machine inspection often employs fixed-path image acquisition and comparison methods. Cameras and other equipment are driven by a preset inspection trajectory to capture images of the electricity meter's exterior. These images are then compared to a standard template to identify defects. However, this method has significant limitations. The fixed inspection sequence cannot adjust the inspection focus based on the probability of defects in different parts of the meter. For example, the edges of some meter casings are more prone to damage due to manufacturing processes, yet traditional machines still inspect them in a uniform order, resulting in wasted inspection resources on low-risk areas while high-risk areas are not adequately checked. Traditional machine inspection also lacks utilization of historical defect data, failing to optimize inspection strategies by combining past defect records for specific types of meters or inspection units, making targeted inspection difficult. Furthermore, the inspection site is susceptible to interference from factors such as lighting changes, vibration, and dust. Traditional machine inspection typically lacks effective interference detection and trajectory correction mechanisms. When interference causes the shooting angle to shift or the image to blur, it directly affects the accuracy of the image comparison results, leading to misjudgments or missed detections, failing to meet the high-precision and highly adaptable requirements for electricity meter appearance defect detection. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting appearance defects in electricity meters based on AI automatic identification, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for detecting appearance defects in electricity meters based on AI automatic identification, the method comprising:

[0007] The electricity meters are grouped into multiple detection units; the initial detection command issued by the user is received and the detection device is started; during the initial detection process, the basic appearance information of the electricity meters of each detection unit is collected, and the historical defect records of each detection unit are retrieved from the storage unit.

[0008] Assess the defect probability level and criticality level of each inspection unit based on basic appearance information and historical defect records; plan the personalized inspection sequence of each inspection unit based on the criticality level and defect probability level.

[0009] When the detection device performs the detection task according to the personalized detection sequence, it captures real-time image information of the electricity meter; it acquires data on interference factors in the detection site and determines whether the detection device is under interference; if the detection device is under interference, it determines whether the detection trajectory has deviated based on the real-time acquired position information and corrects the detection trajectory when it deviates.

[0010] Artificial intelligence algorithms are used to analyze real-time image information to detect defects in the appearance of the electricity meter. When a defect is detected, a defect alarm is generated and the location of the defect and related image information are transmitted to the user's display terminal.

[0011] Preferably, the basic appearance information includes the surface material type and installation orientation information of the electricity meter, and the historical defect record includes the frequency of historical defects and the repair time for each defect.

[0012] Preferably, the process of assessing the probability level and criticality level of defects specifically includes: collecting the historical defect occurrence frequency of each detection unit and obtaining the repair time for each defect in each detection unit; weighting and summing the repair times of each defect in chronological order to obtain the weighted repair time; calculating the defect index of each detection unit based on the historical defect occurrence frequency and the weighted repair time; comparing the defect index with a preset defect index threshold to determine whether the defect probability level of the detection unit is high probability, medium probability, or low probability; simultaneously acquiring the surface material type and installation orientation information of each detection unit, and calculating the criticality index of each detection unit based on material durability and installation complexity; comparing the criticality index with a preset criticality index threshold to determine whether the criticality level of the detection unit is high criticality, medium criticality, or low criticality.

[0013] Preferably, the process of planning a personalized detection sequence is as follows: the detection device uses the criticality level of the detection unit as the primary selection criterion, the defect probability level of the detection unit as the secondary selection criterion, and the path length between the detection device and the detection unit as a reference criterion; starting from the starting point, if there are multiple detection units with the same criticality level nearby, the detection unit with the higher defect probability level is selected for detection; if the criticality level and defect probability level of the detection unit are the same, the detection unit with the shortest path length is selected for detection.

[0014] When the detection device arrives at a detection point, if the point corresponds to only one detection unit, the detection unit is directly tested; if the point corresponds to multiple detection units, the detection unit with the higher criticality level is tested first; if multiple detection units at the point have the same criticality level, the detection unit with the higher defect probability level is tested first; if both the criticality level and the defect probability level are the same, the detection units are selected for testing according to the order of their numbers; if the point is not connected to other detection units and there are untested detection units, the untested detection unit with the shortest path length is tested first.

[0015] When all detection units have been detected, the detection device returns to the starting point along the optimal path; a personalized detection sequence for the detection device is generated according to the above rules; while the detection device moves in accordance with the personalized detection sequence, the detection dwell time is dynamically adjusted according to the criticality level and defect probability level of the detection unit.

[0016] Preferably, the interference factor data includes the spatial coordinates of the interference source, the baseline value of the interference intensity, and the interference diffusion coefficient;

[0017] The spatial coordinates of the interference source are the location parameters of the interference source, the basic value of the interference intensity is the initial interference capability quantification of the interference source, and the interference diffusion coefficient is the interference propagation attenuation characteristic parameter.

[0018] Preferably, the process of determining whether the detection device is in an interference state specifically involves: measuring the Euclidean distance between the real-time coordinates of the detection device and the spatial coordinates of the interference source; obtaining the basic value of the interference intensity and the interference diffusion coefficient of the interference source, and obtaining the effective interference radius of the interference source by multiplying the basic value of the interference intensity by the interference diffusion coefficient; constructing a spherical interference zone of the interference source with the spatial coordinates of the interference source as the center and the effective interference radius as the boundary; when the Euclidean distance is greater than the effective interference radius, the detection device performs normal detection outside the spherical interference zone; when the Euclidean distance is less than or equal to the effective interference radius, the detection device enters the spherical interference zone and generates an interference state signal.

[0019] Preferably, the process of determining whether the detection trajectory has deviated is as follows: define the ideal detection distance range and allowable attitude deviation range of the detection device relative to the energy meter; if the actual distance between the detection device and the energy meter is within the ideal detection distance range and the attitude angle of the acquisition device is within the allowable attitude deviation range, then the detection trajectory meets the requirements; if the actual distance exceeds the ideal detection distance range or the attitude angle of the acquisition device exceeds the allowable attitude deviation range, then it is determined that the detection trajectory has deviated, and a trajectory correction command is generated.

[0020] Preferably, the process of using artificial intelligence algorithms to analyze real-time image information is as follows:

[0021] Visible light images of the electricity meter are acquired; the visible light images are normalized, and noise suppression algorithms are applied to reduce image noise. Texture feature thresholds are set according to the standard appearance model of the electricity meter to segment the electricity meter area; the image is converted to a brightness-saturation space, and a region growing algorithm is used to extract connected regions that conform to the texture features of the electricity meter, preserving the main image of the electricity meter; a contour detection algorithm is used to identify the image contours, and the contour shapes are matched with the reference template of the electricity meter for similarity, eliminating inconsistent parts;

[0022] Clustering analysis algorithms are used to group pixels and filter out regions of interest (ROIs) of the electricity meter. Multi-scale analysis is performed on the ROIs to assess their appearance integrity. Near-infrared images of the electricity meter are acquired simultaneously, and a reflectivity threshold is set based on material reflectivity characteristics to extract regions that meet the threshold. Multi-band fusion technology is used to enhance image details, and contour detection is combined to preserve the meter's edge features. The reflectivity distribution in the near-infrared images is analyzed to identify areas of abnormal reflection. If anomalies are detected in both visible and near-infrared images, a defect confirmation signal is generated. If no anomalies are found in either image, the current state is maintained. If one image shows an anomaly, a re-inspection signal is generated.

[0023] Preferably, the method further includes: after generating a defect alarm, extracting the feature attributes of the defect, including shape category and area ratio, and binding and archiving the feature attributes with image data; the user terminal is configured with a query module for retrieving and displaying defect records according to conditions.

[0024] Preferably, the present invention also includes an AI-based automatic identification system for detecting appearance defects in electricity meters. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the AI-based automatic identification method for detecting appearance defects in electricity meters as described above.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] First, the electricity meters are grouped into testing units. At the initial stage of testing, the basic appearance information and historical defect records of each unit are used to assess the probability and criticality levels of defects, making the testing strategy more targeted. Historical defect records clearly show the past defect occurrences of different testing units. For example, if a batch of electricity meters corresponding to a certain testing unit has repeatedly experienced casing damage, its defect probability level will be correspondingly higher. For parts directly related to subsequent safety, such as installation interfaces, their criticality level can be set to a higher level. Based on this level, a personalized testing sequence is planned, allowing the testing device to prioritize testing units with high defect probability and high criticality, avoiding the ineffective consumption of testing resources on low-risk, low-importance parts. This makes the testing process more closely match the actual defect distribution characteristics, improving the utilization efficiency of testing resources, while ensuring that high-risk, high-importance parts are fully inspected, reducing the omission of critical defects due to unreasonable testing sequences.

[0027] During the inspection process, this method introduces an interference factor judgment and inspection trajectory correction mechanism, which can effectively cope with the complex environment of the inspection site. Interference factors such as changes in lighting, equipment vibration, and dust obstruction at the inspection site can easily lead to blurred images and shooting angle deviations in the images captured by the inspection device, affecting the accuracy of subsequent defect identification. This method can promptly determine whether the inspection device is under interference by acquiring interference factor data. If interference is confirmed, it combines the real-time acquired position information to determine whether the inspection trajectory deviates from the preset path. Once a deviation is detected, it is corrected immediately to ensure that the inspection device always captures real-time image information of the energy meter according to the planned sequence and angle, ensuring the clarity and accuracy of image acquisition, providing a high-quality image data foundation for subsequent AI algorithm analysis, reducing misjudgments and omissions caused by environmental interference, and improving the stability of inspection results.

[0028] Artificial intelligence algorithms are used to analyze real-time image information, providing a stronger defect recognition capability compared to traditional manual visual judgment and fixed template comparison. The AI ​​algorithm can be trained on a large amount of defect image data to learn the characteristics of different types of appearance flaws (such as scratches, deformations, and blurred markings), enabling it to quickly identify subtle and inconspicuous defects in images. Furthermore, the recognition process is unaffected by subjective factors, with consistent judgment standards, effectively avoiding the accuracy decline caused by subjective bias and fatigue in manual inspection. At the same time, the AI ​​algorithm's analysis speed is much faster than manual methods, adapting to the high-speed inspection needs of batch energy meters and further improving overall inspection efficiency.

[0029] Upon discovering cosmetic defects, this method can promptly generate a defect alarm and transmit the defect location and related image information to the user's display terminal. This allows inspection personnel to obtain defect information in real time, clearly identifying the defect location and specific details without manual inspection. This facilitates rapid follow-up actions, such as marking, isolating, or repairing defective electricity meters, shortening the defect handling cycle, and reducing additional costs incurred by defective products flowing to downstream stages. Simultaneously, users can intuitively view defect images on the display terminal, providing a direct basis for defect cause analysis. This helps optimize quality control measures in production and transportation, promoting the improvement of the overall quality control system for electricity meters. Attached Figure Description

[0030] Figure 1 A diagram showing the distribution of electricity meter detection units and the optimal detection path planning.

[0031] Figure 2 A flowchart for determining the interference state of the detection device;

[0032] Figure 3 A flowchart for detecting and correcting trajectory deviations;

[0033] Figure 4 Radar chart for evaluating the image processing quality of electricity meters. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0035] Please see Figure 1This invention provides an AI-based method for detecting appearance defects in electricity meters. The method includes: grouping electricity meters into multiple detection units; receiving an initial detection command from a user and activating a detection device; collecting basic appearance information of each detection unit during the initial detection process; and retrieving historical defect records of each detection unit from a storage unit; evaluating the defect probability level and criticality level of each detection unit based on the basic appearance information and historical defect records; and planning a personalized detection sequence for each detection unit based on the criticality level and defect probability level; capturing real-time image information of the electricity meter while the detection device executes the detection task according to the personalized detection sequence; acquiring data on interference factors in the detection site and determining whether the detection device is under interference; if the detection device is under interference, determining whether the detection trajectory has deviated based on the real-time acquired location information and correcting the detection trajectory if deviation occurs; and using an artificial intelligence algorithm to analyze the real-time image information to detect appearance defects in the electricity meter, generating a defect alarm upon detection, and transmitting the defect location and related image information to the user's display.

[0036] Example 1: The process of assessing defect probability and criticality levels involves collecting historical defect occurrence frequencies for each detection unit and obtaining the repair time for each defect. The repair times for each defect are then weighted and summed in chronological order to obtain a weighted repair time. Based on the historical defect occurrence frequency and weighted repair time, a defect index for each detection unit is calculated. This defect index is compared with a preset defect index threshold to determine the defect probability level of the detection unit as high, medium, or low. Simultaneously, the surface material type and installation orientation information of each detection unit are acquired. A criticality index for each detection unit is calculated based on material durability and installation complexity. This criticality index is compared with a preset criticality index threshold to determine the criticality level of the detection unit as high, medium, or low. Material durability is quantified based on the physical properties of the surface material type, which includes metal casings, engineering plastic casings, and composite material panels. Metal casings have higher hardness and corrosion resistance coefficients than engineering plastic casings, while engineering plastic casings have better weather resistance than composite material panels. Installation orientation information is derived from the spatial coordinates and orientation angle of the electricity meter within the power distribution facility. Installation complexity calculations depend on the offset of the orientation angle, the degree of obstruction by adjacent equipment, and the visibility of the terminal block orientation. Historical defect frequency statistics are compiled monthly, recording the number of defect reports for each detection unit within its operating cycle. The repair time for each defect is calculated from the time of the reported repair to the time of maintenance acceptance. Weighted repair time is calculated using a time decay function, assigning higher weight to recent defect repair times and decreasing weight to older defect repair times. A weighted summation operation produces a time-weighted value reflecting the defect maintenance cost.

[0037] The defect index calculation integrates historical defect occurrence frequency and weighted repair time. Historical defect occurrence frequency is normalized to eliminate dimensional differences, and weighted repair time is multiplied by a cost coefficient to convert it to frequency dimensions. Preset defect index thresholds are set according to power grid operation specifications. High-probability thresholds correspond to energy meters at important transmission nodes, while low-probability thresholds are used for ordinary user-end energy meters. Key index calculation requires establishing a material durability scoring table. Metal casings receive the highest durability score, engineering plastic casings receive a medium durability score, and composite material panels receive a basic durability score. Installation complexity scoring is based on multi-dimensional parameters. Vertical installation has lower complexity than inclined installation, and installations without obstructions score higher than those with pipe obstructions. Preset key index thresholds refer to the functional level of the energy meter in the power grid; high-criticality thresholds are used for metering hub nodes, and medium-criticality thresholds are used for branch line nodes.

[0038] The evaluation process is completed during the initialization phase of the testing device. The testing device downloads the surface material type and installation orientation information of the testing unit from the central database and simultaneously accesses the operation and maintenance database to obtain historical defect records. A three-level classification algorithm is used to determine the defect probability level: a defect index exceeding the upper threshold triggers a high-probability level marker; a defect index in the middle range activates a medium-probability level marker; and a defect index below the lower threshold maintains a low-probability level. Criticality level determination executes a parallel computation process; the comparison result between the critical index and the preset critical index threshold is directly mapped to three categories: high criticality, medium criticality, and low criticality. Material durability scoring incorporates an environmental correction factor: the material score of outdoor-installed energy meters is multiplied by an environmental severity coefficient, while indoor-installed energy meters use a standard scoring method. Installation complexity calculation integrates spatial geometric parameters, using 3D modeling technology to analyze the relative position of the energy meter and surrounding equipment, and outputs the installation operation difficulty value. Historical defect record preprocessing includes data cleaning steps, removing duplicate repair reports and invalid maintenance logs, and filtering out abnormal extreme values ​​for time-consuming data. The decay function for weighted repair time adopts an exponential model. The weight coefficient for defects in the most recent month is set to 1.0, and the weight is multiplied by a decay factor of 0.8 for each additional month. The time decay function ensures that recent data dominates the weighted result. The defect index calculation formula is a linear combination normalized to frequency and weighted time, with the frequency weight coefficient set to 0.6 and the time weight coefficient set to 0.4. The weight coefficient configuration supports dynamic adjustment. The key index calculation adds the material durability score and the installation complexity score, with the material score accounting for 60% and the installation score accounting for 40%. The weighted sum is normalized to a standard range. The preset defect index threshold is dynamically updated according to seasonal changes. The threshold is increased in summer to cope with high temperature and humidity environments, and decreased in winter to adapt to low temperature conditions.

[0039] The grouping logic of the detection units affects the evaluation accuracy. Each detection unit contains a group of physically adjacent electricity meters, and the meters within the group share the same surface material type and installation orientation information. Surface material type identification uses spectral analysis technology. The detection device is equipped with a near-infrared spectral sensor to scan the electricity meter casing, and the spectral characteristics are matched with the material database to determine the material type. Installation orientation information acquisition relies on a combination of lidar and inertial measurement unit navigation to generate the precise coordinates and attitude angles of the electricity meter in three-dimensional space. Historical defect record queries use a structured query language, grouping and aggregating the number of defect occurrences by detection unit number, and calculating the arithmetic mean of repair time data to eliminate random errors. The defect probability level is output as a numerical code: high probability level is coded as 3, medium probability level as 2, and low probability level as 1. The code format facilitates fast decoding by the detection device's processor. Criticality level assessment incorporates grid topology relationships. Electricity meters connected to the main grid are automatically upgraded in criticality level, while electricity meters operating in islanded mode are appropriately downgraded in criticality level. Material durability scoring supplements the surface coating evaluation item. Metal casings with anti-corrosion coatings receive durability bonuses, while engineering plastic casings without protective coatings have some points deducted. The installation complexity calculation incorporates maintainability as an indicator, and the operating space of maintenance tools is included in the scoring system. Meters installed in confined spaces receive higher complexity scores. Preset thresholds for key indices are segmented, with higher thresholds for high-voltage meters than low-voltage meters, and higher thresholds for three-phase meters than for single-phase meters. The evaluation algorithm employs a fault-tolerant mechanism: when historical data is missing, average data from similar meters is used as a substitute; when material information is incomplete, a default durability score is used.

[0040] The results of defect probability level and criticality level are stored in the detection device's cache. Each detection unit's level attribute is associated with a unique identifier, and the level data update cycle is synchronized with the detection task. Intermediate variables generated during the evaluation process include normalized frequency value, weighted time value, material score, and installation score. These intermediate variables are temporarily stored in memory for audit tracking. Floating-point comparison instructions are used for defect index threshold comparison, while integer comparison logic is used for critical index threshold comparison. The comparison result triggers a level state machine switch. The detection device adjusts the detection frequency based on the defect probability level, allocating more detection frequency to high-probability level detection units and reducing detection resource investment for low-probability level units. The criticality level guides the detection priority; high-criticality level detection units are prioritized for detection resource scheduling, while low-criticality level detection units have their detection sequence postponed. The evaluation module communicates in real-time with the detection device's control system. Changes in level trigger dynamic adjustments to the detection strategy, and an increase in defect probability level immediately initiates a re-inspection procedure. The surface material category database is updated online; new materials automatically generate durability scores upon entry into the database, and changes in installation orientation information trigger a recalculation of complexity. The historical defect record import process implements integrity verification. If records are missing, a data repair process is initiated, and outliers in repair time are replaced with the median. Weighted repair time calculation supports multiple time granularities, including daily, weekly, and monthly weighted modes, with the weighting mode selection based on the detection cycle length. The defect index calculation introduces the concept of confidence intervals; detection units with insufficient data use conservative estimates, expanding the judgment boundary for defect probability levels.

[0041] The criticality level assessment integrates manually labeled data, allowing maintenance personnel to correct the automatic scoring results. Correction records are fed back to the assessment algorithm optimization model. Material durability scoring is linked to environmental sensor data; real-time temperature and humidity data participate in the fine-tuning of durability scores, with engineering plastic casings in high-humidity environments having their durability scores lowered. Installation complexity calculation incorporates dynamic obstacle detection; when temporary storage changes the installation environment, the complexity score is recalculated, and changes in complexity scores trigger a reassessment of the criticality level. Preset defect index thresholds and preset criticality index thresholds are stored in configuration files. Threshold parameters support remote configuration updates, and threshold adjustments take effect immediately across all detection units. The assessment process log records the input data and output level of each detection unit; the log file is used for performance analysis and optimization verification of the assessment algorithm. The joint output of defect probability level and criticality level forms the basis of the detection strategy; high-probability, high-criticality level detection units activate an enhanced detection mode, while low-probability, low-criticality level detection units activate a fast detection mode. The assessment algorithm adopts a modular design, with the defect probability assessment module and the criticality assessment module running independently, and parallel computation improving processing efficiency. The accuracy of surface material category identification affects the accuracy of key indices. Spectral analysis algorithms are periodically calibrated against standard samples, and material database updates are synchronized to all testing devices. The frequency of installation orientation information acquisition is matched to the movement speed of the testing device; during high-speed movement, interpolation algorithms are used to compensate for positioning delays, ensuring real-time and accurate installation complexity scoring. Historical defect record queries optimize the index structure, partitioning storage by time range to improve query speed, and compressing and storing time-consuming repair data reduces network transmission load.

[0042] The attenuation factor for weighted repair time is configurable, and adjusting the attenuation factor changes the time sensitivity of the defect probability level, increasing or decreasing the dominance of recent defects. The weighting coefficients in the defect index calculation formula can be adjusted online; changes in frequency and time weights affect the defect probability level distribution. The ratio of material score to installation score in the key index calculation is adjustable, adapting to the specific requirements of different power grid architectures. Preset defect index thresholds are set in segmented intervals, with buffer zones at interval boundaries to prevent frequent level oscillations; the buffer zone width is set based on historical data volatility. Preset key index thresholds differentiate between installation locations; indoor and outdoor energy meter thresholds are set independently, with outdoor environmental thresholds including weather compensation factors. Data validity is verified before the evaluation process: surface material category data verifies spectral signal intensity, and installation orientation information verifies sensor reading consistency. Historical defect records verify timestamp logic, and repair time verifies the reasonableness of start and end times; invalid records are marked as abnormal and excluded from calculation. The defect probability level determination results undergo consistency checks; if the difference from the previous month's results is too large, a manual review process is triggered; and if the key level change exceeds the threshold, a system alarm is generated. The evaluation module has a self-learning function, comparing historical evaluation results with actual defect data to optimize algorithm parameters, and continuously iterating and improving the evaluation model. When the detection device is working offline, it uses locally cached data for evaluation, and re-evaluates with the latest parameters after the network is restored, ensuring the timeliness of the grade evaluation results.

[0043] Example 2: The process of planning a personalized inspection sequence prioritizes the criticality level of the inspection unit, followed by its defect probability level, and then considers the path length between the inspection device and the inspection unit. Starting from the starting point, if multiple inspection units with the same criticality level are nearby, the inspection unit with the higher defect probability level is selected first. If both the criticality level and defect probability level are the same, the inspection unit with the shortest path length is selected. When the inspection device arrives at an inspection point, if the point corresponds to only one inspection unit, inspection is performed directly. If the point corresponds to multiple inspection units, the inspection unit with the higher criticality level is prioritized. If multiple inspection units at the point have the same criticality level, the inspection unit with the higher defect probability level is prioritized. If both the criticality level and defect probability level are the same, inspection units are selected according to their numerical order. If the point is not connected to other inspection units and there are uninspected inspection units, the uninspected inspection unit with the shortest path length is prioritized. After all inspection units have been inspected, the inspection device returns to the starting point along the optimal path. The detection device dynamically adjusts the detection dwell time based on the criticality level and defect probability level of the detection unit as it moves along the personalized detection sequence.

[0044] Path length calculation is based on the positioning system configured in the detection device and high-precision environmental map data. The positioning system employs a fusion positioning technology combining the Global Navigation Satellite System and ultra-wideband base stations. The environmental map includes the two-dimensional plane coordinates and elevation information of the electricity meter installation area. The straight-line distance between the detection device and the detection unit is calculated using the Euclidean distance formula. The actual movement path is planned using the A* algorithm on the environmental map to avoid obstacles. The path length value is updated in real time and stored in the detection device's memory buffer. The detection device's control unit adjusts its movement speed and turning angle based on the path length value. Criticality level priority is mapped to a numerical weight coefficient: high criticality level is assigned a value of 3, medium criticality level is assigned a value of 2, and low criticality level is assigned a value of 1. The defect probability level weight coefficient is set synchronously: high defect probability level is assigned a value of 3, medium defect probability level is assigned a value of 2, and low defect probability level is assigned a value of 1. The path length weight is converted to its reciprocal form and participates in the comprehensive scoring calculation. The comprehensive scoring formula integrates the criticality level weight, defect probability level weight, and path length reciprocal weight. The spatial coordinates of the detection points are pre-entered into the detection device database. Each detection point is associated with a set of electricity meter detection units, and the correspondence between the points and detection units is established through RFID tags or QR codes. The detection device navigation system confirms the precise location of the points through lidar point cloud matching, and triggers a detection ready signal when the arrival error is less than 5 cm. In multi-point detection scenarios, the detection device compares the comprehensive scores of each point. The comprehensive score calculation incorporates real-time environmental parameters. When the light intensity is below the threshold, the priority of indoor points is increased, and when the rain sensor is activated, points with canopy coverage are selected first. The detection unit numbering order is arranged according to the power distribution line topology. The main line detection unit numbering is prefixed with M, and the branch line detection unit numbering is prefixed with B. The numbering order determines the detection order under the same level of conditions.

[0045] The dynamic adjustment of detection dwell time relies on a multi-sensor feedback mechanism. The initial dwell time baseline of the detection device at a designated point is set to 30 seconds. For high-criticality detection units, the system automatically increases the detection time by 50%, and for high-defect-probability detection units, it increases by 30%. When criticality and defect probability levels are superimposed, a time-addition coefficient multiplication rule is used. The detection device's image acquisition system performs multi-angle scanning during the extended period. A visible light camera acquires three sets of images from the front, left side, and right side, while a near-infrared sensor performs material transmission detection. Image clarity indicators are analyzed in real time during detection. If image blur exceeds a threshold, the re-shooting time is automatically extended. If surface reflectivity is too high, a polarization filter adjustment is triggered, and re-detection is initiated. The personalized detection sequence generation algorithm combines a greedy strategy with a backtracking mechanism. The greedy strategy selects the current optimal detection point to advance the detection process, while the backtracking mechanism re-evaluates the comprehensive score of the remaining points after processing high-priority points. The objective function for optimizing the detection sequence is set to minimize the total detection time, with constraints including the detection device's battery life and the task deadline. The sequence generation module outputs an ordered list of locations, an estimated arrival timestamp, and a three-dimensional matrix of planned detection durations. This matrix data is synchronized to the monitoring center via a wireless network. The detection device's movement trajectory is recorded using a combination of odometry and inertial measurement units for positioning. The trajectory data is used to verify path planning efficiency and generate a detection coverage heatmap.

[0046] The optimal path calculation to return to the starting point is transformed into a modified Traveling Salesman Problem model. After completing the task at the last detection point, the detection device calculates the shortest regression path using Dijkstra's algorithm based on the current coordinates and the starting point coordinates. Path planning considers ground flatness parameters; for rugged terrain, detour routes are automatically selected; and differential coefficients are set for the weights of elevator shafts and stairwells. The detection device's battery monitoring module participates in path decision-making; when the battery level is below 20%, an emergency return procedure is initiated, interrupting unfinished detection tasks and directly planning the shortest return route. The dynamic adjustment mechanism for detection dwell time integrates anomaly handling logic; when the detection device's pressure sensor detects obstruction of the robotic arm, the operation time is extended; when the temperature sensor records an ambient temperature exceeding 40 degrees Celsius, the detection time at outdoor points is shortened. The maintenance status of the detection device affects time allocation; when the camera lens contamination index is high, the number of image acquisition repetitions is automatically increased; when the gimbal servo has a large wear coefficient, the rotation speed is reduced and the positioning time is extended accordingly. Updates to critical level data trigger online reconstruction of the detection sequence; when the maintenance system pushes an emergency detection task, it is immediately inserted at the beginning of the sequence; and a preemptive scheduling algorithm is used for temporary task allocation.

[0047] The complexity of the electricity meter installation environment is taken into account in sequence generation. Detection sequences for substation locations are prioritized over those for outdoor substation locations, and the detection time for basement locations includes waiting time for the ventilation system. Different movement modes of the detection device correspond to differentiated time parameters. In wheeled movement mode, the time consumption per unit distance is calculated based on flat ground data, while in tracked movement mode, a terrain difficulty coefficient is introduced to adjust the time consumption model. Collaborative detection device scenarios support multi-machine sequence coordination. After the master detection device generates the global detection sequence, subordinate detection devices synchronize sub-sequence tasks via a local area network to avoid overlapping detection ranges and path intersections. The detection data upload strategy affects the sequence time distribution. Locations with wireless network signal strength greater than -70dBm upload data immediately, while locations with weak signals temporarily store data and delay transmission until areas with good signal strength. Detection device storage space monitoring participates in sequence decision-making. When the remaining storage capacity is less than 10%, priority is given to detecting locations closer to the data transfer station. After data transfer is completed, remote location detection tasks continue. The detection sequence generation algorithm has undergone worst-case testing, and when handling large-scale scenarios with hundreds of detection points, the computation time is controlled within 5% of the total task time. Seasonal factors are considered when adjusting detection sequence parameters. During the rainy season, the detection frequency at points inside power distribution rooms is increased, while in winter, the detection time for outdoor points is advanced to avoid working in dark environments. After updating the detection device software, time parameters are recalibrated, and the new image processing algorithm shortens analysis time, allowing the saved time to be redistributed to more challenging detection points. Evaluation metrics for personalized detection sequences include total path length, total detection time, and priority task completion rate. Evaluation results are fed back to the sequence generation algorithm for continuous optimization.

[0048] Example 3: See Figure 2 The interference factor data includes the spatial coordinates of the interference source, the baseline value of the interference intensity, and the interference diffusion coefficient. The spatial coordinates of the interference source, the baseline value of the interference intensity, and the interference diffusion coefficient refer to the location parameters of the interference source, the quantized value of its initial interference capability, and the interference propagation attenuation characteristic parameters, respectively. The process of determining whether the detection device is under interference involves measuring the Euclidean distance between the real-time coordinates of the detection device and the spatial coordinates of the interference source to obtain the baseline value of the interference intensity and the interference diffusion coefficient. The effective interference radius of the interference source is calculated by multiplying the baseline value of the interference intensity by the interference diffusion coefficient.

[0049]

[0050] in: Indicates the effective interference radius of the interference source. This represents the baseline value of the interference intensity from the interference source. This represents the interference diffusion coefficient (meters per unit intensity). A spherical interference zone is constructed with the spatial coordinates of the interference source as the center and the effective interference radius as the boundary. When the Euclidean distance is greater than the effective interference radius, the detection device can detect normally outside the spherical interference zone. When the Euclidean distance is less than or equal to the effective interference radius, the detection device enters the spherical interference zone and generates an interference status signal.

[0051] The spatial coordinates of the interference sources were obtained through an ultrasonic beacon system deployed at the detection site. Each interference source was equipped with an ultrasonic tag with an active transmission frequency of 40kHz. The microphone array on the detection device calculated the spatial coordinates using a time-of-arrival (TOA) positioning algorithm. The baseline interference intensity was derived from a database of characteristic parameters of the interference sources. The database recorded typical interference intensity values ​​for various types of equipment. The baseline interference intensity for frequency converters was set to 5.0, and for high-power wireless equipment, it was set to 8.0. The interference diffusion coefficient was determined based on the characteristics of the environmental medium. The free space environmental diffusion coefficient was set to 1.0, the concrete wall environmental diffusion coefficient was reduced to 0.7, and the metal shielding environmental diffusion coefficient was further reduced to 0.3. The effective interference radius calculation module was integrated into the digital signal processor of the detection device, and the real-time refresh frequency was synchronized with the motion control cycle of the detection device. The spherical interference zone was constructed using a three-dimensional spatial discretization method, dividing the detection area into 1 cubic meter voxel units, and calculating the Euclidean distance between each voxel unit and the spatial coordinates of the interference source. A voxel unit is marked as an interfering voxel when the Euclidean distance from its spatial coordinates to the interference source is less than or equal to the effective interference radius. The set of all interfering voxels constitutes a complete spherical interference region. The real-time coordinates of the detection device are obtained by fusing data from the Global Positioning System, Inertial Measurement Unit, and Visual Odometry, with a coordinate update frequency of 100Hz to ensure real-time positioning. The Euclidean distance calculation is accelerated using a floating-point arithmetic unit, and the square root operation of the sum of the squares of the three-dimensional differences between the spatial coordinates of the detection device and the spatial coordinates of the interference source is completed in microseconds.

[0052] The generation of interference signals triggers a switching of the detection device's operating mode. In normal detection mode, the device moves along a predetermined trajectory; under interference conditions, a trajectory correction program is activated. Dynamic updates of the interference source's spatial coordinates are achieved through a wireless sensor network. Moving interference sources broadcast their location information every second, and the detection device immediately updates the interference factor data upon receiving this information. Adaptive adjustments to the effective interference radius are based on historical interference records. When the same interference source causes multiple false detections consecutively, the system automatically reduces its interference diffusion coefficient by 10%. A spherical interference zone is visualized on the detection monitoring interface. A red semi-transparent sphere represents the real-time interference range; the detection device icon turns orange when it enters the spherical area, triggering an alarm. Calibration of the interference intensity baseline uses a standard signal source comparison method. The detection device measures the signal attenuation curve near a standard signal source with known interference intensity to infer the actual interference intensity baseline. On-site determination of the interference diffusion coefficient uses a signal strength surveying vehicle. The vehicle measures the interference field strength distribution at grid points in the detection area and fits the data to obtain the diffusion coefficient under specific conditions. The physical basis of the effective interference radius calculation formula is the electromagnetic wave propagation model. The simplified formula is suitable for engineering applications while ensuring calculation accuracy. The detection device's behavior strategy within the spherical interference zone includes reducing its movement speed, enhancing sensor filtering, and activating redundant measurement mechanisms, with multiple measures ensuring detection reliability. The coordinate system for calculating Euclidean distance uses a standardized Cartesian coordinate system, with the origin set at a ground marker in the southwest corner of the detection area. The X-axis points due north, the Y-axis points due east, and the Z-axis is perpendicular to the ground and pointing upwards. Measurement errors in the spatial coordinates of the interference source are eliminated using a weighted average method; the weighted average of measurements from multiple ultrasonic receivers is used as the final coordinates. A 5% safety margin is set for the boundary buffer zone of the effective interference radius, and the actual judgment condition is modified to trigger an interference state when the Euclidean distance is less than or equal to 1.05 times the effective interference radius. Real-time rendering of the spherical interference zone uses the OpenGL graphics library, and the interference zone volume is calculated using voxel counting. A system warning is generated when the interference zone volume abnormally increases.

[0053] The positioning accuracy improvement scheme for the detection device within the spherical interference zone employs ultra-wideband assisted positioning, with an ultra-wideband base station installed at the boundary of the interference zone providing centimeter-level positioning compensation. The storage format for interference factor data uses a timestamp index structure, with each interference source record containing three-dimensional floating-point spatial coordinates, a single-precision floating-point value for the interference intensity baseline, and a single-precision floating-point value for the interference diffusion coefficient. The transmission of interference status signals utilizes a high-priority message queue to ensure the control unit responds to status changes within 10 milliseconds. Outlier detection of interference source spatial coordinates employs a sliding window variance method; when coordinate data mutations exceed a threshold, predicted coordinates are used to replace measured values. The collision detection algorithm optimization for the spherical interference zone utilizes bounding box technology; the intersection point between the detection device's trajectory and the bounding box of the interference zone is pre-calculated, and anti-interference procedures are initiated in advance when an intersection point exists. Environmental correction for the interference intensity baseline considers temperature and humidity factors; the interference intensity baseline for metal equipment is increased by 5% in high-temperature environments and decreased by 3% in high-humidity environments. Dynamic adjustment of the interference diffusion coefficient is based on real-time environmental monitoring; the diffusion coefficient is multiplied by an attenuation factor of 0.8 during rainfall and by an enhancement factor of 1.1 during strong winds. The calculation results of the effective interference radius are smoothed and filtered to avoid the detection device frequently switching states at the boundary of the interference zone due to radius jumps.

[0054] The movement trajectory record of the detection device within the spherical interference zone is encrypted and stored. The trajectory data includes a timestamp, three-dimensional coordinates, interference intensity value, and multi-dimensional information on the device's operating status. Coordinate system transformation for the spatial coordinates of the interference source supports multiple formats, including automatic conversion between WGS84 latitude and longitude coordinates and on-site rectangular coordinates. Precision control for Euclidean distance calculation uses double-precision floating-point arithmetic to ensure a distance error of less than 1 cm. Volume monitoring of the spherical interference zone includes an upper limit alarm; when the effective interference radius of a single interference source exceeds 50 meters, manual verification of the interference's rationality is required. Energy management of the detection device under interference conditions employs dynamic voltage regulation, reducing the main processor frequency to save power and extend battery life. Integrity verification of interference factor data uses cyclic redundancy check codes; retransmission is requested in case of data transmission errors to ensure parameter accuracy. The remote configuration interface for the effective interference radius supports a secure protocol, allowing authorized personnel to modify the interference intensity baseline and interference diffusion coefficient via an encrypted channel. Historical data records of the spherical interference zone support retrospective analysis, allowing querying of the spatial distribution of the interference zone and the interaction records of the detection device at any given time. The detection device switches to a multi-band frequency hopping mode for communication within the spherical interference zone to avoid the blocking effect of interference sources on the wireless link. Long-term monitoring of the spatial coordinates of the interference source generates a trajectory map of the interference source's movement, used to analyze the movement patterns and optimize detection path planning. Interference status signals have higher priority than regular detection tasks; signal triggering immediately interrupts the current detection process and enters obstacle avoidance mode. Hierarchical management of interference intensity baseline values ​​categorizes interference sources into three levels: strong, medium, and weak, with different processing strategies corresponding to different levels. Adaptive learning of the interference diffusion coefficient is based on machine learning algorithms, recording the actual impact range of each interference event to optimize the diffusion coefficient estimate. The effective interference radius is displayed in real-time on the detection device's status panel, with the numerical color gradually changing from green to red as the radius increases. Edge detection of the spherical interference zone uses a gradient calculation method; a sudden change in gradient value when the detection device approaches the interference zone provides an early warning signal.

[0055] Example 4: See Figure 3 The process of determining whether the detection trajectory has deviated defines the ideal detection distance range and allowable attitude deviation range of the detection device relative to the energy meter. When the actual distance between the detection device and the energy meter is within the ideal detection distance range and the attitude angle of the acquisition device is within the allowable attitude deviation range, the detection trajectory meets the requirements. When the actual distance exceeds the ideal detection distance range or the attitude angle of the acquisition device exceeds the allowable attitude deviation range, the detection trajectory is determined to have deviated, and a trajectory correction command is generated. The ideal detection distance range is set based on the energy meter model and specifications; for single-phase energy meters, the ideal detection distance range is set to 0.5 meters to 1.2 meters, and for three-phase energy meters, it is set to 0.8 meters to 1.5 meters. The allowable attitude deviation range is determined based on the optical characteristics of the acquisition device; the pitch angle deviation range is set to ±5 degrees, the yaw angle deviation range is set to ±8 degrees, and the roll angle deviation range is set to ±3 degrees (see Table 1).

[0056] Table 1: Standard Parameters for Electricity Meter Detection Track

[0057]

[0058] The actual distance measurement employs a fusion scheme of a laser rangefinder sensor and a stereo vision system. The laser rangefinder sensor, model LDS-100A, achieves a measurement accuracy of ±1 mm. The stereo vision system uses dual 2-megapixel cameras to form a binocular stereo matching system. Attitude angle measurement relies on the nine-axis inertial measurement unit (IMU) built into the detection device. The IMU includes a three-axis gyroscope, a three-axis accelerometer, and a three-axis magnetometer. The gyroscope range is set to ±2000 degrees / second, the accelerometer range to ±16g, and the magnetometer range to ±8 Gauss. The lower limit of the ideal detection distance range ensures that the image acquisition resolution meets the defect identification requirements, while the upper limit avoids perspective distortion affecting measurement accuracy. The allowable attitude deviation range is set to consider image distortion control requirements. Excessive pitch angle deviation leads to trapezoidal image distortion, excessive yaw angle deviation causes perspective error, and excessive roll angle deviation causes image rotation. When the detection trajectory meets the requirements, the detection device maintains the current motion state and continues to perform the detection task. When the detection trajectory deviates, the trajectory correction command includes distance correction and angle correction. The distance correction is calculated as the difference between the median of the actual distance and the ideal detection distance. The angle correction is decomposed into three components: pitch angle correction, yaw angle correction, and roll angle correction. The trajectory correction command is transmitted to the motion control system of the detection device, which drives the servo motor to adjust the spatial position of the detection device and the orientation of the acquisition equipment. The sampling frequency for actual distance data is set to 100Hz, and the sampling frequency for attitude angle data is set to 200Hz. High-speed sampling ensures real-time detection of trajectory deviations. The ideal detection distance range is dynamically fine-tuned according to ambient lighting conditions. In strong light conditions, the distance range is appropriately reduced to decrease reflection interference, while in low light conditions, the distance range is expanded to compensate for underexposure.

[0059] Calibration within the permissible attitude deviation range is performed using a standard calibration board. The detection device acquires images of the standard calibration board in a laboratory environment, optimizing attitude angle tolerance parameters to ensure image quality. The actual distance measurement values ​​are filtered using a Kalman filter algorithm to eliminate sensor noise and transient fluctuations. Sensor fusion of attitude angle data employs a complementary filtering method, combining the short-term accuracy of the gyroscope and the long-term stability of the accelerometer. The trajectory deviation judgment logic is set with a lag interval to avoid frequent state switching near the boundary; deviation judgment is not triggered if the actual distance is within 0.05 meters of the ideal detection distance boundary. The execution priority of trajectory correction commands is divided into two levels: emergency correction and normal correction. Emergency correction mode is activated when the distance deviation exceeds 30% or the angle deviation exceeds 50%. The motion control system of the detection device parses the trajectory correction commands to generate motor control pulses. The pulse equivalent of the horizontal movement motor is set to 0.1 mm / pulse, and the pulse equivalent of the rotation motor is set to 0.01 degrees / pulse. Temperature compensation for the actual distance measurement values ​​is based on the sensor temperature characteristic curve; the distance reading is corrected by 0.05% for every 10 degrees Celsius change in ambient temperature. The hard magnetic interference compensation for attitude angle measurement employs an ellipse fitting algorithm to eliminate the influence of ferromagnetic materials on the magnetometer at the testing site. The data structure for the ideal detection distance range includes three fields: minimum value, maximum value, and ideal value. The motion control of the detection device uses the ideal value as the target position. Data storage for the allowable attitude deviation range uses a binary bit-field format, saving storage space and improving reading speed.

[0060] The detection trajectory status is graphically displayed on the detection monitoring interface. A green trajectory line indicates a compliant state, while a red trajectory line indicates a deviation. Historical actual distance data records support curve playback, allowing analysis of distance change trends to optimize motion control parameters. Real-time waveform display of attitude angle data aids in diagnosing sensor faults; abnormal waveforms trigger sensor self-test procedures. The remote configuration interface for the ideal detection distance range supports security authentication, allowing authorized engineers to adjust distance parameters via network to adapt to new energy meters. Grouped management of attitude deviation ranges supports multiple detection modes: the standard detection mode uses a default tolerance, while the precision detection mode reduces the tolerance by half to improve detection accuracy. Statistical reports of detection trajectory deviation events are generated, including indicators such as the number of deviations, average deviation, and maximum deviation. This report data is used to evaluate the motion control performance of the detection device. Periodic calibration of the actual distance measurement system uses standard length gauge blocks, performing a full-range calibration monthly to ensure measurement accuracy. Calibration of the attitude angle measurement system uses a 3D turntable with a positioning accuracy of 0.01 degrees, providing a reference angle. Adaptive adjustment of the ideal detection distance range is based on historical detection data, gradually optimizing the distance interval with the most successful detections into a new ideal range. The allowable range of attitude deviations is optimized using image quality assessment feedback, with the attitude angle data that scores the highest in image sharpness used to correct the tolerance range.

[0061] Example 5: The process of analyzing real-time image information using artificial intelligence algorithms involves acquiring visible light images of an electricity meter, normalizing the visible light images, and applying noise suppression algorithms to reduce image noise. Based on the standard appearance model of the electricity meter, texture feature thresholds are set to segment the electricity meter region. The image is converted to a brightness / saturation space, and a region growing algorithm is used to extract connected regions that conform to the texture features of the electricity meter. A contour detection algorithm is used to identify image contours, and the contour shapes are matched with the reference template of the electricity meter for similarity, eliminating inconsistent parts. A clustering analysis algorithm is used to group pixels, filter out regions of interest (ROIs) of the electricity meter, and perform multi-scale analysis on the ROIs to evaluate the appearance integrity. Near-infrared images of the electricity meter are acquired simultaneously, and reflectivity thresholds are set based on material reflectivity characteristics to extract regions in the near-infrared images that meet the reflectivity thresholds. Multi-band fusion technology is used to enhance image details, and contour detection is combined to preserve the edge features of the electricity meter. The reflectivity distribution in the near-infrared images is analyzed to identify areas with abnormal reflection. If both the visible light and near-infrared images detect an anomaly, a defect confirmation signal is generated; if no anomalies are found in either image, the current state is maintained; if one image shows an anomaly, a re-inspection signal is generated. After generating a defect alarm, the defect's characteristic attributes, including shape category and area ratio, are extracted and archived along with the image data. The user terminal is configured with a query module for retrieving and displaying defect records based on specific criteria.

[0062] Visible light image acquisition used a 20-megapixel industrial camera equipped with a polarizing filter to eliminate surface reflections, with an image resolution of 3840×2160 pixels. Normalization linearly transformed the image pixel values ​​to the range of 0-255, and an illumination compensation algorithm adjusted local contrast. Noise suppression employed a non-local mean denoising algorithm, with a search window size of 11×11 pixels and a similarity window size of 5×5 pixels. Texture feature thresholds were set based on typical texture patterns of the electricity meter casing: a grayscale gradient variance greater than 50 for metal casings and a local binary pattern feature number greater than 80 for plastic casings. Brightness and saturation space transformation preserved the H and S components of the HSB color model. The seed point for the region growing algorithm was selected from the image center region, and the growth threshold was set to 15 color levels based on the color difference between adjacent pixels. The contour detection algorithm used the Canny edge detection operator, with a Gaussian filter kernel size of 5×5 and a high-low threshold ratio of 1:3. Similarity matching used Hu moment invariants to calculate contour shape differences, with a matching threshold of 0.2. The clustering analysis algorithm uses the K-means algorithm to classify pixels into three categories: background, shell, and display screen. The number of clusters K=3, and the number of iterations is 100. Multi-scale analysis constructs a Gaussian pyramid with 4 layers and a downsampling rate of 0.5 per layer. Near-infrared image acquisition uses an InGaAs sensor with a wavelength range of 900-1700 nm. The reflectance threshold is set to be greater than 0.6 for metallic materials and less than 0.3 for plastic materials. Multi-band fusion technology registers and weights visible light and near-infrared images, with a weight of 0.7 for visible light and 0.3 for near-infrared images. Anomaly area identification uses adaptive threshold segmentation, with the threshold dynamically calculated based on local image statistical characteristics. Confirmed defect signals trigger a three-level alarm mechanism: Level 1 indicates minor scratches, Level 2 indicates significant cracks, and Level 3 indicates structural damage. Signals awaiting re-inspection initiate the review process, with the detection device automatically adjusting the angle and re-acquiring images. Defect feature attribute extraction includes shape factor calculation. Area ratio = number of defective pixels / total number of pixels.

[0063] The user terminal query module supports SQL queries, allowing searches by time range, defect level, and meter model. Image data archiving uses JPEG2000 compression format with a compression ratio of 20:1, and feature attributes are stored in JSON format. The standard appearance model library contains 3D models of 50 meter models, each providing template images from the front, side, and top views. Texture feature threshold training uses a support vector machine classifier, with training samples including 1000 normal meter images and 500 defective meter images. The connected region area threshold for the region growing algorithm is set to a minimum of 1000 pixels and a maximum of 100,000 pixels, eliminating areas with abnormal areas. The baseline template for contour detection uses an active shape model with 80 feature points, and shape parameters are reduced using principal component analysis. The similarity matching rejection threshold is set to 0.8; matches below this threshold are considered failed. Initial center selection for cluster analysis uses the K-means++ algorithm to avoid getting trapped in local optima. Feature extraction for multi-scale analysis uses SIFT descriptors, with a feature point limit of 500. The variance of regional reflectance is calculated by reflectance distribution analysis. A variance greater than 0.1 is considered an anomaly in reflectance.

[0064] Image registration using multi-band fusion employs SURF feature point matching, requiring more than 50 matching point pairs. Defect confirmation logic requires an overlap of more than 60% between abnormal areas in the visible light and near-infrared images. The processing delay for signals awaiting re-inspection is set to 2 seconds, during which the detection device remains stationary. Defect shape classification utilizes a convolutional neural network with a ResNet-50 network structure, outputting three categories: circular, bar-shaped, and irregular. Area ratio calculation uses pixel statistics, counting the number of pixels in connected components within the defect area. The user terminal display interface adopts a layered design: the bottom layer displays the distribution map of electricity meter locations, the middle layer overlays defect markers, and the top layer displays detailed information. Query results can be exported in Excel format, including image thumbnail links and feature attribute tables. The standard appearance model update mechanism supports online learning; a baseline template is automatically generated after 10 inspections of a new model of electricity meter. The noise suppression algorithm's parameters are adaptively adjusted based on image signal-to-noise ratio (SNR) measurements; the filtering intensity is increased when the SNR is below 30dB. Color quantization in the brightness and saturation space quantizes hue into 360 levels and saturation into 100 levels. The region growing algorithm's stopping condition is set at 100 consecutive pixels failing to meet the growing criteria. Contour approximation for contour detection uses the Douglas-Peucker algorithm with an approximation threshold of 0.01. A multi-template voting mechanism for similarity matching selects the best match from three viewpoint templates. The elbow rule for cluster analysis determines the optimal number of clusters, and contour coefficients evaluate cluster quality. The feature pyramid for multi-scale analysis uses the Laplacian operator to enhance edge information.

[0065] Near-infrared image reflectance calibration uses a standard white board as a reference, and the sensor is calibrated quarterly. Morphological processing of reflective anomaly areas employs opening operations to eliminate noise points, with a 3×3 structuring element size. Priority is set for defect confirmation signals; structural damage defects immediately halt the detection process. The maximum number of retries for signals awaiting review is set to 3; exceeding this limit results in manual review. A B+ tree index is used for the database index of defect feature attributes to accelerate query speed. Visual charts on the user terminal display the statistical distribution of defects, with pie charts showing the proportion of various defects and line charts showing defect trends. Image data backup employs an off-site disaster recovery solution, with daily incremental backups and weekly full backups. Black level correction in normalization processing uses optical black area reference values ​​to eliminate the influence of dark current. Morphological closing operations for texture feature segmentation fill small holes, with a 5×5 structuring element size. The region growing algorithm merges adjacent similar regions, with a merging threshold of less than 10 color levels.

[0066] See Figure 4 This study comprehensively presents the quality assessment results of each stage of image processing in the detection of external defects in electricity meters using radar charts. The six dimensions of the radar chart correspond to key performance indicators such as image quality, noise suppression, texture sharpness, contour matching, reflectivity analysis, and processing efficiency. The image quality dimension comprehensively evaluates the acquisition quality of visible light and near-infrared images, reflecting the imaging capabilities of the detection device under different lighting conditions. The noise suppression dimension shows the system's filtering effect on image noise, ensuring that the accuracy of defect identification is not affected by noise. The texture sharpness index measures the system's ability to segment and extract the texture features of the electricity meter surface, which is crucial for identifying minor scratches and wear. The contour matching dimension evaluates the system's effectiveness in comparing the detected image with a standard template, reflecting the accuracy of shape recognition and contour analysis. The reflectivity analysis dimension demonstrates the ability to detect abnormal areas based on material properties, identifying potential defects by analyzing the reflectivity characteristics of different materials. The processing efficiency dimension reflects the overall image processing speed performance of the system, ensuring that the detection task can be completed efficiently. The closer the score of each dimension of the radar chart is to the outer ring, the better the performance in that aspect. This multi-dimensional comprehensive assessment provides a clear direction for system optimization, helping technicians identify performance bottlenecks and make targeted improvements. Overall, the system maintains a high performance level in all key aspects, meeting the accuracy and efficiency requirements for detecting physical defects in electricity meters.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI automatic identification-based electric energy meter appearance defect detection method, characterized in that, The utility model relates to an electric energy meter detection system, comprising: Grouping electric energy meters to form multiple detection units; receiving initial detection instructions issued by users and starting detection devices; Collecting basic appearance information of electric energy meters in each detection unit during the execution of initial detection processes, and obtaining historical defect records of each detection unit from storage units; Evaluating defect probability levels and criticality levels of each detection unit according to the basic appearance information and the historical defect records; planning individualized detection sequences of detection devices for checking each detection unit according to the criticality levels and the defect probability levels; Capturing real-time image information of electric energy meters when detection devices execute detection tasks according to the individualized detection sequences; Obtaining interference factor data in detection sites and judging whether the detection devices are in an interfered state; If the detection devices are in the interfered state, judging whether the detection trajectories deviate according to real-time collected position information and correcting the detection trajectories when deviation occurs; Analyzing the real-time image information by using artificial intelligence algorithms to find appearance defects of electric energy meters, and generating defect alarms when defects are found and transmitting defect locations and related image information to user display terminals. 2.The AI-based automatic identification-based electric energy meter appearance defect detection method according to claim 1, characterized in that, The basic appearance information includes surface material categories and installation orientation information of electric energy meters, and the historical defect records include historical defect occurrence frequencies and repair time consumption of each defect. 3.The AI-based automatic identification electric energy meter appearance defect detection method of claim 2, wherein, The process of evaluating defect probability levels and criticality levels specifically includes: collecting historical defect occurrence frequencies of each detection unit, obtaining repair time consumption of each defect of each detection unit, weighting and summing the repair time consumption of each defect in chronological order to obtain weighted repair time consumption, calculating defect indexes of each detection unit based on the historical defect occurrence frequencies and the weighted repair time consumption, comparing the defect indexes with preset defect index threshold values to determine defect probability levels of the detection units as high probability levels, medium probability levels or low probability levels, obtaining surface material categories and installation orientation information of each detection unit, calculating criticality indexes of each detection unit through material durability and installation complexity, and comparing the criticality indexes with preset criticality index threshold values to determine criticality levels of the detection units as high criticality levels, medium criticality levels or low criticality levels. 4.The AI-based automatic identification electric energy meter appearance defect detection method of claim 3, wherein, The process of planning individualized detection sequences specifically includes: taking criticality levels of detection units as primary selection criteria, taking defect probability levels of detection units as secondary selection criteria, and taking path lengths between detection devices and detection units as reference criteria; the detection devices start from starting points, preferentially select detection units with higher defect probability levels for detection if multiple detection units with the same criticality levels exist nearby, and select detection units with the shortest path lengths for detection if the criticality levels and the defect probability levels of the detection units are the same. When the detection device reaches the detection point, if the point corresponds to only one detection unit, the detection unit is directly detected; if the point corresponds to multiple detection units, the detection unit with higher criticality level is preferentially detected; if the criticality levels of the multiple detection units are the same, the detection unit with higher defect probability level is preferentially detected; if the criticality level and the defect probability level are the same, the detection unit is selected according to the detection unit number sequence; if the point is not connected to other detection units and there are undetected detection units, the detection unit with the shortest path length is preferentially detected; When all the detection units are detected, the detection device returns to the starting point along the optimal path; the personalized detection sequence of the detection device is generated according to the above rules; during the movement of the detection device, the detection residence time is dynamically adjusted according to the criticality level and the defect probability level of the detection unit. 5.The AI-based automatic identification-based electric energy meter appearance defect detection method according to claim 1, characterized in that, The interference factor data includes spatial coordinates of the interference source, an interference intensity base value and an interference diffusion coefficient; The spatial coordinates of the interference source are position parameters of the interference source, the interference intensity base value is a quantitative initial interference ability of the interference source, and the interference diffusion coefficient is a parameter of interference propagation attenuation characteristics. 6.The AI-based automatic identification and appearance defect detection method of an electric energy meter according to claim 5, characterized in that, The process of judging whether the detection device is in the interference state is specifically: measuring the Euclidean distance between the real-time coordinates of the detection device and the spatial coordinates of the interference source; obtaining the interference intensity base value and the interference diffusion coefficient of the interference source, and obtaining the effective interference radius of the interference source by multiplying the interference intensity base value by the interference diffusion coefficient; taking the spatial coordinates of the interference source as the center and the effective interference radius as the boundary, a spherical interference region of the interference source is constructed; when the Euclidean distance is greater than the effective interference radius, the detection device is normally detected outside the spherical interference region; when the Euclidean distance is less than or equal to the effective interference radius, the detection device enters the spherical interference region and generates an interference state signal.

7. The AI-based automatic identification electric energy meter appearance defect detection method according to claim 6, characterized in that, The process of judging whether the detection trajectory deviates is specifically: defining an ideal detection distance range and an allowable attitude deviation range of the detection device relative to the electric energy meter; if the actual distance between the detection device and the electric energy meter is within the ideal detection distance range and the attitude angle of the acquisition equipment is within the allowable attitude deviation range, the detection trajectory meets the requirements; if the actual distance exceeds the ideal detection distance range or the attitude angle of the acquisition equipment exceeds the allowable attitude deviation range, it is determined that the detection trajectory deviates, and a trajectory correction instruction is generated. 8.The AI-based automatic identification-based electric energy meter appearance defect detection method according to claim 1, characterized in that, The process of analyzing the real-time image information by using the artificial intelligence algorithm is specifically: A visible light image of the electric energy meter is collected; the visible light image is normalized, a noise suppression algorithm is applied to reduce image noise, and a texture feature threshold is set according to a standard appearance model of the electric energy meter to segment the electric energy meter region; the image is converted to a brightness saturation space, a region growing algorithm is used to extract a connected region conforming to the texture feature of the electric energy meter, and the main image of the electric energy meter is retained; a contour detection algorithm is used to identify the image contour, the contour shape is matched with a reference template of the electric energy meter in terms of similarity, and inconsistent parts are removed; Pixel points are grouped by using a clustering analysis algorithm, and a region of interest of the electric energy meter is screened out; multi-scale analysis is performed on the region of interest of the electric energy meter, and appearance integrity is evaluated. The near-infrared image of the electric energy meter is synchronously collected, a reflectivity threshold is set according to the reflection characteristics of materials, and the area meeting the reflectivity threshold in the near-infrared image is extracted; a multi-band fusion technology is used to enhance the image details, and the edge features of the electric energy meter are reserved in combination with contour detection; the reflectivity distribution in the near-infrared image is analyzed, and the abnormal reflection area is identified; if the visible light image and the near-infrared image both detect an abnormality, a confirmation defect signal is generated; if no abnormality is found in the two kinds of images, the current state is maintained; if an abnormality is displayed in one kind of image, a signal to be rechecked is generated. 9.The AI-based automatic identification-based electric energy meter appearance defect detection method according to claim 1, characterized in that, The method further comprises: after generating the flaw alarm, extracting feature attributes of the flaw including a shape category and an area ratio, and binding and archiving the feature attributes with the image data; and a user terminal configuration query module is used to retrieve and display the flaw record according to conditions.

10. An AI automatic identification-based electric energy meter appearance defect detection system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the electric energy meter appearance defect detection method based on AI automatic identification in any one of claims 1 to 9.

Images