Recent years have witnessed the unprecedented development of Industrial 4.0 and Industrial Internet of Things. These two technologies have significantly facilitated data collection from different sources for numerous tasks, such as reconstruction, classification, and prediction, for next-generation applications. However, the effective fusion and interpretation of these multi-source datasets remain challenging, making it a thriving area of research.
Currently, canonical correlation analysis (CCA) is considered a fundamental data fusion technique that preserves the essence of the information in correlated representations. Its extension kernel CCA (KCCA) enables the learning of nonlinear representations. However, its performance is significantly reduced on bigger datasets. Addressing these shortcomings, scientists have proposed deep CCA (DCCA) as a flexible alternative that employs the remarkable representation learning capabilities of deep neural networks (DNNs). However, all three CCA-based methods incorporate correlation into the optimization formulation itself, which may detract from the focus on the task. In this regard, the utilization of canonical correlation as an optimization constraint is promising.
Based on this idea, a team of researchers from China, led by Professor Zhiwen Chen from Central South University, has proposed an innovative deep learning architecture called the canonical correlation guided deep neural network (CCDNN) to learn correlated representations for multi-source data fusion. Joining him in this collaboration were Professors Weihua Gui, Zhaohui Jiang, and Chunhua Yang from Central South University; Professor Steven X. Ding from the University of Duisburg-Essen, Germany; and students Mr. Siwen Mo from Central South University and Mr. Haobin Ke at The Hong Kong Polytechnic University, Hong Kong, China. Their novel findings were published in Volume 13, Issue 3 of the IEEE/CAA Journal of Automatica Sinica on April 1, 2026.
Prof. Chen highlights the most important contribution of their study and shares, "Unlike the linear CCA, KCCA, and DCCA, in our proposed method, the optimization formulation is not restricted to maximizing correlation. Instead, we make canonical correlation a constraint, which preserves the correlated representation learning ability and focuses more on the engineering tasks endowed by optimization formulation, such as reconstruction, classification, and prediction. Furthermore, to reduce the redundancy induced by correlation, a redundancy filter with zero learned parameters is designed."
The team demonstrates CCDNN's data fusion capability through correlated representation learning and excellent performance across diverse engineering tasks. The proposed method demonstrated promising performance when compared to the existing methods. Furthermore, the technique showcased better reconstruction performance in terms of mean squared error (MSE) and mean absolute error (MAE) than DCCA and deep canonically correlated autoencoders in experiments on the MNIST dataset. Specifically, compared to DCCA, MSE and MAE values lowered by 0.43 and 0.42, respectively. Furthermore, the application of CCDNN to industrial fault diagnosis and remaining useful life cases for the classification and prediction tasks accordingly yielded superior performance when compared to existing methods.
"CCDNN can achieve effective data fusion by learning correlated representation via DNNs; hence, how to select appropriate DNNs for a specific engineering task is worth studying. In addition, both views of data are also flexible, which enables CCDNN to deal with multi-source heterogeneous data structures with different industrial applications, for instance, the engineering task of fault diagnosis, in which images give a view, and the other view is given by time-series," concludes Prof. Chen, highlighting the promising potential of their latest innovation.
