Research On Multi-modal Process Ensemble Fault Diagnosis Method

The technology content of modern industrial production process is getting more and more higher,and the requirement of safety production and equipment maintainability is also increasing gradually.At present,the majority of complex chemical production processes belong to multimodal process.Because of its product requirements,production schemes and many other reasons,there will be multiple operating conditions,resulting in a large number of multimodal data.Multimodal data contains many process variables,and there are some characteristics such as non-linearity and non-robustness among the data.If there is equipment trouble in multi-mode industrial production process,the whole production line should be stopped producing related products,which will affect the efficiency and even threaten the personal safety of technicians.Therefore,the development of multimodal process fault diagnosis technology is very important for today's industrial safety production.The main content of this thesis is to acquire data based on the industrial model of Tennessee Eastman process as a simulation model,and to study the collective fault diagnosis method of multimodal process based on deep learning.The multimodal process set fault diagnosis method is divided into four parts:data denoising,data dimension reduction,fault classification and model parameter optimization.A set fault diagnosis method based on deep learning,namely VMD-LLE-CPSO-DBN fault diagnosis method,is proposed in this thesis.The experiment is simulated by TE process simulation model.Variational modal decomposition algorithm is used in data denoising for normal and fault data sets to reduce data noise.Because of the long training time of deep belief network in building high-dimensional data classification model,local linear embedding algorithm is used for high-dimensional samples.This data reduces the number of variables,reduces data redundancy,extracts features more easily and shortens the training time of the model.Chaotic particle swarm optimization is used to optimize the parameters of deep belief network model,which avoids the situation that basic particle swarm optimization is easy to fall into local optimum.Then,taking the dimension-reduced data as the network input of the deep belief network,a deep belief network with five layers of three restricted Boltzmann machines is constructed,and the deep belief network is powerful.Data learning ability is used to classify taults,and a class of deep belief network classifier with data preprocessing mechanism is constructed.Then chaotic particle swarm optimization is used to optimize the parameters of deep belief network model,which avoids the situation that basic particle swarm optimization is easy to fall into local optimum.The comparison proves that this method can improve the learning ability of model samples and the accuracy of model fault diagnosis.Finally,a multimodal process simulation experiment platform is established to verify the effectiveness of VMD-LLE-CPSO-DBN ensemble algorithm.