Anomaly Detection System For Big Data Of Mobile Printed Circuit Board Industry

Data-based PCB detection methods for mobile phones face the following four challenges: large amount of data;high data dimension;unbalanced distribution of data categories;and lack of adaptive ability of detection system.Printed circuit board(PCB)of mobile phone has a high production speed,so it has a large amount of data.The detection system needs 40 or 50 different tests,including the thickness,volume,area and offset of solder paste,which results in high data dimension.In general,the bad rate in production is only 150 PPM,so the data distribution is very unbalanced.The detection system is based on rules,so it lacks adaptive ability.In order to solve the above problems,this paper carries out the following research on anomaly detection of large data in mobile PCB industry:1)Combining Wasserstein distance and mutual information,a feature selection algorithm is proposed to improve the detection speed by reducing the feature dimension of data sets.The basic principle of the algorithm is to measure the importance of features to classification by Wasserstein distance,and to find a subset of features that are important to classification but less redundant to each other by combining the measurement of redundancy between features with mutual information.Experiments show that the algorithm is effective and can reduce the original data set from 45 features to 4,greatly mentioning the detection speed.2)Using GAN to synthesize data to balance the distribution of data sets.For unbalanced data sets,it is difficult for the system to accurately detect scarce bad products.The system can balance the data sets by synthesizing similar scarce data by GAN to improve the detection effect.Experiments show that the detection effect can be effectively improved by balancing data sets.When training about 200 times,high quality rare data can be obtained.3)Design a self-learning large data anomaly detection system for mobile printing board industry.The system adopts modular design,in which the data source module updates the local data continuously,and the anomaly detection module trains multiple standby models in real time according to the local data.The system will switch the off-line training model and the online model according to the change of the detection results,so as to realize the self-learning ability of the system.The results of online operation show that the system can raise the F1 value of test results to about 0.7 and the recall rate to about 0.75.In fact,the results of on-line operation show that the anomaly detection system designed in this paper can effectively improve the detection speed and effect,and realize automatic learning ability through multiple model hot switching,which has a good effect on reducing production costs and improving production efficiency.