Gas-liquid two-phase flow metering method and system based on screw pump working condition self-recognition

Abstract

The invention provides a gas-liquid two-phase flow metering method and system based on screw pump working condition self-recognition, and relates to the field of industrial flow metering. The method comprises the following steps: acquiring operation parameters of the screw pump in real time to obtain an original data flow; performing preprocessing and feature extraction on the original data stream to obtain a feature vector; and obtaining a historical original data stream to construct a multi-level fusion strategy. And performing working condition identification on the feature vectors by using a multi-stage fusion strategy to obtain current working conditions including pure liquid phase, gas-liquid two-phase, cavitation, underload or overload states. And a corresponding model is adaptively selected according to the current working condition, and the gas-liquid two-phase flow, the gas-liquid ratio and the pump efficiency are obtained through mechanism model calculation and data driving model thrust. And based on the gas-liquid two-phase flow, the gas-liquid ratio and the pump efficiency, the stable working condition time period or external calibration data is used for calibration optimization. The technical problems that measurement adaptability is poor, working condition recognition accuracy is insufficient, and accuracy and stability of the screw pump in a gas-liquid two-phase flow scene are difficult to meet are solved.

Description

Technical Field

[0001] This application relates to the field of industrial flow measurement, and in particular to a method and system for measuring gas-liquid two-phase flow based on the self-identification of screw pump operating conditions. Background Technology

[0002] In industrial sectors such as oil extraction and chemical production, screw pumps are core equipment for gas-liquid two-phase flow transportation. The accurate measurement of medium flow rate, gas-liquid ratio, and pump efficiency directly impacts production control and economic benefits. Currently, the mainstream flow measurement methods in the industry mainly rely on external flow meters (such as orifice plate flow meters and turbine flow meters) or single-parameter derivation calculations. External flow meters, based on the traditional single-phase flow measurement principle, calculate flow rate by detecting single physical quantities such as differential pressure and flow velocity. Single-parameter derivation methods, on the other hand, often indirectly estimate key parameters such as torque and flow rate using limited data such as motor current and speed. This leads to instability in flow coefficients in gas-liquid two-phase flow scenarios due to large differences in gas and liquid phase densities and drastic interface fluctuations. Measurement errors increase significantly when the gas phase volume fraction exceeds 20%, making it unsuitable for complex operating conditions with high gas content and varied flow patterns. Furthermore, single-parameter derivation methods are affected by factors such as medium viscosity and mechanical wear, resulting in large deviations between estimated results and actual values, making it difficult to meet the measurement needs under multiple operating conditions. Furthermore, during use, it is impossible to effectively distinguish between different operating states such as pure liquid phase, gas-liquid two-phase, cavitation, underload or overload, and it is especially easy to confuse cavitation with high gas content conditions, leading to inaccurate selection of metering models and thus the risk of equipment damage. Summary of the Invention

[0003] This application provides a gas-liquid two-phase flow metering method and system based on screw pump operating condition self-identification, which solves the technical problems of poor measurement adaptability and insufficient accuracy of operating condition identification in the existing technology, making it difficult to meet the requirements of accurate and stable measurement in the gas-liquid two-phase flow scenario of screw pump.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a gas-liquid two-phase flow measurement method based on screw pump operating condition self-identification includes: real-time acquisition of screw pump operating parameters to obtain a raw data stream, including speed, torque, electrical power, inlet pressure, outlet pressure, temperature, and vibration signals. The raw data stream is preprocessed and feature extracted to obtain feature vectors, which include time-domain features, frequency-domain features, and derived physical features. A multi-level fusion strategy is constructed by acquiring historical raw data streams, including a cavitation risk assessment layer, a multi-sensor fusion classification layer, and a logic verification layer. The multi-level fusion strategy is used to identify the operating conditions of the feature vectors, obtaining the current operating conditions, including pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload states. The corresponding model is adaptively selected based on the current operating condition, and the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two-phase flow are obtained through mechanism model calculation and data-driven model thrust. Based on the gas-liquid two-phase flow rate, gas-liquid ratio, and pump efficiency, the internal parameters of the fusion mechanism model and the data-driven model are calibrated and optimized using stable operating conditions or external calibration data.

[0005] In conjunction with the first aspect mentioned above, one possible implementation involves preprocessing and feature extraction of the original data stream to obtain a feature vector. Specifically, this includes: denoising the synchronized original data stream using an adaptive filtering algorithm based on wavelet transform to remove power frequency interference and random noise; extracting time-domain statistical features from the denoised original data stream, including the mean of rotational speed, the variance of torque, the peak factor of pressure, and the kurtosis of the vibration signal; performing a short-time Fourier transform on the vibration signal to extract frequency-domain features including the dominant frequency energy, bandwidth power, and spectral centroid; calculating derived physical features based on the pump's physical model, including the dimensionless pressure rise coefficient, specific energy efficiency, and net positive suction head (NPSH); constructing a multi-level feature fusion module to weightedly fuse the time-domain features, frequency-domain features, and derived physical features to form a high-dimensional vector; and using principal component analysis to reduce the dimensionality of the high-dimensional vector to obtain a feature vector, which is then used for operating condition identification.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves the process of constructing a multi-level fusion strategy by acquiring historical raw data streams. Specifically, this includes: acquiring historical raw data streams and constructing a cavitation risk assessment layer by setting a safety margin threshold; extracting multi-dimensional feature vectors based on the historical raw data streams and training a multi-sensor fusion classification layer using historical labeled data.

[0007] Based on process context information, a logic verification layer is constructed through a rule engine. The process context information includes valve operation history, separator liquid level data, and pump operation sequence records. A dedicated integration module is used to hierarchically fuse the cavitation risk assessment layer, the multi-sensor fusion classification layer, and the logic verification layer to form a multi-level fusion strategy.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves using a multi-level fusion strategy to identify the operating conditions of the feature vectors, obtaining the current operating conditions including pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload states. Specifically, this includes: the cavitation risk assessment layer calling the inlet pressure from the original data stream. And temperature T, and calculate saturated vapor pressure using the Antoine equation. and import pressure With saturated vapor pressure The difference Preset safety margin threshold and the difference The system compares the results and outputs the cavitation risk indicator judgment. The multi-sensor fusion classification layer receives feature vectors and uses a pre-trained machine learning classifier to analyze them in real time, outputting the probability distributions for pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload states. The logic verification layer verifies the output of the multi-sensor fusion classification layer based on process context information, generating verification results. The dedicated integration module receives the cavitation risk indicator judgment results, probability distributions, and verification results, and uses a weighted fusion algorithm or rule engine for hierarchical fusion to generate the current operating condition status and confidence level.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves adaptively selecting the corresponding model based on the current operating conditions, and obtaining the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two-phase system through mechanistic model calculation and data-driven model thrust. Specifically, this includes: constructing an adaptive model selection module based on the current operating condition's state label and confidence level, and selecting the corresponding mechanistic model and data-driven model from a pre-built model library. The mechanistic model uses the hydraulic and thermodynamic equations of the screw pump to calculate the real-time feature vector, obtaining the basic flow rate, basic gas-liquid ratio, and basic pump efficiency. The data-driven model uses a pre-trained neural network or support vector machine to process the same feature vector, obtaining the flow rate correction value, gas-liquid ratio correction value, and pump efficiency correction value. A digital twin collaborative validator fuses the outputs of the mechanistic model and the data-driven model, calculates the optimal estimate, and generates the gas-liquid two-phase flow rate, gas-liquid ratio, and pump efficiency, along with a confidence index.

[0010] In conjunction with the first aspect mentioned above, one possible implementation of the process of acquiring raw data streams from the real-time operating parameters of the screw pump specifically includes: configuring a multimodal sensor array, which comprises a non-contact digital torque sensor, a true RMS power measurement module, an inlet pressure sensor and an outlet pressure sensor with temperature compensation, a temperature sensor, a high-frequency vibration sensor, and a combination of acoustic emission sensors. A hardware synchronous acquisition unit uses an FPGA chip to drive all sensors to trigger synchronous acquisition at the same sampling frequency. An adaptive filtering algorithm based on wavelet transform is used to perform real-time noise reduction on the raw signal, generating a preprocessed signal. A dedicated integrated module performs real-time physical validity verification on the preprocessed signal; if verification fails, a data quality alarm is triggered and abnormal data is marked. The verified signals are packaged into a time-series raw data stream, with timestamps and quality flags appended.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves calibrating and optimizing the internal parameters of the fusion mechanism model and the data-driven model based on the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two-phase flow, using stable operating conditions or external calibration data. Specifically, this includes: collecting the operating status of the screw pump; when a stable pure liquid phase flow condition is identified, marking the corresponding time period as a stable operating condition period; or obtaining the reference true values ​​of flow rate, gas-liquid ratio, and pump efficiency through an external calibration device. A calibration and optimization module is constructed, comprising a golden moment detection unit, a parameter deviation calculation unit, and an adaptive learning unit. The golden moment detection unit determines the start and end of the stable operating condition period, the parameter deviation calculation unit generates a deviation vector, and the adaptive learning unit estimates and optimizes the state of the parameters of the mechanism model and the data-driven model. A digital twin collaborative verifier is used to simulate the output of the optimized parameters under historical operating conditions, and the rationality of the parameters is verified during the calibration and optimization process.

[0012] Secondly, a gas-liquid two-phase flow metering system based on screw pump operating condition self-identification is provided. This system includes a physical sensing module, a data acquisition and preprocessing module, an operating condition self-identification module, a flow metering core module, and a data storage and calibration module. The physical sensing module, composed of a multi-modal sensor array, is used to acquire the screw pump's operating parameters in real time, obtaining the raw data stream. The data acquisition and preprocessing module performs synchronous, adaptive filtering and noise reduction, and feature extraction on the raw data stream to obtain feature vectors. The operating condition self-identification module, connected to the data acquisition and preprocessing unit, receives the feature vectors and uses a multi-level fusion strategy to identify the operating condition, outputting the current operating condition status and confidence level. The flow metering core module adaptively selects the mechanism model and data-driven model based on the current operating condition status to calculate the gas-liquid two-phase flow rate, gas-liquid ratio, and pump efficiency. The data storage and calibration module stores historical data and applies external calibration data to perform online calibration and optimization of model parameters.

[0013] In conjunction with the second aspect mentioned above, in one possible implementation, the data acquisition and preprocessing module includes a hardware synchronous acquisition subunit, an adaptive filtering and denoising subunit, a feature extraction subunit, and a feature fusion and dimensionality reduction subunit: The hardware synchronous acquisition subunit drives all sensors to synchronously acquire raw signals at the same sampling frequency. The adaptive filtering and denoising subunit uses an adaptive filtering algorithm based on wavelet transform to denoise the raw signals, generating a preprocessed signal. The feature extraction subunit extracts time-domain features, frequency-domain features, and derived physical features from the preprocessed signal. The feature fusion and dimensionality reduction subunit weights and fuses the time-domain features, frequency-domain features, and derived physical features to form a high-dimensional vector, and uses principal component analysis to reduce the dimensionality of the high-dimensional vector to obtain a feature vector. The data acquisition and preprocessing unit performs real-time physical validity checks on the preprocessed signal; if the check fails, a data quality alarm is triggered.

[0014] In conjunction with the second aspect mentioned above, in one possible implementation, the core of flow metering includes an adaptive model selector, a model library, a mechanistic model library, a data-driven model library, and a digital twin collaborative validator. The adaptive model selector selects the corresponding mechanistic model and data-driven model from the pre-built model library based on the current operating condition and confidence level. The model library includes single-phase flow mechanistic models for pure liquid phase conditions, multiphase flow mixing models for gas-liquid two-phase conditions, and degradation models for abnormal operating conditions. The mechanistic model library stores models based on the hydraulic and thermodynamic equations of a screw pump, used to calculate the basic flow rate, basic gas-liquid ratio, and basic pump efficiency based on real-time feature vectors. The data-driven model library stores pre-trained neural network or support vector machine models, used to process feature vectors to obtain flow rate correction values, gas-liquid ratio correction values, and pump efficiency correction values. The digital twin collaborative validator fuses the outputs of the mechanistic model and the data-driven model, calculates the optimal estimates of the gas-liquid two-phase flow rate, gas-liquid ratio, and pump efficiency, and generates a confidence index.

[0015] This application provides a gas-liquid two-phase flow metering method and system based on screw pump operating condition self-identification. It can simultaneously collect multi-dimensional operating parameters such as speed, torque, and electrical power, perform preprocessing noise reduction, and comprehensively extract time-domain, frequency-domain, and derived physical features. This preprocessing noise reduction and comprehensive extraction of time-domain, frequency-domain, and derived physical features overcomes the limitations of traditional external flowmeters that rely on a single sensor signal. It solves the problem of incomplete signal acquisition in scenarios with high gas content and variable flow patterns, ensuring that the raw data fully reflects the operating status of the screw pump and the flow characteristics of the medium. Simultaneously, a multi-level fusion strategy is constructed, including a cavitation risk judgment layer, a multi-sensor fusion classification layer, and a logic verification layer. Through hierarchical judgment based on physical mechanism constraints, multi-feature machine learning classification, and process context verification, it accurately distinguishes various operating conditions such as pure liquid phase, gas-liquid two-phase, and cavitation. In particular, it solves the industry pain point of easily confusing "cavitation" and "gas content," avoiding metering deviations and equipment damage risks caused by misjudgment of operating conditions. Secondly, by using an adaptive selection mechanism model and a data-driven model in conjunction with a collaborative computational structure, the mechanism model can ensure physical rationality and extrapolation, while the data-driven model compensates for errors under complex working conditions. This solves the problems of a single model being unable to adapt to multiple working conditions and insufficient measurement accuracy, and enables high-precision calculation of gas-liquid flow rate, gas-liquid ratio and pump efficiency, thus meeting the core requirement for the accuracy of measurement data in actual production.

[0016] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0017] Figure 1 A system architecture diagram of a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification provided in this application embodiment; Figure 2 A schematic flowchart of a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification provided in this application embodiment; Figure 3 A schematic flowchart of a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification provided in this application embodiment; Figure 4 A schematic flowchart of a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification provided in this application embodiment; Figure 5 A schematic flowchart of a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification provided in this application embodiment; Figure 6 A schematic flowchart of a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification provided in this application embodiment; Figure 7 A schematic flowchart of a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification provided in this application embodiment; Figure 8 A schematic diagram of the gas-liquid two-phase flow metering system based on screw pump operating condition self-identification provided in this application embodiment. Detailed Implementation

[0018] To address the shortcomings of existing external flowmeters that rely on single-phase flow measurement principles, which suffer from unstable flow coefficients and significantly increased measurement errors in high-gas-content and variable-flow-pattern environments such as gas-liquid two-phase flows due to gas-liquid interface fluctuations and large density differences, single-parameter derivation methods rely solely on limited data such as motor current and speed for indirect estimation. These methods are highly susceptible to variations in medium viscosity and composition, failing to meet the measurement needs of complex operating conditions. Furthermore, during the judgment process, it is difficult to accurately distinguish between different operating states such as pure liquid phase, gas-liquid two-phase flow, cavitation, underload, and overload. In particular, there is a problem of confusion between cavitation and high-gas-content conditions, directly leading to inaccurate selection of the measurement model, affecting measurement accuracy and creating potential equipment damage risks.

[0019] Traditional solutions require the installation of multiple independent sensors and wired transmission, making construction difficult and deployment costs high. They also suffer from insufficient flexibility due to complex well site environments. Furthermore, existing metering models lack dynamic calibration mechanisms, failing to address model drift caused by equipment wear and aging, and changes in media properties during production cycles, necessitating frequent manual calibration and further increasing maintenance costs. Finally, in data processing and utilization, overall efficiency is lacking. Not only are multi-source operating parameters such as speed, torque, pressure, and vibration not systematically preprocessed, synchronized, and feature-extracted, but key information in the time domain, frequency domain, and derived physical characteristics is also not fully explored, resulting in wasted data value and an inability to provide reliable support for accurate metering. This severely restricts the accuracy and long-term stability of screw pump gas-liquid two-phase flow metering, making it difficult to meet the technical needs of precise control, cost reduction, and efficiency improvement in industrial production. This application provides a gas-liquid two-phase flow metering method based on screw pump operating condition self-identification.

[0020] like Figure 1As shown in the embodiments of this application, the gas-liquid two-phase flow metering method based on screw pump operating condition self-identification includes: Step 101: Collect the operating parameters of the screw pump in real time to obtain the raw data stream. The raw data stream includes speed, torque, electric power, inlet pressure, outlet pressure, temperature and vibration signal.

[0021] In some implementations, a motor speed sensor directly measures the rotational speed signal, while a torque meter or high-precision current / power sensor acquires torque and electrical power data. Simultaneously, inlet and outlet pressure sensors collect inlet and outlet pressure values, respectively, while a temperature sensor monitors temperature changes in the pump or fluid, and a vibration sensor captures mechanical vibration signals. This allows the data acquisition unit to synchronously sample all sensor signals at a fixed frequency, ensuring timestamp alignment and performing filtering and noise reduction processing (e.g., using low-pass filtering to eliminate high-frequency noise), thereby generating a raw data stream containing rotational speed, torque, electrical power, inlet pressure, outlet pressure, temperature, and vibration signals.

[0022] For example, in an oilfield production scenario, when the screw pump is running, the speed sensor monitors the motor speed as 500 rpm, the torque meter measures the shaft torque as 200 Nm, the power sensor records the input power as 10 kW, the inlet pressure sensor displays 0.5 MPa, the outlet pressure sensor displays 1.5 MPa, the temperature sensor reads 60 degrees Celsius, and the vibration sensor outputs an acceleration signal; the data acquisition system synchronously samples at a frequency of 1 kHz to generate a raw data stream of time series.

[0023] Step 102: Preprocess and extract features from the original data stream to obtain feature vectors.

[0024] Feature extraction refers to calculating representative indicators from preprocessed data. Time-domain features refer to statistics calculated over time (such as mean, variance, and peak value). Frequency-domain features refer to indicators obtained through frequency domain analysis (such as dominant frequency and spectral energy). Derived physical features refer to parameters derived from physical principles (such as pressure rise, pump efficiency, and specific energy).

[0025] In some implementations, during the preprocessing of the raw data stream, the data acquisition unit first synchronizes and aligns the speed, torque, electrical power, inlet pressure, outlet pressure, temperature, and vibration signals to ensure consistent timestamps. This allows for the removal of power frequency interference and high-frequency noise using low-pass filtering or moving average methods, achieving signal smoothing. Time-domain features (such as the mean of speed and the variance of torque) can then be directly calculated, and frequency-domain analysis can be performed, such as Fourier transforming the vibration signal to extract the dominant frequency amplitude. Derived physical features, such as pressure rise (outlet pressure minus inlet pressure), pump efficiency (ratio of output power to input power), and specific energy (ratio of torque to flow rate), are calculated simultaneously. Finally, all features are combined into a feature vector for subsequent operating condition identification.

[0026] Step 103: Obtain historical raw data streams to construct a multi-level fusion strategy. The multi-level fusion strategy includes a cavitation risk assessment layer, a multi-sensor fusion classification layer, and a logic verification layer.

[0027] The cavitation risk assessment layer is defined as follows: the layer that assesses cavitation risk based on a comparison between inlet pressure and saturated vapor pressure; the multi-sensor fusion classification layer that classifies operating conditions by fusing data from multiple sensors; and the logic verification layer that performs logical verification based on timing and context.

[0028] In some implementations, when acquiring historical raw data streams to construct multi-level fusion strategies, By reading historical raw data streams from the data storage module, cavitation risk can be identified by calculating the difference between inlet pressure and saturated vapor pressure and setting a safety margin, thus constructing a cavitation risk assessment layer. A machine learning classifier is then used to train historical feature vectors, outputting operating condition probabilities to construct a multi-sensor fusion classification layer. Simultaneously, logical verification is performed by combining the causes of operating condition changes and process context to construct a logic verification layer. Finally, the cavitation risk assessment layer, multi-sensor fusion classification layer, and logic verification layer can be integrated to form a complete multi-level fusion strategy.

[0029] It should be noted that historical data needs to cover a variety of operating conditions to enhance the generalization of the strategy. Among them, the cavitation risk assessment layer should be executed first to ensure equipment safety; while the multi-sensor fusion classification layer should be updated with new data regularly; at the same time, the logic verification layer can reduce false judgments.

[0030] Step 104: Use a multi-level fusion strategy to identify the operating conditions of the feature vectors and obtain the current operating conditions, including pure liquid phase, gas-liquid two-phase, cavitation, underload or overload conditions.

[0031] The current operating condition refers to the real-time operating status of the screw pump as determined by the operating condition identification module, such as pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload. Cavitation refers to the phenomenon of liquid vaporization caused by excessively low inlet pressure.

[0032] In some implementations, feature vectors are directly invoked, and a multi-level fusion strategy is activated. This strategy first calculates the difference between the inlet pressure and the saturated vapor pressure in the cavitation risk assessment layer based on the feature vectors. If the difference is lower than the safety margin, a cavitation risk alarm is triggered first. Then, a machine learning classifier is used in the multi-sensor fusion classification layer to analyze the feature vectors, outputting the probability distributions of pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload states. Finally, in the logic verification layer, logic verification is performed by combining temporal changes and process context, such as checking valve operation history or separator liquid level, to ultimately determine the current operating condition.

[0033] Step 105: Adaptively select the corresponding model based on the current operating conditions, and calculate the thrust of the gas-liquid two-phase flow rate, gas-liquid ratio and pump efficiency through mechanism model calculation and data-driven model.

[0034] The selectable models include single-phase flow mechanism models for pure liquid phase conditions, multiphase flow mixing models for gas-liquid two-phase conditions, and degradation models for abnormal conditions (such as cavitation). Mechanism models refer to mathematical models established based on physical principles such as the pump's hydraulic equations and thermodynamic equations.

[0035] In some implementations, a model library is first scheduled based on the operating condition identification results. If the operating condition is purely liquid, a single-phase flow mechanism model is called; if it is a gas-liquid two-phase operating condition, a multiphase flow mixing model is called. Then, the mechanism model calculation and data-driven model inference are executed in parallel. The mechanism model uses physical equations to calculate the basic flow rate and efficiency based on characteristics such as rotational speed, torque, and pressure difference. Simultaneously, the data-driven model infers and predicts corrected values ​​for flow rate and efficiency based on the same feature vectors. A hybrid model scheduler can then be used to fuse the outputs of the two models (e.g., through weighted averaging or state estimation algorithms to obtain the optimal estimate). This allows for the calculation of the liquid phase volumetric flow rate, gas phase volumetric flow rate, instantaneous gas-liquid ratio, and current pump volumetric efficiency.

[0036] It should be noted that invalid flags can be directly output through the degradation model under abnormal operating conditions.

[0037] Step 106: Based on the flow rate, gas-liquid ratio and pump efficiency of the gas-liquid two phases, use stable operating conditions or external calibration data to calibrate and optimize the internal parameters of the fusion mechanism model and the data-driven model.

[0038] External calibration data refers to comparative data from external standard measuring devices (such as mass flow meters or standard volume tubes).

[0039] In some implementations, stable operating conditions, such as stable pure liquid flow, are first monitored and identified, or external calibration data is received as a reference value. The deviations of the flow rate, gas-liquid ratio, and pump efficiency calculated by the current model from the reference values ​​can then be compared. Optimization algorithms (such as least squares or extended Kalman filters) are then used to fine-tune the internal parameters of the fusion mechanistic model and the data-driven model (e.g., adjusting the efficiency coefficient of the mechanistic model or the weights of the data-driven model). The optimized parameters can then be directly updated in the model library, completing the calibration optimization.

[0040] Based on the above technical solution, by simultaneously collecting multi-dimensional operating parameters such as speed, torque, and electrical power, and performing preprocessing noise reduction and comprehensive extraction of time-domain, frequency-domain, and derived physical features, this approach overcomes the limitations of traditional external flowmeters that rely on a single sensor signal. It also solves the problem of incomplete signal acquisition in scenarios with high gas content and variable flow patterns, ensuring that the raw data fully reflects the screw pump's operating status and medium flow characteristics. Simultaneously, a multi-level fusion strategy is constructed, including a cavitation risk assessment layer, a multi-sensor fusion classification layer, and a logic verification layer. Through hierarchical judgment based on physical mechanism constraints, multi-feature machine learning classification, and process context verification, it accurately distinguishes between various operating conditions such as pure liquid phase, gas-liquid two-phase, and cavitation. In particular, it addresses the industry pain point of easily confusing "cavitation" with "gas content," avoiding measurement deviations and equipment damage risks caused by misjudgments of operating conditions. Secondly, by using an adaptive selection mechanism model and a data-driven model in conjunction with a collaborative computational structure, the mechanism model can ensure physical rationality and extrapolation, while the data-driven model compensates for errors under complex working conditions. This solves the problems of a single model being unable to adapt to multiple working conditions and insufficient measurement accuracy, and enables high-precision calculation of gas-liquid flow rate, gas-liquid ratio and pump efficiency, thus meeting the core requirement for the accuracy of measurement data in actual production.

[0041] In one possible implementation of the embodiments of this application, combined with Figure 1-2 As shown, the process of preprocessing and feature extraction of the original data stream to obtain feature vectors can be achieved through the following steps 201 to 206, which are explained in detail below: Step 201: Use an adaptive filtering algorithm based on wavelet transform to denoise the synchronized original data stream, removing power frequency interference and random noise.

[0042] In some implementations, a GPS synchronization module or the IEEE 1588 Precise Time Protocol (PTP) is used as a unified time reference source before its use. This allows the GPS synchronization module to obtain high-precision time by receiving satellite signals. The IEEE 1588 protocol is suitable for network environments and can achieve microsecond or millisecond-level time synchronization within a local area network. These clock sources are integrated into the data acquisition hardware to provide synchronous trigger signals for all sensors (such as speed, torque, pressure, and vibration sensors), ensuring that all sensors in the raw data stream are sampled at the same time point. The db4 wavelet basis function can then be selected to perform multi-scale decomposition of the signal, separating it into high-frequency details and low-frequency approximations to separate noise and useful signals. Soft thresholding is then applied to threshold the decomposed wavelet coefficients, treating coefficients below the threshold as noise and setting them to zero, while reducing coefficients above the threshold by the threshold value. This smooths noise while preserving signal characteristics. Finally, the thresholded wavelet coefficients are reconstructed to obtain the denoised signal, effectively removing power frequency interference and random noise, thus improving data quality.

[0043] For example, in an oilfield production scenario, during screw pump operation, a speed sensor monitors the motor speed at 500 rpm, a torque meter measures the shaft torque at 200 Nm, a pressure sensor collects the inlet pressure at 0.5 MPa and the outlet pressure at 1.5 MPa, and a vibration sensor outputs an acceleration signal. Therefore, a GPS synchronization module can be directly used as a high-precision clock source to perform hardware-level synchronized acquisition of all sensor data, ensuring that the timestamps of speed, torque, pressure, and vibration signals are aligned to the millisecond level (e.g., all data are sampled synchronously at a frequency of 1 kHz). Then, an adaptive filtering algorithm based on wavelet transform is applied to the synchronized vibration signal. A 5-level decomposition is performed using the db4 wavelet basis function, and a soft threshold is applied with a threshold set to 0.1 to remove 50Hz power frequency interference and random noise. Finally, the signal is reconstructed to obtain smooth vibration data.

[0044] Step 202: Extract time-domain statistical features from the denoised raw data stream, including the mean of rotational speed, the variance of torque, the peak factor of pressure, and the kurtosis of the vibration signal.

[0045] In some implementations, a fixed-size window can be defined and slid across the rotational speed time series to calculate the arithmetic mean of the data points within each window as the mean of the rotational speed. Simultaneously, the variance is calculated using the standard deviation formula (i.e., first calculating the standard deviation of the torque data as the square root of the variance, then squaring the square root of the variance to obtain the torque variance). Then, the peak value (maximum value) of the pressure signal is identified, and the root mean square value of the pressure is calculated as the denominator. Dividing the peak value by the root mean square value yields the peak factor. Finally, the fourth moment of the vibration signal (the average of the fourth power of the difference between the data and the mean) is calculated, and the square of the variance of the vibration signal is calculated as the denominator. Dividing the fourth moment by the square of the variance yields the kurtosis value. Thus, the mean of the rotational speed, the variance of the torque, the peak factor of the pressure, and the kurtosis of the vibration signal are directly summarized to obtain the time-domain statistical characteristics.

[0046] For example, if the noise-reduced rotational speed data is a sequence of 500 revolutions per minute, the mean can be calculated using a sliding window of 10 data points to obtain an average rotational speed of 500 revolutions per minute. For a sequence of torque data of 200 Nm, the variance is calculated using the standard deviation formula, assuming that the torque fluctuation is small and the variance is close to 0. Simultaneously, for a sequence of pressure data with an inlet pressure of 0.5 MPa, the peak factor is calculated as 1.1 by the ratio of the peak value of 0.55 MPa to the root mean square value of 0.5 MPa. Finally, the kurtosis of the vibration signal sequence is calculated as 100 by the ratio of the fourth moment of 0.01 to the square of the variance of 0.0001, which enables subsequent operating condition identification.

[0047] Step 203: Perform a short-time Fourier transform on the vibration signal to extract frequency domain features including main frequency energy, bandwidth power, and spectral centroid.

[0048] The Short-Time Fourier Transform (SFT) is a time-frequency analysis method that divides a signal into short time intervals and performs Fourier transforms to analyze frequency changes over time. The dominant frequency energy is the energy value corresponding to the main frequency components in the signal spectrum, reflecting the strength of the dominant frequency. Band power is the sum of the signal's power within a specific frequency range, used to assess the energy distribution across different frequency bands. The spectral centroid is the energy-weighted average frequency of the spectrum, used to represent the location of concentrated energy in the spectrum.

[0049] In some implementations, the vibration signal is divided into multiple short time periods, and a Fourier transform is applied to each period to obtain the time spectrum, thereby analyzing the frequency components' changes over time. This can be achieved by identifying peak frequency points in the spectrum and calculating the corresponding energy value as the dominant frequency energy. Furthermore, a specific frequency range (such as a low-frequency band or a high-frequency band) can be directly defined, and the spectral power within this range is integrated and summed to obtain the band power. Simultaneously, by multiplying the energy value of each frequency point by its frequency value, summing the results, and then dividing by the total energy value, the spectral centroid is obtained as a frequency domain feature.

[0050] For example, the denoised vibration signal is an acceleration sequence. A short-time Fourier transform (SFT) with a duration of 0.1 seconds is first applied to it to obtain the time spectrum. This allows for the direct identification of a peak frequency of 100 Hz and an energy value of 50 units, extracting the dominant frequency energy. Then, integration over the 50-150 Hz range yields a bandwidth power of 200 units. Finally, the spectral centroid is calculated, assuming a weighted average energy distribution of 110 Hz.

[0051] Step 204: Calculate derived physical characteristics based on the pump's physical model. The derived physical characteristics include dimensionless pressure rise coefficient, specific energy efficiency, and net positive suction head (NPSH).

[0052] The dimensionless pressure rise coefficient is a dimensionless parameter obtained by dividing the difference between the pump outlet pressure and the inlet pressure by the inlet pressure or dynamic pressure. It is used to standardize the pressure rise effect to eliminate the influence of dimensions. Specific energy efficiency is the ratio of the pump's output hydraulic power to its input mechanical power, used to reflect energy conversion efficiency. Net positive suction head (NPSH) is the difference between the liquid pressure at the pump inlet and its saturated vapor pressure, used to assess cavitation risk.

[0053] In some implementations, based on real-time collected screw pump operating parameters including inlet pressure, outlet pressure, temperature, speed, and torque, the pressure difference is directly obtained by subtracting the inlet pressure from the outlet pressure, and then divided by the inlet pressure. Alternatively, the pressure difference is calculated by combining the medium density and speed using a dynamic pressure formula to obtain a dimensionless pressure rise coefficient. Then, the specific energy efficiency is calculated by the ratio of the pump's output hydraulic power to its input mechanical power. The output hydraulic power is estimated based on the pressure difference and theoretical flow rate, while the input mechanical power is calculated based on torque and speed. Finally, the effective value, i.e., the net positive suction head (NPSH), is obtained by subtracting the saturated vapor pressure from the inlet pressure and then dividing by the medium density and gravitational acceleration, thus completing feature extraction.

[0054] For example, by calling the operating parameters of a screw pump, such as a speed of 500 rpm, torque of 200 Nm, inlet pressure of 0.5 MPa, outlet pressure of 1.5 MPa, and temperature of 60 degrees Celsius, the pressure difference can be calculated as 1.0 MPa. Assuming the medium is water with a density of 1000 kg / m³, the dimensionless pressure rise coefficient of 2.0 is obtained by dividing the pressure difference by the inlet pressure. Simultaneously, the specific energy efficiency is estimated by multiplying the output hydraulic power by the pressure difference and the theoretical flow rate, and the input mechanical power by the torque and speed. Assuming a theoretical flow rate of 10 m³ / h, the calculated efficiency is approximately 85%. The net positive suction head (NPSH) is calculated using the saturated vapor pressure of water at 60 degrees Celsius, which is 0.02 MPa, yielding (0.5 - 0.02). 1000 / (1000 9.81)≈0.049 meters, thus forming a feature vector for working condition identification.

[0055] Step 205: Construct a multi-level feature fusion module to weight and fuse time-domain features, frequency-domain features, and derived physical features to form a high-dimensional vector.

[0056] In some implementations, weight coefficients are assigned to different types of features based on feature importance or expert experience. This allows for the direct weighted fusion of time-domain features, frequency-domain features, and derived physical features through a multi-level feature fusion module. The weighted feature values ​​are then concatenated to form a high-dimensional vector containing all feature components.

[0057] For example, by acquiring extracted time-domain features such as average rotational speed of 500 rpm and torque variance close to 0, frequency-domain features such as main frequency energy of 50 units and bandwidth power of 200 units, and derived physical features such as dimensionless pressure rise coefficient of 2.0, specific energy efficiency of 85%, and cavitation margin of 0.049 m, weights can be directly assigned to various features, such as time-domain feature weight of 0.3, frequency-domain feature weight of 0.4, and derived physical feature weight of 0.3. After weighted calculation, all components are concatenated to finally form a high-dimensional vector [500 0.3,0 0.3,50 0.4,200 0.4, 2.0 0.3, 0.85 0.3, 0.049 0.3].

[0058] Step 206: Use principal component analysis to reduce the dimensionality of the high-dimensional vector to obtain the feature vector, which is used for working condition identification.

[0059] In some implementations, the high-dimensional vector obtained from the multi-level feature fusion module is used as input data. The high-dimensional vector is then standardized to ensure consistent feature scales, allowing for the calculation of the covariance matrix and the determination of its eigenvalues ​​and eigenvectors. In this case, k can be determined directly using the cumulative contribution rate (e.g., retaining 95% variance), and the top k principal components are selected based on their eigenvalues. The original high-dimensional vector is then projected onto these selected principal components to obtain the dimensionality-reduced eigenvectors.

[0060] For example, a high-dimensional vector such as [150,0,20,80,0.6,0.255,0.0147] can be obtained and principal component analysis can be performed in combination with historical datasets. Assuming the cumulative contribution rate threshold is set to 95%, the first three principal components are selected, and the real-time high-dimensional vector is projected to obtain a dimensionality-reduced feature vector such as [0.5,-0.2,1.1]. This vector can then be directly input into the working condition recognition model to determine whether the current state is a pure liquid phase or a gas-liquid two-phase system.

[0061] Based on the above technical solutions, by constructing a complete technical chain of noise reduction processing, multi-dimensional feature extraction, feature fusion, and dimensionality reduction optimization, the precise removal of power frequency interference and random noise can be achieved directly through multi-scale decomposition and soft thresholding. This solves the problem of sensor signals being easily affected by environmental interference and data distortion in industrial scenarios, ensuring that the original data can truly reflect the operating status of the screw pump and the flow characteristics of the medium. Simultaneously, frequency domain features such as the dominant frequency energy of the vibration signal are obtained through short-time Fourier transform, and derived physical features such as the dimensionless pressure rise coefficient are calculated based on the pump body physical model. This breaks through the limitations of traditional single-parameter or limited feature analysis, comprehensively mining key information in the data reflecting changes in operating conditions at the time, frequency, and physical mechanism levels. This solves the problem of insufficient basis for operating condition identification due to incomplete feature capture under complex operating conditions. Furthermore, the three types of features are weighted and fused to form a high-dimensional vector, achieving complementary enhancement of multi-dimensional features. This allows the high-dimensional vector to be reduced to a feature vector using principal component analysis, simplifying the data dimension while retaining core information, reducing the computational complexity of subsequent operating condition identification models, and improving identification efficiency. This effectively improves the accuracy of operating condition identification, especially providing comprehensive and reliable feature support for distinguishing easily confused operating conditions such as cavitation and high gas content. It avoids measurement deviations and equipment damage risks caused by misjudgment of operating conditions, and provides key technical support for the accurate measurement of flow rate, gas-liquid ratio and pump efficiency of screw pump gas-liquid two-phase flow, meeting the actual needs of accurate measurement data and safe operation of equipment in industrial scenarios such as oil extraction and chemical production.

[0062] In one possible implementation of the embodiments of this application, combined with Figure 1-3 As shown, the process of constructing a multi-level fusion strategy by acquiring historical raw data streams can be achieved through the following steps 301 to 304, which are explained in detail below: Step 301: Obtain historical raw data streams and construct a cavitation risk assessment layer by setting a safety margin threshold.

[0063] In some implementations, historical raw data streams are read from the historical data storage module, and the inlet pressure and temperature values ​​are retrieved. The saturated vapor pressure can then be calculated using the Antoine equation based on the temperature value. Finally, the difference between the inlet pressure and the saturated vapor pressure is calculated to obtain the difference value. At the same time, a safety margin threshold should be set within the range of 20-50 kPa based on the media type and operating experience. You can directly use and If a comparison is made, Less than If the cavitation risk priority alarm is triggered, then no alarm will be triggered; otherwise, no alarm will be triggered. Finally, the cavitation risk assessment layer outputs a binary status flag to complete the construction.

[0064] It should be noted that when constructing the cavitation risk assessment layer, the safety margin threshold... The settings need to be adjusted according to the characteristics of the medium; for high-temperature media, the threshold should be reduced to ensure sensitivity. Inlet pressure data must come from a calibrated sensor to avoid drift errors. Simultaneously, the saturated vapor pressure calculation must use the correct Antoine equation coefficients to reflect the liquid properties.

[0065] For example, if the historical raw data stream contains an inlet pressure of 0.5 MPa and a temperature of 60°C, the saturated vapor pressure can be calculated as 0.02 MPa based on the Antoine equation at a temperature of 60°C. =0.48MPa; Setting It is 0.03 MPa, because Greater than No alarm was triggered, and the cavitation risk assessment layer outputs a binary status flag of 0, indicating no cavitation risk.

[0066] Step 302: Based on the historical raw data stream, extract multi-dimensional feature vectors and use historical labeled data to train a multi-sensor fusion classification layer.

[0067] In some implementations, historical raw data streams are preprocessed and feature extracted to obtain multidimensional feature vectors based on these streams. Time-domain features are calculated from the rotational speed signal; for example, the mean rotational speed is obtained through a sliding window arithmetic mean, the torque variance is calculated using the standard deviation formula, the pressure peak factor is obtained by the ratio of the maximum pressure signal value to the root mean square value, and the vibration kurtosis is obtained by the ratio of the fourth moment of the vibration signal to the square of its variance. Then, a short-time Fourier transform is performed on the vibration signal, and frequency-domain features, such as the dominant frequency energy, are extracted by identifying the energy values ​​corresponding to the peak frequencies of the spectrum. The bandwidth power is obtained by integrating and summing over a specific frequency range, while the spectral centroid is calculated using the energy-weighted average frequency. Finally, derived physical features are calculated based on the pump's physical model: for example, the dimensionless pressure rise coefficient is obtained by the ratio of the difference between the outlet pressure and the inlet pressure to the inlet pressure; the specific energy efficiency is calculated by the ratio of the pump's output hydraulic power to its input mechanical power; and the net positive suction head (NPSH) is obtained by the difference between the inlet pressure and the saturated vapor pressure. This allows for the direct combination of time-domain features, frequency-domain features, and derived physical features into a multi-dimensional feature vector.

[0068] At this point, historical labeled data can be directly used as the training set. Multi-dimensional feature vectors are used as input for preprocessing, including digitizing the class labels and standardizing numerical features to ensure a mean of 0 and a variance of 1. This ensures all features have the same scale during training, preventing certain features with large values ​​from dominating the training process. A random forest classifier is then initialized, and its parameters are set. =100 (meaning there are 100 decision trees in the forest). =10 (limits the maximum depth of each decision tree to 10 layers to prevent overfitting). Other default but crucial hyperparameters are also used, such as setting the Gini coefficient as the criterion for splitting nodes, and bootstrap=True is enabled (indicating the use of bootstrap sampling to build a training subset for each tree). This initiates the training loop.

[0069] The training loop is a parallel loop that constructs training subsets for each of the 100 decision trees and trains each tree independently. During training, samples are first randomly drawn with replacement from the total training set to form a subset the same size as the original training set. Each subset contains approximately 63.2% of the original samples, with the remaining approximately 36.8% becoming "out-of-bag data" used to train the current tree. Then, when constructing a single decision tree, starting from the root node, at each node, instead of considering all features, a subset of features (e.g., the square root of the total number of features) is randomly selected. This subset is then traversed through all features and their possible split points, calculating the reduction in the Gini coefficient after splitting. Finally, the feature and split point that maximizes the reduction in the Gini coefficient (i.e., the largest decrease in data impurity) are selected, and the data at the current node is divided into two subsets.

[0070] The above training process is repeated recursively until the maximum depth of 10 layers is reached, or the number of samples contained in a node is less than a certain minimum, which constitutes the termination condition. Once all 100 decision trees have been trained, they together form the final multi-sensor fusion classification layer model.

[0071] For example, if the cavitation risk assessment layer outputs no alarm, and the historical raw data stream contains data such as rotational speed of 500 rpm, torque of 200 Nm, inlet pressure of 0.5 MPa, outlet pressure of 1.5 MPa, temperature of 60℃, and vibration signal sequence, then time-domain features such as mean rotational speed of 500 rpm, torque variance of 0.1, pressure peak factor of 1.1, and vibration kurtosis of 100, frequency-domain features such as dominant frequency energy of 50 units, bandwidth power of 200 units, and spectral centroid of 110 Hz, and derived physical features such as dimensionless pressure rise coefficient of 2.0, specific energy efficiency of 85%, and cavitation margin of 0.049 m can be combined into a multi-dimensional feature vector. Then, using labels known to be pure liquid phase from the historical labeled data, a random forest model is trained, so that after learning, the model outputs a working condition probability distribution of 80% for pure liquid phase and 20% for gas-liquid two-phase.

[0072] Step 303: Based on the process context information, construct a logic verification layer through the rule engine. The process context information includes valve operation history, separator liquid level data, and pump operation sequence records.

[0073] The process context information refers to the process parameters and historical data related to the screw pump operation, providing the environmental background for system operation. Valve operation history records the time-series data of valve opening and closing states, reflecting process control actions. Separator level data refers to the measured liquid height in the separator, used to assess gas-liquid separation efficiency and liquid supply stability. Pump operation sequence records include time-series information such as pump start-up, shutdown, and speed changes, used to analyze operating modes.

[0074] In some implementations, historical valve operation data, including valve opening / closing status and timestamps, is collected in real time, along with separator liquid level data and pump operation sequence records such as start-up and stop events. A rule engine can then define logical rules (e.g., if a sudden drop in inlet pressure is detected after the inlet valve is closed) to trigger cavitation risk verification, or indicate a high gas content condition if the separator liquid level remains below a threshold and the pump operation sequence shows stable operation. This allows the rule engine to match feature vectors with contextual information and apply rule sets for reasoning, such as checking whether changes in operating conditions are consistent with valve operation history, or verifying whether the pump operation sequence supports the current operating condition judgment. The logic verification layer then outputs verification results to correct or confirm the operating condition identification output of the multi-sensor fusion classification layer, forming a closed loop of hierarchical fusion.

[0075] For example, when the feature vector identifies the current operating condition as a risk of cavitation, the logic verification layer activates the rule engine, queries the valve operation history and finds a recent record of the inlet valve being closed, and at the same time the separator liquid level data shows a sudden drop in liquid level, and the pump operation sequence record shows that the pump is in a stable operating state. Based on this, the rule engine determines that the risk of cavitation is credible, and thus corrects the operating condition output to the cavitation state, completing the verification.

[0076] Step 304: The cavitation risk assessment layer, the multi-sensor fusion classification layer, and the logic verification layer are fused hierarchically through a dedicated integration module to form a multi-level fusion strategy.

[0077] In some implementations, a dedicated integration module first receives the output of the cavitation risk assessment layer (such as cavitation risk status flags). This output is then synchronized with the operating condition probability output of the multi-sensor fusion classification layer, and processed by the integration module according to priority (i.e., the cavitation risk assessment layer executes first to ensure safety). If cavitation risk is identified, an alarm is triggered directly; otherwise, the data is passed to the multi-sensor fusion classification layer for operating condition classification. This allows the integration module to call the logic verification layer to verify the classification results based on contextual information such as valve operation history, separator level data, and pump operation sequence records (e.g., checking whether changes in operating conditions are consistent with valve operations). After verification, the integration module integrates the three layers of output through weighted or rule-based fusion to form a final output that includes a multi-level fusion strategy for operating condition status and confidence level.

[0078] For example, when the cavitation risk assessment layer outputs a no-cavitation-risk indicator based on the comparison between the inlet pressure of 0.5MPa and the saturated vapor pressure, and the multi-sensor fusion classification layer outputs a pure liquid phase probability of 80%, the dedicated integration module will activate the logic verification layer to check the valve operation history for no abnormalities, the separator liquid level for stability, and the pump operation sequence for normal operation. After the verification is passed, the integration module fuses the outputs of the three layers and finally forms a multi-level fusion strategy to confirm that the current operating condition is a pure liquid phase.

[0079] Based on the above technical solution, by calculating the difference between the inlet pressure and the saturated vapor pressure and setting a reasonable safety margin through the cavitation risk assessment layer, a safety barrier for cavitation protection can be established at the physical mechanism level. This solves the problem that traditional methods cannot predict cavitation risks in advance, effectively avoiding production accidents such as screw pump impeller damage and sudden drop in conveying efficiency caused by cavitation, and ensuring the safety of equipment operation and production continuity. At the same time, by extracting multi-dimensional feature vectors and using a multi-sensor fusion classification layer trained with labeled data, time-domain, frequency-domain, and derived physical features can be integrated. This allows machine learning algorithms such as random forests to accurately classify operating conditions such as pure liquid phase, gas-liquid two-phase, underload, and overload, breaking through the limitations of traditional single-parameter judgment of operating conditions. It solves the technical pain point of ambiguous operating condition identification in scenarios with high gas content and variable flow patterns, providing a reliable basis for the accurate selection of subsequent flow metering models and improving the basic accuracy of metering data. The logical verification layer, built based on process context information (valve operation history, separator level data, and pump operation sequence records), uses a rule engine to perform time-series tracing and scenario verification of the classification results. This solves the problem of misjudgment that easily occurs when relying solely on sensor data. Combining valve operation records with separator level changes clarifies the causes of operating conditions, further improving the reliability of operating condition identification and avoiding inaccurate production control due to misjudgment. Thus, by utilizing the hierarchical fusion of the three-layer structure, a complete multi-level fusion strategy is formed. This ensures the priority of cavitation risk assessment and achieves high accuracy and robustness in operating condition identification through the collaborative correction of the classification and verification layers. It fundamentally solves the problems of poor adaptability and insufficient accuracy in traditional operating condition identification, laying a solid foundation for the accurate measurement of flow rate, gas-liquid ratio, and pump efficiency in screw pump gas-liquid two-phase flow, effectively assisting industrial production in achieving precise control, cost reduction, and efficiency improvement.

[0080] In one possible implementation of the embodiments of this application, combined with Figure 1-4 As shown, the process of identifying the current operating condition by using a multi-level fusion strategy to identify the feature vector, including pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload conditions, can be achieved through the following steps 401 to 405, which are explained in detail below: Step 401: The cavitation risk assessment layer calls the inlet pressure from the original data stream. And temperature T, and calculate saturated vapor pressure using the Antoine equation. and import pressure With saturated vapor pressure The difference .

[0081] Among them, import pressure This refers to the pressure value collected in real time by a pressure sensor at the inlet of the screw pump. (Saturated vapor pressure) This refers to the pressure at which a liquid begins to vaporize at a specific temperature T, primarily calculated using the Antoine equation. (Difference) Import pressure With saturated vapor pressure The arithmetic difference is a physical quantity used to directly reflect the risk of cavitation.

[0082] In some implementations, the cavitation risk assessment layer obtains the inlet pressure in real time from the synchronously acquired raw data stream. By obtaining the measured value of temperature T, the specific constants A, B, and C of the medium can be obtained, allowing temperature T to be directly substituted into the Antoine equation. Calculated saturated vapor pressure The specific value. Then perform the subtraction operation. This will give you the difference result used for subsequent judgments.

[0083] For example, based on import pressure Given a pressure of 0.5 MPa and a temperature of T = 60 degrees Celsius, the saturated vapor pressure of water can be calculated using the Antoine equation. ≈0.02 MPa, then calculate the difference. =0.5-0.02=0.48 MPa, and this result can be used as a direct input for cavitation risk assessment.

[0084] Step 402: Preset safety margin threshold and the difference The results of the cavitation risk assessment are compared and output.

[0085] In some implementations, a preset safety margin threshold is set based on the media type and operational experience. (Safety margin threshold) Typically within the range of 20-50 kPa, the difference calculated in real time can be obtained. With threshold Perform a direct comparison and generate a judgment logic based on the comparison result: If Greater than or equal to If so, it is determined that there is no risk of cavitation, and the risk flag is output as "0"; if Less than If the risk of cavitation is detected, a high-priority alarm will be triggered immediately, and the risk flag will be output as "1".

[0086] For example, calling the already calculated The value is 0.48 MPa, with a preset safety margin threshold. The value is 0.03 MPa; the system performs a comparison and judgment, and since 0.48 MPa is much greater than 0.03 MPa, it meets the requirement. ≥ The conditions are met; therefore, the final output of the cavitation risk assessment layer is "0", meaning there is currently no cavitation risk.

[0087] Step 403: The multi-sensor fusion classification layer receives feature vectors and uses a pre-trained machine learning classifier to perform real-time analysis on the feature vectors, outputting the probability distribution of pure liquid phase, gas-liquid two-phase, cavitation, underload or overload states.

[0088] In some implementations, the multi-sensor fusion classification layer receives feature vectors containing time-domain, frequency-domain, and derived physical features, and calls a machine learning classifier that has been pre-trained based on historical labeled data. At this time, the machine learning classifier uses historical datasets containing various sensor data such as speed, torque, pressure, and vibration and their corresponding working condition labels during the training phase, and builds the model by learning the complex mapping relationship between features and working conditions.

[0089] Therefore, in its real-time analysis phase, the pre-trained classifier can quickly calculate and recognize patterns on the input feature vector, allowing each decision tree within it to make independent judgments and output a working condition category (e.g., tree 1 judges it as "pure liquid phase", tree 2 judges it as "gas-liquid two-phase", etc.). All judgment results can then be statistically output through a soft voting mechanism, and the voting ratio for each working condition category can be calculated to form a working condition probability distribution. At this point, the probability distribution will represent in numerical form the probability that the current equipment operating state belongs to one of the following working conditions: pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload.

[0090] For example, when the multi-sensor fusion classification layer receives the feature vector [500, 200, 10, 0.5, 1.5, 60, 100, 1.0], it can activate the pre-trained random forest classifier to perform real-time analysis. At this time, the random forest classifier will make a voting decision through multiple decision trees inside, and the final output probability distribution is: pure liquid phase probability 80%, gas-liquid two-phase probability 15%, cavitation probability 3%, underload probability 2%, and overload probability 0%. This result indicates that the system determines that the current working condition is most likely to be pure liquid phase.

[0091] Step 404: The logic verification layer verifies the output of the multi-sensor fusion classification layer based on the process context information and generates a verification result.

[0092] In some implementations, the logic verification layer acquires process context information in real time, including checking whether there are recent records of inlet valve closure or reduced opening in the valve operation history. This allows for querying whether the separator liquid level data shows a continuous decline or falls below the safety threshold, and analyzing whether there are sudden changes in pump speed or start-stop operations in the pump operation sequence record. The context information is then correlated with the operating condition probability distribution (e.g., checking whether an increase in cavitation probability in the probability distribution coincides with an inlet valve closure event in the valve operation history, or whether a high gas content probability is consistent with a downward trend in the separator liquid level), with a focus on checking the rationality of the temporal logic and causal relationship of the operating condition changes. Then, the pre-set rule engine makes a judgment: if the change in operating conditions can be found to have a corresponding process operation reason (e.g., the probability distribution shows an increased risk of cavitation, and the context information shows a recent operation to close the inlet valve), it is judged as consistent, and the original probability distribution is confirmed and output; if the change in operating conditions cannot be reasonably linked to the context information (e.g., the probability distribution shows a high gas content condition, but the separator level is stable and there is no gas inlet operation), it is judged as inconsistent. The logic verification layer then uses the context information to correct the original probability distribution, such as reducing the probability weight of operating condition categories that contradict the current process state and increasing the probability weight of operating condition categories that match the process state. Finally, the verified and corrected operating condition probability distribution is output.

[0093] For example, the multi-sensor fusion classification layer outputs a probability distribution of 80% for pure liquid phase, 15% for gas-liquid two-phase phase, and 5% for other phases. The logic verification layer is then activated, and querying the process context information reveals that: valve operation history shows the inlet and outlet valve openings remained stable over the past hour; separator level data remained at a stable high level of 65%; and pump operation sequence records show the speed remained stable at 500 rpm. Verifying this context with the initial judgment that "pure liquid phase is a high-probability event" confirms stable operating conditions and the absence of external factors that could induce gas-liquid two-phase phase or cavitation, and that the timing logic is consistent. Therefore, the output of the multi-sensor fusion classification layer is confirmed to be effective, and the final output probability distribution remains unchanged, still 80% for pure liquid phase and 15% for gas-liquid two-phase phase, etc.

[0094] Step 405: The dedicated integration module receives the cavitation risk indicator judgment results, probability distribution and verification results, and uses a weighted fusion algorithm or rule engine to perform hierarchical fusion to generate the current operating status and confidence level.

[0095] In some implementations, the dedicated integration module first synchronously receives outputs from three levels (cavitation risk indicators from the cavitation risk assessment layer, operating condition probability distributions from the multi-sensor fusion classification layer, and the verification results of the probability distributions from the logic verification layer). It then selects either a weighted fusion algorithm or a rule engine for hierarchical fusion based on a preset strategy: if a weighted fusion algorithm is used, fixed weights are assigned to each level's output (e.g., the cavitation risk indicator has the highest weight), and the probability distributions and risk indicators are integrated through weighted calculations to generate the final operating condition status and confidence level; if a rule engine is used, preset logical rules are executed (e.g., priority rules: if the cavitation risk indicator is "1", it is directly determined to be a cavitation state; otherwise, the operating condition corresponding to the highest probability in the probability distribution corrected by the logic verification layer is adopted), thus generating the final operating condition status and confidence level, and outputting a clear current operating condition status label and its confidence level.

[0096] For example, the dedicated integrated module receives the input cavitation risk flag = 0 (no risk), probability distribution (pure liquid phase 80%, gas-liquid two-phase 15%, etc.), and verification result = confirmed valid. The rule engine can then process this. Since the cavitation risk flag is 0, no cavitation alarm is triggered. Simultaneously, the verification result is confirmed, so the original probability distribution is adopted. The operating condition corresponding to the highest probability (80%), "pure liquid phase," is selected as the final state. When calculating the confidence level, the consistent output height of each layer is considered, thus generating the final result: the current operating condition is "pure liquid phase," with a confidence level of 95%.

[0097] Based on the above technical solution, the saturated vapor pressure and the difference between the two are accurately calculated using the Antoine equation. This is then compared with a preset safety margin threshold to output a risk indicator, constructing a safety barrier for cavitation protection at the physical mechanism level. This solves the problem that traditional methods cannot predict cavitation risks in advance, effectively avoiding production accidents such as screw pump impeller damage and sudden drops in conveying efficiency caused by cavitation, ensuring equipment operation safety and production continuity. Simultaneously, a pre-trained machine learning classifier is used to analyze and output the probability distribution of various operating conditions in real time, overcoming the limitations of traditional single-parameter judgment of operating conditions. This solves the technical pain point of ambiguous operating condition identification in scenarios with high gas content and variable flow patterns, providing a reliable basis for the accurate selection of subsequent flow metering models and improving the basic accuracy of metering data. Finally, the output of the multi-sensor fusion classification layer is verified based on process context information. By associating information such as valve operation history, separator level data, and pump operation sequence records, the temporal logic and causal relationship of operating condition changes are checked, solving the problem of misjudgment easily caused by relying solely on sensor data. This further improves the reliability of operating condition identification and avoids inaccurate production control due to misjudgment. This allows for the effective use of weighted fusion algorithms or rule engines to perform hierarchical fusion of the outputs from the first three layers, generating a clear current operating condition and confidence level. This not only ensures the priority of cavitation risk assessment but also achieves high accuracy and robustness in operating condition identification through the collaborative correction of the classification and verification layers. It fundamentally solves the problems of poor adaptability and insufficient accuracy in traditional operating condition identification, laying a solid foundation for the accurate measurement of flow rate, gas-liquid ratio, and pump efficiency of screw pumps, and helping industrial production achieve precise control, cost reduction, and efficiency improvement.

[0098] In one possible implementation of the embodiments of this application, combined with Figure 1-5 As shown, the process of adaptively selecting the corresponding model based on the current operating conditions, and obtaining the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two phases through mechanism model calculation and data-driven model thrust can be achieved through the following steps 501 to 504, which are explained in detail below: Step 501: Construct an adaptive model selection module based on the current working condition status label and confidence level, and select the corresponding mechanism model and data-driven model from the pre-built model library.

[0099] In some implementations, the adaptive model selection module initializes itself upon receiving the current operating condition status label and confidence level. At this point, the adaptive model selection module has a built-in model selection strategy (such as setting priority rules based on a confidence level threshold). That is, when the status label is pure liquid phase and the confidence level is higher than a preset threshold (e.g., 90%), a single-phase flow mechanism model is called from a pre-built model library for calculation; if the status label is gas-liquid two-phase or the confidence level is low, a data-driven model or a hybrid model is preferentially selected. Thus, by querying the model library index, matching the model instance corresponding to the operating condition label, and loading the relevant parameters, the selected mechanism model and data-driven model instance can be output, ensuring a high degree of model-condition fit.

[0100] It should be noted that the mechanistic models in the model library include baseline models based on pump physics equations. Data-driven models, on the other hand, utilize AI models trained on historical data for inference.

[0101] For example, if the operating condition identification module outputs the current operating condition as "pure liquid phase" with a confidence level of 95%, the adaptive model selection module, after receiving the state label and confidence level, will query and select a single-phase flow mechanism model from the pre-built model library since the confidence level is higher than the 90% threshold. At the same time, it will load the corresponding hydraulic equation parameters, while the data-driven model will select a neural network model trained based on historical pure liquid phase data.

[0102] Step 502: The mechanism model uses the hydraulic and thermodynamic equations of the screw pump to calculate the real-time feature vector, and obtains the basic flow rate, basic gas-liquid ratio and basic pump efficiency.

[0103] In some implementations, when the mechanistic model is invoked, predefined parameters from the screw pump's hydraulic and thermodynamic equations are loaded simultaneously. The model then receives a real-time feature vector containing characteristic values ​​such as rotational speed, torque, and pressure as input. Using the pressure difference and rotational speed parameters in the feature vector, the theoretical flow rate relationship can be calculated to obtain the base flow rate. Simultaneously, the energy balance is calculated based on the temperature and pressure parameters in the feature vector to obtain the base pump efficiency. Finally, the gas-liquid ratio is derived from the medium properties in the hydraulic equations, thus integrating all calculation results to output the base flow rate, base gas-liquid ratio, and base pump efficiency.

[0104] For example, when the real-time feature vector includes values ​​such as rotational speed of 500 rpm, torque of 200 Nm, inlet pressure of 0.5 MPa, and outlet pressure of 1.5 MPa, the mechanism model calls the hydraulic equation, uses a pressure difference of 1.0 MPa and rotational speed to calculate the basic flow rate as 10 m³ / h, and the thermodynamic equation calculates the basic pump efficiency as 85% based on a temperature of 60°C, and derives the basic gas-liquid ratio as 0, thus connecting with the identification results of the pure liquid phase working condition.

[0105] Step 503: The data-driven model uses a pre-trained neural network or support vector machine to process the same feature vector to obtain the flow correction value, gas-liquid ratio correction value, and pump efficiency correction value.

[0106] In some implementations, a data-driven model is invoked and loaded with a pre-trained neural network or support vector machine model. This allows the model to receive the same real-time feature vectors as the mechanistic model as input. The pre-trained neural network or support vector machine then performs forward propagation or decision function calculations on the feature vectors using internal weight matrices and algorithms, directly outputting predicted values ​​for flow rate, gas-liquid ratio, and pump efficiency. These predicted values ​​are then compared with the baseline values ​​calculated by the mechanistic model to calculate the deviation. The deviation is then used to directly generate corrected values ​​for flow rate, gas-liquid ratio, and pump efficiency, completing the real-time inference process and yielding the corrected values ​​for these parameters.

[0107] For example, when the real-time feature vector contains values ​​such as 500 rpm and 200 Nm of torque, the data-driven model can use a pre-trained neural network to process the vector and output a flow rate correction value of -0.1 m³ / h, a gas-liquid ratio correction value of 0, and a pump efficiency correction value of -1%, which connects with the basic calculation values ​​of the mechanism model under pure liquid phase conditions.

[0108] Step 504: The outputs of the mechanism model and the data-driven model are fused using a digital twin collaborative verifier to calculate the optimal estimate and generate the gas-liquid two-phase flow rate, gas-liquid ratio and pump efficiency, along with confidence indexes.

[0109] Among them, the digital twin collaborative verifier is a software component based on digital twin technology, used to simulate the real operating state of a screw pump and verify the rationality of the model output.

[0110] In some implementations, a digital twin co-verifier is directly activated, receiving the baseline flow rate, baseline gas-liquid ratio, and baseline pump efficiency calculated by the mechanistic model, as well as the flow rate correction, gas-liquid ratio correction, and pump efficiency correction values ​​output by the data-driven model. The verifier then uses a built-in fusion algorithm, such as a weighted average or extended Kalman filter, to co-calculate the outputs of the mechanistic and data-driven models, with the weights dynamically adjusted based on the model's historical accuracy or real-time confidence level. Finally, the verifier integrates the baseline and correction values ​​to obtain the optimal estimates of the gas-liquid two-phase flow rate, gas-liquid ratio, and pump efficiency. Simultaneously, the verifier generates a confidence index based on residuals or consistency analysis during the fusion process and appends it to the output.

[0111] The fusion process of digital twin collaborative verifiers can effectively balance the physical rationality of the mechanistic model and the adaptability of the data-driven model, but it depends on the accuracy of the digital twin model, and the confidence calculation needs to take into account the quality of the sensor data.

[0112] For example, the mechanistic model outputs a basic flow rate of 10 m³ / h, a basic gas-liquid ratio of 0, and a basic pump efficiency of 85%. The data-driven model outputs a flow rate correction of -0.2 m³ / h, a gas-liquid ratio correction of 0, and a pump efficiency correction of -1%. At this point, the digital twin collaborative verifier will calculate the optimal estimate through weighted fusion, obtaining a two-phase flow rate of 9.9 m³ / h, a gas-liquid ratio of 0, and a pump efficiency of 84%. Based on model consistency, it generates a confidence index of 95%, thus connecting the complete processing flow of the pure liquid phase condition.

[0113] Based on the above technical solutions, by accurately matching mechanistic models and data-driven models from a pre-built model library using operating condition labels and confidence levels, the technical problem of poor adaptability of traditional metrology models to complex operating conditions is solved. This eliminates reliance on a single fixed model, allowing for dynamic selection of suitable models based on different operating conditions such as pure liquid phase, gas-liquid two-phase, and cavitation. This ensures a high degree of consistency between the metrology model and actual operating conditions in complex scenarios such as high gas content and variable flow patterns. Simultaneously, the mechanistic model calculates basic flow rate, gas-liquid ratio, and pump efficiency based on the hydraulic and thermodynamic equations of a screw pump, guaranteeing the physical rationality and extrapolation of the metrology results and avoiding deviations from actual physical laws that may occur with purely data-driven models. Furthermore, the structure of neural networks or support vector machines processing the same feature vector effectively compensates for the error shortcomings of mechanistic models when dealing with complex factors such as changes in media properties and equipment wear. Their synergistic effect solves the problem of single models struggling to balance physical accuracy and adaptability to complex scenarios, improving the basic accuracy of metrology parameters. The digital twin collaborative verifier fuses the outputs of two types of models and generates a final result with confidence index. It not only achieves the optimal estimation of measurement data, but also clearly reflects the reliability of the feedback data. It solves the problem that traditional measurement methods cannot quantify the credibility of the results, and provides accurate and reliable data support for process control and equipment operation and maintenance in industrial production. It effectively improves production efficiency, reduces operation and maintenance costs, and promotes the intelligent upgrade of screw pump gas-liquid two-phase flow measurement from passive measurement to active adaptation and precise control.

[0114] In one possible implementation of the embodiments of this application, combined with Figure 1-6 As shown, the process of acquiring the raw data stream by real-time acquisition of the operating parameters of the screw pump can be achieved through the following steps 601 to 605, which are explained in detail below: Step 601: Configure a multimodal sensor array, which includes a non-contact digital torque sensor, a true RMS power measurement module, an inlet pressure sensor and an outlet pressure sensor with temperature compensation, a temperature sensor, a combination of a high-frequency vibration sensor and an acoustic emission sensor.

[0115] Among them, the sampling frequency of the high-frequency vibration sensor is no less than 10kHz, and the frequency range of the acoustic emission sensor is 100kHz-1MHz, which are used to capture the subtle features of cavitation and high gas content conditions.

[0116] In some implementations, before use, a non-contact digital torque sensor can be installed on the motor shaft end to measure shaft torque, based on the screw pump's structure and operating characteristics. A true RMS power measurement module is connected to the power input to monitor power. Temperature-compensated pressure sensors are installed at the pump's inlet and outlet flanges to collect inlet and outlet pressure data. A PT100 temperature sensor is installed on the pump body surface to detect temperature. High-frequency vibration and acoustic emission sensors are installed near the pump bearing housing to capture vibration and acoustic emission signals. The sampling frequency of the high-frequency vibration sensor is set to at least 10kHz to ensure high-frequency vibration characteristics are captured, while the acoustic emission sensor's frequency range is configured to cover the typical frequency bands for cavitation and high gas content conditions (100kHz-1MHz). Calibration and debugging ensure synchronized triggering of all sensor signals, forming a complete sensor array for real-time acquisition of the screw pump's operating parameters.

[0117] It should be noted that when configuring a multimodal sensor array, it is necessary to ensure that the sensor installation position avoids strong electromagnetic interference and mechanical vibration. Non-contact torque sensors should be installed in strict alignment to reduce measurement errors. Pressure sensors should be in direct contact with the medium to avoid blockage of the pressure tapping tube. Vibration and acoustic emission sensors should be close to the pump bearing housing to capture effective signals.

[0118] Step 602: Using the hardware synchronous acquisition unit, the FPGA chip drives all sensors to trigger synchronous acquisition at the same sampling frequency.

[0119] The clock synchronization adopts the IEEE 1588 precision time protocol (a network clock synchronization protocol that can achieve sub-microsecond time synchronization), with time synchronization accuracy reaching the microsecond level, ensuring the timestamp alignment of multi-source data.

[0120] In some implementations, a hardware synchronous acquisition unit is built using a Xilinx Artix-7 series FPGA chip. This allows for the configuration of multiple parallel acquisition channels on the chip, each connected to a different sensor signal. Therefore, during operation, a synchronization signal can be obtained from the network master clock via the IEEE 1588 precision time protocol, and clock domain synchronization is achieved within the FPGA, ensuring that all acquisition channels use a unified 10kHz sampling clock. Simultaneously, a synchronous trigger mechanism is used; upon receiving a trigger signal, all channels begin acquisition simultaneously, adding a timestamp accurate to the microsecond level to the data at each sampling point. Finally, the timestamped multi-channel data is transferred to the storage area via DMA to complete data alignment.

[0121] Step 603: Use an adaptive filtering algorithm based on wavelet transform to perform real-time noise reduction on the original signal to generate a preprocessed signal.

[0122] The wavelet basis function chosen is db4 to ensure tight support and regularity, and the decomposition layer is 5. Soft thresholding is used to remove power frequency interference and random noise by setting a threshold to zero or reduce small coefficients.

[0123] In some implementations, an adaptive filtering algorithm based on wavelet transform is used to perform real-time noise reduction on the original signal. Then, the db4 wavelet basis function is selected as the transform basis, and the number of decomposition levels is set to 5, which directly decomposes the signal into approximate coefficients and detail coefficients. Soft thresholding is then applied to threshold the detail coefficients, removing power frequency interference and random noise, allowing the reconstructed thresholded wavelet coefficients to generate the preprocessed signal.

[0124] Step 604: Perform real-time physical rationality verification on the preprocessed signal through a dedicated integrated module. If the verification fails, trigger a data quality alarm and mark the abnormal data.

[0125] In some implementations, the dedicated integrated module receives the preprocessed sensor signal and can directly call the Antoine equation to calculate the saturated vapor pressure of the current medium based on the real-time temperature value. Then, three parallel verification channels are used: channel one compares whether the difference between the inlet pressure and the saturated vapor pressure is greater than the safety threshold; channel two checks whether the outlet pressure is always higher than the inlet pressure; and channel three verifies whether the mechanical efficiency converted from torque and electrical power is within the preset range of 90%-110%. If any verification fails, a data quality alarm is immediately triggered, and an anomaly flag is added to the data packet header. The verification results can then be packaged and output together with the original data.

[0126] Step 605: Pack the verified signals into a time series raw data stream and attach a timestamp and quality flag.

[0127] In some implementations, the verified sensor signals are packaged into data packets according to the acquisition time sequence. Each data packet contains synchronous readings from multiple sensors. The hardware synchronous acquisition unit then obtains the corresponding timestamp of each data packet. Combined with the quality flag generated from the verification results, a timestamp accurate to the microsecond level and a quality flag (such as "valid" or "abnormal") are added to each data packet. This allows the data packets to be directly written sequentially into a circular buffer. Finally, the data in the circular buffer is output in real time for subsequent feature extraction modules to calculate time-domain and frequency-domain features, and for the operating condition identification module to determine states such as pure liquid phase or gas-liquid two-phase.

[0128] It should be noted that the buffer size is set to 1000 data packets, and the oldest data will be automatically overwritten when the buffer is full.

[0129] Based on the above technical solution, by accurately matching the corresponding mechanistic model and data-driven model from the pre-built model library using the current operating condition status label and confidence level, the technical problem of statistical models being difficult to adapt to the complex and variable operating conditions of screw pumps can be solved. This ensures that the optimal calculation model can be selected under different operating conditions such as pure liquid phase, gas-liquid two-phase, and cavitation. The mechanistic model guarantees the physical rationality and extrapolation of the basic flow rate, basic gas-liquid ratio, and basic pump efficiency results, avoiding measurement deviations that deviate from actual physical laws and meeting the core requirement of scientific measurement data in industrial production. The data-driven model accurately captures complex error factors not covered by the mechanistic model and outputs corresponding correction values, making up for the accuracy shortcomings of the pure mechanistic model when dealing with complex scenarios such as changes in media properties and slight equipment wear, further improving the accuracy of measurement data. In collaboration with a digital twin collaborative verifier, not only are the optimal estimates of gas-liquid two-phase flow rate, gas-liquid ratio, and pump efficiency obtained, but also confidence indexes are added, solving the problem that traditional measurement methods cannot quantify the reliability of results, and providing accurate and reliable data support for production control and equipment operation and maintenance.

[0130] In one possible implementation of the embodiments of this application, combined with Figure 1-7 As shown, the process of calibrating and optimizing the internal parameters of the fusion mechanism model and the data-driven model based on the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two-phase system using stable operating conditions or external calibration data can be achieved through the following steps 701 to 703, which are explained in detail below: Step 701: Collect the operating status of the screw pump. When a stable pure liquid phase flow condition is identified, mark the corresponding time period as the stable operating condition period, or obtain the reference true values ​​of flow rate, gas-liquid ratio and pump efficiency through an external calibration device.

[0131] In some implementations, a data acquisition unit acquires real-time multi-sensor data streams of the screw pump, including inlet pressure, outlet pressure, speed, torque, and vibration signals. This allows for moving average processing of the data within a continuous time window, calculating the volatility of key parameters relative to their average values. When the volatility of inlet pressure, outlet pressure, speed, and torque consistently falls below preset thresholds, a preliminary determination of a stable state is made. At this point, a built-in physical constraint rule library is invoked for verification. Rules include verifying whether the outlet pressure is consistently higher than the inlet pressure, whether the calculated effective net positive suction head (NPSH) is significantly higher than the required NPSH to avoid cavitation, and whether the dominant frequency energy of the vibration signal is within the typical low-frequency range of a pure liquid phase. If all verifications pass, the system is ultimately determined to be in a stable pure liquid phase flow condition. The start and end timestamps of this period, along with all corresponding sensor data, are packaged and marked as a stable operating period. Simultaneously, if the system is connected to an external calibration device, high-frequency sampling is synchronously triggered during this stable operating period to acquire and record the reference true value time series data of flow rate, gas-liquid ratio, and pump efficiency within that period.

[0132] Step 702: Construct a calibration optimization module including a golden moment detection unit, a parameter deviation calculation unit, and an adaptive learning unit. The golden moment detection unit is used to determine the start and end of the stable operating period, the parameter deviation calculation unit is used to generate the deviation vector, and the adaptive learning unit is used to perform state estimation and optimization of the parameters of the mechanism model and the data-driven model.

[0133] In some implementations, the golden moment detection unit continuously receives real-time time series data of inlet pressure, outlet pressure, speed, and torque from the acquisition unit. It calculates whether the ratio of the standard deviation to the mean of each parameter within the sliding time window is simultaneously lower than a set threshold to determine the consistency of the time series. When the consistency condition is met, it is marked as the golden moment. At the beginning of the stable operating period, if any parameter fluctuation exceeds the threshold, it is marked as the end of the period, and the parameter deviation calculation unit is activated. It receives the time series data of gas-liquid two-phase flow rate, gas-liquid ratio, and pump efficiency calculated within the current stable period, and reads the reference true value time series data provided by the external calibration device within the same period. It performs a subtraction operation between the calculated value and the true value for each sampling point to obtain the flow rate deviation, gas-liquid ratio deviation, and pump efficiency deviation, and combines these three deviations into a deviation vector for output. The adaptive learning unit can then be activated. Its internally pre-set extended Kalman filter algorithm defines the volumetric efficiency coefficient in the mechanistic model and a set of key neural network weight parameters in the data-driven model as the state vector. It uses the deviation vector output by the parameter deviation calculation unit as the observation vector. The filter performs recursive calculations through two core steps: prediction and update. The prediction step predicts the state based on the assumption that the parameters change slowly, while the update step uses the latest deviation observations to correct the state predictions. This achieves optimized estimation of the volumetric efficiency coefficient and neural network weights, making the model's output as close as possible to the reference true value.

[0134] Step 703: Use a digital twin collaborative verifier to simulate the output of the optimized parameters under historical operating conditions, and verify the rationality of the parameters by calibration and optimization.

[0135] The verifier selects a representative set of historical operating condition data (such as data covering pure liquid phase, gas-liquid two-phase, etc.) from the historical operating condition database and uses it as input for simulation. Simultaneously, it loads the optimized parameters to be verified and runs its internal high-fidelity screw pump physical model and data-driven model to calculate the simulated output time series of flow rate, gas-liquid ratio, and pump efficiency at the corresponding historical moments. The verifier then compares these simulated outputs point-by-point with the verified true values ​​of the corresponding historical outputs stored in the historical database and calculates the overall consistency index (e.g., calculating the root mean square error of all data points and converting it to a consistency score of 0-100%). If the calculated consistency score is higher than a preset threshold (e.g., 95%), the optimized parameters are deemed reasonable, effective, and have good generalization ability. A parameter update command can then be sent to the dedicated integration module, securely and quickly writing the verified new parameters into the pre-built model library in a transactional manner, overwriting the original parameters, and notifying the flow metering core to start using the new parameters in the next calculation cycle.

[0136] Based on the above technical solution, by constructing a calibration optimization module, the golden moment detection unit accurately defines the start and end of the stable operating period, making the calibration timing more precise. The parameter deviation calculation unit clearly quantifies the difference between the model output and the reference true value by generating a deviation vector. The adaptive learning unit specifically estimates and optimizes the parameters of the mechanistic model and the data-driven model. The three work together to solve the problems of lack of systematic model parameter adjustment and inaccurate deviation correction, realizing the targeted optimization of model parameters and improving the adaptability of the metrology model in practical applications. Finally, a digital twin collaborative verifier is used to simulate the output of the optimized parameters under historical operating conditions and verify their rationality, forming a closed-loop calibration optimization mechanism. This solves the problems of lack of comprehensive verification after parameter optimization and potential adaptability limitations, ensuring that the optimized parameters can play a stable role in various historical operating conditions. This not only significantly improves the long-term stability and accuracy of the measurement of the flow rate, gas-liquid ratio, and pump efficiency of the screw pump gas-liquid two-phase flow, providing reliable data support for industrial production control, but also reduces the frequency of manual calibration, lowers operation and maintenance costs, and promotes the intelligent and efficient upgrading of metrology management in the production process.

[0137] When using integrated units, Figure 8 This diagram illustrates a possible structural schematic of the gas-liquid two-phase flow metering system based on screw pump operating condition self-identification involved in the above embodiments. The system includes a physical sensing module, a data acquisition and preprocessing module, an operating condition self-identification module, a flow metering core module, and a data storage and calibration module. The physical sensing module, composed of a multi-modal sensor array, is used to acquire the operating parameters of the screw pump in real time, obtaining the raw data stream. The data acquisition and preprocessing module performs synchronous, adaptive filtering and noise reduction, and feature extraction on the raw data stream to obtain feature vectors. The operating condition self-identification module, connected to the data acquisition and preprocessing unit, receives the feature vectors and uses a multi-level fusion strategy to identify the operating condition, outputting the current operating condition status and confidence level. The flow metering core module adaptively selects the mechanism model and data-driven model based on the current operating condition status to calculate the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two phases. The data storage and calibration module stores historical data and applies external calibration data to perform online calibration and optimization of model parameters.

[0138] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A gas-liquid two-phase flow metering method based on screw pump operating condition self-identification, characterized in that, include: The operating parameters of the screw pump are collected in real time to obtain a raw data stream, which includes speed, torque, electric power, inlet pressure, outlet pressure, temperature and vibration signal; The original data stream is preprocessed and features are extracted to obtain a feature vector, which includes time-domain features, frequency-domain features, and derived physical features. A multi-level fusion strategy is constructed by acquiring historical raw data streams. The multi-level fusion strategy includes a cavitation risk assessment layer, a multi-sensor fusion classification layer, and a logic verification layer. The multi-level fusion strategy is used to identify the operating conditions of the feature vector to obtain the current operating conditions, including pure liquid phase, gas-liquid two-phase, cavitation, underload or overload conditions. Based on the current operating conditions, the corresponding model is adaptively selected, and the flow rate, gas-liquid ratio and pump efficiency of the gas-liquid two phases are obtained by calculating the thrust of the mechanism model and the data-driven model. Based on the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two phases, the internal parameters of the fusion mechanism model and the data-driven model are calibrated and optimized using stable operating conditions or external calibration data.

2. The gas-liquid two-phase flow metering method based on screw pump operating condition self-identification according to claim 1, characterized in that, The process of preprocessing and feature extraction of the original data stream to obtain feature vectors specifically includes: An adaptive filtering algorithm based on wavelet transform is used to denoise the synchronized original data stream to remove power frequency interference and random noise. Extract time-domain statistical features, including the mean of rotational speed, the variance of torque, the peak factor of pressure, and the kurtosis of the vibration signal, from the original data stream after noise reduction. The vibration signal is subjected to short-time Fourier transform to extract frequency domain features including main frequency energy, bandwidth power, and spectral centroid. Derived physical characteristics are calculated based on the physical model of the pump, including dimensionless pressure rise coefficient, specific energy efficiency, and net positive suction head (NPSH). A multi-level feature fusion module is constructed to weight and fuse the time-domain features, frequency-domain features, and derived physical features to form a high-dimensional vector; Principal component analysis is used to reduce the dimensionality of the high-dimensional vector to obtain a feature vector, which is used for working condition identification.

3. The gas-liquid two-phase flow metering method based on screw pump operating condition self-identification according to claim 2, characterized in that, The process of acquiring historical raw data streams to construct a multi-level fusion strategy specifically includes: Obtain historical raw data streams and construct a cavitation risk assessment layer by setting safety margin thresholds; Based on the historical raw data stream, multi-dimensional feature vectors are extracted, and a multi-sensor fusion classification layer is trained using historical labeled data; Based on process context information, a logic verification layer is constructed through a rule engine. The process context information includes valve operation history, separator liquid level data, and pump operation sequence records. The cavitation risk assessment layer, multi-sensor fusion classification layer, and logic verification layer are fused hierarchically through a dedicated integration module to form a multi-level fusion strategy.

4. The gas-liquid two-phase flow metering method based on screw pump operating condition self-identification according to claim 3, characterized in that, The process of using the multi-level fusion strategy to identify the operating conditions of the feature vector to obtain the current operating conditions, including pure liquid phase, gas-liquid two-phase, cavitation, underload, or overload states, specifically includes: The cavitation risk assessment layer calls the inlet pressure in the original data stream. And temperature T, and calculate saturated vapor pressure using the Antoine equation. and import pressure With saturated vapor pressure The difference ; Preset safety margin threshold and the difference Compare the results and output the cavitation risk indicator assessment results; The multi-sensor fusion classification layer receives the feature vector, uses a pre-trained machine learning classifier to perform real-time analysis on the feature vector, and outputs the probability distribution of pure liquid phase, gas-liquid two-phase, cavitation, underload or overload state. The logic verification layer verifies the output of the multi-sensor fusion classification layer based on process context information and generates a verification result. The dedicated integration module receives the cavitation risk indicator judgment result, probability distribution and verification result, and performs hierarchical fusion using a weighted fusion algorithm or rule engine to generate the current operating status and confidence level.

5. The gas-liquid two-phase flow metering method based on screw pump operating condition self-identification according to claim 4, characterized in that, The process of adaptively selecting the corresponding model based on the current operating conditions, and obtaining the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two-phase system through mechanism model calculation and data-driven model thrust, specifically includes: An adaptive model selection module is constructed based on the status label and confidence level of the current working condition, and the corresponding mechanism model and data-driven model are selected from the pre-built model library. The mechanism model uses the hydraulic and thermodynamic equations of the screw pump to calculate the real-time feature vector, and obtains the basic flow rate, basic gas-liquid ratio and basic pump efficiency. The data-driven model uses a pre-trained neural network or support vector machine to process the same feature vector to obtain flow correction value, gas-liquid ratio correction value and pump efficiency correction value. The outputs of the mechanistic model and the data-driven model are fused by a digital twin collaborative validator to calculate the optimal estimate and generate the gas-liquid two-phase flow rate, gas-liquid ratio and pump efficiency, along with confidence indexes.

6. The gas-liquid two-phase flow metering method based on screw pump operating condition self-identification according to claim 5, characterized in that, The process of obtaining the raw data stream by real-time acquisition of the operating parameters of the screw pump specifically includes: A multimodal sensor array is configured, which includes a non-contact digital torque sensor, a true RMS power measurement module, an inlet pressure sensor and an outlet pressure sensor with temperature compensation, a temperature sensor, a combination of a high-frequency vibration sensor and an acoustic emission sensor; The hardware synchronous acquisition unit uses an FPGA chip to drive all sensors to trigger synchronous acquisition at the same sampling frequency. An adaptive filtering algorithm based on wavelet transform is used to perform real-time noise reduction on the original signal to generate a preprocessed signal; The preprocessed signal is physically validated in real time by a dedicated integrated module. If the validation fails, a data quality alarm is triggered and abnormal data is marked. The verified signals are packaged into a raw time series data stream and timestamps and quality flags are added.

7. The gas-liquid two-phase flow metering method based on screw pump operating condition self-identification according to claim 6, characterized in that, The process of calibrating and optimizing the internal parameters of the fusion mechanism model and the data-driven model based on the flow rate, gas-liquid ratio, and pump efficiency of the gas-liquid two phases using stable operating conditions or external calibration data specifically includes: The operating status of the screw pump is collected. When a stable flow condition of pure liquid phase is identified, the corresponding time period is marked as a stable operating condition period, or the reference true values ​​of flow rate, gas-liquid ratio and pump efficiency are obtained through an external calibration device. A calibration optimization module is constructed, comprising a golden moment detection unit, a parameter deviation calculation unit, and an adaptive learning unit. The golden moment detection unit is used to determine the start and end of a stable operating period. The parameter deviation calculation unit is used to generate a deviation vector. The adaptive learning unit is used to perform state estimation and optimization of the parameters of the mechanism model and the data-driven model. The optimized parameters are simulated under historical operating conditions using a digital twin collaborative verifier, and the rationality of the parameters is verified through calibration and optimization.

8. A gas-liquid two-phase flow metering system based on screw pump operating condition self-identification, characterized in that... The gas-liquid two-phase flow measurement method based on screw pump operating condition self-identification as described in any one of claims 1-9 includes a physical sensing module, a data acquisition and preprocessing module, an operating condition self-identification module, a flow measurement core module, and a data storage and calibration module. The physical sensing module consists of a multimodal sensor array, used to collect the operating parameters of the screw pump in real time and obtain the raw data stream; The data acquisition and preprocessing module is used to perform synchronous, adaptive filtering and noise reduction, and feature extraction on the raw data stream to obtain a feature vector; The working condition self-identification module is connected to the data acquisition and preprocessing unit, and is used to receive the feature vector, perform working condition identification using a multi-level fusion strategy, and output the current working condition status and confidence level. The flow metering core module is used to adaptively select the mechanism model and data-driven model according to the current operating conditions, and calculate the flow rate, gas-liquid ratio and pump efficiency of the gas-liquid two phases. The data storage and calibration module is used to store historical data and external calibration data, and to perform online calibration and optimization of model parameters.

9. The gas-liquid two-phase flow metering system based on screw pump operating condition self-identification according to claim 8, characterized in that, The data acquisition and preprocessing module includes a hardware synchronous acquisition subunit, an adaptive filtering and noise reduction subunit, a feature extraction subunit, and a feature fusion and dimensionality reduction subunit. The hardware synchronous acquisition subunit drives all sensors to synchronously acquire raw signals at the same sampling frequency. The adaptive filtering and denoising subunit uses an adaptive filtering algorithm based on wavelet transform to denoise the original signal and generate a preprocessed signal. The feature extraction subunit is used to extract time-domain features, frequency-domain features, and derived physical features from the preprocessed signal; The feature fusion and dimensionality reduction subunit is used to weight and fuse the time-domain features, frequency-domain features, and derived physical features to form a high-dimensional vector, and then use principal component analysis to reduce the dimensionality of the high-dimensional vector to obtain the feature vector. The data acquisition and preprocessing unit performs real-time physical validity checks on the preprocessed signals. If the check fails, a data quality alarm is triggered.

10. The gas-liquid two-phase flow metering system based on screw pump operating condition self-identification according to claim 9, characterized in that, The core of the flow metering system includes an adaptive model selector, a model library, a mechanistic model library, a data-driven model library, and a digital twin collaborative validator. The adaptive model selector is used to select the corresponding mechanism model and data-driven model from the pre-built model library according to the current working condition and confidence level; The model library includes single-phase flow mechanism models for pure liquid phase conditions, multiphase flow mixing models for gas-liquid two-phase conditions, and degradation models for abnormal conditions. The mechanism model library stores models based on the hydraulic and thermodynamic equations of screw pumps, which are used to calculate the basic flow rate, basic gas-liquid ratio and basic pump efficiency based on real-time feature vectors. The data-driven model library stores pre-trained neural network or support vector machine models, which are used to process the feature vectors to obtain flow correction values, gas-liquid ratio correction values, and pump efficiency correction values. The digital twin collaborative verifier is used to fuse the outputs of the mechanistic model and the data-driven model, calculate the optimal estimates of the gas-liquid two-phase flow rate, gas-liquid ratio and pump efficiency, and generate a confidence index.

Images