Complex Natural Resonance-Based Chipless RFID Multi-Tag Detection Using One-Dimensional Convolutional Neural Networks

This paper proposes a chipless radio frequency identification (RFID) multi-tag detection system. The one-dimensional convolutional neural network (1D CNN) was employed as an intelligent classifier of the proposed system, which was fed by complex natural resonances (CNRs). Experiments with contexts of a single chipless RFID tag were conducted to collect input datasets for training and validating the 1D CNN. The CNRs and natural frequencies only extracted from the individual tag’s responses by using the short-time matrix pencil method (STMPM) and then obtained after performing data augmentation were separately fed to the 1D CNN for training and validation in order to compare their performance. The accuracy obtained from the training and validation of the 1D CNN fed by the CNRs was significantly higher than that of the 1D CNN fed by the frequencies only. The performance matrices in terms of precision, recall, and F1-score also confirmed the superiority of the use of CNRs over that of frequencies only. In order to verify the performance of real-time multi-tag detection utilizing the proposed system, experiments with contexts of multiple tags were carried out, and the experimental results have shown that the system using the 1D CNN fed by the frequency only failed to detect multiple tags. In contrast, the proposed system was able to deliver 100% accurate multi-tag detection. However, as demonstrated by the experimental results, the proposed chipless RFID multi-tag detection system was restricted to a resolution of 3 cm.


I. INTRODUCTION
Chipless RFID tags are a cost-effective and flexible alternative to traditional RFID tags, as they do not rely on integrated circuits or chips for data storage.These tags encode information using variations in electromagnetic properties or resonant frequencies, offering advantages such as lower cost, smaller form factors, and greater design flexibility.
The associate editor coordinating the review of this manuscript and approving it for publication was Mohamed Kheir .
The importance of chipless RFID tags lies in their potential to revolutionize data collection, automation, and decisionmaking processes across industries.They seamlessly integrate into existing systems and offer insights into tagged items' movements, locations, and conditions.By leveraging chipless RFID technology, businesses can streamline operations, reduce costs, prevent losses, and enhance customer experiences.Continued research and development in this field hold promise for expanding the potential applications of chipless RFID technology, unlocking new possibilities in areas such as inventory management, security, and overall operational efficiency.Overall, chipless RFID tags present a compelling solution for efficient identification, tracking, and management of items in various sectors.
Various decoding methods for chipless RFID applications have been explored, each with their advantages and limitations.Threshold setting methods [1], [2] are simple but sensitive to distance variations, while adaptive energy detection is more robust but may not be suitable for all scenarios.The maximum likelihood (ML) [3], [4] and the signal space representation (SSR) [5] methods offer improved accuracy but can be computationally intensive for high bit densities.Recent improvements, such as logarithmic SSR (LSSR) and window-based SVD (WB-SVD), have addressed computational complexity at the cost of a slight increase in error probability.
The extraction of complex natural resonances (CNRs) offers a promising approach for chipless RFID applications, particularly in sensing and identification scenarios [6], [7], [8], [9], [10].CNRs are advantageous due to their aspect independence, meaning that they remain consistent regardless of the angle of arrival or observation point.Several methods for CNR extraction have been explored, including the matrix pencil method (MPM), short-time matrix pencil method (STMPM), and spectrogram method.Among these, the former two are optimized for estimating and extracting CNRs described by the singularity expansion method (SEM) theory.The MPM, for instance, involves solving a generalized eigenvalue problem, making it suitable for chipless RFID tag responses.However, it requires multiple measurements, is distance-dependent, and faces challenges in selecting the appropriate time window [6], [7], [11].Additionally, it struggles to differentiate reflections from background objects and tags.In order to address these issues, researchers have proposed separating CNRs based on their time and direction of arrival.The MPM's inclusion of singular value decomposition assists in the mitigation of noise effects.Moreover, techniques such as the autocorrelation function and time-domain averaging approaches have been introduced to further reduce noise effects during MPM implementation.These advancements in CNR extraction methods hold promise for enhancing the accuracy and reliability of chipless RFID systems in both identification and sensing applications.
The STMPM is emerging as a promising alternative to the matrix pencil method, effectively overcoming some of its challenges.The superiority of the STMPM over the MPM lies in its capability to operate without the need for prior knowledge regarding the initiation of the late time.By utilizing a sliding time window and applying MPM at each window location, the STMPM generates a time-frequency plot, enabling the identification of the approximate start time of the tag mode and the calculation of average resonance frequencies over time.These approaches, [3] and [12], exhibit enhanced noise resilience compared to MPM, leading to more accurate CNR extraction.However, the STMPM does have limitations in terms of fixed resolution in time and frequency.
In order to address these limitations and to further boost its performance, we have proposed innovative improvements, such as incorporating a k-nearest neighbor (k-NN) algorithm to establish decision boundaries and to facilitate the decoding of tag responses [13].These advancements show great potential in elevating the capabilities of chipless RFID systems and advancing their application in diverse scenarios.
The limitations or challenges faced in the current methods for bit ID detection in chipless RFID tags include the operation scenario with one reader and one tag.In this scenario, the reader and tag may experience interference or signal collisions, leading to inaccuracies in bit ID detection.For example, Fig. 1, a schematic diagram of the packaging on the conveyor belt scenario for chipless RFID tags illustrates the practical deployment of chipless RFID technology in a packaging environment.The figure shows a conveyor belt system where various packages are being transported.Each package is equipped with a chipless RFID tag, enabling seamless identification and tracking throughout the packaging process.The diagram highlights the integration of chipless RFID tags into the packaging workflow, offering insights into the real-world application and the significance of accurate bit ID detection in ensuring efficient inventory management, supply chain optimization, and enhanced product security.On the conveyor belt, it is common to encounter multiple chipless RFID tags simultaneously within the reading range of the reader.This scenario introduces a significant challenge in terms of accurately identifying and distinguishing the bit IDs of each tag, leading to potential errors and complications in the identification process.Another example scenario is that self-localization or device-based localization often involves objects equipped with RF readers that rely on coded landmarks to determine their positions [14].These innovative landmarks are essential for ensuring the reliable operation of such systems.In [15], an optimization approach was presented for the placement of chipless tags within indoor environments.This approach was designed to guarantee THz band coverage while adhering to specific system constraints and achieving 3D k-coverage with an emphasis on minimizing tag usage.The proposed algorithms were rigorously validated through numerical simulations and ray-tracing analyses across a range of conditions, thereby contributing to the potential effectiveness of such systems in real-world applications.
In recent years, neural networks have revolutionized various fields of research and technology, and their applications have grown exponentially.With advancements in deep learning, neural networks have emerged as powerful tools for solving complex problems, particularly in the domain of pattern recognition and classification tasks.Convolutional neural networks (CNNs), in particular, have proven to be highly effective in image and signal processing tasks due to their ability to automatically learn hierarchical features from the input data.Notably, in the context of chipless RFID, some noteworthy articles have explored the integration of deep learning techniques.For instance, reference [16] presents a method that combines neural networks with frequency responses and spectrograms using 2D CNNs for chipless RFID tasks.Additionally, a deep learning-based security model was proposed to provide a high accuracy of 93%, when classifying a cloned chipless RFID tag from a genuine tag, even in the presence of additive RF interference in real-time [17].These innovative approaches demonstrate the growing interest in applying deep learning to chipless RFID systems, offering promising avenues for enhanced performance and expanded applications.However, the articles mentioned did not specifically focus on multiple tag detection in chipless RFID systems.Because there have been few articles presenting chipless RFID multi-tag detection, the research into this technique is still challenging.
The contribution of this paper lies in the development of a technique to address the challenge of accurate bit ID detection in chipless RFID multi-tag systems.Our proposed approach combines CNR extraction, utilizing the STMPM to extract CNRs, with the power of a 1D CNN architecture for efficient feature processing.By leveraging the discriminative CNR features and the capabilities of the 1D CNN in handling onedimensional data, our technique achieves precise and reliable classification of bit IDs in multi-tag scenarios.Through comprehensive experimentation, our method demonstrates promising results, indicating its potential as a solution for efficient and accurate chipless RFID multi-tag detection.This advancement in multi-tag detection techniques can significantly enhance the performance and applicability of chipless RFID systems in various real-world applications.
The paper is organized into several sections as follows.After the introduction, the details of the proposed technique, covering the CNR extraction by STMPM and the architecture of the 1D CNN, are in Section II.The training process is described in Section III, where the experiments with onetag scenarios were conducted to collect the data used for the training and validation of the 1D CNN.In Section IV, the multi-tag scenarios are presented in order to confirm the potential impact of the proposed technique by the detection of bit IDs from multiple tags.Finally, conclusions are drawn in Section V.

II. PROPOSED TECHNIQUE
In this paper, a technique for the multi-tag detection of the chipless RFID system is proposed.The 1D CNN fed by CNRs, which are often referred to as poles, including natural frequency and the damping factor, were employed to capture the spatial dependencies in the tag data, enabling effective multi-tag detection.Fig. 2 depicts the basic configuration of the proposed chipless RFID multi-tag detection system.The frequency responses of the tags are measured to collect the data and then first processed with preprocessing techniques, including antenna calibration [3], [6], [18], [19], to reduce response degradation due to the unwanted signals and antenna effect.The frequency responses are transformed into timedomain responses by using the inverse Fourier transform.The CNRs are extracted from the time-domain responses of the tags by using the STMPM.The extracted CNRs which are formed as the input data are fed into the 1D CNN in order to detect the tag IDs.In the 1D CNN procedure, the CNRs are used to create a labeled dataset with the corresponding IDs of tags.The labeled dataset is divided into the training set and the validation set.These features in the training set are fed into the 1D CNN for learning to identify patterns in the CNRs in order to predict the IDs of the tags.The validation set evaluated the performance of the trained 1D CNN.Finally, leveraging the power of the trained 1D CNN, the multi-tag IDs will be predicted for the chipless RFID multi-tag detection system.

A. ANTENNA CALIBRATION
In Fig. 2, a bistatic configuration of the chipless RFID system is shown in the part of the data collection.The system operates by transmitting signals with swept frequencies 138080 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
based on stepped-frequency continuous wave radar (SFCW) from a transmitting antenna to the chipless RFID tag and subsequently receiving the backscattered signals through a receiving antenna.In order to address issues arising from the unflat antenna response and surrounding environments, an antenna calibration technique was implemented in [3], [6], [18], and [19].The overall transfer function of the chipless RFID system can be mathematically expressed as: where H tag (ω) denotes the total transfer function of the chipless RFID tag and where the mutual coupling between the antennas is denoted as H C (ω).The transfer functions of the transmitting and receiving antennas are represented as H Tx (ω) and H Rx (ω) in the equation.It is important to note that the right-hand-side term in (1) relates to the measurement in free space without a tag and is expressed as The resulting expression is obtained by subtracting (1) and ( 2) and is given as follows: The purpose of antenna calibration is to obtain the original transfer function of the tag, denoted as H tag (ω), while H Tx (ω) and H Rx (ω) remain present in (3).To address this, a reference transfer function H plate (ω) is needed, which is obtained by replacing the tag with a large metal plate.Since the plate is considered a perfect reflector across all frequency bands, its transfer function is represented as −1.The reference transfer function and −1 are respectively substituted into the overall transfer function H total (ω) and the tag's transfer function H tag (ω), resulting in the following expressions: Therefore, the transfer function of only the tag can be expressed as follows: The antenna calibration process involves using (5) to eliminate unwanted responses, ensuring the extraction of the pure tag response for further processing.

B. CNR EXTRACTION USING STMPM
As mentioned, in this paper, the 1D CNN was employed as an intelligent classifier of the proposed chipless RFID system.The input features of the 1D CNN were CNRs extracted from the time-domain responses of tags by using the STMPM.The time-domain response of a tag is obtained by transforming the transfer function of the tag denoted by H tag (ω) into the timedomain response as expressed by where y e (t) and y l (t) represent the early-time and late-time portions of the time-domain response of the tag, respectively.η(t) denotes an additive noise.Based on the SEM principle, the time-domain response of the tag can be successively divided into the early-time and late-time portions as given in (6).The late-time portion of the tag can be modeled as a summation of the damped exponentials with complex natural frequencies as given by where s i = σ i + jω i indicates the i th pole (CNR), which is a complex natural frequency, comprising damping factor σ i and angular natural frequency ω i .The R i denotes the complex residue, while M is the number of poles.
Regarding the method of CNR extraction, the MPM was amended by sliding the window with an appropriate size along the whole time-domain response.This amended method is called the STMPM.In this method, the window with a time length of T w is slightly slid along the time axis incrementally by T .The windowed response of the late time portion can be written as where The CNRs are extracted from the windowed response by forming the matrix [Y T l ] first as given by where N denotes the total number of samples.The pencil parameter L is usually chosen between N /3 and N /2 in order to filter the noise-contaminated in data.A singularvalue decomposition (SVD) is used to factorize this matrix to be [ , where H denotes the conjugate transpose, and [U] and [V] are unitary matrices, composed of the eigenvectors of The problem of solving for λ = e sit can be expressed as an ordinary eigenvalue problem where In the following section, we delve into the 1D CNN for chipless RFID tag detection.We will discuss the architecture of the 1D CNN model, the training process, and the evaluation metrics employed to assess the performance of the proposed system.

C. 1D CONVOLUTIONAL NEURAL NETWORKS
The 1D CNN is a type of neural network architecture commonly used for processing and analyzing sequential data.It is particularly effective in capturing spatial dependencies and patterns in one-dimensional data, such as time series or signal data.Similar to traditional CNNs, the 1D CNN employs convolutional layers that apply filters to local regions of the input data, followed by pooling layers to downsample and extract key features.This enables the network to automatically learn and discern relevant patterns and structures in the sequential input.By leveraging the power of deep learning and feature extraction, the 1D CNN model has been successfully applied in various applications, such as biometric recognition with multiple sensor data [20], fault diagnosis of rotating machinery using multi-signals [21], multivariate abnormal detection in industrial control systems [22], non-contact medical detection based on sensor data [23], and radar jamming signal classification from a single signal [24].In this paper, the 1D CNN is applied to the intelligent muti-tag detection of the chipless RFID system.Fig. 3 depicts the basic architecture of the 1D CNN employed as an intelligent classifier of the proposed chipless RFID multi-tag detection.It consists of the input layer, convolutional layers with a rectified linear unit (ReLu) layer, pooling layers, fully connected layers, the Softmax layer, and the output layer.In the architecture, the input layer receives input data that are typically one-dimensional sequences.Here, the input data are CNRs often called poles, extracted from the time-domain responses of tags by using the STMPM discussed in the previous section.Typically, the validation dataset is not directly used as input for any specific layer; instead, it is used to evaluate the performance of the trained 1D CNN and to assess its generalization capability.During the training process, the 1D CNN is trained by using the training dataset, while the validation dataset is used to monitor its performance and to make decisions regarding model selection or hyperparameter tuning.After each training epoch that is comprised of a certain number of iterations, the performance of the 1D CNN is evaluated by using the validation dataset.In order to evaluate the performance of the 1D CNN, the confusion matrix should be determined.This matrix provides valuable insights into the performance of the classification of the 1D CNN and presents a comprehensive From the confusion matrix, several performance metrics can be derived.Accuracy measures the overall correctness of the predictions of the 1D CNN, calculated as the ratio of correctly classified instances to the total number of instances.Precision quantifies the proportion of true positive predictions out of all positive predictions, providing an indication of the ability of the 1D CNN to avoid false positives.Recall, also known as sensitivity or true positive rate, measures the proportion of true positive predictions out of all actual positive instances, indicating the ability of the 1D CNN to identify positive cases.The F1-score combines precision and recall into a single metric, providing a balanced measure of the performance of the 1D CNN.The performance matrices, namely accuracy, precision, recall, and F1-score, can be calculated by the equations, respectively, as follows: and By analyzing the values in the confusion matrix and calculating these performance metrics, we can gain a deeper understanding of the strengths and weaknesses of the 1D CNN in correctly classifying the different classes.This information is essential for assessing the overall performance of the 1D CNN and identifying areas for improvement.
Reconsidering Fig. 2, a multi-tag detection technique of the chipless RFID system is proposed in this paper.In this technique, CNRs formed as an input dataset extracted from the time-domain responses of chipless RFID tags are fed into the 1D CNN employed to detect tag IDs.The proposed technique holds promise for improving the accuracy of the multi-tag detection of the chipless RFID system.By leveraging the power of CNR extraction and 1D CNN, our approach aims to enhance the detection process, enabling more efficient and precise tracking of chipless RFID tags.The effectiveness of the proposed technique will be further evaluated and validated in the subsequent experimental section, where we will assess the performance of the technique by comparing two datasets: one composed of CNRs with damping factors and natural frequencies, and another consisting of natural frequencies only.This comparison aims to examine the impact of including damping factors in the dataset on the accuracy of the detection technique.Through this evaluation, we aim to gain insights into the effectiveness and potential benefits of utilizing the complete dataset of CNRs for chipless RFID multi-tag detection.

III. TRAINING PROCESS A. TAG DESIGN
A chipless RFID tag was designed and implemented solely for the purpose of demonstrating the proposed technique.The tag used in this study consists of a rectangular metallic patch loaded with three slot resonators, following the designs outlined in [13] and [25].It is important to note that the proposed technique can be applied to other tags designed based on resonance frequency, where the tag's unique identifier is encoded within its resonant frequency, as seen in examples in [26] and [27].The coding capacity of a chipless RFID system is generally limited by the number of unique resonances that can be encoded in a tag.This limitation is often influenced by the physical size and properties of the tag.Increasing the coding capacity typically involves designing tags with more complex resonance structures or larger physical dimensions.However, the feasibility of such designs depends on specific application requirements.Due to limitations in the operating frequency of our transmitting and receiving antennas during the experiments discussed later, we selected three unique IDs to embed in the tag.In Fig. 4(a) and (b), the layout of the tag examples with IDs ''111'' and ''000'' is depicted.The tag structure is based on a rectangular patch and three-slot resonators.Fabrication of the tag was carried out on a single-layer FR4 substrate with a dielectric constant and loss tangent of 4.4 and 0.014, respectively.By conducting simulations, the size parameters, including L 1 , L 2 , L 3 , L 4 , w, and s, were optimized to achieve three distinct resonant frequencies, namely 3.928, 4.896, and 6.424 GHz.These resonant frequencies were illustrated by notch frequencies in the S 21 , corresponding to those in [13].The size values for L 1 , L 2 , L 3 , L 4 , w, and s, were set at 12, 10.5, 9, 7.5, 0.75, and 0.4 mm, respectively.Referring to Fig. 4, it is possible to alter the tag's behavior by shorting and/or opening the slots at the corners, resulting in the generation of eight different possible IDs.

B. EXPERIMENTAL SETUP
In order to train the 1D CNN for ID detection of chipless RFID systems, initial experiments were conducted to collect the data.These experiments involved one chipless RFID tag, and they were carried out within an anechoic chamber to mitigate the influence of surrounding environments.Fig. 5 illustrates the experimental setup of the chipless RFID system.In this setup, ultra-wideband tapered slot antennas were utilized as both transmitting and receiving antennas [28].They were placed in a face-to-face bistatic configuration.The operating frequency of these antennas ranged from 3.1 to 10.6 GHz, covering all resonant frequencies of the fabricated chipless RFID tag.The distance between the antennas was set at 6 cm, and the tag was positioned at a distance of 10 cm from the center of the transmitting and receiving antennas.The maximum distance at which a tag can be detected depends on several parameters, such as the transmitted power, antenna gain, sensitivity of the receiver, and operating frequency.However, in our scenario, the maximum distance is approximately 20 cm.In order to ensure a controlled environment, both the antennas and the tag were placed on Styrofoam with a dielectric constant of approximately one.This step was taken to minimize any potential interference and to ensure reliable data collection for training the 1D CNN.
The scattering coefficient S 21 was measured using the R&S ZVB20 vector network analyzer (VNA) and was subsequently recorded on the computer.Each measurement involved sweeping the frequency from 2 to 8 GHz, with a step frequency interval of 10 MHz.The total number of data points collected was 6001.In order to ensure accuracy and reliability, the measurement process was repeated 10 times, and the results were then averaged before proceeding with further signal processing procedures.These additional procedures included antenna calibration and target identification.By averaging the measurements, any potential variations or noise in the data were minimized, allowing for more precise and consistent signal processing results.
Antenna calibration was carried out as the initial step in order to mitigate the effects due to antenna mutual coupling and unflat antenna response.Following the procedure in [13], a large metal plate with dimensions of 50 cm × 50 cm served as the reference H plate (ω).This plate was positioned in the same location as the tag, allowing the measurement of the transfer function of the reference H plate (ω).As a result, all incident electromagnetic waves on the plate were perfectly reflected.After removing the chipless RFID tag and plate from the setup, the scattering coefficient S 21 was measured to obtain H no tag (ω).Subsequently, all transfer functions obtained from these measurements were utilized in (5) to derive the response without the antenna effects.By performing this antenna calibration, the issues of antenna mutual coupling and unflat antenna response were effectively mitigated, leading to more accurate data for subsequent signal processing, including target identification.
The input data utilized to train the 1D CNN were gathered within a free-space anechoic chamber.Throughout this data collection process, measurements were taken for the scattering coefficient S 21 of all possible chipless RFID tags.As previously mentioned, the antenna calibration method was subsequently applied to all the collected data from these chipless RFID tags.Following the antenna calibration, the responses obtained in the frequency domain were converted into time-domain signals using the inverse Fourier transform.The subsequent step involved extracting poles, including damping factors and natural frequencies, using the STMPM.These extracted poles are employed in training the 1D CNN, which will be discussed in the following section.

C. CNR EXTRACTIONS
After completing the antenna calibration, the CNRs extracted from the scattering response of eight possible tags will be utilized for training the 1D CNN in the chipless RFID detection system.The individual tag response H tag (ω) in the frequency domain was transformed into the time domain using the inverse Fourier transform.Subsequently, the CNRs were extracted from the time-domain response using the STMPM.It is worth noting that the STMPM does not require  an additional technique to estimate the commencement of the late time [7], [9].In Fig. 6, the trajectories of three natural frequencies of the extracted poles are shown over time.These three natural frequencies correspond to three bits of an individual tag.The commencement of the late-time response was determined by the STMPM, assuming that the CNRs obtained at the late time would be stationary [7], [9].It was observed that the commencement of the late-time response occurred at t LT = 1 ns.The averages of the five successive natural frequencies extracted from the late-time responses of a tag were found to be 4.069, 5.074, and 6.585 GHz, respectively, representing the ID ''111.''In contrast, the IDs ''000'' were represented by averaged natural frequencies of 4.364, 5.447, and 6.673 GHz.These values were significantly different from the natural frequencies representing the ID ''111.''In the figure, ten different colors of the symbols denote the CNRs obtained from ten measurements.For each measurement, CNRs were extracted from a signal obtained from the averaging of measurements repeated 10 times.Fig. 7 presents the trajectories of the damping factors obtained using the STMPM.The average damping factor was calculated by taking the average of five successive damping factors observed at the late time.The averaged damping factors, for ID ''111,'' were -0.306, -0.445, and -0.728 GHz.It should be noted that some damping factors, such as those used to represent ID ''000,'' cannot be averaged due to their instability.In this paper, the 1D CNN in the proposed multi-tag detection chipless RFID systems was trained using the CNRs, which also included the natural frequency and damping factor extracted from the late-time responses of eight chipless RFID tags.However, the extracted CNRs will probably be shifted because of the fabrication of tags.This can be resolved by using extracted CNRs, including the shifted damping factors and natural frequencies to train the 1D CNN.In addition, the performance of the use of CNRs in the 1D CNN-based chipless RFID tag detection system was compared with that of frequency only, and the results of this comparison will be discussed later in this paper.

D. PREDICTIVE MODEL CREATION 1) DATA AUGMENTATION
In order to develop a predictive model for the proposed chipless RFID multi-tag detection, a substantial quantity of data samples is essential for the training and validation of the 1D CNN [29].The CNRs extracted from the late-time responses of eight chipless RFID tags were found to be insufficient to effectively train the 1D CNN.As a result, the utilization of a data augmentation technique became necessary in order to augment the number of CNRs, which subsequently served as the input data for training the 1D CNN.To address this requirement, a sliding segmentation technique was introduced to expand the number of CNRs.The symbol s[n] represents the CNR sequence containing n sample points at the later time portions.Within this framework, the variables L seg , L overlap , and N respectively indicate the length of each segment, the extent of sample overlap between adjacent segments, and the total number of segments.The sequence of CNRs, s[n], was partitioned into N segments, a partitioning scheme that can be formally expressed as FIGURE 8. Data augmentation.
Fig. 8 illustrates the data augmentation process applied to the CNRs sequence, denoted as s [1], s [2], . . ., s[n].The process involves creating sliding segments, referred to as 1 st L seg and 2 nd L seg , which originate from s [1] and s [3] in the CNRs sequence, respectively.The degree of overlap between these two segments is indicated as L overlap .Following this, the subsequent segment commences at s [5] and continues to slide through the sequence, generating segments until the total reaches N th L seg .This process allows us to calculate the total number of CNRs employed in the augmentation process, which is determined using (16).This approach effectively expands the dataset for training and enhances the model's capacity to learn from the CNR data.
Because of the limited duration of the late-time portion, the sequences of CNRs used for data augmentation consisted of only 25 sample points, resulting in a sequence length n of 25.The parameters L seg and L overlap were assigned values of 6 and 5, respectively.These parameter values were chosen to maximize the number of segments obtained.Each sequence of CNRs, corresponding to a specific bit of the tag ID and comprising these defined parameters, was utilized to generate a total of 20 datasets.Repeated measurements, conducted 10 times, were performed for each three-bit chipless RFID tag.A demonstration of the proposed system involved the utilization of eight distinct tags.This collective effort resulted in the attainment of a cumulative total of 1600 input datasets.Out of these, 80% were designated for training purposes, while the remaining subset was allocated for validation.It is essential for each dataset to be appropriately labeled with its respective class.In this study, a three-bit chipless RFID tag was employed, thereby leading to the formulation of the target data as a categorical vector of IDs such as ''000,'' ''001,'' . . ., and ''111,'' each of which corresponds to one of the eight potential IDs.

2) 1D CNN PARAMETERS
The proposed chipless RFID multi-tag detection system employed the 1D CNN as an intelligent classifier.Fig. 9 depicts the architectural components of the utilized 1D CNN.Within the framework of this 1D CNN, the input layer received the sequences of CNRs obtained from the three-bit chipless RFID tags.The input size was defined according to the number of features in the input data, and these data were subsequently organized into a matrix configuration.In the construction of the input layer, each row of the matrix corresponded to a distinct feature of the input data, while the column count of the matrix aligned with the number of CNRs present.Given the utilization of three-bit chipless RFID tags, the input data matrix for the proposed system was structured with 6 rows and 6 columns.In contrast, the chipless RFID multi-tag detection system that employed natural frequencies only as its basis of operation was employed for comparative purposes.In this case, the input data matrix for the frequency-only system comprised 3 rows and 6 columns.Therefore, the 1D CNN fed by CNRs has a higher level of complexity than the 1D CNN fed by natural frequency only.The feature inputs were channeled through two blocks of 1D convolutional layers, ReLU layers, and normalization layers.The convolutional layers utilized a filter size of 3, with the initial and subsequent convolutional layers containing 32 and 64 filter kernels, respectively.Subsequently, a 1D global average pooling layer was implemented to condense the output of the convolutional layers into a singular vector representation.In order to transform the output into a probability vector, a fully connected layer was incorporated, featuring an output size congruent with the number of distinct classes.This was followed by the integration of a Softmax layer and a classification output layer, thus culminating in the classification process.

3) TRAINING AND VALIDATION
In the training process of the 1D CNN, the adaptive moment estimation (Adam) optimizer with a learning rate of 0.001 was used to minimize the loss.The optimizer updated the parameters using a mini-batch size of 9 for 200 epochs.The training network was validated every 50 iterations by predicting the response of the validation loss and calculating the validation loss and accuracy.Fig. 10(a) and (b) depict the overall accuracy and loss over 200 epochs of using the frequencies only and CNRs in chipless RFID detection, respectively.In Fig. 10(a), the validation accuracy of 79.375% was obtained from the 1D CNN trained by frequencies only in the chipless RFID detection, while we achieved the validation accuracy of 100% by using the CNRs for training the 1D CNN of the proposed system as shown in Fig. 10(b).The training and validation accuracy of the 1D CNN trained by the CNRs in the proposed chipless RFID detection system was superior to that of the frequencies only, as can be seen in Fig. 10  The performance of chipless RFID detection using the 1D CNN was also evaluated through the confusion matrix.Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
the 1D CNN completely predicted the ID of the specific tag for all forty input datasets.As depicted in Fig. 11(a), the value of 28 associated with the predicted ID ''000'' aligned with the actual ID ''000'' (the first element in the matrix) reveals that the 1D CNN trained and validated with frequencies only accurately predicted the ID for 28 out of the total 40 input datasets.With the exception of the first element, the other values in the first row added up to 12.This indicates that the TP of detection for ID ''000'' was equal to 28 and the FN was equal to 12. Except for the first row, the first column's values for the predicted ID ''000'' added up to 2. This indicates that the FP of ID ''000'' was 2. The total of the values that appeared outside of the actual ID ''000'' row and the predicted ID ''000'' column was 278.Therefore, the TN was 278.Note that not all IDs can be accurately detected by chipless RFID detection using the 1D CNN trained and validated with the natural frequencies only.As seen in Fig. 11(b), all values along the diagonal of the matrix were equal to 40.This indicates that all IDs can be completely detected using the proposed chipless RFID tag detection using the 1D CNN trained and validated with the CNRs.This provides preliminary evidence that the 1D CNN fed by the CNRs performs better for chipless RFID detection than that fed by the frequencies only.
It is important to establish the performance matrices, which include precision, recall, and F1-score.In order to derive the precision, recall, and F1-score, the TP, TN, FP, and FN previously calculated and retrieved from the confusion matrix were substituted into (13)- (15).The precision, recall, and F1score of the ID ''000'' with chipless RFID detection using a 1D CNN fed by the frequencies only were, respectively, 0.933, 0.700, and 0.800.The precision, recall, and F1-score of the ID ''000,'' on the other hand, were all one, as determined by the detection using the 1D CNN fed by the CNRs.The collections of performance matrices for ID detection using 1D CNN fed by the natural frequencies only and the CNRs are shown in Table 1.With detection using the 1D CNN fed by the CNRs, the table shows that the tag IDs were completely detected with a 100% success rate because all performance matrices had a value of one.In contrast, the detection system using the 1D CNN fed by the frequencies only had lower success rates in the overall performance matrices, with the exception of the precision of ID ''010'' and the recall of ID ''111,'' both achieving a perfect score of 100%.
In order to apply the proposed system in real-world scenarios, one of the initial challenges we encountered was frequency shift, often caused by factors such as fabrication tolerance or variations in itemization.For instance, when tags were attached to parcel boxes or plastic containers, both components of CNRs were shifted [13].As mentioned, it was observed that the 1D CNN fed by CNRs outperformed that fed by frequency only for chipless RFID detection.In order to investigate the effect of frequency shifts, the 1D CNN fed by CNRs was exclusively determined.The damping factor and natural frequency, which are components of CNRs, were systematically varied with a step size of 20 MHz.This  allowed us to identify the frequency breakdown point for each component.The accuracy percentages obtained from different CNRs are shown in Table 2, where σ and f denote the shift of damping factor and natural frequency, respectively.There were three columns in which the accuracy was less than 100%, specifically for the cases of  3.  According to the obtained breakdown points, it is evident that the proposed system using the 1D CNN fed by CNRs can effectively operate with natural frequency shifts of up to 160 MHz.Considering the design of the chipless RFID tag, where each ''1'' bit is spaced approximately 1 GHz apart, it becomes apparent that we could potentially increase the code capacity by approximately 5 bits within this frequency range.However, the challenge lies in the design complexity.The experimental results confirm that the proposed system can reliably detect tags as long as the frequency shift remains within the limits of σ = 40 MHz for damping and f = 160 MHz for the natural frequency.

4) REAL-TIME DETECTION OF ONE TAG
In order to verify the superiority of the chipless RFID system using the 1D CNN fed by the CNRs over that fed by the frequencies only, the real-time detection for the tag IDs was determined.The CNRs and natural frequencies only extracted from the whole response of individual tags were separately fed into the trained 1D CNN in order to evaluate the performance of real-time detection.The final output of 138088 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
the system was the predicted IDs generated by the trained 1D CNN.Fig. 13(a) and (b) depict the real-time detection results of eight possible tag IDs.Following [30], the commencement of late-time response was estimated in order to make the proposed chipless RFID detection automatic.The estimated late time was t L = 1 ns as specified by a red vertical dotted line in the figures.Thus, the predicted appearing at the late time were therefore considered as real results of ID detection.Fig. 13(a) depicts the one-tag ID detection using the 1D CNN fed by the input dataset of the frequency only.In the figure, the predicted IDs obtained at the late time which was denoted by a red dotted line were detection results.The predicted IDs of ''000,'' ''010,'' ''011,'' ''100,'' and ''110'' did not coincide with the actual tag IDs.However, correct detection was achieved only for the tag IDs ''001,'' ''101,'' and ''111.''This implies that the chipless RFID onetag detection using the 1D CNN fed by the frequencies only struggles to accurately detect the tag IDs as evidenced by the lack of coincidence between the predicted IDs and the actual tag IDs.The true positive detection percentage of 37.5% further emphasizes the limitations of the 1D CNN fed by the frequencies only.On the other hand, Fig. 13(b) illustrates the performance of the real-time one-tag detection using the 1D CNN fed by the CNRs.In this case, all predicted IDs obtained from the 1D CNN fed by the CNRs perfectly coincide with the actual eight tag IDs.The true positive detection percentage of 100% signifies the ability of the proposed system to accurately detect all tag IDs.This stark contrast highlights the superiority of the 1D CNN fed by the CNRs over that fed by the frequencies only in terms of accuracy.These results clearly indicate that the inclusion of CNRs as an input dataset significantly improves the performance of the chipless RFID tag detection system.The use of the CNRs demonstrates a remarkable capability to detect tag IDs, making it a superior choice compared to the use of the frequencies only.
In order to investigate the effects of frequency shifts on the 1D CNN fed by CNRs, we varied the damping factor and natural frequency of the CNRs were systematically varied with a step size of 20 MHz.This allowed us to identify the frequency at which the performance of the 1D CNN began to degrade.The result of the real-time detection for one tag using the 1D CNN fed by CNRs under damping factor and natural frequency shifts with σ = 60 MHz and f = 180 MHz is shown in Fig. 14.In the figure, all predicted IDs from the 1D CNN trained with CNRs perfectly coincided with the actual eight tag IDs.As discussed earlier, we determined that the breakdown frequencies for the damping factor and natural frequency shifts are σ = 60 MHz and f = 180 MHz, respectively.However, even at these points, the accuracy rate remained high at 99.69%, ensuring a perfect match between predicted and actual tag IDs.

IV. CHIPLESS RFID MULTI-TAG DETECTION
The proposed chipless RFID system using the 1D CNN trained and validated with the CNRs was applied to multi-tag A. EXPERIMENTAL SETUP Fig. 15 depicts the experimental setup of the proposed chipless RFID multi-tag detection system.The experiments were conducted in an anechoic chamber in order to avoid the effect of surrounding environments.The configuration of the transmitting and receiving antennas was similar to that of the antennas in the previous experiments, whose results were used in the training process of the 1D CNN.The ultra-wideband tapered slot antennas were employed as the transmitting and receiving antennas as well.The distance between the antennas was 6 cm.The antennas and two chipless RFID tags were placed on the Styrofoam whose dielectric constant was approximately one.In the experiments, the scattering coefficient S 21 was measured by using the VNA with the swept frequency from 2 to 8 GHz.The data collection was performed as in the previous experiments for the training process.In order to investigate the performance of chipless RFID multi-tag detection, the experiments were divided into two main scenarios.First, the distance between the two tags was fixed as d = 5 cm.Two tags were used to demonstrate the proposed chipless multitag detection system.In practical terms, the number of tags can be increased.However, the maximum number of tags depends upon several parameters, including the transmitter's power, sensitivity of the receiver, antenna beamwidth, and distance between tag and reader.The first and second tags, respectively called tag 1 and tag 2, were placed at 10 and 15 cm away from the center of the two antennas, respectively, as seen in the figure.Tag 1 had a specific ID of ''100,'' while tag 2 was systematically changed to all eight possible IDs.The purpose of these experiments was to determine whether the proposed system could effectively detect tag 1 while also accurately detecting and differentiating the different IDs of tag 2. Second, the spacing between the two tags was adjusted to d = 1, 2, 3, 5, and 10 cm.The distance between the center of the antennas and the first tag remained 10 cm.IDs of tag 1 and tag 2 were fixed at ''100'' and ''111,'' respectively.The aim of these experiments was to investigate the limitations of the distance between the tags.

B. REAL-TIME DETECTION OF TWO TAGS
The measure scattering coefficients S 21 obtained from the two-tag scenario H total (ω), empty room H no tag (ω), and large metal plate H plate (ω) were substituted into (5) to perform the antenna calibration and to mitigate the effect due to surrounding environments and mutual coupling between antennas.The results of the antenna calibration were transformed into time-domain signals.The CNRs, including damping factors and natural frequencies, were extracted from the time-domain signals by using the STMPM.From the firstscenario experiments, the CNRs of two chipless RFID tags were fed into the 1D CNN trained with the CNRs, while the natural frequencies were separately fed into the 1D CNN trained with the frequencies only.Both of these methods were employed for comparison.Fig. 16(a) and (b) illustrate the real-time detection results of the two-tag scenario experiments using the 1D CNN fed by the frequencies only and the CNRs, respectively.In the figure, the vertical red dotted and black dashed lines denote the commencements of the late time for tag 1 and tag 2, respectively.The commencements of the late time of tag 1 and tag 2 were at t LT 1 = 1 ns and t LT 2 = 1.37 ns, respectively.As indicated by the blue stars in Fig. 16(a), although the predicted IDs ''011'' and ''111'' of tags 1 and tag 2 appear at the late time correctly, the actual IDs were ''100'' and ''111'' for tags 1 and tag 2, respectively.This implies that tag 2's ID detection was correct whereas tag 1's was incorrect.At the late time of tag 2, only two predicted IDs, including ''100,'' and ''111'' of tag 2, coincided with the actual IDs.Unfortunately, there was no predicted ID for tag 1, which coincided with the actual ID, at the late time of the tag.This confirms that the chipless RFID system using the 1D CNN fed by the frequencies only completely fails to detect multiple tags.On the other hand, Fig. 16(b) depicts the results of realtime detection using the 1D CNN fed by the CNRs.At the late time of tag 1 and tag 2, all predicted IDs of tag 1 and tag 2 coincided with the actual IDs.The results confirm again that the performance of using the CNRs for chipless RFID multitag detection is superior to that of using the frequencies only.
According to the second-scenario experiments, the IDs of tag 1 and tag 2 were fixed at ''100'' and ''111,'' respectively.The spacing between the tags was varied as d = 1, 2, 3, 5, and 10 cm.Fig. 17(a) depicts the results of real-time detection using the 1D CNN fed by the frequencies only.At the late time of tag 1, the tag ID should be predicted to be ''100,'' however, when d = 1 and 2 cm, ''111'' was predicted, and when d = 3, 5, and 10 cm, ''000,'' ''011,'' and ''011,'' respectively, were predicted.This implies that the chipless Fig. 18(a) depicts the results of real-time detection using the 1D CNN fed by the CNRs.When d = 3, 5, and 10 cm, the predicted ID ''100'' at the late time of tag 1 was accurate.However, for d = 1 and 2 cm, the predicted IDs at the late time of tag 1 were ''111'' and ''111,'' leading to incorrect detection.The predicted IDs ''100'' and ''111'' of tag 1 and tag 2 are shown in Fig. 18(b) (zoomed-in version) at various distances from the tags.The figure demonstrates that when d = 3, 5, and 10 cm, all of the ID detections for tag 1 and tag 2 were accurate.This reveals that the proposed chipless RFID system can accurately detect multiple tags despite its detection resolution restriction of 3 cm.This demonstrates once more that the proposed chipless RFID system, which uses a 1D CNN fed by CNRs, has better multi-tag detection than the traditional one fed by frequencies only.The range resolution mainly depends upon the bandwidth of the system [31].It can be reduced by increasing the bandwidth or vice versa.In practical terms, the range resolution cannot be directly calculated from the simple equation because of the unflat antenna response.However, the range resolution of 3 cm is sufficient for smart packaging applications because the typical size of packaging such as a postal box is more than 3 cm.
From the previous investigation of the effect of the frequency shifts, the breakdown point for the frequency shift was determined to be σ = 60 MHz and f = 180 MHz.Figures 19 and 20 depict the real-time detection results for the first and second scenario experiments, respectively.In the first scenario, all predicted IDs at the late time of tag 1 were ''100,'' while at the late time of tag 2, all predicted IDs perfectly coincided with the actual eight tag IDs, as depicted in Fig. 19.In the second scenario (Fig. 20), when distance (d) was set to 3, 5, and 10 cm, the ID detections for both tag 1 and tag 2 were accurate.However, for d = 1 and 2 cm, the predicted IDs at the late time of tag 1 were ''111,'' leading to incorrect detection.This highlights that the proposed chipless RFID system can accurately detect multiple tags, even though it has a detection resolution limitation of 3 cm.In summary, based on the 100% accuracy demonstrated in Table 2, one can infer that the ID detection of the proposed system is expected to maintain its accuracy even in the face of increasing frequency shifts.However, it is important to note that we did not further increase the frequency shifts beyond the breakdown point ( = 60 MHz and f = 180 MHz) because the results in Table 2 consistently demonstrated 100% accuracy at this threshold.To this end, Table 4 provides a comparison of the proposed chipless RFID detection technique with six existing techniques [3], [8], [13], [16], [17].The techniques were categorized based on the number of bits, whether single-tag or multi-tag testing was conducted, and the reported accuracy.Using the k-NN fed by CNRs, we recently introduced the robust chipless RFID detection method [13].In experiments using a single tag, accuracy was 100%.From the table, it is evident that most of the articles focused on single-tag testing and achieved high accuracy rates in the range of 97% to 100% [3], [13], [16], [17].One notable article in [8] employed the space-time-frequency anticollision technique, which is the only existing technique that allows for a comparison of multitag detection.However, the intelligent classifier was not taken into account.Thus, there was no classification for bit IDs.Also, it did not provide the percentage of detection accuracy for multi-tag scenarios.We also can conclude from the table that the proposed chipless RFID system can deliver 100% accurate multi-tag detection.However, as demonstrated by the experimental results, the proposed chipless RFID multitag detection system was restricted to a resolution of 3 cm.

V. CONCLUSION
In this paper, a chipless RFID multi-tag detection system has been proposed.The proposed system utilized the 1D CNN as an intelligent classifier for chipless RFID multi-tag 138092 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
detection.The CNRs extracted from the tag's responses by using the STMPM have also been introduced as input datasets of the 1D CNN.Experiments with a context of a single tag were conducted to collect the CNRs and frequencies only, which were separately fed into the 1D CNN for training and validation.The experimental results obtained from the validation of the 1D CNN have demonstrated the superior performance of the use of the CNRs, achieving a remarkable accuracy of 100%, compared to the use of the frequencies only, which achieves 79.375% accuracy in tag ID detection.Experiments with a context of multiple tags were conducted in order to validate the performance of real-time multitag detection using the proposed system.The experimental results have shown that the proposed system can accurately detect multi-tags while the system using the 1D CNN fed by the frequencies only completely fails to detect multitags.However, the experimental results have shown that the proposed system was still limited to a multi-tag detection resolution of 3 cm.The future publication will showcase the performance of the proposed multi-tag detection system in more complex and real-world environments.This involves scenarios where tags are attached to containers and operated in dynamic settings.

FIGURE 1 .
FIGURE 1. Schematic diagram of the packaging on the conveyor belt scenario for chipless RFID tags.

FIGURE 2 .
FIGURE 2. Basic configuration of the proposed chipless RFID multi-tag detection system.
[I] is an identity matrix and [Y 1 ] † is the Moore-Penrose pseudoinverse of [Y 1 ].The matrix [Y 1 ] and [Y 2 ] are defined by deleting the last column and the first column from the matrix [Y T l ], respectively.The extracted CNRs are associated with corresponding bit IDs, creating a labeled dataset.Each set of CNRs is labeled with the corresponding bit ID to indicate the desired prediction output.

FIGURE 3 .
FIGURE 3. Basic architecture of the 1D CNN.
summary of the predictions made by the 1D CNN, comparing them to the actual class labels.The matrix consists of four key elements: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).True positives represent the cases where the 1D CNN has correctly predicted the positive class, while true negatives indicate correct predictions of the negative class.False positives represent the instances where the 1D CNN incorrectly predicts the positive class and false negatives represent the instances where the 1D CNN incorrectly predicts the negative class.

FIGURE 5 .
FIGURE 5. Experimental setup of the chipless RFID system in the anechoic chamber.

FIGURE 6 .
FIGURE 6. Natural frequencies of extracted CNRs representing tag IDs.

FIGURE 7 .
FIGURE 7. Damping factors of extracted CNRs representing tag IDs.

FIGURE 9 .
FIGURE 9. Structural components of the 1D CNN.
(a) and (b).Moreover, the time of training and validation is also related to processing speed.Forty epochs are sufficient to train the 1D CNN fed by the CNRs, as shown in Fig. 10(b).Up to 20% of the training and validation time can be reduced.However, training and validation time does not affect the processing time of tag detection.

FIGURE 10 .
FIGURE 10.Training accuracy and loss of the 1D CNN trained and validated with (a) the frequencies only and (b) the CNRs.

Fig. 11 (
a) and (b) illustrate the confusion matrices for the classification of IDs with eight classes, namely ''000,'' ''001,'' . . ., and ''111'' by using the 1D CNN trained by frequencies only and the CNRs in chipless RFID detection, respectively.For each tag, the total number of the input datasets employed to validate the 1D CNN in the chipless RFID detection was 40.The values along the matrix's diagonal highlighted in blue represent the TP of ID detection.If the values on the matrix's diagonal were equal to 40, 138086 VOLUME 11, 2023

FIGURE 11 .
FIGURE 11.Confusion matrix for the eight IDs of the 1D CNN trained and validated with (a) the frequencies only and (b) the CNRs.
σ = 60 MHz and f = 60 MHz, σ = 60 MHz and f = 40 MHz, and σ = 40 MHz and f = 180 MHz.These results indicate that the breakdown point for the damping factor frequency shift was σ =40 MHz, while for the natural frequency shift it was f = 160 MHz.The confusion matrices for varying damping factor shifts and natural frequency shifts are depicted in Fig. 12. Notably, the cases of σ = 60 MHz and f = 60 MHz, and σ = 60 MHz and f = 40 MHz, yield identical results, as shown in Fig. 12(a) and (c), specifically the FN occurrence for ID ''010'' with corresponding precision and recall values of 0.975 and 0.987, respectively.Conversely, in the case of σ = 40 MHz and f = 180 MHz, an FN event arises at ID ''110,'' impacting the performance metrics presented in Table

FIGURE 13 .
FIGURE 13.Real-time tag detection results obtained from the 1D CNN fed by (a) the frequencies only and (b) the CNRs.

FIGURE 14 .
FIGURE 14. Real-time detection of one tag using 1D CNN fed by CNRs under damping factor shift of σ = 60 MHz and natural frequency shift of f = 180 MHz.

FIGURE 15 .
FIGURE 15.Experimental setup of chipless RFID multi-tag detection system.

FIGURE 16 .
FIGURE 16.Real-time detection of the first-scenario experiments using the 1D CNN fed by (a) the frequencies only and (b) the CNRs.

FIGURE 19 .
FIGURE 19.Real-time detection of the first-scenario experiments using the 1D CNN fed by the CNRs under varying damping factor shift of σ = 60 MHz and natural frequency shift of f = 180 MHz.

FIGURE 20 .
FIGURE 20.Real-time detection of the second-scenario experiments using the 1D CNN fed by the CNRs under varying damping factor shift of σ = 60 MHz and natural frequency shift of f = 180 MHz (a) different distances (b) predicted IDs ''100'' and ''111.''

TABLE 1 .
Lists of the performance matrices.

TABLE 2 .
Impact of CNR shift (MHz) on accuracy percentage.

TABLE 3 .
Examples of performance matrices under varying damping factor shifts and natural frequency shifts.

TABLE 4 .
Comparison of the proposed chipless RFID multi-tag detection technique with existing approaches.