Hybrid Kalman Filtering and Deep Neural Sequences for Robust Motor Fault Identifications

This study proposes a new motor fault recognition algorithm that integrates extended Kalman filtering and deep neural time series modeling. This algorithm can improve the system’s state perception and fault discrimination capabilities in nonlinear dynamic environments. This method constructs an extended Kalman filter model driven by an error state, which can enhance the adaptability to system disturbances and model uncertainties. At the same time, a long short-term memory network (LSTM) is introduced to achieve dynamic estimation of process noise covariance. This operation can make up for the dependence of traditional filtering methods on static modeling. In terms of feature modeling, a multi-scale one-dimensional convolution structure is designed to perform deep extraction of time series data and combine the residual gated recurrent unit (GRU) module to achieve global representation and non-stationary feature fusion. In order to improve the robustness of discrimination, this study proposes a weighted discrimination mechanism that integrates state estimation and deep features. In the experiment part, a data set containing five working conditions (normal, bearing fault, stator short circuit, rotor bar breakage, and load abnormality) was constructed. The actual permanent magnet synchronous motor platform was used to collect time domain signals, and ±8% load disturbance, ±6°C temperature drift, and ±3% voltage offset were introduced to simulate real industrial interference. The comparative experiment selected three types of models as baselines, including traditional EKF+SVM, CNN+GRU, and BiLSTM+Attention structures. The results show that the method proposed in this study has achieved the best performance in terms of recognition accuracy, F1 score and other indicators. From a numerical perspective, the accuracy reached 96.4%, and the F1 value reached 95.5%.