1. Introduction
The traction drive system (TDS) is the only power source of rail transit vehicles, and its energy consumption accounts for about 40~50% of the total energy consumption of rail transit [
1]. Permanent magnet traction drive systems (PMTDSs) have become the key development direction of next-generation rail transit traction drive systems [
2], due to their advantages of low loss and high efficiency [
1]. However, the PMTDS is a high-order complex system with the multi-dimensional coupling of machine electricity–heat magnetism. Hence, it has also become the main source of faults, with increasing service times [
3].
As shown in
Figure 1, to realize the high-performance closed-loop control of the permanent magnet traction motor, three kinds of sensors are installed in PMTDS to measure the U- and V-phase currents, the intermediate DC voltage, and the rotor position signal of the motor [
4]. Sensors in the traction drive system are susceptible to faults due to mechanical vibration, hot and humid conditions, and strong electromagnetic interferences [
5]. Sensor faults can easily lead to control performance degradation and other derivative faults if not detected promptly. Therefore, the study of real-time diagnostic methods of the relevant sensors in PMTDSs has an important engineering application value for improving the reliability and safety of trains.
Fruitful research results have been achieved for the sensor fault diagnosis of PMTDSs [
6]. According to the type of diagnostic objects, these methods can be divided into two categories: single-sensor fault diagnosis and joint fault diagnosis of multiple types of sensors. Regarding single-sensor faults, current sensor faults are diagnosed using different methods, including current estimation [
7], a phase-locked loop-based current reconstruction technique [
8], an adaptive observer [
9], and a sliding-mode observer [
10]. In terms of motor speed or position sensors, diagnostic methods have been proposed based on q-axis current fault characteristics’ analysis [
11], look-up table [
12], Kalman filter [
13], and the speed estimation model [
14].
Compared to single-type sensor fault diagnosis, multi-sensor joint fault diagnosis is relatively less researched, because a more complex multivariate estimation model is required [
15]. Multiple independent observers-based methods [
1,
16], the signal processing-based method [
17], and the data-driven method [
18] have been proposed to achieve the joint diagnosis of intermediate DC voltages, motor currents, and speed sensor faults. However, the above methods have the disadvantages of their computational burden, reliance on the analytic redundancy of three-phase currents, and poor generalization ability, which are not implementation-friendly. The recently proposed convolutional vector fusion network [
19], the semi-supervised matrixed graph-embedding machine [
20], and the non-parallel bounded-support matrix machine [
21] suffer from the interpretability issue. In this paper, the model-based multi-sensor fault joint diagnosis of PMTDSs is investigated to improve its engineering level further.
Structural analysis is a model-based method [
22] that decomposes a complex system into several subsystems. Following this step, the diagnosis of the related faults in a system by mining the set of analytically redundant relations in the system is carried out [
23,
24]. Zhang J. et al. studied the joint diagnosis problem for eight sensor faults in the permanent magnet drive system of electric vehicles, including inverter output three-phase voltage, motor output three-phase current, a motor position sensor, and a vehicle speed sensor [
23]. Ebrahimi S. H. et al. realized the joint diagnosis of the position sensor and the motor [
24]. The diagnosis of 10 sensor signals was investigated, including inverter input DC voltage, inverter output three-phase voltage, motor three-phase current, motor speed, motor position, and load torque, based on the inverter output three-phase voltage, the motor output three-phase current, and the motor position sensor signals in a permanent magnet drive system [
25]. The above multi-sensor joint diagnosis methods based on structural analysis [
23,
24,
25] have added many redundant sensors (e.g., hardware redundancy among three-phase voltage sensors, redundancy between motor speed and position sensors, etc.) to simplify the redundancy relationship of sequence residual analysis. In contrast, such redundant sensors have not been arranged in the actual system considering the cost and reliability factors. Therefore, the above diagnostic methods have some limitations in real-time fault diagnosis tasks.
This paper investigates the real-time multi-sensor joint fault diagnosis of PMTDSs based on the structural analysis method by making full use of the intermediate DC voltage, the motor currents of phases A and B, and the rotor position information collected from the closed-loop control of the PMTDS. The main contributions are as follows:
A real-time joint diagnosis method for the faults of the intermediate DC voltage sensor, the A- and B-phase current sensors, and the position sensor in PMTDSs is proposed;
The detectability and isolability of each sensor fault with limited sampling signals are presented, and residuals are generated by the analytic redundancy relationship. Different combinations of residuals are used to realize the fast and effective isolation of all the sensors.
A diagnostic algorithm test verification method based on data recording to reproduce real fault scenarios is proposed, and a relevant test platform is built to verify the effectiveness of the proposed diagnostic method.
5. Conclusions
Based on the typical sensor layout of PMTDSs in practical engineering, the diagnosability of each sensor fault was hereby demonstrated, and a real-time joint diagnosis method for the multi-sensor fault of permanent magnet traction systems was proposed. This method is based on the structured model of permanent magnet traction systems, including sensor faults, and takes full account of the correlation between the sensor signals to establish the residual. It has the advantages of a clear physical concept, simple implementation, and a small amount of calculation and has good engineering application prospects.
However, due to the use of the IGBT status, the motor parameters, and other information in structured modeling, when an IGBT fault occurs in the system, the change in the motor parameters will affect the diagnosis results. Therefore, our future research will add IGBT faults to the model and increase the real-time identification strategy of the motor parameters to improve the adaptability of the proposed method.