Control valve data processing method based on edge computing

Abstract

This invention belongs to the field of data processing, specifically relating to a control valve data processing method based on edge computing. This method constructs a sensing structure by micro-nano-structuring the surface of the control valve, causing it to generate mechanical vibration waves. The original vibration signals are acquired using a piezoelectric thin-film sensor array and converted into a frequency domain spectrum through edge computing nodes. The spectrum is then input into a physical information neural network model with embedded fluid dynamics and solid mechanics constraints to invert the fluid's density, viscosity, and multiphase flow ratio parameters. Finally, the inverted parameters are used as feedforward compensation quantities in the control algorithm to achieve closed-loop adaptive adjustment of the control valve opening. This invention achieves non-invasive monitoring through deep fusion of structure and sensing, eliminating the risk of leakage. The use of a physically guided edge computing model ensures the accuracy and robustness of parameter inversion and achieves precise optimization and single-point autonomy of fluid control with millisecond-level response.

Description

Technical Field

[0001] This invention belongs to the field of data processing, specifically relating to a control valve data processing method based on edge computing. Background Technology

[0002] With the continuous evolution of industrial automation and intelligent manufacturing technologies, fluid control systems play a core role in high-value industrial scenarios such as fine chemicals, pharmaceutical manufacturing, and energy and power. As key actuators for fluid transmission and process regulation, the precise control of control valves affects the safety and stability of the production process and the quality of the finished product. In high-precision fluid management, real-time acquisition of key physical properties such as fluid composition, density, and viscosity has become an important prerequisite for achieving process optimization, energy efficiency improvement, and fault early warning.

[0003] Edge computing-based control valve data processing technology is gradually becoming a research hotspot for improving system perception and response speed. This technology integrates edge computing nodes at the control valve end to achieve real-time monitoring of fluid states and adaptive adjustment of actuators. By processing and extracting features from raw data collected by sensors locally, the system can significantly reduce data transmission latency and attempts to establish a deep correlation between fluid physical states and control commands at the actuator level, thereby improving control accuracy under complex operating conditions.

[0004] Existing technologies largely rely on additional specialized physical property monitoring equipment such as spectrometers, densitometers, or flow meters installed on pipeline systems. This results in high hardware redundancy and increases potential fluid leakage points, compromising the integrity and safety of the pipeline system. Control valve stroke data and fluid property monitoring data are typically collected and processed separately by heterogeneous systems, making it difficult to form an efficient execution sensing closed loop at the edge. This hinders millisecond-level real-time responses to complex nonlinear conditions such as fluid phase changes or uneven mixing. Traditional data processing methods lack in-depth analytical capabilities regarding the fluid-structure interaction vibration characteristics of the valve body itself, making it difficult to achieve accurate inversion of multidimensional physical property parameters non-invasively. This leads to technical challenges that need to be addressed in dynamic process environments, including a single sensing dimension, delayed control feedback, and insufficient system robustness. Summary of the Invention

[0005] The purpose of this invention is to provide a control valve data processing method based on edge computing, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a control valve data processing method based on edge computing, comprising the following specific steps: Step 1: Construct a control valve sensing structure. In the control valve data processing method based on edge computing, the valve core or valve seat surface of the control valve is first processed with micro-nano structures to construct a resonant unit array with specific frequency response characteristics, so that it generates mechanical vibration waves related to fluid properties when fluid passes through. Step 2: Deploy a vibration acquisition module and integrate a piezoelectric thin film sensor array on the external valve body surface of the control valve to acquire in real time the broadband raw vibration time-domain signal generated by the valve core or valve seat under fluid impact; Step 3: Perform edge data conversion. The original broadband vibration time-domain signal is preprocessed by the edge computing node integrated at the valve body end. The time-domain signal is converted into a frequency domain spectrum reflecting the frequency distribution characteristics using the fast Fourier transform algorithm. Step 4: Perform physical property parameter inversion. Input the frequency domain spectrum into the physical information neural network model that has been pre-built and deployed in the edge computing node. Utilize the fluid dynamics constraints and solid mechanics constraints embedded in the physical information neural network model to invert the current density, viscosity, and multiphase flow ratio parameters of the fluid. Step 5: Execute closed-loop adaptive control, and introduce the fluid property parameters obtained by inversion into the proportional-integral-derivative control algorithm as feedforward compensation quantities to correct the opening command of the control valve in real time, so as to achieve precise adjustment of the fluid control process.

[0007] Preferably, the process of micro-nano-structuring the valve core or valve seat surface in step 1 includes: constructing a repetitive periodic structure on the surface of the valve core or valve seat using laser etching or precision machining, including periodic micro-grooves or protrusions. The geometric dimensions of the periodic structure are set according to the sound velocity range of the fluid to be measured, so that the structure can generate a local resonance effect when the fluid flows through it. The local resonance effect converts the microscopic kinetic energy fluctuations of the fluid into macroscopically measurable mechanical vibration energy, and the shift in the resonant frequency has a definite mapping relationship with the mass density and shear viscosity of the fluid.

[0008] Preferably, in step 2, the piezoelectric thin-film sensor array is arranged as follows: multiple piezoelectric sensitive units are respectively arranged at the inlet and outlet ends of the valve body, the valve cover, and the connection point of the actuator. The piezoelectric sensitive units are made of flexible piezoelectric material, which can fit tightly against the irregularly shaped surface of the valve body. The piezoelectric thin-film sensor array possesses high-sensitivity dynamic response characteristics, and its response bandwidth covers the entire frequency band from low-frequency industrial noise to high-frequency fluid-structure interaction vibration.

[0009] Preferably, the logic for edge data conversion in step 3 is as follows: the edge computing node first performs analog-to-digital conversion on the acquired current or voltage signal, and then performs high-pass filtering to remove low-frequency mechanical noise generated by the actuator movement. During the Fast Fourier Transform, the edge computing node uses a sliding window mechanism to segment the continuous vibration sequence. After windowing, the power spectral density of each segment is calculated. The power spectral density extracts the characteristic peak position, characteristic peak width, and quality factor of the resonance peak in the spectrum.

[0010] Preferably, the structure and training process of the physical information neural network model in step 4 include: the model uses a multilayer perceptron network as its main body, its input layer receives a vector containing frequency feature points, the hidden layer extracts features through a nonlinear activation function, and the output layer outputs fluid density, viscosity, and multiphase flow ratio. The core of the physical information neural network lies in the construction of its loss function. In addition to including the mean squared error term between the predicted value and the experimental sample value, the loss function also forcibly includes physical equation constraint terms.

[0011] Preferably, the physical equation constraints specifically include constraints from the Navel Stokes equations and the elasticity equilibrium equations. During model inference, the physical information neural network optimizes the network weights by calculating the momentum and mass conservation residuals in the flow field. This means that even when sensor signals are interfered with by the industrial environment, the derived fluid parameters still follow the fundamental laws of fluid mechanics due to the strong constraints of the physical equations.

[0012] Preferably, in step 4, when inverting the fluid density, the model determines it by identifying the frequency shift of the resonant peaks in the spectrum. The frequency shift is proportional to the equivalent added mass of the resonant unit due to the fluid mass load. When inverting the fluid viscosity, the model calculates based on the increase in bandwidth of the resonant peaks in the spectrum or the decrease in the quality factor, where the decrease in the quality factor reflects the viscous damping dissipation effect of the fluid on mechanical vibrations.

[0013] Preferably, in step 4, when inverting the multiphase flow ratio, the physical information neural network identifies whether there are multiple independent characteristic peaks or distortions in the spectral envelope in the spectrum. When bubbles or solid particles are present in the fluid, the resonant unit will produce a scattering effect, causing a change in the energy distribution of the spectrum in a specific frequency band. The model calculates the air tightness or solid content by analyzing this nonlinear characteristic.

[0014] Preferably, the implementation details of the closed-loop adaptive control in step 5 are as follows: the edge computing node dynamically adjusts the proportional coefficient, integral time constant, and derivative time constant in the proportional-integral-derivative (PID) control algorithm based on the inverted real-time density and viscosity. When an increase in fluid density or viscosity causes system control lag, the algorithm automatically increases the proportional coefficient to accelerate the response speed. Simultaneously, the rate of change of physical property parameters is used as a feedforward element, directly acting on the actuator's control end to counteract the disturbance of fluid property fluctuations on flow stability.

[0015] Preferably, the edge computing-based control valve data processing method further includes a sensor health monitoring step: the edge computing node compares the consistency of each sensing unit in the piezoelectric thin film sensor array in real time. If the output signal of a certain sensing unit deviates from a preset range or has a significant logical conflict with other units, the unit is determined to be faulty, and the system automatically switches to a backup sensing unit. Simultaneously, the system monitors the wear of the valve core through the intrinsic noise level in the spectrum diagram. When the intrinsic frequency undergoes irreversible drift, a maintenance warning is sent to the monitoring system.

[0016] Preferably, the edge computing node employs an industrial-grade embedded processing chip, which integrates a neural network acceleration unit. This acceleration unit performs point-based and pruning processing on the weights of the physical information neural network, enabling complex physical model inference to be completed within milliseconds. The edge computing node interacts with the host computer system via an industrial fieldbus or wireless communication network, uploading physical property inversion results and equipment operating status, while simultaneously receiving process setting parameters from the host computer.

[0017] Preferably, before performing the property parameter inversion in step 4, the method further includes an online calibration step. This online calibration step uses vibration data generated when a standard fluid with known properties passes through the control valve to fine-tune the parameters of the physical information neural network model. By comparing the measured spectrum with a preset standard spectrum, the system automatically compensates for measurement deviations caused by temperature fluctuations or valve aging.

[0018] Preferably, in step 5, when correcting the control valve opening command, the pressure drop change caused by the fluid flowing through the valve is also considered. The edge computing node combines the downstream pressure sensor data and the inverted density parameters to calculate the real-time flow coefficient. When the real-time flow coefficient deviates from the preset trajectory, the algorithm compensates for the fluctuation in flow capacity caused by changes in fluid properties by adjusting the stroke of the motor or pneumatic positioner.

[0019] Preferably, the edge computing-based control valve data processing method is applied to fluid environments with corrosive or high-purity requirements. Because the piezoelectric thin-film sensor array is deployed on the outer surface of the valve body, and the micro / nano structure is integrally formed with the valve core, the entire sensing process does not involve any invasive components that penetrate the pipe wall. This non-invasive structure ensures the smoothness of the fluid channel, prevents material residue and cross-contamination, and also eliminates the risk of leakage at traditional sensor seals.

[0020] Preferably, for multiphase flow conditions, the physical information neural network model analyzes the spectrum by frequency band by identifying the acoustic impedance differences under different phase states. In gas-liquid two-phase flow, the model decouples the random vibration components in the high-frequency band from the wave components in the low-frequency band, extracting the continuous characteristics of the liquid phase and the discrete characteristics of the gas phase respectively, thereby achieving accurate measurement of the mixing ratio.

[0021] Preferably, the edge computing node has data caching and network interruption resumption functions. During external network interruptions, the node can independently complete the entire process logic from signal acquisition to closed-loop control, and store the processed structured data in local storage. After communication is restored, the system automatically synchronizes historical data to the cloud server, ensuring the integrity of industrial production data.

[0022] Preferably, the edge computing-based control valve data processing method further includes a self-learning evolution mechanism. The edge computing nodes periodically upload the typical operating condition spectrum diagrams and inversion results they have processed to the central server. The central server uses massive amounts of data to retrain and optimize the physical information neural network, and then distributes the updated model parameters to each edge computing node, enabling the system to continuously improve its adaptability to unknown and complex operating conditions.

[0023] Preferably, the valve core surface after micro / nano-structure processing is coated with an ultra-hard anti-corrosion coating. This ultra-hard anti-corrosion coating protects the microgroove structure from fluid erosion and wear, while its physical thickness and mass density are incorporated into the solid mechanics constraint model in step 4. Precise modeling of the coating parameters ensures that the model's analysis of vibration signals maintains a high degree of physical realism throughout long-term operation.

[0024] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves deep integration of structure and sensing through micro-nano structure design of the control valve core or seat, eliminating reliance on independent physical property sensors. This integrated design reduces system hardware costs and installation complexity, solves the leakage risks and fluid resistance problems caused by traditional invasive sensors, and improves the structural integrity of the pipeline system.

[0025] 2. This invention utilizes edge computing combined with a physical information neural network model to achieve non-intrusive real-time inversion of fluid density, viscosity, and multiphase flow ratio. Compared to a purely data-driven model, the physical information neural network, through embedded physical equation constraints, ensures that the inversion results conform to physical laws under various extreme conditions, thereby improving the accuracy and robustness of the sensing.

[0026] 3. This invention constructs a millisecond-level control closed loop from fluid property inversion to actuator. By directly introducing the inverted fluid property parameters into the control algorithm, the system can make instantaneous responses to minute fluctuations in the fluid state, realizing adaptive dynamic optimization of the control process. This solves the response lag problem caused by the separation of sensing and execution data in traditional methods, and improves the process stability and product quality consistency in high-value fluid control scenarios.

[0027] 4. This invention employs edge computing nodes for on-site data processing, reducing the computational load and network bandwidth pressure on the central server. By performing signal preprocessing, spectrum conversion, and physical modeling at the edge, the system possesses strong real-time processing capabilities and single-point autonomy, ensuring the continuous and reliable operation of the fluid control system in harsh industrial environments.

[0028] 5. The method of this invention has broad industrial applicability and self-learning capabilities. Its non-invasive sensing method is particularly suitable for fields with extremely high requirements for hygiene and safety, such as pharmaceuticals and fine chemicals. Simultaneously, through continuous online calibration and model evolution, the system can continuously adapt to valve aging and operational condition changes, resulting in a long technical lifecycle and low maintenance costs. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram of the core principle framework for fluid property parameter inversion based on a physical information neural network model according to the present invention; Figure 3 This is a flowchart illustrating the logical flow of edge data transformation and frequency domain feature extraction according to the present invention. Figure 4 This is a flowchart of the closed-loop adaptive control based on feedforward compensation of physical property parameters according to the present invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow according to the present invention. Detailed Implementation

[0030] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0031] In the edge computing-based control valve data processing method, step 1 involves precise physical modification of the core mechanical components of the control valve. The process of constructing the control valve's sensing structure first determines the valve core or seat surface as the carrier for sensing excitation. When performing micro / nano structure processing on the valve core or seat surface, an ultrafast pulsed laser etching process with wavelengths in the range of 248 nm to 1064 nm is used, or a precision micromachining method with an accuracy within 1 micrometer is employed to construct highly repeatable periodic structures on the surface of the valve core or seat, including: a periodic array of microgrooves or a periodic array of protrusions.

[0032] In step 1 above, the geometric dimensions of the periodic structure, including the depth and width of the grooves and the center distance between two adjacent grooves, are precisely set according to the sound velocity range of the fluid under preset operating conditions. The setting principle is to couple the characteristic dimensions of the micro / nano structure with the wavelength of mechanical waves in the fluid medium, so that the structure can generate a local resonance effect when the fluid flows through it and generates surface friction and impact. This local resonance effect, through fluid-structure interaction at the microscopic level, converts the microscopic kinetic energy fluctuations and turbulent pressure pulsations of the fluid into macroscopically measurable mechanical vibration energy.

[0033] The design of the resonant unit array follows the physical laws of metamaterials, and its equivalent mass and equivalent stiffness are modulated in real time by the fluid properties. When the fluid density changes, the additional mass block effect generated by the fluid adhering to the surface of the micro / nanostructure causes a change in the total inertia of the resonant system, resulting in a shift in the resonant frequency. Similarly, when the fluid viscosity changes, the shear damping effect of the fluid inside the micro / nanostructure alters the quality factor of the resonant system. The amount of shift in the resonant frequency exhibits a definite positive or negative correlation with the fluid's mass density, while the broadening of the resonant peak exhibits a definite correlation with the fluid's shear viscosity.

[0034] Specifically, the quantitative mapping relationship between the resonant frequency shift and the fluid mass density satisfies:

[0035] in, This is the resonant frequency shift, which is the difference between the no-load resonant frequency and the load resonant frequency. This is the unloaded resonant frequency of the micro / nano structure in the absence of fluid. It is determined by the inherent properties of the valve core substrate and coating. The load resonant frequency of the micro / nano structure when it is in fluid mode; The equivalent stiffness of the micro / nano structure; The equivalent mass of the valve core substrate and the anti-corrosion coating; The additional mass coefficient is related to the geometry and surface roughness of the micro / nano structure, and its value ranges from 0.2 to 0.8. This is the shear viscosity of the fluid.

[0036] The quantitative mapping relationship between the full width at half maximum (FWHM) of the resonance peak and the fluid shear viscosity satisfies:

[0037] in, It is the full width at half maximum (FWHM) of the resonance peak when there is fluid, that is, the frequency width when the energy of the resonance peak drops to half of the peak value; The unloaded full width at half maximum (FWHM) of the micro / nano structure in the absence of fluid is determined by the structure's own damping. It is the viscous damping coupling coefficient, which is related to the surface area of ​​the micro / nano structure and the thickness of the fluid boundary layer, and its value ranges from 1e-4 to 1e-2. This is the shear viscosity of the fluid.

[0038] Following the micro / nano structure processing in step 1, an ultra-hard anti-corrosion coating, such as a diamond-like carbon coating or a tungsten carbide coating, is applied to the valve core surface to ensure long-term stability. This ultra-hard anti-corrosion coating protects the microgroove structure from fluid erosion and wear, while its physical thickness and mass density are incorporated as constants into the subsequent mechanical constraint model. Precise modeling of the coating parameters ensures that the system's analysis of vibration signals maintains a high degree of physical realism throughout long-term operation.

[0039] For the sensing structure constructed above, step 2 involves deploying the vibration acquisition module. A piezoelectric thin-film sensor array is integrated onto the outer surface of the control valve body. Its core function is to acquire internal mechanical waves using a non-invasive method. The piezoelectric thin-film sensor array consists of multiple independent piezoelectric sensitive units, which are made of flexible polyvinylidene fluoride material with a high piezoelectric constant, allowing them to fit tightly against the outer surface of the valve body, which has an irregular radius of curvature.

[0040] In terms of specific arrangement, the piezoelectric sensing units are respectively installed at the inlet flange, outlet flange, top of the valve cover, and the connection between the actuator and the valve stem of the valve body. This spatially distributed multi-point acquisition mode can form a comprehensive vibration observation grid. The piezoelectric thin-film sensor array has extremely high dynamic response characteristics, and its frequency response bandwidth covers the range from low-frequency industrial environmental noise of 10 Hz to high-frequency fluid-structure interaction resonant band of 500 kHz. During the acquisition process, the sensor array converts mechanical strain into weak electrical signals in real time.

[0041] In step 2 above, to eliminate electromagnetic interference from the external environment, each piezoelectric sensing unit is equipped with an independent shielded wire, which is connected to a signal conditioning circuit located near the valve body. The signal conditioning circuit performs charge amplification and anti-aliasing filtering on the original weak signal. The broadband original vibration time-domain signal is transmitted to an edge computing node integrated at the valve body end.

[0042] Step 3 performs edge data conversion, which begins with the digital reconstruction of the signal. The edge computing node has a built-in high-sampling-rate analog-to-digital converter that converts the analog voltage signal into a digital signal stream with a bit depth of no less than 16 bits. After digitization, the edge computing node first executes a digital high-pass filtering algorithm to remove signal components below the cutoff frequency, thereby filtering out low-frequency mechanical noise generated by valve actuator movements, pipeline support vibrations, and external mechanical impacts.

[0043] In one specific embodiment, the specific logic of the edge computing node performing the sliding window fast Fourier transform is as follows: the edge computing node denotes the continuous broadband original vibration time-domain signal sequence as... Set the frame length of the sliding window to The number of overlapping sampling points between two adjacent frames is (usually taken) To achieve 50% overlap). For the first... Frame signal Applying the Hanning window function To suppress spectral leakage, the windowed signal is Subsequently, the edge computing nodes perform a Fast Fourier Transform on the frame signal and calculate the one-sided power spectral density:

[0044] in, Indicates the first Frame signal at frequency The power spectral density value at that location; This indicates the sampling frequency of the analog-to-digital converter at the edge computing node; The normalized energy factor of the window function is expressed by the following formula: ; Represents the imaginary unit; This represents a discrete-time index.

[0045] Furthermore, the edge computing nodes employ a peak search algorithm to extract multi-dimensional feature vectors. Specifically, the node searches for the global maximum point in the power spectral density curve. The frequency corresponding to this point is the center frequency of the main resonance peak. ;exist Searching for power spectral density decreases on both sides Two frequency points and Calculate the full width at half maximum (FWHM) Finally, calculate the quality factor. This feature vector As input tensors for subsequent physical information neural network models.

[0046] During frequency domain transformation, the edge computing node executes the Fast Fourier Transform (FFT) algorithm. To achieve near real-time feature extraction, the algorithm employs a sliding window mechanism, dividing the continuous vibration sequence into multiple frames with overlapping portions. Each frame sequence undergoes Hanning or Hamming windowing before transformation to suppress spectral leakage. The edge computing node calculates the power spectral density of each frame signal.

[0047] In the obtained frequency domain spectrum, the edge computing nodes automatically extract key feature parameters through a peak search algorithm. These parameters include: first, extracting the center frequency position of the main resonant peak with the most concentrated energy in the spectrum; second, extracting the frequency width corresponding to the point where the energy of the resonant peak drops to half of the peak value, i.e., the full width at half maximum (FWHM); and third, calculating the ratio of the center frequency of the main resonant peak to the FWHM, i.e., the quality factor Q. These feature parameters constitute a multidimensional feature vector reflecting the fluid properties, providing data support for subsequent inference and inversion.

[0048] Step 4, which involves the inversion of physical property parameters, is the core logic of the entire data processing method. The feature vector is then input into a pre-constructed physical information neural network model deployed on edge computing nodes. This model is not a purely data-driven black box; the weight updates of its multilayer perceptron network structure are constrained by physical laws.

[0049] Specifically, to address the mismatch between frequency domain feature inputs and physical field coordinates, the Physical Information Neural Network (PINN) model internally constructs an equivalent fluid-structure interaction parameterized network. This network receives frequency domain feature vectors. The equivalent density of the fluid is mapped using a multilayer perceptron. Equivalent viscosity and multiphase flow ratio Meanwhile, a virtual local equivalent flow field module based on spectral characteristics was constructed in parallel within the network, thereby enabling the application of physical equation constraints.

[0050] The composite loss function of the physical information neural network model It is composed of a weighted combination of data-driven loss terms and physical equation constraint loss terms:

[0051] in, This represents the total loss value that needs to be minimized during model training or inference. and This represents the preset hyperparameter weighting coefficients used to balance various losses.

[0052] The data-driven loss term Using mean square error form:

[0053] in, This indicates the number of samples calculated in the current batch (during online inference). ); These represent the predicted density, predicted viscosity, and predicted multiphase flow ratio, respectively, output by the neural network during the current forward propagation. This represents the corresponding true label value (in the online unsupervised inversion phase, this value degenerates to zero or is replaced by a soft label provided by the online calibration phase).

[0054] The first term in the constraint loss term of the physical equation is the solid mechanics equivalent constraint term. Its equivalent mechanical impedance model is based on micro / nano resonant units:

[0055] in, , indicating the use of predicted density The theoretical resonant frequency is calculated using the resonant equivalent mass model. The equivalent stiffness of micro / nano structures The equivalent mass of the valve core matrix. This refers to the additional mass coefficient related to the geometric parameters of the micro / nano structure. The actual center frequency extracted from the preceding steps; This indicates the use of predicted viscosity. The calculated theoretical quality factor, The structural damping coefficient is... The viscous damping coupling coefficient; This is the actual quality factor extracted in the preceding steps. This term enforces the requirement that the physical mapping from frequency domain characteristics to material property parameters must follow the laws of mechanical vibration.

[0056] The second term in the constraint loss term of the physical equation is a fluid dynamics conservation constraint term. Edge computing nodes utilize automatic differentiation techniques to construct virtual velocity tensors within the network. (Depend on Measured pressure difference before and after the valve (Generated using a Bernoulli equation proxy model), partial derivatives are taken with respect to virtual space coordinates to construct momentum conservation residuals and mass conservation residuals:

[0057] in, This represents the number of collocation points sampled in the virtual equivalent flow field; Represents the residuals of the mass conservation equation (continuity equation); This represents the momentum conservation residuals of the Navel-Stokes equations after neglecting body forces; Indicates the proportion of multiphase flow. Corrected effective dynamic viscosity; This represents the pressure field at the collocation point. The residual term is backpropagated through the automatic differential chain rule, forcing the neural network output to... and It is self-consistent in the sense of fluid mechanics.

[0058] The physical information neural network model consists of an input layer, multiple hidden layers, and an output layer. The input layer receives a vector containing the aforementioned frequency feature points. The hidden layers map the input features into a high-dimensional space using a nonlinear activation function. The output layer is responsible for outputting the current density value, viscosity value, and multiphase flow ratio parameters of the fluid.

[0059] During model inference, the loss function of the physical information neural network possesses unique physical guidance characteristics. In addition to including the mean squared error term between the predicted and experimental sample values, the loss function also forcibly incorporates constraints from the Navier-Stokes equations in fluid mechanics and the equilibrium equations of elasticity in solid mechanics. During computation, edge computing nodes utilize automatic differentiation techniques to calculate the partial derivatives of the output parameters with respect to the input coordinates, constructing momentum conservation residuals and mass conservation residuals.

[0060] When fluid flows through the micro / nanostructure at the valve core, its flow state must obey the law of conservation of mass. The physical information neural network verifies whether the current solution satisfies physical constraints by substituting the real-time inverted density and viscosity into the embedded fluid dynamics equations. If the inversion result deviates from physical laws, even if the error in the data fitting term is small, the total loss function will still output a large value, prompting the network weights to adjust in a direction consistent with physical logic. This mechanism enables the system to provide reasonable predictions of physical properties based on physical laws, even when sensor signals are subjected to high-intensity industrial noise interference or partial failure.

[0061] When inverting fluid density, the model focuses on identifying the resonant peak frequency shift in the spectrum caused by the added mass effect. As the fluid density increases, the equivalent fluid mass around the micro / nano structure increases, causing the system's intrinsic frequencies to shift towards lower frequencies. The model accurately calculates the magnitude of this shift and, combined with preset solid structure stiffness parameters, inverts the absolute density of the fluid.

[0062] When inverting fluid viscosity, the model calculates based on the increase in bandwidth of the resonance peak in the spectrum. Higher fluid viscosity results in stronger damping and dissipation of vibrations in micro / nano structures, manifested as a flatter resonance curve and a decrease in the quality factor. The model derives the dynamic viscosity parameters of the fluid by identifying the decay gradient of the quality factor.

[0063] For multiphase flow conditions, the physical information neural network model performs in-depth spectrum analysis by identifying acoustic impedance differences in different phase states. In gas-liquid two-phase flow, due to the significant difference in acoustic impedance between gas and liquid, the vibration signals generated by micro / nano structures exhibit a bimodal distribution or specific spectral envelope distortion. The model decouples the high-frequency random vibration components from the low-frequency wave components, extracting the continuous characteristics of the liquid phase and the discrete characteristics of the gas phase, respectively. By analyzing the energy distribution weights in different frequency bands, the air tightness or solids content in the mixed fluid is calculated.

[0064] Before performing the property parameter inversion in step 4 above, the method also includes a crucial online calibration step. Edge computing nodes use reference vibration data generated when a standard fluid with known properties passes through the control valve to incrementally fine-tune the weights of the neural network. The system automatically calculates compensation coefficients by comparing the measured spectrum with the preset standard spectrum to eliminate measurement deviations caused by drastic fluctuations in ambient temperature, long-term fatigue aging of the valve body material, or slight scaling on the surface of the micro / nano structure.

[0065] Step 5 executes closed-loop adaptive control, transforming the sensing results into actionable actions. The edge computing node dynamically adjusts the internal control parameters of the proportional-integral-derivative (PID) control algorithm based on the inverted real-time density and viscosity parameters. Specifically, when the inverted fluid viscosity increases, leading to increased flow resistance in the pipe and consequently system response lag, the algorithm automatically increases the proportional coefficient to strengthen the initial adjustment and simultaneously decreases the integral time constant to accelerate the elimination of residual error.

[0066] In a preferred embodiment, the specific mathematical expression of introducing the inverted fluid property parameters as feedforward compensation quantities into the proportional-integral-derivative control algorithm in step 5 is as follows: Edge computing nodes construct the basic output of time-series-based incremental PID control laws. :

[0067] in, Indicates the first The increment of the control quantity output by the PID algorithm within one control cycle; Indicates the first The flow deviation setting value for each cycle; These represent the proportional coefficient, integral coefficient, and differential coefficient, respectively.

[0068] The dynamic adjustment The logic is implemented through two-dimensional bilinear interpolation. Edge computing nodes pre-store density-based... x-axis represents viscosity The grid matrix of PID parameters is shown on the y-axis. , When the currently inverted physical property parameters are obtained... At that time, find its four neighboring nodes in the grid, so as to For example, the interpolation formula is:

[0069] in, To enclose in the grid The values ​​of adjacent density nodes; To surround The adjacent viscosity node values; , This formula allows for real-time calculation of the fluid properties adapted to the current fluid state. , , .

[0070] Simultaneously, a feedforward compensation term for the rate of change of physical properties is constructed. To counteract the disturbance of flow stability caused by fluctuations in fluid properties:

[0071] in, Indicates the feedforward compensation control quantity; and The feedforward gain coefficients (pre-determined through system identification) represent the density change rate and viscosity change rate, respectively. and This represents the differential rate of change of density and viscosity obtained by the edge computing node through two adjacent inversion cycles.

[0072] Finally, the control valve opening command is executed after real-time correction. for:

[0073] in, This represents the control output baseline value of the previous cycle. This instruction is directly sent to the actuator, completing the full data loop from spectrum sensing -> property inversion -> physical constraint verification -> feedforward and feedback fusion control.

[0074] Simultaneously, the method uses the rate of change of physical property parameters as a feedforward compensation amount, directly applied to the control signal terminal of the actuator. This fusion of feedforward and feedback control allows the control valve to detect changes in fluid properties in advance and pre-adjust the opening before flow fluctuations occur. Furthermore, the edge computing node combines data from the pressure sensor installed downstream of the valve with the inverted density parameters to calculate the valve's actual flow capacity coefficient in real time. When the calculated real-time flow coefficient deviates from the preset trajectory, the algorithm compensates for flow capacity fluctuations caused by changes in fluid properties by precisely adjusting the stroke of the motor or pneumatic positioner, achieving precise adaptive adjustment of the fluid control process.

[0075] The edge computing-based control valve data processing method also integrates sensor health monitoring functionality. The edge computing node periodically compares the consistency of the output signals of each sensitive unit in the piezoelectric thin-film sensor array. If the signal amplitude of a sensitive unit is detected to be consistently below a preset threshold, or if its spectral characteristics logically conflict with adjacent units, the edge computing node determines that the unit has experienced a hardware failure and immediately blocks the data of that channel at the algorithm level, switching to a preset backup sensor unit signal source.

[0076] In addition, the system utilizes the noise level of a specific high-frequency band in the spectrum, which reflects the intrinsic characteristics of the valve core, to perform long-term monitoring of the wear status of the valve core. When it is detected that the intrinsic frequency of the valve core structure has irreversibly drifted towards a higher frequency, and the amount of drift exceeds the preset safety boundary, the edge computing node determines that the micro-nano structure may have suffered severe erosion wear, and automatically sends a maintenance warning signal to the monitoring system.

[0077] The edge computing node employs a high-performance industrial-grade embedded processing chip, which integrates an acceleration unit specifically optimized for neural network operations. To adapt to the limited computing resources at the edge, the acceleration unit performs fixed-point processing and neuron pruning optimization on the weight parameters of the physical information neural network.

[0078] The edge computing nodes possess a robust data interaction mechanism. Under normal communication conditions, the nodes maintain synchronization with the host computer system via industrial fieldbus or high-speed wireless communication networks, uploading material property inversion results, valve opening status, and equipment health indices in real time. Simultaneously, the nodes also feature data caching and network outage resumption capabilities. During unexpected external network interruptions, the edge computing nodes independently complete the entire process logic from signal acquisition to closed-loop control using local computing power, storing each processed structured data in a local non-volatile memory with cyclic overwrite functionality. Once communication is restored, the system automatically synchronizes historical data packets to the cloud, ensuring the temporal integrity of production data.

[0079] The method in this embodiment also features a self-learning evolution mechanism. Edge computing nodes periodically select representative spectrum diagrams of complex operating conditions and upload them, along with the corresponding inversion results, to the central server. The central server utilizes the collected operating data from multiple valves to perform large-scale retraining and global optimization of the physical information neural network. The optimized model parameters are periodically distributed to each edge node, enabling the entire control system to continuously improve its accuracy in identifying unknown and extreme operating conditions and its control stability as operating time increases.

[0080] Example 2: Building upon Example 1, this example further refines the specific implementation details in an extremely corrosive fluid environment. In such an environment, any invasive sensor faces an extremely high risk of damage. This example employs a non-invasive sensing architecture, with the piezoelectric thin-film sensor array entirely deployed on the outer surface of the valve body, utilizing an integrated micro / nano structure and valve core design for sensing.

[0081] In step 1, special geometric optimization was performed on the micro / nano structure of the valve core surface for highly corrosive fluids. The periodic microgrooves were designed as trapezoidal cross-sections with self-cleaning properties. The large end of the trapezoidal cross-section faces the direction of fluid flow. This structure can utilize the micro-vortex effect when the fluid passes over it to automatically remove tiny particles that may be deposited at the bottom of the grooves, preventing the micro / nano structure from losing its resonant characteristics due to fouling.

[0082] In step 2 above, to improve the reliability of the sensor under high temperature and high corrosion conditions, the piezoelectric thin film sensor array uses high-temperature resistant polyimide as the substrate. The sensor unit is coupled to the valve body surface via a high-conductivity, high-temperature resistant epoxy resin, ensuring that internal vibration waves are transmitted to the sensor array with minimal attenuation. When processing signals, the edge computing node adds a temperature-compensated gain adjustment stage. The edge computing node reads data from the thermistor integrated on the valve body, searches a preset sensitivity-temperature correction table in real time, and dynamically compensates for the acquired vibration amplitude to eliminate the influence of temperature fluctuations on the piezoelectric constant.

[0083] In the physical property parameter inversion step 4, an energy conservation constraint term is introduced into the physical information neural network model to address the frequent nonlinear and drastic changes in fluid density under highly corrosive environments. The loss function, based on the original momentum and mass conservation, is further constrained by the first law of thermodynamics. By calculating the enthalpy change of the fluid before and after passing through the valve core, the inverted density and viscosity parameters are verified.

[0084] Specifically, in step 5, the adaptive control addresses the frictional fluctuation problem of the actuator in highly corrosive environments by adding an online frictional compensation stage to the proportional-integral-derivative (PID) control algorithm. Edge computing nodes analyze specific harmonic components in the vibration spectrum that reflect dry friction characteristics to calculate the equivalent resistance of the current valve stem movement. When outputting the opening correction command, the algorithm automatically increases the amplitude of the starting pulse to overcome static friction, ensuring that the valve can still achieve micron-level precise positioning even in high-viscosity, highly corrosive media.

[0085] Example 3: This example provides a specific application implementation scheme in a multiphase flow mixing process in fine chemical industry. In this application scenario, it is necessary to accurately decouple the physical properties of the gas, liquid, and solid three-phase mixed fluid.

[0086] In step 1, the arrangement of the micro / nano structures is designed as a multi-scale composite array. Specifically, the valve core surface simultaneously contains resonant units of both micrometer and submicrometer scales. The micrometer-scale structures primarily generate resonant responses to collisions of solid particles in the fluid, while the submicrometer-scale structures are more sensitive to the viscous damping of fluid molecules. This multi-scale design allows the original vibration signal to contain rich decoupling information of physical properties.

[0087] In the edge data conversion process of step 3 above, the edge computing nodes use wavelet packet decomposition algorithm instead of traditional fast Fourier transform to obtain higher time-frequency resolution. Wavelet packet decomposition decomposes the original signal into multiple independent frequency band subspaces. The edge computing nodes calculate the energy entropy value of each subspace. When the content of solid particles in the fluid increases, the energy entropy of the corresponding high-frequency subspace will increase; while when the proportion of gas phase increases, the characteristic peaks of the mid-frequency subspace will exhibit unique intermittent fluctuations.

[0088] In step 4, the physical information neural network model employs a multi-task learning architecture. The network is divided into a common feature extraction layer and three independent inversion branches, responsible for outputting density, viscosity, and solids content, respectively. In the loss function, a mixed fluid equivalent acoustic model constraint is introduced to address the physical characteristics of multiphase flow. The model verifies the logical consistency of the inversion results by calculating the sound velocity propagation characteristics under different phase distributions. For example, when the inverted gas phase proportion increases, the model checks whether the corresponding decrease in equivalent density conforms to a preset mixing criterion.

[0089] In step 5 above, to address the valve cavitation erosion problem easily caused by multiphase flow, the edge computing node monitors the cavitation feature vector in the spectrum in real time while performing adaptive control. Once the inversion results show that the fluid is in a metastable state prone to cavitation, the algorithm immediately intervenes in the actuator by adjusting the valve opening or sending a signal to the upstream pumping system to actively change the pressure difference before and after the valve, allowing the fluid to avoid the cavitation region. This ensures process control accuracy while extending the service life of the control valve.

[0090] Example 4: This example focuses on the application of this method in high-purity fluid control in biopharmaceuticals. In this scenario, the system achieves the highest requirements for the smoothness and absence of dead zones in the fluid channels. Because this invention uses non-invasive sensing, it does not require drilling holes in the pipeline to install sensors, thus naturally possessing a high level of hygiene.

[0091] In step 1, the micro-nano structure on the valve core surface is treated with laser polishing to ensure that the roughness of the groove edges reaches the nanometer level. This treatment method preserves the resonant characteristics and also prevents protein molecules or bacteria in the fluid from being trapped inside the microstructure, thus meeting the requirements for online cleaning and online sterilization.

[0092] In step 2 above, the piezoelectric thin-film sensor array employs a fully sealed structural design. The entire array is encased in an extremely thin medical-grade stainless steel housing, which is fixed to the outer wall of the valve body using a laser spot welding process. The edge computing nodes are installed in explosion-proof and waterproof independent housings and connected to the sensors via a sanitary dedicated interface.

[0093] In step 4, considering the non-Newtonian fluid characteristics typically found in biopharmaceutical fluids, the physical information neural network model incorporates specific fluid constitutive equations for constraint. The model no longer assumes constant fluid viscosity, but instead treats it as a function of shear rate. During the inversion process, edge computing nodes calculate the real-time shear rate based on the current valve opening and the inverted flow velocity data, and adjust the branch weights for viscosity inversion accordingly.

[0094] In step 5, the closed-loop adaptive control algorithm places particular emphasis on minimizing flow fluctuations. To prevent excessive shear forces from being generated during mixing and thus damaging bioactivity, an acceleration limiting mechanism is incorporated when correcting the opening command. Edge computing nodes calculate the impact of opening changes on the fluid shear field in real time, ensuring that all adjustments are performed smoothly within a preset safe shear range.

[0095] Meanwhile, the system utilizes an online calibration process to perform benchmark calibration on each batch of materials. When changing production batches, the system automatically guides the user into calibration mode, using the vibration characteristics of the standard buffer solution to reset the initial parameters of the physical model, ensuring consistency in sensing accuracy between different batches of products.

[0096] Example 5: This example describes the distributed application of this method in the context of a large-scale industrial internet.

[0097] In this implementation, hundreds or thousands of control valves integrating edge computing nodes are distributed throughout the chemical plant area. Each edge computing node is not only responsible for local property inversion and control, but also acts as a node in a distributed sensor network.

[0098] In steps 3 and 4, the edge computing nodes execute an algorithm called collaborative inversion. When the confidence level of a single valve's sensor signal decreases due to local electromagnetic interference, the node requests data from neighboring valve nodes via the industrial wireless network. The neighboring nodes send their inverted fluid property trends as prior information. The local edge computing node incorporates this prior information as a soft constraint into the loss function of the physical information neural network, improving the perception robustness of individual nodes through distributed information sharing.

[0099] Specifically, distributed cooperative inversion is achieved by adding a cooperative loss term to the original composite loss function. The modified total loss function is as follows: Among them, collaborative loss item The expression is:

[0100] in, This is the loss term for collaborative inversion, used to incorporate prior information from adjacent nodes; The collaborative loss weighting coefficient ranges from 0.1 to 0.5 and is adjusted based on network communication quality and the similarity of working conditions between nodes. To connect with the current edge computing nodes The set of adjacent nodes; Adjacent nodes The confidence weights, ranging from 0 to 1, are determined by the nodes. The sensor's health status, historical retrieval accuracy, and communication latency all contribute to its determination. For the current node The local inversion result vector; Adjacent nodes The inversion result vector at the same time; It is an L2 norm.

[0101] In the control logic of step 5 above, the edge computing nodes implement a global optimization strategy. The nodes not only consider the current valve's flow stability but also the real-time flow status of upstream and downstream valves. When a significant disturbance in the upstream fluid density is detected, the edge computing modules of each downstream node calculate their respective optimal compensation time based on the propagation speed of the disturbance wave. By coordinating and adjusting the valve opening, the impact of fluid fluctuations on the entire production line is minimized.

[0102] Furthermore, the self-learning evolution mechanism in this embodiment is manifested as swarm learning. Edge computing nodes distributed throughout the plant periodically aggregate the processed, anonymized feature vectors to the plant-level edge server. The edge server uses transfer learning algorithms to quickly distribute neural network parameters optimized for a specific complex operating condition (such as a rare chemical reaction fluctuation) to all control valve nodes under similar conditions. This swarm evolution capability enables the fluid control system of the entire plant to rapidly adapt to adjustments in the production process.

[0103] At the maintenance level, edge computing nodes perform joint analysis using inverted physical property parameters and the energy consumption of actuators. If the system detects that the inverted viscosity is normal, but the actuator's drive current has increased significantly, it determines that mechanical jamming has occurred inside the actuator, rather than a change in fluid properties. This multi-dimensional cross-validation greatly improves the accuracy of predictive maintenance and avoids unnecessary downtime for inspections.

[0104] Example 6: This example refines the internal computational logic of the physical information neural network model.

[0105] The model receives the feature vector processed in step 3, which contains information such as the principal resonant frequency, second-order resonant frequency, quality factor, and skewness of the spectral envelope. In the model's input layer, these features are initially linearly combined using a set of learnable weight matrices. Subsequently, the model enters a deep mapping network consisting of four hidden layers. Each hidden layer contains 64 neurons, using the hyperbolic tangent function as the activation function to ensure the continuity and differentiability of partial differential calculations in the physical constraints.

[0106] In the inversion calculation, the construction of the loss function demonstrates extremely high rigor. The first part of the loss function is a supervision term based on labeled data, which calculates the Euclidean distance between the predicted density, viscosity, and multiphase flow ratio and the true values ​​in a pre-stored standard operating condition library. The second part of the loss function is the physical residual term. For density inversion, the algorithm extracts the functional relationship between the resonant frequency displacement and the equivalent mass, and calculates the additional mass tensor of the fluid on the surface of the micro / nano structure based on the potential flow theory of fluid dynamics. If the predicted density value does not match the frequency displacement corresponding to this mass tensor, the physical residual term will increase.

[0107] For viscosity inversion, the algorithm utilizes an energy dissipation model of fluid-structure interaction vibration. The loss function calculates the viscous work dissipated by the laminar boundary layer due to the predicted viscosity and matches it with the experimentally extracted spectral bandwidth increment. The physical information neural network ensures that the final output fluid parameters strictly adhere to physical laws while satisfying the data distribution by minimizing this total loss function, which incorporates multiple physical definitions. This computational process runs entirely within dedicated hardware acceleration modules on edge computing nodes, simplifying complex simulation calculations that originally required high-performance workstations into millisecond-level real-time inference through highly parallel vector operation units.

[0108] In the proportional-integral-derivative (PID) control algorithm correction in step 5, the adjustment of the proportional coefficient is not a simple linear scaling. The edge computing nodes employ a parameter mapping table based on a property space. This table is generated through offline large-scale fluid dynamics simulations and stores the optimal controller parameters for different combinations of density and viscosity. After obtaining the property parameters through real-time inversion, the edge computing nodes use a bilinear interpolation algorithm to find the current recommended control coefficients in the parameter mapping table and fine-tune them based on the rate of change of the current deviation.

[0109] This online controller parameter tuning method based on physical property inversion solves the problem of control quality degradation caused by fluid condition migration in traditional control. Even under drastic transient conditions where the fluid changes from a pure liquid phase to a multiphase flow containing gas, the valve control process can still maintain extremely high stability because the edge computing nodes can identify the appearance of gas phase characteristics in the spectrum in advance and recalculate the system gain according to the physical constraints of the multiphase flow.

[0110] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A control valve data processing method based on edge computing, characterized in that, Includes the following steps: Step 1: Construct a control valve sensing structure. In the edge computing-based control valve data processing method, the valve core or valve seat surface of the control valve is processed with micro-nano structures to construct a resonant unit array with frequency response characteristics. The resonant unit array generates mechanical vibration waves related to the fluid properties when the fluid passes through it. Step 2: Deploy a vibration acquisition module and integrate a piezoelectric thin film sensor array on the external valve body surface of the control valve to acquire the broadband original vibration time-domain signal generated by the valve core or valve seat under fluid impact; Step 3: Perform edge data conversion. The broadband original vibration time-domain signal is preprocessed by the edge computing node integrated at the valve body end. The broadband original vibration time-domain signal is converted into a frequency domain spectrum reflecting the frequency distribution characteristics by using the fast Fourier transform algorithm. Step 4: Perform physical property parameter inversion. Input the frequency domain spectrum into the physical information neural network model deployed in the edge computing node. Utilize the fluid dynamics constraints and solid mechanics constraints embedded in the physical information neural network model to invert the current density, viscosity, and multiphase flow ratio parameters of the fluid. Step 5: Execute closed-loop adaptive control, and introduce the fluid property parameters obtained by inversion as feedforward compensation into the proportional-integral-derivative control algorithm to correct the opening command of the control valve in real time.

2. The control valve data processing method based on edge computing according to claim 1, characterized in that: The process of performing micro-nano structure processing on the surface of the valve core or valve seat in step 1 includes constructing repetitive periodic structures on the surface of the valve core or valve seat using pulsed laser etching or precision machining, including: periodic micro-groove arrays or protrusion arrays. The geometric dimensions of the periodic structure are set according to the sound velocity range of the fluid under the preset working conditions, so that the periodic structure generates a local resonance effect when the fluid flows through and generates surface friction. The local resonance effect transforms the microscopic kinetic energy fluctuations and turbulent pressure pulsations of the fluid into macroscopically measurable mechanical vibration energy through fluid-structure interaction. The equivalent mass and equivalent stiffness of the resonant unit array are modulated in real time by the fluid properties, wherein the shift of the resonant frequency has a definite mapping relationship with the mass density of the fluid, and the broadening of the resonant peak has a definite mapping relationship with the shear viscosity of the fluid.

3. The control valve data processing method based on edge computing according to claim 2, characterized in that: In step 2, the piezoelectric thin film sensor array is arranged such that multiple piezoelectric sensitive units are arranged at the inlet end, outlet end, valve cover, and actuator connection of the valve body. The piezoelectric sensing unit is made of flexible piezoelectric material and is tightly attached to the irregularly shaped valve body surface; The piezoelectric thin film sensor array has dynamic response characteristics, and its response bandwidth covers the frequency range from low-frequency industrial environmental noise to high-frequency fluid-structure interaction vibration. During the acquisition process, the piezoelectric sensing unit converts mechanical strain into an electrical signal in real time, and then performs charge amplification and anti-aliasing filtering through the signal conditioning circuit to generate the broadband original vibration time-domain signal.

4. The control valve data processing method based on edge computing according to claim 3, characterized in that: The logic of edge data conversion in step 3 is as follows: the edge computing node first performs analog-to-digital conversion on the acquired raw signal, and then executes a digital high-pass filtering algorithm to filter out low-frequency mechanical noise generated by the movement of the actuator and the external environment. During frequency domain conversion, the edge computing node employs a sliding window mechanism to segment and window the continuous vibration sequence in order to suppress spectral leakage. The edge computing node calculates the power spectral density of each frame of signal and automatically extracts multi-dimensional feature vectors from the spectrum using a peak search algorithm. The multidimensional feature vector includes the center frequency position of the main resonance peak, the full width at half maximum (FWHM) of the resonance peak, and the quality factor of the resonance peak.

5. The control valve data processing method based on edge computing according to claim 4, characterized in that: The structure and training process of the physical information neural network model in step 4 include that the physical information neural network model consists of an input layer, multiple hidden layers and an output layer. The input layer receives a vector containing frequency feature points, the hidden layer performs feature extraction and high-dimensional space mapping through a nonlinear activation function, and the output layer outputs fluid density, viscosity, and multiphase flow ratio. The weight update of the physical information neural network model is constrained by a weighted composite loss function; The composite loss function includes a mean squared error term between the predicted value and the experimental sample value, as well as a mandatory physical equation constraint term. During model training and online calibration, the edge computing nodes use automatic differentiation technology to calculate the partial derivatives of the output parameters with respect to the input coordinates, construct momentum conservation residuals and mass conservation residuals, and perform physical-guided optimization adjustments on the network weights to ensure that the inversion results still follow the laws of fluid dynamics even when the sensor signals are disturbed.

6. The control valve data processing method based on edge computing according to claim 5, characterized in that: The physical equation constraints specifically include constraints from the Navier-Stokes equations and constraints from the elasticity equilibrium equations. When inverting fluid density, the physical information neural network model determines it by identifying the frequency shift of the resonant peak in the spectrum, and the frequency shift is proportional to the equivalent additional mass of the resonant unit due to the fluid mass load. When inverting fluid viscosity, the physical information neural network model calculates based on the increase in bandwidth of the resonance peak in the spectrum or the decrease in quality factor. The decrease in quality factor reflects the viscous damping dissipation effect of the fluid on mechanical vibration. The valve core surface after micro-nano structure processing is also coated with an ultra-hard anti-corrosion coating. The physical thickness and mass density of the ultra-hard anti-corrosion coating are incorporated into the solid mechanics constraint model as known constants to compensate for the influence of the coating on vibration signal analysis.

7. The control valve data processing method based on edge computing according to claim 6, characterized in that: In step 4, when inverting the multiphase flow ratio, the physical information neural network model identifies whether there are multiple independent characteristic peaks or distortions in the spectral envelope in the spectrum. For gas-liquid two-phase flow or solid-liquid two-phase flow conditions, the physical information neural network model analyzes the spectrum by frequency band by identifying the acoustic impedance differences under different phase states. The edge computing node decouples the random vibration component of the high-frequency band from the wave component of the low-frequency band, extracts the continuous characteristics of the liquid phase and the discrete characteristics of the non-liquid phase respectively, and calculates the air tightness or solid content in the fluid by analyzing the energy distribution weight in different frequency bands. Before performing the inversion of physical property parameters, the method also includes an online calibration step, which uses the reference data generated when a standard fluid with known physical properties passes through the control valve to incrementally fine-tune the weights of the physical information neural network model in order to automatically compensate for measurement deviations caused by environmental temperature fluctuations, material aging, or surface fouling.

8. The control valve data processing method based on edge computing according to claim 7, characterized in that: The implementation details of the closed-loop adaptive control in step 5 are as follows: the edge computing node dynamically adjusts the proportional coefficient, integral time constant and derivative time constant in the proportional-integral-derivative control algorithm according to the real-time density and viscosity obtained by inversion through bilinear interpolation lookup table method. When the fluid viscosity increases, causing the system response to lag, the algorithm automatically increases the proportional coefficient and decreases the integral time constant to strengthen the initial adjustment and accelerate the elimination of residual error. The rate of change of physical property parameters is converted into a feedforward compensation quantity as a feedforward link, which is directly applied to the control end of the actuator to counteract the disturbance of fluid property fluctuations on flow stability. The edge computing node also calculates the real-time flow coefficient by combining the valve downstream pressure data with the inverted density parameters. When the real-time flow coefficient deviates from the preset trajectory, the flow capacity fluctuation caused by the change in fluid properties is compensated by adjusting the stroke of the actuator.

9. The control valve data processing method based on edge computing according to claim 8, characterized in that: The method also includes a sensor health monitoring step; The edge computing node compares the consistency of each sensing unit in the piezoelectric thin film sensor array in real time. If the output signal of a certain sensing unit deviates from the preset range or has a significant logical conflict with other units, the unit is determined to be faulty and automatically switched to the backup sensing unit. The system monitors the wear of the valve core by analyzing the intrinsic noise level in the spectrum. When the intrinsic frequency undergoes irreversible drift and exceeds the safety boundary, it determines that the micro-nano structure is subjected to erosion wear and sends a maintenance warning to the monitoring system. The edge computing node uses a processor chip with an integrated neural network acceleration unit. The neural network acceleration unit performs point-based and pruning processing on the weights of the physical information neural network, so that the physical model inference is completed within a preset real-time processing cycle.

10. The control valve data processing method based on edge computing according to claim 9, characterized in that: The edge computing node has data caching and network interruption resume functions. During the external network interruption, it independently executes signal acquisition, physical property inversion and closed-loop control logic, and stores structured data in local memory. After communication is restored, it synchronizes historical data to the cloud server. The control valve data processing method based on edge computing also includes a self-learning evolution mechanism. The edge computing nodes periodically upload the processed operating condition spectrum diagram and inversion results to the central server. The central server uses massive data to retrain and globally optimize the physical information neural network model, and then distributes the updated model parameters to each edge computing node. In distributed application scenarios, multiple edge computing nodes perform a collaborative inversion algorithm through a wireless network, incorporating the inversion trends of neighboring nodes as prior information into the local loss function, thereby improving the perceptual robustness of individual nodes through information sharing.

Images