A small sample data under ship section manufacturing time prediction method

Abstract

A kind of ship section manufacturing time prediction method under small sample data, in order to solve the problem that the accuracy of ship section manufacturing time prediction under small sample data is difficult to guarantee, the section with a large amount of historical data is regarded as source field in the present application, and the section under small sample data set is regarded as target field, a prediction model of source field section manufacturing time is constructed by machine learning technology, the network weight and bias value thereof are extracted and migrated to the prediction model of target field section, the similarity and difference characteristics of source field and target field section are analyzed, according to the analysis result, the model parameters of similarity characteristics are frozen and the model parameters of difference characteristics are fine-tuned, and the construction of target field section manufacturing time prediction model is completed.The time prediction method of the present application not only effectively solves the problem of ship section manufacturing time prediction under small sample data, but also shows significant advantages in precision and stability.The present application belongs to the technical field of time prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of time prediction, and particularly relates to a ship block manufacturing time prediction method under small sample data. BACKGROUND

[0002] Under the modern shipbuilding mode, the ship block as an important intermediate product, its manufacturing time prediction is very important for the resource allocation, material procurement and overall production planning of the shipyard, directly affects the efficiency and cost control of shipbuilding, and determines the progress and stability of ship delivery. However, since shipbuilding is usually based on orders, each order requires high personalization, and different ship blocks have significant differences, and the manufacturing workload is heavy, which makes the prediction of ship block manufacturing time particularly difficult, especially in the case of lack of sufficient historical data for some ship types, the accuracy of prediction is more difficult to guarantee. Therefore, under the condition of small sample data, it is necessary to improve the accuracy of ship block manufacturing time prediction, which has become one of the key problems in the current shipbuilding industry.

[0003] Prior art document 1 "A knowledge discovery and reuse method for time estimation in ship block manufacturing planning using DEA", Li et al., Advanced Engineering Informatics, 2019, 39:25-40, proposes a knowledge discovery and reuse method for ship block manufacturing time prediction, however, this method does not consider the case of small sample data set. Technical document 2 "Gear contact fatigue life prediction based on transfer learning", Li et al., International Journal of Fatigue, 2023, 173:12, proposes a gear contact fatigue life prediction method based on transfer learning and machine learning, however, this method does not consider the similarity and difference characteristics of source domain data set and target domain data set. Therefore, it is still necessary to further study the method suitable for ship block manufacturing time prediction under small sample data. SUMMARY

[0004] The present application is to solve the problem of small sample data, ship block manufacturing time prediction difficulty and accuracy difficult to guarantee, and further proposes a ship block manufacturing time prediction method under small sample data.

[0005] The technical scheme of the present application is a ship section manufacturing time prediction method under small sample data, which constructs a source field section manufacturing time prediction model through machine learning technology, and extracts its network weight and bias value. The weight and bias value of the source field section manufacturing time is migrated to the prediction model of the target field section, the similarity and difference characteristics of the source field and the target field section are analyzed, according to the analysis result, the network parameters of the similarity characteristics are frozen to prevent their change in subsequent training, and the network parameters of the difference characteristics are fine-tuned to adapt to the characteristics of the target field data. On the basis of the above, the construction of the target field section manufacturing time prediction model is completed, so that it has the ability to accurately predict the target field section manufacturing time under small sample data. The specific steps are as follows:

[0006] Step one, source field section time prediction model construction based on machine learning;

[0007] Firstly, the methodology selection of the model is determined. The neural network model based on particle swarm algorithm (PSOBPNN) is selected for section manufacturing time prediction, each particle represents a set of parameters of the neural network model, that is, the weight and bias value of the model, the dimension of the particle is determined by the structure of the model, which depends on the number of input layer nodes, the number of nodes of each hidden layer and the number of output layer nodes. The particle dimension calculation formula is:

[0008]

[0009] Where, Dim is the particle dimension, i represents the index value of the hidden layer, L represents the index value of the maximum number of hidden layers, n in is the number of input layer nodes, is the number of nodes of the i-th hidden layer, n out is the number of output layer nodes.

[0010] Secondly, the input and output variables of the model are determined. The section weight, projected area, small group number, weld length and steel plate thickness are selected as the input characteristics of PSOBPNN, and the section manufacturing time is selected as the output. In order to improve the training efficiency and stability of the model, all input and output data are processed by decimal scaling normalization. The decimal scaling normalization formula is:

[0011]

[0012] k=log 10 (max(x)) (3)

[0013] Where, x' is the normalized data value, x is the data value before normalization, k is the logarithmic function value corresponding to the maximum value of x with base 10.

[0014] Again, the prediction accuracy index is determined. The present application adopts the mean absolute percentage error (MAPE) as the relevant precision index, and the MAPE value can represent the average error degree of the prediction.

[0015]

[0016] wherein j represents the index value of the sample, y j represents the actual operation time of the jth sample, represents the predicted operation time of the jth sample, and n represents the maximum index value of the sample.

[0017] Next, the hyperparameter configuration of the neural network model is determined. The hyperparameters of the neural network model include the number of hidden layers, the number of hidden layer nodes, the hidden layer activation function, the output layer activation function, the loss function, the number of training times, and the training target error value. By conducting experiments under different combinations of hyperparameters, the influence of each set of hyperparameters on the prediction performance of the model is evaluated, and the optimal hyperparameter configuration is finally selected.

[0018] Finally, the hyperparameter configuration of the particle swarm algorithm is determined. The hyperparameters of the particle swarm algorithm include individual and population learning factors, maximum iteration times, particle number, minimum inertia weight value, maximum inertia weight value, particle velocity range, and particle position range. By systematically adjusting and testing these hyperparameters, the influence of the hyperparameters on the prediction results of the model is analyzed, and the optimal hyperparameter configuration is finally selected.

[0019] Based on the above steps, the source domain segmented manufacturing time prediction model is constructed and trained, and the network weight and bias value are extracted. The model will serve as the basis for transfer learning and further application to the target domain segmented manufacturing time prediction.

[0020] Step two, target domain segmented time prediction model construction based on transfer learning and machine learning;

[0021] First, the transfer strategy of the weight and bias value is established. The weight and bias value of the source domain segmented manufacturing time prediction model is transferred to the target domain segmented time prediction model (TL-PSOBPNN). By analyzing the similarity and difference characteristics of the source domain segmentation and the target domain segmentation, the following strategy is determined: for the weight of the input layer to the hidden layer between the source domain and the target domain, take the freezing action; for other weights and bias values, take the fine-tuning action.

[0022] Second, the target domain segmented time prediction model is trained. The TL-PSOBPNN is trained using the target domain segmented data, and the weights and biases to be fine-tuned in the model are optimized to ensure that the model adapts to the characteristics of the target domain segmentation, thereby forming a TL-PSOBPNN suitable for target domain segmented time prediction.

[0023] Finally, the method validity verification. By selecting the actual production case, using the mean absolute percentage error (MAPE) as the precision index, comparing the prediction effect of the migration learning model, thereby verifying the effectiveness of the application.

[0024] The beneficial effects of the application are:

[0025] 1. The application constructs a source field segmented manufacturing time prediction model through machine learning technology, mines the correlation knowledge between the source field segmentation features and the manufacturing time, and applies these knowledge to the target field segmentation through transfer learning technology. The time prediction method of the application not only effectively solves the problem of ship segment manufacturing time prediction under small sample data, but also shows significant advantages in precision and stability;

[0026] 2. The ship segment manufacturing time prediction method under small sample data proposed by the application has an average value of 8.62% of the test result MAPE on the test set, so as to more accurately predict the ship segment manufacturing time under small sample data, improve the efficiency and cost control of shipbuilding, and improve the reliability and stability of ship delivery progress. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a method flowchart of the application;

[0028] Figure 2 is a target field segment time prediction result MAPE value comparison chart with and without migration learning;

[0029] A axis is the ship segment manufacturing prediction operation time, and B axis is the ship segment manufacturing actual operation time. DETAILED DESCRIPTION

[0030] In combination Figure 1 and Figure 2 The specific embodiments of the application are described, and the ship segment manufacturing time prediction method under small sample data described in the embodiment includes the following steps: Figure 1 as shown, comprising the following steps:

[0031] Step one: source field segment time prediction model construction based on machine learning;

[0032] The implementation process of step one is as follows:

[0033] Firstly, the method of determining the model is selected. The neural network model based on particle swarm optimization algorithm (PSOBPNN) is selected for segmented manufacturing time prediction. Each particle represents a set of parameters of the neural network model, i.e. the weight and bias value of the model, and the dimension of the particle is determined by the structure of the model, which depends on the number of nodes in the input layer, the number of nodes in each hidden layer and the number of nodes in the output layer. The particle dimension calculation formula is:

[0034]

[0035] wherein Dim is the particle dimension, i represents the index value of the hidden layer, L represents the index value of the maximum number of hidden layers, n in is the number of nodes in the input layer, is the number of nodes in the i-th hidden layer, n out is the number of nodes in the output layer.

[0036] The neural network model structure adopted in the application includes one hidden layer, five input layer nodes, four hidden layer nodes and one output layer node, and the particle dimension is 83.

[0037] Secondly, the input and output variables of the model are determined. The segmented weight, projected area, small group number, weld length and steel plate thickness are selected as the input features of the PSOBPNN, and the segmented manufacturing time is selected as the output. In order to improve the training efficiency and stability of the model, all input and output data are processed by decimal scaling normalization. The decimal scaling normalization formula is:

[0038]

[0039] k=log 10 (max(x)) (3)

[0040] wherein x' is the normalized data value, x is the data value before normalization, and k is the logarithmic function value corresponding to the maximum value of x with base 10.

[0041] Thirdly, the prediction accuracy index is determined. The mean absolute percentage error (MAPE) is used as the relevant accuracy index in the application, and the MAPE value can represent the average error degree of the prediction.

[0042]

[0043] wherein j represents the index value of the sample, y j represents the actual operation time of the j-th sample, represents the predicted operation time of the j-th sample, and n represents the maximum index value of the sample.

[0044] Next, the hyperparameter configuration of the neural network model is determined, including the number of hidden layers, the number of hidden layer nodes, the hidden layer activation function, the output layer activation function, the loss function, the number of training times, and the training target error value, and the hyperparameters of the neural network model adopted by the application are selected as follows:

[0045]

[0046] Finally, the hyperparameter settings of the particle swarm algorithm are determined, including individual and population learning factors, maximum iteration times, particle number, minimum inertia weight value, maximum inertia weight value, particle velocity range, and particle position range, and the hyperparameters of the particle swarm algorithm adopted by the application are selected as follows:

[0047]

[0048] Based on the above steps, the source field segmented manufacturing time prediction model is constructed and trained, and the network weight and bias value are extracted. The model will serve as the basis for transfer learning and further application to the target field segmented manufacturing time prediction.

[0049] Step two, target field segmented time prediction model construction based on transfer learning and machine learning;

[0050] In step two, first, the transfer strategy of weight and bias value is established. The weight and bias value of the source field segmented manufacturing time prediction model is transferred to the target field segmented time prediction model (TL-PSOBPNN). By analyzing the similarity and difference characteristics of the source field segmentation and the target field segmentation, the following strategy is determined: for the weight of the input layer to the hidden layer between the source field and the target field, take the freezing action; for other weights and bias values, take the fine-tuning action.

[0051] Second, the target field segmented time prediction model is trained. The TL-PSOBPNN constructed is trained by using the target field segmentation data, and the weights and biases to be fine-tuned in the model are optimized to ensure that the model adapts to the characteristics of the target field segmentation, thereby forming a TL-PSOBPNN suitable for target field segmented time prediction.

[0052] Finally, the effectiveness of the method is verified. By selecting actual production cases, bulk cargo ship segments are taken as source field segments, and cruise ship segments are taken as target field segments, and the main feature difference between the two types of segments is the thickness of the steel plate. The mean absolute percentage error (MAPE) is used as the precision index to compare the prediction effect of the model with and without transfer learning, thereby verifying the effectiveness of the application.

[0053] Appendix Figure 2A comparison chart of MAPE values of the target field segmented time prediction results with and without transfer learning is shown in the figure. The predicted time and actual job time of 40 test set samples are described in the figure. Each data point in the figure is composed of an interval, which represents the mean, minimum and maximum values of the results of 100 times of training for each data point. When the data point is closer to the 45° symmetry line, the predicted time of the model is more consistent with the actual job time. The mean values of MAPE of the TR model and the non-TR model are 8.62% and 14.97% respectively. At the same time, the mean values of 80% of the test samples under the TR model are controlled within 1.5 times of the dispersion band, and only 70% of the test samples under the non-TR model are controlled within 1.5 times of the dispersion band. Moreover, the error bars of the non-TR model are all larger than those of the TR model. Therefore, it is shown that the TR model has better prediction accuracy and prediction stability than the non-TR model.

[0054] In summary, the ship segment manufacturing time prediction method under small sample data proposed by the present application has a mean value of 8.62% of the test result MAPE obtained on the test set, so as to more accurately predict the ship segment manufacturing time under small sample data, improve the efficiency and cost control of shipbuilding, and improve the reliability and stability of ship delivery progress.

[0055] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and make equivalent embodiments with equivalent changes. Any simple modification, equivalent replacement and improvement of the above embodiments, as long as it does not deviate from the technical solution of the present application, is within the scope of protection of the present application.

Claims

1. A method for predicting ship block manufacturing time under small sample data, characterized in that, The prediction method comprises the following steps: Step 1, source field segment time prediction model construction based on machine learning, specifically comprising: Step 1.1, determining the methodology selection of the model: Select a neural network model based on a particle swarm algorithm to predict the segment manufacturing time, each particle represents a set of parameters of the neural network model, that is, the weight and bias value of the model, the dimension of the particle is determined by the structure of the model, which depends on the number of input layer nodes, the number of nodes of each hidden layer and the number of output layer nodes; the particle dimension calculation formula is: （1） wherein, is the particle dimension, represents the implicit layer number index value, represents the implicit layer maximum number index value, is the number of nodes of the input layer, is the number of nodes of the implicit layer, is the number of nodes of the output layer; Step 1.2, determining the input and output variables of the model: Select the segment weight, projected area, small group number, weld length and steel plate thickness as the input features of the neural network model, and the segment manufacturing time as the output, and perform decimal scaling normalization processing on all input and output data, based on the decimal scaling normalization formula: （2） （3） wherein, is the normalized data value, is the data value before normalization, is the maximum value of the logarithm value with base 10 corresponding to the maximum value of Step 1.3, determining the prediction accuracy index: Step 1.4, determining the hyperparameter configuration of the neural network model: Determine the hyperparameters of the neural network model, including the number of hidden layers, the number of hidden layer nodes, the hidden layer activation function, the output layer activation function, the loss function, the number of training times and the training target error value, by performing experiments under different hyperparameter combinations, evaluate the influence of each group of hyperparameters on the prediction performance of the model, and finally select the optimal hyperparameter configuration; Step 1.5, determining the hyperparameter configuration of the particle swarm algorithm: Determine the hyperparameters of the particle swarm algorithm, including the individual and population learning factors, the maximum number of iterations, the number of particles, the minimum inertia weight value, the maximum inertia weight value, the particle velocity range and the particle position range, by systematically adjusting and testing these hyperparameters, analyze their influence on the prediction results of the model, and finally select the optimal hyperparameter configuration; Step 2, target field segment time prediction model construction based on transfer learning and machine learning, specifically comprising: Step 2.1, establishing a transfer strategy for weight and bias value: Migrate the weight and bias value of the source field segment manufacturing time prediction model to the target field segment time prediction model, by analyzing the similarity and difference characteristics between the source field segment and the target field segment, determine the following strategy: for the weight from the input layer to the hidden layer of the similarity characteristics between the source field and the target field, take the freezing action; for other weights and bias values, take the fine-tuning action; Step 2.2, training the target field segment time prediction model: Train the target field segment time prediction model using the target field segment data to optimize the weights and biases to be fine-tuned in the model; Step 2.3, method effectiveness verification: By selecting actual production cases, using the mean absolute percentage error as the precision index, comparing the prediction effects of the models with and without transfer learning, to verify the effectiveness of the model.

2. The method according to claim 1, wherein, The step 1.3 comprises: Use the mean absolute percentage error as the relevant precision index, and the mean absolute percentage error value can represent the average error degree of the prediction; （4） wherein, represents an index value of a sample, represents an actual working time of the th sample, represents a predicted working time of the th sample, represents a maximum index value of a sample.

Images