Abstract: Microelectromechanical systems (MEMS) inertial measurement units (IMUs), renowned for their compact size, low cost, and low power consumption, have found extensive applications in military, industrial, and civilian sectors. However, single MEMS IMUs are limited by measurement accuracy and reliability, making them inadequate for high-precision navigation requirements. IMU array technology, which integrates multiple MEMS inertial sensors with information fusion methods, offers an effective solution to enhance navigation system performance. This paper systematically reviews the development history and current research status of MEMS IMU array technology, providing an in-depth analysis of key technologies across three dimensions: error calibration, data fusion, and fault detection. Furthermore, it outlines future directions including high-precision, low-cost calibration; large-scale array integration; intelligent fault diagnosis; and industrial applications, offering valuable insights for advancing IMU array technology in high-precision navigation systems.
Keywords: IMU array; data fusion; error calibration; fault detection
Einführung
With the rapid development of microelectromechanical systems (MEMS) technology, inertial measurement units (IMUs) have become smaller, more cost-effective, and more power-efficient. MEMS IMUs are widely used in aerospace, drone navigation, autonomous systems, and other applications. However, individual MEMS IMUs still face challenges such as noise, bias instability, and environmental sensitivity, which limit their performance in high-precision navigation.
To address these limitations, IMU array technology integrates multiple MEMS inertial sensors with advanced data fusion algorithms. By combining redundant sensor outputs, IMU arrays can effectively reduce noise, suppress drift, and improve navigation accuracy and reliability compared with single IMUs. With continuous advances in calibration, sensor fusion, and fault detection technologies, IMU arrays have gradually developed from theoretical concepts into practical solutions for high-performance navigation systems.
This article reviews the evolution, key technologies, and future trends of MEMS IMU array technology. At Andelu, we focus on advancing inertial sensing and fusion technologies to develop reliable INS solutions for applications that require stable and accurate navigation performance.
Overview of IMU Array Technology Development
The Conceptual Evolution from Virtual Gyroscope to Array Fusion
The theoretical origins of IMU array technology can be traced back to the “virtual gyroscope” concept proposed by David S. Bayard at NASA laboratories. From an information fusion perspective, Bayard demonstrated that integrating outputs from multiple MEMS gyroscopes significantly enhances system accuracy—simulation results showed a 173-fold improvement. The underlying mechanism is that when multiple gyroscopes are arranged in an array, independent noise from each sensor cancels out upon integration, substantially reducing drift errors.
Building on this foundation, Chinese scholar Hu Min proposed a Kalman-filter-based fusion scheme for gyro arrays, integrating multiple MEMS gyroscopes into a high-precision virtual gyroscope system that achieved approximately a 233-fold improvement in drift performance through experimental testing. Subsequently, researchers successively introduced concepts such as MEMS virtual gyroscope arrays, accelerometer arrays, inertial sensor arrays, and IMU arrays, gradually establishing a comprehensive IMU array technology framework.
Expansion of Array Size and Methodological Evolution
The evolution of IMU array technology has exhibited a clear progression from small to large array sizes. Early studies predominantly employed compact arrays comprising 4–6 sensors, primarily to validate the effectiveness of fusion algorithms. As demands for higher accuracy and reliability increased, array sizes continued to expand. A research team at the University of Michigan integrated 72 MEMS gyroscopes onto a three-layer development board and utilized a Markov model for fusion, achieving a 50% improvement in accuracy compared to Kalman filtering.
Meanwhile, fusion methods have evolved from simple weighting to sophisticated estimation techniques, and from linear to nonlinear approaches. Methods such as the weighted average method, least squares method, Kalman filtering and its extensions, and factor graph optimization have been successively introduced. In recent years, the integration of artificial intelligence technologies has revitalized IMU array fusion; hybrid architectures combining neural networks with Kalman filtering demonstrate significant advantages in fault tolerance and accuracy improvement.
From Laboratory Research to Engineering Applications
IMU array technology has achieved a pivotal transition from theoretical exploration to engineering implementation. In terms of technological maturity, FPGA-based vector processor designs have resolved scalability challenges in array fusion; at the application level, high-precision integrated navigation systems, micro/nanosatellite attitude measurement systems, and pedestrian navigation systems have been successively developed. These practical applications demonstrate that IMU arrays not only excel in high-precision navigation but also hold significant potential for wearable devices and portable systems.
Key Technologies of IMU Arrays
Error Analysis, Modeling and Calibration
MEMS inertial sensors, particularly low-cost consumer-grade models, are susceptible to processing errors in their micro-mechanical structures (such as mass blocks and cantilever beams) due to manufacturing limitations. These errors can lead to non-ideal orthogonality of the sensitive axis and shifts in resonant frequency. Consequently, precise error analysis and modeling of IMU arrays are critical for data fusion. The primary sources of error in IMU arrays include inertial device errors (e.g., zero bias, noise, and proportionality factor errors) and installation errors—the deviation between the actual mounting positions of gyroscopes and accelerometers and their designed positions.
In error modeling, research on IMU array error modeling has advanced toward higher precision. Taking the first-order Markov stochastic process model as an example, this approach describes time-dependent random error components such as bias instability and temperature drift. The model effectively captures the random drift characteristics of IMUs by employing time-series modeling techniques to precisely characterize the stochastic evolution of sensor zero bias over time. Additionally, the Allan variance serves as a time-domain statistical method for analyzing and quantifying random error characteristics in inertial sensors (e.g., MEMS gyroscopes and accelerometers). It separates and identifies various random noise components in sensor outputs—including angular random walk, zero-bias instability, and rate random walk—by examining noise at different temporal scales. Furthermore, researchers have developed innovative methods for IMU array error modeling, such as utilizing novel universal multi-sensor integration strategies and integrated error analysis frameworks in discrete Kalman filtering. These approaches enable characterization of individual sensor errors (deviation/proportional factor errors) within the IMU array, apply variance component estimation to array data, and employ separate variance components to evaluate the performance of each IMU’s gyroscope and accelerometer. This methodology facilitates time-varying estimation of errors across all individual sensors and enables rigorous fault detection and mitigation for anomalous measurements across the entire sensor array.
The calibration process forms the foundation for ensuring the final measurement accuracy of IMU arrays. At its core, calibration involves determining a set of parameters by comparing measurements against known reference inputs, thereby aligning the outputs of inertial devices with these references. Calibration methods are generally categorized into turntable-based and turntable-free approaches. Turntable-based methods typically employ high-or low-precision turntables to calibrate gyroscopes and accelerometers through velocity tests and multi-position static measurements. In contrast, turntable-free calibration relies on natural signals such as Earth’s rotational angular velocity and gravitational acceleration, utilizing both static and dynamic data to estimate calibration parameters—a method particularly suitable for cost-effective applications.
Data Fusion Technology
The fundamental methods for IMU array data fusion include the weighted average method, least squares method, and Kalman filtering. The weighted average method assigns different weights to individual sensors to reflect their reliability, but weight determination relies on prior knowledge. The least squares method and its weighted sum recursive variant improve fusion performance to some extent, yet they are inherently batch processing approaches that struggle to adapt to dynamic environments. Kalman filtering, due to its recursive structure and optimal estimation properties, has become the most critical and widely adopted method for IMU array fusion. Its core principle involves predicting state estimates using a dynamic model and updating them with real-time observations to achieve optimal fusion results. Research indicates that analyzing noise correlations (particularly negative correlations) in MEMS gyroscope arrays enables optimization of the covariance matrix parameters in Kalman filtering, significantly enhancing the performance metrics of gyroscope array systems.
To address noise suppression in dynamic environments, an efficient estimation algorithm based on wavelet compression fusion employs multi-scale wavelet transforms to compress the original signal and applies thresholding techniques to eliminate noise, thereby improving the gyroscope’s zero-bias instability, angular random walk, and rate slope by varying degrees.
The introduction of artificial intelligence technology has opened new avenues for IMU array data fusion. The integrated algorithm combining neural networks and Kalman filtering employs LSTM networks to calculate gyroscope confidence levels while using BP networks to identify and eliminate faulty sensor data. The LSTM-based error calibration and compensation method applied to the 16-IMU array reduces bias instability and angular random walk by over 50%. These studies demonstrate that AI approaches, through temporal modeling to capture long-term dependencies and implement online dynamic parameter optimization strategies, significantly enhance the fusion accuracy and adaptive capabilities of IMU arrays in complex noisy and time-varying environments.
Fault Detection and Isolation
The IMU array consists of multiple inertial sensors, and sensor failures are difficult to completely avoid. Once a failure occurs, it results in unreliable measurement data, thereby compromising navigation performance. Accurate identification of sensor failure types and timely diagnosis and isolation are crucial for enhancing system stability and reliability.
The development of fault detection and isolation technologies has evolved from traditional decision-theoretic approaches to intelligent algorithms. Early methods included direct comparison, generalized likelihood ratio, optimal parity testing, and singular value decomposition; however, these techniques exhibit high computational complexity when handling large-scale arrays and demonstrate limited adaptability to dynamic environments.
For large-scale IMU arrays, the KNN-based FDIR architecture rapidly identifies abnormal data by comparing real-time sensor measurement differences and eliminates faulty sensors through tolerance settings. The adaptive gradient fault-tolerance method based on fuzzy decision-making detects faults via fuzzy evaluation of odd-even residual quality, tolerating gradient failures in up to two gyroscopes.
In recent years, fault diagnosis methods integrated with deep learning have garnered significant attention. The IMU fault diagnosis approach based on deep residual networks significantly enhances the accuracy and efficiency of fault isolation by learning sensor characteristics under both normal and faulty conditions. The adaptive fault detection system combining autoencoders and support vector machines automatically extracts nonlinear features and classifies faults, enabling effective handling of various fault patterns.
Future Development Trends and Prospects
High-Precision, Low-Cost Calibration Technology
Future calibration technologies will evolve toward higher precision, lower costs, and greater adaptability. Self-calibration techniques require further integration with dynamic characteristics to optimize calibration algorithms, enabling real-time online dynamic calibration. Meanwhile, low-cost calibration solutions for consumer-grade applications hold significant engineering value—achieving acceptable calibration accuracy without relying on high-precision turntables is key to expanding their application scope.
Deep Optimization of Data Fusion Algorithms
As the core of IMU arrays, data fusion requires further advancement in the following areas: first, hybrid fusion strategies that integrate Kalman filtering and its extended methods with neural networks to enhance accuracy and robustness; second, efficient fusion algorithms for large-scale arrays that balance computational complexity with existing fusion performance; third, batch processing techniques such as factor graph optimization evolving toward online and lightweight implementations to meet real-time application demands.
Intelligent Fault Diagnosis and Isolation
The data complexity and information redundancy inherent in large-scale arrays impose higher demands on fault detection technologies. Future efforts should focus on developing diagnostic methods characterized by high stability, precision, and rapid response capabilities, with particular emphasis on: end-to-end fault diagnosis models based on deep learning; hybrid diagnostic architectures that integrate physical prior knowledge; and adaptive fault detection mechanisms utilizing online incremental learning. By establishing large-scale fault databases and model training frameworks, the accuracy of detection and real-time response capabilities can be significantly enhanced.
From Single-Point Breakthrough to Systematic Collaboration
The future evolution of IMU array technology will exhibit a shift from isolated breakthroughs to a systematic paradigm: bridging the computational power gap through synergistic optimization of algorithms and hardware, mitigating the generalization crisis via model-data fusion-driven intelligent approaches, and overcoming integration limitations by leveraging comprehensive multi-source opportunity signals. The ultimate goal is to establish a trustworthy spatiotemporal benchmark for complex near-ground scenarios, providing high-precision, highly reliable navigation support for autonomous driving, the low-altitude economy, and smart manufacturing sectors.
References
[1] Chen Yihao, Liu Jieyu, Wang Lixin, et al. Research Progress on Inertial Measurement Unit Array Technology for Micro-Electromechanical Systems [J]. Sensors and Microsystems, 2025,44(3):422-430.
[2] Hang Yijun. Research on Error Compensation and Information Fusion Techniques for Micro-Navigation Systems in Multi-Rotor Drones [D]. Nanjing: Nanjing University of Aeronautics and Astronautics, 2017.
[3] Guo Jian, Huang Junru, Li Juntao, et al. Advances in Composite Navigation Technology: Synergistic Enhancement of Heterogeneous Complementarity and Homogeneous Redundancy [J]. Journal of Transportation Engineering, 2026.
[4] Bayard D S. Reduction of drift in gyro arrays[C]. AIAA Guidance, Navigation, and Control Conference, 2003.
[5] Wang J, Olson E. High-performance inertial measurements using a redundant array of inexpensive gyroscopes (RAIG)[C]. IEEE International Conference on Multisensor Fusion and Integration, 2015: 71-76.
[6] Xue L, Wang X G, Yang B, [6] Xue L, Wang X G, Yang B, et al. Analysis of correlation in MEMS gyroscope array and its influence on accuracy improvement for the combined angular rate signal[J]. Micromachines, 2018, 9(1): 22.
[7] Miao Lingjuan, Zhang Weixiao, Zhou Zhiqiang, et al. Fusion algorithm for gyro array based on neural networks and Kalman filtering [J]. China Journal of Inertial Technology, 2023,31(5):501-509.
[8] Zhang T S, Yuan M, Wang L Q, [8] Zhang T S, Yuan M, Wang L Q, et al. A robust and efficient IMU array/GNSS data fusion algorithm[J]. IEEE Sensors Journal, 2024, 24(16): 26278-26289.
[9] Mounier E, Karaim M, Korenberg M, [9] Mounier E, Karaim M, Korenberg M, et al. Multi-IMU system for robust inertial navigation: Kalman filters and differential evolution-based fault detection and isolation[J]. IEEE Sensors Journal, 2025, 25(6): 9998-10014.