Digital twinning and neural network three-cylinder pump pressure measurement method and system and medium

Abstract

The invention discloses a three-cylinder pump pressure measurement method and system based on a digital twinning and neural network and a medium, and the method comprises the steps: inputting the time sequence data of pump parameters when a three-cylinder pump works into a trained physical information neural network model, and obtaining the pressure field information of an outlet manifold cavity; the training process of the physical information neural network model is as follows: constructing an analog simulation model mapped with an entity three-cylinder pump by using a digital twinning technology so as to obtain time sequence data comprising pump parameters in normal and abnormal states, and constructing a training set after preprocessing; embedding a partial differential equation term in the physical information neural network model as a physical loss function of physical constraint, training the physical information neural network model by using the training set, and adjusting trainable parameters through reverse updating; the three-cylinder pump pressure measurement method and system and the medium conform to the constraint of physical knowledge, can reflect the change rule of field data in real time, and enhance the reliability and stability of pressure prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of triplex pump pressure measurement, and in particular to a triplex pump pressure measurement method, system, and medium based on digital twin and neural network. Background Art

[0002] As a key fluid machinery in the industrial field, the pressure field prediction of the three-cylinder pump is crucial for the efficient operation of the equipment and fault warning. Traditional methods mainly rely on physical modeling or data-driven technology: physical modeling solves the pressure distribution based on fluid mechanics equations, but requires precise knowledge of fluid parameters and boundary conditions, has high computational complexity and is difficult to adapt to complex working conditions such as multiphase flow; data-driven methods use machine learning to fit the relationship between pressure and working parameters, but rely on a large amount of high-quality data, and the model lacks physical interpretability, and the prediction results may deviate from the actual physical laws. In addition, in actual applications, the measured data of three-cylinder pumps is often limited by sensor deployment, and there are data missing or noise problems. Relying solely on simulation data generated by digital twins faces the challenge of insufficient physical modeling accuracy, which makes it difficult for traditional technologies to achieve high-precision real-time predictions in multi-variable coupling scenarios.

[0003] To address these issues, the existing disconnect between data-driven and physical modeling leads to a conflict between accuracy and real-time performance in triplex pump pressure prediction. Physical models are computationally expensive, making them difficult to meet online monitoring requirements. While lightweight data models improve speed, their lack of physical constraints can easily lead to prediction errors. Effectively integrating physical prior knowledge with data features to achieve high-precision, interpretable predictions of triplex pump pressure fields has become a technical challenge in the field of fluid machinery. Summary of the Invention

[0004] Based on the technical problems existing in the background technology, the present invention proposes a three-cylinder pump pressure measurement method, system and medium based on digital twin and neural network, which not only conforms to the constraints of physical knowledge, but also can reflect the changing laws of field data in real time, thereby enhancing the reliability and stability of pressure prediction.

[0005] The digital twin and neural network triplex pump pressure measurement method proposed in this invention inputs the time series data of the pump parameters of the triplex pump during operation into a trained physical information neural network model to obtain the pressure field information of the outlet manifold cavity; The training process of the physical information neural network model is as follows: Using digital twin technology, a simulation model mapped to the physical triplex pump was constructed to obtain time series data of pump parameters under normal and abnormal conditions. After preprocessing, a training set was constructed. Partial differential equations are embedded in the physical information neural network model as the physical loss function of physical constraints. The physical information neural network model is trained using the training set, and the trainable parameters are adjusted through reverse updating.

[0006] Furthermore, the time series data of the pump parameters include the pressure of the inlet manifold chamber, the pressure of the outlet manifold chamber, the motor current, the mass flow rate of the liquid in the inlet manifold chamber, and time data of corresponding moments.

[0007] Furthermore, after obtaining the time series data, the time series data is preprocessed as follows: Remove outliers and duplicate values ​​of adjacent data points from the time series data, and divide the data into a training set, a validation set, and a test set in chronological order; Standardize the data except time data in time series data; The value of the time data in the time series data is reduced by a set multiple so that the reduced time data is only used for the partial differential equation term.

[0008] Furthermore, the partial differential equation term includes a partial derivative component term, and the physical loss function The formula is as follows: ; in, is the volume of the outlet manifold cavity, is the equivalent density of the fluid mixture in the triplex pump, is the pressure of the outlet manifold cavity, is the time value, is the liquid mass flow rate in the inlet manifold cavity, is the sample index, is the total number of samples.

[0009] Furthermore, when the triplex pump is in pure oil mode, The calculation formula is as follows: ; in, is the density of oil at atmospheric pressure, is the bulk modulus of the oil, is the current pressure of the outlet manifold chamber, is atmospheric pressure.

[0010] Furthermore, when the triplex pump is in the oil-air mixture mode, The calculation formula is as follows: ; in, is the density of the mixture, is the bulk modulus of the oil, is atmospheric pressure, is the air volatility index, is the initial mixture density, is the normalized air volume fraction.

[0011] Furthermore, the The calculation formula is as follows: ; in, is the density of oil at atmospheric pressure, is the gas density at atmospheric pressure, The air fraction indicates the proportion of air mixed in the liquid.

[0012] Furthermore, the physical information neural network model is trained using a total loss function, which includes a physical loss function and a mean absolute error loss function. Specifically: ; in, is the mean absolute error loss function, is the physical loss function, is the physical loss weight.

[0013] A computer system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the method described above in the computer program.

[0014] A computer-readable storage medium stores a plurality of classification programs, wherein the plurality of classification programs are used to be called by a processor and execute the method described above.

[0015] Those skilled in the art will understand that all or part of the steps of implementing the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk, etc. Various media that can store program codes.

[0016] The advantages of the digital twin and neural network three-cylinder pump pressure measurement method, system and medium provided by the present invention are: using digital twin technology to construct an analog simulation model to obtain time series data, alleviate the problem of data acquisition, embed the partial differential equations that the modeling object conforms to into the physical information neural network model, and use the physical information neural network model to integrate physical prior knowledge and data characteristics, while ensuring physical interpretability, improving the pressure field prediction accuracy and real-time performance, and using the embedded partial differential equation terms as physical constraints to train and adjust the hyperparameter optimization model, ultimately obtaining a three-cylinder pump pressure measurement method that not only conforms to the constraints of physical knowledge but can also reflect the changing laws of field data in real time, thereby enhancing the reliability and stability of pressure prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart of the training process of the physical information neural network model. DETAILED DESCRIPTION

[0018] The technical solutions of the present invention are described in detail below through specific embodiments. Numerous specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0019] like Figures 1 to 2 As shown, the digital twin and neural network triplex pump pressure measurement method proposed in the present invention inputs the pump parameter time series data of the triplex pump during operation into the trained physical information neural network model to obtain the outlet manifold cavity pressure field information; The training process of the physical information neural network model is as follows: Step 1: Use digital twin technology to build a simulation model that maps to the physical three-cylinder pump, obtain time series data including pump parameters under normal and abnormal conditions, and construct a training set after preprocessing; Step 2: Embed partial differential equation terms in the physical information neural network model as the physical loss function of physical constraints, use the training set to train the physical information neural network model, and adjust the trainable parameters through reverse updating.

[0020] This embodiment addresses the problems in the existing technology, such as complex physical modeling calculations, difficulty in adapting to complex working conditions, and data-driven reliance on a large amount of high-quality data, as well as the separation of the two leading to a contradiction between accuracy and real-time performance. Digital twin technology is used to construct a simulation model to obtain time series data and alleviate the data acquisition problem. The partial differential equations that the modeling object conforms to are embedded in a physics-informed neural network (PINN) model. The PINN model is used to integrate physical prior knowledge and data features, thereby improving the pressure field prediction accuracy and real-time performance while ensuring physical interpretability. The embedded partial differential equation terms are used as physical constraints to train and adjust the hyperparameter optimization model. Ultimately, a three-cylinder pump pressure measurement method is obtained that not only conforms to the constraints of physical knowledge but can also reflect the changing laws of field data in real time, thereby enhancing the reliability and stability of pressure prediction.

[0021] In one embodiment, step 1 is to use digital twin technology to obtain time series data of pump parameters under normal and abnormal conditions through a simulation model mapped to a physical three-cylinder pump, and to construct a training set after preprocessing, specifically: (a1) Construct a simulation (Simulink) model of the three-cylinder pump. For the simulation model, various existing methods can be used, such as the Simulink simulation model of the Matlab series, the multi-physics field coupling model built by general finite element analysis software such as ANSYS, and the flow field simulation model built by professional CFD software such as CFX and Fluent.

[0022] (a2) By running the simulation model, the time series data of the pump parameters under normal and abnormal conditions are obtained, and the data are divided into a training set, a validation set, and a test set in sequence; Among them, abnormal conditions include leakage, blockage, and bearing wear. The time series data of the pump parameters when the three-cylinder pump is running include the pressure of the inlet manifold cavity, the pressure of the outlet manifold cavity, the motor current, the liquid mass flow rate of the inlet manifold cavity, and the time data of the corresponding moments.

[0023] (a3) Standardize the data except time data in the time series data; ; (1) in, is the standardized data, is the data before normalization, They are The mean and standard deviation of the attribute set; (a4) reducing the value of the time data in the time series data by a set multiple so that the reduced time data is only used for the partial differential equation term; The value of the time data is reduced by a set multiple (for example, 10,000 times) to reduce its weight in the physical information neural network model, so that the time value data is only used for the derivation of the partial differential equation terms without affecting the training of the physical information neural network model.

[0024] (a5) After data preprocessing from (a1) to (a4), the training set is fed into the physical information neural network model for model training.

[0025] This embodiment uses a simulation model to obtain time series data to solve the problem of missing abnormal data of the three-cylinder pump. The training set, validation set, and test set can be divided as needed. This embodiment preferably configures the volume of the training set to account for more than 70% of the volume of the total data set so that the training set can reflect the characteristics of the total data set.

[0026] In one embodiment, step 2 is to embed partial differential equations in the physical information neural network model as a physical loss function of physical constraints, train the physical information neural network model using a training set, and adjust the trainable parameters through reverse updating, specifically: (b1) Embedding partial differential equations as physical loss functions for physical constraints in the physical information neural network model; First, a theoretical physical model for triplex pump pressure prediction is constructed, and the formula is as follows: , (2); in, is the volume of the outlet manifold cavity, is the equivalent density of the fluid mixture in the triplex pump, is the current pressure of the outlet manifold chamber, is the time value, is the liquid mass flow rate in the inlet manifold cavity, is the sample index, is the total number of samples.

[0027] The value of is known and constant, The value of is calculated by automatic differentiation in PyTorch, an open source deep learning framework developed by Facebook AI Research (FAIR). It is the input feature of the model and is also known, that is, the training set. The calculation formula is as follows: When the triplex pump is in pure oil mode: , (3); in, is the density of oil at atmospheric pressure, is the bulk modulus of the oil, is the current pressure of the outlet manifold chamber, is atmospheric pressure.

[0028] When the triplex pump is in oil-air mixture mode: , (4); in, is the mixture density, is the bulk modulus of the oil, is atmospheric pressure, is the air volatility index, is the initial mixture density, is the normalized air volume fraction.

[0029] against The calculation formula is as follows: , (5); in, is the density of oil at atmospheric pressure, is the gas density at atmospheric pressure, The air fraction indicates the proportion of air mixed in the liquid.

[0030] Normalized air volume fraction , which is calculated as follows: , (6); in, is the air fraction, is atmospheric pressure, is the current pressure of the outlet manifold chamber, is the proportion of air entrained in the liquid.

[0031] (b2) Constructing a physical information neural network model; The training set used as input for the physical information neural network model includes the pressure of the inlet manifold cavity, the pressure of the outlet manifold cavity, the motor current, the liquid mass flow rate of the inlet manifold cavity, and the time data at the corresponding moments; The physical information neural network model consists of an input layer, a fully connected layer, an output layer, and a loss function layer. The input layer is used to receive the time data of the triplex pump during operation. , motor current , pressure in the inlet manifold cavity , Current outlet manifold pressure , Liquid mass flow rate of the inlet manifold cavity ; The hidden layer in the fully connected layer uses Swish as the activation function; the output layer is used to output The pressure of the triplex pump outlet manifold chamber at the moment , relative to The derivatives of t are calculated using automatic differentiation, which makes it easier to put them into the loss function. to calculate the physical loss function in back propagation.

[0032] It should be noted that the key concept of the physical information neural network model in this embodiment is to embed physical constraints into the neural network, enabling the network to learn the behavior of physical systems and satisfy physical equations. In this embodiment, the fully connected layer consists of six hidden layers, each containing 258 neurons, and uses Swish as the activation function.

[0033] because When the physical information is input into the neural network model, it is scaled by a set multiple (for example, 10,000 times). According to the following formula, the value after automatic differentiation needs to be divided by 10,000 times to get the actual : The time it takes to pass physical information into the neural network model is actually ,and ; According to the chain rule: , (7); And because: , (8); therefore: , (9); Based on the above records, this embodiment proposes the total loss function of the physical information neural network model as follows: , (10); in, is the mean absolute error loss function, is the physical loss function, is the physical loss weight.

[0034] in, , (11); , (12); Where, For the The pressure prediction result of each sample, that is, the predicted value (the outlet manifold cavity pressure field information predicted by the physical information neural network model), is the pressure simulation data of the i-th sample, that is, the true value (the real outlet manifold cavity pressure field information), is the total number of samples.

[0035] It can intuitively reflect the data error between the actual value and the predicted value of the physical information neural network model, and is used to determine whether the accuracy of the prediction result meets the requirements of actual use. The training of the physical information neural network model can be constrained so that the physical information neural network model can learn the behavior of the physical system and satisfy the physical equations (Formula (2)) as much as possible.

[0036] The physical information neural network model is trained based on the above total loss function and the input training set, and a physical information neural network model with better prediction effect is found by adjusting the trainable parameters. The trainable parameters include batch size, hidden layer size, number of hidden layers, learning rate size, and physical loss weight.

[0037] The physical information neural network model outputs the magnetic induction intensity and the pressure of the manifold cavity at the outlet of the triplex pump during the forward propagation process. , which is relative to and The derivatives at the moment are calculated using automatic differentiation, so that they can be substituted into the physical loss function (Formula (11)) for back propagation when optimizing the model to ensure that the model meets the constraints of the physical equations. Automatic differentiation is accomplished through the automatic differentiation (Autograd function) of the PyTorch framework. Its principle is to dynamically construct a computational graph and automatically calculate the gradient of each parameter during back propagation using the chain rule, so that the model can consider both data errors (data-driven losses, i.e., ) also considers the errors in the physical equations (loss of physical drive, i.e. ).

[0038] When training the physical information neural network model, the total loss function can be made to The value of approaches 0. When the total loss function approaches 0, the model can be considered an approximation of the triplex pump. This completes the PINN network training. During the training process, the validation set is used to verify whether the trained physical information neural network model meets the required performance requirements. Finally, the trained physical information neural network model is tested on the test set. Finally, the trained physical information neural network model is used to predict the pressure of the triplex pump.

[0039] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A three-cylinder pump pressure measurement method based on digital twin and neural network, characterized in that: The time series data of the pump parameters when the triplex pump is working is input into the trained physical information neural network model to obtain the pressure field information of the outlet manifold cavity; The training process of the physical information neural network model is as follows: Using digital twin technology, a simulation model mapped to the physical triplex pump was constructed to obtain time series data of pump parameters under normal and abnormal conditions. After preprocessing, a training set was constructed. Partial differential equations are embedded in the physical information neural network model as the physical loss function of physical constraints. The physical information neural network model is trained using the training set, and the trainable parameters are adjusted through reverse updating.

2. The triplex pump pressure measurement method according to claim 1, characterized in that: The time series data of the pump parameters include the pressure of the inlet manifold chamber, the pressure of the outlet manifold chamber, the motor current, the liquid mass flow rate of the inlet manifold chamber, and the time data of the corresponding moments.

3. The triplex pump pressure measurement method according to claim 1, characterized in that: After obtaining the time series data, the preprocessing process of the time series data is as follows: Remove outliers and duplicate values ​​of adjacent data points from the time series data, and divide the data into a training set, a validation set, and a test set in chronological order; Standardize the data except time data in time series data; The value of the time data in the time series data is reduced by a set multiple so that the reduced time data is only used for the partial differential equation term.

4. The triplex pump pressure measurement method according to claim 1, characterized in that: The partial differential equation term includes a partial derivative component term, and the physical loss function The formula is as follows: in, is the volume of the outlet manifold cavity, is the equivalent density of the fluid mixture in the triplex pump, is the current pressure of the outlet manifold chamber, is the time value, is the liquid mass flow rate in the inlet manifold cavity, is the sample index, is the total number of samples.

5. The triplex pump pressure measurement method according to claim 4, characterized in that: When the triplex pump is in pure oil mode, The calculation formula is as follows: in, is the density of oil at atmospheric pressure, is the bulk modulus of the oil, is atmospheric pressure.

6. The triplex pump pressure measurement method according to claim 4, characterized in that: When the triplex pump is in oil-air mixture mode, The calculation formula is as follows: in, is the density of the mixture, is the bulk modulus of the oil, For current pressure, is atmospheric pressure, is the air volatility index, is the initial mixture density, is the normalized air volume fraction.

7. The triplex pump pressure measurement method according to claim 6, characterized in that: described The calculation formula is as follows: in, is the density of oil at atmospheric pressure, is the gas density at atmospheric pressure, The air fraction indicates the proportion of air mixed in the liquid.

8. The triplex pump pressure measurement method according to claim 1, characterized in that: The physical information neural network model is trained using the total loss function, which includes the physical loss function and the mean absolute error loss function. Specifically: in, is the mean absolute error loss function, is the physical loss function, is the physical loss weight.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of classification programs, which are used to be called by a processor and execute the method according to any one of claims 1 to 8.

Images