A multi-domain data-driven and multiscale 1DCNN-based ensemble learning method for condition monitoring

Condition monitoring is essential for ensuring smooth and efficient production. With advancements in information and communication technologies, condition monitoring has increasingly been facilitated into the data-driven paradigm, particularly the deep learning (DL) methods. As a representative of DL, one-dimensional convolutional neural network (1DCNN) has attracted much attention due to its strong learning ability and has been applied to one-dimensional sensory data and industrial time-series measurement. Nevertheless, conventional 1DCNN suffers from insufficient feature extraction with a single kernel, limited data fusion and integration from a single domain, and unstable model performance with a single model instance. To address these problems, this research proposes a multi-domain data-driven and multiscale 1DCNN-based ensemble learning method for condition monitoring. In particular, a multiscale parallel convolution structure is introduced to capture feature representations at various resolutions by simultaneously employing kernels of different sizes. An attention-based multi-branch fusion mechanism is developed to effectively combine information from multiple domain sources. Additionally, an ensemble learning framework is applied to aggregate predictions from several model learners, thereby improving accuracy while minimising bias and uncertainty. The proposed method is validated through a real-world machining process case study supported by extensive experimental evaluations.