Raw process data may also be used for training. Many manufacturing companies already have available both process parameter data and quality inspection/product test data in their database. In order to use them for NN training, it must be possible to relate the two for an individual product—in other words, product traceability is required. This is frequently the case, for example in the aerospace and silicon wafer industries. However these data were usually collected as quality control records rather than for use in process modelling, and hence the process parameters set might not cover the full range possible, and might not include the optimum conditions. Also, the data collection and storage may not be tightly controlled and this will result in data integrity problems. However, such historical data reflect real process conditions and if they are available, no further data collection cost will occur. If large amounts of raw data are available, using all of it might be unnecessary and time consuming. Random selection of the required training and testing sets from the available data can be employed but it is important that the selected data cover the entire population of interest evenly. Inference tests can be use to ensure that the statistics of the samples used for training and testing are the same as the entire population. On the other hand if the available data are too sparse, resampling from a limited data set could be performed. It may, however, be undesirable to use all of the data available, for example because some reflects abnormal or outdated process conditions. In such cases, data blocking may be used to achieve specific training requirements, or to eliminate particular process behaviours for which it is not desired to train the network.
Raw process data may also be used for training. Many manufacturing companies already have available both process parameter data and quality inspection/product test data in their database. In order to use them for NN training, it must be possible to relate the two for an individual product—in other words, product traceability is required. This is frequently the case, for example in the aerospace and silicon wafer industries. However these data were usually collected as quality control records rather than for use in process modelling, and hence the process parameters set might not cover the full range possible, and might not include the optimum conditions. Also, the data collection and storage may not be tightly controlled and this will result in data integrity problems. However, such historical data reflect real process conditions and if they are available, no further data collection cost will occur. If large amounts of raw data are available, using all of it might be unnecessary and time consuming. Random selection of the required training and testing sets from the available data can be employed but it is important that the selected data cover the entire population of interest evenly. Inference tests can be use to ensure that the statistics of the samples used for training and testing are the same as the entire population. On the other hand if the available data are too sparse, resampling from a limited data set could be performed. It may, however, be undesirable to use all of the data available, for example because some reflects abnormal or outdated process conditions. In such cases, data blocking may be used to achieve specific training requirements, or to eliminate particular process behaviours for which it is not desired to train the network.
การแปล กรุณารอสักครู่..
