Description:FIELD OF THE INVENTION
Embodiments of the present disclosure relates Automatic Modulation Classification for OFDM Systems using Bi-stream and Attention-based CNN-LSTM Model.
BACKGROUND
Generally, Orthogonal Frequency Division Multiplexing (OFDM) is a widely used spectrum-efficient scheme and known to be a promising technology for fifth-generation (5G) communication and beyond 5G due to its high data rate communication and its effectiveness in the mitigation of multi-path effects. In an adaptive OFDM transceiver system, the transmitter selects the modulation format based on the circumstances at the transmitter end, and the receiver may need to identify the modulation format to decode the data.

SUMMARY
Embodiments of the present disclosure relate to a method and system for determining a modulation scheme associated with a signal received at a receiver. An embodiment includes receiving a signal at a receiver, wherein the signal includes an orthogonal frequency division multiplexed signal including at least in-phase, quadrature, amplitude and phase samples. A further embodiment includes determining a modulation scheme for the received signal by a module classification unit based on a spatial attribute and/or a temporal attribute wherein the modulation scheme belong to at least one of a BPSK, a QPSK, a 8-PSK, a MSK, a 16-QAM, a 64-QAM, a 128-QAM, a 512-QAM, and a 1024-QAM. Other embodiments with respect to the present disclosure are also disclosed.

BRIEF DESCRIPTION OF THE DRAWINGS
The detailed description is described with reference to the accompanying figures. Features, aspects, and advantages of the subject matter of the present disclosure will be better understood with regard to the following description and the accompanying drawings. The figures are intended to be illustrative, not limiting, and are generally described in context of the embodiments, and it should be understood that it is not intended to limit the scope of the disclosure to these particular embodiments. In the figures, the same numbers may be used throughout the drawings to reference features and components. In order that the present disclosure may be readily understood and put into practical effect, reference will now be made to exemplary embodiments as illustrated with reference to the accompanying figures. The figures together with detailed description below, are incorporated in and form part of the specification, and serve to further illustrate the embodiments and explain various principles and advantages.
Figure 1 illustrates an exemplary OFDM system in accordance with embodiments of the present disclosure.
Figure 2, which illustrates an exemplary case to facilitate effective classification, especially when dealing with higher-order modulation classification, a global attention module is incorporated in accordance with the embodiment of the present disclosure.
Figure 3A is an exemplary method for determining the modulation scheme in accordance with embodiments of the present disclosure.
Figure 3B is an exemplary method for determining the modulation scheme in accordance with the embodiments of the present disclosure.
Figure 3C is an exemplary method for determining the modulation scheme in accordance with embodiments of the present disclosure.
Figure 3D is an exemplary method for determining the modulation scheme in accordance with the embodiments of the present disclosure.
Figure 3E is an exemplary method for determining the modulation scheme in accordance with embodiments of the present disclosure.
Figure 3F is an exemplary method for determining the modulation scheme in accordance with the embodiments of the present disclosure.
Figure 4 is an exemplary illustration of the classification performance of the proposed bi-stream and attention-based CNN-LSTM method is compared with the existing DL-based AMC technique TRNN, dual stream CNN-LSTM, and model of the present disclosure without attention mechanism in accordance with embodiments of the present disclosure.
Figure 5 illustrates the impact of variations in CFO, STO, and phase offsets on the classification performance using simulation data across all SNRs in accordance with embodiments of the present disclosure.
Figure 6 illustrates a confusion matrix depicted in pertaining to Rician fading channel with a sample size of M = 2048 and SNR of 5 dB for simulation dataset in accordance with embodiments of the present disclosure.
Figure 7 illustrates a confusion matrix obtained after testing the proposed model on the real-time dataset generated through the radio frequency (RF) test bed at 10 dB SNR in accordance with embodiments of the present disclosure.
Throughout the drawings, identical reference numbers designate similar, but not necessarily identical elements. The figures as disclosed herein are not necessarily to scale, and the size of some parts may be exaggerated to more clearly illustrate the example shown. Moreover, the drawings are meant to only be provided as examples and/or implementations consistent with the description, and the description may not be limited to the examples and/or implementations provided in the drawings.

DETAILED DESCRIPTION
The following describes technical solutions in exemplary embodiments of the subject matter of the present disclosure with reference to the accompanying drawings. In this application as disclosed herein, “at least one” means one or more, and “a plurality of” means two or more. The term “and/or” describes an association relationship for describing associated objects and represents that three relationships may exist. For example, A and/or B may represent the following cases: Only A exists, both A and B exist, and only B exists, where A and B may be singular or plural. The character “/” usually indicates an “or” relationship between the associated objects. “At least one item (piece) of the following” or a similar expression thereof means any combination of the items, including any combination of singular items (piece) or plural items (pieces). For example, at least one item (piece) of a, b, or c may represent a, b, c, a and b, a and c, b and c, or a, b, and c, where a, b, and c each may be singular or plural.
It should be noted that in this application articles “a”, “an” and “the” are used to refer to one or to more than one (i.e., to at least one) of the grammatical object of the article. The terms “comprise” and “comprising” are used in the inclusive, open sense, meaning that additional elements may be included. It is not intended to be construed as “consists of only”. Throughout this specification defined above, unless the context requires otherwise the word “comprise”, and variations such as “comprises” and “comprising”, will be understood to imply the inclusion of a stated element or step or group of elements or steps but not the exclusion of any other element or step or group of elements or steps. The term “including” is used to mean “including but not limited to”. “Including” and “including but not limited to” are used interchangeably. In the structural formulae given herein and throughout the present disclosure, the following terms have been indicated meaning, unless specifically stated otherwise.
Unless otherwise defined, all terms used in the disclosure, including technical and scientific terms, have meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. By means of further guidance, term definitions are included for better understanding of the present disclosure. The term ‘about’ as used herein when referring to a measurable value such as a parameter, an amount, a temporal duration, and the like, is meant to encompass variations of ±10% or less, preferably ±5% or less, more preferably ±1% or less and still more preferably ±0.1% or less of and from the specified value, insofar such variations are appropriate to perform the present disclosure. It is to be understood that the value to which the modifier ‘about’ refers is itself also specifically, and preferably disclosed.
It should be noted that in this application, the term such as “example” or "for example" or “exemplary” is used to represent giving an example, an illustration, or descriptions. Any embodiment or design scheme described as an “example” or “for example” in this application should not be explained as being more preferable or having more advantages than another embodiment or design scheme. Exactly, use of the word such as “example” or “for example” is intended to present a related concept in only a specific manner.
It should be understood that in the embodiments of the present subject matter that “B corresponding to A” indicates that B is associated with A, and B can be determined based on A. However, it should be further understood that determining B based on A does not mean that B is determined based on only A. B may alternatively be determined based on A and/or other information.
In the embodiments of this application, “a plurality of” means two or more than two. Descriptions such as ‘first”, “second” in the embodiments of this application are merely used for indicating and distinguishing between described objects, do not show a sequence, do not indicate a specific limitation on a quantity of devices in the embodiments of this application, and do not constitute any limitation on the embodiments of this application.
Exemplary embodiments of the present disclosure related to a method and a system for determining a modulation scheme associated with a signal received at a receiver. An exemplary case includes receiving a signal 105 at a receiver 110. In an exemplary case, the signal 105 includes an orthogonal frequency division multiplexed signal including at least in-phase, quadrature, amplitude and phase samples. An exemplary case, determining a modulation scheme for the received signal by a module classification unit 130 based on a spatial attribute and/or a temporal attribute wherein the modulation scheme belong to at least one of a BPSK, a QPSK, a 8-PSK, a MSK, a 16-QAM, a 64-QAM, a 128-QAM, a 512-QAM, and a 1024-QAM.
An exemplary case includes splitting the received signal 105 at the receiver 110 is into two streams by a splitter 120. In an exemplary case, the two streams include a first part of the signal 112 and a second part of the signal 114. In an exemplary case, the first part of the signal 112 includes in-phase and quadrature (IQ) samples, and the second part of the signal 114 includes amplitude and phase (AP) sample. In an exemplary case, the two streams comprising the first part of the signal 112 and the second part of the signal 114 being kept identical.
In an exemplary case, the modulation classification unit 130 includes selecting, by an upper stream network (USN) unit 130A, wherein the first part of the signal 112 is provided as input data to the USN unit 130A. In an exemplary case, the modulation classification unit 130 includes selecting, by a lower stream network (LSN) unit 130B, wherein the second part of the signal 114 is provided as input data to the LSN unit 130B. An exemplary case includes at the USN unit, extracting spatial features associated with the first part of the signal 112 and the second part of the signal 114. An exemplary case further includes normalizing the extracted spatial features obtaining a multi-dimensional vector. An exemplary case further includes converting the multi-dimensional vector of the first part of the signal 112 and the second part of the signal 114 into a one-dimensional vector. An exemplary case at the LSN unit includes extracting temporal features associated with the first part of the signal 112 and the second part of the signal 114. An exemplary case further includes normalizing the extracted temporal features obtaining a multi-dimensional vector. An exemplary case further includes converting the multi-dimensional vector of the first part of the signal 112 and the second part of the signal 114 into a one-dimensional vector.
An exemplary case includes computing a weight associated with the received signal based on the first part of the signal 112 and the second part of the signal 114 by the module classification unit 130. An exemplary case includes acquiring the spatial features from the first part of the signal 112 and the second part of the signal 114. In an exemplary case, the spatial features include a first module with 64 kernels and a kernel size of 1×3 and a second module with 16 kernels and a kernel size of 2 ×3. An exemplary case includes performing a batch normalization on the multidimension-dimensional vector. An exemplary case includes flattening the spatial feature to obtain is a multi-dimensional vector. An exemplary embodiment includes converting a multi-dimensional vector into a one-dimensional vector.
An exemplary case includes acquiring the temporal features from the first part of the signal 112 and the second part of the signal 114, the temporal features extracted from the two LSTM module comprising: a first module with 64 memory unit and second module with 16 memory unit. A further exemplary case includes performing batch normalization on the temporal feature which is extracted from LSTM module. Yet a further exemplary case includes an attention module is used to assign higher weights to more discriminative features, which is capable for distinguishing between higher order modulation schemes. A further exemplary case includes again flattening the features to obtain is a one-dimensional tensor from multidimension tensor.
A further exemplary case includes concatenating the spatial features and the temporal features from the USN unit 130A and the LSN unit 130B and providing the concatenated features to a decision-making module 135, wherein the decision-making module is a part of the module classification unit 130. A further exemplary case includes the decision-making module 135 being configured to predict the modulation classification.
A further exemplary case includes the modulation classification unit 130 configured for extracting, by a convolutional neural network (CNN) block of the USN unit and the LSN unit, the spatial features from the first part and the second part, and normalizing the spatial features. Yet a further exemplary case includes extracting, by a long short-term memory network (LSTM) layer, temporal features from the first part and from the second part, and normalizing the temporal features. Yet a further exemplary case includes computing, by an attention mechanism layer, weights to be assigned to the input signals, wherein more weightage is assigned to most relevant feature based on the input data. Yet a further exemplary case includes providing an output based on the computed weights to a decision-making module. A further exemplary case includes a real-time dataset generated via a radio frequency (RF) testbed was employed for predicting the modulation schemes.
In an exemplary case, orthogonal frequency division multiplexing (OFDM) may be found to be widely used in spectrum-efficient scheme is pioneering technology for fifth-generation (5G) and beyond due to its high data rate communication and its effectiveness in the mitigation of multi-path effects. In an exemplary case, an adaptive OFDM transceiver system, the transmitter selects the modulation format based on the pre-defined circumstances, and the receiver identifies the modulation format to decode the data. In an exemplary case, automatic modulation classification (AMC) will be crucial in sixth-generation (6G) communication, where the transceiver may need to be adaptive in nature. In an exemplary case, to provide more adaptiveness to the system with a high data rate, an increasing number of modulation schemes, along with higher-order modulation formats, need to be used.
In an exemplary case, in 6G, AMC may be a backbone structure enabling ultra-low latency communication, more efficient communication, and improved security, while adaptation and intelligent communication by reducing overhead/control signals should be a primary focus. In an exemplary case, designing an AMC algorithm may be difficult due to channel impairments such as channel state information (CSI), carrier frequency offset (CFO), symbol timing offset (STO), and phase offset. In an exemplary case, these impairments cause non-orthogonality among the subcarriers and introduce inter-symbol interference (ISI), which results in degrading the performance of AMC for the OFDM-modulated signal.
In an exemplary case, there are several AMC techniques available for OFDM systems. In an exemplary case, discrete Fourier transform (DFT) and normalized higher-order cumulant- based method have been proposed for the classification of lower-order digital modulation formats for OFDM modulated signal. In an exemplary case, a convolutional neural network (CNN) including two convolutional layers and three fully connected layers may be used to handle in-phase and quadrature (IQ) samples for OFDM modulated signal and achieve relatively accurate modulation classification. In an exemplary case, the method does not consider the synchronization issues for the data generation, rendering it impractical for practical applications. In an exemplary case, the impact of CFO on OFDM modulated signal may be addressed and removed to enhance the modulation classification accuracy. In an exemplary case, under severe channel conditions, it may still exhibit unsatisfactory classification accuracy. In an exemplary case, in accordance with the embodiments of the present disclosure, an AMC technique for real-time OFDM signals has been designed that combines a deep residual network with a triple-skip residual stack (TRNN) to achieve better classification results under dynamic fading channel conditions. In an exemplary case, this technique considers only phase-shift keying (PSK) modulation formats for the classification.
In an exemplary case, several deep learning (DL) based AMC techniques for single-carrier systems may be used, wherein the decentralized learning-based AMC (DecentAMC) method and the neural architecture search (NAS)-based AMC employ distributed learning to reduce communication overhead and enhance overall system performance. In another exemplary case, the method includes the ResNet-based AMC which has demonstrated strong performance on the RadioML 2018.01A dataset designed for single-carrier systems. In an exemplary case, there is a gap in the existing techniques and methods for designing an AMC for OFDM systems, particularly for higher-order modulation classes relevant to Wi-Fi 6 and 6G applications, which addressed all three critical synchronization issues, including STO, CFO, phase offset with unknown CSI, and signal parameters, simultaneously.
In an exemplary case, a AMC algorithm is proposed for an M-ary quadrature amplitude modulation (M-QAM) for asynchronous OFDM-modulated signals. In the exemplary case, the key contributions of the proposed technique are as follows:
(a) The use of multimodal information such as imaginary component, real component, amplitude, and phase extracted from the received signal to carry out the AMC of the OFDM system.
(b) A bi-stream and attention-based convolutional neural network and long short-term memory (CNN-LSTM) based AMC to extract special and temporal features from the IQ samples and the amplitude and phase (AP) of the received signal.
(c) Work in the presence of STO, CFO, and phase offsets over frequency-selective fading channels without prior knowledge of CSI and signal parameters.
(d) Advanced communication systems use modulation schemes such as binary phase-shift keying (BPSK), quadrature PSK (QPSK), 8-PSK, minimum shift keying (MSK), and M-QAM for M = 16, 64, 128, 512, and 1024, which can be classified using the methodology implemented herein.
(e) Simulation results demonstrate that the multimodal technique outperforms conventional single-modality methods, even when classifying higher-order digital modulation types, and
(f) Finally, the model has been validated on the real-time dataset generated by the radio frequency (RF) testbed.
Reference is now made to Figure 1, which illustrates an exemplary OFDM system in accordance with embodiments of the present disclosure. Receiver 110 is configured to receive signals 105, that are transmitted from a transmission end, wherein the signal contains information that needs to be decoded at the receiver end. Transmitter may use any of the modulation schemes and it is imperative that the receiver needs to decode the modulation scheme to decode the information received before presenting the information to a user. System 120 contains a splitter wherein the received signal is split into two streams, a first part of the signal 112 and a second part of the signal 114 by splitter 120. The modulation unit 130 processes the upper stream signal 112 and the lower stream signal 114 to ascertain features and feed the features extracted to a de-coding unit 135, which is configured to predict the modulation scheme used by the transmitter. Figure 1 will be explained elaborately below.
The received baseband OFDM signal has K data subcarrier after passing through a frequency-selective fading channel with impulse response, g[l] being the length L, and oversampled by a factor of ? = N/K, which may be represented as

where s ¯_? is the OFDM signal with cyclic prefix (CP), ? ? {BPSK, QPSK, 8-PSK, MSK, 16-QAM, 64-QAM, 128-QAM, 512-QAM, 1024-QAM}, ? represents the normalized CFO, ? represents the phase offset, ? represents the STO, Ns is the length of the OFDM symbol with CP, Ns = N + Ncp, where Ncp is greater than or equal to L, and v[n] represents the additive white Gaussian noise (AWGN) characterized by a mean of zero and a variance s^2.
Modulation classification methods typically rely on a single mode of information, with IQ data being the preferred choice due to its ability to fully characterize the signal content. However, in order to better leverage the capabilities of CNNs for learning signal characteristics, additional modalities of the received OFDM signals are also considered, including AP components. In order to facilitate comparison of the processing gains, the components have been classified into two distinct modalities. The first modality is referred to as M_1^IQ that includes the real and imaginary parts of the received signals. The second modality is referred to as M_2^AP, which encompasses AP of the received signals. By dividing the components into these two modes, the processing gains may be easily assessed and analysed without being burdened on any extraneous factors.
The components within each modality have been normalized and then combined into a matrix, which serves as the input for a CNN. Specifically, the matrices for each modality may be expressed as follows
,

where

is the real, imaginary, amplitude and phase of the received signal

and .
In an exemplary case, a bi-stream and CNN-LSTM-based architecture with an attention mechanism (CNN-LSTM-AM), as depicted in Figure 1 for AMC of OFDM signals processing has been considered. The architecture consists of two different streams for extracting features from two different modalities of the OFDM signal. CNN is not effective at capturing the temporal aspects of the time series data, but LSTM is capable of fusing the time characteristics of signals to a greater extent. Therefore, combining CNN with LSTM is an approach to extracting spatial as well as temporal features from the received signal data. The architecture of the bi-stream CNN-LSTM-AM network used for AMC of OFDM signals is explained below.;
Exemplary architecture of the bi-stream CNN-LSTM model based on the attention mechanism (AM) for AMC is illustrated in Figure 1. The architecture depicts two steams, which are for two different modalities of the received OFDM signals and are trained in an end-to-end fashion. The upper steam network (USN) takes IQ samples as input, whereas the lower stream network (LSN) takes AP samples of the received OFDM signals as input.
Figure 1 further illustrates an exemplary case of the stream of the proposed network kept identical, which enables shared representation learning, parameter efficiency, transfer learning, intermodal fusion, and interpretability in the context of multimodal deep learning for multi-class classification in accordance with the embodiment of the present disclosure. The CNN block of the USN is used for the extraction of the spatial features from the input IQ sample data, whereas the CNN block of the LSN is used for the spatial feature extraction from the AP sample data. The LSTM layer, connected after the CNN layer in both streams, uses the feature maps to extract the temporal features. The AM layer, connected after the CNN-LSTM block, calculates the weight based on the input data.
The network architecture that incorporates the AM into the CNN-LSTM structure is illustrated in Figure 1. The CNN block consists of two CNN layers connected back-to-back to efficiently learn the spatial features from the input data. They learn spatial characteristics from various modulation signal representations by employing a collection of learnable filters and activation functions. The convolutional layers of the CNN block contain 64 and 16 kernels, each having kernel sizes of 1×3 and 2×3, respectively. Subsequently, the feature maps obtained after the convolutional layers undergo the rectified linear unit (ReLU) activation function to introduce non-linearity to the network. A batch normalization (BN) layer is added after the convolutional block to avoid overfitting. In order to exploit the temporal correlation, two LSTM layers are incorporated subsequent to the convolutional block. The feature maps collected from the end of CNN modules must be flattened before passing it through the LSTM module. So, a flattened layer is added after the BN layer, which converts the multidimensional tensor into a one-dimensional vector.
The architecture contains two LSTM modules to keep track of long-term dependencies because modulation schemes often exhibit temporal patterns that span over multiple time steps. The LSTM module used for building our final architecture contains 64 and 16 memory units, respectively. The hyperbolic tangent activation function follows both modules. After that, features from the LSTM block stack are passed through the BN layers and applied as input to the attention layer.
Reference is now made to Figure 2, which illustrates an exemplary case to facilitate effective classification, especially when dealing with higher-order modulation classification, a global attention module is incorporated. This module assigns higher weights to more discriminative features, significantly enhancing our ability to distinguish between higher-order modulation classes. The equations that correspond to the intermediate and final outputs of the attention layer are represented as

where eatt is the feature vector used as input to the dense layer via attention, aatt, catt, and matt represent the intermediate outputs of the attention layer operations, and represent the Hadamard product. After that, the features obtained from each attention layer are concatenated, allowing for effective intermodal fusion to exploit the combined information from each modality and to boost the model’s understanding and decision-making capability, and then given to the final classification layer. The classification layer has softmax as the final activation, which predicts the modulation classes. The determination of optimal hyperparameters, including the kernel size, the number of filters in the convolution layer, the activation function at each layer, the number of hidden layers, and the number of LSTM units, is accomplished through search of the architecture space utilizing the Bayesian method.
In an exemplary case, data generation includes Dataset with IQ samples, AP of the received baseband OFDM signal, and the corresponding label indicating the modulation class. The modulation schemes under consideration include BPSK, QPSK, 8-PSK, MSK, 16-QAM, 64-QAM, 128-QAM, 512-QAM, and 1024-QAM. For each modulation, examples are generated for data subcarriers with K = 16 and signal-to-noise ratio (SNR) values ranging from -10 dB to 20 dB in 5 dB steps. There are 4096 examples for each SNR value for each modulation class. The dataset used is split into two sets: 70% of the data is dedicated to training, while the remaining 30% is designated for validation. The test set contains 512 examples for each modulation class per SNR value. To improve the dataset’s representation, an oversampling factor of ? = 4 was applied, and a CP of NCP = 16 samples is used. The propagation environment is simulated using Rician fading channels with the ITU-R power delay profile, designed to replicate outdoor-to-indoor and pedestrian test channels, aiming to provide a more realistic simulation. Distribution of synchronization parameter normalized and phase offset
The experimental validation of the proposed bi-stream and attention-based CNN-LSTM model has been tested on the real-time dataset generated through the RF testbed. The experimental setup for AMC using a bi-stream and attention-based CNNLSTM model is well known. The configuration of the transmitter and receiver includes a direct line-of-sight (LOS) path, which allows it to be modelled as a Rician distribution. A real-time sample length of 2048 for every modulation scheme is generated. All other parameters are kept the same as in the simulation dataset.
In the training process, the model is trained using data samples with SNR values of -10 dB to 20 dB for 150 epochs. The optimization algorithm employed is Adam, which leverages the strength of both AdaGrad and RMSProp, rendering it particularly suitable for handling noisy or sparse gradients. Initially, a learning rate of 0.01 is adopted for the first 50 epochs, and then it is reduced to 0.001 for the subsequent epochs. The training is conducted with a batch size of 1024, while the model architecture remains independent of the number of samples in the input signal, ensuring consistency across signals of length M = 2048 samples. The categorical cross-entropy is used as the loss function for classification tasks involving multiple classes because of their computational effectiveness and fast convergence. The loss function measures the difference between the predicted and actual probability distributions of the output classes. By minimizing the categorical cross-entropy, the model’s accuracy can be improved during the training process. The loss function can be expressed as

Where ? represents the number of modulation classes, Nb represents the number of training examples in the dataset. The model is tested on signals with length M of 2048 samples.
The classification performance of the proposed bi-stream and attention-based CNN-LSTM method is compared with the existing DL-based AMC technique TRNN, dual stream CNN-LSTM, and model of the present disclosure without attention mechanism, as depicted Figure 4, wherein the results clearly indicate that the method of the present disclosure achieves superior performance in terms of classification accuracy compared to other techniques. Specifically, the proposed method has improved accuracy by 25.4% and 6%, respectively, at 10 dB SNR, compared to the TRNN and dual stream CNN-LSTM model. As illustrated, the proposed bi-stream CNN-LSTM approach with global attention outperforms the model with channel attention, achieving classification accuracies of 93.1% and 89.75% at an SNR of 20 dB, respectively.
Figure 3A is an exemplary method for determining the modulation scheme in accordance with embodiments of the present disclosure. Step 310 includes receiving a signal at a receiver. The signal being received at a receiver 110, wherein the signal is an orthogonal frequency division multiplexed signal, the signal comprising in-phase, quadrature, amplitude and phase samples. Step 320 includes determining the modulation scheme associated with the signal. Determining a modulation scheme for the received signal by a module classification unit 135, based on a spatial attribute and/or a temporal attribute wherein the modulation scheme belong to at least one of a BPSK, a QPSK, a 8-PSK, a MSK, a 16-QAM, a 64-QAM, a 128-QAM, a 512-QAM, and a 1024-QAM. Details of this has been explained with previously and with respect to Figures 4 – 7.
Figure 3B is an exemplary method for determining the modulation scheme in accordance with the embodiments of the present disclosure. In step 330, the received signal is split into two streams, a first part of the signal 112 and a second part of the signal 114. The first part of the signal comprises in-phase and quadrature (IQ) samples, and the second part of the signal comprises amplitude and phase (AP) sample. The two streams including the first part of the signal and the second part of the signal are kept identical. In step 332, the first part of the signal is provided to the upper stream network for processing and in step 334, the second part of the signal is provided to the lower stream network for processing. In the exemplary case, selecting, by an upper stream network (USN) unit 130A, wherein the first part of the signal 112 is provided as input data to the USN unit 130A, and selecting, by a lower stream network (LSN) unit 130B, wherein the second part of the signal 114 is provided as input data to the LSN unit 130B. Details of this has been explained with previously and with respect to Figures 4 – 7.
Figure 3C is an exemplary method for determining the modulation scheme in accordance with embodiments of the present disclosure. Step 342 includes at the USN unit 130A extracting spatial features associated with the first part of the signal 112 and the second part of the signal 114, wherein the spatial features are extracted using CNN block. Step 344 includes normalizing the extracted spatial features obtaining a multi-dimensional vector. Step 346 includes converting the multi-dimensional vector of the first part of the signal 112 and the second part of the signal 114 into a one-dimensional vector. Details of this has been explained with previously and with respect to Figures 4 – 7.
Figure 3D is an exemplary method for determining the modulation scheme in accordance with the embodiments of the present disclosure. Step 352 includes extracting temporal features associated with the first part of the signal 112 and the second part of the signal 114, wherein the temporal features are extracted using the LSTM block. Step 354 includes normalizing the extracted temporal features obtaining a multi-dimensional vector. Step 356 includes converting the multi-dimensional vector of the first part of the signal 112 and the second part of the signal 114 into a one-dimensional vector. Details of this has been explained with previously and with respect to Figures 4 – 7.
Figure 3E is an exemplary method for determining the modulation scheme in accordance with embodiments of the present disclosure. Step 362 includes applying the attention module to obtain the most relevant features by processing the spatial features by computing weights associated with the signal based on the first part of the signal and the second part of the signal, acquiring the spatial features which includes a first module with 64 kernels of size 1×3 and a second module with 16 kernels of size 2×3, and processing the temporal features by computing weights associated with the signal based on the first part of the signal and the second part of the signal, acquiring the temporal features which includes 64 memory unit and a 16 memory unit, then performing a batch normalization on the multi-dimensional vector, and Step 364 includes flattening the spatial features and the temporal features of the multi-dimensional vector and converting the multi-dimensional vector to a single dimensional vector. Details of this has been explained with previously and with respect to Figures 4 – 7.
Figure 3F is an exemplary method for determining the modulation scheme in accordance with the embodiments of the present disclosure. Step 366 includes concatenating the features from both the upper stream and the lower stream and Step 368 includes modulation prediction by the classification unit. Details of this has been explained with previously and with respect to Figures 4 – 7. Concatenating the spatial features and the temporal features from the USN unit 130A and the LSN unit 130B and the concatenated features are provided to a decision-making module 135, wherein the decision-making module is a part of the module classification unit 130, wherein the classification unit is configured for predicting the modulation classification scheme. Details of this has been explained with previously and with respect to Figures 4 – 7.
As illustrated in Figure 4, the model of the present disclosure outperforms ResNet and feature-based (FB) AMC. The reason for this is that ResNet was not designed to perform effectively with higher order modulation schemes such as 512-QAM and 1024-QAM. Furthermore, FB approach exhibits limited generalization of extracted features when additional modulation classes are introduced. Moreover, it is evident that both the classification accuracies of the signal stream-IQ and the single stream-AP are inferior to the method of the present disclosure. This superior performance is achieved through the synergistic interaction of features extracted from two distinct modalities of the received signal.
It can be observed from Figure 4 that the performance the model of the present disclosure on a real-time dataset exhibits a slightly lower classification accuracy compared to the simulation dataset. This is because of the channel environment, which differs somewhat between simulations and measurements, as well as the different RF hardware constraints, such as IQ gain imbalance, quadrature skew, antenna gains, and DC offsets. The experimental results depicted in Figure 4 demonstrate that the proposed bi-stream attention-based CNN-LSTM model outperforms the existing DL-based AMC method TRNN method, dual stream CNN-LSTM in the real-world indoor environment. At 10 dB SNR, the method of the present disclosure achieves a 26.65% higher accuracy compared to the TRNN and 11% higher accuracy compared to the dual-stream CNN-LSTM model. Therefore, the bi-stream attention-based CNN-LSTM method of the present disclosure can be applied in real-world deployments of AMC in future generations of wireless communication systems.
Reference is now made to Figure 5, which illustrates the impact of variations in CFO, STO, and phase offsets on the classification performance using simulation data across all SNRs. The impact of individual impairments on classification accuracy was evaluated. In the exemplary case, one impairment was isolated while setting the values of the other impairments to zero. The outcomes reveal that, for each specific impairment, the variation in classification accuracy remains well within the range of ±2%. Furthermore, in datasets containing only STO, the model of the present disclosure exhibits a marginally lower classification accuracy compared to datasets that include CFO and phase offset.
Reference is now made to Figure 6, which illustrates a confusion matrix depicted in pertaining to Rician fading channel with a sample size of M = 2048 and SNR of 5 dB for simulation dataset. It may be observed that all modulation classes achieve over 82% accuracy at 5 dB SNR, excluding 128-QAM, 512-QAM, and 1024-QAM. This is because the 128-QAM constellation is included in the larger QAM modulations, resulting in confusion with 128-QAM, 512-QAM, and 1024-QAM due to increased noise at lower SNR.
Reference is now made to Figure 7, which illustrates a confusion matrix obtained after testing the proposed model on the real-time dataset generated through the RF testbed at 10 dB SNR. It can be observed that the proposed model in accordance with the present disclosure shows slightly lower classification accuracy for the real-time dataset due to varied channel conditions and RF hardware constraints. Additionally, it may be observed that only higher-order QAMs get slightly confused with their subclasses. The classification accuracy can be improved by the statistical behaviour of the user’s data rate and SNR estimation technique, as higher-order QAM modulations exhibit different SNR thresholds for ensuring reliable communication.
In an exemplary case, the comparative computational complexity between the proposed AMC model and the existing DL-based AMC approach is presented in Table I below.
TABLE 1

The computational complexity for each convolutional layer is expressed as O(Zi Fi M di), where Zi denotes the filter size, M × di represents the size of the input, and Fi signifies the number of filters utilized in the ith convolutional layer. The variable p corresponds to the total number of convolutional layers in the model. The computational complexity of the ith LSTM unit, denoted as O(Wi), depends on the total number of parameters represented by W. The variable q indicates the number of LSTM layers in the architecture. The computational complexity of each attention module can be expressed as O(n2d), where n represents the length of the input sequence, and d represents the depth of the attention module. The variable r represents the number of hidden dense layers, while ci × ei denotes the input feature map for each hidden dense layer. The use of a dash notation ()’, ()’’, and ()’’’ in the complexity term of CNN-LSTM. TRNN-based method and ResNet corresponds to the same interpretation as outlined in the description of the proposed model. CNN-LSTM uses 3 convolution layers, i.e., p’ > p and F_i^'> F_i resulting in an increase in trainable parameters. Consequently, CNN-LSTM is more computationally and time complex than the proposed model of the present disclosure. Additionally, since r’’ >> r and t = 7, number of residual stacks makes the TRNN-based method computationally more complex compared to the proposed model. ResNet contains p’’’ = 32 (>>p) convolutional layers and 128 hidden neurons in two hidden layers (r’’’> r), which also makes it more computationally complex than the proposed method of the present disclosure.
The proposed method in accordance with the present disclosure has 69,609 trainable parameters, whereas TRNN has 124,183 parameters, and CNN-LSTM employs a massive 10 million parameters. This demonstrates that the bi-stream attention-based CNN-LSTM method as disclosed in accordance with the resent disclosure offers lower computational complexity and requires fewer trainable parameters compared to TRNN, CNN-LSTM, and ResNet, resulting in reduced time complexity. While the FB approach has a computational complexity less than our proposed model of the present disclosure, it relies on handcrafted features that do not scale effectively with additional modulation classes. Thus the proposed model in accordance with the present disclosure can lead to faster processing and improved memory efficiency during both training and inference, offering a contrast to existing DL-based techniques
Embodiments of the present disclosure relate to a novel AMC technique for a higher-order modulation format for OFDM signals using a bi-stream CNN-LSTM-AM. The signals are pre-processed into temporal IQ and AP representations to improve feature classification. This data-driven technique can recognize signals with CFO, STO, and phase offset even without knowledge of signal or channel parameters. The proposed technique performs better in terms of classification accuracy and computational complexity than traditional DL-based methods. The proposed model has been validated by simulation and the real-time dataset generated by using the RF testbed.
Although the present disclosure has been described with reference to several preferred embodiments, it should be understood that the present disclosure is not limited to the preferred embodiments disclosed here. Embodiments of the present disclosure are intended to cover various modifications and equivalent arrangements within the spirit and scope of the appended claims. Although the foregoing disclosure has been described in some detail for purposes of clarity of understanding, it will be apparent that certain changes and modifications may be practised within the scope of the appended claims. Examples of the present disclosure have been described in language specific to structural features and/or methods. It should be noted that there are many alternative ways of implementing both the process and apparatus of the present invention. Accordingly, embodiments of the present disclosure are to be considered illustrative and not restrictive, and the invention is not to be limited to the details given herein but may be modified within the scope and equivalents of the appended claims. It should be understood that the appended claims are not necessarily limited to the specific features or methods described. Rather, the specific features and methods are disclosed and explained as examples of the present disclosure , C , Claims:WE CLAIM:
A method for determining a modulation scheme associated with a signal, the method comprising:
receiving a signal 105 at a receiver 110, wherein the signal 105 comprises an orthogonal frequency division multiplexed signal including at least in-phase, quadrature, amplitude and phase samples;
determining a modulation scheme for the received signal by a module classification unit 130 based on a spatial attribute and/or a temporal attribute wherein the modulation scheme belong to at least one of a BPSK, a QPSK, a 8-PSK, a MSK, a 16-QAM, a 64-QAM, a 128-QAM, a 512-QAM, and a 1024-QAM.
The method as claimed in claim 1, the method comprising:
splitting the received signal 105 at the receiver 110 is into two streams by a splitter 120 wherein the two streams comprises:
a first part of the signal 112 and a second part of the signal 114,
the first part of the signal 112 comprises in-phase and quadrature (IQ) samples, and the second part of the signal 114 comprises amplitude and phase (AP) sample, and
the two streams comprising the first part of the signal 112 and the second part of the signal 114 being kept identical.
The method as claimed in claim 1, wherein the modulation classification unit 130 comprises:
selecting, by an upper stream network (USN) unit 130A, wherein the first part of the signal 112 is provided as input data to the USN unit 130A;
selecting, by a lower stream network (LSN) unit 130B, wherein the second part of the signal 114 is provided as input data to the LSN unit 130B.
The method as claimed in claim 1, the method comprising:
performing at the module classification unit 130 by the USN unit 130A
extracting spatial features associated with the first part of the signal 112 and the second part of the signal 114;
normalizing the extracted spatial features obtaining a multi-dimensional vector; and
converting the multi-dimensional vector of the first part of the signal 112 and the second part of the signal 114 into a one-dimensional vector.
The method as claimed in claim 1, the method comprising:
performing at the module classification unit 130 by the LSN unit 103B
extracting temporal features associated with the first part of the signal 112 and the second part of the signal 114;
normalizing the extracted temporal features obtaining a multi-dimensional vector; and
converting the multi-dimensional vector of the first part of the signal 112 and the second part of the signal 114 into a one-dimensional vector.
The method as claimed in claims 4 and 5, the method comprising:
computing a weight associated with the signal based on the first part of the signal 112 and the second part of the signal 114 by the module classification unit 130.
The method as claimed in claim 4, the method comprises:
acquiring the spatial features from the first part of the signal 112 and the second part of the signal 114, the spatial features comprising:
a first module with 64 kernels and a kernel size of 1×3 and a second module with 16 kernels and a kernel size of 2×3,
The method as claimed in claim 7, the method comprising:
performing batch normalization on the multidimension-dimensional vector.
The method as claimed in claim 8, the method comprising:
flattening the spatial feature to obtain is a multi-dimensional vector; and
converting a multi-dimensional vector into a one-dimensional vector.
The method as claimed in claim 5, the method comprising:
acquiring the temporal features from the first part of the signal 112 and the second part of the signal 114, wherein the temporal features extracted from the two LSTM modules comprising: a first module with 64 memory unit and second module with 16 memory unit.
The method as claimed in claim 10, the method comprising:
performing batch normalization on the temporal feature extracted from LSTM module ; and
assigning higher weights to more discriminative features at the attention module, wherein the attention module is capable for distinguishing between higher order modulation schemes
The method as claimed in claim 11, the method comprising:
flattening the features to obtain is a one-dimensional tensor from multidimension tensor.
The method as claimed in claim the preceding claims, the method comprises:
concatenating the spatial features and the temporal features from the USN unit 130A and the LSN unit 130B and providing the concatenated features to a decision-making module 135, wherein the decision-making module is a part of the module classification unit 130.
The method as claimed in claim 10, wherein the decision-making module 135 is configured to predict the modulation classification.
The method as claimed in claim 1, wherein the modulation classification unit 130 comprises:
extracting, by a convolutional neural network (CNN) block of the USN unit and the LSN unit, the spatial features from the first part and the second part, and normalizing the spatial features;
extracting, by a long short-term memory network (LSTM) layer, temporal features from the first part and from the second part, and normalizing the temporal features;
computing, by an attention mechanism layer, weights to be assigned to the input signals, wherein more weightage is assigned to most relevant feature based on the input data; and
providing an output based on the computed weights to a decision-making module.
The method as claimed in claims 1 - 15, wherein a real-time dataset generated via a radio frequency testbed was employed for predicting the modulation schemes.
A system configured to perform the method as claimed in any of the preceding claims 1 – 16.

Dated this 10th day of April 2024 Indian Institute of Science
By their Agent & Attorney

Dr. Eric W B Dias/Reg No 1058
of Khaitan & Co