Adaptive and cross-attention Vision Transformer-based transfer network for elevator fault diagnosis toward unbalanced samples

Abstract

As a large mechanical equipment frequently used in urban life, elevators are prone to various types of mechanical failures, affecting the comfort and safety of passengers. Due to the variable models and operating conditions of elevators, the traditional manual empirical diagnosis based on signal or statistical analysis is difficult to achieve accurate fault identification in the case of scarcity of vibration signal fault samples. In the actual operation of the elevator, guideway, boot liners, rope wheel, and other parts are prone to wear, resulting in vibration and noise of the elevator car. To solve these problems, a new adaptive and cross-attention Vision Transformer-based transfer network (ACAformer) for elevator fault diagnosis is proposed. This method designs a transfer network based on a Vision Transformer (ViT) for a long time series, achieving better equipment and working condition domain adaptation of elevator system under unbalanced samples. Two feature extraction branches with different scales are designed, realizing feature extraction and fusion of rich vibration signals from multiple sensors. Each branch is designed with adaptive attention based on feature resampling, focusing on deep fault features. Cross attention based on the exchange of classification information at different scales is designed, realizing the complementary multiscale fault features of two branches. Ultimately, the fault classification of target domain samples is realized by fine-tuning training of model parameters. For validating the proposed model, fault experiments were done on actual elevators, and the vibration signals were collected. Comparative experiments show that in equipment and working condition transfer tasks, ACAformer improves target domain classification accuracy of 18.2%, 27.0%, 29.6%, 14.1%, 20.3%, and F1-score of 0.299, 0.516, 0.628, 0.274, 0.413 compared with squeeze-excitation residual network, deep convolution neural networks, ViT, ViT with cross attention, ViT with adaptive attention.