A General Intelligent Fault Diagnosis Framework Based on Incomplete Vibration Signals With Missing Values

Rolling bearings are essential components in the operation of machinery. In real-world scenarios, the collected vibration signals are usually incomplete, leading to incorrect fault diagnosis results. Existing methods typically involve two separate steps: signal recovery and fault classification. Moreover, signal repair errors further affect the fault diagnosis performance. To address these limitations and simplify the fault diagnosis process, this article proposes a general intelligent bearing fault diagnosis framework based on incomplete vibration signals. Specifically, it consists of two components. The first component is a locally connected neural network (LC-NN) block that enables local connections between the input and hidden layers, which allows for extracting local shift-invariant features. The second component is a feature optimization (FO) block that adaptively adjusts the local connection degree, reducing the impact of missing values on the extracted features. By integrating these two components, the proposed framework empowers existing fault diagnosis models to mitigate the sensitivity to missing values, thus improving the fault diagnosis performance without additional signal recovery. Extensive experiments are conducted to verify the effectiveness of the proposed method.