Research On Whole Process Fault Prediction Method Based On Multi-Time Series

With the development of industrial production technology and the rising production requirements,production system has developed toward large-scale and complexity. In order to ensure the safety and reliability of system operation, and avoid property lost and casualties caused, the system faults must be detected and isolated before they occur. Fault prediction is considered as one of the main research focus, which has been widely concerned by many scholars. Neural network method is taken as an effective data-driven modeling approach, which is widely used in fault diagnosis and fault prediction. However, at present, for many large production systems, large data will generated during the system running period. On one hand, excessive data will bring much difficulty to system monitoring, and it cost much time. On the other hand, excessive data will also produce redundant information, which will lead to interference on system modeling and model performance. The main research goal of this paper is to extract critical data from large amounts of multi-time series data, and use neural network method to predict system fault effectively and accurately. The concrete research content can be divided into several parts:First, K-nearest neighbor mutual information method is used to reduce variable dimension of multi-time series and calculate the correlation of variables, then the relevant variables are selected. Through selecting the target variable, the K-nearest neighbor MI method is used to calculate the mutual information for each variable with the target variable. Then, the threshold of relevant variables is determined by threshold optimization method, so as to acquire the relevant variable set under the same operation condition. Thus, the dimension reduction for multi-variable is realized.Second, the real-time monitoring for system condition is studied based on trend extraction, and the system condition is divided effectively. Aiming at the specific research object, the threshold parameters in the trend extraction method is simplified and improved.The setting of the model parameters is solved effectively, and the analysis accuracy is enhanced. Thus, the monitoring of running condition for feature variables is realized.Third, aiming at the potential fault stage, the differential evolution extreme learning machine (ELM) method is used to predict the system fault. During the research process, the differential evolution algorithm is used to encode the neural network structure of ELM. Then, a fitness function is constructed to optimize the neural network structure. Finally, condition prediction is made on the feature variables acquired by K-nearest neighbor MI method.The proposed fault prediction method is verified by penicillin fermentation process. From the simulation experiments, it is illustrated that the proposed method has better fault prediction effect.