| CPC G06Q 30/0202 (2013.01) [G06N 3/04 (2013.01); G06N 3/08 (2013.01)] | 4 Claims |
|
1. A customized product performance prediction method based on heterogeneous data difference compensation fusion, comprising the following steps:
(1) collecting and obtaining data samples with a configuration parameter of a customized product as an input feature and performance of the customized product to be predicted as an output feature; collecting actual measurement performance data of an existing product, and constructing a historical actual measurement data set for performance prediction of the customized product; establishing a virtual simulation model of the customized product by using computer simulation software, obtaining performance data through simulation analysis, and constructing a calculation simulation data set for performance prediction of the customized product;
(2) performing data preprocessing on the historical actual measurement data set and the calculation simulation data set;
(3) performing difference compensation correction on the calculation simulation data set on the basis of the historical actual measurement data set: encoding the historical actual measurement data set and the calculation simulation data set on the basis of a depth auto-encoder, mapping the data samples from an input space into a feature space, so as to express key features of the data samples, denoting the encoded historical actual measurement data set as ESeth, and denoting the encoded calculation simulation data set as ESets; dividing, through random sampling, the data set ESeth into a training sample set, a verification sample set and a test sample set, which are denoted as a historical actual measurement training set ESethtrain, a historical actual measurement verification set ESethvalid, and a historical actual measurement test set ESethtest, respectively; and finally performing associated connection on the data set ESets and the data set ESethtrain by using a neighborhood association method, performing difference compensation correction on the data set ESets by using the data set ESethtrain by means of a similarity difference compensation method, and denoting the data set ESets after the difference compensation correction as MSets;
(4) selecting a BP neural network model as a performance prediction model of the customized product, and taking the input feature and the output feature selected in the step (1) as the input and output of the prediction model; using the calculation simulation data set after the difference compensation correction as the training sample set, and training and constructing an optimal BP neural network model combined with a tabu search algorithm; and then testing the model by using the historical actual measurement test set ESethtest, so as to obtain a final performance prediction model of the customized product; and
(5) for a data sample to be predicted, firstly performing data preprocessing according to the processing of the calculation simulation data set in the step (2), and then inputting the data sample into the depth auto-encoder constructed in the step (3) for encoding, and finally inputting the encoded sample to be predicted into the prediction model constructed in the step (4) for prediction and obtaining the product performance of the customized product under different configuration parameter conditions;
in the step (3), a neural network model is trained by using the historical actual measurement data set and the calculation simulation data set to serve as the depth auto-encoder for the data samples; the depth auto-encoder is composed of an input layer, an encoder, a feature expression layer, a decoder and an output layer, and both the encoder and the decoder comprise three hidden layers; the input and output of the depth auto-encoder are input feature vectors of the data samples, the layers are fully connected, an activation function between the input layer and the hidden layer, and an activation function between the hidden layer and the hidden layer are relu function, and an activation function between the hidden layer and the output layer is tanh function;
in the step (3),associated connection on the encoded calculation simulation data set ESets and the encoded historical actual measurement training set ESethtrain is performed by using the neighborhood association method, the specific association process is: initializing an empty mark set for each data sample in the data set ESets; randomly selecting a data sample Samplek in the data set ESethtrain, taking the data sample Samplek as the center, taking a neighborhood threshold ε as the radius, adding tags of data samples in the data set ESets within this neighborhood range into the mark set, wherein the added mark is the serial number of the data sample Samplek, and at the same time, setting the access attribute of the data sample Samplek as visited; traversing all data samples whose access attributes are unvisited in the data set ESethtrain, and repeatedly adding marks for the data samples within the neighborhood range until the access attributes of all the data samples in the data set ESethtrain are visited; and
in the step (3), difference compensation correction on the encoded calculation simulation data set ESets is performed by using the historical actual measurement training set ESethtrain on the basis of the similarity difference compensation method, the similarity difference compensation method is: traversing the data set ESets, and performing difference compensation correction on the output feature of the data sample for each data sample Samplel whose mark tag is not empty according to the following formula:
 where, ŷFEAl represents an output feature vector of the data sample Samplel after the difference compensation correction; yFEAl represents the output feature vector of the data sample Samplel before the difference compensation correction, that is, the output feature vector obtained through simulation analysis; M represents the number of marks in the mark set of the data sample Samplel, that is, the number of data samples in the data set ESethtrain associated with the data sample Samplel; Sz represents the Euclidean distance between the data sample Samplel and the data sample in the zth data set ESethtrain associated with the data sample Samplel, which measures the similarity between the two data samples; Δyz represents an absolute difference between the output feature vector of the data sample Samplel and the output feature vector of the data sample in the zth data set ESethtrain associated with the data sample Samplel; yrealz represents the output feature vector of the data sample in the zth data set ESethtrain associated with the data sample Samplel; and α and β are hyperparameters.
|