Water meter flow real-time correction method and system based on intelligent sensor

Abstract

The invention relates to the technical field of intelligent sensor application, and discloses a water meter flow real-time correction method and system based on an intelligent sensor, and the system comprises a flow signal collection module, a signal feature decomposition module, a frequency domain feature fusion module, an error compensation decision module and a correction instruction execution module. The method comprises the following steps: acquiring multi-dimensional vibration waveform data of a water meter pipeline through a multi-band sensor, performing time-frequency domain decomposition and energy spectrum feature extraction to obtain a local frequency domain feature tensor set, and performing dynamic noise suppression and band correlation optimization to obtain a local optimization feature tensor set; and performing global feature reconstruction based on frequency domain energy distribution to generate an error compensation decision feature map, and finally generating a real-time correction control signal. Through multi-dimensional signal acquisition and intelligent feature processing, the precision and dynamic adaptability of flow metering in a complex pipe network environment are improved, and the method is suitable for high-precision flow correction of an intelligent water meter and has the advantages of being high in anti-interference capacity, accurate in compensation and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent sensor applications, and specifically to a real-time correction method and system for water meter flow based on intelligent sensors. Background Technique

[0002] With the advancement of smart city construction, intelligent water meters, as the core equipment for water resource measurement, their measurement accuracy and stability directly affect the refinement level of water resource management. Traditional water meter flow measurement mainly relies on mechanical structures or single sensors to collect signals. In a complex pipe network environment (such as water flow pulsation, pipe vibration, fluid viscosity change, etc.), it is easily affected by noise interference and multi-factor coupling, resulting in large measurement errors. Specifically, the following prominent problems exist in the existing technologies: The limitations of traditional signal acquisition methods are significant. Single-frequency band or single-type sensors are difficult to comprehensively capture the complex hydrodynamic characteristics in the pipeline. For example, the vibration generated when water flow passes through valves and elbows contains multi-frequency band energy distributions. Single sensors can only obtain signal data in local dimensions and cannot reflect the complete spectral characteristics of the flow state, resulting in a lack of comprehensive information support for subsequent error compensation.

[0003] The robustness of signal processing methods is insufficient. Traditional time-frequency domain analysis methods (such as Fourier transform, wavelet transform) are difficult to dynamically suppress the interference of environmental noise and pipeline inherent vibration when processing non-stationary signals. For example, when a short-term impact flow appears in the pipe network, the noise energy may cover the frequency domain characteristics of the real flow signal, resulting in deviation in feature extraction and further affecting the accuracy of error compensation.

[0004] The defects of the frequency domain feature fusion mechanism are obvious. Existing technologies usually adopt a feature fusion strategy with fixed weights and cannot adapt to the changes in the operating state of the pipe network. For example, under different flow conditions (such as small flow trickle state and large flow turbulent state), the energy correlation between each frequency band is significantly different. Fixed weights cannot dynamically optimize the coupling relationship between frequency bands, resulting in the fused features being unable to accurately represent the actual flow state.

[0005] The degree of intelligence of the error compensation strategy is relatively low. Traditional methods mostly perform error correction based on empirical formulas or simple linear regression models, lacking in-depth exploration of the frequency domain energy distribution law. For example, when the inner wall of the pipeline becomes scaled, resulting in changes in the fluid boundary layer, the frequency domain energy distribution of the flow signal will change non-linearly. Traditional models are difficult to capture such complex changes, resulting in lag or inaccuracy in compensation decisions.

[0006] Finally, there are deficiencies in real-time performance and dynamic adaptability. When faced with sudden traffic fluctuations or changes in the pipe network topology, the existing system cannot quickly complete the entire process of signal acquisition, feature analysis, and compensation instruction generation. For example, during the flow recovery stage after a water pipe burst repair in the water supply network, the traditional system may take a long time to adjust to the accurate metering state, affecting the real-time monitoring and scheduling of water resources. Summary of the Invention

[0007] The purpose of the present invention is to provide a real-time correction method and system for water meter flow based on intelligent sensors to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solutions: A real-time correction system for water meter flow based on intelligent sensors, the system includes: A flow signal acquisition module, used to obtain multi-dimensional vibration waveform data of the water meter pipeline through multi-band sensors; A signal feature decomposition module, used to perform time-frequency domain decomposition and energy spectrum feature extraction on the multi-dimensional vibration waveform data to obtain a set of local frequency domain feature tensors of the flow state; A frequency domain feature fusion module, used to perform dynamic noise suppression and frequency band correlation optimization on the set of local frequency domain feature tensors of the flow state to obtain a set of locally optimized feature tensors of the flow state; An error compensation decision module, used to perform global feature reconstruction based on the frequency domain energy distribution on the set of locally optimized feature tensors of the flow state to generate a flow error compensation decision feature map; A correction instruction execution module, used to generate a real-time flow correction control signal according to the flow error compensation decision feature map.

[0009] Preferably, the signal feature decomposition module includes: A waveform data segmentation unit, used to dynamically divide the multi-dimensional vibration waveform data according to a preset time window to obtain a set of waveform data segments; A feature extraction unit, used to input each segment of data in the set of waveform data segments into a frequency domain encoder based on a multi-scale decomposition network to obtain a set of local frequency domain feature tensors of the flow state.

[0010] Preferably, the frequency domain feature fusion module includes: A frequency domain tensor expansion unit, used to perform matrix expansion on the local frequency domain feature tensors of the flow state along the frequency band dimension to obtain a set of local frequency domain feature vectors of the flow state; A frequency band energy correlation calculation unit, used to calculate the frequency domain correlation coefficient between any two feature vectors in the set of local frequency domain feature vectors of the flow state to generate a frequency band energy correlation matrix; A dynamic noise suppression unit, configured to perform weight correction on the band energy correlation matrix according to the frequency-domain difference between adjacent time windows in the set of local frequency-domain feature vectors of the traffic state to obtain a band energy constraint correlation matrix; A feature optimization unit, configured to perform cross-band convolution fusion on the band energy constraint correlation matrix and the set of local frequency-domain feature vectors of the traffic state to obtain the local optimized feature tensor of the traffic state.

[0011] Preferably, the band energy correlation calculation unit includes: A frequency-domain projection sub-unit, configured to map each feature vector in the set of local frequency-domain feature vectors of the traffic state to an orthogonal frequency-domain space to obtain a set of projected frequency-domain feature vectors; A frequency-domain correlation analysis sub-unit, configured to calculate the spectral energy similarity between any two projected feature vectors in the set of projected frequency-domain feature vectors to generate the band energy correlation matrix composed of multiple band energy correlation values.

[0012] Preferably, the feature optimization unit is specifically implemented as: Perform dilated convolution processing on the band energy constraint correlation matrix to obtain a band energy constraint feature matrix; Input the set of local frequency-domain feature vectors of the traffic state and the band energy constraint feature matrix into a bidirectional recurrent encoding network to obtain a set of local context feature vectors of the traffic state; Perform matrix reconstruction on the set of local context feature vectors of the traffic state to obtain the local optimized feature tensor of the traffic state.

[0013] Preferably, the error compensation decision module includes: A local feature compression unit, configured to perform mean pooling processing on each tensor in the set of local optimized feature tensors of the traffic state in the frequency-domain dimension to obtain a set of local optimized feature vectors of the traffic state; A frequency-domain energy entropy calculation unit, configured to calculate the energy entropy value of each feature vector in the set of local optimized feature vectors of the traffic state to generate a set of traffic dynamic energy entropy; A reference band determination unit, configured to select the local optimized feature vector of the traffic state corresponding to the minimum energy entropy value in the set of traffic dynamic energy entropy as the initial compensation reference vector; A dynamic compensation weight calculation unit, configured to calculate the dynamic compensation weight of each feature vector according to the frequency-domain distance between each feature vector in the set of local optimized feature vectors of the traffic state and the initial compensation reference vector and the energy entropy value of each feature vector to generate a set of dynamic compensation weights; A global reconstruction unit for weighted superposition of the set of locally optimized feature vectors of the traffic state by using the set of dynamic compensation weights to generate the traffic error compensation decision feature map.

[0014] Preferably, the frequency-domain energy entropy calculation unit is specifically implemented as: Calculate the median vector and the range vector of the locally optimized feature vectors of the traffic state; Perform element-wise difference calculation between the locally optimized feature vectors of the traffic state and the median vector, and perform a cubic power operation on the difference result to obtain a traffic feature energy difference vector; Calculate the overall mean of the traffic feature energy difference vector; perform a ratio operation between the mean and the cube value of the range vector, and input the result into a normalization function to obtain the energy entropy value.

[0015] Preferably, the dynamic compensation weight calculation unit is specifically implemented as: Multiply the energy entropy value of the locally optimized feature vectors of the traffic state by the energy entropy value of the initial compensation reference vector and a first compensation coefficient to obtain a first dynamic compensation factor; Multiply the absolute value of the Manhattan distance between the locally optimized feature vectors of the traffic state and the initial compensation reference vector by a second compensation coefficient to obtain a second dynamic compensation factor; Perform a product operation on the first dynamic compensation factor and the second dynamic compensation factor to obtain the dynamic compensation weight.

[0016] Preferably, the correction instruction execution module is specifically implemented as: Input the traffic error compensation decision feature map into a correction signal generator based on a random forest to obtain the real-time traffic correction control signal, which is used to indicate the calibration strategy of the flow measurement device.

[0017] Preferably, the present invention further includes a method for real-time correction of water meter flow based on an intelligent sensor, and the method includes: Collect multi-dimensional vibration waveform data of the water meter pipeline through a multi-band sensor; Perform time-frequency domain decomposition and energy spectrum feature extraction on the multi-dimensional vibration waveform data to obtain a set of locally frequency-domain feature tensors of the traffic state; Perform dynamic noise suppression and frequency band correlation optimization on the set of locally frequency-domain feature tensors of the traffic state to obtain a set of locally optimized feature tensors of the traffic state; Perform global feature reconstruction based on the frequency-domain energy distribution on the set of locally optimized feature tensors of the traffic state to generate a traffic error compensation decision feature map; Generate a real-time traffic correction control signal according to the traffic error compensation decision feature map.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: At the signal acquisition and feature processing level, the application of multi-band sensors breaks through the limitations of traditional single sensors, and can comprehensively capture multi-dimensional vibration signals generated by the fluid movement in the pipeline, covering the energy distribution characteristics of different frequency bands. Through time-frequency domain decomposition and energy spectrum feature extraction, the system can transform the original vibration data into a set of local frequency domain feature tensors of the flow state, realizing the refined characterization of the flow signal and providing rich feature information for subsequent error compensation. Compared with the traditional single-dimensional signal processing method, this method can more accurately reveal the internal relationship between flow changes and vibration characteristics, effectively improving the comprehensiveness and accuracy of feature extraction.

[0019] In the frequency domain feature fusion link, dynamic noise suppression and frequency band correlation optimization technologies realize the adaptive suppression of noise components and the dynamic optimization of the coupling relationship between frequency bands by constructing a frequency band energy correlation matrix and performing weight correction. Specifically, by calculating the frequency domain correlation coefficient between feature vectors and combining the frequency domain differences of adjacent time windows for weight adjustment, the system can filter out the interference of environmental noise and pipeline inherent vibration in real time, while strengthening the energy correlation between effective frequency bands. This data-driven feature fusion mechanism enables the locally optimized feature tensor of the fused flow state to more realistically reflect the actual flow state, significantly enhancing the robustness and anti-interference ability of the features.

[0020] In the error compensation decision-making process, the global feature reconstruction technology based on frequency domain energy distribution realizes the intelligent analysis and precise compensation of flow errors by introducing key parameters such as energy entropy value and dynamic compensation weight. By calculating the energy entropy value of each feature vector to determine the initial compensation reference vector and dynamically adjusting the compensation weight according to the frequency domain distance and energy entropy value, the system can adapt to the changes in the frequency domain energy distribution under different flow conditions, avoiding the limitations of the traditional fixed weight compensation strategy. This compensation decision-making mechanism based on global feature reconstruction can more accurately capture the non-linear change law of the flow signal, thereby generating a flow error compensation decision feature map that better meets the actual needs, significantly improving the accuracy and dynamic adaptability of error compensation.

[0021] At the level of system real-time performance and engineering applications, the calibration instruction execution module generates real-time flow calibration control signals by introducing the random forest algorithm, achieving a fast response from feature analysis to compensation instruction generation. This algorithm has efficient non-linear modeling capabilities and parallel computing characteristics, enabling the processing of feature maps and the generation of control signals within a short time, meeting the strict real-time requirements of intelligent water meters. In addition, the entire system adopts a modular design, and each functional module (such as signal acquisition, feature decomposition, feature fusion, error compensation, and instruction execution) can be independently optimized and upgraded, facilitating engineering deployment and maintenance. It has strong versatility and scalability and is applicable to different types of water supply pipe networks and flow measurement scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is the working principle diagram of the water meter flow real-time calibration system based on intelligent sensors according to the present invention; Figure 2 is the working principle diagram of the frequency-domain feature fusion module; Figure 3 is the working principle diagram of the error compensation decision module; Figure 4 is the working principle diagram of the frequency-domain energy entropy calculation unit. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0024] Please refer to Figures 1-4 , a water meter flow real-time calibration system based on intelligent sensors according to the present invention, the system includes the following cooperating functional modules to achieve real-time calibration of the water meter flow, and the specific implementation steps are as follows: Flow signal acquisition module: Through multi-band sensors deployed on the water meter pipeline, continuously obtain multi-dimensional vibration waveform data caused by water flow in the pipeline. This sensor can cover the acquisition of vibration signals in different frequency ranges, for example, simultaneously acquire vibration signals in the low-frequency, medium-frequency, and high-frequency bands to comprehensively reflect the dynamic characteristics of water flow in the pipeline. The multi-dimensional vibration waveform data includes, but is not limited to, vibration signals in the radial, axial, and circumferential directions of the pipeline, and the synchronous acquisition of vibrations in different directions is achieved through the spatial arrangement of multiple sensors.

[0025] Signal Feature Decomposition Module: It performs time-frequency domain decomposition and energy spectrum feature extraction on the acquired multi-dimensional vibration waveform data. Specifically, first, the continuous vibration waveform data is segmented into multiple waveform data segments according to a preset time window. The length of each time window can be set according to the water flow characteristics and subsequent processing requirements, for example, set to a value between 0.1 second and 1 second. Then, each waveform data segment is input into the frequency domain encoder based on the multi-scale decomposition network. Through multi-layer convolution and pooling operations, the encoder performs time-frequency domain decomposition at different scales on the waveform data segment, extracts the energy spectrum features of each frequency band, and finally obtains a set of local frequency domain feature tensors of the flow state. Each feature tensor contains the energy distribution information corresponding to the time window and frequency band.

[0026] Frequency Domain Feature Fusion Module: For the set of local frequency domain feature tensors of the flow state, it performs dynamic noise suppression and frequency band correlation optimization. By analyzing the energy correlation between different frequency bands and the differences in the time series, it suppresses noise interference and enhances the correlation between effective frequency bands, thereby obtaining a set of locally optimized feature tensors of the flow state. The optimized feature tensors can more accurately reflect the real flow state and reduce the influence of noise and redundant information.

[0027] Error Compensation Decision Module: It performs global feature reconstruction based on the frequency domain energy distribution on the set of locally optimized feature tensors of the flow state. By synthesizing the frequency domain energy information in each local feature, it generates a flow error compensation decision feature map. This feature map intuitively shows the error distribution and compensation direction existing in the current flow measurement, providing a decision basis for subsequent calibration.

[0028] Calibration Instruction Execution Module: According to the generated flow error compensation decision feature map, it generates a real-time flow calibration control signal. This signal is transmitted to the flow measurement device, instructing it to execute the corresponding calibration strategy to achieve real-time calibration of the flow measurement error of the water meter.

[0029] The present invention will be further described below in conjunction with Embodiments 1 to 5: Embodiment 1: In this embodiment, the specific structure and working process of the signal feature decomposition module are described in detail. The signal feature decomposition module includes a waveform data segmentation unit and a feature extraction unit, which work together to achieve time-frequency domain decomposition and energy spectrum feature extraction of multi-dimensional vibration waveform data.

[0030] The core function of the waveform data segmentation unit is to dynamically divide the multi-dimensional vibration waveform data according to a preset time window. The setting of the preset time window needs to comprehensively consider the water flow characteristics and the real-time requirements of subsequent processing. For example, in the industrial water use scenario, if the water flow velocity in the pipeline is high and the state changes rapidly, the time window can be set to 0.2 seconds to ensure that the segmented waveform data segments can capture the transient changes of water flow vibration in a timely manner; while in the conventional scenario of domestic water use, the water flow state is relatively stable, and the time window can be appropriately increased to 0.8 seconds to reduce the number of data segments and the subsequent calculation amount. During the dynamic segmentation process, the waveform data segmentation unit slides and intercepts the original vibration signal with continuous time windows, and each waveform data segment within a time window contains complete multi-dimensional vibration information (such as vibration signals in the radial, axial, and circumferential directions), forming an ordered set of waveform data segments. It should be noted that an overlapping interval (such as 10% of the window length) can be set between adjacent time windows as needed to avoid the loss of boundary information caused by signal truncation.

[0031] The main role of the feature extraction unit is to input each segment of data in the set of waveform data segments into the frequency domain encoder based on the multi-scale decomposition network to obtain a set of local frequency domain feature tensors of the flow state. The multi-scale decomposition network adopts a hierarchical convolutional structure design, and its core is composed of a combination of convolutional layers, activation layers, and pooling layers with different scales. Specifically, the input single-segment waveform data first enters the first convolutional layer, which is equipped with a relatively large-sized convolutional kernel (such as 5×5) for preliminary feature extraction of the input data, focusing on capturing the global vibration characteristics in the low-frequency band, such as the overall low-frequency vibration trend of the pipeline. After the convolutional operation, the output feature map is passed to the activation layer (such as the ReLU activation function) to introduce non-linearity and enhance the model's ability to express complex signals. Subsequently, the feature map is downsampled through the pooling layer (such as max pooling) to reduce the feature dimension and retain the main features.

[0032] The second convolutional layer uses a medium-sized convolutional kernel (such as 3×3) to extract local features in the middle frequency band, such as the intermediate frequency vibration components generated by the water flow impacting the pipeline joint. After the same convolutional, activation, and pooling operations, the output feature map is further passed to the subsequent levels. The subsequent convolutional layers can use smaller-sized convolutional kernels (such as 1×1) as needed to capture the detailed features in the high-frequency band, such as the high-frequency vibration noise caused by water flow turbulence. The stride parameters of each layer of convolutional kernels can be adjusted according to the feature extraction requirements. For example, the stride is set to 2 in the first layer to quickly reduce the dimension, and the stride is set to 1 in the subsequent layers to retain more details.

[0033] Through this multi-scale hierarchical processing, the frequency-domain encoder can perform multi-resolution time-frequency domain decomposition on the input waveform data segments. The feature maps output by each convolutional layer correspond to the energy spectrum features of different frequency scales. For example, the first layer corresponds to low-frequency features, the second layer corresponds to medium-frequency features, and the subsequent layers correspond to high-frequency features. These feature maps are stacked in the depth dimension to form a multi-dimensional feature tensor. For each waveform data segment, the frequency-domain encoder finally outputs a local frequency-domain feature tensor of the flow state that includes the time dimension, frequency dimension, and energy dimension, which records the energy distribution of different frequency bands within this time window. After all waveform data segments are processed by the frequency-domain encoder, a set of local frequency-domain feature tensors of the flow state is formed.

[0034] In the specific implementation, the network parameters of the frequency-domain encoder need to be optimized using training data. During the training process, the historical vibration waveform data collected and its corresponding true flow error are used as the input and label, and the parameters such as the weights and biases of the convolutional kernels are adjusted through the backpropagation algorithm, enabling the network to accurately extract the frequency-domain features related to the flow error from the vibration signal. For example, for the waveform data segments with flow measurement errors, the network needs to learn to enhance the feature frequency bands related to the error (such as abnormal vibration energy at specific frequencies) and suppress the irrelevant noise frequency bands.

[0035] In addition, the feature extraction unit can set multiple parallel frequency-domain encoder branches according to actual needs. Each branch corresponds to a different multi-scale decomposition network structure to achieve independent feature extraction of multi-dimensional vibration data (such as vibration signals in different directions). For example, independent encoder branches are designed for the radial vibration signal and the axial vibration signal respectively. Each branch outputs the corresponding local frequency-domain feature tensor, and finally the tensors of each branch are concatenated in the dimension to form a comprehensive tensor set containing multi-directional features. This parallel processing method can make full use of the spatial information of multi-dimensional data and improve the comprehensiveness of feature extraction.

[0036] The signal feature decomposition module realizes the conversion from the original vibration signal to the structured frequency-domain feature tensor through the dynamic time window division of the waveform data segmentation unit and the multi-scale frequency-domain encoding of the feature extraction unit. This process not only discretizes the continuous time signal into processable segments but also realizes the refined analysis of different frequency components in the vibration signal through the hierarchical feature extraction of the multi-layer convolutional network, providing the key basic data for subsequent frequency-domain feature fusion and error compensation decision-making.

[0037] Embodiment 2: This embodiment elaborates in detail the composition and working process of the frequency-domain feature fusion module. The frequency-domain feature fusion module includes a frequency-domain tensor expansion unit, a band energy correlation calculation unit, a dynamic noise suppression unit, and a feature optimization unit. Each unit realizes the dynamic noise suppression of the frequency-domain features and the optimization of the band correlation through a series of operations.

[0038] The function of the frequency-domain tensor expansion unit is to perform matrix expansion on the local frequency-domain feature tensor of the traffic state along the frequency band dimension. The local frequency-domain feature tensor of the traffic state is a multi-dimensional data structure output by the feature extraction unit, usually including multiple dimensions such as time, frequency, and energy. When expanding along the frequency band dimension, this unit arranges the eigenvalue of each frequency band in the frequency band dimension of each tensor in sequence and converts it into a one-dimensional feature vector. For example, if the frequency band dimension of a certain tensor contains 100 frequency bands, then each tensor is expanded into a feature vector with a length of 100, and each element in the vector corresponds to the eigenvalue of a frequency band. Through this operation, the multi-dimensional feature tensor is converted into a two-dimensional matrix form, forming a set of local frequency-domain feature vectors of the traffic state. Each row of the matrix represents the feature vector corresponding to a time window, and each column represents the eigenvalue distribution of a frequency band in different time windows.

[0039] The band energy correlation calculation unit is used to calculate the frequency-domain correlation coefficient between any two feature vectors in the set of feature vectors and generate a band energy correlation matrix. This unit first maps each feature vector to an orthogonal frequency-domain space through a frequency-domain projection sub-unit. The mapping of the orthogonal frequency-domain space can be achieved through an orthogonal transformation. For example, using the Fourier transform to convert the feature vector from the time domain to the frequency domain, so that different frequency bands satisfy orthogonality mathematically, thereby eliminating the redundant correlation between frequency bands. In the set of projected frequency-domain feature vectors after mapping, each dimension (frequency band) of each vector is independent of each other, which is convenient for subsequent correlation analysis.

[0040] The frequency-domain correlation analysis sub-unit calculates the spectral energy similarity between any two projected feature vectors based on the projected feature vectors. The spectral energy similarity can be measured by methods such as vector inner product or cosine similarity. Its essence is to evaluate the similarity of the energy distribution between two frequency bands. For example, for two projected feature vectors A and B, calculate their cosine similarity: if the similarity value is close to 1, it indicates that the energy distributions of these two frequency bands are highly similar and may correspond to the same physical phenomenon (such as the vibration generated by water flow impacting the same component); if the similarity value is close to 0, it indicates that the energy distributions of the two are quite different and may belong to different signal components (such as the effective signal and noise). Arrange the similarity values of all pairwise feature vectors in sequence to form an N×N band energy correlation matrix (N is the number of frequency bands). The element in the i-th row and j-th column of the matrix represents the energy correlation strength between the i-th frequency band and the j-th frequency band.

[0041] The dynamic noise suppression unit corrects the weights of the band energy correlation matrix according to the frequency-domain differences between adjacent time windows in the set of eigenvectors. The frequency-domain differences between adjacent time windows reflect the stability of the signal in the time series: if the eigenvalues of a certain frequency band change violently in adjacent windows (such as a sudden increase or decrease in energy), it indicates that this frequency band may be interfered by noise (such as instantaneous vibration of the pipeline, external impact), and its correlation weight needs to be suppressed; on the contrary, if the eigenvalue change is gentle, it is considered that this frequency band belongs to a stable effective signal, and the weight can be maintained or enhanced. The specific correction method is as follows: calculate the absolute value of the eigenvalue difference of each frequency band in adjacent time windows, set a threshold (such as twice the average difference determined by statistical analysis of historical data), and for the frequency bands whose differences exceed the threshold, reduce their weights in the correlation matrix according to the difference ratio (for example, the larger the difference, the larger the weight attenuation coefficient). In this way, a band energy constrained correlation matrix is generated, which suppresses the correlation intensity of the noise frequency bands and retains the correlation of the effective frequency bands.

[0042] The feature optimization unit performs cross-band convolution fusion on the band energy constrained correlation matrix and the set of eigenvectors. First, perform dilated convolution on the band energy constrained correlation matrix. Dilated convolution expands the receptive field of the convolution kernel by inserting zero values between the weights of the standard convolution kernel, so as to capture more extensive band correlation information without increasing the computational complexity. For example, using a 3×3 convolution kernel with a dilation rate of 2, its actual receptive field is equivalent to that of a 5×5 convolution kernel, but the number of parameters is still 3×3. After the dilated convolution process, the correlation matrix is converted into a band energy constrained feature matrix, which contains the long-range correlation information between different frequency bands.

[0043] Input the set of eigenvectors and the band energy constrained feature matrix into a bidirectional recurrent encoding network. The bidirectional recurrent encoding network consists of a forward recurrent layer and a backward recurrent layer: the forward recurrent layer processes the eigenvector sequence from front to back to capture the influence of past time windows on the current window; the backward recurrent layer processes from back to front to capture the dependence of future time windows on the current window. Through this bidirectional processing, the network can learn the context dependence between eigenvectors, such as whether the feature change of a certain frequency band is related to the feature changes of other frequency bands in the front and back time windows. The output of the bidirectional recurrent layer is a set of local context feature vectors of the flow state, and each vector fuses the band correlation information of the current time window and the front and back windows.

[0044] Finally, the feature optimization unit performs matrix reconstruction on the context feature vector set. Matrix reconstruction is to convert the two-dimensional feature vector set (time window × frequency band dimension) back into a multi-dimensional tensor form, restoring dimension information such as time and frequency band, and forming a locally optimized feature tensor of the traffic state. This tensor not only retains the time-frequency domain information of the original features, but also enhances the reliability and semantic information of the features through frequency band correlation optimization and noise suppression, providing better input for subsequent error compensation decisions.

[0045] In specific implementation, the matrix expansion method of the frequency domain tensor expansion unit needs to be strictly matched with the output dimension of the feature extraction unit to ensure that the eigenvalue of each frequency band corresponds accurately. The orthogonal projection and similarity calculation of the frequency band energy correlation calculation unit can be implemented through existing mathematical libraries (such as the NumPy library in Python), avoiding complex formula derivation. The bidirectional recursive encoding network can adopt structures such as long short-term memory network (LSTM) or gated recurrent unit (GRU). By setting appropriate hidden layer dimensions and the number of iterations, the calculation efficiency and feature fusion effect can be balanced.

[0046] Embodiment 3: This embodiment focuses on the specific implementation of the error compensation decision module. The error compensation decision module includes a local feature compression unit, a frequency domain energy entropy calculation unit, a reference frequency band determination unit, a dynamic compensation weight calculation unit, and a global reconstruction unit. Each unit generates a traffic error compensation decision feature map through compression, analysis, and reconstruction of the frequency domain features.

[0047] The function of the local feature compression unit is to perform mean pooling processing on each tensor in the set of locally optimized feature tensors of the traffic state in the frequency domain dimension. The locally optimized feature tensor of the traffic state is a multi-dimensional data structure, including dimensions such as time, frequency band, and energy. The mean pooling processing is targeted at the frequency domain dimension, calculating the average value of all frequency band eigenvalues of each tensor in the frequency band dimension, and compressing the multi-dimensional tensor into a one-dimensional feature vector. For example, if a tensor has 200 frequency bands in the frequency band dimension, a mean value vector with a length of 200 is generated after mean pooling. This vector retains the overall energy distribution information of the tensor in the frequency domain dimension and eliminates the detailed fluctuations of local frequency bands. Through this operation, all tensors are converted into a set of locally optimized feature vectors of the traffic state, and the feature dimension is reduced from multi-dimensional to one-dimensional, simplifying the subsequent calculation complexity while retaining the main frequency domain energy features.

[0048] The frequency-domain energy entropy calculation unit is used to calculate the energy entropy values of each eigenvector in the eigenvector set and generate a traffic dynamic energy entropy set. The specific implementation process is as follows: First, calculate the median vector and the range vector for each eigenvector. The median vector is formed by taking the middle value after sorting the elements in the eigenvector by numerical size, reflecting the central tendency of the eigenvalue; the range vector is the difference between the maximum and minimum values in the eigenvector, reflecting the fluctuation range of the eigenvalue. Then, perform an element-by-element difference calculation between the eigenvector and the median vector to obtain the deviation degree of each element from the median, and then perform a cubic power operation on the difference result to amplify the larger deviation values and highlight the unevenness of the energy distribution, forming a traffic feature energy difference vector. For example, if the difference between an element and the median is 2, it becomes 8 after the cubic power operation, and if the difference is -1, it becomes -1. Through this non-linear transformation, the influence of outliers is enhanced.

[0049] Subsequently, calculate the overall mean of the traffic feature energy difference vector, which reflects the average degree of deviation of the elements in the eigenvector from the median. Perform a ratio operation between the mean and the cube value of the range vector to obtain a dimensionless value, which combines the deviation degree and the fluctuation range of the eigenvalue. Finally, input this value into a normalization function (such as a normalization function) to map it to the interval [0,1] to obtain the energy entropy value. The larger the energy entropy value, the more uneven the energy distribution of the eigenvector, and it may contain more noise or abnormal features; the smaller the energy entropy value, the more concentrated and stable the energy distribution, and it is more likely to correspond to the characteristics of the real traffic state.

[0050] The reference frequency band determination unit selects the eigenvector corresponding to the minimum energy entropy value from the traffic dynamic energy entropy set as the initial compensation reference vector. The minimum energy entropy value means that the energy distribution of this eigenvector is the most uniform and stable, so it is used as the reference for error compensation. For example, if there are 100 eigenvector entropy values in the energy entropy set, and the entropy value of the 35th vector is the smallest, then select this vector as the initial compensation reference vector, representing the most reliable frequency-domain feature under the current traffic state.

[0051] The dynamic compensation weight calculation unit calculates the dynamic compensation weights according to the frequency-domain distance and energy entropy value between each eigenvector and the initial compensation reference vector, and generates a dynamic compensation weight set. The specific steps are as follows: First, calculate the product of the energy entropy value of the eigenvector and the energy entropy value of the initial compensation reference vector, and multiply it by the first compensation coefficient to obtain the first dynamic compensation factor. The first compensation coefficient is an adjustable parameter (such as set to 0.5), which is used to control the influence degree of the energy entropy value on the compensation weight. If the energy entropy value of a certain eigenvector is less than that of the reference vector, it means that its energy distribution is more stable, and the first dynamic compensation factor will increase, and vice versa.

[0052] Secondly, calculate the absolute value of the Manhattan distance between the feature vector and the initial compensation reference vector. This distance measures the absolute difference between the two vectors in the frequency domain space. The Manhattan distance is calculated as the sum of the absolute values of the differences of the corresponding elements. For example, for vector A = [a1, a2,..., an] and vector B = [b1, b2,..., bn], the Manhattan distance is |a1 - b1| + |a2 - b2| +... + |an - bn|. Multiply the absolute value of the distance by the second compensation coefficient (such as set to 0.1) to obtain the second dynamic compensation factor. This factor reflects the similarity between the feature vector and the reference vector. The smaller the distance, the smaller the factor value, indicating a higher similarity.

[0053] Finally, multiply the first dynamic compensation factor by the second dynamic compensation factor to obtain the dynamic compensation weight. For example, if the first dynamic compensation factor of a certain feature vector is 0.8 and the second dynamic compensation factor is 0.3, then the dynamic compensation weight is 0.24. The dynamic compensation weight combines the two factors of energy entropy value and frequency domain distance: the lower the energy entropy value and the smaller the frequency domain distance of the feature vector, the greater the weight, and the higher the proportion in the error compensation; conversely, the smaller the weight, indicating lower reliability or greater difference from the reference, and less contribution to the compensation.

[0054] The global reconstruction unit uses the set of dynamic compensation weights to perform weighted superposition on the set of feature vectors to generate a flow error compensation decision feature map. The specific operation is as follows: Multiply each element of each feature vector by the corresponding dynamic compensation weight to obtain the weighted feature vector, and then add all the weighted feature vectors element by element to form the final feature map. For example, if there are M feature vectors and each vector has a length of N, a feature map of dimension N is generated after weighted superposition, where the value of each element is the sum of the products of the elements at the corresponding positions of all feature vectors multiplied by the weights. Through weight assignment, this feature map highlights the contributions of feature vectors with high reliability and similarity to the reference, suppresses the influence of noise and abnormal features, and thus intuitively reflects the distribution and compensation direction of errors in the current flow measurement.

[0055] In the specific implementation, the mean pooling of the local feature compression unit can be efficiently implemented through matrix operations, such as using the mean function of the tensor library to calculate along the specified dimension. The median vector, range vector, and normalization function of the frequency domain energy entropy calculation unit can be implemented through existing data processing tools to avoid complex formula derivation. The Manhattan distance and coefficient product in the calculation of the dynamic compensation weight can be completed through the vector operation library to ensure calculation efficiency. Attention should be paid to the alignment of vector dimensions in the weighted superposition operation of the global reconstruction unit to ensure that the elements of each feature vector are correctly corresponding to the weights.

[0056] Example 4: This embodiment details the working principle and implementation method of the calibration instruction execution module. The core function of the calibration instruction execution module is to convert the flow error compensation decision feature map into a real-time flow calibration control signal, which is used to instruct the flow measurement device to execute the calibration strategy, and its specific implementation depends on the calibration signal generator based on random forest.

[0057] The flow error compensation decision feature map is a multi-dimensional data structure output by the error compensation decision module, usually represented in the form of a matrix or vector, where each element corresponds to the error compensation information in the frequency domain or time dimension. Since the random forest algorithm requires the input to be a one-dimensional feature vector, the feature map needs to be preprocessed. The preprocessing process includes dimensionality compression and normalization: Dimensionality compression converts the multi-dimensional feature map into a one-dimensional vector through a flattening operation. For example, a matrix of M×N is expanded row by row or column by column into a vector of length M×N; Normalization maps the element values in the vector to a specific range (such as [0,1]) to avoid affecting the model training and prediction results due to the scale difference of the feature values.

[0058] The calibration signal generator based on random forest consists of multiple decision trees, and these decision trees improve the generalization ability and robustness of the model through the way of ensemble learning. In the training stage, a historical training data set needs to be prepared, and each piece of data contains two parts: one is the preprocessed feature vector (corresponding to the historical samples of the flow error compensation decision feature map), and the other is the corresponding label (that is, the actual calibration control signal to be generated). The form of the label is determined according to the calibration strategy of the flow measurement device. For example, for an electromagnetic water meter, the label can be defined as the discretized values of the sensitivity adjustment coefficient (such as -3, -2, -1, 0, +1, +2, +3, corresponding to different degrees of gain or attenuation); for a mechanical water meter, the label can represent the adjustment gear of the gear transmission ratio.

[0059] During the training process, the construction of each decision tree adopts the Bootstrap method: Randomly select some samples (usually 60%-80% of the original data set) and features (usually the square root of the total number of features) from the training data for training to reduce the risk of overfitting and increase the diversity of the model. Each node of the decision tree selects the optimal splitting feature and splitting point through metrics such as Gini Impurity or information gain, and divides the samples into different sub-nodes until the preset maximum depth or minimum number of samples is reached. For example, a certain node selects the threshold T of feature X as the splitting condition, divides the samples with feature values less than T into the left subtree, and the samples greater than or equal to T into the right subtree, and constructs a complete tree structure through recursive splitting.

[0060] After all decision trees are trained, the calibration signal generator has the prediction ability. For the input real-time traffic error compensation decision feature map, first perform the same preprocessing operations as the training data to obtain a one-dimensional normalized feature vector. Then, input this vector into the random forest model, and each decision tree makes predictions independently, outputting a discrete calibration signal value. The final real-time traffic calibration control signal is determined by the voting method: for classification tasks, select the category predicted by the majority of decision trees as the result; for regression tasks, take the average of the predicted values of all decision trees. For example, if 4 out of 7 decision trees predict the calibration signal to be +2, 2 predict +1, and 1 predicts +3, then the final signal is +2.

[0061] After the real-time traffic calibration control signal is generated, it needs to be transmitted to the flow measurement device to execute the calibration strategy. Different types of flow measurement devices correspond to different calibration mechanisms: Electromagnetic water meter: Its working principle is based on Faraday's law of electromagnetic induction. The calibration strategy can be achieved by adjusting the excitation current of the sensor or the gain of the signal amplification circuit. For example, when the calibration signal is +1, it indicates increasing the excitation current by 10% to improve the sensitivity of the sensor to low-flow signals; when the signal is -2, it indicates reducing the signal amplification gain by 20% to suppress signal saturation at high flows.

[0062] Mechanical water meter: It mainly measures the flow through a mechanical transmission mechanism. The calibration strategy can be achieved by adjusting the gear ratio between the impeller shaft and the counter. For example, if the calibration signal is +3, it indicates replacing the gear set with a larger gear ratio (such as adjusting from 1:100 to 1:103), so that the counter records more pulses at the same flow rate, thereby compensating for negative errors; if the signal is -1, it indicates reducing the gear ratio (such as adjusting from 1:100 to 1:97) to compensate for positive errors.

[0063] Ultrasonic water meter: It measures the flow based on the time difference of ultrasonic waves propagating in the fluid. The calibration strategy can be achieved by adjusting the signal filtering parameters or the sound speed compensation coefficient. For example, when the calibration signal indicates a positive error, increasing the sound speed compensation coefficient reduces the calculated flow velocity value, thereby reducing the measurement result.

[0064] In the specific implementation, the parameters of the random forest model need to be adjusted according to actual needs, such as the number of decision trees (usually 50 - 200), the maximum depth (usually 5 - 20 layers), the minimum number of samples for splitting (usually 2 - 10), etc. Parameter adjustment can be completed by the cross-validation method, that is, dividing the validation set from the training data, testing the prediction accuracy of the model under different parameter combinations, and selecting the optimal parameter combination. In addition, the calibration signal generator can be deployed in edge computing devices or cloud servers, communicating with the front-end feature generation module and the back-end flow measurement device through real-time data interfaces to ensure the real-time and reliability of the calibration instructions.

[0065] Example 5: This example focuses on the adaptive calibration mechanism of the intelligent water meter system, which realizes intelligent adaptation to different water flow environments by dynamically adjusting parameter weights and calibration periods. The adaptive calibration mechanism includes an environmental feature extraction unit, a parameter weight adjustment unit, a calibration period decision unit, and a feedback verification unit, and each unit works together to optimize the real-time calibration performance of the system.

[0066] The environmental feature extraction unit continuously collects pipeline environmental parameters, including information such as temperature, pressure, water quality hardness, and pipeline material. These parameters are obtained through sensors distributed at different positions of the pipeline and converted into digital signals for transmission to the system. The temperature sensor uses a platinum resistance thermometer, whose resistance value changes linearly with temperature. By measuring the resistance value and converting it into a digital signal, accurate water temperature data can be obtained. The pressure sensor is based on the piezoresistive effect principle, converting pressure changes into electrical signal changes, and obtaining the pressure value after amplification and digitization processing. The water quality hardness sensor is based on the ion-selective electrode principle, determining the water quality hardness level by detecting the calcium and magnesium ion concentration in water. The pipeline material information is obtained through pre-configuration or self-learning methods. For example, at the system initialization stage, the user inputs the pipeline material type, or the system automatically identifies the pipeline material by analyzing the characteristics of the water flow signal.

[0067] The parameter weight adjustment unit dynamically adjusts the weights of each calibration parameter based on environmental features. This unit first preprocesses the environmental features, including data cleaning, normalization, and feature selection. Data cleaning removes outliers and noise in the data collected by the sensors. For example, the temperature and pressure data are smoothed through a sliding window filtering algorithm. Normalization maps feature values in different ranges to a unified interval, such as [0,1], to eliminate the influence of feature scale differences on weight adjustment. Feature selection screens out key features that have a greater impact on calibration parameters from the original environmental features. For example, in some scenarios, temperature and pressure are the main factors affecting flow measurement, while the impact of water quality hardness is relatively small, so the water quality hardness feature can be ignored.

[0068] Subsequently, the parameter weight adjustment unit inputs the preprocessed environmental feature vector into a pre-trained weight mapping model. The model can be expressed as: where, is the calibration parameter weight vector, is the preprocessed environmental feature vector, Represents the weight mapping function. The weight mapping function is trained through historical data, and its specific form can be selected according to the actual situation, such as a multi-layer perceptron or a decision tree. For example, when the environmental characteristics show that the water temperature is low, the model will increase the weight of the calibration parameters related to temperature to compensate for the impact of low temperature on flow measurement. In practical applications, the weight mapping model adopts an incremental learning method to continuously update the model parameters according to the newly collected data to adapt to environmental changes.

[0069] The calibration cycle decision unit dynamically adjusts the calibration cycle according to the water flow stability. The water flow stability is evaluated by analyzing the fluctuation amplitude and frequency of the flow signal. The specific steps are as follows: First, a sliding time window is set in the calibration cycle decision unit to intercept a section of the flow signal. Then, the standard deviation of the flow signal in this time period is calculated as an indicator of the fluctuation amplitude. The larger the standard deviation, the more drastic the fluctuation of the flow signal and the worse the water flow stability.

[0070] In order to analyze the frequency components of the flow signal, the calibration cycle decision unit performs Fourier transform on the intercepted flow signal. Fourier transform converts the time domain signal into a frequency domain signal, so that the strength of different frequency components in the signal can be determined. By analyzing the frequency domain signal, the main fluctuation frequency of the flow signal is determined. If the main fluctuation frequency is high, it means that the flow changes rapidly and the water flow state is unstable.

[0071] In the calibration cycle decision unit, the fluctuation amplitude threshold and the main fluctuation frequency threshold are preset. When the calculated fluctuation amplitude exceeds the fluctuation amplitude threshold, or the main fluctuation frequency exceeds the main fluctuation frequency threshold, it means that the water flow state is unstable. At this time, the calibration cycle decision unit will shorten the calibration cycle and increase the calibration frequency in order to track the changes in the water flow state in a timely manner and ensure the accuracy of flow measurement.

[0072] On the contrary, when the fluctuation amplitude and the main fluctuation frequency are both lower than their respective thresholds, it means that the water flow state is relatively stable. In the calibration cycle decision unit, a longer calibration cycle is set for a stable water flow state. This is because when the water flow is stable, the error of flow measurement changes slowly and frequent calibration is not required. Extending the calibration cycle can reduce the energy consumption of the system, reduce the impact of calibration operations on normal flow measurement, and also extend the service life of each component in the system.

[0073] The sliding time window length of the calibration period decision unit can be adjusted according to the actual application scenario. For example, in the residential water use scenario, the water flow fluctuation is relatively small and changes slowly. The sliding time window length can be set to 10 minutes, so that the change trend of the water flow can be captured more comprehensively. In the industrial water use scenario, the water flow fluctuation may be large and change rapidly. In order to detect the change of the water flow state in time, the sliding time window length can be set to 5 minutes. In this way, the calibration period decision unit can flexibly adjust the calibration period according to different application scenarios and water flow characteristics, while ensuring the measurement accuracy, optimizing the operation efficiency of the system.

[0074] The feedback verification unit verifies the calibration result in real time to ensure the effectiveness of the calibration. This unit compares the calibrated flow data with the reference standard and calculates the deviation value. If the deviation value exceeds the acceptable range, the re-calibration process is triggered and the calibration parameters are corrected. The reference standard can be obtained through a high-precision flowmeter or historical data statistical analysis. For example, in the industrial water use scenario, an electromagnetic flowmeter can be used as the reference standard to compare and verify the calibration result regularly to ensure the long-term stability of the system.

[0075] In the specific implementation, the sensors of the environmental feature extraction unit need to be calibrated regularly to ensure the data accuracy. The calibration process usually uses standard substances or known parameters for comparison, and adjusts the output value of the sensor to be consistent with the standard value. The weight mapping model of the parameter weight adjustment unit can adopt the incremental learning method and continuously update the model parameters according to the newly collected data. The sliding time window length of the calibration period decision unit can be adjusted according to the actual application scenario. For example, it is set to 10 minutes in the residential water use scenario and 5 minutes in the industrial water use scenario. The deviation threshold of the feedback verification unit can be set according to the measurement accuracy requirements, for example, set to ±2%.

[0076] The adaptive calibration mechanism dynamically adjusts the calibration parameters and period through real-time perception and analysis of environmental features, enabling the system to adapt to different water flow environments and working condition changes. This intelligent adaptation ability improves the pertinence and effectiveness of calibration, ensures the accuracy and reliability of flow measurement, and reduces the system energy consumption and maintenance cost. In practical applications, this mechanism can significantly improve the performance of intelligent water meters in complex environments and provide strong support for the accurate measurement and management of water resources.

[0077] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus.

[0078] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time correction system for water meter flow based on intelligent sensors, characterized in that, Comprising: A flow signal acquisition module, configured to obtain multi-dimensional vibration waveform data of a water meter pipeline through a multi-band sensor; A signal feature decomposition module, configured to perform time-frequency domain decomposition and energy spectrum feature extraction on the multi-dimensional vibration waveform data to obtain a set of local frequency domain feature tensors of the flow state; A frequency domain feature fusion module, configured to perform dynamic noise suppression and frequency band correlation optimization on the set of local frequency domain feature tensors of the flow state to obtain a set of locally optimized feature tensors of the flow state; An error compensation decision module, configured to perform global feature reconstruction based on frequency domain energy distribution on the set of locally optimized feature tensors of the flow state to generate a flow error compensation decision feature map; A correction instruction execution module, configured to generate a real-time flow correction control signal according to the flow error compensation decision feature map.

2. The real-time flow correction system of a water meter based on an intelligent sensor according to claim 1, wherein The signal feature decomposition module includes: A waveform data segmentation unit, configured to dynamically divide the multi-dimensional vibration waveform data according to a preset time window to obtain a set of waveform data segments; A feature extraction unit, configured to input each segment of data in the set of waveform data segments into a frequency domain encoder based on a multi-scale decomposition network to obtain the set of local frequency domain feature tensors of the flow state.

3. The real-time flow correction system of a water meter based on intelligent sensors according to claim 2, characterized in that The frequency domain feature fusion module includes: A frequency domain tensor expansion unit, configured to perform matrix expansion on the local frequency domain feature tensors of the flow state along the frequency band dimension to obtain a set of local frequency domain feature vectors of the flow state; A frequency band energy correlation calculation unit, configured to calculate the frequency domain correlation coefficient between any two feature vectors in the set of local frequency domain feature vectors of the flow state to generate a frequency band energy correlation matrix; A dynamic noise suppression unit, configured to correct the weight of the frequency band energy correlation matrix according to the frequency domain difference between adjacent time windows in the set of local frequency domain feature vectors of the flow state to obtain a frequency band energy constraint correlation matrix; A feature optimization unit, configured to perform cross-frequency band convolution fusion on the frequency band energy constraint correlation matrix and the set of local frequency domain feature vectors of the flow state to obtain the locally optimized feature tensor of the flow state.

4. The real-time correction system for water meter flow rate based on intelligent sensors according to claim 3, wherein The frequency band energy correlation calculation unit includes: A frequency domain projection sub-unit, configured to map each feature vector in the set of local frequency domain feature vectors of the flow state to an orthogonal frequency domain space to obtain a set of projected frequency domain feature vectors; A frequency domain correlation analysis sub-unit, configured to calculate the spectral energy similarity between any two projected feature vectors in the set of projected frequency domain feature vectors to generate the frequency band energy correlation matrix composed of multiple frequency band energy correlation values.

5. The real-time flow correction system of a water meter based on intelligent sensors according to claim 4, characterized in that The feature optimization unit is specifically implemented as: Performing dilated convolution processing on the frequency band energy constraint correlation matrix to obtain a frequency band energy constraint feature matrix; Inputting the set of local frequency domain feature vectors of the flow state and the frequency band energy constraint feature matrix into a bidirectional recursive coding network to obtain a set of local context feature vectors of the flow state; Performing matrix reconstruction on the set of local context feature vectors of the flow state to obtain the locally optimized feature tensor of the flow state.

6. The real-time flow correction system of a water meter based on an intelligent sensor according to claim 5, wherein The error compensation decision module includes: A local feature compression unit for performing mean pooling processing on each tensor in the set of local optimized feature tensors of the traffic state in the frequency domain dimension to obtain a set of local optimized feature vectors of the traffic state; A frequency-domain energy entropy calculation unit for calculating the energy entropy values of each feature vector in the set of local optimized feature vectors of the traffic state to generate a set of traffic dynamic energy entropy; A reference frequency band determination unit for selecting the local optimized feature vector of the traffic state corresponding to the minimum energy entropy value in the set of traffic dynamic energy entropy as the initial compensation reference vector; A dynamic compensation weight calculation unit for calculating the dynamic compensation weights of each feature vector according to the frequency-domain distance between each feature vector in the set of local optimized feature vectors of the traffic state and the initial compensation reference vector and the energy entropy values of each feature vector to generate a set of dynamic compensation weights; A global reconstruction unit for performing weighted superposition on the set of local optimized feature vectors of the traffic state by using the set of dynamic compensation weights to generate the traffic error compensation decision feature map.

7. The real-time flow rate correction system of a water meter based on intelligent sensors according to claim 6, wherein The frequency-domain energy entropy calculation unit is specifically implemented as: Calculating the median vector and the range vector of the local optimized feature vectors of the traffic state; Performing element-wise difference calculation between the local optimized feature vectors of the traffic state and the median vector, and performing a cube operation on the difference result to obtain a traffic feature energy difference vector; Calculating the overall mean of the traffic feature energy difference vector; Performing a ratio operation on the mean and the cube value of the range vector, and inputting the result into a normalization function to obtain the energy entropy value.

8. The real-time correction system for water meter flow based on intelligent sensors according to claim 7, characterized in that The dynamic compensation weight calculation unit is specifically implemented as: Multiplying the energy entropy value of the local optimized feature vector of the traffic state and the energy entropy value of the initial compensation reference vector by a first compensation coefficient to obtain a first dynamic compensation factor; Multiplying the absolute value of the Manhattan distance between the local optimized feature vector of the traffic state and the initial compensation reference vector by a second compensation coefficient to obtain a second dynamic compensation factor; Performing a product operation on the first dynamic compensation factor and the second dynamic compensation factor to obtain the dynamic compensation weight.

9. The real-time flow rate correction system of a water meter based on intelligent sensors according to claim 8, characterized in that The correction instruction execution module is specifically implemented as: Inputting the traffic error compensation decision feature map into a correction signal generator based on a random forest to obtain the real-time traffic correction control signal, and the signal is used to indicate the calibration strategy of the flow measurement device.

10. A real-time correction method for water meter flow based on intelligent sensors, characterized in that, Including: Collecting multi-dimensional vibration waveform data of the water meter pipeline through a multi-band sensor; Performing time-frequency domain decomposition and energy spectrum feature extraction on the multi-dimensional vibration waveform data to obtain a set of local frequency domain feature tensors of the traffic state; Performing dynamic noise suppression and frequency band correlation optimization on the set of local frequency domain feature tensors of the traffic state to obtain a set of local optimized feature tensors of the traffic state; Performing global feature reconstruction based on the frequency-domain energy distribution on the set of local optimized feature tensors of the traffic state to generate a traffic error compensation decision feature map; Generating a real-time traffic correction control signal according to the traffic error compensation decision feature map.

Images