Deep neural network and long short-term memory-based successive interference cancellation for non-orthogonal multiple access systems

Abstract

A system for detecting and separating overlapping signals in a non-orthogonal multiple access (NOMA) communication system, comprising a deep neural network (DNN) and a long short-term memory (LSTM) model configured to process the received signal data and perform interference cancellation in real time; where the DNN extracts high-level features from the received signal to aid signal detection and separation, and the LSTM captures temporal dependencies within the signal data, enabling improved performance in dynamic environments.

Description

AREA OF INVENTION

[0001] This invention relates to the detection and suppression of successive interference for non-orthogonal multiple access using deep neural networks (Long Short-Term Memory). The present invention relates to the field of wireless communication systems, in particular the detection and suppression of interference in non-orthogonal multiple access (NOMA) systems. More specifically, the invention utilizes deep neural networks (DNNs) and long short-term memory (LSTM) networks for successive interference suppression (SIC) in multi-input / multi-output (MIMO) NOMA systems. The invention addresses challenges associated with interference in high-density, high-speed communication environments, such as 5G / 6G networks, by providing an efficient real-time solution for detecting and separating overlapping signals from multiple users. BACKGROUND OF THE INVENTION

[0002] The proposed invention addresses the problem of efficiently detecting and separating overlapping signals transmitted by multiple users in modern communication systems. With conventional methods, this process is slow, error-prone, and struggles to keep pace with rapidly changing conditions, particularly in real-time environments such as 5G / 6G networks. This invention overcomes these problems by utilizing Deep Neural Networks (DNNs) and Long Short-Term Memory (LSTM) networks for signal detection in multi-input, multi-output, non-orthogonal multiple access (MIMO-NOMA) systems. The DNN excels at processing complex, non-linear representations of signals, while the LSTM is adept at processing time-sequential data, making it highly effective in dynamic environments.

[0003] The aim of this invention is to improve the speed, accuracy, and reliability of signal detection, thereby overcoming the challenges of high computing power, latency, and performance issues in dynamic environments. Through a more efficient approach, the invention ensures better performance in modern wireless communication systems.

[0004] Several commercially available products utilize Successive Interference Cancellation (SIC) and other traditional methods for signal detection and separation in the case of overlapping user signals. Some notable products include: Huawei 5G C-Band Massive MIMO AAU, Ericsson AIR 6488, Nokia AirScale Base Station, Qualcomm Snapdragon X60 5G Modem RF System, MediaTek Dimensity 1000 Series 5G SoCs

[0005] Existing CNN-LSTM networks for downlink signal detection systems can only process structured spatial inputs. They may have problems with long-term temporal patterns and higher computational costs due to convolutional layers.

[0006] Convolutional Neural Networks (CNNs) are good at capturing spatial patterns but struggle to capture long-term temporal dependencies. When combined with LSTM, CNNs focus only on short-term relationships, potentially failing to fully capture the complexity of time-varying signals in NOMA. DNN-LSTM is best suited for time-series data and temporal dependencies and can process unstructured or low-dimensional data. CNN-LSTM, on the other hand, can only process structured spatial inputs (2D data).

[0007] Convolutional Neural Networks (CNNs) are good at capturing spatial patterns but struggle to capture long-term temporal dependencies. When combined with LSTM, CNNs focus only on short-term relationships, potentially failing to fully capture the complexity of time-varying signals in NOMA. DNN-LSTM is best suited for time-series data and temporal dependencies and can process unstructured or low-dimensional data. CNN-LSTM, on the other hand, can only process structured spatial inputs (2D data). Summary of the invention

[0008] Traditional SiC involves multiple iterations, processing each user in the network sequentially, which increases complexity. Our solution avoids this by using DNN-LSTM, which learns to recognize interference patterns in a single pass. The DNN extracts higher-order features from the received signal, while the LSTM layers capture the temporal dependencies, enabling real-time signal detection without multiple iterations. This results in significantly lower computational overhead compared to traditional SiC. Our proposed solution introduces a deep learning-based SiC framework that uses deep neural networks (DNNs) and long short-term memory (LSTM) networks to mitigate these issues. A DNN-LSTM model is developed to process the received signal data.The DNN layers are responsible for extracting the key features from the signal, while the LSTM layers model the temporal dependencies in the signal data. The model is trained under supervision, enabling it to predict interference patterns and accurately identify the desired signal for each user. During training, the model optimizes its weights to minimize signal detection error, thus enabling good generalization to real-world data. The DNN-LSTM network reduces computational complexity, predicts and eliminates interference for all users simultaneously, and therefore leads to the immediate detection of the desired signal. The sequential processing step is eliminated, resulting in faster signal detection. This algorithmic deep learning approach can replace existing networks, thus enabling improved signal detection. DETAILED DESCRIPTION OF THE INVENTION

[0009] Traditional SiC involves multiple iterations, processing each user in the network sequentially, which increases complexity. Our solution avoids this by using DNN-LSTM, which learns to recognize interference patterns in a single pass. The DNN extracts higher-order features from the received signal, while the LSTM layers capture the temporal dependencies, enabling real-time signal detection without multiple iterations. This results in significantly lower computational overhead compared to traditional SiC. Our proposed solution introduces a deep learning-based SiC framework that uses deep neural networks (DNNs) and long short-term memory (LSTM) networks to mitigate these issues. A DNN-LSTM model is developed to process the received signal data.The DNN layers are responsible for extracting the key features from the signal, while the LSTM layers model the temporal dependencies in the signal data. The model is trained under supervision, enabling it to predict interference patterns and accurately identify the desired signal for each user. During training, the model optimizes its weights to minimize signal detection error, thus enabling good generalization to real-world data. The DNN-LSTM network reduces computational complexity, predicts and eliminates interference for all users simultaneously, and therefore leads to the immediate detection of the desired signal. The sequential processing step is eliminated, resulting in faster signal detection. This algorithmic deep learning approach can replace existing networks, thus enabling improved signal detection.

[0010] DNN-LSTM models are better able to detect complex, nonlinear relationships in the signal, resulting in more precise interference suppression and the ability to handle an increasing number of users without a significant increase in computational complexity.

[0011] The present invention relates to a system for detecting and separating overlapping signals in a NOMA (Non-Orthogonal Multiple Access) communication system using a Deep Neural Network (DNN) and a Long Short-Term Memory (LSTM) model. The system overcomes the challenges of conventional SIC (Successive Interference Cancellation) methods in the efficient processing of multi-user signals, particularly in dynamic environments such as modern 5G and 6G networks. By leveraging the strengths of DNN and LSTM, the invention significantly improves the accuracy, speed, and computing power of signal detection in NOMA systems.

[0012] The core of the system is the DNN, which is responsible for extracting high-level features from the received signal data. These features help identify and isolate the desired signal amidst overlapping signals from multiple users. Complementing the DNN, the LSTM network captures the temporal dependencies within the signal data, enabling the system to effectively process time-series data. This combination of DNN and LSTM allows the system to detect interference patterns and separate user signals with high precision, making it well-suited for real-time communication scenarios.

[0013] One of the main advantages of this system is its ability to process received signal data in a single pass. This is a significant advantage over traditional SiC methods, which rely on multiple iterations to sequentially process and separate each user's signals. This reduces computational overhead, making the system faster and more efficient. The DNN-LSTM model is supervised trained, with its weights optimized to minimize signal detection errors. This training process ensures that the model generalizes well to real-world signal data, thereby improving its robustness and reliability in real-world applications.

[0014] The system is designed to detect and eliminate interference for all users in a NOMA network simultaneously. By processing all user signals in a single pass, the DNN-LSTM model can instantly identify the desired signal and eliminate the delays associated with sequential processing. This results in faster signal detection, which is crucial for applications in dynamic and high-speed wireless communication systems such as 5G and 6G networks.

[0015] Another significant advantage of the invention is its scalability. If the number of users in a NOMA system increases, the DNN-LSTM model can handle the additional complexity without a significant increase in computational load. This scalability makes the system ideal for large networks where a high number of users need to be supported efficiently, for example, in urban or dense network environments.

[0016] In addition to its speed and accuracy, the DNN-LSTM model excels at capturing long-term temporal dependencies. This makes it particularly effective in dynamic communication environments where signal conditions can change rapidly. While conventional Convolutional Neural Network (CNN)-LSTM models effectively capture spatial patterns, they struggle with long-term temporal dependencies. The DNN-LSTM, on the other hand, is optimized for time-series data and can more accurately detect and separate overlapping signals, even under complex, time-varying conditions.

[0017] This system also offers significant improvements over existing CNN-LSTM-based solutions. While CNN-LSTM models are typically limited to processing structured spatial inputs, the DNN-LSTM model can also process unstructured or less dimensional data, making it more versatile and efficient for a wide range of real-world communication scenarios. Its ability to process both spatial and temporal aspects of signal data ensures that the system delivers superior performance in a variety of NOMA applications, from cellular networks to advanced wireless communication systems.

[0018] In summary, the DNN-LSTM-based interference suppression system offers a more efficient and scalable solution for signal detection in NOMA systems. By reducing computational complexity, enabling real-time signal detection, and improving accuracy in dynamic environments, this invention represents a significant advancement in wireless communication. It provides a robust alternative to conventional SIC methods and offers superior performance for modern 5G and 6G networks. Simultaneously, it is capable of handling increasing user numbers without any performance degradation.

Claims

[1] A system for detecting and separating overlapping signals in a NOMA (Non-Orthogonal Multiple Access) communication system, comprising a Deep Neural Network (DNN) and a Long Short-Term Memory (LSTM) model configured to process the received signal data and perform interference suppression in real time; wherein the DNN extracts high-level functions from the received signal to assist in signal detection and separation, and the LSTM captures temporal dependencies within the signal data, enabling improved performance in dynamic environments. [2] System according to claim 1, wherein the DNN-LSTM model processes multi-user signals in a single pass, thereby eliminating the need for multiple iterations as required by conventional successive interference suppression (SIC) systems, thereby reducing computational effort and improving signal detection speed. [3] System according to claim 1, wherein the DNN-LSTM model is supervised trained to optimize its weights and minimize signal detection errors, thereby enabling the model to be effectively generalized to real signal data. [4] System according to claim 1, wherein the DNN-LSTM model is specifically configured for processing time series data and capturing temporal dependencies over long distances, thereby improving the accuracy and reliability of signal detection in time-variable NOMA systems.