Incipient fault detection based on ensemble learning and distribution dissimilarity analysis in multi-feature processes

Abstract

Timely and accurate detection of incipient faults has attracted considerable attention and research interest in recent years, due to its potential for the prevention of serious safety incidents and for supporting preventive maintenance. However, most existing methods use single detection model, neglecting the coexistence of multiple features and the local data distribution information found in industrial scenes. To overcome this problem, an incipient fault detection method named multiple model ensemble and distribution dissimilarity analysis (MME-DISSIM) is proposed. First, various multivariate statistical analysis methods are employed as basic detectors to comprehensively capture the feature information hidden in the process data. Second, distribution dissimilarity analysis is performed to evaluate the dissimilarity between the current sliding window and each training subset. This evaluation allows for the calculation of weighting factors for each basic detector, which helps to preserve the local distribution information of the current sliding window. Third, ensemble learning is utilized to integrate the statistics from all basic detectors into two detection indices to determine the operation status of the system. In addition, two measurement metrics are defined to quantitatively analyze the fault level of incipient faults. Finally, several experiments on a numerical case, Tennessee Eastman process, and actual PROcess NeTwork Optimization are presented to verify the efficacy and superiority of the proposed method.

Unsupervised method for detecting surface defects in steel based on joint optimization of pseudo-labeling and clustering

“…technology [2]. However, labeling the necessary correlation data is quite expensive in supervised learning.…”

“…(2) batch_size (the size of the sample set for running the minibatch K-means algorithm), comparing 64, 128, 256, and 512, respectively, to choose the best batch_size setting value. (3) init_size (set the number of samples that are candidates for the initial value of the center of mass, the value is usually set to 3 times the value of batch_size).The mini-batch K-means feature clustering method's batch size hyperparameters are compared and analyzed, and five different batch sizes, including 64, 128, 256 and 512, are used for comparison experiments.…”