Research On SPC Of Marine Engineering Project Quality Based On Complex System

In recent years the SPC technology application in production process has been more widely, use SPC method using statistical technology on production process of each procedure parameters monitor, can ensure production changed subtly and product quality goal. But the basic premise of traditional SPC application condition is variables, and independent observation value in the process of actual production since each observation often exist in the relevant phenomena, if still use the traditional SPC will increase the probability of wrong judgment.In order to solve this problem, this thesis first focuses on traditional SPC, residual control chart relevant theoretical knowledge, and the single value control diagram and residual simulation analysis and calculation of control chart, according to the time sequence model produce different from the observation data correlation coefficient, respectively in the mean no migration and mean offset two cases comparative study, eventually got related conclusions.In order to get more sensitive, more accurate process control, this thesis secondly the BP neural network is applied to solve the problems relevant process control, seriously study the BP neural network structure and the establishment of the training, and training of BP network is applied in good autocorrelation statistical process simulation experiments, according to set up different number of neurons in hidden layer and different from correlation coefficient, and kinds of the result was analyzed, and based on this, compared the BP neural network and residual control chart of autocorrelation process detection effect, eventually got related conclusions.Finally, this thesis the offshore platform in ocean engineering for example, prefabricated process empirical study hart control charts, the Hugh residual control chart and the BP neural network to the actual statistical process control ability, through comparing found that eventually got related conclusions. The findings and conclusions of the thesis are as follows:1.No offset and respectively in the mean reversion migration of two kinds of situations, a single value of control chart and residual simulation analysis and calculation of control chart, a comparative study concluded that:(1) the positive autocorrelation and mean no offset cases: single value of control chart often leads to data generated a lot of virtual sound the alarm behaviours, and with the increase of related coefficient, creating virtual sound the alarm, eventually mislead the likelihood of actual operation personnel to make the wrong decision; quality management And the residue of control chart can well reflect the real situation.(2) on the positive autocorrelation and mean offset cases: single value of control chart more tally with the actual situation, can give the correct offset direction; And the residue of control chart on abnormal monitoring is not very sensitive, and is, the greater the correlation coefficient of from the lower monitoring sensitivity.(3) in a negative autocorrelation and mean no offset cases: single value of control chart will display control fluctuation boundary scope change, cause will appear leak sound the alarm problem, and along with the negative correlation coefficient of absolute value since become larger, the control range will become larger and larger, more appear leak sound the alarm; And the residue of control chart, better reflect the real situation.(4) in a negative autocorrelation and mean offset cases: single value of control chart will display control fluctuation boundary scope change, cause will appear leak sound the alarm problem, and along with the negative correlation coefficient of absolute value since become larger, the control range will become larger and larger, more appear leak sound the alarm; And the residue of control chart, better reflect the real situation, and the smaller the correlation coefficient, residual control chart sensitivity is higher.(5) single value of control chart can only mean happened in the case, process shift when there is a positive autocorrelation play a role, at other times, not exist virtual sound the alarm that being leak sound the alarm, the serious influence operators designated quality control decision-making, to quality management bring serious adverse effects. And single value of control chart opposite is residual control chart, except in the case occurred offset mean there is a positive autocorrelation, process the moment will appear leak sound the alarm, monitoring the defects of the sensitivity in any other case has good performance.2. Through research in hidden layer number of neurons in the same (not respectively, 5, 8, 10, 12, 15), different mean migration (0σ, 1σ, 2σ), since the correlation coefficient different (0.1,... and 0.9) case, BP neural network (BPN) and residual control chart (TSCC) the comparative study of the network that:BP network network recognition rate than residual control chart with high recognition rate; Another hidden neurons when the number five, 8, 12 to 15 when BP network, the network recognition performance is not very stable and network recognition rate are relatively low, and when the number of hidden neurons for 10, network stability and recognition rate is very high.According to the above conclusions, hidden layer number of 10 neuron network performance is recognition rate is best, so we emphatically summarized the circumstances BP network and residual control chart of autocorrelation process, and the effect that:(1) for different since the correlation coefficient and different mean migration, BP network recognition rate is higher than the residual control chart recognition rate.(2) to discuss in the mean reversion fluctuant circumstance: offset the little change of the shift appear 1, BP network general prep above the recognition rate of 70%, and residual control chart, the recognition that the basic below 40% residual control chart recognition of little change mean offset limited ability, visible when mean offset changes for 1, the recognition of BP network than residual control chart recognition rate. In the mean offset appear 2 changes, the BP network recognition rate is higher than the residual control chart recognition rate.(3) to discuss the process fluctuant circumstance: since correlation coefficient in mean no offset conditions, BP network and residual control chart recognition rate has good performance. But since the correlation coefficient of different, both recognition rate is still have different change trend, the BP network recognition rate will with the correlation coefficient and the increase of increasing constantly; Instead the residual control chart recognition rate will with the correlation coefficient and the increase of gradually reduced; In a tiny mean offset appear changes, the recognition of BP network with the increase of the correlation coefficient and continuously decreased; Instead the residual control chart recognition rate will with the correlation coefficient and the increase of gradually increasing.3. Through the offshore platform in ocean engineering prefabricated process, for example, empirical study hart control charts, the Hugh residual control chart and the BP neural network for engineering pile pipe welding process of relativity statistical process control ability, through to a conclusion, that is:Traditional Hugh hart control chart and residual control chart didn't detect the mean migration happened; And the BP neural network testing out the situation happened mean migration. Can proof, the BP neural network on the relevant statistical process control ability than Hugh hart control charts, residual control chart.