Multi-feature fusion-based TCA-WKNN cross-sensor fault diagnosis method for dynamic weighing

Abstract

The outstanding performance of current machine-learning fault diagnosis methods is mainly attributed to the availability of a large amount of labeled training data. However, in practical dynamic weighing systems, the high costs and variability of operating conditions limit the availability of reliable training data, hampering engineering fault diagnoses in dynamic weighing systems. To address this issue, this study proposes a novel cross-sensor fault diagnosis method based on multi-feature fusion using a transfer component analysis (TCA)-weighted k-nearest-neighbor (WKNN) network. Using this method, time- and frequency-domain features are extracted from a laboratory-simulated set of fault data with small batches of real operational data. Source and target domain features are fused, and TCA is applied to map the source and target domain samples to a latent space using kernel functions to reduce the distribution differences among the samples. Finally, the WKNN is employed as a metric learner to enhance small-sample data matching and classification to improve diagnostic accuracy. The results show that with three samples per support set, the proposed method achieves a diagnostic accuracy of 93.33%. Compared with other approaches, the proposed method exhibits stronger generalizability for diagnostic knowledge transference from sensor to dynamic weighing failure data, effectively improving precision in on-site small-sample environments and reducing sample imbalances.