Flow rate measurement method and device based on AI image recognition

Abstract

The application provides a flow velocity measurement method and device based on AI image recognition; including: collecting fluid images and preprocessing to obtain feature point motion information, acquiring fluid images of different wavelengths through a multispectral image acquisition device, removing high light areas and jitter noise after image registration; using an improved semantic segmentation network to identify the bubble profile in the image and extract the bubble dynamics parameters; combining infrared temperature measurement data with an image gray scale distortion model to reconstruct the temperature field distribution and calculate the refractive index gradient parameters; performing spectral analysis on the continuous image sequence to extract the turbulent characteristic frequency and pressure fluctuation parameters; sequentially passing the initial flow velocity through a bubble influence correction model, a temperature gradient coupling correction model and a turbulent pressure fluctuation correction model for iterative correction, and performing secondary optimization on the correction results through a deep belief network to output the final flow velocity value; the application can comprehensively quantify external interference and realize progressive correction.

Description

Technical Field

[0001] This invention relates to the field of flow velocity measurement methods, and more specifically to a flow velocity measurement method and apparatus based on AI image recognition. Background Technology

[0002] In the field of flow velocity measurement based on AI image recognition, existing technologies typically design correction mechanisms for single interfering factors, making it difficult to address measurement deviations caused by the synergistic effects of multiple factors in complex fluid environments. Specifically, fluids in industrial pipelines, reactors, and other similar scenarios often simultaneously exhibit microbubbles, temperature gradients, and turbulent disturbances: microbubbles scatter light, causing blurring of image feature point edges and inaccurate displacement measurements by AI; nonlinear changes in fluid refractive index caused by temperature gradients result in geometric distortion of the image, further amplifying feature point localization errors; and local pressure fluctuations caused by turbulence distort the flow velocity vector field, causing the initial flow velocity calculated by traditional optical flow methods to deviate from the actual value. Existing solutions often use independent filtering or correction modules to handle these issues separately, failing to consider the dynamic changes in bubble distribution affected by temperature gradients and the interference of turbulence intensity on bubble trajectories, leading to cumulative errors in the correction process. In complex scenarios with multiple coupled factors, the deviation of measurement results often exceeds the allowable range for industrial applications, making it difficult to meet the requirements of high-precision flow monitoring.

[0003] Based on the above problems, there is an urgent need for a technical solution that can comprehensively quantify external interference and achieve progressive correction. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and proposes a flow velocity measurement method based on AI image recognition, comprising:

[0005] The fluid images are acquired and preprocessed to obtain feature point motion information. Fluid images of different wavelengths are acquired through a multispectral image acquisition device. After registration, the acquired images are de-highlighted and removed from jitter noise.

[0006] An improved semantic segmentation network is used to identify bubble contours in images and extract bubble dynamic parameters;

[0007] The temperature field distribution was reconstructed by combining infrared thermometry data with an image grayscale distortion model, and the refractive index gradient parameters were calculated.

[0008] Spectral analysis of continuous image sequences is performed to extract turbulence characteristic frequencies and pressure fluctuation parameters;

[0009] The initial flow velocity is iteratively corrected sequentially using the bubble influence correction model, the temperature gradient coupling correction model, and the turbulent pressure fluctuation correction model. The correction results are then optimized a second time using a deep belief network to output the final flow velocity value.

[0010] Preferably, the multispectral image acquisition device includes three industrial cameras with different wavelengths, the wavelengths of which are selected from the visible light and near-infrared bands, respectively. The three industrial cameras achieve synchronous image acquisition through a synchronous triggering device, which is electrically connected to the image acquisition control module. The triggering signal of the synchronous triggering device is transmitted to the control terminal of the three industrial cameras.

[0011] Further preferred, the improved semantic segmentation network is a U-Net network with a newly added bubble morphological feature extraction layer. The bubble morphological feature extraction layer includes an edge detection operator and a morphological operation unit. The bubble contour information output by the edge detection operator is transmitted to the morphological operation unit. The morphological operation unit outputs the equivalent diameter and motion direction parameters of the bubble. The equivalent diameter and motion direction parameters of the bubble are transmitted to the bubble influence correction model.

[0012] In a further preferred embodiment, the spectrum analysis employs an algorithm combining Fast Fourier Transform (FFT) and wavelet packet decomposition (WPD). The power spectral density data output by the FFT module is transmitted to the WPD module. The WPD module obtains the turbulence characteristic frequencies through multi-level decomposition. The turbulence characteristic frequencies and pressure fluctuation parameters are correlated and calculated through the data fusion module. The output parameters of the data fusion module are transmitted to the turbulence pressure fluctuation correction model.

[0013] Furthermore, the initial flow velocity is calculated using the following formula:

[0014] ;

[0015] in The initial flow velocity, For image scaling coefficients, The pixel displacement of feature points identified by AI. The sampling time interval, The light influence coefficient is... represents the standard deviation of the image grayscale.

[0016] A further preferred calculation formula for the bubble influence correction model is:

[0017] ;

[0018] in, The flow rate after bubble correction. The correction factor for the effect of bubbles, Bubble density, For fluid density, The equivalent diameter of the bubble. The average particle size of the characteristic points. It is the angle between the direction of bubble movement and the mainstream direction.

[0019] A further preferred calculation formula for the temperature gradient coupling correction model is:

[0020] ;

[0021] in The flow rate after temperature gradient correction. For temperature sensitivity coefficient, T represents the temperature gradient, and d represents the distance from the measurement area to the lens. The refractive index at the current temperature, The refractive index at the reference temperature of 20°C is the temperature coefficient of refractive index.

[0022] A flow velocity measurement device based on AI image recognition, applied to the flow velocity measurement method based on AI image recognition as described in any one of the above, comprising:

[0023] The system comprises an image acquisition unit, a preprocessing unit, and a flow rate calculation unit. The output of the image acquisition unit is connected to the input of the preprocessing unit, and the output of the preprocessing unit is connected to the input of the flow rate calculation unit. The system is characterized by further including a multispectral control unit, a bubble analysis unit, a temperature field unit, a turbulence analysis unit, and an AI correction unit. The multispectral control unit is electrically connected to the image acquisition unit. The outputs of the bubble analysis unit, temperature field unit, and turbulence analysis unit are respectively connected to the input of the AI ​​correction unit. The output of the AI ​​correction unit is connected to the input of the flow rate calculation unit, and the feedback output of the flow rate calculation unit is connected to the feedback input of the AI ​​correction unit.

[0024] In a further preferred embodiment, the temperature field unit includes a distributed infrared temperature sensor and a distortion correction module. The signal output terminal of the distributed infrared temperature sensor is connected to the signal input terminal of the distortion correction module. The distortion correction module has a built-in refractive index calculation model. The output parameters of the refractive index calculation model are transmitted to the AI ​​correction unit, and the sampling signal of the distributed infrared temperature sensor is transmitted to the temperature field reconstruction module.

[0025] In a further preferred embodiment, the AI ​​correction unit includes a model storage module and a deep belief network computation module. The model storage module pre-stores a bubble influence correction model, a temperature gradient coupling correction model, and a turbulent pressure fluctuation correction model. The input layer of the deep belief network computation module receives the output parameters of the three correction models. The output layer of the deep belief network computation module is connected to the feedback input of the flow velocity calculation unit. The model storage module and the deep belief network computation module are connected through a data bus.

[0026] Technical effects:

[0027] This invention creatively achieves comprehensive quantitative correction of the combined interference of microbubbles, temperature gradients, and turbulent disturbances through multispectral image acquisition and preprocessing, improved semantic segmentation network extraction of bubble parameters, reconstruction of temperature field by combining infrared data, extraction of turbulence features by spectral analysis, and collaborative design of progressive correction model and deep belief network secondary optimization. It solves the problem of cumulative error caused by the linkage of multiple factors in the background technology, significantly improves the accuracy of flow velocity measurement in complex fluid environments, and meets the requirements of high-precision monitoring. Attached Figure Description

[0028] Figure 1 This is a flowchart of the flow velocity measurement method based on AI image recognition in this application;

[0029] Figure 2 This is a connection block diagram of the flow velocity measurement device based on AI image recognition in this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] Traditional technical solutions have the following technical problems: existing flow velocity measurement methods based on AI image recognition are difficult to accurately extract feature point motion information in complex fluid environments due to image distortion caused by bubble scattering, temperature gradients, and turbulent pressure fluctuations. The initial flow velocity calculation does not incorporate multi-factor correction, which affects the measurement accuracy.

[0032] Based on this, please refer to Figure 1 This embodiment provides a flow velocity measurement method based on AI image recognition, including:

[0033] S1: Acquire fluid images and perform preprocessing to obtain feature point motion information. Obtain fluid images of different wavelengths through a multispectral image acquisition device. After registering the acquired images, remove highlight areas and jitter noise.

[0034] S2: An improved semantic segmentation network is used to identify bubble contours in images and extract bubble dynamic parameters;

[0035] S3: Combine infrared thermometry data with image grayscale distortion model to reconstruct temperature field distribution and calculate refractive index gradient parameters;

[0036] S4: Perform spectral analysis on continuous image sequences to extract turbulence characteristic frequencies and pressure fluctuation parameters;

[0037] S5: The initial flow velocity is iteratively corrected sequentially through the bubble influence correction model, the temperature gradient coupling correction model, and the turbulent pressure fluctuation correction model. The correction results are then optimized a second time using a deep belief network to output the final flow velocity value.

[0038] This technical solution first utilizes a multispectral image acquisition device to obtain fluid images at different wavelengths. Preprocessing steps, including registration, highlight removal, and jitter noise reduction, lay a high-quality image foundation for subsequent feature extraction. Next, an improved semantic segmentation network is specifically designed for bubble contour recognition, extracting bubble dynamic parameters and mitigating the interference of bubble scattering on image analysis.

[0039] By combining infrared thermography data with a grayscale distortion model to reconstruct the temperature field and calculate the refractive index gradient, image distortion caused by temperature gradients is addressed. For the effects of turbulence, characteristic frequencies and pressure fluctuation parameters are extracted through spectral analysis, providing data support for turbulence correction.

[0040] The initial flow velocity is iteratively corrected by three progressive correction models: bubble, temperature gradient, and turbulence, and then further optimized by a deep belief network to form a precise measurement logic with multiple factors working together. The data from each stage are interconnected to ensure the consistency and accuracy of the correction.

[0041] The technical effects of this embodiment include: resolving measurement deviations caused by multi-factor interference and improving the stability and accuracy of flow velocity measurement in complex fluid environments.

[0042] Traditional technical solutions have the following technical problems: the selection of camera wavelengths in existing multispectral image acquisition devices is not targeted enough, and the synchronous acquisition mechanism is imperfect, resulting in poor consistency of image data and affecting the accuracy of subsequent multispectral data fusion analysis.

[0043] Based on this, the multispectral image acquisition device includes three industrial cameras with different wavelengths. The wavelengths of the three industrial cameras are selected from the visible light and near-infrared bands, respectively. The three industrial cameras achieve synchronous image acquisition through a synchronous triggering device. The synchronous triggering device is electrically connected to the image acquisition control module, and the trigger signal of the synchronous triggering device is transmitted to the control terminal of the three industrial cameras.

[0044] In this technical solution, the core of the multispectral image acquisition device lies in the configuration and synchronous acquisition mechanism of three industrial cameras with different wavelengths. The camera wavelengths are selected from the visible light and near-infrared bands, which can specifically capture the response characteristics of different substances in the fluid to different wavelengths of light, thereby enhancing the richness of image information.

[0045] The synchronization triggering device is crucial. Electrically connected to the image acquisition and control module, it receives commands from the module, generates synchronization trigger signals, and transmits them to the control terminals of the three industrial cameras. This ensures that the three cameras acquire images simultaneously, preventing image misalignment due to acquisition time differences and guaranteeing the spatial consistency of the multispectral images. This hardware configuration and signal transmission logic provide highly synchronized and comprehensive spectral data for subsequent image registration and feature extraction, enhancing the reliability of multispectral analysis.

[0046] The technical effects of this embodiment include: improving the synchronization and targeting of multispectral image acquisition, providing high-quality data for subsequent processing, and enhancing the stability of the measurement foundation.

[0047] Traditional technical solutions have the following technical problems: existing semantic segmentation networks have insufficient accuracy in recognizing bubble contours and lack a dedicated bubble morphology feature extraction mechanism, resulting in inaccurate extraction of bubble dynamic parameters and affecting the correction effect of bubbles on flow velocity measurement interference.

[0048] Based on this, the improved semantic segmentation network is a U-Net network with a newly added bubble morphology feature extraction layer. The bubble morphology feature extraction layer includes an edge detection operator and a morphology operation unit. The bubble contour information output by the edge detection operator is transmitted to the morphology operation unit. The morphology operation unit outputs the equivalent diameter and motion direction parameters of the bubble. The equivalent diameter and motion direction parameters of the bubble are transmitted to the bubble influence correction model.

[0049] This technical solution improves the U-Net network by adding a bubble morphology feature extraction layer, specifically designed for accurate bubble feature extraction. This extraction layer includes an edge detection operator and a morphological operation unit. The edge detection operator accurately identifies the bubble's contour boundary, resolving the problem of inaccurate contour extraction caused by blurred boundaries between the bubble and the background. After the extracted contour information is transmitted to the morphological operation unit, the contour is further optimized through morphological processing, and the equivalent diameter and motion direction parameters of the bubble are calculated.

[0050] These parameters are directly transmitted to the bubble influence correction model, providing the model with accurate input data and ensuring that the interference of bubbles on flow velocity measurement is effectively corrected. The improved network structure enhances the targeted identification capability of bubble features and improves the accuracy of parameter extraction.

[0051] The technical effects of this embodiment include: improving the accuracy of bubble contour recognition and parameter extraction, providing reliable data for bubble influence correction, and enhancing the effectiveness of the correction model.

[0052] Traditional technical solutions have the following technical problems: existing spectrum analysis methods are limited and it is difficult to accurately extract turbulence characteristic frequencies and pressure fluctuation parameters from continuous image sequences, resulting in the inability to effectively quantify and correct the interference of turbulence on flow velocity measurement.

[0053] Based on this, the spectrum analysis adopts an algorithm that combines Fast Fourier Transform (FFT) and wavelet packet decomposition (WPD). The power spectral density data output by the FFT module is transmitted to the WPD module. The WPD module obtains the turbulence characteristic frequency through multi-level decomposition. The turbulence characteristic frequency and pressure fluctuation parameters are correlated and calculated by the data fusion module. The output parameters of the data fusion module are transmitted to the turbulence pressure fluctuation correction model.

[0054] This technical solution employs a combined algorithm of Fast Fourier Transform (FFT) and wavelet packet decomposition (WPD) for spectral analysis. The FFT module converts the time-domain signal of a continuous image sequence into a frequency-domain signal, outputting power spectral density data to initially capture the frequency characteristics of turbulence. However, the FFT alone has limited analytical capabilities for non-stationary signals. Therefore, the power spectral density data is transmitted to the wavelet packet decomposition module, which performs multi-level decomposition to achieve fine segmentation of different frequency bands, accurately extracting turbulent characteristic frequencies and solving the frequency analysis challenge of non-stationary turbulent signals. The extracted turbulent characteristic frequencies and pressure fluctuation parameters are correlated and calculated by the data fusion module to establish their quantization relationship. The output parameters are transmitted to the turbulent pressure fluctuation correction model, providing a basis for accurate correction of turbulent interference. The combination of these two algorithms enhances the comprehensiveness and accuracy of the spectral analysis.

[0055] The technical effects of this embodiment include: improving the accuracy of turbulence feature extraction, providing reliable parameters for turbulence pressure fluctuation correction, and enhancing the correction effect on turbulence interference.

[0056] Traditional technical solutions have the following technical problems: the existing initial flow velocity calculation does not fully consider the impact of illumination uniformity on image feature point recognition, resulting in deviations in the initial flow velocity data and affecting the input accuracy of subsequent correction models.

[0057] Based on this, the initial flow velocity is calculated using the following formula:

[0058] ;

[0059] in The initial flow velocity, For image scaling coefficients, The pixel displacement of feature points identified by AI. The sampling time interval, The light influence coefficient is... represents the standard deviation of the image grayscale.

[0060] This formula is used to calculate the initial flow velocity based on the motion of image feature points. The core of this formula is to integrate image scale transformation and illumination interference correction to solve the problem of feature point recognition error caused by ignoring illumination uniformity in traditional flow velocity calculation.

[0061] left side of the formula This is the initial flow rate, which serves as the base value for subsequent adjustments. (Right side, first part) For the physical quantity conversion process: As an image scale coefficient, it realizes the conversion of pixel displacement to actual physical distance and establishes the correspondence between image coordinates and real flow field space; It is the pixel displacement of the feature point identified by AI, reflecting the change in pixel position of the same feature point at adjacent sampling times, and is directly related to the visual representation of fluid motion; The sampling time interval serves as a time reference to convert displacement into velocity dimensions, ensuring the integrity of the physical meaning.

[0062] Part Two The illumination interference correction term is expressed as an exponential function, reflecting the nonlinear characteristics of the illumination effect: It is the standard deviation of image gray levels, which reflects the uniformity of illumination by quantizing the dispersion of pixel gray values. The larger the value, the more uneven the illumination, such as the coexistence of highlight and shadow areas, making the edges of feature points prone to blurring, leading to... Identification bias; The illumination influence coefficient is used to adjust the correction intensity. Its value is determined based on experimental data under different illumination conditions to ensure that interference can be reasonably suppressed in extreme scenarios such as strong light and weak light.

[0063] The decreasing property of the exponential function makes the initial flow velocity The decay rate decreases appropriately as the degree of uneven illumination increases, thus compensating in advance for system errors that are difficult to eliminate in subsequent corrections.

[0064] This formula, through the transformation of physical quantities and the correction of nonlinear disturbances, ensures the dimensional rationality of the flow velocity calculation and achieves quantitative compensation for illumination factors.

[0065] The initial flow velocity calculation formula of this technical solution In this model, each parameter has a clear physical meaning and dimension. k, as the image scaling coefficient, converts pixel displacement into actual physical distance. It is the pixel displacement of the feature points identified by AI, reflecting the movement of the feature points on the image; The sampling time interval is used to convert displacement into a fundamental velocity quantity. The key innovation lies in the introduction of... Item, of which The standard deviation of image grayscale reflects the uniformity of illumination.

[0066] The larger the value, the more uneven the illumination, and the greater the potential error in feature point recognition; α is the illumination influence coefficient, used to quantify the degree of influence of illumination on flow velocity calculation. This formula, by incorporating illumination factors, improves the basic flow velocity... The corrections reduce the initial flow velocity deviation caused by uneven illumination, providing more accurate input for subsequent model corrections.

[0067] The technical effects of this embodiment include: reducing the interference of illumination factors on the initial flow velocity calculation, improving the accuracy of the initial flow velocity data, and laying a good foundation for subsequent corrections.

[0068] Traditional technical solutions suffer from the following problems: existing bubble influence correction methods do not comprehensively consider the synergistic effects of the bubble-to-fluid density ratio, bubble size, and direction of motion; the correction model is simplistic and struggles to accurately quantify the degree of bubble interference with flow velocity measurement. Therefore, the calculation formula for the bubble influence correction model is as follows:

[0069] ,

[0070] in, The flow rate after bubble correction. The correction factor for the effect of bubbles, Bubble density, For fluid density, The equivalent diameter of the bubble. The average particle size of the characteristic points. It is the angle between the direction of bubble movement and the mainstream direction.

[0071] This formula addresses the interference of bubbles in flow velocity measurement during gas-liquid two-phase flow by using multi-physical parameter synergistic quantification correction, thus solving the problem that traditional single-parameter correction cannot reflect the interconnected effects of bubble size, density, and direction of motion.

[0072] left side of the formula The flow rate after bubble correction is the initial flow rate. The first iteration of optimization based on the existing foundation. Right-hand side correction term. As the core innovation, various parameters characterize bubble interference from different dimensions: It is the density ratio of the bubble to the fluid, reflecting the buoyancy characteristics of the bubble in the fluid. The greater the density difference, the more significant the disturbance of the bubble's rising / sinking motion on the surrounding flow field, and the more obvious the deviation of the characteristic point's trajectory. The power function form of 0.3 weakens the excessive influence of extreme density ratios and avoids correction distortion. This demonstrates the relative relationship between bubble size and feature points. The equivalent diameter of the bubble. The average particle size of the feature points is represented by the squared term, which amplifies the influence of size differences. Since larger bubbles have a stronger scattering and blocking effect on light, they can lead to false displacements in feature point recognition. Quantifying the interference of bubble movement direction, Let be the angle between the direction of bubble motion and the mainstream direction, representing the lateral motion component. When approaching 90° The interference with the mainstream velocity measurement is most significant when the value approaches 1. To correct the impact of bubbles, the weights of each parameter were balanced through experiments to ensure that the correction strength matches different bubble concentration scenarios.

[0073] The correction term is in reciprocal form, making It decreases as bubble interference increases, achieving dynamic adaptation where the stronger the interference, the greater the correction magnitude.

[0074] The bubble influence correction model in this technical solution achieves accurate correction of the initial flow velocity v0 through the synergistic effect of multiple parameters. In the formula, The density ratio of the bubble to the fluid reflects the influence of the bubble's buoyancy in the fluid on its trajectory. The greater the density difference, the more significant the disturbance of the bubble to the surrounding fluid. The equivalent diameter of the bubble. θ represents the average particle size of the feature points, and the square ratio of the two reflects the relative size of the bubble to the feature points. Larger bubbles have a stronger occlusion or scattering effect on image feature point recognition. θ is the angle between the bubble's motion direction and the mainstream direction, and its sine value quantifies the interference of the bubble's lateral motion component on the mainstream velocity measurement. The larger the angle, the more significant the lateral interference. β serves as a bubble influence correction coefficient, used to balance the weights of various parameters, making the correction result more closely match the actual flow field characteristics. By integrating these parameters into a correction term and scaling the initial flow velocity v0, a progressive quantitative correction of bubble interference is achieved, overcoming the limitations of single-parameter correction.

[0075] The technical effects of this embodiment include: accurately quantifying the influence of the multi-dimensional characteristics of bubbles on the flow rate, and improving the accuracy of flow rate measurement of fluids containing bubbles.

[0076] Traditional technical solutions have the following technical problems: existing temperature gradient correction does not correlate the spatial distribution of refractive index with temperature sensitivity characteristics, and only uses linear correction, which cannot cope with image distortion caused by nonlinear refractive index changes, affecting the accuracy of flow velocity measurement.

[0077] Based on this, the calculation formula for the temperature gradient coupling correction model is as follows:

[0078] ;

[0079] in The flow rate is the temperature gradient corrected, and γ is the temperature sensitivity coefficient. T represents the temperature gradient, and d represents the distance from the measurement area to the lens. The refractive index at the current temperature, The refractive index at the reference temperature of 20°C is the temperature coefficient of refractive index.

[0080] The temperature gradient coupling correction model of this technical solution achieves secondary correction of the flow velocity v1 after bubble correction through multi-physical quantity coupling.

[0081] This formula addresses the image distortion problem caused by fluid refractive index changes due to temperature gradients. It overcomes the limitations of traditional linear correction in handling nonlinear refractive index distributions by using multi-physics parameter coupling correction.

[0082] left side of the formula The flow rate after temperature gradient correction is the flow rate corrected by bubbles. Secondary optimization based on the existing foundation. Right-hand correction term. Integrating the relationship between temperature field and optical properties: The temperature gradient directly reflects the rate of temperature change in the fluid along space. The larger the gradient, the more uneven the spatial distribution of the refractive index, and the more significant the deflection of the light propagation path. The distance from the lens to the measurement area reflects the cumulative effect of the temperature gradient; the greater the distance, the wider the temperature gradient area that light passes through, and the more severe the image distortion. The relative refractive index, It is the refractive index of the fluid at the current temperature. The refractive index is at the reference temperature, and the ratio quantifies the absolute change in refractive index due to temperature. The temperature coefficient of refractive index characterizes the sensitivity of the refractive index to temperature changes; the larger the value, the more sensitive the fluid's optical properties are to temperature. It is a temperature sensitivity coefficient used to adjust the overall correction intensity and match the temperature response characteristics of different fluids.

[0083] The correction term is in additive form, making It adjusts linearly as the temperature gradient disturbance increases, ensuring effective compensation for nonlinear refractive index changes.

[0084] In the formula, T represents the temperature gradient, which directly reflects the spatial rate of temperature change in the fluid. The larger the gradient, the more uneven the refractive index distribution. d represents the distance from the measurement area to the lens, reflecting the cumulative effect of the temperature gradient on the light propagation path—the farther the distance, the more significant the distortion. This represents the ratio of the refractive index at the current temperature to that at the reference temperature, quantifying the absolute change in refractive index caused by temperature. γ is the temperature coefficient of refractive index, characterizing the sensitivity of a material's refractive index to temperature changes; γ is the temperature sensitivity coefficient, used to adjust the overall correction intensity. These parameters form a product term, comprehensively characterizing the impact of temperature gradients on the displacement measurement of image feature points through refractive index changes, and then... After correction This enables precise compensation for nonlinear temperature distortion.

[0085] The technical effects of this embodiment include: combining temperature gradient and refractive index characteristics to effectively correct temperature-induced measurement deviations and enhance measurement stability under complex temperature fields.

[0086] Traditional technical solutions have the following technical problems: existing flow velocity measurement devices are scattered, multi-factor correction lacks an integrated processing unit, data interaction between modules is not smooth, and it is impossible to achieve coordinated correction of the effects of bubbles, temperature and turbulence.

[0087] Based on this, please refer to Figure 2 This embodiment provides a flow velocity measurement device based on AI image recognition, including an image acquisition unit, a preprocessing unit, and a flow velocity calculation unit. The output end of the image acquisition unit is connected to the input end of the preprocessing unit, and the output end of the preprocessing unit is connected to the input end of the flow velocity calculation unit. It also includes a multispectral control unit, a bubble analysis unit, a temperature field unit, a turbulence analysis unit, and an AI correction unit. The multispectral control unit is electrically connected to the image acquisition unit. The output ends of the bubble analysis unit, the temperature field unit, and the turbulence analysis unit are respectively connected to the input end of the AI ​​correction unit. The output end of the AI ​​correction unit is connected to the input end of the flow velocity calculation unit, and the feedback output end of the flow velocity calculation unit is connected to the feedback input end of the AI ​​correction unit.

[0088] The measurement device of this technical solution achieves multi-factor integrated correction through a modular collaborative architecture. The core lies in the interactive design of the newly added multispectral control unit, bubble analysis unit, temperature field unit, turbulence analysis unit, and AI correction unit: the multispectral control unit regulates the image acquisition unit to acquire multi-wavelength images, providing basic data for subsequent analysis; the bubble analysis unit processes images to extract bubble parameters; the temperature field unit combines infrared data to calculate the refractive index gradient; and the turbulence analysis unit obtains turbulence characteristics through spectral analysis. The output data of the three units are transmitted to the AI ​​correction unit to achieve the convergence of multi-source data.

[0089] As the core processing node, the AI ​​correction unit receives parameters from each unit, executes a progressive correction model, and transmits the correction results to the flow rate calculation unit. Simultaneously, feedback signals from the flow rate calculation unit are sent back to the AI ​​correction unit, forming a closed-loop optimization mechanism to ensure dynamic adjustment during the correction process. Each module achieves end-to-end integration from data acquisition and feature extraction to collaborative correction through clearly defined signal connections.

[0090] The technical effects of this embodiment include: enabling multi-module data collaboration and integrated correction, and improving the device's adaptability to complex fluid environments.

[0091] Traditional technical solutions have the following technical problems: existing temperature field monitoring uses single-point temperature measurement, which is not directly related to image distortion correction, and the refractive index calculation lacks the support of measured data, resulting in insufficient accuracy of temperature gradient correction.

[0092] Based on this, the temperature field unit includes a distributed infrared temperature sensor and a distortion correction module. The signal output terminal of the distributed infrared temperature sensor is connected to the signal input terminal of the distortion correction module. The distortion correction module has a built-in refractive index calculation model. The output parameters of the refractive index calculation model are transmitted to the AI ​​correction unit, and the sampling signal of the distributed infrared temperature sensor is transmitted to the temperature field reconstruction module.

[0093] This technical solution achieves precise temperature-refractive index correlation through the collaboration of hardware and software models in its temperature field unit. The distributed infrared temperature sensor can collect temperature data from multiple points on the fluid, overcoming the limitations of single-point temperature measurement. Its sampled signals are transmitted not only to the distortion correction module but also simultaneously to the temperature field reconstruction module to construct a complete spatial distribution of the temperature field. The distortion correction module incorporates a refractive index calculation model. This model uses the sensor's measured temperature as input and, combined with the fluid's refractive index temperature characteristics, calculates the refractive index n(T) and related parameters of the refractive index gradient at the current temperature.

[0094] These parameters are transmitted as output to the AI ​​correction unit, providing measured data support for the temperature gradient coupled correction model. This ensures data consistency across the entire chain, from temperature acquisition to refractive index calculation and flow velocity correction, solving the problem of disconnect between the theoretical model and the actual flow field in traditional temperature correction. The technical effects of this embodiment include: achieving distributed monitoring of the temperature field and accurate calculation of the refractive index, providing a reliable data foundation for temperature correction.

[0095] Traditional technical solutions have the following technical problems: existing correction models are stored and computed separately, resulting in low model calling efficiency, lack of deep networks for secondary optimization of correction results, and difficulty in eliminating the cumulative error of multiple models.

[0096] Based on this, the AI ​​correction unit includes a model storage module and a deep belief network computation module. The model storage module pre-stores bubble influence correction models, temperature gradient coupling correction models, and turbulent pressure fluctuation correction models. The input layer of the deep belief network computation module receives the output parameters of the three correction models. The output layer of the deep belief network computation module is connected to the feedback input of the flow velocity calculation unit. The model storage module and the deep belief network computation module are connected through a data bus.

[0097] The AI ​​correction unit in this technical solution achieves efficient calling and optimization of the correction model through hardware module collaboration. The model storage module is specifically used to pre-store three progressive correction models to ensure fast reading of the correction logic; the deep belief network operation module serves as the core of the operation, and its input layer receives the output parameters of the three models. Through multi-layer nonlinear mapping, these parameters are fused and optimized to eliminate the cumulative errors that may be generated by independent correction of each model.

[0098] The two modules are connected via a data bus, enabling high-speed transmission of model parameters and computational data to ensure the real-time nature of the correction process. The output of the deep belief network is transmitted to the feedback input of the flow velocity calculation unit, forming a dynamic adjustment mechanism for the correction results. This makes the final flow velocity value more closely match the actual flow field, solving the problem of insufficient optimization in traditional multi-model correction. The technical effects of this embodiment include: improving the efficiency of corrective model invocation, optimizing correction results through deep networks, and reducing the cumulative error of multiple factors.

[0099] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A flow velocity measurement method based on AI image recognition, characterized in that, include: The fluid images are acquired and preprocessed to obtain feature point motion information. Fluid images of different wavelengths are acquired through a multispectral image acquisition device. After registration, the acquired images are de-highlighted and removed from jitter noise. An improved semantic segmentation network is used to identify bubble contours in images and extract bubble dynamic parameters; The temperature field distribution was reconstructed by combining infrared thermometry data with an image grayscale distortion model, and the refractive index gradient parameters were calculated. Spectral analysis of continuous image sequences is performed to extract turbulence characteristic frequencies and pressure fluctuation parameters; The initial flow velocity is iteratively corrected sequentially using the bubble effect correction model, the temperature gradient coupling correction model, and the turbulent pressure fluctuation correction model. The correction results are then optimized a second time using a deep belief network to output the final flow velocity value. The calculation formula for the bubble effect correction model is as follows: ； in, The initial flow velocity, The flow rate after bubble correction. The correction factor for the effect of bubbles, Bubble density, For fluid density, The equivalent diameter of the bubble. The average particle size of the characteristic points. The angle between the direction of bubble movement and the mainstream direction; The calculation formula for the temperature gradient coupling correction model is as follows: ； in The flow rate after temperature gradient correction. For temperature sensitivity coefficient, T represents the temperature gradient, and d represents the distance from the measurement area to the lens. The refractive index at the current temperature, The refractive index at the reference temperature of 20°C is the temperature coefficient of refractive index.

2. The flow velocity measurement method based on AI image recognition according to claim 1, characterized in that, The multispectral image acquisition device includes three industrial cameras with different wavelengths, selected from the visible light and near-infrared bands. The three industrial cameras achieve synchronous image acquisition through a synchronous triggering device, which is electrically connected to the image acquisition control module. The triggering signal of the synchronous triggering device is transmitted to the control terminal of the three industrial cameras.

3. The flow velocity measurement method based on AI image recognition according to claim 1, characterized in that, The improved semantic segmentation network is a U-Net network with a newly added bubble morphology feature extraction layer. The bubble morphology feature extraction layer includes an edge detection operator and a morphological operation unit. The bubble contour information output by the edge detection operator is transmitted to the morphological operation unit. The morphological operation unit outputs the equivalent diameter and motion direction parameters of the bubble. The equivalent diameter and motion direction parameters of the bubble are transmitted to the bubble influence correction model.

4. The flow velocity measurement method based on AI image recognition according to claim 1, characterized in that, The spectrum analysis employs an algorithm combining Fast Fourier Transform (FFT) and wavelet packet decomposition (WPD). The power spectral density data output by the FFT module is transmitted to the WPD module, which obtains the turbulence characteristic frequencies through multi-level decomposition. The turbulence characteristic frequencies and pressure fluctuation parameters are correlated and calculated by the data fusion module, and the output parameters of the data fusion module are transmitted to the turbulence pressure fluctuation correction model.

5. The flow velocity measurement method based on AI image recognition according to claim 1, characterized in that, The initial flow velocity is calculated using the following formula: ； in The initial flow velocity, For image scaling coefficients, The pixel displacement of feature points identified by AI. The sampling time interval, The light influence coefficient is... represents the standard deviation of the image grayscale.

6. A flow velocity measurement device based on AI image recognition, applied to the flow velocity measurement method based on AI image recognition as described in any one of claims 1-5, comprising: The system comprises an image acquisition unit, a preprocessing unit, and a flow rate calculation unit. The output of the image acquisition unit is connected to the input of the preprocessing unit, and the output of the preprocessing unit is connected to the input of the flow rate calculation unit. The system is characterized by further including a multispectral control unit, a bubble analysis unit, a temperature field unit, a turbulence analysis unit, and an AI correction unit. The multispectral control unit is electrically connected to the image acquisition unit. The outputs of the bubble analysis unit, temperature field unit, and turbulence analysis unit are respectively connected to the input of the AI ​​correction unit. The output of the AI ​​correction unit is connected to the input of the flow rate calculation unit, and the feedback output of the flow rate calculation unit is connected to the feedback input of the AI ​​correction unit.

7. The flow velocity measurement device based on AI image recognition according to claim 6, characterized in that, The temperature field unit includes a distributed infrared temperature sensor and a distortion correction module. The signal output terminal of the distributed infrared temperature sensor is connected to the signal input terminal of the distortion correction module. The distortion correction module has a built-in refractive index calculation model. The output parameters of the refractive index calculation model are transmitted to the AI ​​correction unit. The sampling signal of the distributed infrared temperature sensor is transmitted to the temperature field reconstruction module.

8. The flow velocity measurement device based on AI image recognition according to claim 6, characterized in that, The AI ​​correction unit includes a model storage module and a deep belief network computation module. The model storage module pre-stores bubble influence correction models, temperature gradient coupling correction models, and turbulent pressure fluctuation correction models. The input layer of the deep belief network computation module receives the output parameters of the three correction models. The output layer of the deep belief network computation module is connected to the feedback input of the velocity calculation unit. The model storage module and the deep belief network computation module are connected through a data bus.

Images