Self-optimization ultrasonic flow monitoring method based on cloud center neural network

Abstract

The invention provides a self-optimization ultrasonic flow monitoring method based on a cloud center neural network. The self-optimization ultrasonic flow monitoring method comprises the following sub-steps that S1, original data are collected and subjected to lightweight preprocessing; s2, the cloud center receives the data and performs data storage and feature supplementation; s3, training and optimizing a cloud center neural network model; s4, issuing the optimal neural network model and performing local flow calculation; s5, optimizing the model based on an error result; according to the method, the deep neural network of the cloud center is introduced to carry out flow prediction, data are uploaded to the cloud end, the ultrasonic flow meter is not limited by the computing power of equipment any more, the computing precision is improved, the adaptability to complex flow states and working condition changes is improved, and therefore measurement errors are reduced; data are stored through the cloud center and are associated with the global knowledge base, and data and experience are shared among different devices, so that the dependence on human experience is reduced, and the problem that the devices depend on local experience is solved.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic technology for fluid flow measurement, and in particular to a self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network. Background Technology

[0002] Ultrasonic flow meters calculate flow rate by detecting the time difference (or phase difference) of ultrasonic waves propagating in a fluid along and against the flow, combined with fluid sound velocity, pipe parameters, etc., and are widely used in liquid flow monitoring. Ultrasonic flow meters typically rely on a fixed physical model of an embedded device for flow estimation; however, due to the limitations of the device's hardware resources, the flow calculation has the following drawbacks: 1. Limited local computing power leads to large errors: The computing power and storage resources of embedded devices are limited by hardware, making it impossible to run complex nonlinear models, such as advanced technologies like deep neural networks. They rely on simple linear correction or lookup table methods, and their inability to adapt to complex flow conditions leads to large measurement errors. 2. Data silos and experience dependence: Traditional local embedded devices cannot effectively share global data and can only rely on the local experience of a single device to make parameter corrections, resulting in poor generalization ability and inability to adapt to complex and ever-changing actual working conditions. 3. Dynamic optimization lag: Traditional adaptive methods, such as local parameter adjustment based on Kalman filtering, have a slow response speed and cannot learn long-term patterns using historical data, making it difficult for traditional flow meters to make quick and accurate adjustments when faced with sudden changes in operating conditions. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: The self-optimizing ultrasonic flow monitoring method based on cloud-centric neural networks includes the following sub-steps: S1: Collect raw data and perform lightweight preprocessing; S11: Collect raw data according to the set sampling frequency; Multiple ultrasonic flow meters are pre-deployed on the pipeline. Each ultrasonic flow meter has a unique device ID, each pipeline has a unique pipe code, and each pipeline has a corresponding initial calibration value. The ultrasonic flow meter synchronously collects raw data through a sensor group at a set sampling frequency; the sensor group includes an ultrasonic flow sensor, a temperature sensor, a pressure sensor, etc.; the sampling frequency is ≥100Hz / S. The raw data includes ultrasonic propagation data, environmental parameters, equipment status parameters, basic pipeline data, timestamps, etc. The ultrasonic propagation data includes downstream flight time, upstream flight time, echo signal intensity, and flow direction; The environmental parameters include fluid temperature, fluid pressure, pipeline vibration noise, etc. The device status parameters include transducer coupling status, local battery voltage, etc. The basic pipeline data includes pipeline cross-sectional direction, pipeline code, initial calibration value, original inner diameter of the pipeline, pipeline condition, etc. S12: Perform lightweight preprocessing on the collected raw data; The ultrasonic flow meter performs lightweight preprocessing on the collected raw data, including noise reduction, normalization, feature extraction, and data compression. High-frequency noise in the original data is removed using denoising techniques such as moving average filtering; the original data is normalized using methods such as linear scaling. The feature extraction includes extracting basic features such as countercurrent time difference and temperature-pressure ratio; Specifically, the time difference of the countercurrent is denoted as... Downstream flight time is Counter-current flight time , Let the fluid temperature be T, the fluid pressure be P, and the temperature-pressure ratio be T / P. The preprocessed raw data is compressed, and after compression, it is packaged together with the extracted basic features, device ID, pipeline code, and timestamp and uploaded to the cloud center.

[0005] S2: The cloud center receives data and performs data storage and feature supplementation; The cloud center receives data uploaded by multiple ultrasonic flow meters and categorizes and stores the received data according to device ID and timestamp; The cloud center is associated with a global knowledge base, which includes pipeline design drawings, fluid type knowledge base, and historical data of ultrasonic flow meters under different operating conditions. Based on the information in the global knowledge base, the cloud center supplements the received data with features, adding the flow velocity range, temperature / pressure range, pipeline status flow velocity value, and flow rate value of the ultrasonic flow meter under normal operating conditions, high temperature operating conditions, high pressure operating conditions, and abnormal operating conditions to the classified and stored data, forming a multi-dimensional feature vector.

[0006] S3: Training and optimization of the cloud center neural network model; Includes the following sub-steps: S31: Deploy the neural network model and train it; A neural network model is deployed in the cloud center. The neural network model includes an input layer, a first fully connected hidden layer, an attention mechanism hidden layer, a second fully connected hidden layer, and an output layer. The multidimensional feature vector is used as input features into the neural network model. Each dimension of the multidimensional feature vector is a node. The input layer of the neural network model receives the input features and passes them to the first fully connected hidden layer. The first fully connected hidden layer contains multiple nodes, each with a corresponding weight, forming a weight matrix. The first fully connected hidden layer receives the input features from the input layer, multiplies the value of each node in the input features sequentially with the weight matrix, and adds it to the bias term preset by the input layer to obtain a weighted value. The weighted value is then non-linearly transformed using the ReLU activation function. If the weighted value is negative, it is transformed to zero; if it is positive, it remains unchanged. The transformed multi-dimensional vector is output, with each dimension corresponding to a feature. The multidimensional vector is passed to the hidden layer of the attention mechanism, which focuses on the key features in the multidimensional vector. The attention mechanism performs weighted processing on the key features and outputs a weighted multidimensional vector. The weighted multidimensional vector is passed to the second fully connected hidden layer. The weight matrix in the second fully connected hidden layer is multiplied with the weighted multidimensional vector and added to the bias term. The weighted value is then nonlinearly transformed by the ReLU activation function. If the weighted value is negative, it is transformed to zero; if it is positive, it remains unchanged. The final multidimensional vector is then output. The final multidimensional vector is passed to the output layer, which performs weighted processing on the final multidimensional vector through a linear activation function. The weighted output value is the predicted flow rate value, thus obtaining the trained neural network model. Furthermore, the weight matrix and bias terms in the second fully connected hidden layer are different from those in the first fully connected hidden layer; S32: Optimize the trained neural network model to obtain the optimal neural network model; Abnormal operating conditions are assigned higher weights than normal operating conditions using the weighted mean square error method. These abnormal operating conditions include sudden changes in flight time difference or waveform anomalies. The predicted flow velocity is output, and the error between the predicted and actual flow velocity values ​​is minimized using the weighted mean square error loss function. During training, the neural network model is optimized through dynamic weight adjustment, regularization, and other methods until the error is within a preset range, thus obtaining the optimal neural network model.

[0007] S4: Distribute the optimal neural network model and perform local traffic calculation; Includes the following sub-steps: S41: The cloud center sends the trained optimal neural network model to the ultrasonic flow meter; The cloud center pre-sets specific conditions such as the prediction error exceeding a set threshold, including two distribution methods: periodic distribution and triggered distribution. The periodic distribution refers to the distribution of the optimal neural network model to the ultrasonic flow meter according to a set cycle; The trigger-based sending means that when a preset specific condition is detected, the optimal neural network model is immediately sent to the ultrasonic flow meter. S42: Output the predicted flow rate value and obtain the predicted flow rate value based on the predicted flow rate value; Based on the method described in step S2, a real-time multidimensional feature vector is obtained. The multidimensional feature vector is then input into the optimal neural network model issued by the cloud center, and the optimal neural network model outputs the predicted flow rate value. The predicted flow rate is calculated by combining the cross-sectional area of ​​the pipe; The specific calculation method is as follows: ; in, For predicting flow velocity, A is the cross-sectional area of ​​the pipe. The angle between the fluid flow direction and the normal to the pipe cross-section. To predict flow rates; The cross-sectional area A of the pipe is the initial calibration value of the pipe plus the effective correction amount for the inner diameter of the pipe; ; Where v is the flow velocity vector and n is the normal vector of the pipe cross-section. Let |v| and |n| be the dot product of the velocity vector and the normal vector of the pipe cross section, where |v| and |n| are the magnitudes of the velocity vector and the normal vector of the pipe cross section, respectively. The flow direction and predicted flow velocity value obtained by the ultrasonic flow meter constitute the flow velocity vector; the cross-sectional normal of the pipe is perpendicular to the cross-section of the pipe. The cross-sectional normal of the pipe is set as a unit vector, that is, the magnitude is 1, to form the cross-sectional normal of the pipe. Furthermore, in the event of a malfunction or network interruption, the local ultrasonic calculation method is switched: the current flow velocity value is obtained through the time-of-flight method; and through... The formula yields the current flow rate; at this point... This represents the current flow rate.

[0008] S5: Optimize the optimal neural network model based on the error results; Includes the following sub-steps: S51: Perform error analysis and obtain key training samples; The cloud center pre-sets a difference threshold, a reference flow velocity value, and a reference flow rate value. The cloud center calculates the difference between the predicted flow velocity value and the reference flow velocity value, and the predicted flow rate value and the reference flow rate value. If any difference exceeds the set difference threshold, the cloud center combines the raw data collected by sensors and ultrasonic transducers to identify high-error operating conditions. These high-error operating conditions include low-temperature and low-flow-velocity scenarios in specific areas. Corresponding data is extracted from these high-error operating conditions as key training samples. These corresponding data include pipe cross-sectional area, temperature, etc. S52: Incrementally train the optimal neural network model; The optimal neural network model is retrained using full data, which includes the original data, sample data from the global database, and filtered data of the optimal neural network model under various working conditions. The optimized neural network model is then used for subsequent calculations to achieve dynamic self-optimization of the neural network model.

[0009] Preferably, for scenarios with privacy protection requirements, each ultrasonic flow meter performs lightweight model training. The lightweight model includes Long Short-Term Memory (LSTM) networks, one-dimensional convolutional neural networks (1D-CNN), etc., and the specific method is as follows: The image data is obtained by taking pictures with a camera and based on image recognition technology. The original data and image data after lightweight preprocessing in step S12 are augmented by Gaussian noise, time offset, amplitude scaling and other methods. The lightweight model is trained using processed time-series data, and INT8 quantization is simulated during training to reduce the model's storage and computational load without significantly reducing accuracy. The updated lightweight model weights (cross-sectional area, temperature parameters, etc.) are uploaded to the cloud center for aggregation via encrypted communication to ensure data privacy and security.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This method introduces a deep neural network in the cloud center for flow prediction, which has stronger computing power and more complex model training capabilities. By uploading data to the cloud, the ultrasonic flow meter is no longer limited by the device's computing power, improving calculation accuracy and adaptability to complex flow states and changes in operating conditions, thereby reducing measurement errors. Data is stored in the cloud center and linked to the global knowledge base, which can retrieve information related to pipelines, equipment and environment to supplement data features. Data and experience are shared among different devices, which enables flow meters to better adapt to different pipeline or operating conditions, enhances the intelligence and adaptability of the equipment, reduces reliance on human experience, and solves the problem of equipment relying on local experience. It employs a cloud-centric neural network model for learning and dynamic optimization, repeatedly optimizing through training and testing sets to adjust the prediction model in real time. By using a weighted mean square error method to assign higher weights to abnormal operating conditions, it can quickly respond to sudden changes in operating conditions. This enables the flowmeter to make faster and more accurate adjustments when faced with sudden changes in operating conditions, eliminating the need for manual calibration of each unit. It avoids the lag of traditional adaptive methods and improves the dynamic adaptability and response speed of the equipment. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of the self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network according to the present invention. Detailed Implementation

[0012] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0013] like Figure 1 As shown, the self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network includes the following sub-steps: S1: Collect raw data and perform lightweight preprocessing; S11: Collect raw data according to the set sampling frequency; Multiple ultrasonic flow meters are pre-deployed on the pipeline. Each ultrasonic flow meter has a unique device ID, each pipeline has a unique pipe code, and each pipeline has a corresponding initial calibration value. The ultrasonic flow meter synchronously collects raw data through a sensor group at a set sampling frequency; the sensor group includes an ultrasonic flow sensor, a temperature sensor, a pressure sensor, etc.; the sampling frequency is ≥100Hz / S. The raw data includes ultrasonic propagation data, environmental parameters, equipment status parameters, basic pipeline data, timestamps, etc. The ultrasonic propagation data includes downstream flight time, upstream flight time, echo signal strength, and flow direction; the signal-to-noise ratio is evaluated using the echo signal strength. The environmental parameters include fluid temperature, fluid pressure, pipeline vibration noise, etc. The device status parameters include transducer coupling status, local battery voltage, etc. The basic pipeline data includes pipeline cross-sectional direction, pipeline code, initial calibration value, original inner diameter of the pipeline, pipeline condition, etc. Specifically, the ultrasonic flow sensor collects data such as downstream flight time, upstream flight time, and echo signal strength through an ultrasonic transducer; it collects fluid temperature through a temperature sensor; it obtains fluid pressure through a pressure sensor; it collects pipeline vibration noise through a vibration sensor; it monitors the coupling status of the ultrasonic transducer through sensors (judging the coupling effect through the signal attenuation coefficient); and it collects the local battery voltage through a voltage sensor (which can indirectly reflect the health of the equipment based on the battery voltage). Furthermore, the timestamp of the acquired data is generated by the real-time clock (RTC) built into the ultrasonic flow meter; S12: Perform lightweight preprocessing on the collected raw data; The ultrasonic flow meter performs lightweight preprocessing on the collected raw data, including noise reduction, normalization, feature extraction, and data compression. High-frequency noise in the original data is removed using denoising techniques such as moving average filtering; the original data is normalized using methods such as linear scaling. The feature extraction includes extracting basic features such as countercurrent time difference and temperature-pressure ratio; feature extraction can help reduce model computational complexity and improve prediction accuracy. Specifically, the time difference of the countercurrent is denoted as... Downstream flight time is Counter-current flight time , Let the fluid temperature be T, the fluid pressure be P, and the temperature-pressure ratio be T / P. The preprocessed raw data is compressed, and after compression, it is packaged together with the extracted basic features, device ID, pipeline code, and timestamp and uploaded to the cloud center.

[0014] S2: The cloud center receives data and performs data storage and feature supplementation; The cloud center receives data uploaded by multiple ultrasonic flow meters and categorizes and stores the received data according to device ID and timestamp; The cloud center is associated with a global knowledge base, which includes pipeline design drawings, fluid type knowledge base, historical data of ultrasonic flow meters under different operating conditions, etc.; the cloud center supplements the received data with features based on the information in the global knowledge base. Specifically, the cloud center obtains the effective correction amount of the pipe inner diameter and regional environmental characteristics from the global knowledge base based on the pipe code; the cloud center obtains the equipment lifecycle characteristics from the global knowledge base based on the equipment ID; The regional environmental characteristics include regional seasonal temperature fluctuation patterns and flow correlation between adjacent flow meters; The device lifecycle characteristics include the ultrasonic flow meter's operating time and historical calibration records.

[0015] The flow velocity range, temperature / pressure range, flow velocity value and flow rate value of the ultrasonic flow meter under normal operating conditions, high temperature operating conditions, high pressure operating conditions and abnormal operating conditions, as well as the effective correction amount of the pipe inner diameter, regional environmental characteristics and equipment life cycle characteristics, are supplemented into the classified and stored data to form a multi-dimensional feature vector.

[0016] S3: Training and optimization of the cloud center neural network model; Includes the following sub-steps: S31: Deploy the neural network model and train it; A neural network model is deployed in the cloud center. The neural network model includes an input layer, a first fully connected hidden layer, an attention mechanism hidden layer, a second fully connected hidden layer, and an output layer. The multidimensional feature vector is used as input features into the neural network model. Each dimension of the multidimensional feature vector is a node. The input layer of the neural network model receives the input features and passes them to the first fully connected hidden layer. The first fully connected hidden layer contains multiple nodes, each with a corresponding weight, forming a weight matrix. The first fully connected hidden layer receives the input features from the input layer, multiplies the value of each node in the input features sequentially with the weight matrix, and adds it to the bias term preset by the input layer to obtain a weighted value. The weighted value is then non-linearly transformed using the ReLU activation function. If the weighted value is negative, it is transformed to zero; if it is positive, it remains unchanged. The transformed multi-dimensional vector is output, with each dimension corresponding to a feature. The multidimensional vector is passed to the hidden layer of the attention mechanism, which focuses on the key features in the multidimensional vector. The attention mechanism performs weighted processing on the key features and outputs a weighted multidimensional vector. The weighted multidimensional vector is passed to the second fully connected hidden layer. The weight matrix in the second fully connected hidden layer is multiplied with the weighted multidimensional vector and added to the bias term. The weighted value is then nonlinearly transformed by the ReLU activation function. If the weighted value is negative, it is transformed to zero; if it is positive, it remains unchanged. The final multidimensional vector is then output. The final multidimensional vector is passed to the output layer, which performs weighted processing on the final multidimensional vector through a linear activation function. The weighted output value is the predicted flow rate value, thus obtaining the trained neural network model. Furthermore, the weight matrix and bias terms in the second fully connected hidden layer are different from those in the first fully connected hidden layer; S32: Optimize the trained neural network model to obtain the optimal neural network model; Abnormal operating conditions are assigned higher weights than normal operating conditions using the weighted mean square error method. These abnormal operating conditions include sudden changes in flight time difference or waveform anomalies. The predicted flow velocity is output, and the error between the predicted and actual flow velocity values ​​is minimized using the weighted mean square error loss function. During training, the neural network model is optimized through dynamic weight adjustment, regularization, and other methods until the error is within a preset range, thus obtaining the optimal neural network model.

[0017] This method only requires lightweight data acquisition and feature extraction (low computing power requirements), while complex calculations are handled by the cloud center, reducing hardware costs.

[0018] S4: Distribute the optimal neural network model and perform local traffic calculation; Includes the following sub-steps: S41: The cloud center sends the trained optimal neural network model to the ultrasonic flow meter; The cloud center pre-sets specific conditions, such as the prediction error exceeding a set threshold, among which... It includes two distribution methods: scheduled distribution and triggered distribution; The periodic distribution refers to the distribution of the optimal neural network model to the ultrasonic flow meter according to a set cycle; The trigger-based sending means that when a preset specific condition is detected, the optimal neural network model is immediately sent to the ultrasonic flow meter. It combines two model update methods: periodic and triggered. When a specific condition is detected, such as the prediction error exceeding the limit, the optimal model can be quickly sent to the flow meter for real-time adjustment, which improves the flow meter's response speed. Especially in the event of equipment abnormality or network interruption, it can quickly switch to local calculation mode. It improves real-time performance and flexibility, and can quickly correct and adapt to new situations when equipment failure or abnormal operating conditions occur, thereby ensuring that the flow meter always maintains high-precision calculation capabilities. S42: Output the predicted flow rate value and obtain the predicted flow rate value based on the predicted flow rate value; Based on the method described in step S2, a real-time multidimensional feature vector is obtained. The multidimensional feature vector is then input into the optimal neural network model issued by the cloud center, and the optimal neural network model outputs the predicted flow rate value. The predicted flow rate is calculated by combining the cross-sectional area of ​​the pipe; The specific calculation method is as follows: ; in, For predicting flow velocity, A is the cross-sectional area of ​​the pipe. The angle between the fluid flow direction and the normal to the pipe cross-section. To predict flow rates; The cross-sectional area A of the pipe is the initial calibration value of the pipe plus the effective correction amount for the inner diameter of the pipe; ; Where v is the flow velocity vector and n is the normal vector of the pipe cross-section. Let |v| and |n| be the dot product of the velocity vector and the normal vector of the pipe cross section, where |v| and |n| are the magnitudes of the velocity vector and the normal vector of the pipe cross section, respectively. The flow direction and predicted flow velocity value obtained by the ultrasonic flow meter constitute the flow velocity vector; the cross-sectional normal of the pipe is perpendicular to the cross-section of the pipe. The cross-sectional normal of the pipe is set as a unit vector, that is, the magnitude is 1, to form the cross-sectional normal of the pipe. Furthermore, tasks and triggering conditions are set in advance; the triggering conditions are a fault or network interruption, and the task is to switch the local ultrasound calculation mode. In case of failure or network interruption, switch the local ultrasonic calculation method: obtain the current flow velocity value through the time-of-flight difference method; and through... The formula yields the current flow rate; at this point... This represents the current flow rate.

[0019] S5: Optimize the optimal neural network model based on the error results; Includes the following sub-steps: S51: Perform error analysis and obtain key training samples; The cloud center pre-sets a difference threshold, a reference flow velocity value, and a reference flow rate value. The cloud center calculates the difference between the predicted flow velocity value and the reference flow velocity value, and the predicted flow rate value and the reference flow rate value. If any difference exceeds the set difference threshold, the cloud center combines the raw data collected by sensors and ultrasonic transducers to identify high-error operating conditions. These high-error operating conditions include low-temperature and low-flow-velocity scenarios in specific areas. Corresponding data is extracted from these high-error operating conditions as key training samples. These corresponding data include pipe cross-sectional area, temperature, etc. S52: Incrementally train the optimal neural network model; The optimal neural network model is retrained using full data, which includes the original data, sample data from the global database, and filtered data of the optimal neural network model under various working conditions. The optimized neural network model is then used for subsequent calculations to achieve dynamic self-optimization of the neural network model.

[0020] Preferably, for scenarios with privacy protection requirements, each ultrasonic flow meter performs lightweight model training. The lightweight model includes Long Short-Term Memory (LSTM) networks, one-dimensional convolutional neural networks (1D-CNN), etc., and the specific method is as follows: The image data is obtained by taking pictures with a camera and based on image recognition technology. The original data and image data after lightweight preprocessing in step S12 are augmented by Gaussian noise, time offset, amplitude scaling and other methods. Specifically, noise is added to the time-series signal using Gaussian noise to simulate the error of the original data; different sampling times or environmental changes are simulated by using time offset; and data under different measurement conditions are simulated by using amplitude scaling. Normalize the time series data; for long-term time series signals (such as large bubble or empty tube state), divide the time series signal into fixed-length segments by sliding window to reduce the computational complexity of lightweight models and improve training efficiency. The lightweight model is trained using processed time-series data, and INT8 quantization is simulated during training to reduce the model's storage and computational load without significantly reducing accuracy. The updated lightweight model (cross-sectional area, temperature parameters, etc.) weights are uploaded to the cloud center for aggregation via encrypted communication to ensure data privacy and security. By querying and comparing similar signals under highly similar operating conditions (such as the same fluid type, pipe material, etc., which lead to slight differences in flow rate and different fault modes), the operating condition type that adapts to slight parameter changes is detected and analyzed through database query and comparison, and different filtering algorithms are selected to adapt the model. The flow meter model parameters are then queried and matched through database query and model matching, and data is shared in the cloud to accelerate the model convergence of new equipment.

[0021] In scenarios with high privacy protection requirements, flow meters can perform lightweight local model training (such as LSTM, 1D-CNN), reducing reliance on cloud resources; they can also enhance and normalize locally trained data to reduce computational complexity, and perform INT8 quantization to reduce model storage and computation without reducing accuracy; this improves data privacy and security, reduces the computational burden on devices, improves model training efficiency, and ensures performance.

[0022] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network, characterized in that: Includes the following steps: S1: Collect raw data and perform lightweight preprocessing; S11: Collect raw data according to the set sampling frequency; Multiple ultrasonic flow meters are pre-deployed on the pipeline at the site. The ultrasonic flow meters synchronously collect raw data through the sensor group according to the set sampling frequency. S12: Perform lightweight preprocessing on the collected raw data; The raw data after lightweight preprocessing is compressed, and after compression, it is packaged with the extracted basic features and uploaded to the cloud center. S2: The cloud center receives data and performs data storage and feature supplementation; The cloud center is linked to the global knowledge base, and supplements the received data with features based on the information in the global knowledge base to form a multi-dimensional feature vector. S3: Training and optimization of the cloud center neural network model; S31: Deploy the neural network model and train it; S32: Optimize the trained neural network model to obtain the optimal neural network model; S4: Distribute the optimal neural network model and perform local traffic calculation; S41: The cloud center sends the trained optimal neural network model to the ultrasonic flow meter; S42: Output the predicted flow rate value and obtain the predicted flow rate value based on the predicted flow rate value; S5: Optimize the optimal neural network model based on the error results; S51: Perform error analysis and obtain key training samples; S52: Incrementally train the optimal neural network model.

2. The self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network as described in claim 1, characterized in that: The specific details of step S1 are as follows: Multiple ultrasonic flow meters are pre-deployed on the pipeline. Each ultrasonic flow meter has a unique device ID, each pipeline has a unique pipe code, and each pipeline has a corresponding initial calibration value. The ultrasonic flow meters synchronously collect raw data through the sensor group according to the set sampling frequency. The raw data includes downstream flight time, upstream flight time, flow direction, fluid temperature, fluid pressure, pipe cross-sectional direction, pipe code, initial calibration value, original pipe inner diameter, and timestamp. Basic characteristics include countercurrent time difference and temperature-pressure ratio; Let the time difference of the countercurrent be . Downstream flight time is Counter-current flight time , Let the fluid temperature be T, the fluid pressure be P, and the temperature-pressure ratio be T / P.

3. The self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network as described in claim 1, characterized in that: In step S2, the cloud center receives data uploaded by multiple ultrasonic flow meters and classifies and stores the data collected by the ultrasonic flow meters according to the device ID and timestamp. The cloud center is associated with a global knowledge base. Based on the information in the global knowledge base, the cloud center supplements the received data with features, including the effective correction amount of the pipe inner diameter, regional environmental characteristics, equipment life cycle characteristics, and historical data of the ultrasonic flow meter under different operating conditions; thus forming a multi-dimensional feature vector.

4. The self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network as described in claim 1, characterized in that: The specific details of step S3 are as follows: S31: Deploy the neural network model and train it; A neural network model is deployed in the cloud center. The neural network model includes an input layer, a first fully connected hidden layer, an attention mechanism hidden layer, a second fully connected hidden layer, and an output layer. The multidimensional feature vector is used as input features into the neural network model. Each dimension of the multidimensional feature vector is a node. The input layer of the neural network model receives the input features and passes them to the first fully connected hidden layer. The first fully connected hidden layer contains multiple nodes, each with a corresponding weight, forming a weight matrix. The first fully connected hidden layer receives the input features from the input layer, multiplies the value of each node in the input features sequentially with the weight matrix, and adds it to the bias term preset by the input layer to obtain a weighted value. The weighted value is then non-linearly transformed using the ReLU activation function. If the weighted value is negative, it is transformed to zero; if it is positive, it remains unchanged. The transformed multi-dimensional vector is output, with each dimension corresponding to a feature. The multidimensional vector is passed to the attention mechanism hidden layer, and the attention mechanism focuses on the key features in the multidimensional vector. The attention mechanism performs weighted processing on key features and outputs a weighted multidimensional vector; The weighted multidimensional vector is passed to the second fully connected hidden layer. The weight matrix in the second fully connected hidden layer is multiplied with the weighted multidimensional vector and added to the bias term. The weighted value is then nonlinearly transformed by the ReLU activation function. If the weighted value is negative, it is transformed to zero; if it is positive, it remains unchanged. The final multidimensional vector is then output. The final multidimensional vector is passed to the output layer, which performs weighted processing on the final multidimensional vector through a linear activation function. The weighted output value is the predicted flow rate value, thus obtaining the trained neural network model. S32: Optimize the trained neural network model to obtain the optimal neural network model; The abnormal operating conditions are assigned higher weights than the normal operating conditions by using the weighted mean square error method; the error between the predicted and actual flow velocity values ​​is minimized by using the weighted MSE loss function; and the neural network model is optimized during training by dynamic weight adjustment and regularization until the error is within the preset range, thus obtaining the optimal neural network model.

5. The self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network as described in claim 1, characterized in that: Step S41 includes two distribution methods: periodic distribution and trigger-based distribution. The cloud center pre-sets specific conditions under which the prediction error exceeds a set threshold. When no specific conditions are detected, the optimal neural network model is sent to the ultrasonic flow meter according to the set cycle; When a preset specific condition is detected, the optimal neural network model is immediately sent to the ultrasonic flow meter.

6. The self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network as described in claim 1, characterized in that: In step S42, the specific calculation methods for the predicted flow velocity and predicted flow rate are as follows: Based on the method described in step S2, a real-time multi-dimensional feature vector is obtained. The multi-dimensional feature vector is then input into the optimal neural network model issued by the cloud center, and the optimal neural network model outputs the predicted flow rate value. The predicted flow rate is calculated by combining the cross-sectional area of ​​the pipe; The specific calculation method is as follows: ; in, For predicting flow velocity, A is the cross-sectional area of ​​the pipe. The angle between the fluid flow direction and the normal to the pipe cross-section. To predict flow rates; The cross-sectional area A of the pipe is the initial calibration value of the pipe plus the effective correction amount for the inner diameter of the pipe; ； Where v is the flow velocity vector and n is the normal vector of the pipe cross-section. Let |v| and |n| be the dot product of the velocity vector and the normal vector of the pipe cross section, where |v| and |n| are the magnitudes of the velocity vector and the normal vector of the pipe cross section, respectively. The flow direction and predicted flow velocity value obtained by the ultrasonic flow meter constitute the flow velocity vector; the cross-sectional normal of the pipe is perpendicular to the cross-section of the pipe. The cross-sectional normal of the pipe is set as a unit vector, that is, the magnitude is 1, which constitutes the cross-sectional normal of the pipe.

7. The self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network as described in claim 1, characterized in that: The specific details of step S5 are as follows: The cloud center pre-sets a difference threshold, a reference flow rate, and a reference flow rate. The cloud center calculates the difference between the predicted flow rate and the reference flow rate, and between the predicted flow rate and the reference flow rate. If any difference exceeds the set difference threshold, the cloud center combines the raw data collected by the sensors and ultrasonic transducers to identify high-error operating conditions and extracts the corresponding data from the high-error operating conditions as key training samples. S52: Incrementally train the optimal neural network model; The optimal neural network model is retrained using the full dataset, which includes the original data, data from the global database, and filtered data of the optimal neural network model under various working conditions. The optimized neural network model is then used for subsequent calculations, enabling dynamic self-optimization of the neural network model.

8. The self-optimizing ultrasonic flow monitoring method based on a cloud-centric neural network as described in claim 1, characterized in that: For scenarios requiring privacy protection, each ultrasonic flow meter performs lightweight model training, as follows: The image data is obtained by taking pictures with a camera and using image recognition technology. The original data and image data after lightweight preprocessing in step S12 are augmented by Gaussian noise, time offset and amplitude scaling. The lightweight model is trained using the processed time-series data, and INT8 quantization is simulated during training to reduce the model's storage and computational load without significantly reducing accuracy. The updated lightweight model weights are then uploaded to the cloud center for aggregation via encrypted communication.

Images