1. Introduction
Internet of things (IoT) devices have been widely used in many fields, and they are estimated to increase in number to 5.5–12.6 trillion worldwide by 2030 [
1]. As the number of IoT devices exponentially increases, the central station or base station (BS) should efficiently manage limited radio resources. The BS needs to know the channel state information (CSI) of the devices to efficiently manage the radio resources. Hence, the BS requests the devices to report their channel quality indicator (CQI), where the CQI is a 4-bit value that aims to reflect the channel state [
2,
3]. The BS determines the modulation and coding scheme (MCS) based on the CQI reported by the IoT device. The MCS strongly influences throughput because it combines the modulation and code rate according to the channel condition. Many studies have been conducted using adaptive modulation and coding (AMC) to obtain high throughput [
4,
5,
6].
The CQI reporting incurs feedback overhead because the reporting data are transmitted over a low-speed feedback channel. However, if the device does not frequently report its CQI, the BS cannot accurately determine the MCS according to the channel condition, which results in performance degradation. Conversely, if the device frequently reports its CQI, the BS can accurately determine the MCS according to the channel condition, but the feedback overhead is large. Hence, the devices need to report their CQI aperiodically based on channel conditions to achieve a trade-off between the accuracy of the MCS selection and the feedback overhead.
In 5G systems, the two kinds of feedback mode are periodic and aperiodic feedback [
3]. Some researchers have mathematically analyzed the feedback interval [
7,
8,
9]. The authors of [
7] mathematically analyzed the optimal feedback interval that maximizes the average received power. The authors of [
8] mathematically analyzed the feedback interval that maximizes the bidirectional transmission throughput. The authors of [
9] calculated the CSI-dependent interval and used adaptive modulation to obtain higher energy efficiency and throughput. However, the work of [
7,
8,
9] focused on finding intervals of periodic rather than aperiodic feedback. Hence, there is a limit to reduce the number of feedback transmissions.
Machine learning (ML) has been applied to various fields and has led to substantial performance improvements. The 3GPP Release 18 standard applied artificial intelligence (AI)/ML for a new radio (NR) air interface [
10]. Many researchers have focused on compressing the CSI feedback data by using a convolutional neural network (CNN)-based model, where the CNN-based CSI feedback consists of the CSI compression at the terminal and the reconstruction at the BS [
11,
12,
13]. The authors of [
11] introduced a CNN-based feedback model called “CsiNet”, in which an encoder and decoder model were introduced for CSI feedback. The CsiNet model evolved into CRNet and CsiNet+ with improved performance [
12,
13]. The CRNet outperforms the CsiNet without any extra flops by using convolution factorization and the cosine learning rate scheduler. The CsiNet+ improves the performance but it significantly increases the complexity due to the increase in the convolutional kernel size. Other researchers proposed autoencoder-based CSI feedback to improve the accuracy of the CSI feedback with feedback errors [
14,
15,
16]. However, the work of [
11,
12,
13,
14,
15,
16] focused on reducing the amount of feedback data in the periodic feedback. Moreover, the conventional ML structures are too complicated to be used in IoT devices because IoT devices have lightweight processors and small memory capacities.
Some recent studies have been conducted to reduce the complexity of the neural network in CNN-based CSI feedback [
17,
18,
19,
20,
21]. The ENet proposed in [
17] reduced the complexity by exploiting the correlation of the real and imaginary part of CSI while improving the performance. The CLNet proposed in [
18] reduced the complexity by integrating the real and imaginary parts with a spatial-wise attention block. The authors of [
19] proposed two CSI feedback models: One is ConvCsiNet based on a CNN autoencoder, which improves the reconstruction performance, and the other is ShuffleCsiNet based on a lightweight ConvCsiNet, which saves memory space and computing power. The authors of [
20] proposed an adaptive lightweight CNN-based CSI feedback, where the network adaptively finds the compression ratio of feedback data and reduces the complexity of the decoder by about 38.2% in comparison with the CsiNet. The CVLNet proposed in [
21] reduced the complexity by using complex-valued convolutions in the CNN-based CSI feedback. However, the work of [
17,
18,
19,
20,
21] also focused on compressing the feedback data, and moreover they failed to reduce the number of feedback transmissions because the terminal periodically reports its CSI.
Specifically, recurrent neural networks (RNNs) have been applied to CSI feedback in order to exploit the temporal correlation of wireless channel [
22,
23,
24,
25,
26,
27,
28,
29]. Existing work has demonstrated that RNNs can provide efficient CSI feedback and reconstruction for time-varying channels [
22,
23,
24,
25,
26]. However, the number of parameters in RNN layers for CSI compression and reconstruction is generally too large. Although other work attempts to reduce RNN size [
26,
27], most competitive models still require very large parameters. In recent studies, MarkovNet, proposed in [
28], and CoCsiNet, proposed in [
29], improved CSI recovery accuracy with reduced model size. However, the work of [
22,
23,
24,
25,
26,
27,
28,
29] also focused on compressing the feedback data and neglected the feedback delay.
Some researchers predicted the channel state using the past channel state to solve the problem of the CSI quickly becoming outdated due to device mobility of feedback delay [
2,
30,
31,
32]. The authors of [
2,
30] deployed machine learning to resolve the CQI mismatch problem caused by outdated CSI. The authors of [
31] proposed a deep reinforcement learning-based adaptive modulation (DRL-AM) scheme, in which the current CSI is predicted from outdated values. The authors of [
32] integrated an LSTM model and a deep Q-network to overcome the outdated CSI problem for underwater acoustic communication with feedback delay. However, the work of [
2,
30,
31,
32] failed to reduce the number of feedback transmissions because the terminal periodically reports its CSI, although they took into account the feedback delay or outdated CSI. Moreover, the complexity of ML is too high to be implemented in the lightweight IoT device. Recent work of [
33] proposed aperiodic CSI feedback based on the deep neural network (DNN)-based channel prediction, where the terminal decides whether or not to feed back its CSI relying on the DNN-based channel prediction. However, the work of [
33] did not take into account the feedback delay, and moreover, applied a classical DNN rather than an LSTM model in time-series forecasting.
In this paper, we propose a lightweight LSTM-based CQI feedback scheme for IoT devices. Most previous studies have the following limitations. First, most previous studies have focused on compressing the amount of CSI feedback data in the ML-based CSI feedback. Moreover, the number of parameters in the ML model is generally too large. Second, most previous studies have failed to reduce the number of feedback transmissions because they assumed periodic feedback. Third, some studies considered the feedback delay, but most previous studies did not consider the feedback delay. In this paper, the above three problems are overcome. The contributions of this paper are as follows: First, we develop a lightweight LSTM model by decomposing the matrices of the conventional LSTM model using singular value decomposition (SVD) and applying dimensionality reduction. Hence, the proposed LSTM can significantly reduce the complexity. Second, we propose a lightweight LSTM-based CQI feedback scheme for IoT devices, where the IoT device aperiodically reports its CQI to the BS relying on LSTM-based channel prediction. Hence, the proposed LSTM-based aperiodic CQI feedback scheme can dramatically reduce the number of feedback transmissions in comparison with conventional periodic CQI feedback schemes. Third, we evaluate the performance of the proposed LSTM-based CQI feedback scheme under the feedback delay channel. The simulation results show that the proposed lightweight LSTM model shows equivalent performance to the conventional LSTM model despite reducing the complexity by half. The rest of this paper is organized as follows:
Section 2 presents the system model and the AMC scheme.
Section 3 introduces the proposed lightweight LSTM model and the LSTM-based CQI feedback scheme.
Section 4 provides the simulation results and
Section 5 concludes this paper.
5. Conclusions
In this paper, we proposed a lightweight LSTM-based adaptive CQI feedback scheme for IoT devices, where the IoT device aperiodically reports its CQI on the basis of an LSTM-based channel prediction. We developed a lightweight LSTM by using singular value decomposition (SVD) and applying dimensionality reduction. The complexity of the developed LSTM is approximately two times lower than that of the conventional LSTM without sacrificing performance. In the proposed LSTM-based CQI feedback scheme, the IoT device predicts the future channel states relying on the developed lightweight LSTM and determines whether it reports CQI to the BS according to the difference between the last reported CQI and the predicted CQI. In comparison with the conventional periodic CQI feedback scheme, the performance of the proposed lightweight LSTM-based CQI feedback scheme only degrades approximately 4%, but the number of feedback transmissions is reduced by more than approximately 72%.
In practice, IoT devices for sensing and tracking are typically low-cost systems, requiring reduced hardware and software complexity. In comparison with the convention LSTM model, the proposed lightweight LSTM model enables the use of lightweight IoT devices by reducing the complexity by half without deteriorating performance. Moreover, the proposed LSTM-based aperiodic CQI feedback significantly decreases the number of feedback transmissions, resulting in reduced uplink interference. However, because the proposed CQI feedback scheme uses LSTM to predict the future channel state, each IoT device requires learning time and memory to store weights.
For simplicity, no feedback error was assumed in this paper, but for further study, this work can be extended to channel environments with feedback error. Moreover, it is necessary to find an appropriate value of r that determines the SVD dimensionality reduction. Because the number of singular values that we retain in the SVD dimensionality reduction determines the energy, we need to find an appropriate value that compromises on the balance between the channel prediction accuracy and the complexity reduction for different channel environments.