Method and system for observing pressure in a fluid pipe based on a residual neural network

Abstract

The application relates to a residual neural network-based fluid pipeline internal pressure observation method and system. A residual neural network-based fluid pipeline internal pressure observer state space equation is generated by constructing a fluid flow state space equation, residual neural network hyperparameters, and an output and input equation of the network. The data input of the neuron of the residual neural network depends on the parameter input of the fluid flow state space equation, so the residual neural network-based fluid pipeline internal pressure observer state space equation has strong adaptability. When the residual neural network-based fluid pipeline internal pressure observer state space equation is trained and optimized, the network weight matrix of the residual neural network in the transmission process of adjacent residual neural network layers is approximated and fine-tuned with the actual pressure value, so that the pressure change of each section in the pipeline can be accurately observed, and the limitation that traditional fluid pipeline internal pressure observation needs to be repeatedly observed is solved.

Description

Technical Field

[0001] This application relates to the field of fluid pipeline pressure observation technology, and in particular to a method and system for fluid pipeline pressure observation based on residual neural networks. Background Technology

[0002] Fluid pipelines play a vital role in industry, transporting various types of fluids such as liquids, gases, and steam. They are widely used in chemical, oil and gas, food and pharmaceutical, and energy industries. To ensure safe and efficient transportation, and to achieve precise flow control within the pipelines, thereby improving pipeline efficiency, pressure monitoring within fluid pipelines is crucial. Currently, most fluid pipeline pressure monitoring techniques rely on laboratory experiments or on-site sampling measurements during installation. Such methods have several limitations. Using physical monitoring equipment is time-consuming and labor-intensive, requiring multiple measurements for pipelines of different sizes, fluid velocities, and fluid types. Furthermore, human error is significant during actual measurements, and the actual operation of monitoring physical pipelines can cause some damage. Therefore, to overcome the limitations of traditional fluid pipeline pressure monitoring methods that require repeated experiments, this application proposes a fluid pipeline pressure monitoring method and system based on residual neural networks. Summary of the Invention

[0003] Therefore, it is necessary to provide a method and system for observing pressure in fluid pipelines based on residual neural networks, which addresses the limitation of traditional methods that require repeated observation experiments.

[0004] This application provides a method for observing pressure inside a fluid pipeline based on a residual neural network, including:

[0005] Obtain the fluid flow state space parameters of different types collected by the observer at the current moment, and create a fluid flow state space parameter set;

[0006] Construct fluid flow state space equations based on the fluid flow state space parameter set;

[0007] Obtain the parameter set of the residual neural network, and construct the hyperparameters of the residual neural network and the output and input equations of the network based on the parameter set;

[0008] The initial network weight matrix of adjacent residual neural network layers during the transmission process is analyzed. Based on the initial network weight matrix, the fluid flow state space equation, the hyperparameters of the residual neural network, and the output and input equations of the network, the initial state space equation of the fluid pipe pressure observer based on the residual neural network is generated.

[0009] The actual pressure value at the current moment is obtained from the observer. The difference between the pressure observation value and the actual pressure value is input into the hyperparameters of the residual neural network and the output and input equations of the network. The pressure observation value at the current moment is obtained by calculating the initial state space equation of the pressure observer in the fluid pipeline based on the residual neural network.

[0010] After the first preset sampling interval, the network weight matrix is ​​updated, and the actual pressure value collected by the acquisition observer at the current moment is returned until the error between the pressure observation value and the actual pressure value is zero, so as to adjust the state space equation of the fluid pipeline pressure observer based on the residual neural network and output the state space equation of the fluid pipeline pressure observer based on the residual neural network.

[0011] After the second preset sampling interval, the flow state space parameters of different types of fluids collected by the current time of the acquisition observer are returned, and a flow state space parameter set is created until the state space equations of M fluid pipe pressure observers based on residual neural networks are constructed; the second preset sampling interval is greater than the first preset sampling interval.

[0012] Obtain the state space parameter set of the fluid to be measured, determine the state space equation of the fluid pipe pressure observer based on residual neural network that matches the state space parameter set of the fluid to be measured, extract the pressure and velocity data of each pipe segment in the state space parameter set of the fluid to be measured, input the pressure and velocity data of each pipe segment into the state space equation of the fluid pipe pressure observer based on residual neural network that matches the state space parameter set of the fluid to be measured, and generate the observed pressure distribution data inside the pipe.

[0013] This application also provides a fluid pipeline pressure observation system based on residual neural networks, including:

[0014] The host computer is used to execute a method for observing pressure inside a fluid pipeline based on a residual neural network.

[0015] The observer is connected to the host computer for communication.

[0016] This application relates to a method and system for observing pressure in fluid pipelines based on residual neural networks. It constructs a fluid flow state-space equation using a set of fluid flow state-space parameters, and constructs the hyperparameters and output / input equations of the residual neural network using a set of residual neural network parameters. Based on the fluid flow state-space equation, the hyperparameters, and the output / input equations of the residual neural network, a state-space equation for a pressure observer in a fluid pipeline based on a residual neural network with multiple forward propagation layers and a residual layer is generated. The data input of the neurons in the residual neural network depends on the parameter input of the fluid flow state-space equation, thus the state-space equation for the pressure observer in a fluid pipeline based on the residual neural network has strong adaptability. During the training and optimization of the state-space equation for the pressure observer in a fluid pipeline based on the residual neural network, the network weight matrix of the residual neural network is approximated and fine-tuned with the actual pressure value during the transmission process between adjacent residual neural network layers. This enables the state-space equation for the pressure observer in a fluid pipeline based on the residual neural network to accurately observe the pressure changes in each section of the pipeline, overcoming the limitation of traditional fluid pipeline pressure observation requiring repeated observation experiments. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0018] Figure 1 This is a module connection diagram of a fluid pipeline pressure observation method based on residual neural networks provided in an embodiment of this application.

[0019] Figure 2 The diagram shows the activation function of a residual neural network for a fluid pipeline pressure observation method based on a residual neural network, as provided in an embodiment of this application.

[0020] Figure 3 This is a schematic diagram of the observer structure for a fluid pipeline pressure observation method based on residual neural networks, provided in an embodiment of this application.

[0021] Figure 4 This is a module connection diagram of a fluid pipeline pressure observation system based on residual neural networks, provided as an embodiment of this application.

[0022] Figure label:

[0023] 100 - Host computer; 200 - Observer. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] This application provides a method and system for observing pressure inside a fluid pipeline based on a residual neural network.

[0026] like Figure 1 As shown in one embodiment of this application, the fluid pipeline pressure observation method based on residual neural network includes:

[0027] S100: Obtain the flow state space parameters of different types of fluids collected by the observer at the current moment, and create a flow state space parameter set.

[0028] S200, constructs the fluid flow state space equation based on the fluid flow state space parameter set.

[0029] S300: Obtain the parameter set of the residual neural network, and construct the hyperparameters of the residual neural network and the output and input equations of the network based on the parameter set.

[0030] S400 parses the initial network weight matrix of adjacent residual neural network layers during the transmission process, and generates the initial state space equation of the fluid pipe pressure observer based on the initial network weight matrix, the fluid flow state space equation, the hyperparameters of the residual neural network, and the network's output and input equations.

[0031] S500: Obtain the actual pressure value collected by the observer at the current moment, input the difference between the pressure observation value and the actual pressure value into the hyperparameters of the residual neural network and the output and input equations of the network, and calculate the pressure observation value at the current moment through the initial state space equation of the fluid pipeline pressure observer based on the residual neural network.

[0032] S600, after the first preset sampling interval, update the network weight matrix, return the actual pressure value collected by the acquisition observer at the current moment, until the error between the pressure observation value and the actual pressure value is zero, so as to adjust the state space equation of the fluid pipeline pressure observer based on the residual neural network and output the state space equation of the fluid pipeline pressure observer based on the residual neural network.

[0033] S700, after the second preset sampling interval, return to the different types of fluid flow state space parameters collected by the current time of the acquisition observer, create a flow state space parameter set, until the state space equations of M fluid pipeline pressure observers based on residual neural networks are constructed; the second preset sampling interval is greater than the first preset sampling interval.

[0034] S800: Obtain the state space parameter set of the fluid to be measured, determine the state space equation of the fluid pipe pressure observer based on residual neural network that matches the state space parameter set of the fluid to be measured, extract the pressure and velocity data of each pipe segment in the state space parameter set of the fluid to be measured, input the pressure and velocity data of each pipe segment into the state space equation of the fluid pipe pressure observer based on residual neural network that matches the state space parameter set of the fluid to be measured, and generate the observed pressure distribution data inside the pipe.

[0035] Specifically, the fluid flow state space equation is:

[0036]

[0037] Where X(t) is the state variable, specifically expressed as

[0038] [p1(t)…p N (t) v0(t)…v N-1 (t)] T ∈R 2N

[0039] p i (t) represents the pressure value at point i in the pipeline at time t, v i-1 (t) is the flow velocity at point i-1 in the pipeline at time t, where i = 1, ..., N and N is the number of segments in the pipeline, which is set to N = 8 in this embodiment;

[0040] Matrix A is represented as:

[0041]

[0042] And A 12 With A 21 They are respectively:

[0043]

[0044]

[0045] Where c is the wave velocity, ρ is the flow density, Δl = l / N and l is the pipe length, which are set to c = 1200 m / s and ρ = 1000 kg / m in this embodiment. 3 And Δl = l / N = 18.54m / 8 = 2.3175m;

[0046] Matrix B is represented as:

[0047]

[0048] Where f is the Darcy-Weisbach friction coefficient, which is taken as f = 0.03 in this embodiment; D is the diameter of the pipe, which is set to D = 0.1m in this embodiment, and I is an N×N identity matrix; P * The pressure p0(t) at the connection between the container and the pipe is P in this embodiment. * =p0(t)=2×10 5 Pa,U(t) is the control signal, i.e., the flow rate at the valve;

[0049] In addition, matrices a and b are respectively:

[0050]

[0051]

[0052] Matrix C is a 2N×2N identity matrix, the operator o represents the Hadamard operation on matrices, and Y(t)∈R 2 N is the output vector.

[0053] The hyperparameters of the residual neural network and the network's output and input equations are as follows:

[0054]

[0055] Where υ is the set of network hyperparameters, L is the number of network layers, and θ g Let θ be the weight matrix of the g = j = 1, ... L layers of the network's forward propagation part. s Let n be the weight matrix of the residual part. χ Let represent the total number of neurons in the network that contain non-linear activation functions, and let In represent the network input. It is the output of the j-th layer of the forward propagation part, σ [j] (·) is the nonlinear activation output of the j-th layer in the forward propagation part. This is the final output of the residual neural network. Note that when j=1...

[0056]

[0057] State-space equation of a pressure observer in a fluid pipe based on a residual neural network:

[0058]

[0059] K≠0 represents the variable input coefficients of the network;

[0060] In fact, This involves adjusting the parameters between the observed pressure values ​​and the actual pressure values ​​in the state-space equation of a pressure observer in a fluid pipeline based on a residual neural network.

[0061] This embodiment relates to a method for observing pressure in a fluid pipeline based on a residual neural network. A fluid flow state space equation is constructed using a set of fluid flow state space parameters. Hyperparameters and the network's input and output equations are then constructed using the residual neural network parameter set. Based on the fluid flow state space equation, the hyperparameters, and the network's input and output equations of the residual neural network, a state space equation for a pressure observer in a fluid pipeline based on a residual neural network is generated, featuring multiple forward propagation layers plus a residual layer. The data input to the neurons of the residual neural network depends on the parameter input of the fluid flow state space equation; therefore, the state space equation for the pressure observer in a fluid pipeline based on the residual neural network has strong adaptability. During the training and optimization of the state space equation for the pressure observer in a fluid pipeline based on the residual neural network, the network weight matrix of the residual neural network is approximated and fine-tuned with the actual pressure value during the transmission process between adjacent residual neural network layers. This enables the state space equation for the pressure observer in a fluid pipeline based on the residual neural network to accurately observe pressure changes in different sections of the pipeline, overcoming the limitation of traditional fluid pipeline pressure observation methods that require repeated observation experiments.

[0062] In one embodiment of this application, step S100 includes:

[0063] S111 receives pipe length data and pipe diameter data.

[0064] S112 receives fluid flow density data.

[0065] S113, receiving fluid wave velocity.

[0066] S114, Number of segments in the receiving pipeline.

[0067] S115 receives pressure and flow velocity data from each pipe segment.

[0068] S116, receiving the Darcy-Weisbach friction coefficient.

[0069] S117, Create a fluid flow state space parameter set, which includes pipe length and diameter data, fluid flow density data, fluid wave velocity, number of receiving pipe segments, pressure and velocity data of each receiving pipe segment, and Darcy-Weisbach friction coefficient.

[0070] This embodiment involves a set of fluid flow state space parameters. These parameters are the fundamental parameters for constructing the fluid flow state space equations, and the number of pipe segments affects the modeling of the residual neural network.

[0071] In one embodiment of this application, step S200 further includes:

[0072] S210, the function of the flow velocity at the valve in the receiving pipeline as a function of time.

[0073] S220, the pressure constant at the connection between the receiving container and the pipeline.

[0074] S230, using the fluid flow state space parameter set, determine the first parameter matrix of the function of the flow velocity at the pipeline valve changing with time.

[0075] S240, using the fluid flow state space parameter set, determines the second parameter matrix of the pressure constant at the connection between the container and the pipe.

[0076] S250 uses the fluid flow state space parameter set, the flow velocity change function with time at the pipe valve, the pressure constant at the connection between the container and the pipe, the first parameter matrix and the second parameter matrix to construct the fluid flow state space equation.

[0077] This embodiment involves the fluid flow state space equation. In reality, the velocity variation over time at the pipe valve has a significant impact on the overall fluid flow state space equation. Furthermore, even with identical pipe and fluid conditions, replacing the pipe valve will change the velocity variation over time at that valve. It is worth noting that, to facilitate the observation of internal pipe pressure without damaging the installed pipe, all the above parameters can be observed individually under simulation conditions with identical device specifications and requirements, and then integrated into the fluid flow state space equation.

[0078] In one embodiment of this application, step S300 includes:

[0079] S311 receives the parameters set for the forward propagation layer of the residual neural network.

[0080] S312 receives the residual layer setting parameters of the residual neural network.

[0081] S313 receives parameters regarding the number of neurons in the residual layer for the number of pipeline segments.

[0082] S314 creates a residual neural network parameter set, which includes the parameters for the forward propagation layer of the residual neural network, the parameters for the residual layer of the residual neural network, and the parameters for the number of neurons in the residual layer related to the number of pipeline segments.

[0083] This embodiment relates to the state-space equations of fluid flow. The residual neural network parameter set is used to construct the residual neural network framework. More specifically, this embodiment sets up four layers of the residual neural network forward propagation layer and one layer of the residual neural network residual layer. The number of neurons in the forward propagation layer of the residual neural network is distributed as 2N-(2N+4)-(2N+2)-2N, specifically 16-20-18-16; the number of neurons in the residual layer is 2N, specifically 16.

[0084] The input vector In∈R of the residual neural network at each sampling time step 2N for:

[0085]

[0086] Among them The observed pressure value at point i in the pipeline at time t. This represents the observed flow velocity at point i-1 in the pipeline at time t.

[0087] In one embodiment of this application, step S300 further includes:

[0088] S321, receiving time with respect to the pressure and fluid velocity matrix of each pipe segment.

[0089] S322 uses the received time-dependent pressure and fluid velocity matrices for each pipe segment as the input vector for the residual neural network.

[0090] S323, obtain the output vector of the forward propagation layer of the residual neural network, use the output vector of the previous layer of the forward propagation layer of the residual neural network as the input vector of the next layer of the forward propagation layer of the residual neural network, until the output vector of the last layer of the forward propagation layer of the residual neural network is obtained.

[0091] S324 uses the output vector of the last layer of the residual neural network's forward propagation layer as the hyperparameter of the residual neural network and the output vector of the residual neural network's forward propagation layer in relation to the network's output and input equations.

[0092] Specifically, the input vector is fed into the first layer of the residual neural network's forward propagation, where In1∈R. 2N for:

[0093] In1=θ 1 In (1.5)

[0094] Where θ 1 ∈R 2N×2N This is the weight matrix for the first layer;

[0095] Output of the first layer for:

[0096]

[0097] The specific formula for calculating the function tanh(·) is as follows:

[0098]

[0099] The output of the first layer becomes the input of the second layer, In2∈R. 2N+4The calculation formula is:

[0100] Where θ 2 ∈R (2N+4)×2N This is the weight matrix for the second layer;

[0101] The output of the second layer for:

[0102]

[0103] The output of the second layer is fed into the input of the third layer, In3∈R. 2N+2 The calculation formula is:

[0104]

[0105] Where θ 3 ∈R (2N+2)×(2N+4) The weight matrix of the third layer; the output of the third layer. for:

[0106]

[0107] The output of the third layer is fed into the input of the fourth layer, In4∈R. 2N The calculation formula is:

[0108]

[0109] Where θ 4 ∈R 2N×(2N+2) This is the weight matrix for the fourth layer;

[0110] The final output of the fourth layer for:

[0111]

[0112] The specific formula for calculating the ReLU function is as follows:

[0113]

[0114] Where x represents a single element in matrix In4.

[0115] This embodiment involves the hyperparameters of the residual neural network and the output vector of the network's input equation and the output vector of the residual neural network's forward propagation layer. In this embodiment, the final output of the fourth layer is the hyperparameters of the residual neural network and the output vector of the network's input equation and the output vector of the residual neural network's forward propagation layer.

[0116] In one embodiment of this application, step S300 further includes:

[0117] S331, receiving time with respect to the pressure and fluid velocity matrix of each pipe segment.

[0118] S332 uses the received time-dependent pressure and fluid velocity matrices for each pipe segment as the input vector for the residual neural network.

[0119] S333, obtain the output vector of the residual layer of the residual neural network.

[0120] S334 uses the output vector of the residual layer of the residual neural network as the hyperparameter of the residual neural network and the output vector of the residual layer of the residual neural network in the equation of the network's output and input.

[0121] Specifically, the input vector to the input In of the residual layer res ∈R 2N The calculation formula is:

[0122] In res =θ s In (1.15)

[0123] Where θ s ∈R 2N×2N This is the weight matrix of the residual layer;

[0124] Output of residual layer res ∈R 2N The calculation formula is:

[0125] Out res =Sigmoid(In res (1.16)

[0126] The specific formula for calculating the output of the function Sigmoid(·) is as follows:

[0127]

[0128] The final output of the residual neural network is:

[0129]

[0130] This embodiment relates to the hyperparameters of a residual neural network and the output vector of the residual layer in the network's output and input equations. The output vector of the residual layer is a crucial output matrix model in the hyperparameters and output-input equations of the residual neural network. The output vector of the residual layer ensures the good fit between the observed values ​​and the actual pressure values ​​in the state-space equations of the pressure observer within the fluid pipe.

[0131] In one embodiment of this application, step S400 includes:

[0132] S411 is a nonlinear activation function model that receives residual neural network layers.

[0133] S412, converts the operation matrix of the nonlinear activation function model.

[0134] S413 performs nonlinear isolation on the residual neural network layer and applies secondary constraints to the operation matrix of the nonlinear activation function in the residual neural network layer.

[0135] S414, simplifies the operation matrix of the nonlinear activation function in the residual neural network layer.

[0136] S415, Analyze the operation matrix related to the nonlinear activation function in the residual neural network layer.

[0137] S416, obtains the convergence criterion for the operation matrix of the nonlinear activation function in the residual neural network layer.

[0138] S417, use the first positive definite matrix of the operation matrix of the nonlinear activation function in the parsed residual neural network layer to parse the weight matrix of the corresponding residual neural network layer.

[0139] Specifically, the input and output of the j-th layer during the forward propagation of the residual neural network.

[0140] enter This indicates the data received by this layer;

[0141] Output This represents the feature obtained after the input of this layer is processed by a nonlinear activation function. We can obtain the matrix form of the residual neural network operation as follows:

[0142]

[0143] Where O represents an all-zero matrix;

[0144] To simplify the description, we use vectors. and τ σ Let them represent respectively:

[0145]

[0146] Define the following two matrices:

[0147]

[0148] Where E is the identity matrix, introducing the concept of vectors... The linear matrix transformation of is expressed as follows:

[0149]

[0150] After isolating the neural network nonlinearly, we need to impose secondary constraints on the nonlinear activation functions in the network. Specifically, this constraint should hold that for any real number r, the following holds:

[0151] (σ(r)-αr)(βr-σ(r))≥0 (1.22)

[0152] Where σ(r) is a nonlinear activation function that satisfies the condition, the range of the nonlinear activation function is restricted to a sector region [αr βr], where the lower boundary coefficient and the upper boundary coefficient of the sector region are α and β respectively, and α≤β<∞.

[0153] The activation function that meets the conditions is shown in the following formula:

[0154]

[0155] Where γ and κ are both nonlinear coefficients, and the specific graph of the function is shown in the figure. Figure 2 As shown.

[0156] We can use vectors and Let represent the upper and lower boundary coefficients of all nonlinear activation functions in the residual neural network;

[0157] If it exists And λ σ >0;

[0158] Then, for the boundary coefficient vector α of the nonlinear activation function of the residual neural network... σ and β σ We can obtain:

[0159]

[0160] Where matrix ψ σ (α σ ,β σ ) and M σ (λ σ They are respectively:

[0161]

[0162] Where diag(·) represents constructing a diagonal matrix;

[0163] There exists a positive definite matrix P, i.e., P > O;

[0164] The convergence criterion of the observer is obtained by satisfying the following inequalities:

[0165]

[0166] but That is, what is observed by the observer It will definitely approximate the actual Y(t);

[0167] By setting a suitable positive definite matrix P and solving matrix inequality (1.26), matrix H can be obtained. χ and This corresponds to the weight matrix of the residual neural network.

[0168] This embodiment relates to the initial network weight matrix of the residual neural network during the transmission process between adjacent residual neural network layers. The introduction and precise construction of the weight matrix helps to ensure the observation accuracy of the residual neural network observer, while also ensuring the convenience of training the residual neural network observer equation.

[0169] In one embodiment of this application, step S400 includes:

[0170] S421, the initial network weight matrix of the residual neural network is analyzed to explain the transmission process between adjacent residual neural network layers.

[0171] S422, Analyze the hyperparameters of the residual neural network and the network's output and input equations.

[0172] S423, Analyze the state space equations for fluid flow.

[0173] S424 generates the state-space equations for a fluid pipeline pressure observer based on a residual neural network.

[0174] This embodiment relates to the construction of the state-space equation for a pressure observer in a fluid pipeline based on a residual neural network. In fact, the state-space equation for a pressure observer in a fluid pipeline based on a residual neural network, generated after analyzing the hyperparameters of the residual neural network and the network's output and input equations with the fluid flow state-space equation, is a parameter-adjustable pipeline pressure model. During later training, it can be decoupled from the physical model of the actual object.

[0175] In one embodiment of this application, step S600 includes:

[0176] S610 receives the difference matrix between the observed pressure value and the actual pressure value.

[0177] S620, input the difference matrix into the forward propagation layer and residual layer of each residual neural network respectively.

[0178] S630, using the convergence criterion of the operation matrix of the nonlinear activation function in the residual neural network layer, determines the second positive definite matrix of the operation matrix of the nonlinear activation function in the residual neural network layer.

[0179] S640 uses the second positive definite matrix to analyze the weight matrix of the corresponding residual neural network layer.

[0180] S650 overwrites the original weight matrix with the weight matrix of the corresponding residual neural network layer obtained from the analysis.

[0181] This embodiment relates to the fitting training of the state-space equation of a pressure observer in a fluid pipeline based on a residual neural network. Different values ​​of K and different positive definite matrices P are used, according to... Figure 3 Under the observer fitting training structure, updating the network weights at each sampling time will cause the error between the observed and actual values ​​to approach zero. At this point, the state-space equation of the fluid pipeline pressure observer based on the residual neural network can achieve high-precision observation of the pipeline pressure value.

[0182] This application provides a fluid pipeline pressure observation system based on residual neural networks.

[0183] like Figure 4 As shown, in one embodiment of this application, the device includes a host computer 100 and an observer 200. The host computer 100 is used to execute a method for observing pressure inside a fluid pipeline based on a residual neural network. The observer 200 is communicatively connected to the host computer.

[0184] This embodiment relates to a fluid pipeline pressure observation system based on a residual neural network. The observer 200 is used to collect the data parameters and related dynamic curve data required for the state-space equation of the fluid pipeline pressure observer based on the residual neural network. This ensures that the host computer 100 can obtain suitable data to construct and train the state-space equation of the fluid pipeline pressure observer based on the residual neural network.

[0185] The technical features of the embodiments described above can be combined arbitrarily, and the execution order of the method steps is not restricted. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The embodiments described above only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this patent application.

Claims

1. A method for observing pressure inside a fluid pipeline based on a residual neural network, characterized in that, include: Obtain the fluid flow state space parameters of different types collected by the observer at the current moment, and create a fluid flow state space parameter set; Construct fluid flow state space equations based on the fluid flow state space parameter set; Obtain the parameter set of the residual neural network, and construct the hyperparameters of the residual neural network and the output and input equations of the network based on the parameter set; The initial network weight matrix of adjacent residual neural network layers during the transmission process is analyzed. Based on the initial network weight matrix, the fluid flow state space equation, the hyperparameters of the residual neural network, and the output and input equations of the network, the initial state space equation of the fluid pipe pressure observer based on the residual neural network is generated. The actual pressure value at the current moment is obtained from the observer. The difference between the pressure observation value and the actual pressure value is input into the hyperparameters of the residual neural network and the output and input equations of the network. The pressure observation value at the current moment is obtained by calculating the initial state space equation of the pressure observer in the fluid pipeline based on the residual neural network. After the first preset sampling interval, the network weight matrix is ​​updated, and the actual pressure value collected by the acquisition observer at the current moment is returned until the error between the pressure observation value and the actual pressure value is zero, so as to adjust the state space equation of the fluid pipeline pressure observer based on the residual neural network and output the state space equation of the fluid pipeline pressure observer based on the residual neural network. After the second preset sampling interval, the flow state space parameters of different types of fluids collected by the current time of the acquisition observer are returned, and a flow state space parameter set is created until the state space equations of M fluid pipe pressure observers based on residual neural networks are constructed. The second preset sampling interval is greater than the first preset sampling interval; Obtain the state space parameter set of the fluid to be measured, determine the state space equation of the fluid pipe pressure observer based on residual neural network that matches the state space parameter set of the fluid to be measured, extract the pressure and velocity data of each pipe segment in the state space parameter set of the fluid to be measured, input the pressure and velocity data of each pipe segment into the state space equation of the fluid pipe pressure observer based on residual neural network that matches the state space parameter set of the fluid to be measured, and generate the observed pressure distribution data inside the pipe.

2. The method for observing pressure inside a fluid pipeline based on a residual neural network according to claim 1, characterized in that, The process of acquiring different types of fluid flow state space parameters collected by the observer at the current moment and creating a fluid flow state space parameter set includes: Receive pipe length data and pipe diameter data; Receive fluid flow density data; Received fluid wave velocity: Number of segments in the receiving pipeline; Receive pressure and flow velocity data from each segment of the pipeline; Receive the Darcy-Weisbach coefficient of friction; Create a fluid flow state space parameter set by incorporating pipe length and diameter data, fluid flow density data, fluid wave velocity, number of pipe segments, pressure and velocity data of each pipe segment, and Darcy-Weisbach friction coefficient into the fluid flow state space parameter set.

3. The method for observing pressure inside a fluid pipeline based on a residual neural network according to claim 2, characterized in that, The construction of fluid flow state space equations based on the fluid flow state space parameter set includes: The function of flow velocity at the valve in the receiving pipeline as a function of time; The pressure constant at the connection between the receiving container and the pipeline; Using the fluid flow state space parameter set, determine the first parameter matrix of the function of the flow velocity at the pipeline valve as a function of time; The second parameter matrix of the pressure constant at the connection between the container and the pipe is determined using the fluid flow state space parameter set; The fluid flow state space equation is constructed using the fluid flow state space parameter set, the flow velocity change function with time at the pipe valve, the pressure constant at the connection between the container and the pipe, the first parameter matrix, and the second parameter matrix.

4. The method for observing pressure inside a fluid pipeline based on a residual neural network according to claim 3, characterized in that, The process of obtaining the residual neural network parameter set and constructing the hyperparameters and output / input equations of the residual neural network based on the parameter set includes: Receive parameters set for the forward propagation layer of the residual neural network; Receive residual layer setting parameters for the residual neural network; Receive parameters regarding the number of neurons in the residual layer for the number of pipeline segments; Create a residual neural network parameter set, which includes the parameters for the forward propagation layer of the residual neural network, the parameters for the residual layer of the residual neural network, and the parameters for the number of neurons in the residual layer related to the number of pipeline segments.

5. The method for observing pressure inside a fluid pipeline based on a residual neural network according to claim 4, characterized in that, The process of obtaining the residual neural network parameter set, and constructing the hyperparameters of the residual neural network and the network's output and input equations based on the residual neural network parameter set, further includes: The receiving time is related to the pressure and fluid velocity matrix of each pipe segment; The received time-dependent pressure and fluid velocity matrices for each pipe segment are used as the input vector for the residual neural network. Obtain the output vector of the forward propagation layer of the residual neural network, and use the output vector of the previous layer of the forward propagation layer of the residual neural network as the input vector of the next layer of the forward propagation layer of the residual neural network, until the output vector of the last layer of the forward propagation layer of the residual neural network is obtained. The output vector of the last layer of the residual neural network's forward propagation layer is used as the hyperparameter of the residual neural network, and the output of the network is the output vector of the residual neural network's forward propagation layer in the equation of input.

6. The method for observing pressure in a fluid pipeline based on a residual neural network according to claim 5, characterized in that, The steps of obtaining the parameter set of the residual neural network, constructing the hyperparameters of the residual neural network, and the output and input equations of the network further include: Receive the pressure and fluid velocity matrix of each pipeline segment; The pressure and fluid velocity matrices of each pipe segment are used as the input vectors of the residual neural network. Obtain the output vector of the residual layer in the residual neural network; The output vector of the residual layer of the residual neural network is used as the hyperparameter of the residual neural network, and the output of the network is used as the output vector of the residual layer of the residual neural network in the equation of input and output.

7. The method for observing pressure inside a fluid pipeline based on a residual neural network according to claim 6, characterized in that, The initial network weight matrix of the adjacent residual neural network layers during the transmission process is analyzed. Based on the initial network weight matrix, the fluid flow state space equation, the hyperparameters of the residual neural network, and the network's output and input equations, the initial state space equation of the fluid pipeline pressure observer based on the residual neural network is generated, including: A nonlinear activation function model for receiving residual neural network layers; The operation matrix of this nonlinear activation function model is converted; Nonlinear isolation is applied to the residual neural network layers; Apply quadratic constraints to the operation matrix of the nonlinear activation function in the residual neural network layer; Simplify the operation matrix of the nonlinear activation function in the residual neural network layer; Analyze the operation matrix related to the nonlinear activation function in the residual neural network layer; Convergence criteria for obtaining the operation matrix of the nonlinear activation function in the residual neural network layer; The weight matrix of the corresponding residual neural network layer is analyzed using the first positive definite matrix of the operation matrix of the nonlinear activation function in the analyzed residual neural network layer.

8. The method for observing pressure inside a fluid pipeline based on a residual neural network according to claim 7, characterized in that, The method of analyzing the initial network weight matrix of adjacent residual neural network layers during the transmission process, and generating the initial state space equation of the fluid pipe pressure observer based on the initial network weight matrix, the fluid flow state space equation, the hyperparameters of the residual neural network, and the network's output and input equations, further includes: The initial network weight matrix of the residual neural network is analyzed to explain the propagation process between adjacent residual neural network layers. Analyze the hyperparameters of the residual neural network and the network's output and input equations; Analyzing the state-space equations of fluid flow; Generate the state-space equations for a pressure observer inside a fluid pipe based on a residual neural network.

9. The method for observing pressure inside a fluid pipeline based on a residual neural network according to claim 8, characterized in that, The updated network weight matrix includes: Receive the difference matrix between observed pressure values ​​and actual pressure values; The difference matrix is ​​input into the forward propagation layer and the residual layer of each residual neural network, respectively; The second positive definite matrix of the operation matrix of the nonlinear activation function in the residual neural network layer is determined by using the convergence criterion of the operation matrix of the nonlinear activation function in the residual neural network layer. The weight matrix of the corresponding residual neural network layer is analyzed using the second positive definite matrix. The weight matrix of the corresponding residual neural network layer is overwritten with the original weight matrix.

10. A fluid pipeline pressure monitoring system based on residual neural networks, characterized in that, include: A host computer is used to execute the fluid pipeline pressure observation method based on residual neural network as described in any one of claims 1 to 9; The observer is connected to the host computer for communication.

Images