A time-series deep learning algorithm for motor fault diagnosis integrating multisource data such as vibration and current

Abstract

As a core power component in modern industrial production, the stable and reliable operation of electric motors directly determines the safety and operational efficiency of the entire system. The early and accurate identification of faults is key to minimizing downtime losses and ensuring the safety of equipment operation and maintenance. Addressing the engineering challenges posed by traditional diagnostic methods—such as the susceptibility of single vibration or current signals to industrial noise interference, as well as the weak and incomplete nature of fault characteristics, which result in poor generalization capabilities and insufficient diagnostic accuracy under complex operating conditions—this paper proposes a time-series deep learning algorithm for motor fault diagnosis that integrates multi-source data, including vibration and current signals. First, we design a process for the synchronized acquisition and preprocessing of multi-source time-series data. Through detrending, adaptive noise reduction, and timestamp alignment, we address the heterogeneity of vibration and current signals in terms of sampling frequency and physical dimensions, thereby preserving the temporal evolution characteristics of faults; Second, a deep learning framework based on a temporal attention mechanism is constructed. A hybrid structure combining an improved gated recurrent unit (GRU) with a convolutional neural network (CNN) is introduced to capture long-term temporal dependencies and local fault features in the multi-source data, respectively. Through a feature-level fusion strategy, complementary synergies among different data sources are achieved, while interference from redundant information is suppressed; Finally, on a standard motor fault test platform, typical fault conditions such as bearing wear and winding inter-turn short circuits were selected. Comparative experiments were conducted using field measurement data from industrial sites to validate the algorithm’s performance. The experimental results demonstrate that, under complex conditions such as low signal-to-noise ratios and variable loads, the proposed algorithm achieves higher fault diagnosis accuracy compared to methods relying on a single data source.