Research On Method Of Multi-source Text Data Domain Knowledge Mining And Scheme Decision For Supporting Conceptual Design

The massive multi-source text data accumulated on the Internet contains abundant domain knowledge about product.Mining domain knowledge from the above text data and providing decision support for product conceptual design has become an important research goal in conceptual design research community.In view of the existing problems in the theory and practice of product conceptual design under the current multi-source data environment,this paper conducts research on method of conceptual design-oriented multi-source data domain knowledge mining and method of scheme decision for supporting conceptual design.The innovative research work of the thesis mainly includes the following four aspects:The paper studies the process of multi-source text data domain knowledge mining for conceptual design and proposes the method of domain knowledge mining.Based on the research on the conceptual design knowledge features and representation forms contained in the multi-source text data,the process of knowledge mining in conceptual design field is constructed.In this process,aiming at the problem that it is difficult to label the annotation data set for product technical term recognition,this paper proposes a product-specific pre-trained language model-CD-BERT(Conceptual Design-Bidirectional Encoder Representations from Transformers)to learn the basic domain knowledge through the self-supervised pre-training learning strategy of product-specific multi-source text deeping into the syntactic and semantic levels.Combining Conditional Random Field(CRF),this paper further proposes a CD-BERT-CRF technical term set generation model.The constructed CD-BERT-CRF model uses the strategy of transfer learning to inherit the basic domain knowledge learned by CD-BERT model,which greatly improves the effect of product technical term recognition.In view of the feature that the association relationship between technical terms in symbol form is not easy to mine,we use word embedding technology to map the association relationship between symbolized technical points into a continuous real-valued semantic vector space.Aiming at the existing methods of product functional requirements analysis are difficult to cope with the modeling challenges as the lack of word co-occurrence information caused by the sparsity of product short text data content,This paper proposes BCK-TM(Biterm Correlation Knowledge-based Topic Model)to mine product functional requirements from short texts.The constructed model solves the problem of insufficient word co-occurrence information caused by the sparse content of product short text by modeling the global word pair co-occurrence mode.In addition,the problem that the existing methods are difficult to integrate domain knowledge in the process of functional requirements modeling is solved by constructing a domain knowledge fusion mechanism based on Markov Random Field at the latent topic layer of the model,and the product function requirement modeling driven by the integration of data and domain knowledge is realized.The constructed model in this paper can improve the modeling quality of product functional requirements,thereby more effectively supporting R&D personnel to carry out product functional requirements analysis and decision-making.In view of the problem that the existing technology element mining methods which support the decision-making of conceptual design scheme cannot incorporate the prior knowledge in the data and results in the low information content and poor consistency of product technology element mining,this paper proposes a product technology element mining model-MTCK-LDA(Mixed Tec Correlation Knowledge-based Latent Dirichlet Allocation)based on domain knowledge.The knowledge representation form suitable for the mining task of product technical elements is proposed to integrate domain knowledge in the modeling process of technical elements in MTCK-LDA model.By constructing a knowledge fusion mechanism based on Hybrid Markov Random Field in the proposed model,the model is given the ability to fuse related knowledge between technical terms in the modeling process,and further improves the information content and coherence of the generated technical elements,so as to provide more effective support for assisting R&D personnel to make generation decisions about product scheme.Aiming at the problem that the existing quantitative technology prediction methods can not support the evaluation of conceptual design schemes effectively,this paper proposes a product technology evolution prediction model-SE-MVTTP(Scheme Evaluation-Multi View Technology Trend Prediction,SE-MVTTP)based on multi-source text data.The proposed SE-MVTTP model provides support for the decision-making of conceptual design schemes evaluation by accurately predicting the development trend of technology.The proposed SE-MVTTP model uses multi-sequence LSTMs to perform representation learning for the different perspectives of technology evolution.In view of the possible dynamic interaction between different views,a dynamic bidirectional sequential attention mechanism is designed in the model to capture the interaction effects.Finally,the model uses a multi-task learning mechanism to simultaneously learn in the unified deep learning architecture to predict the technology development trend in different technical activity views.Compared with traditional technology prediction methods,the model proposed in this paper can model more objective and accurate knowledge of technology evolution trend to provide reference for R&D personnel to carry out technical scheme evaluation and decision-making.Based on the above research,taking the intelligent numerical control system as an example,the methods proposed in the paper are used to provide support for the decision-making of the conceptual design of the intelligent numerical control system,and the validity of the methods proposed in this paper are verified.