Blind demodulation system of digital signal based on category incremental learning
Through a digital signal blind demodulation system based on category incremental learning, combined with semi-supervised learning and unsupervised learning, the problems of high computational complexity and catastrophic forgetting in the existing technology are solved, and efficient for multiple types of digital communication signals is achieved. Identification and demodulation are suitable for massive data processing and reduce network computing complexity.
Patent Information
- Application Number
- CN202211546253.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-05
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-12-05
AI Technical Summary
The existing digital signal blind demodulation system has problems such as high computational complexity, limited identification performance and catastrophic forgetting in modulation method identification and demodulation, especially when processing massive data and diversified signals, it is difficult to effectively identify and resolve. Adjust.
Using a digital signal blind demodulation system based on category incremental learning, combining semi-supervised learning and unsupervised learning for new and old categories, through category incremental learning and characterization learning, using convolutional neural networks, recurrent neural networks or fully connected neural networks Data feature extraction and symbol prediction are carried out to identify and demodulate unknown modulation methods, and class imbalance problem is handled through balanced fine-tuning methods.
It realizes efficient identification and demodulation of multiple types of digital communication signals, reduces network computing complexity, is suitable for massive data processing, avoids catastrophic forgetting, maintains the recognition performance of old knowledge, and supports multi-platform operation and offline real-time Data transmission.
Smart Images

Figure CN116016070B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of signal processing, and in particular relates to a digital signal blind demodulation system based on category incremental learning. Background Art
[0002] The blind demodulation problem is a basic problem in the field of signal reconnaissance, which is divided into two parts: modulation mode recognition and demodulation. First, regarding modulation mode recognition, there are three types of algorithms: 1) likelihood ratio-based algorithms; 2) feature extraction-based algorithms; and 3) deep learning algorithms.
[0003] The likelihood ratio algorithm rule is to obtain the predicted value by determining the likelihood of the sampling sequence. The disadvantage of this algorithm is that it has high computational complexity. It requires knowing the probability distribution of random variables in the signal model, but this is usually not possible in practice.
[0004] The feature extraction algorithm obtains the prediction value based on the characteristic parameters of the signal, which include: statistical characteristics of instantaneous parameters of the signal, high-order statistical characteristics, cyclostationary characteristics, etc. The disadvantage of this algorithm is that the recognition performance is limited by the classification rules; when the signal itself becomes complex, the design of signal features is difficult.
[0005] Deep learning algorithms have become a hot research topic in recent years and have many practical applications. This is a technology that can improve the performance of computer systems from experience and data. Deep learning algorithms can be divided into: fully connected neural networks (also known as BP neural networks), convolutional neural networks, recurrent neural networks, generative adversarial networks, etc. according to different network structures; deep learning algorithms can be divided into: supervised learning, semi-supervised learning and unsupervised learning according to whether the data has labels; deep learning algorithms can be divided into: joint training, incremental training, transfer learning, category incremental learning, etc. according to different training methods.
[0006] Deep learning is a specific type of machine learning with great capabilities and flexibility. Deep learning also provides a very powerful tool to process massive data and make accurate predictions in the scientific field. Therefore, the research on the direction of arrival estimation algorithm based on digital signal processor has very important theoretical and practical value. In addition, like deep neural networks, deep neural networks are widely used in modulation estimation problems due to their strong learning ability, generalization ability, adaptability and portability. However, deep learning algorithms are generally still in the joint training stage, which requires a lot of time cost, and there are still problems such as catastrophic forgetting through transfer learning and incremental training. Summary of the invention
[0007] In view of this, the present invention provides a digital signal blind demodulation system based on category incremental learning, which can identify unknown modulation modes, save training time through incremental learning, avoid catastrophic forgetting, and demodulate the signal.
[0008] To achieve the above object, the technical solution of the present invention is achieved as follows:
[0009] A digital signal blind demodulation system based on category incremental learning, comprising a modulation mode recognition unit and a demodulation unit;
[0010] The modulation mode recognition unit uses semi-supervised learning / unsupervised learning to identify new and old categories, and then performs category incremental learning;
[0011] The demodulation unit uses the modulation mode as prior knowledge to distribute the data to the corresponding demodulation network to obtain the final predicted symbol data.
[0012] Furthermore, the modulation mode identification unit includes a data preprocessing unit and a network modeling unit.
[0013] Furthermore, the data preprocessing unit is used to divide the data into a training set and a validation set using K-fold cross validation, and perform z-score standardization, perform a deduplication operation through a cyclic comparison algorithm, and perform one-hot encoding on the labels.
[0014] Furthermore, the network modeling unit first uses semi-supervised learning / unsupervised learning to identify new and old classes, and then uses category incremental learning to perform representation learning and classifier learning in sequence. The representation learning corrects the previous feature representation and uses a new feature extractor to train the incoming and memory data to expand it. An auxiliary loss function is used on the new extractor to promote it to learn diverse and distinctive features. In each incremental step, the previously learned representation is frozen and the feature dimension is increased from the new feature extractor to expand it. After representation learning, the classifier is retrained with the currently available data, and the class imbalance is handled by adopting a balanced fine-tuning method.
[0015] Furthermore, the classifier uses Softmax and ReLU.
[0016] Furthermore, the data is baseband signal I / Q path data, and the label is a modulation mode.
[0017] Furthermore, the modulation network is trained using a convolutional neural network, a recurrent neural network or a fully connected neural network.
[0018] Furthermore, the demodulation unit extracts data features through a convolutional neural network, a recurrent neural network, or a fully connected neural network, saves the optimal network weights, and uses a classifier to output codeword prediction results.
[0019] Compared with the prior art, the digital signal blind demodulation system based on category incremental learning of the present invention has the following advantages:
[0020] The system of the present invention is applicable to all kinds of digital communication signals. The algorithm is data-driven and has strong generalization ability. The prediction performance is not affected by parameters such as modulation mode and sampling rate. It has good prediction performance for modulation modes based on phase modulation, frequency modulation and amplitude modulation. The network calculation complexity is low and it is suitable for blind demodulation of massive data.
[0021] The system of the present invention classifies known modulation modes and unknown modulation modes through semi-supervised / unsupervised learning, thereby improving the adaptability of the system. The system can also train data incrementally, saving time costs, and avoiding catastrophic forgetting while learning new knowledge, and can continuously maintain the recognition performance of old knowledge;
[0022] The system of the present invention supports multi-platform operations such as Windows, Linux, MacOS, etc., supports offline real-time data transmission, and the transmission rate can reach millisecond-level response. Its processing flow and network architecture have generalization capabilities and are also universal for other application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:
[0024] Figure 1 is a system block diagram of the present invention;
[0025] Figure 2 A schematic diagram of a modulation mode identification unit of the present invention;
[0026] Figure 3 It is a schematic diagram of the network model structure of the modulation mode identification unit of the present invention. DETAILED DESCRIPTION
[0027] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0028] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.
[0029] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.
[0030] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.
[0031] like Figure 1 As shown, the digital signal blind demodulation system based on category incremental learning proposed in the present invention is divided into a modulation mode recognition unit and a demodulation unit.
[0032] Specifically, the modulation mode recognition unit: uses semi-supervised learning to identify the novelty of the signal. Here, the K-nearest neighbor algorithm, K-means clustering algorithm, etc. can be used to distinguish the old class from the new class. Then, incremental category learning is performed to ensure that the recognition performance of the old class is not reduced while the new class is trained. For details of the modulation mode recognition unit flow chart, see Figure 2 .
[0033] Specifically, the modulation mode identification unit includes a data preprocessing unit and a network modeling unit:
[0034] Data preprocessing unit: Use K-fold cross validation to divide the training set and validation set, perform z-score standardization, perform deduplication operation through a circular comparison algorithm, and perform one-hot encoding on the labels. The data is the baseband signal I / Q path data, and the label is the modulation mode.
[0035] Network modeling unit: that is, the dataset training phase. First, semi-supervised learning / unsupervised learning is used to identify old and new classes. Then, category incremental learning is used to perform representation learning and classifier learning in sequence. Representation learning expands the previous feature representation by correcting it and using a new feature extractor to train the incoming and memory data. An auxiliary loss function is used on the new extractor to encourage it to learn diverse and distinctive features. The algorithm can dynamically expand the representation according to the complexity of the new class through pruning methods. In each incremental step, the previously learned representation is frozen and expanded by adding feature dimensions from the new feature extractor, which enables the algorithm to integrate new knowledge and retain learned knowledge. After representation learning, the algorithm retrains the classifier with the currently available data and handles the class imbalance problem by adopting a balanced fine-tuning method. The classifier can use classifiers such as Softmax and ReLU. See the network model structure of the modulation mode recognition unit. Figure 3 .
[0036] Specifically, the demodulation unit: uses the modulation mode as prior knowledge, distributes the data to the corresponding demodulation network, and obtains the final predicted symbol data. The demodulation network can be trained using a convolutional neural network, a recurrent neural network, or a fully connected neural network.
[0037] Specifically, the demodulation unit includes a data preparation unit and a decoding unit:
[0038] Data preparation unit: Use K-fold cross validation to divide the training set and validation set, perform z-score standardization, perform deduplication through a circular comparison algorithm, and perform one-hot encoding on the labels. The data is the baseband signal I / Q path data, and the label is the code element;
[0039] Decoding unit: that is, the dataset training phase, through convolutional neural network / recurrent neural network / fully connected neural network, data features are extracted, the optimal network weights are saved, and the codeword prediction results are output using a classifier, where the classifier can use Softmax, ReLU and other classifiers.
[0040] In the specific implementation process, after the receiver receives the modulated signal and down-converts it to the baseband, it first performs preprocessing operations on the baseband signal, including: z-score standardization, wavelet transform denoising, median filtering to remove clutter, etc., and fills the default part with 0. For the baseband signal after data cleaning, it is divided into 4 parts according to the time domain amplitude, time domain phase, I-channel FFT transform, and Q-channel FFT transform as a data set. Compared with traditional I / Q two-channel time domain data, the present invention uses a variety of time-frequency transforms to mine signal features from a deeper dimension, which can effectively reduce the network training convergence time and improve classification performance.
[0041] The data set is divided into training set and validation set using K-fold cross validation, and duplicate removal is performed using a cyclic comparison algorithm, and the labels are one-hot encoded.
[0042] The neural network classifier uses a network structure that combines CNN+LSTM+Dropout. The loss function uses Keras's multi-classification cross entropy and Adam optimizer. The network construction process is shown in Figure 3 .
[0043] For open set data, that is, the test data contains real category labels that are not included in the training data, it can be classified through open set recognition. The data processing flow is shown in Figure 1 .
[0044] For the signal with modulation mode, we can Figure 2 The final predicted symbol data is obtained through the processing operation. Compared with the traditional envelope-based demodulation algorithm, the neural network-based decoding algorithm is more robust to noise. It is less affected by the interference of oversampling rate and channel environment.
[0045] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A digital signal blind demodulation system based on category incremental learning, characterized by: It includes a modulation mode identification unit and a demodulation unit; The modulation mode identification unit includes a data preprocessing unit and a network modeling unit; The modulation mode recognition unit recognizes the novelty of the signal through semi-supervised learning, distinguishes old classes from new classes, and then performs category incremental learning; The data preprocessing unit is used to divide the data in the data set into a training set and a validation set using K-fold cross validation, perform z-score standardization, perform a deduplication operation using a circular comparison algorithm, and perform one-hot encoding on the labels; The data is the baseband signal I / Q data, and the label is the modulation mode; The data preprocessing unit also includes that after the receiver receives the modulated signal and down-converts it to the baseband, the baseband signal is first preprocessed. The baseband signal after data cleaning is divided into four parts according to the time domain amplitude, time domain phase, I-channel FFT transformation, and Q-channel FFT transformation as a data set; The demodulation unit uses the modulation mode as prior knowledge to distribute the data to the corresponding demodulation network to obtain the final predicted symbol data; the demodulation network is trained using a convolutional neural network, a recurrent neural network or a fully connected neural network; the network modeling unit first uses semi-supervised learning to identify new and old classes, and then uses category incremental learning to perform representation learning and classifier learning in sequence. The representation learning corrects the previous feature representation and constructs a new feature extractor. The new feature extractor is expanded by training the incoming and memory data with the new feature extractor, and an auxiliary loss function is used on the new extractor to promote the new feature extractor to learn diverse and distinctive features; in each incremental step, the previously learned representation is frozen, and the new feature extractor is expanded by adding feature dimensions from the new feature extractor to integrate new knowledge and retain the learned knowledge; after representation learning, the classifier is retrained with the currently available data, and the imbalance of classes is handled by adopting a balanced fine-tuning method; The demodulation unit includes a decoding unit: in the data set training stage, data features are extracted through convolutional neural network / recurrent neural network / fully connected neural network, network weights are saved, and code element prediction results are output using a classifier; the neural network selected is a network structure combining CNN+LSTM+Dropout, wherein the loss function uses Keras's multi-classification cross entropy and Adam optimizer.
2. The digital signal blind demodulation system based on category incremental learning according to claim 1, characterized in that: The classifier uses Softmax and ReLU.
