Prediction method of target erosion amount by particle collision based on back propagation neural network

Abstract

The present invention relates to a method for predicting the erosion amount of a target material due to particle collision based on a back propagation neural network, comprising the following steps: (1) collecting experimental data on the erosion amount of a planar target material due to particle collision, wherein the collected experimental independent variables include 7, namely, particle density ρ, particle diameter D, particle hardness H p , incident speed V, incident angle α, target hardness H t , the hardness ratio of the particle to the target material H p / H t , the dependent variable is the target material erosion amount; (2) to eliminate the influence of the magnitude difference between different variables on the weight distribution, the collected experimental data are standardized; (3) a back propagation neural network structure is constructed, in which the neuron nodes of the input layer correspond to the experimental independent variables, and one neuron node of the output layer corresponds to the dependent variable, i.e., the target material erosion amount. (4) the evaluation index of the neural network prediction results is determined; (5) the network model training is optimized.

Description

Technical Field

[0001] The present invention relates to the field of oil and gas energy transportation, in which solid particles carried by fluid collide with pipe walls, causing pipeline erosion, and specifically relates to a method for predicting the amount of erosion caused by particle collision with a target material based on an optimized back propagation neural network. Background Art

[0002] Regarding the erosion caused by particle collision with planar target, due to the numerous parameters in the collision process, it is challenging to accurately and quickly predict the particle erosion amount under various circumstances. Previous researchers have conducted a lot of research from the aspects of experimental research, theoretical analysis and numerical simulation. For experimental methods, the coverage is small, it will consume a lot of manpower and material resources, and the experimental cycle is long; in theory, certain assumptions are often made on a certain problem, which may cause deviations from the actual situation; in numerical modeling, there will also be some simplifications, there is a gap between the model and the actual situation, and it will consume a lot of computer resources, and sometimes there will be calculation non-convergence problems. In addition, some empirical or semi-empirical formulas derived from experiments or theories often include more parameters, which need to be determined experimentally and are not easy to operate. At the same time, the prediction accuracy of the formula is limited to the experiment conducted by the experimenter himself, and the formula has poor universality. At present, artificial intelligence has begun to be applied to oil and gas energy transportation problems with its advantages of fast calculation speed, stable results, and the ability to process large amounts of data. However, at present, the use of artificial intelligence methods to address the erosion of planar targets caused by particle collisions only focuses on a certain target material or a certain erosion particle, and the scope of application is small. In addition, in the field of artificial intelligence, back propagation neural network is one of the most mature neural network algorithms, which has good self-learning, self-adaptation, robustness and generalization capabilities. A three-layer back propagation neural network can approximate any nonlinear function with arbitrary precision. Back propagation neural network has been widely used in the fields of pattern recognition, function approximation, image processing, etc., but the back propagation algorithm also has disadvantages such as poor convergence speed and easy to fall into local minimum. Therefore, the present invention needs to optimize the back propagation neural network when using it to adapt to this problem and achieve better prediction effect.

[0003] References

[0004] [1]Oka YI,Okamura K,Yoshida T,Practical Estimation of Erosion DamageCaused by Solid Particle Impact:Part 1:Effects of Impact Parameters On aPredictive Equation[J],WEAR,2005,259(1):95-101.

[0005] [2]Oka YI, Yoshida T, Practical Estimation of Erosion Damage Caused by Solid Particle Impact: Part 2: Mechanical Properties of Materials Directly Associated with Erosion Damage[J], WEAR, 2005, 259(1):102-9.

[0006] [3]Huang C,Chiovelli S,Minev Pet al.,A Comprehensive PhenomenologicalModel for Erosion of Materials in Jet Flow[J],POWDER TECHNOL,2008,187(3):273-9. Summary of the invention

[0007] The purpose of the present invention is to provide a method for predicting the erosion of target materials due to particle collision, which is more adaptable, safer and more reliable, and has higher prediction speed and accuracy. The present invention establishes a prediction model based on an optimized back propagation neural network, which can predict the erosion of planar targets with a larger parameter variation range, rather than just the erosion of a certain particle or a certain target material. To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0008] A method for predicting the erosion amount of a target material due to particle collision based on a back propagation neural network comprises the following steps:

[0009] (1) Collect experimental data on the erosion of the target when particles collide with the plane. The experimental independent variables collected include 7, namely particle density ρ, particle diameter D, particle hardness H p , incident speed V, incident angle α, target hardness H t , the hardness ratio of the particle to the target material H p / H t , the dependent variable is the target erosion amount;

[0010] (2) In order to eliminate the influence of magnitude differences between different variables on weight distribution, the collected experimental data are standardized, including standardization between different independent variables and standardization after taking the logarithm of the dependent variable, and the training and test data sets are divided; for the target erosion amount as the dependent variable, the logarithm is first taken and then standardized; the range of each variable is converted to between 0 and 1;

[0011] (3) Constructing a back-propagation neural network structure, wherein the neuron nodes of the input layer correspond to the experimental independent variables, and one neuron node of the output layer corresponds to the dependent variable, i.e., the target material erosion amount.

[0012] (4) Determine the evaluation index of the neural network prediction results

[0013] The coefficient of determination R 2 As an evaluation indicator of the neural network prediction results:

[0014]

[0015] Where: y i is the experimental value of erosion; p i is the predicted value of erosion; n is the number of data; i is the data label; represents the average value of the experimental value of erosion in the data set;

[0016] (5) Network model training optimization, using the Adam gradient descent algorithm, select the optimal network model based on the evaluation index of the neural network prediction results: the input layer is all the experimental independent variables collected in step (1); the hidden layer activation function is relu, and the initialization method is He initialization. The optimal network model is the target erosion model that is finally predicted to be caused by particle collision.

[0017] Furthermore: Step (2) standardizes the collected experimental data, including standardization between different independent variables and standardization after taking the logarithm of the dependent variable; for the target material erosion amount as the dependent variable, the logarithm is taken and then standardized; the variation range of each variable is converted to between 0 and 1.

[0018] Furthermore: the back propagation neural network structure constructed in step (3) is divided into five layers, and the number of neuron nodes in each layer is 7, 20, 15, 10, and 1 respectively.

[0019] Furthermore: it is characterized in that: the method of step (5) is: fixing the output layer activation function to be sigmoid, comparing the prediction results of the network test set when the hidden layer activation functions are sigmoid, tanh and relu respectively; comparing the network prediction effects when the number of input variables is different; comparing the network prediction effects when different initialization methods are used, wherein the Xavier initialization method is used when the hidden layer activation function is sigmoid or tanh, and the He initialization method is used when the hidden layer activation function is relu, to obtain the optimal network model optimized by the Adam gradient descent algorithm.

[0020] The present invention has the following advantages:

[0021] (1) The present invention optimizes the traditional back propagation neural network from four aspects, namely, adopting the optimization algorithm of gradient descent, changing the activation function, adopting the optimized weight initialization method and increasing the number of independent variables. The back propagation neural network has faster convergence speed and smaller fluctuation, and the weight and bias matrix converge to the global optimal value better. The method of the present invention has fast calculation speed and stable calculation results.

[0022] (3) The method of the present invention can predict the erosion amount of a large range of parameter variations, a variety of erosion particles, and a variety of planar target materials, rather than just the erosion of a certain particle or a certain target material, which is often the shortcoming of traditional experiments and numerical simulations, and has a wide range of applications.

[0023] (4) The target erosion prediction accuracy of the method of the present invention is much higher than that of the existing formula, and is safer and more reliable. Compared with the traditional formula in which many parameters need to be determined experimentally, the application of the present invention is more convenient. Only seven independent variables (including particle density, particle diameter, particle hardness, incident speed, incident angle and target hardness) need to be given to quickly calculate the target erosion amount. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the neural network structure of the present invention

[0025] Figure 2 Loss function convergence graph (a) Convergence graph when the hidden layer activation function is sigmoid; (b) Convergence graph when the hidden layer activation function is tanh; (c) Convergence graph when the hidden layer activation function is relu

[0026] Figure 3 Comparison of the determination coefficient of the test set before and after optimization using the initialization method

[0027] Figure 4 Comparison of the erosion amount predicted by the neural network model of the present invention and the actual experimental value (a) training set data; (b) test set data

[0028] Figure 5 Comparison of prediction results between neural network and formula DETAILED DESCRIPTION

[0029] The method of the present invention is further described below in conjunction with specific embodiments:

[0030] Example

[0031] The present invention provides a method for predicting the erosion amount of a target material caused by particle collision based on an optimized back propagation neural network, which mainly comprises the following steps:

[0032] S1. Collect and obtain previous experimental data;

[0033] The present invention collects the experimental data of the erosion problem caused by particle collision on the plane target and constructs a neural network basic database. The engineering conditions of interest include particle density ρ, particle diameter D, particle hardness H p , incident speed V, incident angle α, target hardness H t , the hardness ratio of the particle to the target material H p / H t , the dependent variable is the target erosion amount E. A total of 629 sets of data were collected, and the range of each variable is shown in Table 1. From the range of independent variables, it can be seen that the collected experimental data basically covers the common particle types, target types and engineering conditions (incident angle, incident speed) in engineering.

[0034] Table 1 Variable range

[0035]

[0036] S2, standardize variables and divide training and test data sets;

[0037] As can be seen from Table 1, due to the differences in magnitude between the ranges of variation of different independent variables, in order to avoid problems in the weight distribution of the network, a standardized method is adopted to convert the range of variation of each variable into between 0 and 1. The standardized method is as follows:

[0038]

[0039] Where: x norm is the standardized variable value, that is, the value actually used for network training; x actual is the actual experimental value collected; x min is the minimum experimental value; x max is the maximum experimental value.

[0040] Since the dependent variable (erosion amount) varies from 0 to 15851.3932 mm 3 / kg, and it can be seen from the average value that in the range of smaller erosion amount, there are more data and the data distribution is uneven, so the erosion amount is first taken logarithm and then standardized.

[0041] 80% of the data (503 groups) of the data set are used as training sets to train the network, and 20% of the data (126 groups) are used as test sets to test the prediction effect of the network. In order to achieve a better training effect of the neural network and make the data representative, the training set and the test set take into account the experimental data of different people. The erosion data are sorted from small to large, and then divided into training sets and test sets. The division method is to select 1 data as the test set data every 4 data from the sorted database, and select 126 groups of data in total, and the remaining data are used as the training set. This ensures that the range of variation of the dependent variables of the test set and the training set is basically consistent, while reflecting the randomness of the data sources of the training set and the test set.

[0042] S3, build a back propagation neural network structure;

[0043] The neural network structure constructed by the present invention is divided into five layers, and the number of neuron nodes in each layer is 7, 20, 15, 10, and 1 respectively, wherein 7 neuron nodes in the input layer correspond to 7 independent variables, and 1 neuron node in the output layer corresponds to the dependent variable (target material erosion amount). The neural network structure is as follows Figure 1 shown.

[0044] S4, network training and learning, and evaluating the prediction results with certain judgment criteria;

[0045] The evaluation criteria for the prediction results of the neural network are determined by the coefficient of determination R 2 , its formula is as follows: (R 2 The closer it is to 1, the better the prediction result. )

[0046]

[0047] Where: y i is the experimental value of erosion; p i is the predicted value of erosion; n is the number of data; i is the data label. Represents the average value of the experimental erosion amount of the data set.

[0048] S5. Network optimization;

[0049] First, the Adam optimization algorithm is used in the back-propagation process. The algorithm combines the advantages of two optimization algorithms: the ability of the AdaGrad algorithm to handle sparse gradients, and the ability of the RMSProp algorithm to handle non-stationary targets. This method is easy to implement and requires very little memory. It has high computational efficiency and can effectively speed up training and improve network learning efficiency. The optimization effect of this algorithm is as follows Figure 2 As shown, Figure 2 It can be seen from (a) and (b) that after using the Adam optimization algorithm, the convergence speed of the neural network is significantly improved compared to the ordinary gradient descent algorithm, and it can converge to a lower error level. Figure 2 In (b), the Adam optimization algorithm is also shown, which performs well in reducing the fluctuations caused by the loss decline. Figure 2 In (c), when the hidden layer activation function is relu, the loss function of the ordinary gradient descent algorithm stops decreasing when it drops to around 0.35, indicating that the loss function falls into a local minimum. However, after using the Adam optimization algorithm, the loss function can converge quickly and will not fall into a local minimum. It can be seen that the optimization effect of the Adam algorithm is obvious.

[0050] The training process of a neural network is essentially the continuous updating of the weight matrix and the bias matrix, so the network weight initialization method is crucial. In addition, gradient vanishing or gradient exploding problems, as well as overfitting problems, often occur during network training, all of which are related to the initialization method. The present invention compares the network prediction results using traditional random initialization, Xavier initialization, and He initialization methods, wherein the initialization method proposed by Xavier is used when the hidden layer activation function is sigmoid and tanh, and the initialization method proposed by He is used when the hidden layer activation function is relu. The optimization results are shown in Figure 2. Figure 3 As shown in the figure, after initialization optimization, the determination coefficient of the test set predicted by the neural network is improved, especially when the hidden layer activation function is relu, the determination coefficient of the test set is significantly improved, indicating that the initialization method proposed by He is effective in improving the prediction ability of the network when the activation function is relu.

[0051] The activation function adds nonlinear factors to the neural network, so that the neural network can adapt to more complex problems. Therefore, the role of the activation function in the neural network is also very important. Since the output variable erosion amount is standardized in the range of 0-1, the output layer activation function is fixed to sigmoid. The network prediction results when the hidden layer activation functions are sigmoid, tanh and relu are compared. In addition, the number of input variables is changed, and the neural network input variables are changed to 7. On the basis of the original six variables such as particle density, particle diameter, particle hardness, incident speed, incident angle and target hardness, the combined variable of particle and target hardness ratio is added. The comparison results are shown in Table 2. For any activation function, the test set results when the input variables are 7 are better than the results when the variables are 6. In addition, Table 2 also shows that when the input variables are 6, when the hidden layer activation function is tanh, the determination coefficient of the test set results is better than when the activation function is sigmoid and relu, and the priority order is: tanh>sigmoid>relu. When the input variables are 7 and the hidden layer activation function is relu, the determination coefficient of the test set is better than the other two activation functions, and the priority is: relu>tanh>sigmoid. It can be seen that different activation functions have a significant impact on the network prediction results. Finally, when the input variables are 7 and the hidden layer activation function is relu, the determination coefficient of the test set results reaches the highest, so it is selected as the optimal model (the learning rate is 0.03 and the number of iterations is 8000).

[0052] Table 2 Neural network evaluation results with different numbers of variables and activation functions

[0053]

[0054] The optimal model prediction effect of the present invention is as follows Figure 4 As shown, Figure 4 (a) is the comparison result of the training set. It can be seen that the network has been well trained and the predicted value of the neural network model proposed in the present invention is relatively close to the true value. Figure 4 (b) is the comparison result of the test set. It can be seen that the neural network model proposed in the present invention can better predict the erosion data of particles impacting the planar target.

[0055] S6, comparison of prediction results between network and formula;

[0056] The prediction results of the model of the present invention are compared with the existing formula, and the comparison formula is the formula proposed by Oka [1,2] The formula proposed by Huang [3] , Oka's formula is as follows:

[0057]

[0058] Where: α is the incident angle, unit is °; E(α) is the target erosion value after the collision of unit mass particles when the incident angle is α, unit is mm 3 / kg; E 90 The target erosion value after the collision of unit mass particles at an incident angle of 90°, in mm 3 / kg; H t is the target material Vickers hardness, in GPa. V is the particle incident velocity, in m / s; V' is the standard incident velocity, according to the literature [1,2] The value of the particle incident velocity in the experiment is in m / s; D is the particle diameter in μm; D' is the standard diameter, which is [1,2] The particle diameter in the experiment is in μm; n1 and n2 are determined by the target hardness and other collision conditions, such as particle characteristics, and are equal to s(HV) q , wherein s and q of the present invention are according to the existing literature [1,2] K, k1, k3: determined by the properties of the particles (types of particles), k1, k3 in the present invention are based on the existing literature [1,2][1,2] k2 is determined by the hardness of the target material and the properties of the particles. [1,2] The value is determined by the method in the formula; a and b are experimentally measured parameters, which are related to the stress relaxation ratio of the target material. The three parameters K, a, and b in the formula are obtained by fitting the formula, and the remaining parameters are obtained by using the literature. [1,2] The corresponding value in .

[0059] Huang's formula is as follows:

[0060] ΔQ≈Amρ 0.15 (Vsinα) 2.3 +Bm 1.1875 D -0.0625 V 2.375 (cosα) 2 (sinα) 0.375

[0061] Where: ΔQ is the target erosion amount after a single particle collision, unit: mm 3 The unit is the same as the unit of the erosion amount E in the present invention (mm 3 / kg) are different. In order to unify the units and make comparisons, a unit conversion is required: E = ΔQ / m. m is the mass of a single particle, in kg; ρ is the particle density, in kg / m 3 A and B are combination coefficients, which are determined by experimental data and obtained by fitting; other parameters are the same as those defined in the Oka formula.

[0062] The comparison results are as follows Figure 5As shown, the predicted value of the neural network model proposed in the present invention is closer to the true value. The prediction results of Oka formula and Huang's formula are acceptable when the erosion amount is relatively small. However, when the erosion amount increases, the predicted values ​​of Oka and Huang formulas deviate more from the true value than the neural network method of the present invention, and the predicted value is smaller than the actual erosion value, which is unconservative and easily causes safety accidents.

[0063] The embodiments disclosed above of the present invention can be easily implemented by professionals in this industry and related industries, and it is not difficult to modify many details in the embodiments. The methods defined herein can also be appropriately modified and used in other specific embodiments. Therefore, the protection scope of the present invention is not limited to the embodiments shown herein, but should be based on the claims and conform to the maximum scope consistent with the methods and features disclosed in the present invention.

Claims

1. A method for predicting the erosion amount of a target material due to particle collision based on a back propagation neural network, comprising the following steps: (1) Collect experimental data on the erosion of the target when particles collide with the plane. The experimental independent variables collected include 7, namely particle density ρ, particle diameter D, particle hardness H p , incident speed V, incident angle α, target hardness H t , the hardness ratio of the particle to the target material H p / H t , the dependent variable is the target erosion amount; (2) In order to eliminate the influence of magnitude differences between different variables on weight distribution, the collected experimental data are standardized, including standardization between different independent variables and standardization after taking the logarithm of the dependent variable, and the training and test data sets are divided; for the target erosion amount as the dependent variable, the logarithm is first taken and then standardized; the range of each variable is converted to between 0 and 1; (3) constructing a back-propagation neural network structure, wherein the neuron nodes of the input layer correspond to the experimental independent variables, and one neuron node of the output layer corresponds to the dependent variable, i.e., the target material erosion amount; (4) Determine the evaluation index of the neural network prediction results The coefficient of determination R 2 As an evaluation indicator of the neural network prediction results: Where: y i is the experimental value of erosion; p i is the predicted value of erosion; n is the number of data; i is the data label; represents the average value of the experimental values ​​of erosion in the data set; (5) Network model training optimization, using the Adam gradient descent algorithm, select the optimal network model based on the evaluation index of the neural network prediction results: the input layer is all the experimental independent variables collected in step (1); the hidden layer activation function is relu, and the initialization method is He initialization. The optimal network model is the target erosion model that is finally predicted to be caused by particle collision.

2. The prediction method according to claim 1, characterized in that: The back propagation neural network structure constructed in step (3) is divided into five layers, and the number of neuron nodes in each layer is 7, 20, 15, 10, and 1 respectively.

3. The prediction method according to claim 1, characterized in that: The method of step (5) is as follows: fix the output layer activation function as sigmoid, compare the prediction results of the network test set when the hidden layer activation functions are sigmoid, tanh and relu respectively; compare the network prediction effects when the number of input variables is different; compare the network prediction effects when different initialization methods are used, among which the Xavier initialization method is used when the hidden layer activation function is sigmoid or tanh, and the He initialization method is used when the hidden layer activation function is relu, to obtain the optimal network model optimized by the Adam gradient descent algorithm.

Images