Optimizing Fault Detection for Big Data Analytics Through Evolutionary Computation

Abstract Intelligent fault detection is promising to deal with big data due to its ability in rapidly and efficiently processing collected signals and providing accurate detection results. In traditional fault detection methods, however, the features are manually extracted depending on prior knowledge and diagnostic expertise, such processes take advantage of human ingenuity but are time-consuming. Inspired by the idea of unsupervised feature learning artificial intelligence techniques are used to learn features from the raw data. As the dimensionality increases, the accuracy of fault identification methods implemented on big data decreases significantly. For supervised learning, large volume of data is needed which leads to high cost and time consuming. In this paper, an unsupervised learning approach is proposed based on weighted softmax regression for fault detection using the power signals. Fault prediction has become an important subject in recent years, as it helps businesses to make significant savings in time and expense by offering successful methods for predictive maintenance. Pre-processing of data was a complicated job to overcome many problems with the dataset, including scale, sparsity, distortion, burst effects and confidence. As pre-monitor signals for failure did not share standard patterns, but were characterized only as non-normal system signals, a predictive error was made using outlier detection. Faults were explained by displaying system characteristics with abnormal values. An experimental assessment was conducted to determine the quality of the solution proposed. Results indicate that high-grade outliers provide successful markers of initial failures. In addition, explanations about irregular characteristic values (responsible for oversight) seem rather expressive. Based on the sliding window technique, the method to detect errors in high dimensional data streams is applied to an on-line mode. The online extension can be adapted to the time changing behaviour of the controlled system by experiments on synthetic datasets and is therefore applicable to the dynamic error detection. To assess the suggested strategy, we contrasted it with engineered data sets created utilizing the LOF (online expansion), SVDD, and KPCA approaches getting more than 90% result. The data exhibit our methodology's accomplishment as far as perfection, productivity, and strength.