Modeling Method For Complex Industrial Process Based On GANs Data Augmentation

Complex process industry is a key component of China's national economy and an important power source of social development,it is of great significance to the intelligent optimization and control of its operation process.In order to realize the intelligent control and optimization of complex industrial process,data-driven modeling method has been widely concerned and applied in the industry.In general,the predictive performance of data-driven models depends on the quality of the modeling data.However,a series of problems caused by a variety of factors such as high sampling cost and variable working conditions,for example insufficient number of modeling samples and uneven distribution of samples,seriously limit the further promotion and application of data-driven methods.So how to establish accurate performance prediction model of key indicators in complex industrial process based on small sample data has important theoretical research and practical application value.In this paper,the problem of small samples in complex industrial process modeling is studied deeply.By combining the data augmentation method based on GANs with the data-driven method,three kinds of small samples modeling methods for complex industrial process are proposed.(1)Aiming at the problem that it is difficult to establish an accurate data-driven prediction model in complex industrial processes due to insufficient modeling samples,this paper proposes a modeling method for complex industrial process based on WGANs-GP data augmentation.The method learns the statistical distribution information of the original modeling samples through WGANs-GP to obtain the generated samples consistent with the distribution of small samples.Considering that low quality samples may have a negative impact on the final prediction model,a data elimination technology based on similarity measurement is adopted to verify all generated samples to eliminate low quality samples.By selectively adding generated samples to the original modeling samples and adaptively increasing the number of generated samples in the training samples,the final data-driven model(such as least square support vector machine)can be trained to achieve the best effect,and the final performance prediction model of complex industrial processes can be established.Simulation results show the effectiveness of the proposed method.(2)Aiming at the problem of “negative transfer” in complex industrial multi processes modeling because of the unbalanced number of modeling samples of multi processes,this paper proposes a multi process modeling method for complex industrial based on CWGANs data augmentation.This method combines the advantages of CWGANs and multi-task learning,learns the distribution of modeling data of multiple industrial processes of multi-task learning by CWGANs.Data augmentation is selectively performed on industrial processes with relatively small amount of modeling data and combine generated data and original data to balance the amount of modeling data across multiple tasks.The industrial process model was regarded as the common feature and the unique feature,and the multi-task least square support vector machine(MTLSSVM)was used to train the model.After the training,several industrial process prediction models were obtained,which further improved the generalization and prediction accuracy of the model.Simulation results show the effectiveness of the proposed method.(3)Aiming at the migration problem of prediction models between similar old and new industrial processes,this paper proposes a transfer modeling method for complex industrial process based on ATL-BMA.This method mainly consists of two transfer processes.The first transfer process is to learn the transformation mapping function between the new process data and the old process data through the ATL,to transform a small amount of new process data into multiple types of “old process data with new process information”.Then we can obtain multiple “old process models with new process information” through the support vector regression algorithm.The second transfer process is to use the multi-model migration strategy to migrate several "old process models with new process information" that have been trained,and combine a small amount of new process data to get the final new process performance prediction model.This method not only improves the efficiency of new process modeling,but also reduces the modeling cost,improves the effect of multi-model migration,and improves the prediction performance and generalization ability of the final prediction model of the new industrial process.Simulation results show the effectiveness of the proposed method.