A robust temporal multivariable fault detection method for blast furnace: Robust temporal convolution detection network

As a complex industrial process, the blast furnace ironmaking process (BFIP) often suffers from faults due to changes in raw materials, variables’ setpoints, reaction conditions, etc. However, because of the nonlinearity, dynamics, and mixture of normal and abnormal outliers in BFIP, the existing methods always meet difficulties in practical application. Therefore, this paper proposes a robust temporal convolution detection network (RTCDN) for BFIP fault detection. The basic TCDN network is constructed by stacking the residual blocks of the temporal convolution network (TCN). The residual block consists of the 1-D dilated causal convolution, which is performed on the time scale of variables, giving the network the ability to extract time-scale information. Since TCDN consists of convolutional networks, it balances the extraction of temporal information and speed compared to RNN-type methods. Furthermore, a robust solution is proposed to overcome the interference of abnormal outliers. The robust solution exploits that TCDN can’t reconstruct abnormal outliers and decomposes the training dataset into the clean part and abnormal outliers. In this work, the proximal approach to optimizing the sparse outlier matrix is also improved to achieve complete separation of abnormal outliers. Finally, TCDN is applied to three BF fault datasets and performs better than the conventional temporal detection methods. RTCDN has also been proven to have good outlier separation capability, maintaining satisfactory detection performance when the proportion of abnormal outliers is less than 20%.