Pilot Contamination in Massive MIMO Systems: Challenges and Future Prospects
^††thanks: This research was partly funded by Palmer Department Chair Endowment Funds at Iowa State University.

Massive multiple input multiple output (M-MIMO) has been widely regarded as an essential component of fifth-generation (5G) and beyond communication technologies that enables several benefits spanning from increased spectral efficiency (SE) to enhanced energy efficiency and increased reliability. However, these benefits are contingent upon the presence of accurate channel state information (CSI) at the base station (BS). Timely acquisition of CSI at the BS is crucial for maximizing network throughput. However, acquiring CSI is often challenging, primarily due to small coherence intervals.

Coherence interval refers to a time duration wherein the channel response remains frequency-flat and time-invariant. Generally, the coherence interval is shorter in scenarios characterized by rapid changes in the wireless channel, such as high mobility, narrow bandwidth, change in the propagation environment, etc. This limitation results in reusing the same pilot sequences in adjacent cells, preventing orthogonal distribution to all users. This introduces coherent interference into channel estimation, as multiple devices transmit the same pilot sequence, making it challenging for the BS to distinguish each device’s pilot sequences. This phenomenon is called pilot contamination. Pilot contamination is considered a critical limiting factor in enhancing SE [1].

Though a comprehensive survey article [2] on mitigating pilot contamination was published in 2015, most discussed schemes are now outdated. Therefore, this paper summarizes the most recent developments in addressing pilot contamination, driven by rapid research progress. It categorizes the pilot decontamination schemes into three broader categories: pilot assignment schemes, advanced signal processing methods, and advanced channel estimation techniques. Pilot assignment schemes intelligently allocate pilot sequences to minimize interference and ultimately maximize SE [3], [5], [9]. Advanced signal processing strategies introduce innovative pilot transmission and signal processing techniques that effectively mitigate interference and enhance throughput [13], [14], [17]. Pilot decontamination through advanced channel estimation techniques, such as deep learning, aids in mitigating the mean square error (MSE) of the channel estimate, eventually improving CSI and enhancing SE. This paper analyzes these techniques, compares and contrasts using key features and performance indicators.

The time-frequency is segmented into coherence intervals/blocks, $\tau_{c}$, during which channel response remains time-invariant and frequency-flat. Within each block, a designated interval, $\tau_{\rho}$, is reserved for transmitting predefined pilot signals during uplink transmission. Pilot signals with predefined patterns are transmitted, enabling the BS to compare the transmitted and received values, as the received pilot signal may contain valuable information, such as multi-path fading, interference, and noise, as it transverses the wireless channel.

This section explores representative pilot assignment strategies to mitigate pilot contamination in M-MIMO systems.

This section reviews techniques aimed at enhancing SE using the latest signal processing advancements.

Deep Learning (DL) offers a promising solution for complex wireless communication challenges, boosting the system’s performance with low computational complexity in tasks like resource allocation, channel decoding, and channel estimation. Xu et al. [18] proposed a DL-based pilot design approach for M-MIMO systems, addressing the challenge of optimizing power allocation for each user’s pilot sequence to minimize the cumulative mean square error in channel estimation. Their solution involved a deep neural network that mapped the input data (large-scale fading coefficients) to output data (pilot power allocation vector). Simulation results claimed that the proposed scheme surpassed conventional pilot decontamination methods while having lower computational complexity. One limitation of the paper was its reliance on unsupervised learning for pilot power allocation, which may not guarantee optimal solutions.

Therefore, Hirose et al. [19] presented two innovative channel estimation techniques: one employed a neural network (NN), and the other utilized a convolutional neural network (CNN). The NN-based approach relied on fully connected layers, focusing on extracting spatial information from the least squares estimated channel. In contrast, the CNN took advantage of the user’s spatial correlation using sliding convolutional filters. Their results demonstrated that CNN-based estimation outperformed the NN-based alternative in accuracy, despite slightly longer training times for datasets. However, the paper might have lacked a thorough investigation into the scalability of the proposed DL-based channel estimation method for M-MIMO systems with a large number of antennas or users.

In [20], the authors proposed a low-complexity channel estimation technique for M-MIMO systems by using a deep neural network (DNN) for denoising before the conventional least squares operation, enhancing performance through a tailored DNN architecture. To maintain low complexity, they reconfigured this architecture’s input and output layers and optimized the number of parameters, substantially reducing it from millions to just a few thousand parameters. While the proposed scheme demonstrated favorable outcomes, the paper might not have fully explored the computational complexity and resource requirements associated with deploying untrained DNNs in real-time M-MIMO systems, which could have been crucial factors for practical implementation.

Lim et al. [21] proposed a method to mitigate pilot contamination in wireless systems using DL. They introduced an NN-based pilot design to minimize mean square error, augmented with unsupervised learning to remove the necessity of using pilot samples during the uplink training phase. Additionally, they utilized deep residual learning for the channel estimator to reduce distortion noise caused by pilot contamination and to reconstruct the original channel. Finally, they introduced a novel pilot design and channel estimator using transfer learning, which outperformed conventional linear estimators, even in the presence of hardware impairments, showing its capability to mitigate interference without prior knowledge.

Pilot contamination is a recognized barrier to improving SE in M-MIMO systems. Table 1 provides a comparative analysis of schemes based on major contribution, system considered, channel assumption, combining/precoding scheme, computational complexity, pilot overhead, and scalability. The major contributions of the proposed schemes include enhancing throughput through intelligent pilot assignment or reuse [3], reducing pilot overhead[13], and improving channel estimation via advanced techniques [18]. Moreover, the proposed schemes are evaluated in two scenarios: cell-free massive MIMO, utilizing distributed antennas without dedicated cell boundaries, and multicell massive MIMO, which employs traditional cell-based structures with coordinated interference control. Furthermore, most papers consider correlated and uncorrelated Rayleigh fading channel assumptions. The key difference is that uncorrelated channels assume an equal signal distribution in all directions, which is impractical, while correlated channel assumptions suggest variations in signal presence among different directions.

Different schemes adopt specific combining/precoding techniques based on their complexity and accuracy requirements. Precoding schemes play crucial roles in signal transmission optimization. Match filter (MF), maximum ratio (MR), and zero-forcing (ZF) methods are foundational techniques. Additionally, regularized zero-forcing (RZF) and minimum mean squared error (MMSE) precoding schemes offer advanced capabilities for improved system performance but exhibit higher complexity than conventional schemes [2]. Generally, channel estimation requires a minimum of $\tau_{p}$ pilot symbols, with $\tau_{p}\geq K$, often reused in neighboring cells. We categorize pilot overhead as ’increased’ for schemes that introduce pilot sequence orthogonality between users of neighboring cells, $\tau_{p}>K$, ’decreased’ for schemes where $\tau_{p}<K$, fewer pilot symbols than $K$ while accommodating all users, and ’not addressed’ means a typical pilot reuse scenario. Lastly, the scalability criterion assesses the scheme’s ability to accommodate a growing number of devices in a cell while considering the limitation imposed by the coherence block length constraint.

Smart pilot assignment schemes [3], [4] offer simplicity and efficiency but come with high computation costs. Moreover, graph coloring schemes [5], [7] minimize the use of orthogonal pilot signals, yet their effectiveness heavily relies on interference graph criteria. Additionally, soft pilot reuse [9] efficiently assigns orthogonal pilot sequences between cells’ center and edge users, enhancing performance, but faces challenges due to high pilot signal overhead. Furthermore, the angle of arrival [11] considers the signal’s direction as a criterion to distinguish pilot signals of multiple users, but in multi-path scenarios, distinguishing between multiple user angles becomes challenging. Overall, in the ongoing efforts to combat pilot contamination within wireless communication systems, pilot assignment strategies have emerged as pioneering solutions and ongoing research is dedicated to refining their effectiveness [8].

Additionally, introducing the superimposed pilot schemes [14], [13] has addressed the issue of pilot signal length by transmitting pilots alongside data, significantly mitigating pilot contamination. Furthermore, RSMA [16], [17] has effectively reduced interference by segregating pilot signals into common and private components, allowing users to decode their data with minimal interference from other users’ pilots. Overall, superimposed pilots and RSMA significantly reduce pilot overhead, ultimately increasing spectral efficiency. However, both schemes require advanced signal processing techniques, consequently necessitating high-end hardware support [16].

Furthermore, the increasing utilization of advanced machine learning techniques like DL and CNN is playing a growing role in enhancing pilot design and channel estimation accuracy. A DL-based pilot design approach [18] showed promise in optimizing power allocation, but concerns persist regarding its reliance on unsupervised learning. Another [19] introduced NN and CNN-based channel estimation techniques, with CNN demonstrating higher accuracy but potential scalability issues. Additionally, a low-complexity DNN-based channel estimation method [20] was presented, with concerns about computational complexity. Lastly, a DL-based approach [21] to mitigate pilot contamination was proposed, showing superior performance even in the presence of hardware impairments, highlighting its interference mitigation potential without prior knowledge. These combined advancements emphasize the dynamic evolution of pilot contamination mitigation strategies, with a clear focus on optimizing SE and efficient resource utilization.

Future research into resource-efficient pilot schemes, such as time-multiplexed pilots [3], [7], superimposed pilots [13], and rate splitting [16], could continue evolving due to their potential to effectively mitigate interference. These schemes offer opportunities to optimize the allocation of resources for both pilot and data transmission, thereby reducing overhead and maximizing throughput. Further exploration of these techniques enables researchers to delve into the integration of the aforementioned schemes and fine-tuning resource management for optimal efficiency, thereby improving SE.

The increasing demand for wireless connectivity, coupled with the need for a growing number of orthogonal pilot sequences for channel estimation, presents a challenge due to the coherence block length constraint [7]. In forthcoming research, it is imperative to prioritize the advancement of intelligent user scheduling schemes [8] to effectively manage the projected exponential growth in the number of devices in the coming years. These intelligent scheduling schemes will reduce pilot overhead, address pilot contamination, and enable scalability, ultimately improving the system’s throughput.

In the future, incorporating reinforcement learning [12] into pilot assignments entails the creation of an adaptive framework where an agent adjusts pilot assignments based on real-time network conditions, including mobility, interference, and channel quality. The agent’s actions, guided by a reward function, aim to optimize throughput and minimize interference. Training methods, such as Q-learning, can be utilized to balance exploration and exploitation, enabling effective pilot assignments.

Furthermore, advanced research should focus on joint pilot design and channel estimation techniques [21], using specialized deep neural network architectures and optimization algorithms to concurrently optimize pilot sequences and estimate channels in dynamic wireless environments. This research would create real-time, adaptable solutions seamlessly integrating pilot optimization and channel estimation to address dynamic fading channels and interference challenges.

M-MIMO is crucial for 5G and beyond technologies and offers increased SE but relies on accurate channel information, while pilot contamination remains a significant challenge. Pilot contamination arises from short coherence intervals in dynamic wireless environments, causing interference to/from neighboring cells and reducing SE. This paper briefly discusses various mitigation strategies, such as intelligent pilot allocation schemes, interference mitigation schemes, and advanced techniques like deep learning, with the aim of reducing interference and enhancing SE. Future research may explore adaptive pilot assignment with reinforcement learning, resource-efficient pilot schemes, joint pilot design, and channel estimation to optimize SE and adapt to dynamic wireless conditions.