1. Introduction
Micro-electro-mechanical system inertial measurement unit (MEMS IMU) has the advantages of being low cost, small volume, and lightweight, but low precision and poor stability are still problems that cannot be ignored. Many aircraft, cars, and satellites are beginning to use MEMS sensors [
1,
2,
3,
4,
5]. Among them, most MEMS accelerators have already met the tactical and navigation grade requirements [
6,
7]. However, only a few MEMS gyros can reach the level of tactical grade [
1], and these tactical grade gyros are too expensive for common application. To improve the performance of MEMS IMU, a series of studies have been conducted to promote MEMS manufacturing technology. On the other hand, the improvement can also be achieved through reasonable designs and algorithms.
Under the restriction of current manufacturing technology, a lot of research has been done to improve the performance of the MEMS system, and redundant design is the most effective approach to improve the accuracy and stability [
8,
9,
10,
11,
12,
13]. Even for the tactical or navigation grade sensors, the redundant configuration can improve the performance substantially [
14]. In the redundant inertial navigation system (RINS), the processing of redundant data is the most important link. It is divided into two parts: processing data anomalies and fusing redundant data. The motivation of this paper is to develop the anomaly diagnosis method, and improve the reliability of the MEMS RINS.
The data anomalies are comprised of sensor faults and outliers. The faults refer to the failure and variation of sensor characteristics, which are probably caused by harsh environmental conditions, such as the moisture, extreme temperature, electromagnetic radiation, impact, and vibration [
15]. These kind anomalies are related to the aging, damage, and performance degradation of the sensors, and they are generally irreversible hard failures. The definition of an outlier is an observation (or subset of observations) which appears to be inconsistent with the remainder of that set of data. The outliers in inertial measurement data arise because of instrument error, natural deviations, or changes in the behavior of systems. The outliers usually appear as the self-recovery and transient mutations [
16].
For further analysis, the anomalies can be divided into more subcategories based on a different baseline. In this paper, anomaly patterns serve as evidence to choose anomaly solution methods. So we classify anomalies as short-duration auto-recovery anomalies and sustained anomalies. The former includes isolated outliers, small outlier patches, transient faults, etc., which refer to the anomalies that sustain a short duration and recover spontaneously. The short-duration auto-recovery anomalies can be corrected well with outlier eliminating methods. The latter includes large outlier patches, hard faults, permanent failures, etc. These kinds of anomalies can be subdivided by their different performance characteristics. The constant drift fault and multiplicative fault (variation of scale factor) can be repaired through re-calibration technology [
17,
18]. For instance, the faulty sensors can be calibrated with the data of other types of navigation information or the fusion estimation signal without anomaly data. However, large outlier patches, noise failure, complete failure, and other hard faults are usually isolated directly. Overall, the reclassification of anomalies is shown in
Figure 1.
At present, most methods only divide the anomalies into faults and outliers to process. Due to the differences between faults and outliers, the countermeasures against them are carried out separately in different ways. To protect the output fusion signal from faults, the fault detection and isolation (FDI) methods can make the system continue working normally [
19,
20,
21,
22]. The solution against the faults is to detect the measurement data in real time, isolate faulty sensors and discontinue using them. A set of FDI methods have already been applied in the RINS successfully, which includes the Chi-square test [
23,
24], the network-based method [
25], and the parity space method. The parity space method includes the optimal parity vector test (OPT) [
26], the generalized likelihood test (GLT) [
27,
28], and the singular value decomposition (SVD) method [
21,
22,
29,
30,
31]. Among these methods, the methods based on the Chi-square test are subjected to the filtering algorithm performance. When the filtering performance is degraded or even diverged, the detection and isolation results are not reliable. The neural network-based methods are restricted by the heavy computing burden and the requirements for large samples. In the main, the parity space method is a good choice for the navigation system, but it cannot handle the situation of multi faults. As for the outliers, there are two approaches commonly used to deal with them: one is to improve filter functions, which alleviate the impact of outliers by setting the weighted matrix or correcting the covariance matrix at every sampling moment [
32,
33,
34,
35]; another one is to replace the detected outlier points with mean values or the points from fitted curves [
36,
37].
The methods introduced above can reduce the effect of anomalies to some degree, but there are still some problems. Firstly, the outliers in the measurement data, especially the outlier patches, are easily misdiagnosed as the faults, which will lead to an increase in the false alarm rate. The outliers will cause the detection function values to exceed the detection threshold, but there is no failure or damage in the sensor with outliers. Such misdiagnosis may induce a well-behaved sensor to be deactivated. Secondly, when the outlier eliminating methods and FDI methods are operating simultaneously, the practical performance may degrade, because the calculation burden will increase and the results of different methods will interfere with each other. For example, the outlier eliminating methods regard every anomaly point as an outlier. Since the deviation of anomaly points caused by sensor faults will be diminished after the outlier eliminating processing, the original statistical characteristics of the faults will change. As a result, the probability of correct detection (PCD) and the probability of correct isolation (PCI) will decline. Thirdly, the FDI methods cannot identify the patterns of anomalies. So the different modes of anomalies have to be handled in the same way. Deactivating sensors with hard faults is reasonable. However, some faulty sensors are not physically damaged and can continue to be used after correction, so traditional strategies can possibly result in the loss of available sensors.
To solve the problems above, this paper proposed a diagnosis method for data anomalies in MEMS RINS. The SVD-based method is briefly reviewed, and the flaws of this method are analyzed. The improved SVD-based method is presented. The disadvantages of this method are analyzed and improved. To demonstrate the feasibility of recognizing a fault pattern by analyzing the detection function, the mathematical relation between the detection function and sensor faults is derived. Five indicators of detection are defined on the basis of the analysis of various anomalies. The anomaly recognition method is proposed according to these five indicators. The proposed method can detect faults and outliers simultaneously, so that, the mutual interference between outlier eliminating and fault detection can be removed. The anomalies can receive suitable and reasonable treatments based on the recognition results. The anomaly processing is simplified too.
This paper is organized as follows: In
Section 2, the SVD method is briefly reviewed, and the disadvantages are discussed and improved. To demonstrate the feasibility of using detection function values as the evidence for anomaly recognition, the relation between the detection function and anomalies is deduced. In
Section 3, the five indicators are designed to analyze different anomalies quantitatively, and the anomaly detection and recognition method are presented. In
Section 4, the simulation experiments are conducted to test the proposed method.
Section 5 gives the conclusion.