An electricity meter auditing method based on AI intelligent identification of electricity meter readings

Abstract

This invention discloses an AI-based intelligent meter reading recognition method for meter auditing, comprising the following steps: S1, Multi-type data collection: Collecting on-site image data of mechanical meters, digital meters, and IoT meters to construct a labeled dataset; S2, Adaptive model training: Based on the labeled dataset, training a recognition model using a unified detection architecture and classification preprocessing mode. The model includes a DB++ text detection network and a lightweight recognition network, and the robustness of the model is optimized through data augmentation and transfer learning; S3, Multi-type meter recognition: Adapting the preprocessing strategy to the meter images to be recognized according to their types, inputting them into the trained recognition model, and outputting structured data containing meter number, current reading, and meter type after specific area detection and text recognition; S4, Dynamic threshold auditing: Generating a dynamic threshold range based on user type and historical electricity consumption patterns, calculating the difference between the current cycle and the previous cycle reading, comparing the threshold to determine anomalies, and generating an alarm.

Description

Technical Field

[0001] This invention relates to the field of intelligent power identification, and in particular to an AI-based intelligent identification method for electricity meter readings. Background Technology

[0002] Currently, China Tower mainly relies on manual on-site inspections, photographic evidence collection, and manual back-end auditing for meter reading and auditing. In practice, the number of meters in various cities is enormous, and manual inspections not only consume a lot of manpower and resources but are also prone to reading errors, increasing the difficulty of auditing. Furthermore, some meters lack historical readings, making reliance solely on manual experience or spot checks inefficient and unable to promptly detect abnormal electricity usage.

[0003] Considering that electricity meter readings can essentially be expressed through image features and pattern recognition, utilizing AI algorithms for intelligent analysis and recognition of electricity meter images can provide effective support for electricity auditing. Currently, the application of image recognition combined with artificial intelligence has become a trend, achieving good results in scenarios such as financial document recognition, license plate recognition, and industrial inspection. However, for automatic OCR recognition and auditing of electricity meter readings, there is still a lack of a method that is adaptable to complex environments, compatible with various meter types, and possesses anomaly detection capabilities.

[0004] Therefore, this invention proposes an AI-based intelligent meter reading recognition method for meter auditing to solve the above problems. Summary of the Invention

[0005] This invention overcomes the shortcomings of existing technologies and provides an AI-based intelligent meter reading recognition method for meter auditing. By combining image processing and deep learning models, it achieves automated recognition and result comparison of meter images, effectively improving the efficiency and accuracy of meter auditing.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a meter auditing method based on AI intelligent recognition of meter readings, comprising the following steps: S1. Multi-type data collection: Collect on-site image data of mechanical meters, digital meters and IoT meters, and construct a labeled dataset. The labeled content includes meter reading frame, meter number frame and meter type label. S2. Adaptive Model Training: Based on the labeled dataset, a unified detection architecture and classification preprocessing mode are used to train the recognition model. The model includes a DB++ text detection network and a lightweight recognition network. Data augmentation and transfer learning are used to optimize the robustness of the model. S3. Multi-type meter recognition: The image of the meter to be recognized is adapted to the type of the preprocessing strategy. The image is input into the trained recognition model. After specific area detection and text recognition, the output is structured data containing meter number, current reading and meter type. S4. Dynamic threshold audit: Based on user type and historical electricity consumption patterns, a dynamic threshold range is generated, the difference between the current cycle and the previous cycle reading is calculated, the threshold is compared to determine anomalies and an alarm is generated, forming an automated closed loop of identification, auditing and alarming.

[0007] In a preferred embodiment of the present invention, the annotation form of the annotation content in step S1 is a rectangle or four-point coordinates, which accurately marks the location of the meter reading and the meter number area.

[0008] In a preferred embodiment of the present invention, the data augmentation in step S2 adopts a combined strategy, including ±15° rotation, 0.8-1.0 ratio random cropping, horizontal flipping, and ±20% illumination adjustment. The lightweight recognition network is any one of MobileNetV3, CRNN, or SVTR.

[0009] In a preferred embodiment of the present invention, the model training in step S2 adopts a transfer learning and two-stage optimization strategy: the model is initialized based on DB++ weights pre-trained on a general text dataset, DBLoss is used as the loss function, the Adam optimizer has an initial learning rate of 0.0005, the warmup strategy is executed for the first two epochs, and the cosine annealing learning rate is scheduled thereafter.

[0010] In a preferred embodiment of the present invention, the preprocessing strategy in step S3 includes: performing pointer region enhancement and noise filtering for mechanical pointer meters, performing character edge sharpening for digital LCD meters, and performing form region segmentation for IoT form-type meters to adapt to the image features of different meters.

[0011] In a preferred embodiment of the present invention, the structured data in step S3 adopts JSON format, and the fields include ["meter_id":"meter number","current_reading":"current reading","meter_type":"meter type","recognition_time":"recognition time"], which is directly connected to the power management system.

[0012] In a preferred embodiment of the present invention, the dynamic threshold range in step S4 is generated through iterative optimization using historical electricity consumption data, and the formula is: , Where α is the lower limit coefficient with a value of 0.3-0.5, β is the upper limit coefficient with a value of 1.5-3.0, k is the current period, and Δx is the difference in historical period readings. It is the average difference of the first k-1 periods.

[0013] In a preferred embodiment of the present invention, the anomaly determination based on the dynamic threshold range and the difference between readings in step S4 includes three sub-types: when Δx < Δmin, it is determined to be a low power anomaly; when Δx > Δmax, it is determined to be an excessive power consumption anomaly; and when Δx < 0, it is determined to be a negative difference anomaly. Different anomaly types correspond to different alarm levels.

[0014] In a preferred embodiment of the present invention, step S4, meter audit analysis, involves identifying the meter reading R. t After matching the corresponding meter number, anomaly detection is performed based on threshold comparison, specifically including: Difference calculation: Δx = R t -R t-1 Among them, R t R represents the meter reading for the current cycle. t-1 This represents the meter reading for the previous cycle, where Δx is the difference in electricity consumption over the cycle. Threshold setting: Based on user type or historical electricity consumption patterns, preset upper and lower threshold ranges: Δmin≤Δx≤Δmax; Anomaly detection: If the difference Δx exceeds the threshold range, it is considered an anomaly. Δx < Δmin or Δx > Δmax ⇒ Abnormal; Output: Meter number, meter reading R t Output is provided for backend auditing or alarm purposes.

[0015] In a preferred embodiment of the present invention, the meter audit analysis algorithm in step S4 is specifically as follows: A threshold ε is set, which is a multiple of the average historical reading error. The identified meter reading is R. t The historical period reading is y; like If the current identification result conforms to the historical electricity consumption pattern and is within the normal threshold, it will be automatically stored in the system. like Even if Δx is in the range of Δmin to Δmax, it will still trigger a manual audit, requiring a second manual audit.

[0016] This invention addresses the shortcomings of the prior art and has the following beneficial effects: This invention presents a comprehensive design encompassing multi-type data collection, adaptive model training, and classification preprocessing. It employs a deep learning OCR architecture comprised of a DB++ text detection network and a lightweight recognition network. Targeting the differentiated image features of mechanical, digital, and IoT meters, it adapts classification preprocessing strategies such as pointer area enhancement, character edge sharpening, and form area segmentation. Combined with combined data augmentation and transfer learning optimization, it achieves automatic detection, segmentation, and recognition of various meter readings. This not only replaces the traditional manual reading method, significantly reducing human error, but also ensures the accuracy and efficiency of reading extraction thanks to the model's full adaptability to multiple meter types and robustness to complex environments. This provides standardized and highly reliable data source support for subsequent auditing work.

[0017] Meanwhile, after obtaining structured identification results such as meter number and current reading, a dynamic threshold range that fits the user type and historical electricity consumption patterns is generated. The difference between the current cycle and the previous cycle reading is calculated and compared with the threshold. Abnormal situations such as low power consumption, excessive power consumption, and negative difference are automatically identified. If the identification result deviates from the historical electricity consumption pattern or exceeds the dynamic threshold range, it will be automatically marked as abnormal and trigger a second manual audit. This forms an automated closed loop of identification-audit-alarm, which supports auditors to quickly locate problematic meters, avoids the inefficient mode of manual verification of each household in traditional audits, significantly improves the pertinence and efficiency of audit work, and reduces operation and maintenance costs. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments; Figure 1 This is a flowchart of a preferred embodiment of the present invention for an electricity meter auditing method based on AI intelligent recognition of electricity meter readings; Figure 2 This is a logic diagram of a preferred embodiment of the electricity meter auditing method based on AI intelligent recognition of electricity meter readings. Figure 3 This is a logic diagram of the meter audit analysis algorithm of a preferred embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] like Figure 1-2 As shown, an electricity meter auditing method based on AI-powered intelligent meter reading recognition includes the following steps: S1. Multi-type data collection: Collect on-site image data of mechanical meters, digital meters and IoT meters, and construct a labeled dataset. The labeled content includes meter reading frame, meter number frame and meter type label. The labeled content is in the form of rectangle or four-point coordinates, accurately marking the location of the meter reading and meter number area.

[0022] S2. Adaptive Model Training: Based on the labeled dataset, a unified detection architecture and classification preprocessing mode are used to train the recognition model. The model includes a DB++ text detection network and a lightweight recognition network. Data augmentation and transfer learning are used to optimize the robustness of the model.

[0023] Preferably, the data augmentation employs a combined strategy, including ±15° rotation, 0.8-1.0 ratio random cropping, horizontal flipping, and ±20% illumination adjustment, with the lightweight recognition network being any one of MobileNetV3, CRNN, or SVTR.

[0024] The model training adopts a transfer learning and two-stage optimization strategy: the model is initialized based on DB++ weights pre-trained on a general text dataset, DBLoss is used as the loss function, the Adam optimizer has an initial learning rate of 0.0005, the warmup strategy is executed in the first 2 epochs, and cosine annealing learning rate scheduling is used thereafter.

[0025] S3. Multi-type meter recognition: The preprocessing strategy is adapted according to the type of the meter image to be recognized. The trained recognition model is input, and after specific area detection and text recognition, the structured data containing meter number, current reading and meter type is output.

[0026] Preferred classification preprocessing strategies include: performing pointer area enhancement and noise filtering for mechanical pointer meters, performing character edge sharpening for digital LCD meters, and performing form area segmentation for IoT form-type meters, adapting to the image features of different meters.

[0027] The structured data uses JSON format, with fields including ["meter_id":"meter number","current_reading":"current reading","meter_type":"meter type","recognition_time":"recognition time"], which can be directly connected to the power management system.

[0028] S4. Dynamic threshold audit: Based on user type and historical electricity consumption patterns, a dynamic threshold range is generated, the difference between the current cycle and the previous cycle reading is calculated, the threshold is compared to determine anomalies and an alarm is generated, forming an automated closed loop of identification, auditing and alarming.

[0029] The dynamic threshold range is generated through iterative optimization using historical electricity consumption data, and the formula is: , Where α is the lower limit coefficient (0.3-0.5), β is the upper limit coefficient (1.5-3.0), k is the current period, and Δx is the difference in historical period readings. It is the average difference of the first k-1 periods.

[0030] The anomaly detection includes three sub-types: when Δx < Δmin, it is judged as low power anomaly; when Δx > Δmax, it is judged as excessive power consumption anomaly; and when Δx < 0, it is judged as negative difference anomaly. Different anomaly types correspond to different alarm levels.

[0031] Electricity meter audit analysis identifies the electricity meter reading R t After matching the corresponding meter number, this method performs anomaly detection based on threshold comparison, specifically including: Difference calculation: Δx = R t -R t-1 Among them, R t R represents the meter reading for the current cycle. t-1 This represents the meter reading for the previous cycle, where Δx is the difference in electricity consumption over the cycle. Threshold setting: Based on user type or historical electricity consumption patterns, preset upper and lower threshold ranges: Δmin≤Δx≤Δmax; Anomaly detection: If the difference Δx exceeds the threshold range, it is considered an anomaly. Δx < Δmin or Δx > Δmax ⇒ Abnormal; Output: Meter number, meter reading R t Output is provided for backend auditing or alarm purposes.

[0032] like Figure 3 As shown, the specific algorithm for electricity meter audit analysis is as follows: Set a threshold ε (ε is a multiple of the average historical reading error), and the identified meter reading is R. tThe historical period reading is y; like If the current identification result conforms to the historical electricity consumption pattern and is within the normal threshold, it will be automatically stored in the system. like Even if Δx is in the range of Δmin to Δmax, it will still trigger a manual audit, requiring a second manual audit.

[0033] Existing technologies often suffer from the limitation that a single identification scheme can only be adapted to specific types of meters, making it difficult to meet the practical application scenarios where mechanical meters, digital meters, and IoT meters coexist in power systems. This invention constructs a labeled dataset by collecting on-site image data from multiple types of meters and designs differentiated classification preprocessing strategies based on the image features of different meters. For mechanical pointer meters, pointer area enhancement and noise filtering are performed; for digital LCD meters, character edge sharpening is performed; and for IoT form-type meters, form area segmentation is performed. Combined with a unified detection architecture recognition model, accurate adaptation to various types of meters is achieved. This design effectively solves the recognition problem caused by the differences in image features among different types of meters, and its applicability covers various metering scenarios in power systems, significantly improving the practicality and versatility of the technical solution.

[0034] This invention employs a combined architecture of a DB++ text detection network and a lightweight recognition network (MobileNetV3, CRNN, or SVTR). The DB++ network provides accurate text region detection, while the lightweight network balances recognition speed and resource consumption. Together, they achieve efficient "detection-recognition" linkage. Simultaneously, a combination of data augmentation strategies, including ±15° rotation, random cropping, and illumination adjustment, enhances the model's adaptability to complex environmental factors such as lighting changes, image noise, and shooting angle deviations. Furthermore, transfer learning and a two-stage optimization strategy (pre-training initialization based on a general text dataset, DBLoss loss function, and cosine annealing learning rate scheduling) further improve the model's robustness. Compared to traditional recognition schemes, this invention significantly improves the accuracy of meter reading recognition in complex scenarios, and the lightweight network design reduces inference latency, meeting the real-time requirements of meter reading collection and providing a reliable data foundation for subsequent auditing work.

[0035] The recognition results are encapsulated into JSON-formatted structured data containing core fields such as meter number, current reading, meter type, and recognition time. This format conforms to the data interaction specifications of the power management system and can be directly connected to the existing power operation and maintenance management platform without additional data format conversion or adaptation development. This solves the problems of inconsistent output data formats, difficult system integration, and high operation and maintenance costs of traditional recognition solutions, greatly improving the engineering implementation efficiency of the technical solution and facilitating its rapid integration into the existing power management system to achieve seamless data flow and sharing.

[0036] Existing auditing techniques often rely on fixed thresholds to determine anomalies, which are difficult to adapt to the differences in electricity consumption habits and historical changes in electricity usage patterns among different users, easily leading to false alarms or missed alarms. This invention generates a dynamic threshold range based on user type and historical electricity consumption data through iterative optimization. , The calculation formula uses α as the lower limit coefficient (0.3-0.5) and β as the upper limit coefficient (1.5-3.0) to ensure that the threshold accurately reflects user electricity consumption characteristics in real time. Anomalies are further categorized into three types: low power consumption (Δx < Δmin), excessive power consumption (Δx > Δmax), and negative difference anomalies (Δx < 0), each with a corresponding differentiated alarm level, enabling precise differentiation and targeted handling of anomaly types. Furthermore, a manual auditing mechanism with an error threshold ε is introduced. Even if the periodic difference is within the threshold range, if the identification result deviates from the historical pattern by more than the error threshold, manual review is still triggered, further reducing the risk of missed anomaly detection and improving the accuracy and rigor of the audit results.

[0037] This invention constructs a fully automated system encompassing multi-type data collection, adaptive model training, multi-type meter identification, dynamic threshold auditing, and anomaly alarms. From meter image acquisition to anomaly detection, no manual intervention is required, transforming the traditional manual meter reading and verification process. This automated closed-loop system not only significantly reduces manpower input and labor costs in power operation and maintenance but also avoids human errors that may occur during manual reading, recording, and verification. It improves the efficiency, precision, and standardization of meter measurement and auditing, providing strong support for the intelligent management of power systems.

[0038] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These are all equivalent modifications and improvements made to the above embodiments based on the essential technology of the present invention, and all of these fall within the protection scope of the present invention.

Claims

1. A method for auditing electricity meters based on AI-powered intelligent recognition of meter readings, characterized in that, Includes the following steps: S1. Multi-type data collection: Collect on-site image data of mechanical meters, digital meters and IoT meters, and construct a labeled dataset. The labeled content includes meter reading frame, meter number frame and meter type label. S2. Adaptive Model Training: Based on the labeled dataset, a unified detection architecture and classification preprocessing mode are used to train the recognition model. The model includes a DB++ text detection network and a lightweight recognition network. Data augmentation and transfer learning are used to optimize the robustness of the model. S3. Multi-type meter recognition: The image of the meter to be recognized is adapted to the type of the preprocessing strategy. The image is input into the trained recognition model. After specific area detection and text recognition, the output is structured data containing meter number, current reading and meter type. S4. Dynamic threshold audit: Based on user type and historical electricity consumption patterns, a dynamic threshold range is generated, the difference between the current cycle and the previous cycle reading is calculated, the threshold is compared to determine anomalies and an alarm is generated, forming an automated closed loop of identification, auditing and alarming.

2. The meter auditing method based on AI intelligent recognition of meter readings according to claim 1, characterized in that: In step S1, the annotation format is a rectangle or four-point coordinates, which accurately marks the location of the meter reading and meter number area.

3. The meter auditing method based on AI intelligent recognition of meter readings according to claim 1, characterized in that: The data augmentation in step S2 employs a combined strategy, including ±15° rotation, 0.8-1.0 ratio random cropping, horizontal flipping, and ±20% illumination adjustment. The lightweight recognition network is any one of MobileNetV3, CRNN, or SVTR.

4. The meter auditing method based on AI intelligent recognition of meter readings according to claim 1, characterized in that: In step S2, the model training adopts a transfer learning and two-stage optimization strategy: the model is initialized based on DB++ weights pre-trained on a general text dataset, DBLoss is used as the loss function, the Adam optimizer has an initial learning rate of 0.0005, the warmup strategy is executed for the first two epochs, and the cosine annealing learning rate is scheduled thereafter.

5. The meter auditing method based on AI intelligent recognition of meter readings according to claim 1, characterized in that: The preprocessing strategy in step S3 includes: performing pointer area enhancement and noise filtering for mechanical pointer meters, performing character edge sharpening for digital LCD meters, and performing form area segmentation for IoT form-type meters to adapt to the image features of different meters.

6. The meter auditing method based on AI intelligent recognition of meter readings according to claim 1, characterized in that: The structured data in step S3 adopts JSON format, with fields including ["meter_id":"meter number","current_reading":"current reading","meter_type":"meter type","recognition_time":"recognition time"], which is directly connected to the power management system.

7. The meter auditing method based on AI intelligent recognition of meter readings according to claim 1, characterized in that: The dynamic threshold range in step S4 is generated through iterative optimization using historical electricity consumption data, and the formula is as follows: , Where α is the lower limit coefficient, ranging from 0.3 to 0.5, β is the upper limit coefficient, ranging from 1.5 to 3.0; k is the current period, and Δx is the difference between historical period readings. It is the average difference of the first k-1 periods.

8. The meter auditing method based on AI intelligent recognition of meter readings according to claim 1, characterized in that: The anomaly determination in step S4 based on the dynamic threshold range and the difference in readings includes three sub-types: when Δx < Δmin, it is determined to be a low power anomaly; when Δx > Δmax, it is determined to be an excessive power consumption anomaly; and when Δx < 0, it is determined to be a negative difference anomaly. Different anomaly types correspond to different alarm levels.

9. The meter auditing method based on AI intelligent recognition of meter readings according to claim 8, characterized in that: Step S4, meter audit analysis, involves identifying the meter reading R. t After matching the corresponding meter number, anomaly detection is performed based on threshold comparison, specifically including: Difference calculation: Δx = R t -R t-1 Among them, R t R represents the meter reading for the current cycle. t-1 This represents the meter reading for the previous cycle, where Δx is the difference in electricity consumption over the cycle. Threshold setting: Based on user type or historical electricity consumption patterns, preset upper and lower threshold ranges: Δmin≤Δx≤Δmax; Anomaly detection: If the difference Δx exceeds the threshold range, it is considered an anomaly. Δx < Δmin or Δx > Δmax ⇒ Abnormal; Output: Meter number, meter reading R t Output is provided for backend auditing or alarm purposes.

10. The meter auditing method based on AI intelligent recognition of meter readings according to claim 9, characterized in that: The specific algorithm for meter auditing and analysis in step S4 is as follows: A threshold ε is set, where ε is a multiple of the average historical reading error. The identified meter reading is R. t The historical periodic reading is y; like If the current identification result conforms to the historical electricity consumption pattern and is within the normal threshold, it will be automatically stored in the system. like Even if Δx is in the range of Δmin to Δmax, it will still trigger a manual audit, requiring a second manual audit.

Images