Cross-Technology Interference-Aware Rate Adaptation in Time-Triggered Wireless Local Area Networks

Abstract

The proliferation of IoT using heterogeneous wireless technologies within the unlicensed spectrum has intensified cross-technology interference (CTI) in wireless local area networks (WLANs). As WLANs increasingly adopt time-triggered transmission methods to support real-time services, this interference affects throughput, packet loss, and latency. This paper presents a CTI-aware rate adaptation framework designed to mitigate interference in WLANs without direct coordination with heterogeneous wireless devices. The framework includes a CTI identification model and CTI-aware rate selection algorithms. Leveraging short-time Fourier transform, the identification model captures the time–frequency–power characteristics of CTI signals, enabling the estimation of the average power of various heterogeneous wireless technologies employed by interfering devices. The rate selection algorithms predict CTI occurrence times and adjust the transmission rate accordingly, enhancing the performance of existing explicit and implicit interference mitigation methods. Experimental results demonstrated that the lightweight CTI identification model accurately estimated the average power of each type with an error margin of ±1.414 dBm, achieving this in under 1 ms on the target hardware. Additionally, applying the proposed framework to explicit interference mitigation enhanced goodput by 20.67%, reduced packet error rate by 2.38%, and decreased the probability of packets exceeding 1 ms latency by 0.932% compared to conventional methods.

1. Introduction

The global availability and the license-free policy of the unlicensed spectrum have drawn various Internet of Things (IoT) devices into the spectrum [1,2]. Consequently, IoT devices employing heterogeneous wireless technologies, even non-communication devices like microwave ovens, now coexist and mutually interfere within the same spectrum [3]. This cross-technology interference (CTI) is expected to intensify, highlighting the need for careful countermeasures in IEEE 802.11-based wireless local area networks (WLANs) [4].

Particularly in the 2.4 GHz industrial, scientific, and medical (ISM) band, WLAN devices are susceptible to interference from devices using ZigBee, Bluetooth, and other wireless technologies that share the same frequency spectrum [5]. Such interference degrades communication quality by causing bit errors and packet losses, leading to reduced throughput [6]. Moreover, with the recent trend of adopting time-triggered transmission methods in IEEE 802.11-based WLANs for real-time services in applications like industrial IoT (IIoT) and extended reality (XR), this interference can also impact latency, necessitating appropriate countermeasures [7,8,9].

Existing interference mitigation strategies that can be used to alleviate interference caused by CTI include the following: (1) reducing interference through direct coordination between communication devices [10,11], (2) performing explicit interference mitigation by measuring the channel state from received signals [12], and (3) conducting implicit interference mitigation based on the success or failure of packet transmissions [13,14].

However, these strategies have several limitations. First, direct coordination increases system complexity and complicates overall network management, especially as the number of devices using different wireless technologies grows [15]. Additionally, this method is ineffective against interference from non-communication devices like microwave ovens [16]. Second, measurement-based explicit interference mitigation offers good performance but often requires increased measurement frequency to select appropriate rates when CTI occurrence is unpredictable. This leads to more control messages, causing a trade-off that can reduce overall network throughput [17]. Third, packet error-based, also referred to statistics-based, implicit interference mitigation is simple to implement but may suffer from delayed responses or incorrect rate selections in complex scenarios where multiple devices cause interference simultaneously with different patterns, degrading mitigation performance [18].

Interference from devices using different wireless technologies manifests in various patterns depending on their spectrum usage characteristics in the frequency, time, and power domains [19]. While these diverse characteristics complicate interference dynamics, they also provide an opportunity to identify the presence of specific wireless technologies in the spectrum [20]. This identification enables rate adaptation that responds to CTI without direct communication with the interfering devices. By analyzing the signal patterns emitted by devices, we can perform CTI-aware interference mitigation, allowing on-demand responses and improving performance by tailoring adaptations to each specific CTI.

In this paper, we propose a CTI-aware rate adaptation (CARA) framework that supports interference mitigation in time-triggered WLANs. The main contributions of this study are as follows:

1.

We propose a framework that effectively identifies CTI by leveraging the distinct spectrum usage characteristics of various wireless technologies and performs interference-aware rate adaptation accordingly. This framework addresses two key challenges in time-triggered WLANs for mitigating interference from cross-technology wireless devices without requiring protocol designs specific to each device. These challenges involve identifying signals in each slot and using this information to select an appropriate transmission rate for the next time slot.

2.

Within the framework, we introduce a MobileNet-based CTI identification model designed for low computational complexity. This model estimates the presence and average power of CTI signals in each slot from the detected spectrum, accurately capturing CTI signal characteristics in the time, frequency, and power domains through the application of short-time Fourier transform (STFT). This approach enables estimation of the power of each CTI signal, supporting efficient CTI-aware rate adaptation.

3.

We adapt our framework to two common WLAN interference mitigation approaches, one based on explicit channel measurements and the other leveraging packet error-based implicit interference mitigation, resulting in the M-CARA and P-CARA algorithms, respectively. Both algorithms utilize information obtained from the CTI identification model to project interference occurrence times according to each CTI type’s pattern, and then adjust their responses based on interference intensity. Simulation results show that both algorithms outperform existing methods in terms of packet error rate, goodput, and latency.

2. Related Work

As mentioned earlier in the introduction, the core of the proposed CTI-aware rate adaptation framework lies in CTI identification and rate adaptation based on the identified results. This section reviews related studies on these two key techniques.

2.1. Cross-Technology Interference Identification

As the frequency spectrum becomes increasingly congested, managing heterogeneous wireless networks and addressing spectrum scarcity have become critical challenges. This has led to research on signal detection and identification methods, such as distinguishing various wireless technologies that support IoT or recognizing modulation schemes [21]. Early research focused primarily on recognition methods that can extract signal features, such as statistical properties [22]. However, due to their high computational complexity and limited accuracy, machine learning (ML)- and deep learning-based recognition techniques have gained significant attention [23]. These approaches leverage the fact that signals generated by devices using heterogeneous wireless technologies exhibit distinct characteristics across various dimensions of the spectrum, such as time, frequency, and power, to achieve accurate signal identification [19].

Recent efforts have aimed to expand the number of identifiable signal types, as handling only one or two types is no longer sufficient given the increasing diversity of heterogeneous devices. In [24], Wi-Fi, LTE-license-assisted access (LTE-LAA), and 5G new radio-unlicensed (NR-U) coexistence signal types were detected and distinguished from received signals without explicit decoding, using in-phase and quadrature (I/Q) samples. The researchers employed a convolutional neural network (CNN) model combined with a long short-term memory (LSTM) network and enhanced classification accuracy by applying the STFT to the I/Q sequences. In another study [25], the YOLO (“You Only Look Once”) CNN model was used to transform signal identification into an object detection problem on spectrograms, achieving over 99% accuracy in identifying Wi-Fi, ZigBee, and LoRa signals. J. Gong et al. [26] proposed a multi-task open-set recognition network, inspired by CountGAN, to classify unknown signal types without prior training. Kim et al. [20] utilized a multi-task learning technique to simultaneously extract multi-dimensional spectrum usage characteristics, such as signal type, modulation scheme, occupied channels, and power levels, achieving superior performance compared to traditional single-task models.

These studies collectively demonstrate that signals can be effectively identified by exploiting their unique spectrum usage characteristics. However, it is crucial to consider which features to utilize for model training and how to structure the model’s input, depending on the specific objectives. Moreover, since the devices requiring identification are often resource-constrained communication devices, computational complexity and memory usage must be carefully considered during model design to ensure feasibility in target devices and systems [27].

2.2. Cross-Technology Interference Mitigation

One method to prevent performance degradation caused by CTI is to directly communicate with heterogeneous devices causing interference to coordinate and reduce CTI. This approach seeks to mitigate interference through inter-technology communication between devices operating under different wireless standards [10]. However, it faces the challenge of requiring separate protocols to manage or avoid interference across different technologies [28]. While this method may be effective in environments with a limited number of CTI sources, an increase in the diversity of CTI types can significantly complicate device configuration, leading to network management difficulties due to control overhead [29] or energy consumption issues [30]. Furthermore, this approach struggles to address interference caused by non-communication devices such as microwave ovens. Although methods encoding CTI mitigation information solely within the WiFi payload while maintaining compatibility with existing WLAN standards have been proposed, these methods remain limited to specific scenarios, such as mitigating interference with ZigBee devices [11].

In WLANs, although no specific technologies are designed for CTI mitigation, two common approaches exist: one based on packet errors [13,14] to reduce interference and another relying on channel state measurements to adapt rates [12] for interference mitigation. Packet error-based methods utilize statistical knowledge of packet transmission success rates. In these methods, the transmitting node sends data packets and calculates the success rate based on acknowledgments received from the receiver, adjusting the transmission rate accordingly to mitigate interference. Although these methods are not dependent on specific interference types and can handle heterogeneous CTI, they are implicit mitigation techniques and thus less responsive to rapid changes in the communication environment [18]. This limitation makes it challenging to achieve effective interference mitigation in complex scenarios where multiple wireless technologies contribute to interference dynamics. Some studies have enhanced interference mitigation performance by incorporating IEEE 802.11 fine timing measurements to estimate the distance to access points (APs) [4] or by learning packet error frequency statistics [31]. However, these approaches are often tied to IEEE 802.11 and do not account for CTI characteristics during training, limiting their effectiveness in environments with diverse CTI sources.

Measurement-based methods offer explicit mitigation by estimating communication quality through metrics like the received signal strength indicator (RSSI) and signal-to-noise ratio (SNR) at the receiver, which are then fed back to the transmitter for rate adaptation [12]. These methods are known to respond better to channel variations than packet error-based techniques because they directly measure channel conditions [32]. However, if the measurement frequency is too low relative to the dynamics of the communication environment, performance can degrade; conversely, too high a measurement frequency can reduce throughput due to control messages. To address this issue, Y. Tao and W. L. Tan [33] proposed ReinRate, which can optimize throughput using reinforcement learning (RL). However, the information collected by the RL agent may not be suitable for time-triggered WLANs, or they focus solely on throughput, resulting in suboptimal performance in interference situations caused by CTI.

2.3. Goals of This Study

The goal of this study is to minimize communication performance degradation caused by various types of CTI in time-triggered WLANs. Specifically, it explores a framework that identifies CTI and mitigates interference using WLAN-based interference mitigation techniques without requiring direct protocol designs for the wireless technologies causing the CTI. The key research questions (RQs) are as follows:

RQ1

In time-triggered WLANs, how can interference mitigation be performed when WLAN devices experience cross-technology interference from multiple types of wireless devices without direct coordination with those devices?

RQ2

To perform interference mitigation based on CTI types, how can the CTI devices in the network be distinguished, and how can interference strength information for those devices be obtained?

RQ3

How can CTI in the network be effectively mitigated using interference mitigation methods in WLANs without direct coordination for CTI?

3. CTI-Aware Rate Adaptation Framework

3.1. Overview of the CARA Framework

This paper aims to mitigate performance degradation caused by CTI when WLAN devices operate in a time-triggered communication framework without direct communication with CTI-causing devices. The scenario where APs and STAs in a time-triggered WLAN are affected by CTI is illustrated in Figure 1. When N STAs engage in time-triggered communication with an AP under access time management, they may experience CTI from devices operating in the same frequency band at specific times, using various wireless technologies. In this scenario, each STA is allocated a time duration, $T_{slot}$, during which it can transmit packets for low-latency communication.

In the 2.4 GHz ISM bands, various devices such as ZigBee, Bluetooth devices, and non-communication devices like microwave ovens operate; we represent them as CTI-A, CTI-B, and CTI-C, respectively. Each device exhibits different interference patterns depending on the type of wireless technology it uses, as shown in

Table 1. Features of wireless technologies.

	Time Domain	Frequency Domain	Power Domain
WLAN [34]	Max 5.5 ms wavelength	2400∼2480 MHz, 14 channels, 20 MHz BW	Max 100 mW (20 dBm)
Bluetooth [35]	0.625 µs slot, 1600 hops/s, AFH	2400∼2480 MHz, 79 channels, 1 MHz BW	100 mW (20 dBm) 2.5 mW (4 dBm) 1 mW (0 dBm)
ZigBee [36]	Max 5 ms wavelength	2400∼2480 MHz, 16 channels 2 MHz BW	1 mW (0 dBm)
MWO [37]	∼5 ms On, ∼15 ms Off repeat 60 Hz	2450∼2465 MHz	60 dBm

.

The proposed framework to achieve the interference mitigation objective (RQ1) in the given network environment is shown in Figure 2. This framework performs CTI identification (addressing RQ2) and uses the identified information to enable CTI-aware rate adaptation (addressing RQ3).

Each STA transmits a WLAN packet during its allocated slot duration, and the AP receives radio frequency (RF) signals from the STAs. The received signal waveform reflects interference effects due to CTI. Through a CTI identification process, the AP distinguishes signals generated by other devices operating in the same band and estimates their interference strength. Based on this information, and considering the CTI patterns mentioned earlier, the AP projects the CTI occurrence times for each device and performs CTI-aware rate selection. Finally, the rate selection results, which account for the impact of CTI, are fed back to each STA via modulation and coding scheme (MCS) feedback messages [34], guiding the STAs to adjust their rate control to mitigate performance degradation due to CTI.

To better understand the operation and effectiveness of the framework, we examine a simple scenario in Figure 3a,b, illustrating cases where our framework is not applied and when it is applied. In this scenario, WLAN STAs sequentially transmit uplink data to the AP during each time-triggered interval. In this scenario, CTI-A periodically generates interference every six slots.

As shown in Figure 3a, when STA-2 transmits a packet to the AP in the first slot ($t_1$–$t_2$), the signal is subject to interference from CTI-A. The AP measures the SNR of the received packet and detects a low SNR. Consequently, it sends feedback to STA-2 via an MCS feedback message, instructing it to lower its transmission rate (Rate ↓). Following this feedback, STA-2 transmits packets at a reduced rate during the second slot ($t_4$–$t_5$), but this time, there is no interference from CTI-A during the slot. This incorrect rate control unnecessarily reduces throughput. As the AP receives packets from STA-2 during the $t_4$–$t_5$ slot, it recognizes the absence of interference during this period and sends feedback to STA-2, prompting it to increase the transmission rate by more than two levels (Rate ↑↑). However, when STA-2 follows this instruction and transmits packets at an increased data rate during the third slot, it again encounters interference from CTI-A. With the increased rate, the signal becomes more vulnerable to interference, resulting in unsuccessful packet reception, packet loss, and delays caused by retransmissions.

In contrast, as shown in Figure 3b, the proposed framework enables effective interference mitigation against CTI. Starting from $t_0$, the AP identifies CTI interference in each slot. Using the identified information, the AP projects the expected occurrence of CTI based on the access patterns of CTI signals detected in previous slots (projected CTI information is represented by red crosshatched boxes). Specifically, the AP monitors interference from CTI-A during the $t_0$–$t_1$ and $t_1$–$t_2$ slots through CTI identification and uses this information to project future occurrences of CTI. Considering the interference pattern of CTI-A, the AP determines that interference is unlikely to occur when STA-2 transmits its packets during the $t_4$–$t_5$ slot. Consequently, the AP sends a control message (Rate ↑↑) to STA-2, instructing it to transmit at a higher data rate during $t_4$–$t_5$. Similarly, for the $t_7$–$t_8$, the AP anticipates interference from CTI-A and prompts STA-2 to use a simpler MCS to ensure transmission stability (Rate ↓↓).

As demonstrated in this scenario, the success of the proposed framework in mitigating CTI interference in time-triggered networks relies on the performance of CTI identification and CTI-aware rate selection. In Section 3.2 and Section 3.3, we will examine in detail the intelligent CTI identification model proposed within the framework and the CTI-aware rate selection algorithm, which integrates our framework’s approach with measurement-based and packet error-based rate adaptation algorithms widely used in traditional WLANs.

3.2. CTI Identification

The primary objective of CTI identification is to estimate the power levels of interfering signals within the target spectrum, considering the types of signals causing the interference. Traditional filter-based detection methods struggle to accurately estimate the power levels of each signal, and decode signal methods are limited to estimating only specific signals. To overcome these limitations, we adopted an ML-based approach capable of accurately estimating the average power of CTI signals present in the received RF waveform. More specifically, it estimates the average power for each type of wireless technology used by these signals. The proposed MobileNet [38]-based CTI identification model, CTINet, is illustrated in Figure 4.

The signal received by the AP is down-converted and digitized into a set of raw I/Q samples. During the time duration $T_{scan}$, the received signal includes both WLAN signals and CTI signals. This raw signal sample can be directly input into the model or preprocessed for signal identification purposes. We use the STFT to effectively capture the time–frequency domain characteristics of CTI signals and to facilitate the estimation of power density within the sample. Since the window size of the STFT determines the time and frequency resolution, an appropriate window size should be selected based on the characteristics of the wireless technology used by the target CTI signals [24]. If other characteristics of CTI signals, such as modulation schemes, need to be considered for interference mitigation, a multi-task learning-based approach that can derive these features simultaneously along with power estimation can be employed. Interested readers are referred to our previous work [20].

The transformed STFT sample is converted into a grayscale image and input into the MobileNet backbone. In our framework, we assume that the power of WLAN signals and noise are estimated using the preamble fields of WLAN [34], rather than by the model itself. As a result, CTINet focuses on identifying the types of wireless technologies causing CTI in the input sample and estimating the average power for each:

$$\widehat{P}=f_{\theta }\left(I\right),$$

(1)

where:

• $\widehat{P}={[{\widehat{P}}^{\left(1\right)},{\widehat{P}}^{\left(2\right)},\dots ,{\widehat{P}}^{\left(M\right)}]}^{\top }$: Predicted average power vector for the M CTI signal types;
• $f_{\theta }$: CTI identification model parameterized by $\theta$;
• I: An image representation obtained by applying the STFT to the received signal over the duration $T_{\mathrm{scan}}$.

The proposed CTINet uses MobileNetV3 [38] as the backbone, providing high performance in resource-constrained environments. The model takes a grayscale STFT image as input and extracts various features of CTI signals in the time, frequency, and power dimensions through multiple convolutional layers. A Flatten layer then converts the multi-dimensional feature maps into a one-dimensional vector, which is passed to a Dense layer. The Dense layer effectively combines the time–frequency features using this vector, learning key patterns necessary to distinguish each signal’s type and predict their respective power levels. Finally, the learned vector is connected to the output layer, performing multi-output regression to simultaneously predict the average power of each signal’s type.

To train the model, a loss function $L\left(\theta \right)$ is defined to minimize the difference between the actual and predicted power levels of the CTI signals:

$$L\left(\theta \right)={\displaystyle \frac{1}{N}}\sum _{i=1}^N\sum _{m=1}^M\left|P_i^{\left(m\right)}-{\widehat{P}}_i^{\left(m\right)}\right|,$$

(2)

where:

• N is the total number of training samples;
• $P_i^{\left(m\right)}$ is the actual average power of CTI signal type m in sample i;
• ${\widehat{P}}_i^{\left(m\right)}$ is the predicted average power of CTI signal type m in sample i.

To perform conservative interference mitigation, when multiple signals of the same type are present within a sample, we set the signal with the highest average power level as the representative average power for that signal type.

In constructing the signal identification model, computational complexity and execution speed must be carefully considered. Since the model runs on systems with limited computing resources, such as AP, simply choosing a model that provides the highest accuracy may not be desirable. An appropriate balance should be found by considering the impact of input sample size, the model parameters, and the size of the output layer (number of CTI signal types to identify) on the model’s processing speed. The computational complexity of the proposed model can be expressed as follows:

$$\begin{array}{cc}\hfill TC=& \underset{\mathrm{Input}\mathrm{Complexity}}{\underbrace{O\left(W\times H\times C\times C_1\times K_1^2\right)}}\hfill \\ \hfill & + \underset{\mathrm{Model}\mathrm{Complexity}}{\underbrace{O\left(\sum _{l=2}^L\left(K_l^2\times C_{l-1}\times C_l\times {\displaystyle \frac{W_l\times H_l}{S_l^2}}\right)\right)}}\hfill \\ \hfill & +\underset{\mathrm{Output}\mathrm{Complexity}}{\underbrace{O\left(F\times N\right)}},\hfill \end{array}$$

(3)

where:

• $W,H,C$: input sample size (width, height, and channels);
• $C_1,K_1$: number of output channels and kernel size in the first layer;
• L: total number of layers;
• $K_l$: kernel size of layer l;
• $C_{l-1}$, $C_l$: number of output channels from the previous layer and current layer, respectively;
• $W_l$, $H_l$: width and height of the output feature map at layer l;
• $S_l$: stride at layer l;
• F: number of parameters in the final fully connected layer;
• N: number of final output labels.

Based on these equations, appropriate parameters need to be selected considering the presence of devices causing interference in the target network and the computing resources of the system where the model runs. This allows balancing the model’s accuracy and execution speed to support real-time signal identification in the target operating environment. For example, in MobileNetV3, the width multiplier ($\alpha$, $0<\alpha \le 1$) can be used to significantly reduce computational load under certain performance constraints. By applying $\alpha$, layers can be made narrower, and the computational cost of the layers with $\alpha$ applied is reduced to

$${\alpha }^2\times K_l^2\times C_{l-1}\times C_l\times {\displaystyle \frac{W_l\times H_l}{S_l^2}}.$$

(4)

Additionally, to reduce the number of output labels when designing the model, similar signals within CTI can be grouped into the same category [26].

3.3. CTI-Aware Rate Selection

In this framework, we propose a CTI-aware rate selection algorithm that supports interference mitigation by utilizing CTI information (power estimates for each CTI signal type) derived from the CTI identification model. These algorithms are presented in Algorithms 1 and 2. Our approach has the advantage of not requiring separate communication protocols for devices using different wireless technologies, making it easy to apply to existing rate adaptation algorithms in WLANs. Specifically, Algorithm 1, measurement-based CARA (M-CARA), applies the CARA framework to an explicit rate adaptation method based on SINR measurements, and Algorithm 2, packet error-based CARA (P-CARA), applies the CARA framework to an implicit approach based on packet errors.

Algorithm 1 Measurement-based CTI-aware rate selection

Require: Number of STAs N; slot duration $T_{\mathrm{slot}}$; signal identification period $T_{\mathrm{scan}}$. Ensure: $MC{S}_{\mathrm{SINR}}^{\left(k\right)}$ for each STA k.  Initialization: CTI projection table; $MC{S}_{\mathrm{SNR}}^{\left(k\right)}$ for all $k=1,2,\dots ,N$.  for slot index $i=0$ to total number of slots do     Determine the STA k assigned to the i-th slot.     Estimate the signal and noise power of STA k and update average values.     if $imod\left(T_{\mathrm{scan}}/T_{\mathrm{slot}}\right)=0$ then         Identify CTI signals.         if CTI signals are identified then             Update the CTI projection table.         end if      end if      Calculate SINR and derive $MC{S}_{\mathrm{SINR}}^{\left(k\right)}$ for STA k’s next transmission slot.      Return $MC{S}_{\mathrm{SINR}}^{\left(k\right)}$.    end for

The MCS index can be sent to the STA via an MCS feedback message. The STA performs rate control based on the MCS index received from the AP, thereby mitigating interference from CTI in each transmission cycle. The main steps of this algorithm are as follows: (1) identify types of CTI signals every $T_{\mathrm{scan}}$ period; (2) if types of CTI signals are identified, predict the times of occurrence of interference based on their patterns in the time domain and update the CTI projection table; and (3) based on the CTI projection table, calculate the SINR reflecting the impact of CTI for the next cycle of the STA transmitted in the current $T_{\mathrm{slot}}$, and derive the MCS index accordingly.

When calculating the SINR for a specific STA at the AP, the signal power and channel noise of that STA’s signal are calculated using the preamble field.

Table 2. MCS index and SNR threshold.

MCS Index	Modulation	Coding	SNR Threshold
0	BPSK	1/2	2
1	QPSK	1/2	5
2	QPSK	3/4	9
3	16-QAM	1/2	11
4	16-QAM	3/4	15
5	64-QAM	2/3	18
6	64-QAM	3/4	20
7	64-QAM	5/6	25
8	256-QAM	3/4	29
9	256-QAM	5/6	31

is used to derive the MCS index based on the calculated SINR. The MCS index is selected according to the minimum SNR threshold required. The projection table is updated when a specific signal type is continuously identified, using the identified time intervals. The time resolution of the projection table is set to the same as the signal identification period $T_{\mathrm{scan}}$.

The P-CARA algorithm mitigates channel noise interference by adjusting the rate on the transmitter side based on ACK messages, while compensating for the impact of CTI using the CARA framework.

Algorithm 2 Packet error-based CTI-aware rate selection

Require: Number of STAs N; slot duration $T_{\mathrm{slot}}$; signal identification period $T_{\mathrm{scan}}$; rate selection step m. Ensure: $fla{g}_{\mathrm{CTI}}^{\left(k\right)}$, $\mathsf{\Delta }MC{S}^{\left(k\right)}$ for each STA k.  Initialization: CTI projection table; set $fla{g}_{\mathrm{CTI}}^{\left(k\right)}=\mathtt{off}$; $\mathsf{\Delta }MC{S}^{\left(k\right)}=0$ for all $k=1,2,\dots ,N$.  for slot index $i=0$ to total number of slots do      if A packet is received from the STA using slot i then         Estimate the signal and noise power of the STA.         Update the average signal and noise power values for the STA.         Transmit an ACK message to the STA.      end if      if $imod\left(T_{\mathrm{scan}}/T_{\mathrm{slot}}\right)=0$ then          Identify CTI signals.          if CTI signals are identified then              Update the CTI projection table.          end if      end if      if CTI is predicted in the next m-th slot then          Determine the STA k assigned to the next m-th slot.          Set $fla{g}_{\mathrm{CTI}}^{\left(k\right)}$ to on.          Estimate the signal power of STA k.          Calculate $MC{S}_{\mathrm{SNR}}$ and $MC{S}_{\mathrm{SINR}}$ for STA k.          Compute $\mathsf{\Delta }MC{S}^{\left(k\right)}=MC{S}_{\mathrm{SNR}}-MC{S}_{\mathrm{SINR}}$.           if $fla{g}_{\mathrm{CTI}}^{\left(k\right)}=\mathtt{on}$ and $\mathsf{\Delta }MC{S}^{\left(k\right)}$ has changed then              Return $MC{S}_{\mathrm{SNR}}$ and $MC{S}_{\mathrm{SINR}}$.           end if      else           Find STA k where $fla{g}_{\mathrm{CTI}}^{\left(k\right)}=\mathtt{on}$.           Set $fla{g}_{\mathrm{CTI}}^{\left(k\right)}$ to off.           Return $fla{g}_{\mathrm{CTI}}^{\left(k\right)}$.      end if  end for

The STA performs rate adaptation against CTI by receiving $\mathsf{\Delta }MC{S}^{\left(k\right)}$, the difference between the MCS index based on channel noise and the MCS index including CTI impact, from the AP and incorporating it into its rate control. Although a separate message format could be designed to send the $\mathsf{\Delta }MC{S}^{\left(k\right)}$, we choose to use the existing MCS feedback message. By utilizing the MFB subfield in the VHT control field, which can convey redundant information about the channel state [39], we send the $\mathsf{\Delta }MC{S}^{\left(k\right)}$ through this message. The receiver includes in the MFB’s MCS subfield the MCS index calculated from the SINR (including CTI impact) and includes the average SNR value for channel noise in the SNR subfield. The transmitter receives these values, calculates $\mathsf{\Delta }MC{S}^{\left(k\right)}$, and performs rate control accordingly.

The P-CARA algorithm mitigates interference from CTI at the receiver side through $\mathsf{\Delta }MC{S}^{\left(k\right)}$, independently of the existing transmitter-side rate adaptation algorithm that addresses interference based on channel conditions. Consequently, interference mitigation is activated only when CTI impact is anticipated. This management is handled using $fla{g}_{\mathrm{CTI}}^{\left(k\right)}$ to perform rate adaptation only when the impact of CTI interference changes.

The proposed algorithms (M-CARA and P-CARA) can effectively mitigate CTI interference through the framework without direct communication with CTI-causing devices. Each algorithm is derived from existing WLAN interference mitigation methods, M-Base and P-Base, retaining their respective advantages while addressing CTI-related interference. (M-Base, a channel state measurement-based algorithm, may be more complex to implement but is better suited for adapting to channel variations and interference. In contrast, P-Base, which relies on packet error-based methods, is relatively simpler to implement but may offer lower interference mitigation performance compared to measurement-based methods).

4. Experiment Setup

To evaluate the CARA framework, we constructed a dataset for training the CTI identification model and developed a simulation-based experimental environment for evaluation. The performance of the CTI identification model is assessed based on the power estimation accuracy (mean absolute error, MAE) for each CTI signal type within the detected spectrum. Since the size of the STFT image affects the accuracy, we also evaluate performance based on varying image sizes.

When designing the model, it is essential to consider the computing performance of the target system (in this experiment, the identification needs to be performed within a 1 ms processing time requirement). Therefore, in addition to accuracy performance, we also analyze the model’s floating point operations (FLOPs) and the number of parameters required to meet these performance constraints.

The performance evaluation of the interference mitigation algorithms integrated with the CARA framework (M-CARA and P-CARA) is conducted in terms of packet error rate (PER), goodput, and latency. The proposed algorithms are compared with the measurement-based algorithm (M-Base) and the packet error-based algorithm (P-Base). The M-Base algorithm employs an explicit feedback approach where the receiver estimates channel conditions by measuring the SNR, selects the appropriate data rate based on these measurements, and communicates the selected MCS value to the transmitter via MCS feedback messages to perform rate control [12]. In contrast, the P-Base algorithm relies on the transmitter selecting an appropriate data rate based on observed packet errors using ACK messages. In this paper, we adopt automatic rate fallback (ARF) [13] as the representative P-Base algorithm.

4.1. Data Generation and Model Training Parameters

In the data generation process, we considered a scenario where WLAN signals generated by STAs in a time-triggered WLAN environment are received by AP under interference from three types of CTI signals: Bluetooth, ZigBee, and microwave oven signals. The frequency band is 2.4 GHz, and the sampling rate is 20 MHz. The waveforms of the WLAN signals were generated using the MATLAB WLAN Toolbox, while the waveforms of Bluetooth and ZigBee signals were generated using the MATLAB Bluetooth Toolbox and Communication Toolbox, respectively. The microwave oven waveform was generated by referring to [37]. Each waveform was sampled according to various parameter settings supported by the standards, as described in

Table 3. Parameter options for waveform generation.

Signal Type	Parameter	Values
WLAN [34] IEEE 802.11	Standards	IEEE 802.11a, b, n, ac, ax
	Modulation Schemes	BPSK, QPSK, 16-QAM, 64-QAM, 256-QAM
	Transmission Rate	Determined by MCS
	Center Frequency	2.452 GHz
	Bandwidth	20 MHz
	Transmission Power	≤20 dBm
ZigBee [36] IEEE 802.15.4 (CTI-A)	Modulation Schemes	OQPSK
	Transmission Rate	250 Kbps
	Center Frequency	2.445, 2.45, 2.455, 2.46 GHz
	Bandwidth	5 MHz
	Transmission Power	≤0 dBm
Microwave oven [37] (CTI-B)	Center Frequency	2.45 GHz
	AC Frequency	60 Hz
	Transmission Power	≤60 dBm
Bluetooth [35] IEEE 802.15.1 (CTI-C)	Modulation Schemes	GFSK, DQPSK, 8DPSK
	Transmission Rate	Determined by Tx Mode 1, 2, 3 Mbps
	Center Frequency	18 channels, 2.443–2.461 GHz
	Bandwidth	1 MHz
	Transmission Power	≤20 dBm

. The diversity in the generation of WLAN signal’s standards adds to the difficulty of CTI signal identification.

Depending on each wireless technology’s access pattern, the CTI waveforms interfere with the WLAN waveforms, and the WLAN waveforms with CTI waveforms are collected over a duration of $T_{\mathrm{scan}}$ to generate baseband raw signal samples. $T_{\mathrm{scan}}$ was set to 1 ms. We collected I/Q samples with SNRs ranging from 0 dB to 20 dB. The sliding window size of the STFT was set to 256. The STFT was converted into three image sizes, 32 × 32, 64 × 64, and 128 × 128. A total of 500,000 data samples were generated and divided into training, testing, and validation sets in a ratio of 80%, 10%, and 10%, respectively.

For training CTINet, we employed various depth multipliers $\left(\alpha \right)$ of 1.0, 0.75, 0.5, and 0.25 to adjust the network’s capacity and complexity. The models were trained using the Adam optimizer with a learning rate of 0.001, for 100 epochs, and a batch size of 64.

4.2. Simulation Setup

In the simulation scenario, we used the identification model that yielded optimal results on the test dataset. We constructed a MATLAB-based simulation test environment for the evaluation scenario. The parameters considered in the simulation settings are presented in

Table 4. Simulation parameters.

Parameter	Value
Simulation time	10 s
Number of APs	1
Number of STAs	3
Tx to Rx distance	10 m
$T_{slot}$	1 ms
Inter-frame space	16 $\mathsf{\mu }$s
Payload size	500 bytes
Ack size	14 bytes
MCS feedback size	20 bytes
Feedback interval (MB)	1 ms
Success threshold (PE)	20
Failure threshold (PE)	5
Packet loss threshold	${10}^{-5}$ bit error

.

The target network is a wireless time-triggered network with a time slot length of $T_{\mathrm{slot}}$, based on the IEEE 802.11ac [34]. Three STAs sequentially obtain repetitive opportunities to transmit uplink data frames to the AP. In each slot, the data rate of the STA is determined based on the MCS selected by the algorithm. The payload size of each packet is set to 500 bytes, and packets are transmitted continuously during the given $T_{\mathrm{slot}}$, with a minimum IFS interval between transmissions. If there is not enough time to complete the packet transmission, the remaining data are sent in the next cycle. In the scenario, it is assumed that MCS feedback control frames and ACK frames are not lost, and to minimize transmission errors, these control frames are always transmitted at the lowest data rate. If the transmitted data packet has a bit error rate (BER) of ${10}^{-5}$ or higher, the packet is considered lost.

The CTI devices are located 3 m away from the receiver, the AP, with transmission power of 0 dBm, 60 dBm, and 20 dBm for ZigBee, MWO, and Bluetooth, respectively. ZigBee transmits data with a 120-byte payload at 30 ms intervals [40], MWO operates with on/off cycles at 60 Hz [37], and Bluetooth uses adaptive frequency hopping at 1600 hops per second [35].

5. Results and Analysis

5.1. Performance of CTI Identification

Table 5. Performance comparison of CTINet based on input size.

SNR	STFT 32 × 32 (3.232 MFLOPs)	STFT 64 × 64 (9.821 MFLOPs)	STFT 128 × 128 (36.178 MFLOPs)
0 dB	6.956	3.041	1.322
5 dB	6.751	3.035	1.308
10 dB	6.424	2.780	1.294
15 dB	5.680	2.624	1.034
20 dB	5.669	2.525	0.999

presents a comparison of the MAE according to the input STFT image size of the CTI identification model (with $\alpha =1$), with each model’s FLOPs indicated in parentheses. The MAE of identification models decreases as the image size increases. The model that takes 128 × 128 images as input achieves an MAE of 1.322 even at 0 dB, meaning it can estimate the average power within ±1.322 dBm. However, as the input size increases, the number of convolution operations using the sliding window approach also grows, leading to a higher FLOPs due to the larger input. In contrast, the model that takes 32 × 32 STFT images has a low FLOPs of 3.232 M but shows poor performance, with an MAE of 6.424 even at an SNR of 10 dB. This could directly lead to degraded rate adaptation performance in the CARA framework, as the error in the SINR calculation may affect the determination of the MCS index during rate adaptation. Regarding the model size, all comparative models have the same size of 0.976 × ${10}^6$ parameters due to the characteristics of model, which mainly consists of convolution and global average pooling operations; thus, the model size was not affected by the input size [41,42].

MobileNetV3 supports hyperparameters for model lightweighting. Among these, $\alpha$, also known as the depth multiplier, adjusts the model’s width (i.e., the number of filters) by a certain ratio. The performance comparison of CTINet according to $\alpha$ is shown in

Table 6. Performance comparison of CTINet based on the $\alpha$.

Models	FLOPs (M)	Params (M)	MAE	CPU (ms)
CTINet-32-1.0	3.232	0.976	6.956	0.202
CTINet-32-0.75	2.309	0.611	7.106	0.144
CTINet-32-0.5	1.176	0.297	7.292	0.074
CTINet-32-0.25	0.594	0.110	7.361	0.037
CTINet-64-1.0	9.821	0.976	3.041	0.614
CTINet-64-0.75	7.293	0.611	3.109	0.456
CTINet-64-0.5	3.760	0.297	3.230	0.235
CTINet-64-0.25	2.032	0.109	3.256	0.127
CTINet-128-1.0	36.178	0.976	1.322	2.261
CTINet-128-0.75	27.231	0.611	1.359	1.702
CTINet-128-0.5	14.101	0.297	1.414	0.881
CTINet-128-0.25	7.787	0.109	1.468	0.487

. The computational complexity of the models mentioned in Section 3.2 is compared using the metric called FLOPs. As $\alpha$ decreases, the FLOPs and model size (number of parameters) of CTINet-128 decrease from 36.178 M and 0.976 M at $\alpha =1$ to 7.787 M and 0.109 M at $\alpha =0.25$. However, we observe that the MAE increases as the model becomes more lightweight from $\alpha =1$ to $\alpha =0.25$. Some commercial WiFi APs are equipped with high-performance processors, such as the Broadcom BCM4916, which has four cores running at 2 GHz. The theoretical peak performance of these APs can reach at least 16 GFLOPS (Giga Floating-point Operations Per Second), assuming a minimum of 2 operations per cycle. Based on this, the processing speed comparison results calculated with the model’s FLOPs are shown in the CPU (ms) column of Table 6. Considering the model’s performance (MAE of 2 or less) and the required processing speed in the network (within $T_{\mathrm{slot}}$), we selected CTINet-128-0.5 as the model to use in the rate adaptation of the next experiments. When designing models, it is important to consider the trade-off between computational complexity and performance according to network requirements.

5.2. Performance of CTI-Aware Rate Selection

5.2.1. Packet Error Rate

The comparison results for PER across the algorithms are shown in Figure 5. Packet errors are defined as cases where the bit error rate of a transmitted frame exceeds ${10}^{-5}$. In the result graph, “Ideal” represents the outcome under ideal rate control, where the type and power of CTI-causing devices are accurately known for each slot, allowing rate selection based on the calculated SINR to mitigate interference effectively.

Overall, algorithms incorporating the CARA framework demonstrate superior reliability by adaptively performing rate adaptation in response to CTI situations. M-CARA shows a decreasing PER trend as WLAN SNR increases, presenting a graph shape that is nearly identical to the ideal case. In contrast, M-Base exhibits a tendency where PER does not decrease in the SNR range of 18–30. This is likely because it fails to reduce the transmission rate appropriately in the presence of CTI, resulting in error accumulation.

Packet error-based algorithms inherently adjust the MCS index based on the success or failure of packet transmission, which results in a relatively higher PER in the SNR range of 8–25 compared to measurement-based algorithms. Specifically, packet error-based algorithms, as shown in the experimental settings in Table 4, rely on success/failure thresholds for rate control. This makes them slower to react to channel state changes than measurement-based algorithms. Furthermore, because these algorithms operate implicitly, they tend to interpret the receipt of an ACK message as an indication of good channel quality. Consequently, they may attempt to select a higher rate, which leads to transmitting packets at rates exceeding the actual channel capacity, ultimately resulting in a worse PER (explained in more detail in Section 5.2.4).

5.2.2. Goodput

Figure 6 illustrates the comparison of algorithms in terms of goodput when receiving CTI under different SNR conditions based on WLAN in the simulation setup. The goodput is averaged over all WLAN STAs. The results show that measurement-based algorithms generally outperform packet error-based algorithms, which are implicit methods, because they perform rate adaptation based on explicit channel state measurements. Moreover, packet error-based algorithms send ACK messages for every packet, resulting in additional control message overhead that reduces goodput compared to measurement-based algorithms.

M-CARA achieves the highest performance across all WLAN SNRs by preemptively identifying potential CTI situations and selecting an appropriate MCS to minimize packet loss. This proactive approach enhances throughput. In addition, M-CARA reduces control overhead by exchanging MCS feedback only during CTI-induced situations, leading to further throughput improvements.

Packet error-based rate adaptation inherently has delayed reactions in adjusting rates, resulting in similar performance among packet error-based algorithms. Moreover, CARA’s rate adaptation in response to CTI tends to select lower rates compared to traditional algorithms when CTI is present, which leads to better performance in terms of reliability. However, this also results in reduced throughput. In algorithms like PE, where the rate adaptation response is slower, the effort to improve throughput by minimizing packet loss is relatively less effective compared to MB algorithms, leading to less improvement in goodput.

5.2.3. Latency

To compare performance in terms of latency, Figure 7 presents latency boxplots for each algorithm at WLAN SNRs of 8, 18, and 30 dB. The measured latency pertains to data packets generated by all WLAN STAs in the network during a 10 s simulation. In our experiment, lost packets are considered retransmitted, causing delays to accumulate until the packet is successfully transmitted without loss; these appear as outliers in the boxplots. In the time-triggered network considered, if an STA fails to complete packet transmission within its assigned time duration, the packet is sent in the next cycle, resulting in additional delays equivalent to $N\times T_{\mathrm{slot}}$, where N is the number of STAs. Since time-triggered networks prioritize not exceeding specific delay thresholds, the proportion of delays exceeding 1 ms is also examined in

Table 7. Percentage of packets exceeding 1 ms latency for each algorithm.

WLAN SNR	M-CARA	M-Base	P-CARA	P-Base
8 dB	7.857%	8.789%	10.912%	11.191%
13 dB	1.040%	1.354%	5.465%	5.556%
18 dB	0.212%	0.267%	3.597%	3.920%
25 dB	0.051%	0.175%	2.077%	2.398%
30 dB	0.023%	0.166%	0.111%	0.254%

.

At 8 dB, the median latency values are [0.354, 0.662, 0.662, 0.5] ms, indicating low delays below 1 ms, but some packets experience delays accumulating up to over 9 ms. Notably, M-CARA, through appropriate MCS selection, has no packets with delays exceeding 7 ms. According to Table 7, the proportion of packets with delays over 1 ms for M-CARA is 7.857%, the lowest among the algorithms, confirming its superior latency performance.

At 18 dB, where the channel condition is relatively good, the total packet transmissions are higher, and the overall PER is lower, resulting in median latency values of 0.146 ms across all algorithms. However, the P-Base algorithm exhibits packets with delays around 6 ms, and a higher proportion of delays exceeding 1 ms at 3.92%, indicating poorer performance.

At 30 dB, the median latency values are [0.106, 0.106, 0.094, 0.094] ms, with packet-based algorithms showing slightly lower latency. This is because, in this condition, packet error-based algorithms transmitted most packets at the data rate corresponding to MCS index 8, whereas measurement-based algorithms used MCS index 7. The proportion of packets with delays exceeding 1 ms is lowest for M-CARA at 0.023%. This indicates that only the M-CARA approach meets the latency requirements of 99.9% compliance in networks where all transmitted packets must satisfy stringent latency constraints (e.g., wireless time-sensitive networking [7]). M-CARA uniquely achieved a compliance rate of 99.977%. M-Base shows a higher proportion of outliers exceeding 1 ms compared to P-CARA, demonstrating that the CARA framework significantly improves latency performance.

5.2.4. Rate Selection Ratios

Figure 8 illustrates the proportion of MCS indices, i.e., the data rates selected during data packet transmissions, along with the PER for each algorithm. The MCS indices range from 0 to 8. In each subplot corresponding to a specific SNR, the most frequently selected MCS index reflects the typical choice of comparison algorithms under the given channel conditions.

As shown in Section 5.2.1, packet error-based algorithms generally exhibit higher PERs than measurement-based algorithms. This is because packet error-based algorithms interpret the receipt of an ACK message as an indication of good channel quality, leading them to continually attempt to select higher MCS indices. Consequently, across all subplots, packet error-based algorithms show a higher frequency of selecting higher MCS indices compared to measurement-based algorithms. Notably, this tendency to select MCS indices higher than the optimal channel condition often results in relatively higher PERs, as shown in Figure 8b–d.

At a WiFi SNR of 25 dB, measurement-based algorithms predominantly select MCS index 6, as it is determined based on the measured SNR under the given channel conditions. In contrast, packet error-based algorithms tend to select MCS indices 7 and 8, depending on the success or failure of packet transmissions.

Algorithms incorporating the CARA framework demonstrate more conservative behavior in selecting MCS indices, as they account for the impact of CTI. Comparing across algorithm types, CARA-based algorithms consistently show a higher frequency of selecting lower MCS indices under all conditions. This difference becomes more pronounced as channel conditions improve. For instance, at a WiFi SNR of 30 dB, the difference in the proportion of lower MCS index selections between CARA-based algorithms and their non-CARA counterparts is 5.16% and 2.84%, respectively (proportions of MCS indices 4–6, highlighted in green). This higher frequency of selecting lower MCS indices explains why CARA-based algorithms achieve lower PERs while maintaining comparable or superior performance in terms of goodput and latency compared to conventional algorithms.

6. Conclusions

In this paper, we presented a CTI-aware rate adaptation framework to address the performance degradation caused by CTI in time-triggered WLANs. The proposed framework includes an intelligent model for CTI signal identification and two algorithms supporting CTI-aware rate adaptation: M-CARA and P-CARA. CTINet, built on a lightweight architecture, can estimate the average power levels of heterogeneous wireless technologies used by CTI signals in systems with constrained computing resources. The M-CARA and P-CARA algorithms mitigate the performance degradation of existing algorithms and minimize network overhead by projecting and responding to interference occurrence times based on CTI identification results and their spectrum usage patterns.

Our signal identification experiments showed that the model in the CARA framework that takes 128 × 128 STFT images as input can distinguish heterogeneous wireless technologies used by CTI signals and estimate the average power of those types within an error of ±1.414 dBm at an SNR of 0 dB. This result is significant in practical terms because it comes from a model that can perform identification within 1 ms on an AP equipped with a processor offering 16 GFLOPS performance.

According to comparative experiments on interference mitigation performance, the application of the CARA framework showed performance improvements over existing rate adaptation algorithms in all aspects of goodput, PER, and latency. In particular, the CARA framework exhibited better synergy with measurement-based algorithms that perform explicit interference mitigation based on channel conditions, improving goodput by up to 20.67%, reducing the PER by up to 2.38%, and decreasing the probability of packets exceeding 1 ms latency by 0.932% compared to existing algorithms.

We expect that the CARA framework will contribute to enhancing the communication performance of real-time applications using time-triggered WLANs by identifying CTI caused by devices employing heterogeneous wireless technologies in unlicensed bands and performing corresponding mitigation strategies.