With developments of deep learning technologies and increasing demands for industrial process feature extractions, researchers pay much attention to deep learning of process historical data. Currently, the convolutional neural network (CNN) has made progress into extracting potential deep features from industrial process plant data. However, the existing sliding window associated with CNNs is rarely concerned with the characteristics of process historical data such as slow time varying and variable correlations. In response to this problem, an improved bar-shaped CNN (IBS-CNN) tailored to industrial process historical data is developed in this paper. Therein, process historical data are formulated as bars before trial-and-error methods are used to determine the optimum range of data for each calculating iteration. As a result, the width of sliding window is consistent with variable numbers during the algorithmic operations, in which the historical data involved in the sliding window are governed by several critical factors: time varying factors, variable similarity factors, and operating trend factors. Consequently, the IBS-CNN can enjoy making effective uses of historical data, helping to promote the development of process historical data driven models. The proposed method is applied to chemical operational optimizations, and a methanol synthesis process is employed as a case study to verify the effectiveness of the proposed method. Comparing experimental results of the IBS-CNN with those of traditional CNNs, bar-shaped CNN, and back-propagation (BP) neural networks, the tangible benefit of our contributions is demonstrated.