Series An Empirical Study Of The Mechanical System Failure Prediction And Control Problems

With the development of microelectronics and manufacturing industry, the structure of devices is becoming more complicated while the cost is becoming higher day after day. Improving the utilizing efficiency of equipment is essential to advance the production efficiency as well as to reduce the production cost. The main study object of this thesis is the equipment failure data from semiconductor packaging factories. This thesis predicts the failure data by analyzing the history failure data, and then provides some suggestions of equipment maintenance strategy to approach the target of the increase of the production efficiency.The failure data studied in this thesis came from different mechanical systems, which are composed of several different devices. This thesis mainly applies time series method to those complicated data to get the forecasting model. The failure data may be regarded as a set of dynamic data in time series. The dynamic data are usually non-stationary and needed to be transformed into stationary data first.The difference method is used to transform the non- stationary data into stationary data, and the result is still perfect. So the ARIMA model is established to make prediction analysis, and the result shows that the method is available.Considering the phase trend of some data, the thesis tries a new congruence forecasting model. In this model, the deterministic trend sequence {mt} can be separated from a stochastic sequence {Yt} in the sequence {Xt}. In demonstration study, failure data pretreated is regarded as a set of dynamic data in time series. The deterministic trend sequence from the dynamic data is used to delete the phase trend of the failure data, and the stochastic sequence shows the stochastic factor of the failure data. The thesis forecasts the deterministic trend sequence with nonlinear autoregressive method, while it forecasts the random sequence by establishing ARMA model. In the end, a forecast of the original data comes out with the integration of those two kinds of forecast results, and the feasibility of the new congruence model is obtained by the prediction result.