Intelligent water meter recognition system based on three-layer edge computing architecture

Abstract

The invention discloses an intelligent water meter recognition system based on a three-layer edge computing architecture, which has the following beneficial effects: (1) based on the three-layer edge computing architecture, the intelligent water meter recognition system gives full play to the advantages of edge computing, the load of a central server is reduced, and the service life of the central server is prolonged; and the service life of the water meter image data acquisition equipment is prolonged. (2) YOLO-METER is improved on the basis of YOLO11 and is used for water meter reading recognition, a FastC3k2 module and an MRFBlock are introduced into the YOLO-METER, and the FastC3k2 module can improve the image feature extraction performance and enhance the low-contrast feature extraction capability of a water meter image by reducing the number of parameters of a model and FLOPs; and the MRFBlock obviously expands the receptive field of the model, enhances the capability of autonomously selecting features, improves multi-scale fusion, and can optimize the feature selection of a digital roller region and improve the positioning capability of a water meter reading region. According to the invention, the calculation cost is reduced while the identification precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of water meter reading recognition, and specifically to an intelligent water meter recognition system based on a three-layer edge computing architecture. Background Art

[0002] Water meters are essential instruments for measuring water supply in various sectors, including residential, hotel, and restaurant settings, as well as a wide range of industrial applications. One type of mechanical water meter uses a digital wheel display. The current meter reading is obtained by reading the numbers on the wheel.

[0003] Water resources are vital to human development and survival and play a key role in global sustainable development. Efficient water resource management not only helps conserve and reduce consumption, but also alleviates environmental burdens, restores ecological balance, and enhances public well-being. For water utilities, optimizing resource management can reduce operating costs, improve pipeline efficiency, minimize leakage, and better address growing urban water demand. An accurate and efficient water meter reading and recognition system is crucial to achieving intelligent water resource management.

[0004] However, traditional manual meter reading methods are still used in many water resource management systems. These methods are inefficient and time-consuming, making them incapable of meeting the accuracy and real-time requirements of modern urban monitoring. To address these issues, advanced water meters, such as electromagnetic, flow, and ultrasonic, have emerged in recent years. These devices feature automatic data collection and wireless transmission capabilities; however, their high cost and the need for infrastructure replacement have limited their widespread adoption. Therefore, upgrading existing water meters with intelligent technologies, without requiring hardware replacement, has become a key research focus.

[0005] In recent years, with the rapid development of the Internet of Things (IoT) and deep learning technologies, meter reading systems based on image recognition have significantly improved. These systems can be roughly divided into two types based on the algorithm deployment location: the first is local recognition, in which the algorithm is deployed on the terminal device, which is responsible for image acquisition, reading recognition, and uploading the results to the server. However, this method is limited by the terminal's computing resources, resulting in low recognition accuracy and a limited device lifespan. The second is cloud-based recognition, in which the algorithm is deployed in the cloud, with the terminal only acquiring images and uploading them to the cloud server for processing. This method can easily overload the cloud server as the number of tasks increases, affecting response speed and system stability, and leading to higher data transmission bandwidth requirements and potential security risks. Furthermore, the recognition accuracy of water meter readings based on image recognition using existing deep learning technologies still has room for improvement. Summary of the Invention

[0006] (1) Technical problems solved

[0007] In response to the shortcomings of the existing technology, the present invention provides an intelligent water meter identification system based on a three-layer edge computing architecture, which can solve at least one of the above technical problems.

[0008] (2) Technical solution

[0009] To solve the above technical problems, the present invention provides the following technical solutions: an intelligent water meter identification system based on a three-layer edge computing architecture, comprising: an Internet of Things layer, an edge layer, and a cloud layer; the Internet of Things layer is used to capture water meter images and transmit them to the edge layer; the edge layer uses the YOLO-METER model to identify water meter readings on water meter images, wherein the YOLO-METER model is an improvement based on the YOLO11 model; the cloud layer is used to receive water meter reading identification data from the edge layer.

[0010] Preferably, the Internet of Things layer is also used to compress the water meter image for further transmission to the edge layer.

[0011] Preferably, the Internet of Things layer is installed on the water meter, and the Internet of Things layer includes a power supply, a camera module and a data transmission module. The camera module is used to capture the water meter image, and the data transmission module is used to transmit the water meter image to the edge layer.

[0012] Preferably, the Internet of Things layer also includes an LED light and a control chip. The LED light is used to provide lighting for the camera module to shoot, and the control chip is used to time the camera module, the LED light, and the data transmission module to start working.

[0013] Preferably, the YOLO-METER model introduces the FastC3k2 module into its backbone network, wherein the FastC3k2 module is an improvement based on FasterNet.

[0014] Preferably, the FastC3k2 module includes a convolution block, a splitting block, an intermediate calculation block and a splicing block.

[0015] Preferably, when the model parameter C3k2 is True, the intermediate calculation block includes multiple FastC3k blocks, wherein FastC3k includes a convolution block, a FasterNet block, and a splicing block. Accordingly, the output Y of the FastC3k2 module is shown in the following equations (1)-(4):

[0016] X1,X2=Split(Conv1(X)) (1)

[0017] Z (1) =FaskC3k(X1) (2)

[0018] Z (i) =FaskC3k(Z (i-1) ) (3)

[0019] Y=Concat(X2,Z (1) ,Z (2) ,...,Z (n) ) (4)

[0020] Among them, X1 and X2 are the two branches that the input X is divided into after convolution.

[0021] Preferably, when the model parameter C3k2 is False, the intermediate computation block includes multiple FasterNet blocks.

[0022] Preferably, the YOLO-METER model introduces MRFBlock in its neck network, wherein MRFBlock includes VSSBlock and PGBlock.

[0023] Preferably, the cloud layer is also used to analyze and process the water meter reading identification data.

[0024] (3) Beneficial effects

[0025] Compared with the prior art, the present invention provides an intelligent water meter identification system based on a three-layer edge computing architecture, which has the following beneficial effects: (1) The intelligent water meter identification system of the present invention is based on a three-layer edge computing architecture, including an Internet of Things layer, an edge layer, and a cloud layer. The intelligent water meter identification system gives full play to the advantages of edge computing, while reducing the load of the central server, it also extends the service life of the water meter image data acquisition equipment; (2) Based on YOLO11, the present invention improves and proposes YOLO-METER for water meter reading identification. YOLO-METER enhances the feature extraction capability of low-contrast areas and optimizes the feature selection of the digital roller area, so that the present invention can improve the accuracy of water meter reading identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a first structural diagram of an intelligent water meter identification system based on a three-layer edge computing architecture according to the present invention;

[0027] Figure 2 This is a second structural diagram of an intelligent water meter identification system based on a three-layer edge computing architecture according to the present invention;

[0028] Figure 3 This is a structural diagram of the YOLO-METER model of the present invention;

[0029] Figure 4 A comparison diagram of the category distribution in the training set, validation set, and test set of the present invention;

[0030] Figure 5 This is the FasterNet structure diagram of the present invention;

[0031] Figure 6 It is the structural diagram of FastC3k of the present invention;

[0032] Figure 7 This is the SS2D structure diagram of the present invention;

[0033] Figure 8 This is a comparison diagram of water meter reading recognition between YOLO-METER and YOLO11 n of the present invention. In this figure, (a) represents the original image, (b) represents the recognition result of YOLO11 n, and (c) represents the recognition result of YOLO-METER;

[0034] Figure 9 This is a visual comparison of the sensory wildness of different models. In this figure, (a) represents the wilderness of YOLO11, and (b) represents the wilderness of YOLO11+MRFBlock.

[0035] Figure 10 Recognition detection heat maps for different models. In this figure, (a) represents the original image, (b) represents the heat map of YOLO11 n, and (c) represents the heat map of YOLO11+FastC3k2;

[0036] Figure 11 This is the result graph of water meter readings predicted by Holt's linear trend method. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] The present invention provides an intelligent water meter identification system based on a three-layer edge computing architecture, including an Internet of Things layer (IoT layer) 11, an edge layer (Edge layer) 12 and a cloud layer (Cloud layer) 13.

[0039] The Internet of Things layer 11 is used to capture water meter images and transmit them to the edge layer 12.

[0040] Preferably, the Internet of Things layer is also used to compress the water meter image for further transmission to the edge layer. The Internet of Things layer compresses the image to reduce the volume and transmits the compressed grayscale image to the edge layer to reduce power consumption.

[0041] Specifically, the Internet of Things layer is installed on the water meter and set close to the water meter dial. The Internet of Things layer includes a power supply, a camera module and a data transmission module. The camera module is used to capture water meter images, and the data transmission module is used to transmit water meter images to the edge layer.

[0042] In addition, preferably, the Internet of Things layer also includes LED lights and a control chip. The LED lights are used to provide lighting for the camera module to shoot, and the control chip is used to time the camera module, LED lights, and data transmission module to start working.

[0043] As you can understand, the IoT layer is responsible for collecting water meter image data, periodically capturing and sending water meter images to the edge layer. The aforementioned control chip can be the HC32F460, and the data transmission module can be a 4G data transmission module (SLM332). The HC32F460 chip is based on the ARMv7-M architecture and features low power consumption. In addition to its run mode, the HC32F460 also has a sleep mode. In sleep mode, only the control chip's timer (RTC) remains powered, while other modules, such as the camera module, are powered off. This ensures very low power consumption, making the HC32F460 ideal for the short-term operations required for water meter data collection. The IoT layer can be powered by a 3.6V lithium battery to power the entire device. To extend battery life, the HC32F460 chip features a timer that outputs a high-level signal at specified intervals. This signal activates the MOSFET, powering other modules, such as the camera module. After powering on, the LEDs on the circuit illuminate, the camera module captures an image of the water meter, and the HC32F460 chip enters run mode to process data, thereby timing the camera module, LEDs, and data transmission module to start up. To ensure image quality, LEDs on both sides of the camera module illuminate during capture, providing sufficient brightness for a clear image. After capturing the image, the HC32F460 chip performs compression operations, converting the water meter image to grayscale and converting it to base64 format for transmission to the edge layer.

[0044] The edge layer uses the YOLO-METER model to identify water meter readings from water meter images. The YOLO-METER model is an improvement on the YOLO11 model, specifically the YOLO11n model. The edge layer is responsible for processing water meter images received from the IoT layer to identify and analyze water meter readings. The YOLO-METER recognition algorithm and its associated database are deployed on the edge layer. The edge layer can utilize a motherboard such as the NCT3568, a high-performance, low-power development board based on the RK3568. The edge layer can be equipped with a 4G module and communicate via HTTP.

[0045] The present invention can be applied to the recognition of readings from digit-wheel water meters. Specifically, the dataset used to train the YOLO-METER model consists of 9,658 water meter images collected through the aforementioned IoT layer. This dataset covers images captured under normal shooting conditions as well as complex environments such as blur, low light, and overexposure. This dataset comprehensively encompasses various scenarios likely to be encountered in practical applications. These images are divided into 5,800 training (train) images, 1,919 validation (val) images, and 1,859 test (test) images. Because digit-wheel water meters have progressive increments, intermediate states exist between readings. To this end, the present invention defines 20 categories: 10 integer categories (0, 1, 2, ..., 9) and 10 decimal categories (0.5, 1.5, 2.5, ..., 9.5) to better address the recognition of water meter readings in intermediate states.

[0046] In order to further analyze the composition of the above dataset, the present invention visualizes the category distribution in the training set, validation set, and test set, as shown in the following figure: Figure 4 As shown, since the number of samples in category 0 is much larger than that in other categories, the present invention performs a logarithmic transformation on the number of samples. This transformation improves the visualization effect of smaller sample categories and provides a clearer comparison overview of each category. Figure 4 As shown, the proportions of each category in different subsets are relatively consistent, which indicates that the dataset maintains a balanced distribution, which is crucial for ensuring fair model training and evaluation.

[0047] The YOLO-METER of the present invention is an improved version of YOLO11, such as Figure 3As shown in the figure, the YOLO-METER model of the present invention introduces the FastC3k2 module into its backbone network. Specifically, the YOLO-METER model of the present invention replaces the C3k2 module in the YOLO11 model backbone network with the FastC3k2 module, wherein the FastC3k2 module is an improvement based on FasterNet (i.e., FasterNet block or FasterNet Block). The FastC3k2 module can improve the image feature extraction performance and enhance the extraction of low-contrast features of water meter images by reducing the number of parameters and FLOPs of the model; the FastC3k2 module includes a convolution block (Conv), a splitting block (Spl it), an intermediate calculation block and a splicing block (Concat).

[0048] The FasterNet mentioned above was proposed by Chen et al. The FastC3k2 module of the present invention is an enhanced version of FasterNet. In order to design a fast neural network, Chen et al. analyzed the relationship between floating point numbers and memory access and proposed a new partial convolution (PConv), based on which they further proposed FasterNet. Figure 5 As shown in the figure, the processing process of FasterNet is as follows: FasterNet first performs partial convolution (PConv), then 1x1 convolution, BN (normalization) layer, activation function uses ReLU, followed by another layer of 1x1 convolution layer, and finally the calculation result is added to the input through the residual connection.

[0049] Specifically, if Figure 3 As shown, for the FastC3k2 module of the YOLO-METER model of the present invention, when the model parameter C3k2 is True (i.e. Figure 3 When C3k2=True as shown in a in , the intermediate calculation block of the FastC3k2 module includes multiple FastC3k blocks, wherein the FastC3k includes a convolution block, the above-mentioned FasterNet block and a splicing block. Accordingly, the output Y of the FastC3k2 module is shown in the following equations (1)-(4):

[0050] X1,X2=Split(Conv1(X)) (1)

[0051] Z (1) =FaskC3k(X1) (2)

[0052] Z (i) =FaskC3k(Z (i-1) ) (3)

[0053] Y=Concat(X2,Z (1),Z (2) ,...,Z (n) ) (4)

[0054] Among them, X1 and X2 are the two branches that the input X is divided into after convolution. Split is the above-mentioned split block, and Concat is the above-mentioned splicing block. It can be understood that for the FastC3k2 module, its input X is first processed by convolution, and then split into two branches X1 and X2 through Split. One of the branches X1 is processed by FastC3k, that is, multiple FastC3k are responsible for calculating the intermediate output, and finally spliced ​​with the other branch X2 to obtain the output Y. In addition, if Figure 6 As shown in the figure, FastC3k includes a convolution block, the aforementioned FasterNet block, and a splicing block. It can be understood that the processing process of FastC3k is as follows: the input is convolved to obtain O, then the FasterNet block operation is performed, the calculation result is spliced ​​with O, and finally another convolution calculation is performed to obtain the output.

[0055] In addition, when the above model parameter C3k2 is False (i.e. Figure 3 In the example, when C3k2=False as shown in b in the figure, the intermediate calculation block of the FastC3k2 module includes multiple FasterNet blocks. That is to say, the FastC3k2 module is correspondingly responsible for calculating the intermediate output by multiple FasterNet blocks at this time, and the other processing procedures of the FastC3k2 module are the same as when the above model parameter C3k2 is True, which is not described here.

[0056] In addition, preferably, the convolution operation is used for the feature fusion component in the original YOLO model such as YOLO11, which results in a limited receptive field and weak adaptive feature selection capability. In order to solve these limitations, Figure 3 As shown, the YOLO-METER model of the present invention introduces a new module MRFBlock in its neck network. Specifically, the YOLO-METER model of the present invention replaces the C3k2 module in the neck network of the YOLO11 model with MRFBlock, wherein, as shown in FIG. Figure 3As shown in c, MRFBlock includes two branches: VSSBlock (visual state space block) and PGBlock. The MRFBlock is based on Mamba (a large language model LLM architecture) and a gating mechanism. VSSBlock takes advantage of the advantages of Mamba, while PGBlock takes advantage of the gating mechanism of the convolutional neural network (CNN). MRFBlock ultimately fuses the features of the two branches. The MRFBlock significantly expands the receptive field of the model, enhances its ability to independently select features, and improves multi-scale fusion. By introducing MRFBlock, the present invention can optimize the feature selection of the digital scroll wheel area and improve the positioning capability of the water meter reading area, thereby improving the detection performance and accuracy of the model of the present invention for water meter readings. For MRFBlock, its input passes through VSSBlock and PGBlock respectively, and the final output is shown in the following formula (5):

[0057] Z l =VSSBlock(Z l-1 )+PGBlock(Z l-1 ) (5)

[0058] Specifically, for the above-mentioned VSSBlock, the 2D selective scanning mechanism (SS2D) is the core component of VSSBlock. SS2D aims to construct a hidden state space for cross-modal feature fusion. The structure of SS2D is as follows: Figure 7 The VSSBlock process is as follows: VSSBlock first passes through a linear layer, then through a depthwise convolution (DWConv) and SILU activation function; then, it passes through SS2D for feature extraction, layer normalization through layer norm, and finally passes through an additional linear layer for processing.

[0059] Specifically, for the above-mentioned PGBlock, PGBlock efficiently filters features through CNN and valves, and combines the above-mentioned PConv to perform feature fusion without introducing excessive computational costs; the structure of PGBlock is as follows Figure 3 As shown in d. PGBlock divides the input into two branches through Split. Further, PGBlock performs the calculation in parallel through two PConv as shown in the following equations (6)-(7):

[0060]

[0061] Furthermore, PGBlock uses ReLU as the activation function to control the information flow, then merges it with the global features through dot products, and uses the global features to mix the channel information; finally, it fuses it with the original input through residual connections. In this way, PGBlock can capture more global features while only slightly increasing the computational cost. The output feature X obtained by PGBlock is l It is defined as follows (8):

[0062]

[0063] Among them, Φ represents the activation function (ReLU), and ⊙ represents the dot product (i.e., element-wise production).

[0064] The PGBlock of the present invention captures more global features with minimal computational overhead. PGBlock combines a gating mechanism with convolution to preserve spatial information and capture detailed image features. Unlike traditional MLPs (Multi-Layer Perceptrons), the PGBlock of the present invention transmits global features at the pixel level, enriching contextual expression and improving feature modeling capabilities, thereby better ensuring the water meter reading recognition accuracy of the YOLO-METER model of the present invention.

[0065] The cloud layer of the present invention is used to receive the water meter reading identification data from the edge layer. In addition, preferably, the cloud layer is also used to analyze and process the water meter reading identification data, including water consumption tracking, water leakage detection, water consumption forecasting and other analysis and processing, so as to better provide user services. The cloud layer can also realize remote management and monitoring of equipment. The cloud layer aggregates all water meter reading identification data transmitted by the above-mentioned edge layer and serves as the central hub for user interaction and system management; users can access real-time water usage data through the software system, view consumption forecasts, analyze historical water usage patterns, and receive alerts for potential water leaks. In addition, the cloud layer can also support remote device management, allowing technicians to monitor the operating status of all edge layer devices through the central server, which includes real-time diagnosis, performance tracking and fault detection to ensure efficient operation of the system. Through the integration of data processing and user interaction, the cloud layer improves intelligent water resource management and system reliability.

[0066] Preferably, the cloud layer can use the Holt linear trend method to perform water consumption forecasting analysis based on the water meter reading identification data, that is, predict the water meter reading. In this analysis, the daily water consumption data of some water meter dials in Panyu District, Guangdong Province were collected and counted, as shown in Table 1 below. The water meters are classified according to the pipe diameter, mainly including different diameters such as 15, 20, 25 and 80, among which 15 and 20 diameters are the most common in residential water use; the average daily water consumption in different months is analyzed to provide users with a valuable benchmark reference; water consumption forecasting analysis relies on the analysis of historical consumption data. The cloud layer uses the Holt linear trend method to predict water consumption, and the smoothing coefficient is set to 0.8 for horizontal prediction and 0.2 for trend prediction. The data of the past 21 days are used for training to predict the current water consumption. The prediction results are as follows: Figure 11 shown.

[0067] Table 1 Average daily water consumption by diameter size

[0068]

[0069] The following is an experimental analysis of the performance of the YOLO-METER model of the present invention. The experimental indicators used precision, recall, mAP, FLOPs, and number of parameters. The hyperparameters of the experiment are set as follows: batch size = 32, training rounds (epochs) = 300, weight decay (weight decay) = 0.0005, initial learning rate (lr0) = 0.01, and final learning rate (lrf) = 0.0001. The variables of the ablation experiment are mainly concentrated on FastC3k2 and MRFBlock of the present invention.

[0070] Comparative experiments: The comparison models are divided into single-stage models and two-stage models. The two-stage model includes Fast R-CNN, which uses ResNet50 as the backbone network; the single-stage models include RTDETR, YOLOV5x, YOLOV6, YOLOV7, YOLOV8, YOLOV9, and YOLOV10. The relevant experimental results are shown in Table 2 below:

[0071] Table 2 Comparison of experimental results of different network models Table 1

[0072]

[0073] As can be seen from Table 2 above, the mAP of YOLO-METER of the present invention reaches 94.9, which is significantly higher than other models. Compared with RTDETR, YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv9 and YOLOv10, YOLO-METER's mAP increases by 7.6%, 2.7%, 1.9%, 3.6%, 2.6%, 1.6% and 2.4% respectively. In addition, YOLO-METER has fewer parameters than other models, 25.93M fewer than RTDETR, 0.13M fewer than YOLOv5, 1.73M fewer than YOLOv6, 34.03M fewer than YOLOv7, and 0.53M fewer than YOLOv8. Although YOLOv9 and YOLOv10 have fewer parameters than YOLO-METER, their accuracy is not as good as YOLO-METER and the floating-point operations exceed YOLO-METER. Overall, the YOLO-METER of the present invention achieves an effective balance between accuracy and speed, making it very suitable for deployment on edge devices.

[0074] Ablation experiment: In order to evaluate the impact of activation function on the performance of PGBlock, the present invention tested ReLU, Leaky ReLU, PReLU and GELU in various network settings. As shown in Table 3, the choice of activation function significantly affects the accuracy and efficiency. ReLU performs best in mAP and recall, proving to be very effective in fine-grained feature extraction and powerful representation learning. In addition, as shown in Table 4, replacing C3K2 with FastC3k2 of the present invention can reduce 11.9% of parameters and 13.4% of FLOPs, while improving mAP by 0.7%. Although MRFBlock adds some computational overhead, it improves mAP by 1.2%. Combining the two modules of FastC3k2 and MRFBlock further reduces parameters and FLOPs, while improving mAP by 2.4%. As Figure 8 As shown in Figure 3, YOLO-METER can accurately detect the bounding box and effectively avoid the interference of surrounding digits even in blurry images.

[0075] Table 3 Comparison of experimental results of different network models Table 2

[0076]

[0077] Table 4 Comparison of experimental results of different network models Table 3

[0078]

[0079] In addition, if Figure 9As shown in the figure, the present invention draws a visual comparison of the sensory wildness of different models. It can be seen that MRFBlock significantly improves the sensory wildness of the model. In addition, in order to illustrate the ability of FastC3k2 of the present invention in detecting character box features, Grad-CAM++ is used to generate a heat map, as shown in the figure. Figure 10 As shown in the figure, these heat maps visualize the areas that have the greatest impact on the model prediction, where the red area indicates stronger attention and the blue area indicates weaker attention. It can be seen from the figure that FastC3k2 helps the model focus on the key details of the image more effectively, highlighting its advantages in feature extraction.

[0080] It can be understood that the present invention designs a fully automated multifunctional smart water meter identification system, which consists of three layers: the Internet of Things layer, the edge layer and the cloud layer, thereby ensuring the system's efficiency and flexibility in data collection, processing and storage; wherein, the Internet of Things layer is responsible for water meter image capture, compression and upload; the edge layer deploys the YOLO-METER algorithm, which improves the accuracy of water meter image recognition through the FastC3k2 and MRFBlock modules, significantly enhances feature extraction and detection accuracy, and reduces the parameters of the model; the cloud layer aggregates data from the edge layer, enabling users and administrators to effectively monitor water use and manage the system. The above-mentioned layered architecture of the present invention ensures efficiency, reliability and optimized power usage. The present invention has the following beneficial effects: (1) The smart water meter identification system of the present invention is based on a three-layer edge computing architecture. The smart water meter identification system fully utilizes the advantages of edge computing, while reducing the load on the central server, it also extends the service life of the equipment used for water meter image data collection. (2) Based on YOLO11, the present invention proposes YOLO-METER for water meter reading recognition. The YOLO-METER of the present invention introduces the FastC3k2 module and MRFBlock: wherein, the FastC3k2 module can improve the image feature extraction performance and enhance the extraction capability of low-contrast features of water meter images by reducing the number of parameters and FLOPs of the model; and MRFBlock significantly expands the receptive field of the model, enhances its ability to independently select features, and improves multi-scale fusion. By introducing MRFBlock, the present invention can optimize the feature selection of the digital roller area and improve the positioning capability of the water meter reading area; the YOLO-METER of the present invention can improve the accuracy of water meter reading recognition by introducing the FastC3k2 module and MRFBlock. Experimental results show that the YOLO-METER of the present invention improves mAP50 by 2.4%, reduces the number of parameters by 6%, and reduces the computational cost while improving the recognition accuracy; the intelligent water meter recognition system of the present invention helps to realize the efficient monitoring of water resources, thereby improving operational efficiency and promoting the development of intelligent resource management.

[0081] It should be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not preclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0082] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A smart water meter identification system based on a three-layer edge computing architecture, characterized in that: include: IoT layer, edge layer, and cloud layer; The Internet of Things layer is used to capture water meter images and transmit them to the edge layer; The edge layer uses the YOLO-METER model to identify the water meter reading on the water meter image, wherein the YOLO-METER model is an improvement based on the YOLO11 model; The cloud layer is used to receive the water meter reading identification data of the edge layer.

2. The smart water meter identification system based on the three-layer edge computing architecture according to claim 1 is characterized by: The Internet of Things layer is also used to compress the water meter image for further transmission to the edge layer.

3. The smart water meter identification system based on the three-layer edge computing architecture according to claim 1 is characterized in that: The Internet of Things layer is installed on the water meter. The Internet of Things layer includes a power supply, a camera module and a data transmission module. The camera module is used to capture the water meter image, and the data transmission module is used to transmit the water meter image to the edge layer.

4. The smart water meter identification system based on the three-layer edge computing architecture according to claim 3 is characterized by: The Internet of Things layer also includes an LED light and a control chip. The LED light is used to provide lighting for the camera module to shoot, and the control chip is used to time the camera module, the LED light, and the data transmission module to start working.

5. The smart water meter identification system based on the three-layer edge computing architecture according to claim 1 is characterized in that: The YOLO-METER model introduces the FastC3k2 module into its backbone network, wherein the FastC3k2 module is an improvement based on FasterNet.

6. The smart water meter identification system based on the three-layer edge computing architecture according to claim 5 is characterized by: The FastC3k2 module includes a convolution block, a splitting block, an intermediate calculation block and a splicing block.

7. The smart water meter identification system based on the three-layer edge computing architecture according to claim 6 is characterized in that: When the model parameter C3k2 is True, the intermediate computation block includes multiple FastC3k blocks, wherein the FastC3k blocks include convolution blocks, FasterNet blocks, and concatenation blocks. Accordingly, the output Y of the FastC3k2 module is as shown in the following equations (1)-(4): X1,X2=Split(Conv1(X)) (1) WITH (1) =FaskC3k(X1) (2) WITH (i) =FaskC3k(Z (i-1) ) (3) Y=Concat(X2,Z (1) ,Z (2) ,...,Z (n) ) (4) Among them, X1 and X2 are the two branches that the input X is divided into after convolution.

8. The smart water meter identification system based on the three-layer edge computing architecture according to claim 7 is characterized in that: When the model parameter C3k2 is False, the intermediate calculation block includes multiple FasterNet blocks.

9. The smart water meter identification system based on a three-layer edge computing architecture according to claim 1 is characterized in that: The YOLO-METER model introduces MRFBlock in its neck network, wherein the MRFBlock includes VSSBlock and PGBlock.

10. The smart water meter identification system based on a three-layer edge computing architecture according to claim 1 is characterized in that: The cloud layer is also used to analyze and process water meter reading identification data.

Images