INTELLIGENT ROLLING BEARING FAULT DIAGNOSIS UNDER SENSOR CHANNEL FAILURE USING IMAGE FUSION AND CNN–BILSTM DEEP LEARNING NETWORKS

Abstract

Rolling bearings are essential components in large-scale and complex mechanical systems, where their operational reliability directly affects overall equipment safety. Under harsh working conditions, bearing fault diagnosis is often challenged by weak fault characteristics in one-dimensional vibration signals and performance degradation caused by sensor channel failures. To overcome these limitations, this study proposes an intelligent and adaptive rolling bearing fault diagnosis framework that integrates multi-channel image fusion with deep learning techniques. Initially, a threshold-driven channel assessment mechanism is introduced to identify and eliminate abnormal or failed sensing channels, followed by a bootstrap aggregation strategy to select representative and high-quality signal channels. Subsequently, vibration signals are transformed into two-dimensional feature representations using Gramian Angular Field encoding, and a multi-channel image fusion scheme is employed to enhance fault-related information. For efficient feature learning and classification, a convolutional neural network is adopted to extract discriminative spatial features and perform dimensionality reduction, while a bidirectional long short-term memory network is utilized to capture temporal dependencies and establish the relationship between extracted features and fault categories. The effectiveness and robustness of the proposed framework are evaluated through extensive simulation studies based on a benchmark rolling bearing vibration dataset with simulated sensor channel failure scenarios. Simulation results indicate that the proposed method achieves superior diagnostic accuracy and robustness compared with conventional CNN-, LSTM-, BiLSTM-, and CNN–LSTM-based approaches. These findings demonstrate the potential of the proposed framework for reliable rolling bearing fault diagnosis in complex sensing environments with incomplete or unreliable measurement channels.