Inter-turn short-circuit (ITSC) faults are the most common fault types of brushless DC (BLDC) motors used in industry. Fault diagnosis of BLDC motors is of critical importance due to their wide spread applications. Existing diagnosis methods are based on voltage and current analysis which is useful but difficult to identify early faults and fault locations, as these suffer from external disturbances. This paper proposes a new offline method for fault detection based on magnetic leakage flux (MLF) and backpropagation neural network (BPNN) for improving the level of fault diagnosis. The ITSC and MLF are modeled and analyzed theoretically. Then they are verified by finite element analysis (FEM) and experimental tests. Hall sensors are used to form an array to collect MLF signals at different positions outside the test motor. The frequency-domain characteristic matrix of MLF signals is analyzed by BPNN models. The experimental results show that the proposed method can effectively detect ITSCs, and estimate the fault degree and the location of the fault. The method is a promising technology as it is non-intrusive and accurate.

Predicting the degradation of mechanical components, such as rolling bearings is critical to the proper monitoring of the condition of mechanical equipment. A new method, based on a long short-term memory network (LSTM) algorithm, has been developed to improve the accuracy of degradation prediction. The model parameters are optimized via improved particle swarm optimization (IPSO). Regarding how this applies to the rolling bearings, firstly, multi-dimension feature parameters are extracted from the bearing’s vibration signals and fused into responsive features by using the kernel joint approximate diagonalization of eigen-matrices (KJADE) method. Then, the between-class and within-class scatter (SS) are calculated to develop performance degradation indicators. Since network model parameters influence the predictive accuracy of the LSTM model, an IPSO algorithm is used to obtain the optimal prediction model via the LSTM model parameters’ optimization. Finally, the LSTM model, with said optimal parameters, was used to predict the degradation trend of the bearing’s performance. The experiment’s results show that the proposed method can effectively identify the trends of degradation and performance. Moreover, the predictive accuracy of this proposed method is greater than that of the extreme learning machine (ELM) and support vector regression (SVR), which are the algorithms conventionally used in degradation modeling.

A simulation approach was proposed in this work for monitoring fault status of cracked gear systems. The approach calculates the dynamic response signals from a dynamic model of the cracked gear system. First, time-varying meshing stiffness (TVMS) of the cracked gear pair (CGP) was computed by slicing method considered the coupling effect. Second, a novel 12 DOF dynamic model was constructed by using TVMS and tooth non-equilibrium moment. Subsequently, fault features of the cracked gears were extracted by investigation of the dynamic response. The results demonstrate the proposed method effectively monitors cracked gear system condition, and the condition indicators were sensitive to the cracked status and system resonance.