Statistics-Informed Deep Learning Method for Parameter Drift Diagnosis in Fused Deposition Modeling

Fused deposition modeling (FDM), a prevalent additive manufacturing (AM) technique, is known for its low maintenance costs and safe system operation. However, parameter drift due to equipment aging can occur during the manufacturing process, adversely affecting product quality. Effective diagnosis of equipment-related faults can provide operators with corrective measures and promote quality improvement. The advancement of computer vision is driving the development of deep learning (DL) fault diagnosis with image data. Statistical information helps to measure the uncertainty of the manufacturing process and the model; therefore, this paper proposes a statistics-informed deep learning method. It comprises two parts: (1) a statistics-based explainable network (SEN) to provide a prior physics model; and (2) a statistics-injected network (SIN) to introduce the statistics-aware adaptively features into DL networks for the classification of fault labels. Manufacturing-process-related statistical information referring to the generalized likelihood ratio is constructed in this paper. Then, this information is integrated into ResNet18 through bilinear interpolation. The proposed method in conjunction with Resblk3 achieves the highest diagnostic accuracy of 89.44%, surpassing both the baseline and other learnable fusion modules. Experimental results demonstrate that incorporating statistical information helps to improve fault diagnosis.