| With the advent of Industry 4.0 era,a variety of production data have been produced in the industrial field.It is one of the most important applications of factory intelligence to monitor production status by learning data and mastering its internal rules.Learning historical data by machine learning and deep learning algorithms,detecting and predicting real-time data,and early warning of abnormal production data are of great significance for improving product quality and production efficiency.As a flexibly configurable,high-parallel,low-latency computing platform,FPGA is the preferred platform for implementing machine learning and deep learning models in the industrial field.Aiming at the demand of intelligent data anomaly detection in the industrial field,this paper studies the data anomaly detection algorithms based on the combination of prediction and classification,and implements the data prediction-classification anomaly detecting based on FPGA.The main work of the paper is as follows:(1)A FPGA implementation mechanism of multiple long short-term memory network models based on the barrel effect is designed,which can improve the reusability of computing resources,reduce the usage of computing resources,and reduce the requirements of computing resources of the FPGA platform without increasing the prediction time of the system.(2)Aiming at the shortage of FPGA memory resources,this paper designs an FPGA implementation method of decision tree based on data coding splicing and address index,and designs and implements an extreme gradient boosting algorithm FPGA implementation method based on the number of decision trees in parallel,which reduces the memory required to deploy the extreme gradient boosting models on the FPGA platform and improves the computing speed.(3)Aiming at the difficulty of identifying abnormal samples,a joint anomaly detection algorithm based on long short-term memory network and extreme gradient boosting is designed.The long short-term memory network algorithm is used to learn historical data,and the extreme gradient boosting model is jointly trained with the prediction results of the long short-term memory network model and the original data,which effectively improves the recognition rate of the extreme gradient boosting algorithm for abnormal samples. |
