A moisture flow measurement method and device based on multi-sensor feature fusion and physical constraint LightGBM model

Abstract

The present application relates to the technical field of multiphase flow measurement of oil and gas fields, and discloses a wet gas flow measurement method and device based on multi-sensor feature fusion and a physically constrained LightGBM model. The method comprises the following steps: data acquisition; feature data processing; constructing a physically constrained LightGBM model through training, taking the volume flow coefficient and the mass flow coefficient as the training target, and embedding the obtained model in the form of a function into a single-chip microcomputer program; embedded reasoning. The device comprises a double-differential pressure mechanism and an ultrasonic mechanism connected in series in the same pipeline. The present application effectively senses the change of flow pattern through sensor combination; the model taking the flow coefficient as the target has excellent generalization ability; the landing from the algorithm to the embedded product is completed, the requirements of low cost, low power consumption, small size and high reliability of the instrument in the industrial field are met, and the conditions for large-scale industrial application are met.

Description

Technical Field

[0001] This invention relates to the field of multiphase flow measurement technology in oil and gas fields, specifically to a method and device for measuring wet gas flow based on multi-sensor feature fusion and a physically constrained LightGBM model. Background Technology

[0002] Multi-sensor fusion is a common approach in the field of multiphase flow (e.g., patent CN105222831A). However, existing technologies generally suffer from the following fundamental defects, resulting in poor performance in the complex condition of wellhead wet gas metering: 1. Sensor combinations fail to accurately characterize flow pattern changes: For example, there are measurement schemes based on in-pipe separation or guidance. For instance, patent CN105222831A mentions using a cyclone separator to forcibly adjust a complex flow pattern into a ring flow before measurement. On the one hand, this relies on a specific guiding structure, introducing problems such as pressure loss, structural complexity, and response hysteresis; on the other hand, it is difficult to adapt to rapid dynamic changes in flow patterns. For example, when the flow pattern is a slug flow (one stream of liquid and one stream of gas), it cannot form a ring flow as intended, but rather a stream of pure liquid and a stream of pure gas, and the changes are rapid. Another type uses a method without front-end rectification, directly measuring with a multi-sensor array, but data processing and calculation still rely on traditional formula derivation and correction. This traditional approach is largely wishful thinking. For example, patent CN104155471A mentions using ultrasonic attenuation rate to calculate gas phase content; however, ultrasonic attenuation in a humid environment is affected by many factors, such as droplet scattering and refraction at the gas-liquid interface. Different flow rates and flow patterns can result in attenuation rates that are not significantly different. Moreover, various influencing factors are often interrelated, and a simple mathematical formula cannot accurately express this relationship, which can lead to huge calculation errors.

[0003] 2. "Black box" modeling with poor generalization ability: Most solutions using AI algorithms such as neural networks and support vector machines typically establish an end-to-end mapping from sensor features to the final flow rate (Q_g, Q_l) (e.g., CN118069975A). This type of model heavily relies on the distribution of training data. When the field conditions (such as gas-liquid ratio, flow pattern) exceed the range of experimental data samples, the model may give physically unreasonable and highly erroneous predictions, lacking extrapolation and generalization ability.

[0004] 3. Reliance on high-performance computing platforms: Complex AI models typically run on industrial control computers or servers, making it difficult to deploy them in low-cost, low-power embedded instruments, which limits the industrial application of the technology (e.g., CN120510454A). Summary of the Invention

[0005] To address the technical problems in existing wellhead wet gas measurement technologies, such as the inability of sensor combinations to accurately characterize flow pattern changes, poor model generalization ability, and difficulty in embedded deployment, this invention provides a wet gas flow measurement method and device based on multi-sensor feature fusion and a physically constrained LightGBM model. The technical solution is as follows: A method for measuring moisture flow based on multi-sensor feature fusion and a physically constrained LightGBM model includes the following steps: (1) Data collection Sensor data collected includes the following sources: dual differential pressure signals ΔP1 and ΔP2, bidirectional propagation time difference Δt_h and Δt_v of dual ultrasonic waves and signal strength data, and temperature and pressure compensation signals for gas volume, including pressure P and temperature T. With a high-performance MCU as its core, it is responsible for signal acquisition, feature calculation, model inference, and result output; (2) Feature data processing High-value features are extracted from the raw data based on the flow pattern mechanism of the gas-liquid two-phase flow at the wellhead, forming a feature vector; (3) Construct a physically constrained LightGBM model through training The training objective is to output the volumetric flow rate coefficient k_v and the mass flow rate coefficient k_m; The feature data and target values ​​k_v1, k_v2, k_m1, and k_m2 are converted to integers, and the training and testing sets are divided. The LightGBM regression algorithm is used for training to obtain the prediction model. The obtained model is converted into C language code and embedded into the microcontroller program in the form of a function. (4) Embedded reasoning Inference is performed using the LightGBM model deployed in the MCU, outputting real-time k_v1, k_v2, k_m1, and k_m2; Qv and Qm are calculated by combining the predicted k_v1, k_v2, k_m1, and k_m2 with the real-time measured ΔP1 and ΔP2; and the gas volume Q_g, liquid volume Q_l, and gas-liquid ratio GLR are calculated based on Qv, Qm, and gas-liquid density.

[0006] Furthermore, in step (1), the dual differential pressure signals ΔP1 and ΔP2 are two response characteristics with different flow states provided by the Venturi tube structure section and the V-cone structure section connected in series in the same pipeline.

[0007] Furthermore, in step (1), the bidirectional propagation time difference Δt_h and Δt_v of the dual ultrasonic waves comes from two sets of sensors installed in a through-beam configuration, one set being a horizontal path and the other a vertical path.

[0008] Furthermore, the high-value characteristics of step (2) include the following aspects: A. From differential pressure: instantaneous values, moving average, variance, peak-to-trough values, and the ratio of ΔP1 / ΔP2 of ΔP1 and ΔP2; B. From ultrasound: Horizontal downstream time difference Δt_h1; Horizontal upstream time difference Δt_h2; Vertical downstream time difference Δt_v1; Vertical upstream time difference Δt_v2; Bidirectional time difference sum: Δt_h1+Δt_h2, Δt_v1+Δt_v2; Bidirectional time difference: Δt_h2-Δt_h1, Δt_v2-Δt_v1; Signal attenuation intensity α: Transmission is at a fixed intensity, only the reception intensity needs to be detected; C. Derived from temperature and pressure: temperature T and pressure P; D. Combinatorial characteristics: including ΔP1 / (Δt_h1+Δt_h2)².

[0009] Furthermore, the sources of the volumetric flow rate coefficient k_v and the mass flow rate coefficient k_m in step (3) are: based on , , make: , ， In the above formula: It is volumetric flow rate. It refers to mass flow rate, where k is a coefficient related to flow pattern, Reynolds number, pipe diameter, and throttling device factors. For the differential pressure before and after the throttling device, The density is the medium.

[0010] Furthermore, in step (3), the training data is less than 10,000 records, the number of trees is 600, the number of leaves is 12, the depth is 8, and the learning rate is 0.05.

[0011] Furthermore, the calculation formulas for Qv and Qm in step (4) are as follows: ; ; ; ; ; .

[0012] A moisture flow measurement device, which is compatible with the moisture flow measurement method based on multi-sensor feature fusion and physical constraint LightGBM model described above, includes a dual differential pressure mechanism and an ultrasonic mechanism connected in series in the same pipeline.

[0013] Furthermore, the dual differential pressure mechanism includes a Venturi structure section and a V-cone structure section connected in series in the same pipeline. The Venturi structure section is provided with a first differential pressure tap H and a second differential pressure tap L, and the V-cone structure section is provided with a second differential pressure tap H and a second differential pressure tap L. The ultrasonic mechanism includes ultrasonic transducers A, B, C, and D, a temperature sensor, and a pressure sensor. Ultrasonic transducers A and B are vertically mounted facing each other, while ultrasonic transducers C and D are horizontally mounted facing each other.

[0014] Furthermore, the ultrasonic transducer A and ultrasonic transducer B are both at an angle of 30° to the axis.

[0015] To enable those skilled in the art to better understand the present invention, the relevant basic principles are briefly described below: 1. To address the challenges of high gas-liquid ratios and variable flow patterns in wellhead humid environments, a dual differential pressure sensing unit consisting of a Venturi tube and a V-cone in series, along with a horizontally and vertically bidirectional ultrasonic sensing unit, was designed. This combination is not merely a simple arrangement of sensors; rather, through analysis of the wellhead humid flow pattern mechanism, it enables a synergistic effect between the output characteristics of different sensors, providing the model with complementary and high-value information that profoundly reflects flow pattern changes. The Venturi tube exhibits high sensitivity at high gas phase velocities, but its measured values ​​fluctuate significantly under drastic flow pattern changes. The V-cone performs better in measuring low liquid phase content and flow pattern stability, but its sensitivity may decrease at extremely high gas flow velocities. The series connection of these two sensors ensures data sensitivity and stability over a wide flow range. The ultrasonic transit time and signal attenuation along the horizontal and vertical paths allow for the perception of anisotropy in phase distribution across the fluid cross-section. For example, in laminar flow, the measured values ​​along the vertical and horizontal paths show significant differences, while the differences are smaller in bubbly or mist-like flow. This dual-path comparative information, which a single ultrasonic sensor cannot provide, is crucial for identifying complex flow patterns. Differential pressure sensors provide macroscopic, total flow information related to the square of the flow velocity, while ultrasonic sensors provide microscopic, local information related to the phase distribution along the sound wave propagation path. This invention fuses this information at these two different scales using LightGBM, enabling the model not only to estimate flow rate but also to sense and adapt to changes in flow patterns, thus maintaining high accuracy even during drastic flow pattern changes.

[0016] 2. Breaking away from the "black box" paradigm of AI models directly regressing flow rate, the LightGBM model innovatively outputs "flow rate coefficients" (k_v, k_m) with clear physical meaning. These coefficients are then embedded into the classic differential pressure formula, transforming the model's learning objective from "flow rate" to a "correction factor for the physical formula." This confines data-driven learning within the framework of physical laws, significantly enhancing the model's generalization ability under unknown conditions and the physical rationality of its output. In simpler terms, if "flow rate" is directly used as the training objective, a large training sample size is needed to ensure that each combination of feature vectors yields a relatively accurate flow rate value, thus requiring the model to derive a more accurate value. However, regardless of the size of the training sample, these samples are discrete data and cannot cover all situations. Furthermore, once the model exceeds a certain order of magnitude, it becomes unusable in embedded environments. Moreover, the final required flow rate value is not entirely devoid of physical laws from the sensor data. Therefore, combining physical formulas with the intelligent model reduces the model's resource requirements while improving its generalization ability.

[0017] 3. A closed loop has been completed, from complex algorithms to industrially usable embedded products. Through targeted lightweight model design and optimization, the LightGBM model has been successfully deployed on resource-constrained STM32 series MCUs, achieving miniaturization, low power consumption, and low cost of high-performance algorithms, and resolving the contradiction between "easy to use" and "usable".

[0018] 4. The reasoning behind setting the volumetric flow rate coefficient k_v and the mass flow rate coefficient k_m as the training targets is as follows: First, flow meters based on the differential pressure principle always follow physical formulas: ① or ② It is volumetric flow rate. It refers to mass flow rate, where k is a coefficient related to factors such as flow pattern, Reynolds number, pipe diameter, and throttling device. For the differential pressure before and after the throttling device ( or ), Let k be the density of the medium. In single-phase flow, k varies within a relatively small range. In multiphase flow metering, k is a variable that changes drastically with the flow regime, and the primary reason for this drastic change is the drastic change in the gas-liquid ratio, which in turn is reflected in the overall density of the medium. Above. That is, we can connect k with... At the same time, let , ③ then: ④ ⑤ Obtain the total volumetric flow rate Then, combine the gas-liquid density and (Determined by the composition of the gas well's output) The gas-liquid flow rate (under operating conditions) can be calculated: ⑦ ⑧ in, This represents the gas volumetric flow rate. This refers to the gas volumetric flow rate (operating condition). The density of the gas under operating conditions is given by the following relationship with its standard density: 9 in For standard gas density, Standard atmospheric pressure The standard thermodynamic temperature is... P For operating pressure, T The operating temperature is used as the reference temperature. After temperature and pressure compensation, the standard flow rate can be obtained from the operating flow rate. ⑩ The liquid volume does not change significantly under operating and standard conditions, so it can be assumed that: ⑪ In summary, once the flow coefficients k_v and k_m are obtained, the gas-liquid flow rate can be calculated using a small number of physical formulas.

[0019] 5. Regarding training preparation: (a) In the laboratory, various flow patterns (bubble, slug, annular, laminar, mist, etc.) and various flow combinations are simulated to collect sensor data and accurately measure the actual gas volume. , liquid volume (Standard volumetric flow rate) and record the data through a data acquisition system.

[0020] (b) Based on the above flow rate values , and gas-liquid density combined with operating temperature T Pressure value P Calculate Qv and Qm (operating flow rate). ⑫ ⑬ Since gas flow rate varies significantly between operating and standard conditions, while liquid flow rate does not, equation ⑫ only requires conversion between standard and operating conditions for gas volumetric flow rate. Then, combining the measured ΔP1 or ΔP2, the actual volumetric and mass flow rate coefficients k_v1 (through ΔP1), k_v2 (through ΔP2), and k_m1 (through ΔP1), k_m2 (through ΔP2) under that operating condition can be calculated.

[0021] Compared with the prior art, the present invention has the following main advantages: 1. The carefully designed sensor combination of this invention provides rich and complementary feature information, especially the synergy of dual-path ultrasonic waves and dual differential pressure, which can effectively sense changes in flow patterns.

[0022] 2. By outputting a "flow coefficient" and using physical formulas as constraints in the output layer, this invention ensures that even in new operating conditions not covered by training data, the model's prediction results are strictly constrained within a reasonable range of physical laws, and the prediction deviation is significantly lower than that of models that directly use flow as the target, demonstrating excellent generalization ability.

[0023] 3. The LightGBM model of this invention can efficiently learn the complex nonlinear mapping between feature information and flow coefficient, thereby maintaining high measurement accuracy under a wide range of gas-liquid ratio and flow pattern changes.

[0024] 4. This invention has successfully translated the algorithm into an embedded product. The optimized model can run in real-time on a low-cost MCU, meeting all the requirements of industrial applications for low cost, low power consumption, small size, and high reliability. It is ready for large-scale industrial application and possesses true industrial practicality.

[0025] 5. This invention creatively proposes and implements a new paradigm of "deep integration of physical mechanisms and AI intelligent algorithms," the core of which lies in: using physical formulas The AI ​​model is defined with a learning objective (k) rather than a final answer (Q). A multi-sensor collaborative hardware solution capable of deep flow pattern perception is specifically designed to better predict k. This is a system-level innovation with tightly coupled components, fundamentally solving the challenges of accuracy, generalization, and deployment in moisture measurement, providing a novel and highly competitive technological solution for this field.

[0026] 6. This invention is an online, real-time, high-precision flow measurement method for wet gas conditions with large variations in gas-liquid ratio and drastic changes in flow pattern. It is applicable to wellhead flow measurement of natural gas wells or coalbed methane wells.

[0027] 7. This invention proposes a hybrid modeling paradigm of "flow coefficient mapping based on physical constraints" and designs a hardware scheme of "mechanism-oriented feature collaborative perception" for this purpose. Ultimately, it realizes the collaborative optimization of "model-hardware" and solves the problems of adaptability, accuracy, generalization and embedded deployment of moisture measurement. Attached Figure Description

[0028] Figure 1 This is a MAPE plot showing the interpolation error when the target value is the flow rate, as shown in the comparative example of the present invention. Figure 2 This is a MAPE plot showing the interpolation error when the flow coefficient is the target in the comparative example of the present invention. Figure 3 This is a MAPE diagram showing the extrapolation error when the target value is the flow rate, as a comparative example of the present invention. Figure 4 This is a MAPE diagram showing the extrapolation error when the flow coefficient is the target in the comparative example of the present invention. Figure 5 Vertical cross-sectional schematic diagram of the dual differential pressure mechanism and ultrasonic mechanism of the present invention; Figure 6 A horizontal cross-sectional schematic diagram of the invention of the dual differential pressure mechanism and ultrasonic mechanism; Figure 5 and Figure 6 In the diagram: 1-Ultrasonic transducer A; 2-Ultrasonic transducer B; 3-Temperature sensor; 4-Pressure sensor; 5-First differential pressure tap H; 6-First differential pressure tap L; 7-Venturi structure section; 8-V-cone structure section; 9-Second differential pressure tap H; 10-Second differential pressure tap L; 11-Ultrasonic transducer C; 12-Ultrasonic transducer D; 13-Pipeline. Detailed Implementation

[0029] The present invention will now be described in detail with reference to embodiments, comparative examples and figures. Example 1

[0030] See Figure 5 and Figure 6 A method for measuring moisture flow based on multi-sensor feature fusion and a physically constrained LightGBM model includes the following steps: (1) Data collection The sensor data collected includes the following sources: dual differential pressure signals ΔP1 and ΔP2, dual ultrasonic bidirectional propagation time differences Δt_h and Δt_v and signal strength data, and temperature and pressure compensation signals for gas volume, including pressure P and temperature T. The dual differential pressure signals ΔP1 and ΔP2 are two response characteristics with different flow states provided by the Venturi tube structure section 7 and the V-cone structure section 8 connected in series in the same pipe 13. The dual ultrasonic bidirectional propagation time differences Δt_h and Δt_v come from two sets of sensors installed in opposite directions, one set being a horizontal path and the other a vertical path.

[0031] With a high-performance MCU as its core, it is responsible for signal acquisition, feature calculation, model inference, and result output; (2) Feature data processing High-value features are extracted from the raw data based on the flow pattern mechanism of the gas-liquid two-phase flow at the wellhead, forming a feature vector; High-value characteristics include the following aspects: A. From differential pressure: instantaneous values, moving average, variance, peak-to-trough values, and the ratio of ΔP1 / ΔP2 of ΔP1 and ΔP2; B. From ultrasound: Horizontal downstream time difference Δt_h1; Horizontal upstream time difference Δt_h2; Vertical downstream time difference Δt_v1; Vertical upstream time difference Δt_v2; Bidirectional time difference sum: Δt_h1+Δt_h2, Δt_v1+Δt_v2; Bidirectional time difference: Δt_h2-Δt_h1, Δt_v2-Δt_v1; Signal attenuation intensity α: Transmission is at a fixed intensity, only the reception intensity needs to be detected; C. Derived from temperature and pressure: temperature T and pressure P; D. Combinatorial characteristics: including ΔP1 / (Δt_h1+Δt_h2)².

[0032] (3) Construct a physically constrained LightGBM model through training The training objective is to output the volumetric flow rate coefficient k_v and the mass flow rate coefficient k_m; The feature data and target values ​​k_v1, k_v2, k_m1, and k_m2 are converted to integers, and the training and testing sets are divided. The LightGBM regression algorithm is used for training to obtain the prediction model. The obtained model is converted into C language code and embedded into the microcontroller program in the form of a function. The sources of the volumetric flow rate coefficient k_v and the mass flow rate coefficient k_m are: based on , , make: , ， In the above formula: It is volumetric flow rate. It refers to mass flow rate, where k is a coefficient related to flow pattern, Reynolds number, pipe diameter, and throttling device factors. For the differential pressure before and after the throttling device, The density is the medium.

[0033] Training data < 10,000 records, number of trees 600, number of leaves 12, depth 8, learning rate 0.05.

[0034] (4) Embedded reasoning Inference is performed using the LightGBM model deployed in the MCU, outputting real-time k_v1, k_v2, k_m1, and k_m2; Qv and Qm are calculated by combining the predicted k_v1, k_v2, k_m1, and k_m2 with the real-time measured ΔP1 and ΔP2; and the gas volume Q_g, liquid volume Q_l, and gas-liquid ratio GLR are calculated based on Qv, Qm, and gas-liquid density.

[0035] The formulas for calculating Qv and Qm are as follows: ; ; ; ; ; . Example 2

[0036] See Figure 5 and Figure 6 A moisture flow measurement device, matching the moisture flow measurement method based on multi-sensor feature fusion and a physically constrained LightGBM model described in Example 1, includes a dual differential pressure mechanism and an ultrasonic mechanism connected in series in the same pipe 13. The dual differential pressure mechanism includes a Venturi section 7 and a V-cone section 8 connected in series in the same pipe 13. The Venturi section 7 has first differential pressure taps H5 and L6, and the V-cone section 8 has second differential pressure taps H9 and L10. The ultrasonic mechanism includes ultrasonic transducers A1, B2, C11, and D12, a temperature sensor 3, and a pressure sensor 4. Ultrasonic transducers A1 and B2 are vertically mounted facing each other, while ultrasonic transducers C11 and D12 are horizontally mounted facing each other. The angle between ultrasonic transducers A1 and B2 and the axial direction is 30°.

[0037] Comparative example Taking training data of (50–120) m³ / h and (180–260) m³ / h as examples, when the actual flow rate is in the range of (130–170) m³ / h (interpolation) or the flow rate is >260 m³ / h (extrapolation), the prediction errors of the two models can be compared. Figures 1 to 4 The MAPE error shown in the figure is listed in Table 1: Table 1 Comparison of Prediction Errors of Different Models , Table 1 (comparison table) clearly shows that: regarding interpolation error, the error is 17.46% when using flow rate as the model target, and 12.02% when using flow coefficient as the model target, the latter being 5.44% lower than the former; regarding extrapolation error, the error is 18.71% when using flow rate as the model target, and 10.25% when using flow coefficient as the model target, the latter being 8.46% lower than the former. Therefore, whether for interpolation or extrapolation prediction, using "flow coefficient" instead of "flow rate" as the model target can significantly reduce the error.

Claims

1. A method for measuring moisture flow rate based on multi-sensor feature fusion and a physically constrained LightGBM model, characterized in that, Includes the following steps: (1) Data collection Sensor data collected includes the following sources: dual differential pressure signals ΔP1 and ΔP2, bidirectional propagation time difference Δt_h and Δt_v of dual ultrasonic waves and signal strength data, and temperature and pressure compensation signals for gas volume, including pressure P and temperature T. With a high-performance MCU as its core, it is responsible for signal acquisition, feature calculation, model inference, and result output; (2) Feature data processing High-value features are extracted from the raw data based on the flow pattern mechanism of the gas-liquid two-phase flow at the wellhead, forming a feature vector; (3) Construct a physically constrained LightGBM model through training The training objective is to output the volumetric flow rate coefficient k_v and the mass flow rate coefficient k_m; The feature data and target values ​​k_v1, k_v2, k_m1, and k_m2 are converted to integers, and the training and testing sets are divided. The LightGBM regression algorithm is used for training to obtain the prediction model. The obtained model is converted into C language code and embedded into the microcontroller program in the form of a function. (4) Embedded reasoning Inference is performed using the LightGBM model deployed in the MCU, outputting real-time k_v1, k_v2, k_m1, and k_m2; Qv and Qm are calculated by combining the predicted k_v1, k_v2, k_m1, and k_m2 with the real-time measured ΔP1 and ΔP2; and the gas volume Q_g, liquid volume Q_l, and gas-liquid ratio GLR are calculated based on Qv, Qm, and gas-liquid density. The dual differential pressure signals ΔP1 and ΔP2 in step (1) are two response characteristics with different flow states provided by the Venturi tube structure section and the V-cone structure section connected in series in the same pipeline. In step (1), the bidirectional propagation time difference Δt_h and Δt_v of the dual ultrasonic waves come from two sets of sensors installed in a through-beam configuration, one set being a horizontal path and the other a vertical path. The high-value characteristics of step (2) include the following aspects: A. From differential pressure: instantaneous values, moving average, variance, peak-to-trough values, and the ratio of ΔP1 / ΔP2 of ΔP1 and ΔP2; B. From ultrasound: Horizontal downstream time difference Δt_h1; Horizontal upstream time difference Δt_h2; Vertical downstream time difference Δt_v1; Vertical upstream time difference Δt_v2; Bidirectional time difference sum: Δt_h1+Δt_h2, Δt_v1+Δt_v2; Bidirectional time difference: Δt_h2-Δt_h1, Δt_v2-Δt_v1; Signal attenuation intensity α: Transmission is at a fixed intensity, only the reception intensity needs to be detected; C. Derived from temperature and pressure: temperature T and pressure P; D. Combinatorial characteristics: including ΔP1 / (Δt_h1+Δt_h2)²; The sources of the volumetric flow rate coefficient k_v and mass flow rate coefficient k_m in step (3) are: based on , , make: , ， In the above formula: It is volumetric flow rate. It refers to mass flow rate, where k is a coefficient related to flow pattern, Reynolds number, pipe diameter, and throttling device factors. For the differential pressure before and after the throttling device, The density of the medium; The calculation formulas for Qv and Qm in step (4) are as follows: ； ； ； ； ； 。 2. The moisture flow measurement method based on multi-sensor feature fusion and a physically constrained LightGBM model according to claim 1, characterized in that, The training data in step (3) is less than 10,000 records, the number of trees is 600, the number of leaves is 12, the depth is 8, and the learning rate is 0.

05.

3. A moisture flow measurement device, characterized in that, The device is compatible with the moisture flow measurement method based on multi-sensor feature fusion and physical constraint LightGBM model as described in claim 1 or 2, and includes a dual differential pressure mechanism and an ultrasonic mechanism connected in series in the same pipeline.

4. The moisture flow measuring device according to claim 3, characterized in that, The dual differential pressure mechanism includes a Venturi structure section and a V-cone structure section connected in series in the same pipeline. The Venturi structure section is provided with a first differential pressure tap H and a second differential pressure tap L, and the V-cone structure section is provided with a second differential pressure tap H and a second differential pressure tap L. The ultrasonic mechanism includes ultrasonic transducers A, B, C, and D, a temperature sensor, and a pressure sensor. Ultrasonic transducers A and B are vertically installed facing each other, while ultrasonic transducers C and D are horizontally installed facing each other.

5. A moisture flow measuring device according to claim 4, characterized in that, The ultrasonic transducer A and ultrasonic transducer B are both 30° apart from the axis.

Images