Dynamic Soft Sensor Modeling for Time Sequence Data Based on Convolution Filters

Dynamic responses, measurement noise, and time delays in industrial processes can affect the accuracy of soft sensing models. Our goal is to create a dynamic soft sensor modeling method that is robust to the abovementioned problems. Aiming at time sequence data, a feature extraction-based model is constructed by combining convolution filtering techniques with a backpropagation neural network. In addition, training strategies based on gradient descent are proposed to ensure the efficiency and accuracy of the training process. Convolution parameters are trained with model parameters synchronously without prior knowledge. The performance is verified through simulations and industrial experiments in terms of measurement noise, sensor drifting, and dynamic modeling problems.