Objective Rotational speed information is a key parameter for the operational condition monitoring and fault diagnosis of traction gearboxes in urban rail transit trains. Due to the compact mechanical structure of train systems and the limited space for sensor deployment, traditional speed sensors often face problems such as high cost, installation difficulty, or even infeasibility in practical applications. It is therefore necessary to study effective methods for extracting rotational speed information, in order to achieve intelligent extraction of rotational speed information from multi-channel vibration signals.

Method The relevant theories of CNN (convolutional neural network) and LSTM (long short-term memory) networks are briefly introduced. An intelligent method for extracting the rotational speed information of train gearboxes based on a WAMF (weight-adaptive multi-scale feature fusion) network is proposed. The rotational speed information extraction process of the WAMF network is analyzed, and the specific structures and functions of the WAP (weight-adaptive) layer and MS (multi-scale feature fusion) module are described in detail. Using a helical gearbox dataset and a fixed-axis gearbox dataset as test cases, FCNN (fully connected neural network), BILSTM (bidirectional long short-term memory) network, CNN, LSTM network, and BILSTM-LSTM-FT network are selected as comparison models, and the rotational speed information extraction results of each model are compared.

Result & Conclusion  The WAMF network can accurately extract rotational speed information from vibration signals and demonstrates significant accuracy and robustness in comparative experiments with five other methods.