Incipient fault detection based on double Kullback–Leibler divergence KLDattW improved by a self-attention mechanism

The precise identification of incipient faults in industrial processes presented a significant challenge, as traditional methods based on principal component analysis (PCA) exhibit unsatisfactory detection rates. Kullback–Leibler divergence (KLD) detection improves fault detection capabilities to a certain extent, but it processes all the statistical components in the same way: diminishing or obscuring essential data that are pertinent to faults. This paper presents a self-attention-based double KLD detection technique in which the first stage of KLD is combined with the local outlier factor (LOF) to quantify the severity of faults. The second KLD stage calculates a new statistic, $KL{D}_{att}^W$, on the basis of the fault-weighted scores obtained from the self-attention mechanism. Additionally, control limits are determined via the kernel density estimation (KDE) method. $KL{D}_{att}^W$ validated this method by applying it to three types of incipient sensor faults induced during the continuous stirred tank heater (CSTH) process and two incipient faults induced during the Tennessee Eastman (TE) process, demonstrating its superior fault detection rates (FDRs) and efficacy compared to the existing methods in all evaluated cases.