Intelligent meter reading method, device and equipment based on coprocessing of large and small models

Abstract

The invention provides an intelligent meter reading method, device and equipment based on large and small model co-processing, relates to the technical field of artificial intelligence, in particular to the field of high-precision intelligent identification realized through deep learning by combining a large model and a small model, and can be used for scenes such as intelligent meter reading and electric power inspection. According to the specific implementation scheme, the method comprises the steps of obtaining original dial plate image data of an instrument collected in a target area; processing the original image data of the dial plate through a pre-trained lightweight image analysis model to obtain the equipment type and the current reading of the instrument; inputting the equipment type and the current reading into a pre-trained large-scale anomaly detection model, analyzing the current reading by combining the large-scale anomaly detection model with the equipment type, and generating an anomaly probability value and an anomaly type of the instrument; and in response to determining that the abnormal probability value exceeds the preset risk threshold, outputting a detection result containing the abnormal probability value and the abnormal type. According to the scheme, the meter reading accuracy and efficiency can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and particularly to the fields of achieving high-precision intelligent recognition through deep learning by combining large models and small models, which can be used in scenarios such as intelligent meter reading and power inspection. Specifically, it relates to an intelligent meter reading method, device, and equipment based on collaborative processing of large and small models. Background Art

[0002] The power grid energy field involves a lot of meter reading work, and the reading methods of different monitoring instruments are also different. Therefore, in actual work, in order to ensure the normal operation of the entire power grid energy system, manual meter reading is mostly used. However, the manual meter reading method cannot meet the high-frequency data collection requirements, and cross-regional meter reading requires coordination of personnel scheduling, resulting in low meter reading efficiency and unable to guarantee the meter reading accuracy. Summary of the Invention

[0003] The present disclosure provides an intelligent meter reading method, device, and equipment based on collaborative processing of large and small models.

[0004] According to a first aspect of the present disclosure, there is provided an intelligent meter reading method based on collaborative processing of large and small models, including: obtaining original image data of the dial of an instrument collected in a target area; processing the original image data of the dial through a pre-trained lightweight image analysis model to obtain the device type and current reading of the instrument; inputting the device type and current reading into a pre-trained large-scale anomaly detection model, so that the large-scale anomaly detection model analyzes the current reading in combination with the device type to generate an anomaly probability value and an anomaly type of the instrument; and in response to determining that the anomaly probability value exceeds a preset risk threshold, outputting a detection result including the anomaly probability value and the anomaly type.

[0005] According to a second aspect of the present disclosure, there is provided an intelligent meter reading device based on collaborative processing of large and small models, including: an image acquisition module for obtaining original image data of the dial of an instrument collected in a target area; a reading determination module for processing the original image data of the dial through a pre-trained lightweight image analysis model to obtain the device type and current reading of the instrument; a probability generation module for inputting the device type and current reading into a pre-trained large-scale anomaly detection model, so that the large-scale anomaly detection model analyzes the current reading in combination with the device type to generate an anomaly probability value and an anomaly type of the instrument; and an anomaly processing module for, in response to determining that the anomaly probability value exceeds a preset risk threshold, outputting a detection result including the anomaly probability value and the anomaly type.

[0006] According to a third aspect of the present disclosure, there is provided an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute any of the methods in the embodiments of the present disclosure.

[0007] According to a fourth aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to execute any of the methods in the embodiments of the present disclosure.

[0008] According to a fifth aspect of the present disclosure, there is provided a computer program product, including a computer program, which when executed by a processor, implements any of the methods in the embodiments of the present disclosure.

[0009] Adopting the solution of the present disclosure can improve the accuracy and efficiency of meter reading.

[0010] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become easily understandable through the following description. Description of the Drawings

[0011] The drawings are used to better understand the solution and do not constitute a limitation to the present disclosure. Among them:

[0012] Figure 1 is a schematic flowchart of an intelligent meter reading method based on the collaborative processing of large and small models according to an embodiment of the present disclosure;

[0013] Figure 2 is a schematic architecture diagram of an intelligent meter reading system based on the collaborative processing of large and small models according to an embodiment of the present disclosure;

[0014] Figure 3 is a schematic structural diagram of an intelligent meter reading device based on the collaborative processing of large and small models according to an embodiment of the present disclosure;

[0015] Figure 4 is a schematic scenario diagram of an intelligent meter reading method based on the collaborative processing of large and small models according to an embodiment of the present disclosure;

[0016] Figure 5 is a structural diagram of an electronic device used to implement the intelligent meter reading method based on the collaborative processing of large and small models according to an embodiment of the present disclosure. Detailed Embodiments

[0017] The following describes exemplary embodiments of the present disclosure with reference to the accompanying drawings. Various details of the embodiments of the present disclosure are included to facilitate understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0018] As used herein, the term "and / or" merely describes an association relationship between associated objects and indicates that three relationships can exist. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. As used herein, the term "at least one" means any one of a plurality or any combination of at least two of a plurality. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set composed of A, B, and C. As used herein, the terms "first" and "second" are used to refer to multiple similar technical terms and distinguish them, and do not mean to limit the order or limit to only two. For example, the first feature and the second feature refer to two categories / two features. The first feature can be one or more, and the second feature can also be one or more.

[0019] In addition, to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. Those skilled in the art should understand that the present disclosure can be implemented without some specific details. In some instances, methods, means, elements, and circuits well-known to those skilled in the art are not described in detail to highlight the gist of the present disclosure.

[0020] Before introducing the technical solutions of the embodiments of the present disclosure, further explanations are made on the technical terms that may be used in the present disclosure:

[0021] Instruments: They are divided into electrical instruments and non-electrical instruments. Among them, electrical instruments include voltmeters, ammeters, watt-hour meters, power meters, etc.; non-electrical instruments include thermometers, barometers, flow meters, etc.

[0022] In the related art, generally, specialized personnel read and record specialized instruments on-site. In case of abnormal situations, the on-site situation is remotely reported to remote experts, waiting for the experts' solutions, which may miss the best processing time and increase potential safety hazards. In addition, manual meter reading is limited by physical operations and cannot meet the requirements of high-frequency data acquisition (such as minute-level monitoring). Cross-regional meter reading requires coordinating personnel scheduling, which is time-consuming and laborious. After manual recording, it needs to be entered into the system a second time, resulting in a delay in data update and affecting real-time decision-making. Moreover, paper records are prone to loss or damage, and data integrity cannot be guaranteed.

[0023] To at least partially address one or more of the above problems and other potential problems, the present disclosure proposes an intelligent meter reading method based on collaborative processing of large and small models, which can improve the accuracy and efficiency of meter reading.

[0024] An embodiment of the present disclosure provides an intelligent meter reading method based on collaborative processing of large and small models. Figure 1 FIG. is a schematic flowchart of the intelligent meter reading method based on collaborative processing of large and small models according to an embodiment of the present disclosure. This intelligent meter reading method based on collaborative processing of large and small models can be applied to an intelligent meter reading device based on collaborative processing of large and small models. The intelligent meter reading device based on collaborative processing of large and small models is located in an electronic device. The electronic device includes, but is not limited to, fixed devices and / or mobile devices. For example, fixed devices include, but are not limited to, servers, and the server can be a cloud server or a general server. For example, mobile devices include, but are not limited to, intelligent meter reading devices, and the intelligent meter reading device can be a mobile phone, a tablet computer, etc. In some possible implementation manners, this intelligent meter reading method based on collaborative processing of large and small models can also be implemented by a processor calling computer-readable instructions stored in a memory.

[0025] As Figure 1 shown, this intelligent meter reading method based on collaborative processing of large and small models includes:

[0026] S101. Obtain the original image data of the meter dial collected in the target area;

[0027] S102. Process the original image data of the meter dial through a pre-trained lightweight image analysis model to obtain the device type and the current reading of the meter;

[0028] S103. Input the device type and the current reading into a pre-trained large-scale anomaly detection model, so that the large-scale anomaly detection model analyzes the current reading in combination with the device type to generate an anomaly probability value and an anomaly type of the meter;

[0029] S104. In response to determining that the anomaly probability value exceeds a preset risk threshold, output a detection result including the anomaly probability value and the anomaly type.

[0030] In an embodiment of the present disclosure, the target area is the area where the meters to be monitored are located. Here, the target area includes, but is not limited to, areas such as power plants, substations, or power consumption terminals.

[0031] In the embodiments of the present disclosure, the lightweight image analysis model is a model used to process the original image data of the dial to obtain the device type and the current reading of the instrument. The performance indicators of the lightweight image analysis model meet at least one of the following requirements: the number of parameters is less than the first threshold, the computing power is less than the second threshold, and the latency is less than the third threshold. For example, the number of parameters is less than 1 megabyte; the computing power < 100 million floating-point operations per second, and the latency < 50 milliseconds.

[0032] In some embodiments, the lightweight image analysis model can adopt any one of the following: the MobileNetV3 network based on the lightweight deep learning architecture; the ResNet18-D network compressed by the knowledge distillation technology; the network optimized by the Neural Architecture Search (NAS). The above is only an exemplary illustration and does not limit all possible architectures of the lightweight image analysis model, but exhaustive listing is not done here.

[0033] In some embodiments, the lightweight image analysis model can be trained in the following ways: using the dial image dataset for transfer learning, where the training data includes electromagnetic interference noise samples in the target area under high-voltage environment; adopting the Focal Loss function to solve the problem of unbalanced sample numbers of mechanical / electronic meters; quantizing and training through the edge device simulator. It should be noted that the present disclosure does not limit the training methods and training approaches of the lightweight image analysis model.

[0034] In some embodiments, the large-scale anomaly detection model is any one of the following architectures: the multi-modal time series analysis model based on the Transformer; the hybrid architecture of the graph neural network and the temporal convolutional network (Graph Neural Network - Temporal Convolutional Network, GNN-TCN); the cascaded model integrating the Extreme Gradient Boosting algorithm (XGBoost) and the deep residual network. The above is only an exemplary illustration and does not limit all possible architectures of the large-scale anomaly detection model, but exhaustive listing is not done here.

[0035] In some embodiments, the training data of the large-scale anomaly detection model can include: cross-regional power load data covering the power consumption patterns of different regions; the instrument failure case library containing various types of anomaly samples such as meter tampering, communication interruption, and mechanical aging; extreme scenario data including voltage dips and zero-value drifts under severe weather conditions. It should be noted that the present disclosure does not limit the training methods and training approaches of the large-scale anomaly detection model.

[0036] In the embodiments of the present disclosure, the device types of the meters are at least divided into electrical meters and non-electrical meters. Among them, the electrical meters are specifically divided into voltmeters, ammeters, watt-hour meters, power meters, etc.; the non-electrical meters are specifically divided into thermometers, barometers, flow meters, etc.

[0037] In the embodiments of the present disclosure, the current reading refers to the value read from the meter. Exemplarily, for a mechanical pointer electric meter, the current reading refers to the physical quantity value after polar coordinate mapping of the scale value pointed by the pointer tip, including the unit; for an electronic digital electric meter, the current reading refers to the continuous digital sequence recognized in the digital display area, including the integer part and the decimal part.

[0038] In the embodiments of the present disclosure, the abnormal types are divided into data abnormal types and device failure types.

[0039] In the embodiments of the present disclosure, the data abnormal types include at least one of the following: Sudden increase type anomaly: The deviation between the current reading and the average value of the previous N cycles exceeds a certain percentage, and the continuous duration ≥ 2 acquisition cycles; Sudden decrease type anomaly: The current reading is lower than the theoretical minimum value (theoretical minimum value = historical baseline value × 0.7); Zeroing type anomaly: The readings are zero for 3 consecutive times and do not match the device operating state; Fluctuation type anomaly: The standard deviation within the sliding window exceeds the threshold (threshold = 3 times the historical standard deviation). The above is only an exemplary description and does not limit all possible situations of the data abnormal types, and only exhaustive listing is not done here.

[0040] In the embodiments of the present disclosure, the device failure types include at least one of the following: Mechanical needle jamming failure: The angle of the pointer of the mechanical electric meter remains unchanged for 3 consecutive cycles and is contradictory to the change trend of the load current; Communication interruption failure: The smart meter does not return a data packet within the preset timeout window; Tampering trace failure; Among them, the tampering trace failure refers to detecting one of the following features in the dial image: The seal point of the meter case is damaged; The reflection of the dial glass is abnormal; Battery low power failure: The digital display of the electronic electric meter flashes at a frequency greater than a certain threshold. The above is only an exemplary description and does not limit all possible situations of the device failure types, and only exhaustive listing is not done here.

[0041] In this way, by using the lightweight image analysis model to identify and process the reading of the original dial image data, the speed and accuracy of meter reading can be improved. By analyzing the current reading with the large-scale anomaly detection model combined with the device type, the anomaly probability value of the meter can be obtained, which can improve the calculation accuracy of the anomaly probability value. When the anomaly probability value exceeds the preset risk threshold, the detection result including the abnormal type is output, which helps to improve the emergency response speed. By the collaborative division of labor of the lightweight image analysis model and the large-scale anomaly detection model to process the original dial image data, the data transmission volume can be reduced, the call times of the large-scale anomaly detection model can be reduced, and the consumption of computing resources can be reduced, thereby improving the accuracy and efficiency of meter reading in the target area.

[0042] In some embodiments, the original dial image data is processed by a pre-trained lightweight image analysis model to obtain the device type and the current reading of the instrument, including: extracting the appearance features of the instrument from the original dial image data; matching the preset classification rules based on the appearance features of the instrument to determine the device type of the instrument; and calling the algorithm corresponding to the device type to extract the current reading of the instrument.

[0043] In the embodiments of the present disclosure, the appearance features of the instrument include but are not limited to the dial shape, the layout of scale marks, and the position of the digital display area.

[0044] In some embodiments, the appearance features of the instrument in the original dial image data can be extracted based on region detection technology. In some other embodiments, the appearance features of the instrument in the original dial image data can be extracted based on image segmentation and boundary detection technology. The above is only an exemplary illustration and does not limit all possible implementation manners of extracting the appearance features of the instrument in the original dial image data, and only exhaustive listing is not done here.

[0045] In some embodiments, the preset classification rules can be matched based on the appearance features of the instrument by using the method of traditional image processing + rule engine, or the method of feature engineering + classifier, or the method of end-to-end classification, or the method of template matching + similarity calculation to determine the device type of the instrument. The above is only an exemplary illustration and does not limit all possible implementation manners of matching the preset classification rules based on the appearance features of the instrument to determine the device type of the instrument, and only exhaustive listing is not done here.

[0046] In some embodiments, the algorithm corresponding to the device type can be called based on the preset correspondence between the device type and the algorithm. In some embodiments, the algorithm matching the device type can be screened online from the database. The above is only an exemplary illustration and does not limit all possible implementation manners of calling the algorithm corresponding to the device type, and only exhaustive listing is not done here.

[0047] In this way, by extracting the appearance features of the instrument from the original dial image data, it helps to improve the success rate of image feature extraction; by matching the preset classification rules based on the appearance features of the instrument to determine the device type of the instrument, it can improve the recognition speed and recognition accuracy of the device type; by calling the algorithm corresponding to the device type to extract the current reading of the instrument, it can improve the speed and accuracy of the reading.

[0048] In some embodiments, the extraction of the appearance features of the instrument and the determination of the device type are realized through a preset target detection algorithm, which specifically includes: extracting multi-scale dial features of the instrument based on the cross-stage feature fusion network of the preset target detection algorithm; outputting the device type label and the dial position coordinates through the adaptive classification head of the preset target detection algorithm; wherein, the instrument appearance features include multi-scale dial features and dial position coordinates.

[0049] In the embodiments of the present disclosure, the device type label is divided according to the functions and types of the instrument. Its functions include but are not limited to voltmeters, ammeters, watt-hour meters, power meters, etc., thermometers, barometers, and flow meters. Its types include but are not limited to mechanical, electronic, or intelligent types.

[0050] In the embodiments of the present disclosure, the dial position coordinates are the normalized bounding box coordinates of the dial area in the image, defined as a quadruple (x min , y min , x max , y max ). X min , y min : The proportional coordinates of the upper left corner of the bounding box relative to the width and height of the image (ranging from 0 to 1); X max , y max : The proportional coordinates of the lower right corner of the bounding box relative to the width and height of the image (ranging from 0 to 1). x min = the abscissa of the upper left corner of the image width box / the image width; y min = the ordinate of the upper left corner of the image height box / the image width. X max = the abscissa of the upper right corner of the image width box / the image width; y max = the ordinate of the upper right corner of the image height box / the image width.

[0051] In some embodiments, extracting multi-scale dial features based on the cross-stage feature fusion network includes: the input image generates multi-layer feature maps through the backbone network (such as scales of 80×80, 40×40, 20×20, 10×10); using the Path Aggregation Network (PAN) structure for top-down + bottom-up feature fusion. Exemplarily, the feature fusion formula is: Where C onv1×1 represents the channel dimension transformation of the 1×1 convolution kernel, is element-wise addition, F up is the upsampled high-level feature map, F down is the downsampled low-level feature map, and F out is the fused feature.

[0052] In some embodiments, the device type label and the dial position coordinates are output through an adaptive classification head, including: performing global average pooling on the input feature map; mapping it to a class probability vector through a fully connected layer; when the maximum probability value ≥ a certain value, output the corresponding type label, otherwise mark it as an unknown type. Predict the position offset based on a preset anchor box, calculate based on the anchor box offset regression, and filter through non-maximum suppression to obtain the dial position coordinates (x min , y min , x max , y max ).

[0053] In this way, the multi-scale dial features of the instrument are extracted by the cross-stage feature fusion network based on the preset object detection algorithm, which can reduce the dial positioning error; the device type label and the dial position coordinates are output by the adaptive classification head of the preset object detection algorithm, which can reduce the device type misclassification rate and reduce the post-processing complexity, thereby improving the accuracy and efficiency of the meter reading form.

[0054] In some embodiments, the algorithm corresponding to the device type is called to extract the current reading of the instrument, including: if the device type is an electronic digital instrument, the preset digital recognition algorithm is used to extract the current reading. Specifically, it includes: using the lightweight text detection network of the preset digital recognition algorithm to locate the digital display area in the electronic digital instrument; correcting the digital tilt angle through the direction classifier of the preset digital recognition algorithm; outputting a digital sequence based on the ultra-lightweight text recognition network of the preset digital recognition algorithm; and optimizing the digital sequence in combination with the context semantics to obtain the current reading.

[0055] In some embodiments, the implementation process of using the lightweight text detection network to locate the digital display area can use MobileNetV3 as the backbone network, combined with a Differentiable Binarization (DB) text detection head to output a pixel-level text area probability map P, perform binarization processing on P with a threshold T to obtain a binary mask B; perform connected component analysis on B to output a set of candidate bounding boxes for the digital display area. The threshold T can be a value dynamically calculated by the algorithm or fixed to a certain value.

[0056] In some embodiments, the implementation process of the direction classifier correcting the digital tilt angle includes: First, design the classifier. For example, 4-direction classification (0°, 90°, 180°, 270°). Then, perform geometric correction. When tilting is detected, rotate the image through an affine transformation matrix. Finally, perform random rotation augmentation to improve robustness. In this way, the discrete angle classification can reduce the model complexity; through data augmentation + affine transformation, the digital integrity can be retained.

[0057] In some embodiments, the implementation process of the ultra-lightweight text recognition network outputting an optimized digital sequence is as follows: First, the ultra-lightweight text recognition network structure is a Convolutional Recurrent Neural Network (CRNN) architecture, specifically: three-layer convolutional neural network (each layer contains a 3×3 convolutional kernel, with a stride of 1) + bidirectional long short-term memory network (the hidden layer dimension is 64); the output layer uses Connectionist Temporal Classification Loss (abbreviated as CTC loss), and the character set contains 12 categories including "0-9" and ".". Second, semantic optimization. A context finite state machine (FSM) is constructed based on the instrument reading rules, and the illegal sequence is corrected through the Viterbi algorithm (e.g., "12.345" → "123.45"). Finally, quantization deployment. For example, the model is compressed to 0.6 megabytes, and its power consumption when running on the processor is <0.5 watts. In this way, through the joint optimization of CRNN+FSM, the accuracy of the digital sequence can be improved. By forcing grammar constraints through the state machine, the misrecognition rate of the decimal point can be reduced. Through depthwise separable convolution + parameter quantization, the model size can be reduced.

[0058] In this way, by extracting the current reading through the preset digital recognition algorithm, the accuracy of the reading of the electronic digital instrument can be improved.

[0059] In some embodiments, calling the algorithm corresponding to the device type to extract the current reading of the instrument includes: If the device type is a mechanical pointer instrument, call the preset image segmentation algorithm to implement dial pointer detection and reading calculation. Specifically, it includes: locating the main axis center line and the tip area of the pointer based on the preset image segmentation algorithm's rotated bounding box, and outputting the rotation angle and spatial coordinates of the pointer; extracting the angle between the tip of the pointer and the zero scale line; correcting the perspective distortion of the dial plane through perspective transformation, and calculating the actual tilt angle of the dial; correcting the angle according to the actual tilt angle, and obtaining the corrected angle; and outputting the current reading according to the mapping relationship between the corrected angle and the instrument range.

[0060] In some embodiments, the implementation process of locating the main axis and tip area of the pointer by the rotated bounding box is as follows: First, segment the network architecture, and then optimize the tip positioning. Among them, an improved instance segmentation network (PP-YOLOE-SEG) model is used, and its Mask branch outputs the parameters of the rotated rectangle bounding box of the pointer: (x c , yc, w, h, θ), θ ∈ [-π / 2, π / 2]; where (x c , y c ) is the center point, (w, h) is the width and height, and θ is the rotation angle.

[0061] Among them, optimizing the tip positioning includes: extracting the pointer skeleton line based on the segmentation mask, and detecting the tip coordinates through the curvature extreme points; dynamically adjusting the angle of the rotation box. For example, calculating the pixel coverage at intervals of 5°, and taking the maximum value as the final angle. In this way, the rotation box parameterizes and directly outputs the angle, improving the positioning accuracy; the skeleton line assists the tip positioning, improving the anti-occlusion ability; optimizing the single-stage segmentation network architecture, improving the inference speed.

[0062] In some embodiments, extracting the angle between the pointer tip and the zero scale line is divided into two steps: scale line detection and angle calculation.

[0063] Among them, scale line detection includes: searching for peaks in the polar coordinate space (p, θ), and limiting the θ search range to ±10° (to avoid false detection); setting the scale line length threshold to 1 / 5 of the dial radius.

[0064] Among them, angle calculation includes: establishing a polar coordinate system with the center of the dial as the origin; calculating the pointer vector and the zero scale vector angle

[0065] In some embodiments, perspective transformation for correcting the tilt of the dial includes: obtaining the outer contour of the dial through semantic segmentation, fitting the ellipse equation Ax 2 +Bxy+Cy 2 +Dx+Ey+F = 0, calculating the ellipse inclination angle φ = 1 / 2 arctan(B / (A - C)); selecting the 4 corner points of the circumscribed rectangle of the dial as the source points {P i}, defining the target rectangle corner points {Q i}, and solving the homography matrix H: Q i = H·P i . In this way, partial occlusion can be resisted through ellipse fitting; accurate mapping can be achieved through the homography matrix.

[0066] In some embodiments, angle correction and reading mapping include: according to the actual tilt angle φ after perspective transformation, correcting the pointer angle The compensation amount threshold is set to ±5° (to prevent overcorrection). Among them, represents the corrected pointer angle. Reading the dial range R,

[0067] In this way, by calling the preset image segmentation algorithm to implement dial pointer detection and reading calculation, even in extreme scenarios of tilt + partial occlusion, the overall reading error can be reduced and the reading accuracy can be improved.

[0068] In some embodiments, the large-scale anomaly detection model generates an anomaly probability value based on the deviation probability of the current reading from the dynamic threshold interval, including: constructing a dynamic threshold interval based on the historical power consumption data, real-time environmental parameters, and device type of the meter; calculating the deviation probability of the current reading from the dynamic threshold interval through time-series residual analysis to generate an anomaly probability value.

[0069] In some embodiments, features can be constructed first based on multi-dimensional data such as historical power consumption data, real-time environmental parameters, and device type. Then, a Temporal Fusion Transformer (TFT) is used to determine the threshold interval. Specifically, the static feature encoder in the TFT processes invariant features such as device type, the temporal feature encoder extracts periodic patterns (day / week / month), and the predictor outputs the dynamic threshold interval.

[0070] In some embodiments, the anomaly probability calculation formula is: P anomaly = 1 - exp(-|r t | / σ), where σ represents the residual scaling coefficient optimized by historical data; among them, P anomaly represents the anomaly probability, |r t | represents the absolute value of the residual of the current value y t .

[0071] In some embodiments, the anomaly probability value P anomaly is dynamically weighted to obtain the final anomaly probability value P final . Among them, P tinal = 0.7P anomaly + b, where the current value y t does not belong to the dynamic threshold interval [T low , T high , T low represents the minimum value of the dynamic threshold interval, T high represents the maximum value of the dynamic threshold interval, and b represents the dynamic bias term related to environmental parameters and device type.

[0072] In this way, generating the anomaly probability value according to the dynamic threshold interval can improve the accuracy of anomaly detection.

[0073] In some embodiments, the large-scale anomaly detection model combines the associated knowledge analysis of the retrieval enhancement generation module to generate an anomaly probability value, including: inputting the current reading into the Retrieval-Augmented Generation (RAG) module to obtain the knowledge fragments related to the current reading output by the retrieval enhancement generation module; inputting the current reading and the knowledge fragments into the large-scale anomaly detection model to obtain the anomaly probability value and anomaly type label output by the large-scale anomaly detection model.

[0074] In some embodiments, a knowledge base is constructed based on multiple data sources such as industry standard documents, historical fault case libraries, and equipment technical manuals. Exemplarily, multi-dimensional document vectors are generated, and an inverted index is established. Each knowledge fragment is labeled with its source, timeliness, and applicable equipment type.

[0075] In some embodiments, the current reading and its context features (such as sudden change amount, environmental parameters) are input into the retrieval-enhanced generation module, and the retrieval-enhanced generation module generates a query vector based on the current reading and its context features; a hybrid retrieval strategy of keyword retrieval and vector retrieval is adopted, and a summary is generated for the retrieval results to obtain a structured text within a preset number of characters as a knowledge fragment.

[0076] In some embodiments, the large-scale anomaly detection model converts the current reading into numerical features and the knowledge fragment into knowledge features; the numerical features and knowledge features are fused to obtain fused features; the fused features are respectively input into the anomaly detection head and the type classification head to obtain an anomaly probability value and an anomaly type label.

[0077] In some embodiments, in order to further improve the accuracy of the anomaly probability, the anomaly probability value can be corrected according to the confidence of the knowledge fragment. Exemplarily, P final = P anomaly ·(1 + λS k ), where λ is the knowledge confidence weighting coefficient, and its value is dynamically adjusted according to the semantic similarity between the retrieval result and the current reading; S k is the highest similarity of the retrieval result.

[0078] In some embodiments, the large-scale anomaly detection model also generates an analysis report based on the anomaly probability value and the anomaly type label. The report includes: the original text of the industry standard clauses matched; the handling records of similar historical cases; a summary of the basis for anomaly determination.

[0079] In this way, through the collaborative architecture of knowledge retrieval - multi-modal decision-making, when outputting the anomaly report, associating specific knowledge fragments can improve the interpretability of the detection results; since the knowledge base can be dynamically updated, the accuracy of the detection results can be further improved.

[0080] In some embodiments, the method further includes: generating an equipment maintenance work order and pushing it to the operation and maintenance terminal. The equipment maintenance work order includes the anomaly type and a priority identifier.

[0081] In the embodiments of the present disclosure, the anomaly types are divided into data anomaly types (such as sudden increase, sudden decrease, zeroing) and equipment failure types (such as pointer jamming on the dial, communication module damage).

[0082] In the embodiments of the present disclosure, the priority identifier can be divided into multiple levels of warnings such as first-level warning and second-level warning. Among them, the smaller the warning level, the more urgent the time for maintenance.

[0083] In this way, by associating the exception type with the priority identifier, the maintenance response speed for high-risk faults can be improved; based on the equipment maintenance work order, the maintenance resources are dynamically allocated (for example, a two-person maintenance team is automatically triggered for a first-level alarm), which can reduce the operation and maintenance costs while ensuring the operation and maintenance speed and quality.

[0084] In some embodiments, the method further includes: marking the abnormal instrument as a high-risk node, and increasing the image acquisition frequency of the high-risk node to a preset frequency threshold in the subsequent meter reading process, where the high-risk node is an instrument whose detection result includes an abnormal type.

[0085] In the embodiments of the present disclosure, the strategy for increasing the image acquisition frequency can be set or adjusted according to the instrument type and the target area. Exemplarily, the strategy for increasing the image acquisition frequency includes: if the high-risk node is an instrument in a power plant or a distribution plant, the acquisition frequency is increased to once per minute; if the high-risk node is an instrument in an end-user power network, the acquisition frequency is increased to once every ten minutes.

[0086] In this way, by differentially adjusting the acquisition frequency of abnormal instruments according to the instrument type and the target area, the scene adaptability can be improved.

[0087] In some embodiments, obtaining the original dial image data of the instruments collected in the target area includes: collecting instrument images through edge computing devices deployed in the target area, and preprocessing the original dial image data to obtain the original dial image data.

[0088] In the embodiments of the present disclosure, the preprocessing at least includes one of the following: denoising using a non-local means filtering algorithm; generating a correction matrix based on a checkerboard calibration method to perform distortion correction; adjusting the illumination uniformity through histogram equalization.

[0089] In the embodiments of the present disclosure, the edge computing device can be configured with a supercapacitor module to maintain the image acquisition and local storage capabilities for at least 8 hours after the power grid is powered off.

[0090] In this way, through edge preprocessing such as denoising and distortion correction, while reducing the amount of image transmission, the quality of the transmitted image data is improved, not only reducing the overall processing delay, but also increasing the success rate of extracting key features of the preprocessed image, thereby improving the accuracy of meter reading.

[0091] Figure 2 Shows the architecture diagram of an intelligent meter reading system based on the collaborative processing of large and small models, as Figure 2As shown, the architecture is mainly composed of four modules: a dial recognition module, a pointer recognition module, an Artificial Intelligence (AI) reading module, and a RAG module.

[0092] Among them, the dial recognition module is used to construct a comprehensive and detailed feature model by learning a large number of dial samples of different types using a preset object detection algorithm. These types cover traditional circular pointer dials, square digital dials, and various industrial dials with special designs. When faced with an image containing a dial, the preset object detection algorithm can quickly analyze the texture, shape, color, and other feature information in the image and compare it with the learned model to accurately identify the specific type of the dial. Once the type of the dial is recognized, it will bring a lot of prior knowledge for subsequent work. For example, if it is a traditional circular pointer dial, it can be known that its scales are usually evenly distributed around the circumference, and the reading method is to determine according to the scale position pointed by the pointer; if it is a digital dial, it can be clear that its reading is directly obtained through the displayed numbers. These prior knowledge play a key guiding role in the smooth operation of the subsequent AI reading module, which can greatly improve the accuracy and efficiency of reading.

[0093] In terms of dial position detection, the preset object detection algorithm uses target localization technology to accurately determine the position coordinates of the dial in a complex image scene. Whether the dial is in the center of the image, at the edge corner, or partially occluded, the algorithm can accurately outline the boundary of the dial by analyzing the key information such as the target contour and edge in the image. Detecting the position of the dial in the image is of great significance. According to these position information, the area where the dial is located can be accurately cropped out from the image, and then the cropped area can be magnified. In this way, the detailed information of the dial can be presented more clearly, providing higher-resolution image data for the pointer recognition module, which can significantly improve the accuracy of pointer recognition, thus enhancing the accuracy and reliability of the processing of dial-related information.

[0094] The pointer recognition module is used to accurately detect the pointer area in the image using a preset image segmentation algorithm according to the characteristics of the circular pointer dial, such as the shape and movement trajectory of the pointer, so as to obtain the information indicated by the pointer. Once it is determined to be a digital dial, considering that its information presentation form is direct numbers and there is no need to perform detection operations similar to the pointer area, the relevant detection process can be directly skipped to improve the overall image processing efficiency.

[0095] In practical applications, in the field of complex image recognition and processing, for images containing dials, a precise and efficient processing process needs to be constructed. First, the type of the dial needs to be recognized and judged. When the dial in the image is recognized as a circular pointer dial, a preset image segmentation algorithm will be enabled. The preset image segmentation algorithm has powerful feature extraction and segmentation capabilities. It will deeply analyze many features of the circular pointer dial. For example, the unique slender shape of the pointer, the various movement trajectories it presents at different times due to the change of the indicated time, and the differences in color and texture between the pointer and other elements of the dial. Based on these features, the algorithm can accurately outline the contour of the pointer area in the image, achieve precise detection of the pointer area, and then accurately obtain key information such as the time indicated by the pointer.

[0096] When the dial recognition module determines that it is a circular pointer dial, the AI reading module will start a precise calculation process. It will calculate the angle formed by the pointer and the zero position of the dial through complex trigonometric operations based on the pointer position information provided by the pointer recognition module. Then, combined with the prior knowledge of the dial scale, such as the scale range of one circle of the dial and the value represented by each scale, according to the angle ratio relationship, it will accurately calculate the specific reading indicated by the circular pointer dial.

[0097] If the dial recognition module determines that it is a digital dial, the reading process is relatively simple at this time. The AI reading module will directly use the preset digital recognition algorithm. This algorithm has efficient character recognition capabilities. It can quickly recognize the numbers displayed on the digital dial. Its working principle is to extract and classify the features of the digital image and compare them with the pre-trained digital templates, so as to accurately read the numbers on the dial and complete the reading task.

[0098] The RAG module is used to comprehensively search in the professional knowledge document library according to the characteristics and key information of the data. This document library covers a large amount of content such as industry standards, historical data cases, and professional research results. The RAG module screens out the knowledge content closely related to the current data through precise matching. Subsequently, the retrieved knowledge and the original data are input into the large model together. The large model uses its powerful data analysis and pattern recognition capabilities to deeply judge the data and identify whether the data is abnormal. If the large model determines that the data is normal, the system will immediately return a normal signal, indicating that the data is within the normal fluctuation range and the business process can continue to run smoothly.

[0099] Once the large model detects anomalies in the data, the system will immediately activate the alarm mechanism. It will send alarm messages to the duty room in the first instance. The alarm content will detail key information such as the source of the abnormal data, the occurrence time, and the abnormal manifestations, so that the duty personnel can quickly understand the situation. Meanwhile, the system will output a solution to the abnormal situation based on the previously retrieved professional knowledge and the built-in abnormal handling logic. This solution may include data repair suggestions, emergency operation guides, notifications to relevant responsible persons, etc., providing strong support for quickly handling the anomaly and minimizing the impact of the abnormal data on the business.

[0100] Through the above architecture, the inspector takes pictures of the dial and uploads them. After dial recognition and pointer recognition, the AI reading module integrates and processes the information, and finally obtains accurate values. Then, the automatically read values are stored in the database, and at the same time, through the RAG module, automatic analysis and judgment are performed based on knowledge in the professional field. If it is determined to be an abnormal situation, the best solution for handling the abnormal situation will be immediately returned, improving the efficiency of meter reading and inspection.

[0101] It should be understood that Figure 2 the schematic diagrams shown are merely exemplary and not restrictive, and they are extensible. Those skilled in the art can make various obvious changes and / or substitutions based on Figure 2 the examples, and the resulting technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0102] Embodiments of the present disclosure provide an intelligent meter reading device based on collaborative processing of large and small models. As Figure 3 shown, the device may include: an image acquisition module 301 for acquiring the original image data of the dial of the instrument collected in the target area; a reading determination module 302 for processing the original image data of the dial through a pre-trained lightweight image analysis model to obtain the device type and the current reading of the instrument; a probability generation module 303 for inputting the device type and the current reading into a pre-trained large-scale anomaly detection model, so that the large-scale anomaly detection model analyzes the current reading in combination with the device type to generate an anomaly probability value and an anomaly type of the instrument; and an anomaly processing module 304 for outputting a detection result including the anomaly probability value and the anomaly type in response to determining that the anomaly probability value exceeds a preset risk threshold.

[0103] In some embodiments, the reading determination module 302 includes: an extraction sub-module for: extracting the appearance features of the instrument in the original image data of the dial; determining the device type of the instrument based on matching the appearance features of the instrument with a preset classification rule; and a call sub-module for calling an algorithm corresponding to the device type to extract the current reading of the instrument.

[0104] In some embodiments, an extraction sub-module is configured to: extract the appearance features of the instrument and determine the device type through a preset target detection algorithm. Specifically, the extraction sub-module is configured to: extract the multi-scale dial features of the instrument based on the cross-stage feature fusion network of the preset target detection algorithm; output the device type label and the dial position coordinates through the adaptive classification head of the preset target detection algorithm; wherein, the instrument appearance features include multi-scale dial features and dial position coordinates.

[0105] In some embodiments, a call sub-module is configured to: if the device type is an electronic digital instrument, extract the current reading of the electronic digital instrument through a preset digital recognition algorithm, specifically including: locating the digital display area in the electronic digital instrument by using the lightweight text detection network of the preset digital recognition algorithm; correcting the digital tilt angle through the direction classifier of the preset digital recognition algorithm; outputting a digital sequence based on the ultra-lightweight text recognition network of the preset digital recognition algorithm; optimizing the digital sequence in combination with the context semantics to obtain the current reading.

[0106] In some embodiments, a call sub-module is configured to: if the device type is a mechanical pointer instrument, call a preset image segmentation algorithm to extract the current reading of the mechanical pointer instrument, specifically including: locating the main axis center line and the tip area of the pointer based on the rotated bounding box of the preset image segmentation algorithm, and outputting the rotation angle and spatial coordinates of the pointer; extracting the angle between the pointer tip and the zero scale line; correcting the perspective distortion of the dial plane through perspective transformation, and calculating the actual tilt angle of the dial; correcting the angle according to the actual tilt angle, and obtaining the corrected angle; obtaining the current reading according to the mapping relationship between the corrected angle and the instrument range.

[0107] In some embodiments, the probability generation module 303 includes a first generation sub-module, which is configured to: construct a dynamic threshold interval based on the historical power consumption data, real-time environmental parameters and device type of the instrument; calculate the deviation probability between the current reading and the dynamic threshold interval through time series residual analysis; generate an abnormal probability value and an abnormal type according to the deviation probability.

[0108] In some embodiments, the probability generation module 303 includes a second generation sub-module, which is configured to: input the current reading into the retrieval enhanced generation module to obtain the knowledge fragment related to the device type output by the retrieval enhanced generation module; input the current reading and the knowledge fragment into a large-scale anomaly detection model to obtain the abnormal probability value and the abnormal type output by the large-scale anomaly detection model.

[0109] In some embodiments, the device further includes:

[0110] A work order generation module ( Figure 3 not shown in the figure) is configured to generate a device maintenance work order and push it to the operation and maintenance terminal, and the device maintenance work order includes the abnormal type and the priority identifier.

[0111] In some embodiments, the device further comprises:

[0112] a frequency adjustment module ( Figure 3 not shown in the figure) for determining the location of the abnormal meter as a high-risk node and increasing the image acquisition frequency of the high-risk node to a preset frequency in the subsequent meter reading process, where the abnormal meter is a meter whose detection result includes an abnormal type.

[0113] In some embodiments, the image acquisition module 301 is specifically configured to: collect meter images through edge computing devices deployed in the target area and preprocess the meter images to obtain original dial image data.

[0114] For the specific functions and examples of the modules and sub-modules of the device in the embodiments of the present disclosure, reference may be made to the relevant descriptions of the corresponding steps in the above method embodiments, which will not be elaborated here.

[0115] The intelligent meter reading device based on the collaborative processing of the large and small models in the embodiments of the present disclosure can improve the accuracy and efficiency of meter reading.

[0116] The embodiments of the present disclosure provide a schematic diagram of a scenario of an intelligent meter reading method based on the collaborative processing of the large and small models, as Figure 4 shown.

[0117] As described above, the intelligent meter reading method provided by the embodiments of the present disclosure is applied to an electronic device. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers.

[0118] Specifically, the electronic device can specifically perform the following operations:

[0119] Obtain the original dial image data of the meters collected in the target area;

[0120] Process the original dial image data through a pre-trained lightweight image analysis model to obtain the device type and the current reading of the meter;

[0121] Input the device type and the current reading into a pre-trained large-scale anomaly detection model, so that the large-scale anomaly detection model analyzes the current reading in combination with the device type to generate an anomaly probability value and an anomaly type of the meter;

[0122] In response to determining that the anomaly probability value exceeds a preset risk threshold, output a detection result including the anomaly probability value and the anomaly type.

[0123] It should be understood that Figure 4The scene diagram shown is merely illustrative and not restrictive. Those skilled in the art can make various obvious changes and / or substitutions based on Figure 4 the examples, and the resulting technical solutions still fall within the scope of the disclosure of the embodiments of the present disclosure.

[0124] In the technical solutions of the present disclosure, the acquisition, storage, and application of the user's personal information involved all comply with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0125] According to the embodiments of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0126] Figure 5 FIG. shows a schematic block diagram of an exemplary electronic device 500 that can be used to implement the embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as, for example, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, for example, a personal digital assistant, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely exemplary and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0127] As Figure 5 shown, the device 500 includes a computing unit 501, which can execute various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 502 or a computer program loaded from a storage unit 508 into a random access memory (RAM) 503. In the RAM 503, various programs and data required for the operation of the device 500 can also be stored. The computing unit 501, the ROM 502, and the RAM 503 are connected to each other via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0128] A plurality of components in the device 500 are connected to the I / O interface 505, including: an input unit 506, such as a keyboard, a mouse, etc.; an output unit 507, such as various types of displays, speakers, etc.; a storage unit 508, such as a disk, an optical disc, etc.; and a communication unit 509, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 509 allows the device 500 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0129] The computing unit 501 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 executes the various methods and processes described above, such as the intelligent meter reading method based on the cooperation of large and small models. For example, in some embodiments, the intelligent meter reading method based on the cooperation of large and small models can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as the storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 500 via the ROM 502 and / or the communication unit 509. When the computer program is loaded into the RAM 503 and executed by the computing unit 501, one or more steps of the intelligent meter reading method based on the cooperation of large and small models described above can be executed. Alternatively, in other embodiments, the computing unit 501 can be configured to execute the intelligent meter reading method based on the cooperation of large and small models by any other suitable means (e.g., by means of firmware).

[0130] The various embodiments of the systems and techniques described above in this specification can be implemented in digital electronic circuitry, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application-specific standard products (ASSPs), system on chip (SOC) systems, complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include: being implemented in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which may be a special-purpose or general-purpose programmable processor that receives data and instructions from, and transmits data and instructions to, a storage system, at least one input device, and at least one output device.

[0131] The program code for implementing the methods of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program codes, when executed by the processor or controller, cause the functions / operations specified in the flowchart and / or block diagram to be implemented. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

[0132] In the context of this disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0133] In order to provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, speech input, or tactile input).

[0134] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0135] A computer system may include a client and a server. The client and the server are generally far from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server incorporating a blockchain.

[0136] It should be understood that various forms of processes shown above can be used, steps can be reordered, added, or deleted. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and no limitation is imposed herein.

[0137] The above specific embodiments do not constitute a limitation on the protection scope of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the principles of this disclosure shall be included within the protection scope of this disclosure.

Claims

1. An intelligent meter reading method based on collaborative processing of a small model and a large model, comprising: Obtaining the original dial image data of meters collected in a target area; Processing the original dial image data through a pre-trained lightweight image analysis model to obtain the device type and the current reading of the meter; Inputting the device type and the current reading into a pre-trained large-scale anomaly detection model, so that the large-scale anomaly detection model analyzes the current reading in combination with the device type to generate an anomaly probability value and an anomaly type of the meter; In response to determining that the anomaly probability value exceeds a preset risk threshold, outputting a detection result including the anomaly probability value and the anomaly type.

2. The method according to claim 1, wherein The processing the original dial image data through a pre-trained lightweight image analysis model to obtain the device type and the current reading of the meter comprises: Extracting the appearance features of the meter from the original dial image data; Based on the appearance features of the meter, matching a preset classification rule to determine the device type of the meter; Invoking an algorithm corresponding to the device type to extract the current reading of the meter.

3. The method according to claim 2, wherein, The method further comprises: Implementing the extraction of the appearance features of the meter and the determination of the device type through a preset object detection algorithm, specifically comprising: Extracting multi-scale dial features of the meter based on the cross-stage feature fusion network of the preset object detection algorithm; Outputting a device type label and dial position coordinates through the adaptive classification head of the preset object detection algorithm; wherein, the appearance features of the meter include the multi-scale dial features and the dial position coordinates.

4. The method according to claim 2, wherein, The invoking an algorithm corresponding to the device type to extract the current reading of the meter comprises: If the device type is an electronic digital meter, extracting the current reading of the electronic digital meter through a preset digital recognition algorithm, specifically comprising: Locating the digital display area in the electronic digital meter by using the lightweight text detection network of the preset digital recognition algorithm; Correcting the digital tilt angle through the direction classifier of the preset digital recognition algorithm; Outputting a digital sequence based on the ultra-lightweight text recognition network of the preset digital recognition algorithm; Optimizing the digital sequence in combination with the context semantics to obtain the current reading.

5. The method according to claim 2, wherein The invoking an algorithm corresponding to the device type to extract the current reading of the meter comprises: If the device type is a mechanical pointer meter, invoking a preset image segmentation algorithm to extract the current reading of the mechanical pointer meter, specifically comprising: Locating the main axis center line and the tip area of the pointer based on the rotated bounding box of the preset image segmentation algorithm, and outputting the rotation angle and spatial coordinates of the pointer; Extracting the angle between the pointer tip and the zero scale line; Correcting the perspective distortion of the dial plane through perspective transformation and calculating the actual tilt angle of the dial; Correcting the angle according to the actual tilt angle to obtain a corrected angle; Obtaining the current reading according to the mapping relationship between the corrected angle and the meter range.

6. The method according to claim 1, wherein, The analysis of the current reading by the large-scale anomaly detection model in combination with the device type to generate the anomaly probability value and anomaly type of the meter includes: Construct a dynamic threshold interval based on the historical power consumption data, real-time environmental parameters of the meter, and the device type; Calculate the deviation probability between the current reading and the dynamic threshold interval through time series residual analysis; Generate the anomaly probability value and the anomaly type according to the deviation probability.

7. The method according to claim 1, wherein The analysis of the current reading by the large-scale anomaly detection model in combination with the device type to generate the anomaly probability value and anomaly type of the meter includes: Input the current reading into the retrieval-enhanced generation module to obtain the knowledge fragments related to the device type output by the retrieval-enhanced generation module; Input the current reading and the knowledge fragments into the large-scale anomaly detection model to obtain the anomaly probability value and the anomaly type output by the large-scale anomaly detection model.

8. The method according to claim 1, wherein, The method further includes: Generate a device maintenance work order and push it to the operation and maintenance terminal, where the device maintenance work order includes the anomaly type and a priority identifier.

9. The method according to claim 1, wherein The method further includes: Determine the location of the abnormal meter as a high-risk node, and increase the image acquisition frequency of the high-risk node to a preset frequency in the subsequent meter reading process, where the abnormal meter is a meter whose detection result includes the anomaly type.

10. The method according to claim 1, wherein The obtaining of the original dial image data of the meter collected in the target area includes: Collect meter images through edge computing devices deployed in the target area, and preprocess the meter images to obtain the original dial image data.

11. An intelligent meter reading device based on collaborative processing of large and small models, including: An image acquisition module for obtaining the original dial image data of the meter collected in the target area; A reading determination module for processing the original dial image data through a pre-trained lightweight image analysis model to obtain the device type and the current reading of the meter; A probability generation module for inputting the device type and the current reading into a pre-trained large-scale anomaly detection model, so that the large-scale anomaly detection model analyzes the current reading in combination with the device type to generate the anomaly probability value and anomaly type of the meter; An anomaly processing module for outputting a detection result including the anomaly probability value and the anomaly type in response to determining that the anomaly probability value exceeds a preset risk threshold.

12. An electronic device, including: At least one processor; And A memory communicatively connected to at least one processor; wherein, The memory stores instructions executable by at least one processor, and the instructions are executed by at least one processor so that at least one processor can execute the method according to any one of claims 1-10.

13. A non-transitory computer-readable storage medium storing computer instructions, wherein, Computer instructions are used to cause a computer to execute the method according to any one of claims 1-10.

14. A computer program product, including a computer program stored on a storage medium, where the computer program, when executed by a processor, implements the method according to any one of claims 1-10.

Images