Method and system for verifying meter reading accuracy of water meter readers based on image recognition

Abstract

The present invention relates to the field of image processing technology, and discloses a method and system for verifying the meter reading accuracy of a water meter reader based on image recognition. The method comprises: performing multimodal data acquisition on a water meter to obtain an original data set; correcting a water meter reading photo to obtain an optimized water meter image; performing feature extraction and classification on the optimized water meter image to obtain an image recognition reading, a digital confidence score, and an abnormal mark; performing data standardization processing on manual meter reading data to obtain standardized meter reading data; performing time-frequency analysis and feature extraction on water flow sound data and pipeline vibration data to obtain an instantaneous flow estimation value and a historical average flow at the time of acquisition; performing multi-dimensional cross-validation on image recognition readings, standardized meter reading data, instantaneous flow estimation values, and historical average flow to obtain meter reading verification data and an abnormal situation report. The present invention improves the efficiency and accuracy of existing meter reading verification methods and effectively reduces human errors.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method and system for verifying the meter reading accuracy of a water meter reader based on image recognition. Background Art

[0002] Water meter reading is an important part of the daily operation of water companies. The traditional manual meter reading method relies on the experience and responsibility of the meter reader, which is prone to reading errors, recording errors or data tampering. With the development of technology, some water companies have begun to adopt automatic meter reading systems, such as remote meter reading or smart water meters. These systems automatically read water meter data through electronic devices and transmit the data to the central system for processing, which has improved the efficiency and accuracy of meter reading to a certain extent.

[0003] However, even automatic meter reading systems have some limitations. For example, electronic devices may not work properly due to failure or battery exhaustion, resulting in missing or inaccurate data. In addition, automatic systems can usually only collect a single type of data (such as water meter readings), lack a comprehensive perception of the meter reading environment and water flow conditions, and have difficulty effectively identifying and verifying abnormal situations. At the same time, existing meter reading verification methods often rely on simple data comparison or statistical analysis, and cannot fully utilize multi-source data for in-depth cross-validation, which can easily miss some potential problems, resulting in low accuracy of existing meter reading verification methods. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and system for verifying the meter reading accuracy of a water meter reader based on image recognition, which is used to improve the efficiency and accuracy of existing meter reading verification methods.

[0005] The present invention provides a method for verifying the meter reading accuracy of a water meter reader based on image recognition, comprising: performing multimodal data collection on a water meter to obtain an original data set, wherein the original data set includes a water meter reading photo, manual meter reading data, water flow sound data, and pipeline vibration data; performing noise reduction, light balance, and geometric correction on the water meter reading photo to obtain an optimized water meter image; performing feature extraction and classification on the optimized water meter image through a deep convolutional neural network to obtain an image recognition reading, a digital confidence score, and an abnormal mark; performing data standardization processing on the manual meter reading data to obtain standardized meter reading data; performing data standardization processing on the water flow sound data to obtain standardized meter reading data; The water flow sound data and the pipeline vibration data are subjected to time-frequency analysis and feature extraction to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection. The step of performing time-frequency analysis and feature extraction on the water flow sound data and the pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection comprises: performing high-pass filtering on the water flow sound data to obtain sound data with environmental noise removed, and performing frame processing on the sound data with environmental noise removed to obtain a sound frame sequence; performing fast Fourier transform on the sound frame sequence to obtain a spectrum sequence, and performing Mel frequency transform on the spectrum sequence to obtain a Mel spectrum; The pipeline vibration data is low-pass filtered to obtain smooth vibration data, and the smooth vibration data is divided into time windows to obtain vibration data segments; the vibration data segments are discrete wavelet transformed to obtain multi-scale vibration coefficients, and the multi-scale vibration coefficients are energy calculated to obtain vibration energy distribution; the Mel spectrum and the vibration energy distribution are feature spliced ​​to obtain a joint acoustic-vibration feature vector, and the joint acoustic-vibration feature vector is standardized to obtain a normalized feature; each component in the normalized feature is integrated by weighted summation to obtain a flow indication parameter, and according to a preset parameter mapping relationship The flow indication parameter is converted to obtain the instantaneous flow estimation value at the collection time; the previously acquired historical flow data is subjected to sliding window processing to obtain a flow data window sequence, and the average value is calculated for each window in the flow data window sequence to obtain a window average flow sequence; the window average flow sequence is subjected to exponential weighted average calculation to obtain the historical average flow; based on the digital confidence score and the abnormal mark, the image recognition reading, the standard meter reading data, the instantaneous flow estimation value and the historical average flow are subjected to multi-dimensional cross-validation to obtain meter reading verification data and an abnormal situation report.

[0006] The present invention also provides a system for verifying the meter reading accuracy of water meter readers based on image recognition, comprising:

[0007] An acquisition module is used to perform multimodal data acquisition on the water meter to obtain an original data set, wherein the original data set includes water meter reading photos, manual meter reading data, water flow sound data, and pipeline vibration data;

[0008] A correction module, used for performing noise reduction, illumination balance and geometric correction on the water meter reading photo to obtain an optimized water meter image;

[0009] An extraction module, for performing feature extraction and classification on the optimized water meter image through a deep convolutional neural network to obtain image recognition readings, digital confidence scores and abnormal markings;

[0010] A processing module, used for performing data standardization processing on the manual meter reading data to obtain standardized meter reading data;

[0011] The analysis module is used to perform time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection. The steps of performing time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection include: performing high-pass filtering on the water flow sound data to obtain sound data with environmental noise removed, and performing frame processing on the sound data with environmental noise removed to obtain a sound frame sequence; performing fast Fourier transform on the sound frame sequence to obtain a spectrum sequence, and performing Mel frequency transform on the spectrum sequence to obtain a Mel spectrum; performing low-pass filtering on the pipeline vibration data to obtain smooth vibration data, and performing time window division on the smooth vibration data to obtain vibration data segments; Perform discrete wavelet transform to obtain multi-scale vibration coefficients, and perform energy calculation on the multi-scale vibration coefficients to obtain vibration energy distribution; perform feature splicing on the Mel spectrum and the vibration energy distribution to obtain a sound-vibration joint feature vector, and perform standardization processing on the sound-vibration joint feature vector to obtain a normalized feature; synthesize the various components in the normalized feature by weighted summation to obtain a flow indication parameter, and convert the flow indication parameter according to a preset parameter mapping relationship to obtain an instantaneous flow estimation value at the acquisition time; perform sliding window processing on the historical flow data obtained in advance to obtain a flow data window sequence, and calculate the average value of each window in the flow data window sequence to obtain a window average flow sequence; perform exponential weighted average calculation on the window average flow sequence to obtain the historical average flow;

[0012] A verification module is used to perform multi-dimensional cross-validation on the image recognition reading, the standard meter reading data, the instantaneous flow estimation value and the historical average flow based on the digital confidence score and the abnormal mark to obtain meter reading verification data and an abnormal situation report.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] In the technical solution provided by the present invention, multimodal data acquisition integrates water meter reading photos, manual meter reading data, water flow sound data and pipeline vibration data, providing a rich source of information for subsequent analysis, significantly increasing the comprehensiveness and reliability of verification. Secondly, the water meter reading photos are subjected to noise reduction, light balance and geometric correction, which effectively improves the image quality and lays the foundation for accurate identification. Feature extraction and classification by deep convolutional neural network can not only accurately identify the readings, but also output digital confidence scores and abnormal marks, standardize the manual meter reading data, and ensure the consistency and comparability of the data. Time-frequency analysis and feature extraction are performed on water flow sound data and pipeline vibration data to obtain instantaneous flow estimation values ​​and historical average flow, providing an independent verification path for meter reading data. Finally, based on digital confidence scores and abnormal marks, multi-dimensional cross-validation is performed on various data, which can not only identify potential meter reading errors, but also classify and evaluate abnormal situations. This multi-angle and multi-level verification method significantly increases the comprehensiveness and reliability of verification, and improves the accuracy of water meter identification, thereby improving the efficiency and accuracy of existing meter reading verification methods, and effectively reducing human errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 It is a flow chart of a method for verifying the meter reading accuracy of a water meter reader based on image recognition in an embodiment of the present invention;

[0017] Figure 2 The present invention is a schematic diagram of a system for verifying the meter reading accuracy of a water meter reader based on image recognition in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention will be described clearly and completely below 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 creative work are within the scope of protection of the present invention.

[0019] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0020] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0021] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , Figure 1 is a flow chart of a method for verifying the meter reading accuracy of a water meter reader based on image recognition according to an embodiment of the present invention. Figure 1 As shown, the following steps are included:

[0022] S101, performing multimodal data collection on a water meter to obtain an original data set, wherein the original data set includes water meter reading photos, manual meter reading data, water flow sound data, and pipeline vibration data;

[0023] S102, performing noise reduction, illumination balance and geometric correction on the water meter reading photo to obtain an optimized water meter image;

[0024] S103, extracting features and classifying the optimized water meter image through a deep convolutional neural network to obtain image recognition readings, digital confidence scores, and abnormal markings;

[0025] S104, performing data standardization processing on the manual meter reading data to obtain standardized meter reading data;

[0026] S105, performing time-frequency analysis and feature extraction on the water flow sound data and the pipeline vibration data to obtain an instantaneous flow estimation value and a historical average flow at the time of collection;

[0027] S106. Based on the digital confidence score and the abnormal mark, the image recognition reading, the standard meter reading data, the instantaneous flow estimation value and the historical average flow are cross-validated in multiple dimensions to obtain the meter reading verification data and the abnormal situation report.

[0028] It should be noted that multimodal data collection is performed on water meters to obtain the original data set. This includes taking photos of water meter readings with a high-resolution camera, recording manual meter reading data, collecting water flow sound data with a high-sensitivity microphone, and collecting pipeline vibration data with a vibration sensor. These multimodal data provide a comprehensive information basis for the subsequent verification process. Next, the water meter reading photos are processed. First, the image noise is reduced by methods such as median filtering, then the overall brightness and contrast of the image are improved by using histogram equalization technology, and finally geometric correction methods such as affine transformation are applied to ensure that the position and angle of the water meter in the image are standardized. These processing steps significantly improve the image quality and lay the foundation for subsequent image recognition. Subsequently, the optimized water meter images are analyzed using a deep convolutional neural network. The network structure adopts the improved EfficientNet-B7, extracts image features through multi-layer convolution and pooling operations, and then obtains the final image recognition reading through a fully connected layer and a softmax classifier. At the same time, the network also outputs the confidence score of each number and generates anomaly tags based on the preset threshold. This information provides an important basis for subsequent cross-validation.

[0029] For manual meter reading data, the system performs a series of standardization processes. First, the data format is unified, and then the unit conversion and decimal place normalization are performed to ensure that all data have a consistent representation. In addition, time series analysis methods are applied to detect and correct outliers, and finally obtain high-quality standardized meter reading data. Water flow sound data and pipeline vibration data are subjected to time-frequency analysis and feature extraction to estimate instantaneous flow and calculate historical average flow. Specifically, the sound data is subjected to fast Fourier transform to obtain spectral features, and the vibration data is subjected to wavelet transform to extract multi-scale features. These features are fused and mapped to obtain instantaneous flow estimates. At the same time, the historical average flow is calculated by sliding window processing and exponential weighted averaging of historical flow data.

[0030] Finally, based on the data obtained from the above processing, multi-dimensional cross-validation is performed. The system comprehensively considers image recognition readings, standard meter reading data, instantaneous flow estimates and historical average flow, combined with digital confidence scores and abnormal markings, and obtains the final meter reading verification data and abnormal situation report through a series of calculations and comparisons.

[0031] For example, during a meter reading process, the value read by the image recognition system from the water meter photo is 10035, with a confidence level of 0.98, the manually entered meter reading data is 10036, the instantaneous flow estimation value is 0.15 m³ / h, and the historical average flow rate is 0.14 m³ / h. First, calculate the difference between the image recognition reading and the manual meter reading data. Assuming that the allowable error range is set to ±2 units, the standardized difference index is (10036-10035) / 2=0.5. Next, calculate the flow change index, that is, the logarithm of the ratio of the instantaneous flow estimation value to the historical average flow rate: log(0.15 / 0.14)≈0.0693. Then, perform a weighted fusion on these two indicators. Assuming that the standardized difference index is given a weight of 0.6 and the flow change index is given a weight of 0.4, the comprehensive deviation index is obtained: 0.60.5 + 0.40.0693 ≈ 0.3277.

[0032] The preset judgment thresholds are as follows: 0-0.3 is normal, 0.3-0.6 is slightly abnormal, 0.6-1 is moderately abnormal, and greater than 1 is severely abnormal. In this example, 0.3277 falls into the slightly abnormal range. Considering the high confidence of image recognition (0.98), the comprehensive deviation index is fine-tuned, but it remains in the slightly abnormal range. The final verification result is generated: the meter reading data is basically credible (because the comprehensive deviation index is close to the normal range), but it is marked as slightly abnormal, and manual review is recommended in the abnormal situation report to ensure the absolute accuracy of the data.

[0033] By executing the above steps, multimodal data collection integrates water meter reading photos, manual meter reading data, water flow sound data and pipeline vibration data, providing a rich source of information for subsequent analysis, significantly increasing the comprehensiveness and reliability of verification. Secondly, noise reduction, light balance and geometric correction are performed on water meter reading photos, which effectively improves the image quality and lays the foundation for accurate recognition. Feature extraction and classification through deep convolutional neural network can not only accurately identify readings, but also output digital confidence scores and abnormal marks, standardize manual meter reading data, and ensure data consistency and comparability. Time-frequency analysis and feature extraction of water flow sound data and pipeline vibration data are performed to obtain instantaneous flow estimation values ​​and historical average flow, providing an independent verification path for meter reading data. Finally, based on digital confidence scores and abnormal marks, multi-dimensional cross-validation of various data can not only identify potential meter reading errors, but also classify and evaluate abnormal situations. This multi-angle and multi-level verification method greatly improves the verification efficiency and reliability of meter reading accuracy and effectively reduces human errors.

[0034] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0035] (1) Take a high-resolution photo of the water meter dial to obtain a photo of the water meter reading, and extract metadata from the photo of the water meter reading to obtain shooting parameter information;

[0036] (2) Extracting the manually entered information to obtain manual meter reading data;

[0037] (3) Collecting sound from the surrounding environment of the water meter to obtain original sound data, and then reducing noise and enhancing the original sound data to obtain water flow sound data;

[0038] (4) Collect vibration signals from the water meter connection pipe to obtain raw vibration data, and perform frequency domain analysis on the raw vibration data to obtain pipe vibration data;

[0039] (5) Based on the shooting parameter information, the water meter reading photos, manual meter reading data, water flow sound data, and pipeline vibration data are time-synchronized to obtain time-aligned multimodal data. The time-aligned multimodal data are then fused to obtain the original data set.

[0040] Specifically, the water meter dial is photographed at high resolution using a digital camera with at least 12 megapixels to ensure that the clear details of the water meter reading are captured. After the shooting is completed, the shooting parameter information, including shooting time, aperture value, shutter speed, and ISO sensitivity, is immediately extracted from the EXIF ​​(Exchangeable Image File Format) data of the photo. At the same time, the meter reader manually enters the reading through a specially designed mobile application, which automatically records the input time and geographic location information to form manual meter reading data.

[0041] For the collection of water flow sound data, a high-sensitivity microphone is used to record around the water meter, and the sampling rate is set to 44.1kHz to capture various sound features including water flow, bubbles, and possible leakage sounds. Subsequently, an adaptive filtering algorithm is applied to the original sound data for noise reduction to remove environmental background noise, and the sound signal related to the water flow is enhanced through dynamic range compression technology. The vibration signal collection of the water meter connecting pipe relies on a high-precision acceleration sensor, and the sampling frequency is set to 1kHz to fully capture the subtle changes in pipe vibration. The collected raw vibration data is converted to the frequency domain through fast Fourier transform (FFT) to obtain a spectrum diagram, from which the main frequency components and corresponding amplitude information are extracted, which reflects the water flow state and pipeline conditions.

[0042] Based on the shooting timestamp extracted from the water meter reading photos, the system performs time synchronization processing on all collected data. First, the timestamps of manual meter reading data, water flow sound data, and pipeline vibration data are aligned with the photo shooting time, and the data with different sampling frequencies are processed by linear interpolation method to ensure that all data correspond to the same time point. Then, the multimodal data after time alignment is fused, and the weighted average method is used to integrate the information of various data sources to generate the final original data set.

[0043] For example: During a meter reading operation, the meter reader took a photo of the water meter at 10:30:00 on September 15, 2023. The EXIF ​​data showed an aperture value of f / 2.8, a shutter speed of 1 / 60 second, and an ISO of 400. At the same time, the meter reader entered the reading 12345 through the application and recorded the input time as 10:30:05. The sound collection device recorded 10 seconds of sound data from 10:29:55 to 10:30:05, and the vibration sensor recorded 7 seconds of vibration data from 10:29:58. During data processing, a high-pass filter was first applied to the sound data, with the cutoff frequency set to 100Hz, effectively removing low-frequency environmental noise. After the vibration data was converted by FFT, two obvious peaks were found at 500Hz and 1000Hz, corresponding to amplitudes of 0.5g and 0.3g (g is the acceleration of gravity), respectively. During time synchronization processing, the other data were aligned based on the photo shooting time of 10:30:00. For example, the manually entered reading time was adjusted from 10:30:05 to 10:30:00, and the sound and vibration data were linearly interpolated to obtain an estimated value at 10:30:00. Finally, the visual information in the photo, the manually entered reading 12345, and the processed sound and vibration features were combined to form a complete original data set for this meter reading operation.

[0044] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0045] (1) Grayscale the water meter reading photo to obtain a grayscale water meter image, and perform histogram equalization on the grayscale water meter image to obtain an initial equalized image;

[0046] (2) Performing median filtering on the initial equalized image to obtain a denoised image, and performing adaptive threshold binarization on the denoised image to obtain a binarized image;

[0047] (3) Perform edge detection on the binary image to obtain an edge image, and perform Hough transform on the edge image to obtain a binary water surface contour image;

[0048] (4) Performing affine transformation on the denoised image according to the binary water surface contour image to obtain a corrected image, and performing gamma correction on the corrected image to obtain a brightness adjusted image;

[0049] (5) performing local contrast enhancement on the brightness adjusted image to obtain a contrast enhanced image, and performing sharpening processing on the contrast enhanced image to obtain a clear image;

[0050] (6) Dehaze the cleared image to obtain an optimized water meter image.

[0051] Specifically, the color water meter reading photo is converted into a grayscale image by calculating the weighted average of the red, green and blue channels of each pixel. The obtained grayscale water meter image only contains brightness information. Then, the grayscale water meter image is histogram equalized. By redistributing the grayscale value distribution of the image, the overall contrast of the image is improved, making the water meter scale and pointer more clearly visible. Subsequently, the median filter is applied to the initial equalized image. This nonlinear filtering method can effectively remove salt and pepper noise while retaining the edge information of the image. The window size of the median filter is usually set to 3x3 or 5x5, depending on the degree of noise in the image. The filtered image is further processed by adaptive threshold binarization. This method dynamically adjusts the threshold according to the grayscale distribution of the local area of ​​the image, converts the image into black and white, and effectively separates the key structural information of the water meter.

[0052] The binarized image is then subjected to edge detection. Common methods include the Sobel operator or the Canny edge detector, which highlights the edge features of the water meter contour and scale lines. The resulting edge image is then processed by the Hough transform, which can detect straight and circular structures in the image, thereby accurately locating the outer contour of the water meter. Based on the detected binary water meter contour image, the denoised image is affine transformed to correct the position and angle of the water meter in the image, ensuring that the water meter is centered and horizontally placed in the image. The corrected image is then gamma corrected to adjust the overall brightness and contrast of the image. The gamma value is usually selected between 0.5 and 2.0, depending on the initial brightness of the image.

[0053] Next, the brightness-adjusted image is subjected to local contrast enhancement. This technique improves the clarity of low-contrast areas in a targeted manner by analyzing the local area features of the image, making the water meter scale and pointer more prominent. The enhanced image is then sharpened. Common methods include unsharp masking or Laplace sharpening, which further improves the image details and makes the water meter readings clearer and more discernible. Finally, the clarified image is dehazed to eliminate the effects of possible image blur or water vapor condensation. Common methods include the dark channel prior dehazing algorithm, which estimates the depth information of the image to restore the details obscured by fog, thereby obtaining the final optimized water meter image.

[0054] For example, the original image obtained from a water meter shooting has a resolution of 3000x2000 pixels. The water meter in the image is slightly tilted, and due to insufficient light, the overall image is dark. First, the image is converted to a grayscale image with a pixel value range from 0 to 255. After histogram equalization, the average brightness of the image is increased from the original 85 to 127, an increase of 42%. A 5x5 median filter window is applied to remove noise, and then the adaptive threshold method is used for binarization. The threshold is set to 95% of the average grayscale value of the local 11x11 pixel area. The Canny algorithm is used for edge detection, with the low threshold set to 50 and the high threshold set to 150. The Hough transform detects the outer circle of the water meter, with a radius of about 40% of the image height. The affine transformation adjusts the water meter to the center of the image and corrects the tilt angle of 2 degrees. The gamma correction value is set to 1.2 to brighten the image. The local contrast enhancement uses the CLAHE (contrast-limited adaptive histogram equalization) algorithm, and the grid size is set to 8x8. The sharpening uses the Laplacian operator with a kernel size of 3x3. The final dehazing process uses the dark channel prior algorithm and the atmospheric light value is estimated to be 220.

[0055] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0056] (1) Performing data enhancement processing on the optimized water meter image to obtain an enhanced image set, and normalizing the enhanced image set to obtain a standardized image set;

[0057] (2) Extract features from the standardized image set using the EfficientNet-B7 network to obtain a primary feature map, and then perform an adaptive pooling operation on the primary feature map to obtain a fixed-size feature map;

[0058] (3) Perform spatial transformation on the fixed-size feature map to obtain an aligned feature map, and then weight the importance of the aligned feature map through the attention mechanism to obtain a weighted feature map;

[0059] (4) Perform global average pooling on the weighted feature map to obtain a feature vector, and then perform dimensionality reduction processing on the feature vector through a fully connected layer to obtain a low-dimensional feature representation;

[0060] (5) Perform softmax classification on the low-dimensional feature representation to obtain the image recognition reading, and calculate the classification probability of each digit to obtain the digit confidence score;

[0061] (6) Analyze the numerical confidence score by setting the confidence threshold and obtain the abnormal mark.

[0062] Specifically, the optimized water meter images are subjected to data augmentation, and multiple image variants are created through a variety of transformation methods, including rotation, scaling, flipping, brightness adjustment, etc. For example, the original image is rotated 5 degrees and 10 degrees clockwise, flipped horizontally, and the brightness is increased by 10%, thereby obtaining an enhanced image set. This enhancement process effectively expands the diversity of training samples and improves the generalization ability of the model. Subsequently, the enhanced image set is normalized, and the pixel values ​​are scaled to between 0 and 1, eliminating the brightness and contrast differences between different images, and obtaining a standardized image set. Next, the EfficientNet-B7 network is used to extract features from the standardized image set. EfficientNet-B7 is an efficient convolutional neural network architecture that achieves excellent feature extraction capabilities at a low computational cost by balancing the network depth, width, and resolution. The network contains multiple convolutional layers, batch normalization layers, and activation functions, which extract low-level to high-level features of the image layer by layer, and finally output a primary feature map. An adaptive pooling operation is performed on the primary feature map to unify feature maps of different sizes to a fixed size, such as 7x7, to obtain a fixed-size feature map, ensuring the consistency of subsequent processing.

[0063] Then, the fixed-size feature map is spatially transformed. This operation automatically adjusts the spatial layout of the feature map by learning the affine transformation parameters, making the key features more aligned and prominent, and obtaining an aligned feature map. Next, the aligned feature map is importance-weighted through the attention mechanism. The attention mechanism calculates the importance score of each feature position and uses these scores to weight the features, highlighting key information and suppressing irrelevant information, thereby obtaining a weighted feature map. The weighted feature map is globally average pooled to compress the two-dimensional feature map into a one-dimensional feature vector, greatly reducing the number of parameters while retaining key feature information. Subsequently, the feature vector is dimensionality reduced through a fully connected layer to further compress the feature space and obtain a low-dimensional feature representation. This low-dimensional feature representation contains the core information of the water meter image, laying the foundation for subsequent classification.

[0064] Finally, the low-dimensional feature representation is softmax classified to obtain the image recognition reading. The Softmax function maps the features to the probability distribution of each digital category, and the category corresponding to the highest probability is the recognition result. At the same time, the classification probability of each number is calculated to obtain the digital confidence score. The digital confidence score is analyzed by setting a confidence threshold (such as 0.9). If the confidence of a certain number is lower than the threshold, it is marked as abnormal and an abnormal mark is obtained.

[0065] For example, in a water meter reading recognition, the original optimized water meter image size is 224x224 pixels. The data augmentation step generates 10 variant images, including ±5°, ±10° rotation, horizontal flip, and brightness adjustment of ±10%. Normalization scales the pixel value from 0-255 to 0-1. The EfficientNet-B7 network extracts features to obtain a 14x14x1280 feature map, which is converted to 7x7x1280 after adaptive pooling. After spatial transformation and attention mechanism processing, the feature dimension remains unchanged, but the importance is adjusted. Global average pooling compresses the features into a 1280-dimensional vector, and the fully connected layer further reduces the dimension to 128 dimensions. Softmax classification obtains a probability distribution of 5 numbers, such as [0.98, 0.99, 0.97, 0.95, 0.92], corresponding to the reading 12345. Since the last digit 0.92 is lower than the set threshold of 0.95, it is marked as an anomaly.

[0066] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0067] (1) Intelligent format recognition is performed on manual meter reading data to obtain multi-format meter reading data, and the multi-format meter reading data is processed uniformly through an adaptive parsing algorithm to obtain standard format data;

[0068] (2) Perform a spatiotemporal consistency check on the standard format data to obtain spatiotemporal labeled data, and correct the abnormal items in the spatiotemporal labeled data to obtain logically consistent data;

[0069] (3) Dynamically convert logically consistent data to obtain unified unit data, and standardize non-standard expressions in the unified unit data to obtain standardized expression data;

[0070] (4) Conduct historical trend analysis on the standard statement data to obtain trend characteristic data, and screen the trend characteristic data to obtain credibility score data;

[0071] (5) Perform weighted averaging on the credibility score data to obtain comprehensive score data, and then correct the comprehensive score data through an automatic calibration algorithm to obtain standardized meter reading data.

[0072] Specifically, the manual meter reading data is intelligently formatted, and regular expressions and natural language processing techniques are used to identify data inputs in different formats, such as "1234.5", "1,234.5" or "1234.5 m³", etc., to obtain multi-format meter reading data. Subsequently, these multi-format data are unified through an adaptive parsing algorithm, which can automatically adjust the parsing rules according to the identified format features and convert the data in different formats into a unified standard format, such as a unified numerical format of "1234.5". Next, the standard format data is checked for spatiotemporal consistency, which involves checking whether the meter reading time and geographic location information are consistent with expectations. For example, check whether the meter reading time is within a reasonable working time range and whether the geographic location is consistent with the scheduled meter reading route. This process generates spatiotemporal tag data, marking any data items that do not meet expectations. For these abnormal items, the system will try to make corrections, such as inferring the correct time or location based on adjacent normal data, so as to obtain logically consistent data.

[0073] Subsequently, the logically consistent data is dynamically converted to units. This is intended to handle possible unit inconsistencies, such as converting different units (e.g., cubic meters, gallons, etc.) into standard units (e.g., cubic meters). The conversion process takes into account specific conversion factors for different regions or types of water meters. At the same time, non-standard expressions in the unified unit data are standardized, such as converting "1.2 thousand cubic meters" to "1,200 cubic meters", so as to obtain standardized expression data. Historical trend analysis of standardized expression data uses time series analysis techniques, such as moving average or exponential smoothing, to calculate water consumption trends over a period of time (e.g., 3 months). The analysis results form trend feature data, including indicators such as average water consumption and water consumption growth rate. Subsequently, the trend feature data is screened to compare the deviation between the current meter reading data and the historical trend. If the deviation exceeds a preset threshold (e.g., ±20%), its credibility score is reduced. This process results in credibility score data.

[0074] Finally, the credibility score data is weighted averaged. Different scoring indicators (such as spatiotemporal consistency, unit standardization, historical trend compliance, etc.) are assigned different weights and calculated together to obtain the comprehensive score data. For example, historical trend compliance may be assigned a higher weight. Then, the comprehensive score data is finally corrected through the automatic calibration algorithm. This algorithm takes into account additional information such as seasonal factors and special events (such as holidays), fine-tunes the score, and finally obtains standardized meter reading data.

[0075] For example, the raw data entered by the meter reader is "1,234.56 m³", the time is "2023-09-15 14:30:00", and the GPS coordinates are "(37.7749, -122.4194)". Intelligent format recognition parses it into the value 1234.56, in cubic meters. The spatiotemporal consistency check found that the time was normal, but the GPS position deviated from the expected route by 50 meters, which was confirmed as valid data after correction. Dynamic unit conversion converts cubic meters into the system standard unit cubic decimeter, and obtains 1234560 cubic decimeters. Historical trend analysis shows that the average monthly water consumption of this user in the past three months is 1,200,000 cubic decimeters, with a standard deviation of 100,000 cubic decimeters. The deviation between the current reading and the historical trend is 2.88%, which is within the acceptable range. Taking all factors into consideration, the final credibility score is 0.95 (out of 1.0). The automatic calibration algorithm took into account the current peak water usage period in summer and fine-tuned the final score to 0.97.

[0076] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0077] (1) High-pass filtering is performed on the water flow sound data to obtain sound data with environmental noise removed, and the sound data with environmental noise removed is processed into frames to obtain a sound frame sequence;

[0078] (2) Perform fast Fourier transform on the sound frame sequence to obtain a spectrum sequence, and perform Mel frequency transform on the spectrum sequence to obtain a Mel spectrum;

[0079] (3) Low-pass filtering is performed on the pipeline vibration data to obtain smooth vibration data, and the smooth vibration data is divided into time windows to obtain vibration data segments;

[0080] (4) Performing discrete wavelet transform on the vibration data segments to obtain multi-scale vibration coefficients, and performing energy calculation on the multi-scale vibration coefficients to obtain vibration energy distribution;

[0081] (5) Perform feature concatenation on the Mel spectrum and vibration energy distribution to obtain a joint acoustic-vibration feature vector, and perform normalization on the joint acoustic-vibration feature vector to obtain a normalized feature;

[0082] (6) The components in the normalized features are integrated by weighted summation to obtain flow indication parameters, and the flow indication parameters are converted according to the preset parameter mapping relationship to obtain the instantaneous flow estimation value at the time of collection;

[0083] (7) Perform sliding window processing on the historical traffic data acquired in advance to obtain a traffic data window sequence, and calculate the average value of each window in the traffic data window sequence to obtain a window average traffic sequence;

[0084] (8) Perform exponentially weighted averaging on the window average flow sequence to obtain the historical average flow.

[0085] Specifically, the water flow sound data is high-pass filtered, and a Butterworth high-pass filter with a cutoff frequency of 100Hz is used to effectively remove low-frequency environmental noise and obtain sound data without environmental noise. Subsequently, these data are framed, with each frame length of 20ms and an overlap rate of 50%, to obtain a sound frame sequence. The sound frame sequence is subjected to a fast Fourier transform (FFT) to convert the time domain signal into a frequency domain representation and obtain a spectrum sequence. The calculation formula of FFT is:

[0086]

[0087] in, is the input signal, is the output spectrum, is the number of FFT points, usually 4096. is an imaginary unit, = -1, It is pi.

[0088] Next, the mel frequency transform is performed on the spectrum sequence to convert the linear frequency scale into a mel scale that is more consistent with human ear perception, and the mel spectrum is obtained. The calculation formula of the mel frequency is:

[0089]

[0090] in, is the linear frequency, is the corresponding Mel frequency.

[0091] At the same time, the pipeline vibration data is low-pass filtered using a Butterworth low-pass filter with a cutoff frequency of 500Hz to obtain smooth vibration data. The smooth vibration data is divided into time windows, each window length is 1 second, and the overlap rate is 30%, to obtain vibration data fragments. The vibration data fragments are subjected to discrete wavelet transform, and the Daubechies wavelet with 5-level decomposition is used to obtain multi-scale vibration coefficients. Then, the energy of the multi-scale vibration coefficients is calculated to obtain the vibration energy distribution. The energy calculation formula is:

[0092]

[0093] in, It is The energy of the scale, It is Scale coefficients, is the number of coefficients for this scale.

[0094] The Mel spectrum and vibration energy distribution are feature spliced ​​to obtain the joint acoustic-vibration feature vector. The vector is normalized and the normalized feature is obtained using the Z-score normalization method. The components in the normalized feature are combined by weighted summation to obtain the flow indication parameter. The weight is determined based on the pre-determined experimental data, for example, the sound feature weight is 0.6 and the vibration feature weight is 0.4. Then, according to the preset parameter mapping relationship, the flow indication parameter is converted into the actual flow value to obtain the instantaneous flow estimation value at the time of collection.

[0095] The previously acquired historical traffic data is processed by sliding windows, with a window size of 30 days and a step size of 1 day, to obtain a traffic data window sequence. The average value is calculated for each window to obtain a window average traffic sequence. Finally, the window average traffic sequence is subjected to exponential weighted average calculation to obtain the historical average traffic. The calculation formula for exponential weighted average is:

[0096]

[0097] in, yes The smoothed value at time, yes The actual observed value at time, is the smoothing coefficient, usually 0.2, yes Smoothed value at time.

[0098] For example, during a meter reading, the 10-second water flow sound data collected was high-pass filtered and divided into 998 frames. FFT was performed on these frames to obtain a spectrum sequence of 4096 points. The spectrum was converted into 40 Mel bands to obtain a Mel spectrum. At the same time, the vibration data was low-pass filtered and windowed to obtain 13 data fragments. After wavelet transform and energy calculation, a 5-scale energy distribution was obtained. The 40-dimensional Mel spectrum and the 5-dimensional vibration energy distribution were spliced ​​to obtain a 45-dimensional sound-vibration joint feature vector. After standardization, the flow indication parameter 0.75 was obtained by weighted summation. According to the preset mapping relationship, it was converted into an instantaneous flow estimate of 0.5 m³ / h. The historical data shows that the average daily flow in the past 30 days is [0.4, 0.45, 0.42, ...,0.48] m³ / h, and the historical average flow is 0.46 m³ / h after exponential weighted average calculation.

[0099] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0100] (1) Threshold segmentation is performed on the digital confidence score to obtain high and low confidence intervals, and the abnormal labels are classified and counted to obtain the abnormal type distribution;

[0101] (2) Calculate the difference between the image recognition reading and the standard meter reading data to obtain the reading difference value, and normalize the reading difference value to obtain the standardized difference index;

[0102] (3) Calculate the ratio of the instantaneous flow estimate to the historical average flow to obtain the flow change ratio, and perform logarithmic transformation on the flow change ratio to obtain the flow change index;

[0103] (4) Perform weighted fusion on the standardized difference index and the flow change index to obtain a comprehensive deviation index, and perform piecewise function mapping on the comprehensive deviation index to obtain an accuracy evaluation value;

[0104] (5) Adjust the accuracy assessment value according to the high and low confidence intervals to obtain a corrected accuracy index, and set an alarm threshold for the corrected accuracy index to obtain a preliminary abnormal sign;

[0105] (6) Based on the distribution of exception types, the preliminary exception signs are refined to obtain detailed exception categories, and the detailed exception categories are prioritized to obtain the exception handling order;

[0106] (7) The modified accuracy index is combined with the exception handling order to obtain meter reading verification data, and descriptive statistical results are generated based on the meter reading verification data to obtain an exception report.

[0107] It should be noted that the digital confidence score is thresholded, the high confidence threshold is set to 0.9, the low confidence threshold is set to 0.7, and the score is divided into three confidence intervals: high, medium, and low. At the same time, the abnormal marks are classified and counted, including blur, occlusion, damage, and other types, to obtain the abnormal type distribution. Next, the difference between the image recognition reading and the standard meter reading data is calculated to obtain the reading difference value. For example, if the image recognition reading is 1234.5 and the standard meter reading data is 1235.0, the difference value is 0.5. The reading difference value is then normalized, and the maximum allowable error (such as 1.0) is used as the normalization factor to obtain the standardized difference index. In this example, the standardized difference index is 0.5 / 1.0=0.5.

[0108] The ratio of the instantaneous flow estimate to the historical average flow is calculated to obtain the flow change ratio. Assuming that the instantaneous flow estimate is 0.6 m³ / h and the historical average flow is 0.5 m³ / h, the flow change ratio is 0.6 / 0.5=1.2. In order to balance the sensitivity of changes in large and small flows, the flow change ratio is logarithmically transformed to obtain the flow change index. In this example, the flow change index is log(1.2)≈0.0792. Subsequently, the standardized difference index and the flow change index are weighted and fused, assuming that the weights are 0.6 and 0.4 respectively, to obtain the comprehensive deviation index. Using the values ​​of the above example, the comprehensive deviation index is 0.6×0.5+0.4×0.0792≈0.3317. The comprehensive deviation index is mapped to a piecewise function, setting 0-0.2 as high accuracy, 0.2-0.5 as medium accuracy, and above 0.5 as low accuracy to obtain the accuracy assessment value. In this example, 0.3317 falls into the medium accuracy interval.

[0109] The accuracy assessment value is adjusted according to the high and low confidence intervals obtained previously. If the digital confidence score falls in the high confidence interval (>0.9), the accuracy assessment value is slightly increased; if it falls in the low confidence interval (<0.7), the accuracy assessment value is reduced. Assuming that the digital confidence score in this example is 0.95, which belongs to the high confidence interval, the accuracy assessment value is increased from "medium" to "medium-high" to obtain the modified accuracy index. An alarm threshold is set for the modified accuracy index. If it is lower than the "medium" level, it is marked as a preliminary abnormal sign. Based on the previously statistical distribution of abnormal types, the preliminary abnormal signs are refined. For example, if the abnormal type is mainly concentrated in "blurry", the abnormal category is refined into "poor image quality". The refined abnormal categories are prioritized, and the priorities may be: severe damage > numerical abnormality > poor image quality > slight deviation, and the abnormal processing order is obtained. Finally, the modified accuracy index is combined with the abnormal processing order to obtain the meter reading verification data. For example, "accuracy: medium-high; abnormal type: poor image quality; processing priority: 3". Descriptive statistical results are generated based on these meter reading verification data, including accuracy distribution, abnormal type frequency, average deviation, etc., to form the final abnormal situation report.

[0110] For example, the image recognition reading of a water meter is 10035.6, and the confidence score is 0.92; the standard meter reading data is 10036.0; the instantaneous flow estimation value is 0.55 m³ / h, and the historical average flow is 0.50 m³ / h. First, 0.92 belongs to the high confidence interval. The reading difference value is 0.4, which is 0.4 / 1.0=0.4 after standardization. The flow change ratio is 0.55 / 0.50=1.1, which is log(1.1)≈0.0414 after logarithmic transformation. The weighted fusion obtains the comprehensive deviation index 0.6×0.4+0.4×0.0414≈0.2566, which is mapped to medium accuracy. Considering the high confidence, the accuracy assessment value is adjusted to a medium-high level. The distribution of abnormal types shows slight blurring. Finally, the meter reading verification data is generated: "Accuracy: medium-high; abnormal type: slight image blur; processing priority: 4".

[0111] The embodiment of the present invention also provides a system for verifying the accuracy of meter reading by water meter readers based on image recognition, such as Figure 2 As shown, the system for verifying the meter reading accuracy of water meter readers based on image recognition specifically includes:

[0112] The acquisition module 201 is used to perform multimodal data acquisition on the water meter to obtain an original data set, wherein the original data set includes water meter reading photos, manual meter reading data, water flow sound data and pipeline vibration data;

[0113] A correction module 202, used for performing noise reduction, illumination balance and geometric correction on the water meter reading photo to obtain an optimized water meter image;

[0114] An extraction module 203 is used to extract features and classify the optimized water meter image through a deep convolutional neural network to obtain image recognition readings, digital confidence scores and abnormal marks;

[0115] The processing module 204 is used to perform data standardization processing on the manual meter reading data to obtain standardized meter reading data;

[0116] The analysis module 205 is used to perform time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection. The step of performing time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection includes: performing high-pass filtering on the water flow sound data to obtain sound data with environmental noise removed, and performing frame processing on the sound data with environmental noise removed to obtain a sound frame sequence; performing fast Fourier transform on the sound frame sequence to obtain a spectrum sequence, and performing Mel frequency transform on the spectrum sequence to obtain a Mel spectrum; performing low-pass filtering on the pipeline vibration data to obtain smooth vibration data, and performing time window division on the smooth vibration data to obtain vibration data segments; The fragment is subjected to discrete wavelet transformation to obtain multi-scale vibration coefficients, and the energy of the multi-scale vibration coefficients is calculated to obtain vibration energy distribution; the Mel spectrum and the vibration energy distribution are subjected to feature splicing to obtain a sound-vibration joint feature vector, and the sound-vibration joint feature vector is subjected to standardization processing to obtain a normalized feature; each component in the normalized feature is integrated by weighted summation to obtain a flow indication parameter, and the flow indication parameter is converted according to a preset parameter mapping relationship to obtain an instantaneous flow estimation value at the time of collection; the historical flow data acquired in advance is subjected to sliding window processing to obtain a flow data window sequence, and the average value of each window in the flow data window sequence is calculated to obtain a window average flow sequence; the window average flow sequence is subjected to exponential weighted average calculation to obtain the historical average flow;

[0117] The verification module 206 is used to perform multi-dimensional cross-validation on the image recognition reading, the standard meter reading data, the instantaneous flow estimation value and the historical average flow based on the digital confidence score and the abnormal mark to obtain meter reading verification data and an abnormal situation report.

[0118] Through the collaborative work of the above modules, multimodal data collection integrates water meter reading photos, manual meter reading data, water flow sound data and pipeline vibration data, providing a rich source of information for subsequent analysis, significantly increasing the comprehensiveness and reliability of verification. Secondly, the water meter reading photos are subjected to noise reduction, light balance and geometric correction, which effectively improves the image quality and lays the foundation for accurate recognition. Through feature extraction and classification by deep convolutional neural network, not only can the readings be accurately identified, but also digital confidence scores and abnormal marks can be output, and manual meter reading data can be standardized to ensure data consistency and comparability. Time-frequency analysis and feature extraction of water flow sound data and pipeline vibration data are performed to obtain instantaneous flow estimation values ​​and historical average flow, providing an independent verification path for meter reading data. Finally, based on digital confidence scores and abnormal marks, multi-dimensional cross-validation of various data can not only identify potential meter reading errors, but also classify and evaluate abnormal situations. This multi-angle and multi-level verification method greatly improves the verification efficiency and reliability of meter reading accuracy and effectively reduces human errors.

[0119] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0120] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0121] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for verifying the accuracy of water meter readers based on image recognition, characterized in that: include: Perform multimodal data collection on the water meter to obtain an original data set, wherein the original data set includes water meter reading photos, manual meter reading data, water flow sound data, and pipeline vibration data; Performing noise reduction, illumination balance and geometric correction on the water meter reading photo to obtain an optimized water meter image; Extracting features and classifying the optimized water meter image through a deep convolutional neural network to obtain image recognition readings, digital confidence scores, and abnormal markings; Performing data standardization processing on the manual meter reading data to obtain standardized meter reading data; Performing time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection, the step of performing time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection comprises: performing high-pass filtering on the water flow sound data to obtain sound data with environmental noise removed, and performing frame processing on the sound data with environmental noise removed to obtain a sound frame sequence; performing fast Fourier transform on the sound frame sequence to obtain a spectrum sequence, and performing Mel frequency transform on the spectrum sequence to obtain a Mel spectrum; performing low-pass filtering on the pipeline vibration data to obtain smooth vibration data, and performing time window division on the smooth vibration data to obtain vibration data segments; performing separation processing on the vibration data segments Scattered wavelet transform is performed to obtain multi-scale vibration coefficients, and energy calculation is performed on the multi-scale vibration coefficients to obtain vibration energy distribution; feature splicing is performed on the Mel spectrum and the vibration energy distribution to obtain a sound-vibration joint feature vector, and the sound-vibration joint feature vector is standardized to obtain a normalized feature; each component in the normalized feature is integrated by weighted summation to obtain a flow indication parameter, and the flow indication parameter is converted according to a preset parameter mapping relationship to obtain an instantaneous flow estimation value at the acquisition time; sliding window processing is performed on the historical flow data obtained in advance to obtain a flow data window sequence, and the average value of each window in the flow data window sequence is calculated to obtain a window average flow sequence; exponential weighted average calculation is performed on the window average flow sequence to obtain the historical average flow; Based on the digital confidence score and the abnormal mark, the image recognition reading, the standard meter reading data, the instantaneous flow estimation value and the historical average flow are cross-validated in multiple dimensions to obtain meter reading verification data and an abnormal situation report.

2. The method for verifying the meter reading accuracy of water meter readers based on image recognition according to claim 1 is characterized in that: The step of performing multimodal data collection on the water meter to obtain an original data set, wherein the original data set includes water meter reading photos, manual meter reading data, water flow sound data and pipeline vibration data, includes: Taking a high-resolution photo of the water meter dial to obtain the water meter reading photo, and extracting metadata from the water meter reading photo to obtain shooting parameter information; Extracting the manually input information to obtain the manual meter reading data; Collecting sound from the surrounding environment of the water meter to obtain original sound data, and performing noise reduction and enhancement on the original sound data to obtain the water flow sound data; Collecting vibration signals of a water meter connection pipe to obtain original vibration data, and performing frequency domain analysis on the original vibration data to obtain the pipe vibration data; Based on the shooting parameter information, the water meter reading photo, the manual meter reading data, the water flow sound data and the pipeline vibration data are subjected to time synchronization processing to obtain the original data set.

3. The method for verifying the meter reading accuracy of water meter readers based on image recognition according to claim 1 is characterized in that: The step of performing noise reduction, illumination balance and geometric correction on the water meter reading photo to obtain an optimized water meter image includes: Gray-scaling the water meter reading photo to obtain a gray-scale water meter image, and histogram equalizing the gray-scale water meter image to obtain an initial equalized image; Performing median filtering on the initial equalized image to obtain a denoised image, and performing adaptive threshold binarization on the denoised image to obtain a binarized image; Performing edge detection on the binary image to obtain an edge image, and performing Hough transform on the edge image to obtain a binary water surface contour line image; Performing an affine transformation on the denoised image according to the binary water surface contour image to obtain a corrected image, and performing gamma correction on the corrected image to obtain a brightness adjusted image; Performing local contrast enhancement on the brightness-adjusted image to obtain a contrast-enhanced image, and performing sharpening processing on the contrast-enhanced image to obtain a clear image; The cleared image is defogged to obtain the optimized water meter image.

4. The method for verifying the meter reading accuracy of water meter readers based on image recognition according to claim 1, characterized in that: The step of extracting features and classifying the optimized water meter image by a deep convolutional neural network to obtain image recognition readings, digital confidence scores and abnormal markings includes: Performing data enhancement processing on the optimized water meter image to obtain an enhanced image set, and performing normalization processing on the enhanced image set to obtain a standardized image set; Performing feature extraction on the standardized image set through an EfficientNet-B7 network to obtain a primary feature map, and performing an adaptive pooling operation on the primary feature map to obtain a fixed-size feature map; Performing spatial transformation on the fixed-size feature map to obtain an aligned feature map, and weighting the importance of the aligned feature map through an attention mechanism to obtain a weighted feature map; Performing global average pooling on the weighted feature map to obtain a feature vector, and performing dimensionality reduction processing on the feature vector through a fully connected layer to obtain a low-dimensional feature representation; Performing softmax classification on the low-dimensional feature representation to obtain the image recognition reading, and calculating the classification probability of each digit to obtain the digit confidence score; The digital confidence score is analyzed by setting a confidence threshold to obtain the abnormal mark.

5. The method for verifying the meter reading accuracy of water meter readers based on image recognition according to claim 1, characterized in that: The step of performing data standardization processing on the manual meter reading data to obtain standardized meter reading data includes: Intelligently identifying the format of the manual meter reading data to obtain multi-format meter reading data, and uniformly processing the multi-format meter reading data through an adaptive parsing algorithm to obtain standard format data; Performing a time-space consistency check on the standard format data to obtain time-space tag data, and correcting abnormal items in the time-space tag data to obtain logically consistent data; Dynamically convert the logically consistent data into units to obtain unified unit data, and standardize the non-standard expressions in the unified unit data to obtain standardized expression data; Performing historical trend analysis on the standard expression data to obtain trend characteristic data, and screening the trend characteristic data to obtain credibility score data; The credibility score data is weighted averaged to obtain comprehensive score data, and the comprehensive score data is corrected by an automatic calibration algorithm to obtain the standard meter reading data.

6. The method for verifying the meter reading accuracy of water meter readers based on image recognition according to claim 1, characterized in that: The step of performing multi-dimensional cross-validation on the image recognition reading, the standard meter reading data, the instantaneous flow estimation value and the historical average flow based on the digital confidence score and the abnormal mark to obtain meter reading verification data and abnormal situation report includes: Performing threshold segmentation on the digital confidence score to obtain high and low confidence intervals, and classifying and counting the abnormal marks to obtain abnormal type distribution; Calculating the difference between the image recognition reading and the standard meter reading data to obtain a reading difference value, and normalizing the reading difference value to obtain a standardized difference index; Calculating the ratio of the instantaneous flow estimation value to the historical average flow to obtain a flow change ratio, and performing logarithmic transformation on the flow change ratio to obtain a flow change index; Performing weighted fusion on the standardized difference index and the flow change index to obtain a comprehensive deviation index, and performing piecewise function mapping on the comprehensive deviation index to obtain an accuracy evaluation value; The accuracy evaluation value is adjusted according to the high and low confidence intervals to obtain a modified accuracy index, and an alarm threshold is set for the modified accuracy index to obtain a preliminary abnormality mark; Refining the preliminary exception flag based on the exception type distribution to obtain detailed exception categories, and prioritizing the detailed exception categories to obtain an exception handling order; The modified accuracy index is combined with the abnormality processing sequence to obtain the meter reading verification data, and descriptive statistical results are generated according to the meter reading verification data to obtain the abnormality report.

7. A system for verifying the accuracy of meter reading by a water meter reader based on image recognition, used to execute the method for verifying the accuracy of meter reading by a water meter reader based on image recognition as claimed in any one of claims 1 to 6, characterized in that: include: An acquisition module is used to perform multimodal data acquisition on the water meter to obtain an original data set, wherein the original data set includes water meter reading photos, manual meter reading data, water flow sound data, and pipeline vibration data; A correction module, used for performing noise reduction, illumination balance and geometric correction on the water meter reading photo to obtain an optimized water meter image; An extraction module, for performing feature extraction and classification on the optimized water meter image through a deep convolutional neural network to obtain image recognition readings, digital confidence scores and abnormal markings; A processing module, used for performing data standardization processing on the manual meter reading data to obtain standardized meter reading data; The analysis module is used to perform time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection. The steps of performing time-frequency analysis and feature extraction on the water flow sound data and pipeline vibration data to obtain an estimated instantaneous flow rate and a historical average flow rate at the time of collection include: performing high-pass filtering on the water flow sound data to obtain sound data with environmental noise removed, and performing frame processing on the sound data with environmental noise removed to obtain a sound frame sequence; performing fast Fourier transform on the sound frame sequence to obtain a spectrum sequence, and performing Mel frequency transform on the spectrum sequence to obtain a Mel spectrum; performing low-pass filtering on the pipeline vibration data to obtain smooth vibration data, and performing time window division on the smooth vibration data to obtain vibration data segments; Perform discrete wavelet transform to obtain multi-scale vibration coefficients, and perform energy calculation on the multi-scale vibration coefficients to obtain vibration energy distribution; perform feature splicing on the Mel spectrum and the vibration energy distribution to obtain a sound-vibration joint feature vector, and perform standardization processing on the sound-vibration joint feature vector to obtain a normalized feature; synthesize the various components in the normalized feature by weighted summation to obtain a flow indication parameter, and convert the flow indication parameter according to a preset parameter mapping relationship to obtain an instantaneous flow estimation value at the acquisition time; perform sliding window processing on the historical flow data obtained in advance to obtain a flow data window sequence, and calculate the average value of each window in the flow data window sequence to obtain a window average flow sequence; perform exponential weighted average calculation on the window average flow sequence to obtain the historical average flow; A verification module is used to perform multi-dimensional cross-validation on the image recognition reading, the standard meter reading data, the instantaneous flow estimation value and the historical average flow based on the digital confidence score and the abnormal mark to obtain meter reading verification data and an abnormal situation report.

Images