Research And Application On Multi-process Manufacturing Quality Based On Intelligent Predictive Modeling

With the increasingly competitive global market, product quality gets unprecedented attention. The key technologies and implementation methods of intelligent predication for manufacturing process quality are systematically researched in this paper, and it has the important academic value and practical significance to advanced manufacturing technology and intelligent prediction theory.The mechanism of manufacturing quality fluctuation is analyzed in detail, and predictive control theory of manufacturing quality is fully illustrated, then the research activities and trends of quality prediction are further discussed from three key problems, which are lower-dimensional quality information extraction, intelligent predictive modeling of manufacturing quality and quality information integration. Thus, research and application on multi-process manufacturing quality based on intelligent predictive modeling is discussed in this paper, and the main work includes the following parts:(1) Study on the multi-dimensional data modeling of manufacturing quality for key process is carried out. Establishing a KQCs(Key Quality Characteristics)mapping model of product structure tree, which indirectly realize KQCs mapping to the product structure tree through the functional decomposition in manufacturing quality BOM. For quality characteristics to each component, it not only embodies the common KQCs, but also expresses the personal KQCs. Then a multi-criteria decision ways combinding qualitative and quantitative factors is proposed, which can avoid limitations arose by subjective interpretation or incomplete consideration. Study on the online extraction method of different types of time series data sample in manufacturing process, and a multi-dimensional quality data model is established by the introduction of dimension, dimension stratification and measurement, which make the describing of manufacturing quality data more structured and standardized.(2) Deep study on the lower-dimensional quality information extraction from higher-dimensional quality information. A method combining the principle component analysis and sliding data window sampling, which adapt the PCA model to the changes in data sample, and it is useful to extract the quality small number characteristic including large amount of information. Further, a recursive KPCA is proposed and mapped to linearly separable high-dimensional feature space, and a recursive algorithm is given subsequently. Simulation results show that recursive algorithm is better than non-recursive algorithm, analyzing 500 samples, the SPE results based on PCA exceeded 95% and 99% control lines are reduced by 10% and 8% in order, correspondingly, the T 2 results based on KPCA exceeded 95% and 99% control lines are reduced by 10% and 7% than the former PCA model, while the SPE results are reduced by 15% and 4% accordingly.(3) Deep study on intelligent quality prediction based on Elman neural network. With increasingly complex and dynamic manufacturing quality control, various quality variables associated with the uncertainty and complexity are further increased, while the standard Elman neural network model only effective for the low-level static system, then a new OHIF Elman is proposed in this paper, associated layer 2 using the factor ? from the previous output values feedback to the input layer, while associated layer 3 using the factor ? from the previous output values feedback to the hidden layer. Being a multi-channel feedback link, more historical information can be used to improve prediction accuracy. However, the complexity of the model structure in turn affects the algorithm convergence and convergence rate, therefore, LM-CGD mixed algorithm is proposed to speed up the convergence, and the key is adopt CDG to solve the canonical equations J k of the matrix Jacobi , further, the computational complexity can be reduced to O (2 N 2), while the original method for solving the computational complexity is O ( N 3/ 6). The traditional Sigmoid activation function is difficult to establish a quantitative relationship between the network size and resolution scale, therefore, a wavelet OHIF Elman neural network model is proposed in this paper, which full use of the neural network weights of the linear distribution and learning convex objective function, so it can avoid the local optimal nonlinear optimization problems. Simulation results show that the wavelet Elman OHIF Elman network decreased 12.9 percent compared with OHIF Elman network which used the sigmoid activation function in hidden layer, and it also decreased 36.4 percent than RBF neural network.(4) Study on the design and implementation of information integration system for on-line manufacturing quality. The three-layer architecture is discussed including field monitoring, enterprise monitoring and remote monitoring layer, and the data transfer agreement is designed between the USB?RS232/422/484?analog and CAN bus, then a combination scheme of field bus, Ethernet, TCP/IP and OPC is designed to avoid unnecessary resources competition and enhance the security of information transfer. Subsequently, the quality data communication mode based on XML is put forward, and the proceeding of XML data modeling is divided into three parts, which include UML notional modeling and mapping from UML to XML Schema and generation of XML measurement & control data, all this can meet the demand of seamless transfer and interoperability among different structure system. According to the 985 project of south china university of technology, an on-line experimental platform of manufacturing quality detection is constructed and verify the above mentioned contents.(5) Study on the application of intelligent prediction modeling in piston-ring seal quality control. Analyzing the cause of light not sealed, further, construct a multi-criteria decision system with AHP, then the nitride, stereotype and microhoning process are selected as the first, second and third key process of quality control. For nitride process, six principle components are extracted with KPCA, and the nitride temperature, nitride time and activator are selected as the input of the prediction model, while the nitride rigidity is selected as the output of the prediction model. Then the quality prediction model based on an improved Elman neural network is successfully applied in nitride process, which inspects any odd change in the process and predict the process characteristics. The predictive accuracy of wavelet OHIF Elman and OHIF Elman neural network are correspondingly increased by 22.3 and 21.7 percent. Finally, the intelligent quality prediction model is applied in piston ring enterprise, and the ratio of superiorquality is increased to 87 percent from the original 75 percent.