Search

An adaptive optimization scheduling AQ (Assembly Q-learning) algorithm of knowledgeable manufacture is proposed for an aero-engine assembly line to address the scheduling of an aero-engine assembly workshop in an uncertain manufacture system, which combines the real-time feature of Q-learning with the self-adaptation feature of knowledgeable manufacture system. An adaptive scheduling model of aero-engine assembly based on Q-learning is developed for the purpose of minimizing earliness penalty and completion time cost. New production scheduling rules are proposed to deal with the aeroengine assembly scheduling problem. Addressing the characteristics of aero-engine assembly, four system-state features are defined, and a proper reward is designed as a reward function. Coherence of the reward function and the scheduling objective is proved by theorem. Simulation tests show that the developed algorithm is superior to other scheduling rules in various situations. Particularly, in the ever-changing assembling environment, better results are guaranteed by the desirable adaptivity of the proposed algorithm. An adaptive optimization scheduling algorithm of an aero-engine assembly workshop in a knowledgeable environment is designed, whereby a problem-solving method based on Q-learning and process simulation is activated. The coherence of the reward function and the scheduling objective is proven by theorem. The developed algorithm is superior to the other scheduling rules in the vast majority of cases. Thus, through self-learning, the aero-engine assembly workshop has much higher efficiency and adaptability than the others. [ABSTRACT FROM AUTHOR]

Design time forecast is characterized by problems of small samples and heteroscedastic noise. For better control of product development process, the forecast method is expected to yield not only highly precise forecast values but valid forecast variances. For this end, relative entropy kernel regression (REKR) is proposed based on the combination of kernel-based regression and Gaussian margin machines (GMM). REKR maintains a Gaussian distribution over weight vectors for the kernel-based regression to minimize the combination of the relative entropy and the negative log probability densities of the targets. To simplify the formulation of REKR, the covariance matrix of Gaussian distribution is set as a diagonal matrix. The simplified optimization problem of REKR is solved based on particle swarm optimization (PSO). Having inherited the benefits of GMM, REKR can simultaneously offer the desirable forecast value and forecast variance. Experiments on the time forecast of plastic injection mold design by both synthetic and real datasets do verify the feasibility and validity of REKR. [ABSTRACT FROM AUTHOR]

In view of the low generalization capacity of standard support vector machine for some types of noises existing in time series such as white noise, singularities, and biggish magnitude noises, a robust loss function is designed to inhibit (penalize) the above hybrid noises. The structure risk minimization (SRM) based on a geometrical interval with a special hyper-plane is theoretically proved. On the basis of that, a new support vector machine called R v -SVM (robust v -support vector machine) that meets SRM is proposed to handle sale series. Moreover, by modifying the standard v -SVM formulation, R v -SVM is characterized with its simpler dual optimization and fewer output parameters. With the help of the above, a sale forecasting method based on R v -SVM and its relevant parameter-choosing algorithm is formulated and applied to car sales forecasting, the convincing results of which definitely confirm the feasibility and validity of the forecasting method. Compared with the standard v -SVM and traditional model, the R v -SVM method is of better estimating precision and higher generalization capacity. [ABSTRACT FROM AUTHOR]

A new approach to combining forecasts by rules based on reliability and validity is proposed in terms of the basic principle of the personnel scrutiny method. Robust models that fit in current forecasting problems are chosen by means of matching rules based on reliability and validity, forming the participating model sets of historical case set and current case set. Then, these models are assigned the degrees of satisfaction according to their forecasting accuracy on the above-mentioned two case sets and endowed with corresponding weights. A rational experiment on validating the combination model is designed. The result shows the proposed method not only attains a forecasting accuracy comparable to that of the theoretically optimal combining methods but also provides some valuable insights into combination methodology, which verifies that the combining forecast upgrades the forecasting performance. [ABSTRACT FROM AUTHOR]

The manufacturing enterprise needs price forecasting in order to attain maximum profit and guild production. However, it is difficult to trace the changing rules of their historical data since product price may fluctuate abruptly. Therefore, a price forecasting method based on correlative price net and neural networks is proposed in the paper. Firstly, by analyzing the main factors to product price changes, a model of correlative price net (CPN) that connects many products whose prices affect each other is set up. Then, the theoretical proof that the whole CPN can be substituted by its correlative price sub-net for price forecasting is provided. Based on the correlative price sub-net, a set of BP neural networks are introduced for building a new network, named the price forecasting net (PFN), which can reflect the factors of price fluctuation in the correlative price sub-net. Finally, since the change of factors in the correlative price sub-net relates to the variation of product price caused by it, the product price can be forecasted by the change of factors in the correlative price sub-net by means of PFN. The theoretic analysis and simulation experiment show that the proposed method adapts well to product price forecasting, especially when prices change abruptly. [ABSTRACT FROM AUTHOR]

With the growth in customer demand diversification, research on mixed-model assembly lines have been given increasing attention in the field of management. Sequencing decisions are crucial for managing mixed-model assembly lines. To improve production efficiency, the product sequencing problem with skip utility work strategy and sequence-dependent setup times is focused on in this study, and its mathematical model is established, whereby the idle cost, the utility cost and the setup cost are to be optimized simultaneously. A necessary condition for skip policy of the system is set, and a lower bound of utility work cost is given and theoretically proved. Strong NP-hardness of the problem is confirmed. Addressing the main features of the problem, a hybrid EMVNS (electromagnetism-like mechanism-variable neighborhood search) algorithm is developed. To enhance the local search ability of EM, a VNS algorithm is employed and five neighborhood structures are designed. With the aid of the VNS algorithm, the fine neighbour search of the optimum individual is made available, thus improving the solution to a certain extent. Simulation results demonstrate that the algorithm is feasible and valid. Significance: This paper presents a hybrid EMVNS algorithm to solve the product sequencing problem of a mixed-model assembly line with skip utility work strategy and sequence-dependent setup times. The simulation results demonstrate that the proposed method is feasible and valid. [ABSTRACT FROM AUTHOR]