A method for predicting penetration depth based on multi-source data fusion

Abstract

The present invention discloses a penetration depth prediction method based on multi-source data fusion, firstly, the evaluation intervals for evaluating engineering algorithms are divided; in each evaluation interval, several preferred engineering algorithms are selected for each different evaluation interval; batch sample calculations are performed using the selected engineering algorithms to obtain engineering calculation simulation data; then, a dimensionless penetration depth prediction model BP_exp based on a GA-BP neural network is established using experimental data; a dimensionless penetration depth prediction model BP_cand based on a GA_BP neural network is established using all engineering calculation simulation data; finally, an attention mechanism is used to weightedly fuse the outputs of the BP_exp and BP_cand models, and finally a dimensionless penetration depth prediction output is generated. The present invention ensures that the data has a certain accuracy, and at the same time, a large amount of engineering calculation simulation data is used to avoid the influence of the lack of experimental data on the deep learning model, and a prediction model that can exceed the accuracy of existing engineering algorithms is generated.

Description

Technical Field

[0001] The present invention belongs to the field of information processing technology, and in particular relates to a penetration depth prediction method based on multi-source data fusion. Background Art

[0002] The penetration mechanism of protective material media is revealed by empirical algorithms established by a large amount of test data. The penetration and destruction problem is a very complex physical process. The existing methods are difficult to accurately and comprehensively restore the actual situation. Therefore, in the actual process, engineering algorithms still occupy an important position. This type of engineering algorithm is relatively simple to use. Only a few parameters need to be given to obtain an estimated dimensionless penetration depth value. The existing engineering algorithm for penetration depth is a calculation formula generated by the inventor based on the experimental data he owns through fitting and other methods. Limited by the experimental environment, economic conditions, etc., the test data used can often only cover a part of the parameter range, and most of them are scaled tests, which results in each engineering algorithm being used only within a part of the parameter range and cannot meet the full parameter range of penetration depth analysis. At the same time, there is no clear measurement standard for the scope of application of the engineering algorithm, and it is not clear. Therefore, how to use the test data in the hands of each worker in combination with the existing engineering algorithm to establish a penetration depth prediction model, improve the accuracy of the prediction, and reduce the difficulties in selecting many engineering algorithms in actual engineering applications has become an important direction of penetration research. Summary of the invention

[0003] Purpose of the invention: In order to overcome the deficiencies in the prior art and applications, the present invention provides a penetration depth prediction method based on multi-source data fusion, which gives full play to the advantages of test data and the advantages of different engineering algorithms in each parameter interval, combines the two by data fusion, and improves the accuracy of penetration depth prediction.

[0004] Technical solution: The present invention provides a penetration depth prediction method based on multi-source data fusion, which specifically includes the following steps:

[0005] (1) Based on domain expert knowledge and experimental data clustering, the evaluation interval for engineering algorithm evaluation is divided;

[0006] (2) Select an engineering algorithm with better calculation accuracy for each evaluation interval. If there is no test data in the evaluation interval, select an engineering algorithm with better calculation accuracy based on the domain expert knowledge; if there is test data, perform performance analysis on the engineering algorithm based on the test data to select an engineering algorithm with better calculation accuracy;

[0007] (3) In each evaluation interval, the selected engineering algorithm is used to perform batch sample calculations to obtain engineering calculation simulation data;

[0008] (4) Using experimental data, a dimensionless penetration depth prediction model BP_exp based on GA_BP neural network is established. The network structure adopts BP neural network, and the genetic algorithm is used to optimize the hyperparameters of the network structure;

[0009] (5) Using engineering calculation simulation data, a dimensionless penetration depth prediction model BP_cand based on GA_BP neural network is established;

[0010] (6) The attention mechanism is used to perform weighted fusion on the outputs of the BP_exp and BP_cand models to generate the GA_Atten_BP_Fusion model, which is used to ultimately generate a dimensionless penetration depth prediction output.

[0011] Furthermore, when dividing the evaluation interval according to the domain expert knowledge in step (1), the expert determines the main interval division parameters and gives the segmentation of the parameters, and then the segmentation of different parameters is combined to form the evaluation interval; when dividing the evaluation interval according to the clustering of test data, the correlation coefficient between the feature and the dimensionless penetration depth is first calculated, and the top k feature quantities with the highest correlation coefficient are selected to cluster the test data. The expert evaluates or fine-tunes the clustered class range and finally determines the evaluation interval.

[0012] Furthermore, the implementation process of evaluating the algorithm according to the test data in step (2) is as follows:

[0013] For the evaluation interval R, extract the test data within the evaluation interval R to form the test sample set S R , assuming that R selects g engineering algorithms to be evaluated a1, a2, …, a g , using S R and a1, a2, …, a g Calculate the dimensionless penetration depth and compare it with the actual value to calculate the algorithm accuracy; use the mean absolute percentage error (MAPE) as the evaluation standard for the calculation accuracy of the engineering algorithm:

[0014]

[0015] Among them, y i is the true dimensionless penetration depth value, p i is the calculated dimensionless penetration depth value, n is the total number of samples in the evaluation interval;

[0016] Assume that MAPE1, MAPE2, …, MAPE g Respectively represent engineering algorithms a1, a2, ..., a g The mean absolute percentage error in the parameter interval analyzed is the MAPE i(i=1,…,g), sort them from low to high, and take the algorithms corresponding to the first k errors as the better algorithms for this parameter interval.

[0017] Furthermore, the implementation process of step (3) is as follows:

[0018] The parameters required for the engineering algorithm calculation are discretized within the evaluation interval, and then combined to form a large number of input vectors. Then, the engineering algorithm of the evaluation interval is used for calculation to obtain engineering calculation simulation data. Assume that the current algorithm requires m input parameters, and the discretized values ​​of each parameter are p1, p2, ..., p m Then all the combinations are That is, in the current evaluation interval, the batch calculation of an engineering algorithm can produce simulation data; for parameters with real values, by determining the maximum and minimum values ​​of the parameter and the step size of the change, all possible values ​​of the parameter can be discretized; for parameters of enumeration type, the enumerable values ​​determine the possible values ​​of the parameter.

[0019] Furthermore, the step (4) comprises the following steps:

[0020] (41) Construct a GA-BP neural network data set. The engineering calculation simulation data set is constructed in matrix form. The specific format is as follows:

[0021]

[0022] Among them, the last column is the model output, that is, the dimensionless penetration depth, and the other columns are the input of the model, and n is the number of data; the input matrix is ​​normalized, and each column of the original input matrix is ​​transformed into the interval [0,1];

[0023] (42) Determine the input feature vector of the GA_BP neural network model, the output is the dimensionless penetration depth, and the international units are uniformly adopted;

[0024] (43) Chromosome encoding and initial population generation of BP neural network:

[0025] According to formula (3), the number of hidden layers H is determined, and the number of nodes in each hidden layer and the hyperparameters Batch_size and Epoch are determined by genetic algorithm to ensure the optimal performance of the network:

[0026] H=log2 n (3)

[0027] Among them, H represents the number of hidden layers, n represents the number of nodes in the input layer, and the number of neurons contained in each hidden layer is determined by genetic algorithm;

[0028] There are H+2 parameters to be encoded in the GA_BP neural network, that is, there are H+2 populations, and each population contains r individuals;

[0029] After determining the chromosome composition, the initial population is encoded, and it starts to iterate and evolve continuously until the preset number of iterations t is reached, the algorithm stops, and the optimal solution in the t-generation population is output; the initialization method is to randomly encode the initial population, that is, each gene position of the r individuals in the initial population is randomly composed of 0 or 1;

[0030] (44) Design of fitness function of GA_BP neural network:

[0031] The fitness f of the i-th individual i Defined as the error err of the neural network solution corresponding to the individual on the test set i Difference from 1:

[0032] f i =1-err i (4)

[0033] (45) Genetic operation of GA-BP neural network:

[0034] Adopt adaptive genetic algorithm to carry out genetic operation: The mutation probability and crossover probability of each individual are dynamically determined by the fitness of the individual in each round; when the fitness of the individual is lower, the probability of crossover and mutation of the individual needs to be increased to update and produce offspring; when the fitness of the individual is higher, the probability of crossover and mutation of the individual needs to be reduced to retain it.

[0035] Furthermore, the implementation process of step (6) is as follows:

[0036] (61) The input InputFeatures of GA_Atten_BP_Fusion is a 2D feature vector, which is formed by the vector concatenation of the dimensionless penetration depth output by the BP_exp neural network model and the BP_cand neural network model;

[0037] (62) The attention mechanism includes a fully connected layer and a Softmax layer. The corresponding attention weight distribution is obtained after the fully connected layer. The formula of the Softmax layer is as follows:

[0038] Attn(InputFeatures)=Softmax(MLP(InputFeatures) (5)

[0039] Among them, MLP represents the multi-layer perceptron of the fully connected layer; after the Softmax normalization operation, the attention weight of the sample satisfies:

[0040]

[0041] Where n represents the dimension of the input vector, and Attn i ∈[0,1];

[0042] The dimensionless penetration depth fusion value output by weighted fusion is as follows:

[0043]

[0044] Among them, pred BP_exp 、pred BP_cand are the predicted values ​​of the BP_exp neural network model and the BP_cand neural network model respectively, and Attn1 and Attn2 are their corresponding attention weights respectively.

[0045] Furthermore, the step (45) comprises the following steps:

[0046] (451) The proportional method is used as the selection operator, that is, the probability of each individual in the population being selected to reproduce offspring is equivalent to the ratio of its fitness function value. The probability of the i-th individual in the population being selected is P i , the higher the individual's fitness, the higher the probability of being selected:

[0047]

[0048] (452) The crossover operator simulates the process of chromosome exchange during embryonic development. Single-point crossover is used as the crossover operator. Only one position is crossed each time. The crossover probability Pci of the i-th individual is:

[0049]

[0050] Among them, f max is the highest fitness in the group, f min is the lowest fitness in the group, k1 is the crossover probability when the individual fitness is the lowest, and k2 is the crossover probability when the individual fitness is the highest. Ensure k1>k2, f i is the fitness of the i-th individual;

[0051] For the i-th individual, firstly, the crossover position is randomly determined, and then the crossover is paired with the parent individual according to the probability Pci. The parent and the child individuals exchange some information to generate two new offspring individuals, and the first offspring individual is retained as the new individual in the offspring population. When the length of the binary code string of the individual is l, there are l-1 optional crossover positions between the parent individuals. At this time, the single-point crossover method can generate l-1 different new individuals.

[0052] (453) For the i-th individual, according to the mutation probability Pm iTo mutate, randomly select the mutation position of the individual gene. If the position mutates, the gene will mutate to 1 if it is 0, and vice versa. The mutation probability Pmi of the i-th individual is as follows:

[0053]

[0054] Among them, f max is the highest fitness in the group, f min is the lowest fitness in the group, k3 is the mutation probability when the individual fitness is the lowest, k4 is the mutation probability when the individual fitness is the highest, ensuring k3>k3, f i is the fitness of the ith individual.

[0055] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are: relative to a single engineering algorithm, the present invention can achieve accurate prediction in a wider parameter range; compared with the existing machine learning models and deep learning models for in-depth prediction based only on experimental data, the present invention can adopt a large amount of engineering calculation simulation data with higher credibility to improve the accuracy of the deep learning model; compared with some prediction methods that fuse the output of engineering algorithms, the present invention can avoid large errors in the output of certain engineering algorithms in different parameter regions, thereby affecting the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the present invention;

[0057] Figure 2 MAPE diagram for different experimental models on the test set. DETAILED DESCRIPTION

[0058] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0059] like Figure 1As shown, the present invention discloses a penetration depth prediction method based on multi-source data fusion, which divides the evaluation interval for evaluating engineering algorithms in combination with domain expert knowledge; in each evaluation area, select k engineering algorithms with higher calculation accuracy in this interval; in each evaluation interval, use the selected engineering algorithm to perform batch sample calculations to obtain engineering calculation simulation data; use experimental data to establish a dimensionless penetration depth prediction model BP_exp based on the GA-BP neural network; use all engineering calculation simulation data to establish a dimensionless penetration depth prediction model BP_cand based on the GA_BP neural network; use the attention mechanism to fuse the outputs of BP_exp and BP_cand, and establish the final dimensionless penetration depth prediction output. The present invention uses different engineering algorithms in different parameter intervals to generate simulation data, which not only ensures the accuracy of the data, but also avoids the influence of the lack of experimental data on the deep learning model through a large amount of engineering algorithm simulation data, and generates a prediction model that can exceed the accuracy of existing engineering algorithms.

[0060] Step 1: Divide the engineering algorithm evaluation interval based on domain expert knowledge or test data clustering. The predicted output is dimensionless penetration depth, and the input feature quantities mainly include many feature quantities such as projectile mass, projectile velocity, target impact angle, target density, etc., and different engineering algorithms use different feature quantities. The division of the evaluation interval is generally based on these important feature quantities.

[0061] When dividing the evaluation interval according to the domain expert knowledge, the experts give the main interval division features and give the feature segmentation, and then combine the different feature segments to form the evaluation interval. For example, first select the division feature quantity and divide the feature quantity into intervals. For example, the speed is divided into low speed [0,340), medium speed [340,650), and high speed [650,) in units of m / s; the mass of the projectile is divided into low mass [0,50), medium mass [50,500) and high mass [500,) in units of kilograms. The intervals of these feature quantities are combined to divide the evaluation space into different areas. For example, according to the speed and mass of the projectile, different evaluation intervals such as "low speed and low mass", "medium speed and medium mass" and "high speed and high mass" are generated.

[0062] When dividing the evaluation interval based on the clustering of test data, the correlation coefficient between the input feature and the dimensionless penetration depth is first calculated, and the first k features with the highest correlation coefficient are selected to cluster the test data. The clustering results are evaluated and fine-tuned by experts, and each clustered range finally corresponds to an evaluation interval. For example, when analyzing the penetration depth of concrete, the first two features with the highest correlation coefficients are the target velocity and the projectile mass, and these two features are then used to cluster the test data.

[0063] The evaluation interval combination must be complete and cover all areas of the selected feature quantity combination.

[0064] Step 2: Evaluate the engineering algorithm and select the engineering algorithm. In each evaluation interval, select k engineering algorithms with high calculation accuracy in the evaluation interval. The selection method includes two steps:

[0065] (1) Based on the existing application practice data, the applicable ranges of various engineering algorithms are analyzed. For example, the Young formula for the concrete penetration depth algorithm generally works better in the medium speed range.

[0066] (2) Under the premise that (1) cannot obtain sufficient information, use experimental data to analyze the algorithm performance. For the evaluation interval R, extract the experimental data within the evaluation interval R to form the experimental sample set S R , assuming that R selects g engineering algorithms to be evaluated a1, a2, …, a g , using S R and a1, a2, …, a g Calculate the dimensionless penetration depth and compare it with the actual value to calculate the algorithm accuracy. The mean absolute percentage error (MAPE) is used as the evaluation standard for the calculation accuracy of the engineering algorithm:

[0067]

[0068] Among them, y i is the true dimensionless penetration depth value, p i is the calculated dimensionless penetration depth value, and m is the total number of samples in the evaluation interval.

[0069] Assume that MAPE1, MAPE2, …, MAPE g Respectively represent engineering algorithms a1, a2, ..., a g The mean absolute percentage error in the parameter interval analyzed is the MAPE i (i=1,…,g), sort them from low to high, and take the algorithms corresponding to the first k errors as the better algorithms for this parameter interval.

[0070] Step 3: In each evaluation interval, use the selected engineering algorithm to perform batch sample calculations to obtain engineering calculation simulation data.

[0071] In each evaluation interval, batch engineering calculations are performed to obtain engineering calculation simulation data. The parameters required for the engineering algorithm calculation are discretized within the evaluation interval, and then combined to form a large number of input vectors. Then, the engineering algorithm of the evaluation interval is used for calculation to obtain engineering calculation simulation data. Assume that the current algorithm requires m parameters, and the discretized values ​​of each parameter are p1, p2, ..., p mThen all the combinations are That is, in the current evaluation interval, the batch calculation of an engineering algorithm can produce There are two types of simulation data.

[0072] 1) Real number value range. For parameters with real number value range, determine the maximum and minimum values ​​of the parameter, as well as the step size, and then automatically generate all possible values ​​of the parameter. For example, in the medium speed range [340,650), the minimum value is 340 and the maximum value is 650. Assuming the step size is set to 10, 31 values ​​can be generated.

[0073] 2) Enumeration value. For example, the bullet shape includes flat-head bullet, oval-shaped bullet, pointed-head bullet, etc. The value is set as an enumeration type, including three values: 0.72, 0.8, and 1.14.

[0074] The engineering calculation simulation data is obtained through batch engineering calculation. Since the engineering algorithm is also generated by fitting a large amount of test data, the simulation data obtained by using the engineering algorithm for calculation in the interval with high accuracy of the engineering algorithm has high accuracy. The engineering simulation data obtained by the engineering algorithm calculation in multiple intervals are mixed as the data set for training the model.

[0075] Step 4: Using the test data, establish the dimensionless penetration depth prediction model BP_exp based on GA_BP neural network. The network structure adopts BP neural network, and the genetic algorithm is used to optimize the hyperparameters of the network structure. It mainly includes the following steps:

[0076] (1) Constructing the GA-BP neural network data set. The engineering calculation simulation data set is constructed in the form of a matrix. The specific format is as follows:

[0077]

[0078] Among them, the last column is the model output, that is, the dimensionless penetration depth, and the other columns are the input of the model, and n is the number of data; the dimensions of different features are quite different, so it is necessary to normalize the input matrix. The present invention uses the min-max method to transform each column of the original input matrix into the interval [0,1]. The conversion method is as follows:

[0079]

[0080] in, is the data after min-max normalization, x is the original value without transformation, x min 、x max are their minimum and maximum values ​​respectively. The data set is divided into training set and test set in a ratio of 8:2.

[0081] (2) Input feature vector of GA_BP neural network model. The input includes eight feature quantities, including target impact velocity, projectile mass, projectile diameter, target compressive strength, target material density, warhead CRH, warhead shape factor, and warhead length. The output is dimensionless penetration depth, and international units are used uniformly.

[0082] (3) Chromosome encoding and initial population generation of BP neural network.

[0083] Increasing the number of hidden layers in the BP neural network can help improve the accuracy of the network, but it may lead to adverse effects such as overfitting. The number of hidden layers H is determined according to formula (4), and the number of nodes in each hidden layer and the hyperparameters Batch_size and Epoch are determined by the genetic algorithm to ensure the optimal performance of the network.

[0084] H=log2 n (4)

[0085] Wherein, H represents the number of hidden layers, and n represents the number of nodes in the input layer, which in the present invention is 8. The number of neurons contained in each hidden layer is determined by a genetic algorithm.

[0086] There are H+2 parameters to be encoded in the GA_BP neural network, that is, there are H+2 populations, and each population contains r individuals (chromosomes).

[0087] After determining the chromosome composition, the initial population is encoded, and it starts to iterate and evolve continuously until the preset number of iterations t is reached, the algorithm stops, and the optimal solution in the t-generation population is output. The initialization method is to randomly encode the initial population, that is, each gene position of the r individuals in the initial population is randomly composed of 0 or 1.

[0088] (4) Design of fitness function of GA_BP neural network.

[0089] The smaller the error of the BP neural network model on the test set, the better the performance of the model and the higher the fitness of the problem to be solved. Therefore, the fitness of the i-th individual f i Defined as the error err of the neural network solution corresponding to the individual on the test set i The difference from 1 is as follows:

[0090] f i =1-err i (5)

[0091] (5) Genetic Operation of GA-BP Neural Network

[0092] The genetic operation of the genetic algorithm includes three operators: selection, crossover and mutation. The present invention adopts an adaptive genetic algorithm to perform genetic operations. The mutation probability and crossover probability of each individual are dynamically determined by the fitness of each round of individuals. The overall idea is: when the fitness of an individual is lower, the individual needs to increase the probability of crossover and mutation, and update to produce offspring; when the fitness of an individual is higher, the individual needs to reduce the probability of crossover and mutation and be retained.

[0093] 1) The selection operator selects individuals with good performance from the parent population by simulating the natural selection phenomenon in nature. The greater the fitness of an individual, the higher the probability that it will reproduce offspring individuals during the selection process. The present invention adopts the most common proportional method (roulette method) as the selection operator, that is, the probability of each individual in the population being selected to reproduce offspring is equivalent to the ratio of its fitness function value. The probability P of the i-th individual in the population being selected is i It can be expressed as the following formula: the higher the individual's fitness, the higher the probability of being selected.

[0094]

[0095] 2) The crossover operator simulates the process of chromosome exchange during embryonic development. The present invention uses single-point crossover as the crossover operator, and only crosses one position at a time. The crossover probability Pci of the i-th individual is as follows:

[0096]

[0097] f max is the highest fitness in the group, f min is the lowest fitness in the group, k1 is the crossover probability when the individual fitness is the lowest, and k2 is the crossover probability when the individual fitness is the highest, ensuring k1>k2. i is the fitness of the ith individual.

[0098] For the i-th individual, the crossover position is first randomly determined, and then it is cross-paired with the parent individual according to the probability Pci. The parent and child individuals exchange some information to generate two new offspring individuals, and the first offspring individual is retained as a new individual in the offspring population. When the length of the binary code string of the individual is l, there are l-1 optional crossover positions between the parent individuals. At this time, the single-point crossover method can generate l-1 different new individuals.

[0099] 3) The mutation operator simulates the gene mutation phenomenon that may occur during embryonic development. For the i-th individual, according to the mutation probability Pm i The mutation position of the individual gene is randomly selected. For the binary encoding method adopted by the present invention, if the position mutates, the gene mutates to 1 if it is 0, and vice versa. The mutation probability Pmi of the i-th individual is as follows:

[0100]

[0101] f max is the highest fitness in the group, f min is the lowest fitness in the group, k3 is the mutation probability when the individual fitness is the lowest, and k4 is the mutation probability when the individual fitness is the highest, ensuring k3>k3. i is the fitness of the ith individual.

[0102] In genetic operations, the main hyperparameters include: population size r, maximum number of iterations t. In actual use, the larger the population size r and the maximum number of iterations t, the closer the potential solution searched by the genetic algorithm is theoretically to the optimal solution. However, the larger the population size and the maximum number of iterations are, the more time-consuming it is for the genetic algorithm to search the solution space. Therefore, it is necessary to find a balance point.

[0103] Step 5: Based on the engineering calculation simulation data and the GA_BP neural network model, the BP_cand model is trained.

[0104] Step 6: Using the experimental data, the attention mechanism is used to establish a weight-based fusion of the outputs of the BP_cand model and the BP_exp model, thereby establishing the Attn_GA_BP_Fusion fusion model, and the output of this model is used as the final output.

[0105] The process of dimensionless penetration depth prediction is as follows: the input is generated through BP_cand and BP_exp respectively, and then fused through the GA_Attn_BP_Fusion model to generate the final dimensionless penetration depth value.

[0106] (1) Input features InputFeatures is a two-dimensional feature vector, which is formed by vector concatenation of the dimensionless penetration depths output by the BP_exp neural network model and the BP_cand neural network model.

[0107] (2) The attention mechanism includes a fully connected layer and a Softmax layer. The corresponding attention weight distribution is obtained after the fully connected layer. The Softmax layer can make the weights of relatively important inputs more prominent while performing normalization. The formula is as follows, where MLP represents the multi-layer perceptron of the fully connected layer.

[0108] Attn(InputFeatures)=Softmax(MLP(InputFeatures) (9)

[0109] After the Softmax normalization operation, the attention weight of the sample satisfies:

[0110]

[0111] Among them, n represents the dimension of InputFeatures, and Attn i ∈[0,1], which meets the requirements for weights in weighted data fusion theory. Therefore, the attention weights corresponding to different data sources are used as weights in the weighted data fusion process to fuse the dimensionless penetration depth decisions of different networks.

[0112] (3) After weighted fusion, the dimensionless penetration depth fusion value of the measured data is output, which can be expressed as:

[0113]

[0114] Among them, pred BP_exp 、pred BP_cand are the predicted values ​​of the BP_exp neural network model and the BP_cand neural network model respectively, and Attn1 and Attn2 are their corresponding attention weights respectively.

[0115] In order to verify the performance of the penetration depth prediction method based on multi-source data fusion, an experiment was conducted using concrete penetration depth prediction as an example to compare the present invention with existing algorithms. The algorithms involved in the comparison include: Simple Average (SA), BP neural network algorithm, and existing empirical algorithms. Among them, the simple average algorithm directly averages the dimensionless penetration depth decisions output by the BP_exp and BP_cand neural network models as the fusion value of the dimensionless penetration depth; the fusion method based on the BP neural network algorithm uses the BP neural network to solve the weights to fuse the dimensionless penetration depth decisions output by the BP_exp and BP_cand neural network models.

[0116] For the convenience of comparison, the experimental data set is divided into a training set and a test set in a ratio of 8:2. The engineering calculation simulation data set is used as the training set of the BP_cand model. The performance difference between the method proposed in the present invention and the existing method is compared on the test set, and the mean absolute percentage error MAPE is used as the evaluation index.

[0117] First, the evaluation interval is divided. According to the speed and quality, the concrete penetration depth engineering algorithm is divided into four evaluation intervals, as shown in Table 1.

[0118] Table 1 Concrete penetration depth engineering algorithm evaluation interval table

[0119]

[0120] According to the existing algorithm documents and test data, engineering algorithms are selected in four intervals, and four engineering algorithms are selected in each interval, as shown in Table 2.

[0121] Table 2 Candidate list of engineering algorithms for each evaluation interval

[0122]

[0123] In each interval, batch engineering calculations are performed, and each parameter is discretized according to Table 3 to generate combined samples. The candidate engineering algorithms in each evaluation interval are used to generate engineering calculation simulation data.

[0124] Table 3 Parameters for generating pseudo data sources using empirical algorithms

[0125]

[0126] In the process of training the GA-BP neural network, the number of individuals in each genetic population is set to r = 30, the maximum number of iterations is t = 10, the crossover probability is k1 = 0.9, k2 = 0.1; the mutation probability is k3 = 0.1, k4 = 0.01, and the solution space is defined as shown in Table 4:

[0127] Table 4 GA-BP solution space definition table

[0128]

[0129] After 10 rounds of iterations, the GA-BP parameters are shown in Table 5.

[0130] Table 5 GA-BP parameters and errors after 5 rounds of iterations

[0131]

[0132] The MAPE error of each model on the test set is as follows: Figure 2 As shown in the figure, the dotted line indicates the lowest MAPE error in the traditional empirical algorithm. It can be seen from the figure that the MAPE errors of the models based on BP neural network (BP-exp, BP-cand) and data fusion methods (SA, BPF, GA-Attn-BP-Fusion) on the test set are generally better than those of the models without data fusion methods. The MAPE error of the BP-exp model based on the experimental data source is slightly lower than that of the BP-cand model based on the pseudo data source of the empirical algorithm; among the data fusion methods, the SA model based on the simple average of the prediction results of BP-exp and BP-cand has the highest error, and the weighted fusion model based on the genetic algorithm and attention mechanism has the lowest MAPE error.

Claims

1. A penetration depth prediction method based on multi-source data fusion, characterized in that: The steps include: (1) Based on domain expert knowledge and experimental data clustering, the evaluation interval for engineering algorithm evaluation is divided; (2) Select an engineering algorithm with better calculation accuracy for each evaluation interval. If there is no test data for the evaluation interval, select an engineering algorithm with better calculation accuracy based on the knowledge of domain experts. If there is test data, perform performance analysis on the engineering algorithm based on the test data and select an engineering algorithm with better calculation accuracy. (3) In each evaluation interval, the selected engineering algorithm is used to perform batch sample calculations to obtain engineering calculation simulation data; (4) Using experimental data, a dimensionless penetration depth prediction model BP_exp based on GA_BP neural network is established. The network structure adopts BP neural network, and the genetic algorithm is used to optimize the hyperparameters of the network structure; (5) Using engineering calculation simulation data, a dimensionless penetration depth prediction model BP_cand based on GA_BP neural network is established; (6) The attention mechanism is used to perform weighted fusion on the outputs of the BP_exp and BP_cand models to generate the GA_Atten_BP_Fusion model, which is then used to perform dimensionless penetration depth prediction. The step (4) comprises the following steps: (41) Construct a GA-BP neural network data set. The engineering calculation simulation data set is constructed in matrix form. The specific format is as follows: Among them, the last column is the model output, that is, the dimensionless penetration depth, and the other columns are the input of the model, and n is the number of data; the input matrix is ​​normalized, and each column of the original input matrix is ​​transformed into the interval [0,1]; (42) Determine the input feature vector of the GA_BP neural network model, the output is the dimensionless penetration depth, and the international units are uniformly adopted; (43) Chromosome encoding and initial population generation of BP neural network: According to formula (3), the number of hidden layers H is determined, and the number of nodes in each hidden layer and the hyperparameters Batch_size and Epoch are determined by genetic algorithm to ensure the optimal performance of the network: H=log2 m (3) Among them, H represents the number of hidden layers, m represents the number of nodes in the input layer, and the number of neurons contained in each hidden layer is determined by genetic algorithm; There are H+2 parameters to be encoded in the GA_BP neural network, that is, there are H+2 populations, and each population contains r individuals; After determining the chromosome composition, the initial population is encoded, and it starts to iterate and evolve continuously until the preset number of iterations t is reached, the algorithm stops, and the optimal solution in the t-generation population is output; the initialization method is to randomly encode the initial population, that is, each gene position of the r individuals in the initial population is randomly composed of 0 or 1; (44) Design of fitness function of GA_BP neural network: The fitness f of the i-th individual i Defined as the error err of the neural network solution corresponding to the individual on the test set i Difference from 1: f i =1-err i (4) (45) Genetic operation of GA-BP neural network: Adopt adaptive genetic algorithm to carry out genetic operation: The mutation probability and crossover probability of each individual are dynamically determined by the fitness of the individual in each round; when the fitness of the individual is lower, the probability of crossover and mutation of the individual needs to be increased to update and produce offspring; when the fitness of the individual is higher, the probability of crossover and mutation of the individual needs to be reduced to retain it.

2. The penetration depth prediction method based on multi-source data fusion according to claim 1 is characterized in that: When dividing the evaluation interval according to the domain expert knowledge in step (1), the expert determines the main interval division parameters and gives the segmentation of the parameters, and then combines the segmentation of different parameters to form the evaluation interval; when dividing the evaluation interval according to the clustering of test data, first calculate the correlation coefficient between the feature and the dimensionless penetration depth, select the top k feature quantities with the highest correlation coefficient to cluster the test data, and the expert evaluates or fine-tunes the clustered class range to finally determine the evaluation interval.

3. The penetration depth prediction method based on multi-source data fusion according to claim 1 is characterized in that: The implementation process of the algorithm evaluation based on the test data in step (2) is as follows: For the evaluation interval R, extract the test data within the evaluation interval R to form the test sample set S R , assuming that R selects g engineering algorithms to be evaluated a1, a2, …, a g , using S R and a1, a2, …, a g Calculate the dimensionless penetration depth and compare it with the actual value to calculate the algorithm accuracy; use the mean absolute percentage error (MAPE) as the evaluation standard for the calculation accuracy of the engineering algorithm: Among them, y i is the true dimensionless penetration depth value, p i is the calculated dimensionless penetration depth value, n is the total number of samples in the evaluation interval; Assume that MAPE1, MAPE2, …, MAPE g Respectively represent engineering algorithms a1, a2, ..., a g The mean absolute percentage error in the parameter interval analyzed is the MAPE i Sort from low to high and take the algorithms corresponding to the first k errors as the better algorithms for this parameter interval, i = 1,…,g.

4. The penetration depth prediction method based on multi-source data fusion according to claim 1 is characterized in that: The implementation process of step (3) is as follows: The parameters required for the engineering algorithm calculation are discretized within the evaluation interval, and then combined to form a large number of input vectors. Then, the engineering algorithm of the evaluation interval is used for calculation to obtain engineering calculation simulation data. Assume that the current algorithm requires m input parameters, and the discretized values ​​of each parameter are p1, p2, ..., p m Then all the combinations are That is, in the current evaluation interval, the batch calculation of an engineering algorithm can produce simulation data; for parameters with real values, by determining the maximum and minimum values ​​of the parameter and the step size of the change, all possible values ​​of the parameter can be discretized; for parameters of enumeration type, the enumerable values ​​determine the possible values ​​of the parameter.

5. The penetration depth prediction method based on multi-source data fusion according to claim 1 is characterized in that: The implementation process of step (6) is as follows: (61) The input features of GA_Atten_BP_Fusion are 2D feature vectors, which are the dimensionless penetration depths output by the BP_exp neural network model and the BP_cand neural network model. (62) The attention mechanism includes a fully connected layer and a Softmax layer. The corresponding attention weight distribution is obtained after the fully connected layer. The formula of the Softmax layer is as follows: Attn(InputFeatures)=Softmax(MLP(InputFeatures) (5) Among them, MLP represents the multi-layer perceptron of the fully connected layer; after the Softmax normalization operation, the attention weight of the sample satisfies: Where n represents the dimension of the input vector, and Attn i ∈[0,1]; The dimensionless penetration depth fusion value output by weighted fusion is as follows: Among them, pred BP_exp 、pred BP_cand are the predicted values ​​of the BP_exp neural network model and the BP_cand neural network model respectively, and Attn1 and Attn2 are their corresponding attention weights respectively.

6. The penetration depth prediction method based on multi-source data fusion according to claim 1 is characterized in that: The step (45) comprises the following steps: (451) The proportional method is used as the selection operator, that is, the probability of each individual in the population being selected to reproduce offspring is equivalent to the ratio of its fitness function value. The probability of the i-th individual in the population being selected is P i , the higher the individual's fitness, the higher the probability of being selected: (452) The crossover operator simulates the process of chromosome exchange during embryonic development. Single-point crossover is used as the crossover operator. Only one position is crossed at a time. The crossover probability P of the i-th individual is ci : Among them, f max is the highest fitness in the group, f min is the lowest fitness in the group, k1 is the crossover probability when the individual fitness is the lowest, and k2 is the crossover probability when the individual fitness is the highest. Ensure k1>k2, f i is the fitness of the ith individual; For the i-th individual, firstly, the crossover position is randomly determined, and then the crossover is paired with the parent individual according to the probability Pci. The parent and the child individuals exchange some information to generate two new offspring individuals, and the first offspring individual is retained as the new individual in the offspring population. When the length of the binary code string of the individual is l, there are l-1 optional crossover positions between the parent individuals. At this time, the single-point crossover method can generate l-1 different new individuals. (453) For the i-th individual, according to the mutation probability P mi To mutate, randomly select the mutation position of the individual gene. If the position mutates, the gene will mutate to 1 if it is 0, and vice versa. The mutation probability P of the i-th individual mi as follows: Among them, f max is the highest fitness in the group, f min is the lowest fitness in the group, k3 is the mutation probability when the individual fitness is the lowest, and k4 is the mutation probability when the individual fitness is the highest. Ensure k3>k4, f i is the fitness of the ith individual.

Images