Wireless signal modulation mode and signal type recognition method based on multi-task learning

Abstract

The present invention provides a method for identifying wireless signal modulation modes and signal types based on multi-task learning. The method constructs a satellite communication protocol and a modulation data set, wherein the satellite communication protocol and modulation data set include a plurality of composite sample data and label information for marking the composite sample data. Each composite sample data includes time-domain complex data and power spectrum density frequency-domain data, and the label information includes signal type and modulation mode. Based on the satellite communication protocol and modulation data set, a multi-task learning neural network model is trained. The baseband signal received by the radio frequency system is preprocessed to obtain a time-domain complex signal and a frequency-domain characteristic signal to be measured. These signals are transmitted to the pre-trained multi-task learning neural network model to obtain an identification result of the model output signal type and modulation mode. In this way, the present invention can reduce resource overhead while ensuring recognition accuracy, and is suitable for resource-constrained satellite scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless signal processing, and in particular to a wireless signal modulation mode and signal type recognition method based on multi-task learning. Background Art

[0002] With the release of the 6G technology vision, the space-ground integrated network paradigm will be a key component of the future space-ground integrated network. This paradigm embodies the convergence of space-based and ground-based communication infrastructure. Within this space-ground integrated network, satellite communications have become a key component, transcending the traditional forwarding communication functions carried by satellites to include enhanced communication capabilities such as baseband data processing. This evolution aligns with the growing demand for ubiquitous connectivity. In future networks, satellites will not only be passive relays providing transparent forwarding, but will become active nodes capable of hosting satellite base stations and even in-orbit data centers. This transformation requires a significant increase in satellite onboard processing capabilities to perform complex tasks such as data fusion and accommodate diverse and heterogeneous communication scenarios involving ground gateways and mobile terminals. The ability to handle heterogeneous communication protocols will be a critical requirement for satellite capabilities in future space-ground integrated networks, ensuring seamless interoperability and efficient data processing between different platforms and users.

[0003] The main characteristic of traditional satellites is their single function, with each satellite dedicated to a specific mission such as remote sensing or communications. This specialization means that most satellites can only process a single communication protocol, which limits their versatility in meeting different communication needs. In the ever-evolving satellite network, the ability to dynamically meet the needs of various services is becoming increasingly important, and this limitation poses a huge challenge. Therefore, the concept of satellite communications capable of processing heterogeneous communication protocols is gaining increasing attention; this capability will enable satellites to adaptively handle different types of data transmission. The key challenge in achieving this capability lies in baseband data processing capabilities, and different communication protocols have different baseband data processing modes. In order to process data from different communication protocols, satellites must be able to accurately identify their signal protocols and distinguish the signal modulation method.

[0004] Currently, technologies for signal type and modulation identification fall into three main categories. The first involves determining the signal type (communication protocol) and automatically identifying the modulation method. The second involves simply identifying the signal type without further analysis or identification of the modulation method. Conventional communication systems often rely on fixed protocols for communication, so simply identifying the protocol and adaptive modulation is sufficient to effectively support baseband processing. Consequently, these two approaches account for the majority of related work. However, in future integrated space-ground networks, satellites will need to process signals from ground stations, mobile ground equipment, and other satellites using heterogeneous communication protocols. Simply identifying the signal type or modulation method corresponding to the protocol is insufficient to effectively process diverse baseband signals using the same RF baseband. A third category of technologies currently exists that simultaneously identifies both signal type and modulation method, but these technologies primarily distinguish between radar and terrestrial wireless signals. The characteristics of these two types of signals are often distinct and easy to distinguish. Furthermore, the design scenarios for these approaches often involve conventional channels affected by additive white Gaussian noise (AWGN). However, AWGN channels are difficult to directly describe for satellite communication. Furthermore, these approaches often assume that the constellation points corresponding to the modulation type are ideal, meaning that the baseband signal is identified without filtering through a shaping filter. Under these conditions, the baseband constellation mapping signal is ideal and has infinite bandwidth, which is inconsistent with the operation of actual communication systems. In satellite communications, channel characteristics are typically much more demanding, making these methods difficult to directly apply to integrated space-ground networks.

[0005] Therefore, there is an urgent need for a method that can identify the heterogeneous signal types and modulations of satellite communications in a space-ground integrated network. Summary of the Invention

[0006] The purpose of the present invention is to provide a wireless signal modulation mode and signal type identification method based on multi-task learning, which can identify the signal type and modulation mode of signals of different communication protocols.

[0007] To achieve the above objectives, the present invention provides a wireless signal modulation mode and signal type identification method based on multi-task learning, comprising:

[0008] Constructing a satellite communication protocol and modulation data set; wherein the satellite communication protocol and modulation data set includes a plurality of composite sample data and label information for marking the composite sample data, each composite sample data includes time domain complex data and power spectrum density frequency domain data, and the label information includes signal type and modulation mode;

[0009] Based on the satellite communication protocol and modulation data set, a multi-task learning neural network model is trained; the multi-task learning neural network model includes multiple input layers, shared layers, and multiple output layers; the multiple input layers are used to receive multiple input data and extract corresponding feature parameters, the shared layers are used to fuse the feature parameters and extract multi-task common features based on multi-task output constraints, and the multiple output layers are used to output multi-task recognition results based on the multi-task common features;

[0010] The baseband signal received by the radio frequency system is preprocessed to obtain the time domain complex signal and the frequency domain characteristic signal to be measured, and are transmitted as input data to the pre-trained multi-task learning neural network model to obtain the recognition results of the model output signal type and modulation method.

[0011] Optionally, the power spectrum density frequency domain data in the composite sample data is calculated based on the corresponding time domain complex data using a Welch overlapping piecewise average estimator;

[0012] The preprocessing of the baseband signal received by the radio frequency system to obtain the time domain complex signal to be measured and the frequency domain characteristic signal to be measured includes:

[0013] Extracting the time-domain complex signal to be measured based on the baseband signal received by the radio frequency system;

[0014] The power spectrum density of the time-domain complex signal to be measured is calculated by using a Welch overlapping segmented average estimator to obtain a corresponding frequency-domain characteristic signal to be measured.

[0015] Optionally, the multi-input layer includes a first group of neuron structures and a second group of neuron structures, the first group of neuron structures is used to extract features of the time domain complex signal to be measured, and the second group of neuron structures is used to extract features of the frequency domain characteristic signal to be measured;

[0016] The first group of neuron structures includes a first convolution module, a first dropout layer, a first reconstruction layer, a first maximum pooling layer and a first flattening layer;

[0017] The second group of neuron structures includes a second convolution module, a second Dropout layer, a second maximum pooling layer and a second flattening layer.

[0018] Optionally, the first convolution module and the second convolution module are both composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu;

[0019] The characteristic size of the time-domain complex signal to be measured after being reshaped by the first reconstruction layer and the first maximum pooling layer is consistent with the characteristic size of the frequency-domain characteristic signal to be measured after being reshaped by the second maximum pooling layer.

[0020] Optionally, the shared layer includes:

[0021] A fusion layer, configured to fuse the feature parameters output by the multiple input layers;

[0022] The second reconstruction layer is used to reconstruct the dimension of the fused feature data;

[0023] A third convolution module is used to extract the fusion features of the time domain complex signal and the frequency domain feature signal according to the reconstructed feature data; the third convolution module is composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu;

[0024] The third Dropout layer is used to randomly disable some features of the fusion feature;

[0025] The third maximum pooling layer is used to reduce the dimension of the fusion feature to output the multi-task common feature.

[0026] Optionally, the multi-output layer includes a first task branch structure for classifying and identifying signal types and a second task branch structure for classifying and identifying modulation modes;

[0027] The first task branch structure includes a fourth convolution module, a fourth Dropout layer, a first dense module, a fifth Dropout layer, a third flattening layer, and a first dense layer with a SoftMax activation function;

[0028] The second task branch structure includes a fifth convolution module, a sixth Dropout layer, a second dense module, a seventh Dropout layer, a fourth flattening layer, and a second dense layer with a SoftMax activation function;

[0029] Among them, the fourth convolution module and the fifth convolution module are both composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu.

[0030] Optionally, the first task branch structure obtains the multi-task common features, and extracts a first target feature corresponding to the signal type recognition task from the multi-task common features through the fourth convolution module and the first dense module, so that the first dense layer outputs a first recognition result of the signal type based on the first target feature;

[0031] The second task branch structure obtains the common features of the multiple tasks, and extracts the second target features of the corresponding modulation mode identification task from the common features of the multiple tasks through the fifth convolution module and the second dense module, so that the second dense layer outputs the second recognition result of the modulation mode based on the second target feature.

[0032] Optionally, the training of a multi-task learning neural network model based on the satellite communication protocol and the modulation data set includes:

[0033] Dividing the satellite communication protocol and modulation data set into a training set, a validation set, and a test set according to a preset ratio;

[0034] Constructing a multi-task learning neural network model, and iteratively training the multi-task learning neural network model using the training set;

[0035] Performing model adjustment and model parameter determination on the multi-task learning neural network model according to the verification set;

[0036] The test set is input into the multi-task learning neural network model to obtain corresponding output data to evaluate the model performance.

[0037] Optionally, the constructing of a satellite communication protocol and modulation data set includes:

[0038] Select several communication protocols;

[0039] Based on the modulation and coding mechanisms of the selected communication protocols, generate a plurality of time-domain complex signals, and mark the time-domain complex signals with label information, wherein the label information includes a signal type and a modulation method corresponding to the generation of the time-domain complex signals;

[0040] Outputting the time domain complex signal through at least one preset channel and signal-to-noise ratio;

[0041] Performing power spectral density calculation on the time-domain complex signal using a Welch overlapping segmented average estimator to obtain a corresponding frequency-domain characteristic signal;

[0042] The time domain complex signal, the frequency domain characteristic signal and the corresponding label information are constructed to generate the composite sample data, and stored in a preset satellite communication protocol and modulation data set.

[0043] Optionally, the at least one channel includes an AWGN channel and an LMS channel;

[0044] The communication protocols include S2 protocol, S2X protocol, RCS2 protocol, CCSDS TM protocol and CCSDS TC protocol;

[0045] The modulation and coding mechanisms of the communication protocol include BPSK, PI / 2BPSK, QPSK, 8PSK, 16APSK and 16QAM.

[0046] The multi-task learning-based wireless signal modulation and signal type identification method described in this invention is targeted at satellite baseband processing of radio signals with different communication protocols. It utilizes a multi-task learning model to simultaneously identify signal type and modulation, reusing a single network model to identify multiple signal types and modulations, thereby improving signal processing efficiency. Furthermore, the dataset used to train this multi-task learning model includes multiple signal types and modulations, taking into account the impact of satellite-to-ground channels on wireless signals. Compared to traditional methods, the method provided by this invention eliminates the need for mandatory pre-agreed or polling-based determination of the received signal type and modulation. After identifying the signal type and modulation, it can more efficiently select a demodulation algorithm, thereby improving signal processing speed and hardware resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flowchart of the steps of the wireless signal modulation mode and signal type identification method based on multi-task learning provided in one embodiment of the present invention;

[0048] Figure 2 A flowchart of the steps for training the multi-task learning neural network model for the wireless signal modulation mode and signal type recognition method based on multi-task learning provided in one embodiment of the present invention;

[0049] Figure 3 A flowchart of the steps for constructing the satellite communication protocol and modulation data set by the wireless signal modulation mode and signal type identification method based on multi-task learning provided in one embodiment of the present invention;

[0050] Figure 4 A neural network structure diagram of a multi-task learning neural network model for the wireless signal modulation mode and signal type recognition method based on multi-task learning provided in one embodiment of the present invention;

[0051] Figure 5a-5b A schematic diagram showing how the recognition accuracy of the multi-task learning neural network model of the wireless signal modulation mode and signal type recognition method based on multi-task learning provided in one embodiment of the present invention varies with the signal-to-noise ratio;

[0052] Figure 6a to Figure 6d A schematic diagram of a confusion matrix for signal type and modulation mode identification of the multi-task learning neural network model of the wireless signal modulation mode and signal type identification method based on multi-task learning provided in one embodiment of the present invention when the signal-to-noise ratio is 4dB. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0054] It should be noted that references to "one embodiment," "an embodiment," "an example embodiment," etc., in this specification indicate that the described embodiment may include specific features, structures, or characteristics, but not every embodiment must include these specific features, structures, or characteristics. Furthermore, such references do not necessarily refer to the same embodiment. Furthermore, when specific features, structures, or characteristics are described in conjunction with an embodiment, whether or not explicitly described, it is understood that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0055] In addition, certain words are used in the specification and subsequent claims to refer to specific components or parts. It should be understood by those with ordinary knowledge in the relevant field that manufacturers may use different nouns or terms to refer to the same component or part. This specification and subsequent claims do not use differences in names as a way to distinguish components or parts, but rather use differences in the functions of components or parts as the criteria for distinction. The words "including" and "comprising" mentioned throughout the specification and subsequent claims are open-ended terms and should be interpreted as "including but not limited to". In addition, the word "connect" here includes any direct and indirect electrical connection means. Indirect electrical connection means include connection through other devices.

[0056] Figure 1 A wireless signal modulation mode and signal type identification method based on multi-task learning provided by an embodiment of the present invention is shown, including the following steps:

[0057] S101: Construct a satellite communication protocol and modulation data set; wherein the satellite communication protocol and modulation data set includes a number of composite sample data and label information for marking the composite sample data, each composite sample data includes time domain complex data and power spectrum density frequency domain data, and the label information includes signal type and modulation method; the constructed satellite communication protocol and modulation data set is used to train, verify and test the neural network model in the subsequent process; the composite sample data in the data set can be specifically generated through multiple protocol types and their digital modulation methods.

[0058] Because most existing datasets, such as DeepSig's RADIOML 2018.01A, primarily focus on modulation as the sole labeling criterion, they clearly fail to address the complexity of more comprehensive multi-task learning scenarios, where both signal type and modulation mode are relevant in the context of integrated space-ground networks. An example of an attempt to bridge this gap is the RadarCom dataset, which annotates data using both signal type and modulation mode. However, RadarCom's scope is limited to radar and wireless signals, making it difficult to meet the requirements of satellite communications in integrated space-ground networks.

[0059] In view of the above limitations, an optional implementation of the satellite communication protocol and modulation data set constructed in this embodiment will be introduced below. Figure 3 As shown, step S101 further includes:

[0060] S301: Select several communication protocols; in this step, the type of communication protocol to be targeted can be determined according to specific needs. Preferably, in this example, the S2, S2X, and RCS2 protocols specified by the Digital Video Broadcasting (DVB) organization and the satellite tracking, telemetry and command (TT&C) system communication protocol specified by the Consultative Committee for Space Data Systems (CCSDS) can be selected, including TM (Telemetry) and TC (Telecommand) signals.

[0061] S302: Based on the modulation and coding mechanisms of the selected communication protocols, multiple time-domain complex signals are generated, and labeling information is added to the time-domain complex signals. The labeling information includes the signal type and modulation scheme corresponding to the generation of the time-domain complex signals. In a specific implementation, the modulation and coding mechanism of the communication protocol determined above is selected to generate a corresponding waveform. For example, the corresponding modulation and coding mechanisms include BPSK, PI / 2BPSK, QPSK, 8PSK, 16APSK, and 16QAM. The generated waveform is then labeled with the communication protocol and modulation scheme. Thus, based on the above example, the generated waveform includes five signal types and six modulation schemes. Of course, this embodiment does not limit the type and number of signal types and modulation schemes, and can be adjusted according to the needs of the application environment.

[0062] S303: Outputting the time-domain complex signal through at least one preset channel and signal-to-noise ratio; after generating a plurality of waveform data through a plurality of signal types and modulation modes in the aforementioned steps, further in step S303, selecting different channels and signal-to-noise ratios to output the signal; the optional channels in this embodiment include AWGN channels and LMS (Land Mobile Satellite) channels, etc.; in specific implementation, the signal is output in the AWGN channel according to the corresponding selected signal-to-noise ratio, while the LMS channel generates the signal affected by the corresponding channel ionosphere and rain loss, and then sends the signal to the AWGN channel of the corresponding signal-to-noise ratio and outputs the signal.

[0063] The time domain complex signal of this embodiment is specifically an in-phase and quadrature complex (IQ) signal that can represent time domain characteristics. The IQ complex signal will be used for explanation below.

[0064] S304: The power spectral density of the time-domain complex signal is calculated using the Welch overlapping piecewise average estimator to obtain the corresponding frequency-domain characteristic signal. After the IQ complex signal is output through the above-selected channel, the power spectral density is further calculated based on the time-domain IQ complex signal using the Welch overlapping piecewise average estimator, and the calculation result is the corresponding frequency-domain characteristic signal. The frequency-domain characteristic signal of this embodiment is specifically the power spectral density (PSD) frequency-domain data used to characterize the frequency-domain characteristics, and PSD data will be used for explanation below.

[0065] S305: Composite sample data is constructed from the time-domain complex signal, frequency-domain characteristic signal, and corresponding label information, and stored in a preset satellite communication protocol and modulation data set. That is, each pair of generated time-domain complex signals and frequency-domain characteristic signals is combined to form a data unit in the data set, which is the composite sample data. The corresponding label information can be stored together with the composite sample data in the data set, or the label information can be composited with the composite sample data for label extraction during subsequent data processing. The composite sample data in this embodiment's data set is generated based on the multiple communication protocols and modulation methods described above.

[0066] The satellite communication protocol and modulation dataset constructed by the above-mentioned implementation method is different from many existing public datasets in that it includes a variety of communication protocols related to satellite communications, such as the digital video broadcasting protocol and the satellite tracking and control protocol of the Consultative Committee for Space Data Systems, including five protocol types and six digital modulation methods. More importantly, the dataset also includes the impact of AWGN channels and satellite-to-ground channels on radio signals, which is a factor often overlooked by existing datasets. The dataset constructed by this embodiment is closer to the actual situation of satellite communications in the context of a broader integrated space-ground network; the LMS channel scenario takes into account not only the signal-to-noise ratio, but also factors unique to satellite communications such as ionospheric scintillation and rain attenuation. Multiple channel scenarios enable the dataset to simulate more realistic land mobile satellite channels, thereby providing more suitable and realistic resources for the development and testing of automatic identification algorithms for satellite communication signals.

[0067] The composite sample data in the constructed satellite communication protocol and modulation dataset contains IQ complex data used to characterize time domain characteristics and power spectral density frequency domain data PSD used to characterize frequency domain characteristics. Among them, the IQ complex data reflects the time domain characteristics of the baseband signal received by the RF front-end, while the PSD data describes the frequency domain characteristics of the corresponding signal more clearly, which is crucial for the iterative learning process of the subsequent multi-task learning neural network (MTL) model to be trained.

[0068] The method for generating PSD data is specifically to use the average windowed power spectrum estimation method through IQ complex data to generate smooth frequency domain feature data.

[0069] Identifying signal types in the satellite communication protocol and modulation dataset is one of the tasks of the proposed MTL model. Therefore, it is necessary to rationally construct dataset labels that include multiple signal types and modulation schemes. Signal types primarily consider the possible protocol types in a space-ground integrated network. This embodiment specifically includes three DVB protocols (S2, S2X, and RCS2) and two CCSDS protocols (telemetry and remote control). This protocol diversity ensures that the dataset effectively represents the convergence capabilities required to handle different communication protocols.

[0070] The satellite communication protocol and modulation data set contain modulation schemes that are generated according to the modulation and coding scheme (MCS) specified in the protocol. Therefore, the six modulation methods included in the data set are BPSK, PI / 2BPSK, QPSK, 8PSK, 16QAM, and 16APSK. Furthermore, in actual communication systems, all transmit and receive signals must be filtered by a shaping filter before further processing. That is, the constellation diagram of the actual signal exhibits a certain degree of dispersion rather than an ideal single-point state when not affected by the channel. The IQ complex data and PSD frequency domain data in the data set of this embodiment are also filtered by a shaping filter according to the roll-off coefficient specified in the protocol. Specifically, after determining the modulation scheme based on the modulation and coding mechanism of each satellite protocol and generating an IQ complex signal that meets the selected protocol type and modulation scheme, a shaping filter is generated according to the protocol, and an IQ complex signal filtered by the shaping filter is generated. The IQ complex signal filtered by the shaping filter is further used to generate an IQ complex signal affected by the channel from a plurality of selected wireless channel scenario types, so that the IQ complex signal in the sample data is closer to the actual application scenario.

[0071] After the satellite communication protocol and the modulation data set are constructed, the process proceeds to step S102.

[0072] S102: Based on the satellite communication protocol and modulation data set, a multi-task learning neural network model is trained; the multi-task learning neural network model includes multiple input layers, shared layers, and multiple output layers; the multiple input layers are used to receive multiple input data and extract corresponding feature parameters, the shared layers are used to fuse the feature parameters and extract multi-task common features based on the constraints of the multi-task outputs, and the multiple output layers are used to output multi-task recognition results based on the multi-task common features. Combined with the above, it can be seen that the composite sample data in the data set is composed of IQ complex data and PSD data. Therefore, in the training of the multi-task learning neural network model in this step S102, the IQ complex data and PSD data are used as the model's multiple inputs; that is, the multi-input part of the multi-task learning neural network model constructed in this embodiment is composed of two inputs, the first input being IQ complex data and the second input being PSD data. The multiple inputs highlight the time domain constellation features of the IQ complex signal and the frequency domain features of the spectrum data, thereby providing rich raw information for feature extraction of the neural network.

[0073] Multi-task learning can uniquely solve multiple classification or regression problems within a single framework. The basic logic of MTL is that there is a potential dependency between seemingly different tasks, and this dependency stems from a shared data generation process; this premise holds that although the tasks appear unrelated on the surface, their underlying generation processes may be related to each other. Therefore, the unified neural network model in the MTL model can effectively utilize these commonalities to identify multiple subtasks. This embodiment utilizes the MTL model to perform multiple subtasks within the same framework with IQ complex data and PSD data as two inputs; the multiple subtasks in this embodiment specifically refer to signal type recognition tasks and modulation mode recognition tasks, that is, the IQ complex data and PSD data of the signal are analyzed by the MTL model to output corresponding signal type recognition results and modulation mode recognition results.

[0074] Multiple input layers can intuitively extract time and frequency domain features from the input data. The differentiated waveform characteristics of signals corresponding to different signal protocols (i.e., signal types) are often more obvious in the frequency domain than in the time domain. Therefore, the purpose of using IQ complex signals and power spectrum data as multiple inputs is to provide the MTL model with richer and more diverse feature data, thereby enhancing its ability to effectively process and extract features.

[0075] The shared layer before each task extracts features is a key feature of multi-task learning. It corresponds to a network model that first extracts common features from multiple tasks and then learns each task's features based on these common features. Compared to learning multiple tasks separately to form multiple single-task learning models, multi-task learning is characterized by high method efficiency and low computational overhead. Therefore, multi-task learning models are more suitable for satellite scenarios with limited computing and storage resources. The main function of the shared layer portion of this embodiment is to extract common features of multiple inputs and multiple task outputs from the output of the multi-input layer. The shared layer utilizes the common features of different tasks to improve the learning efficiency of the overall model; its shared feature extractor's ability to fuse and interpret multiple output features plays a key role in discovering the inherent representational relationships between different tasks, thereby enriching the neural network model's understanding of the data.

[0076] The multi-input layer of this embodiment includes a first group of neuron structures and a second group of neuron structures, wherein the first group of neuron structures is used to extract features of the time domain complex signal to be measured, and the second group of neuron structures is used to extract features of the frequency domain characteristic signal to be measured; Figure 4 As shown, the first group of neuron structures includes a first convolution module, a first Dropout layer, a first reconstruction layer, a first maximum pooling layer and a first flattening layer; the second group of neuron structures includes a second convolution module, a second Dropout layer, a second maximum pooling layer and a second flattening layer.

[0077] In specific implementations, the baseband time-domain complex signal and power spectral density data received by the RF front end are used as two types of multi-input signals in the multi-task learning neural network model and enter the multi-input layer of the neural network. This multi-input layer specifically includes a first convolution module, a first dropout layer, a first reconstruction layer, a first maximum pooling layer, and a first flattening layer for feature extraction of the time-domain signal; and a second convolution module, a second dropout layer, a second maximum pooling layer, and a second flattening layer for feature extraction of the power spectral density data.

[0078] Among them, the first convolution module and the second convolution module are both composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu; in order to fuse the IQ complex two-channel data with the PSD data, it is necessary to pay attention to adjusting the size of the Reshape and maximum pooling layers to ensure the consistency of the size of the multi-input layer when it is imported into the shared layer, so as to be used for the multi-input feature fusion of the shared part of the neural network; to this end, the feature size of the time domain complex signal to be measured in this embodiment is shaped by the first reconstruction layer and the first maximum pooling layer and the feature size of the frequency domain feature signal to be measured is consistent with the feature size of the second maximum pooling layer; that is, the time domain complex signal and the power spectral density data are sized after the corresponding reconstruction layer and the maximum pooling layer to obtain features with consistent size as the output of the multi-input layer, so as to be used for the multi-input feature fusion of the shared layer of the neural network.

[0079] See also Figure 3 , the preferred shared layers in this embodiment include a fusion layer, a second reconstruction layer, a third convolution module, a third Dropout layer, and a third maximum pooling layer, wherein:

[0080] The fusion layer is used to fuse the feature parameters output by multiple input layers; the second reconstruction layer is used to reconstruct the dimension of the fused feature data; the third convolution module is used to extract the fusion features of the time domain complex signal and the frequency domain feature signal based on the reconstructed feature data; the third convolution module is composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu; the third Dropout layer is used to randomly invalidate some features of the fusion features; the third maximum pooling layer is used to reduce the dimension of the fusion features to output multi-task common features.

[0081] Specifically, the shared layer first uses the fusion layer to merge the two types of input (IQ complex signals and PSD data) into one channel of data; in order to meet the convolution kernel size of the convolution layer, the fused feature data is reconstructed through the second reconstruction layer to reconstruct the feature data dimension, and then the reconstructed feature data is transmitted to the third convolution module. The composition of the third convolution module of the shared layer is the same as the corresponding convolution module in the multi-input layer, consisting of a convolution layer, a batch normalization layer and a ReLu activation function; the third convolution module extracts the fusion features from the IQ time domain signal and the frequency domain feature signal, and extracts some features in the subsequent third Dropout layer and the third maximum pooling layer to avoid overfitting during training; the third Dropout layer randomly invalidates some features. Finally, the feature data processed by the third maximum pooling layer further reduces the dimension and becomes the overall output of the shared layer, that is, the multi-task common feature.

[0082] The multi-task common features output by the shared layer are transmitted to the multi-output layer, which extracts output features and performs recognition and classification for two types of tasks: signal type and modulation mode.

[0083] See also Figure 3 , the preferred multi-output layer of this embodiment includes a first task branch structure for classifying and identifying signal types and a second task branch structure for classifying and identifying modulation modes; the first task branch structure includes a fourth convolution module, a fourth Dropout layer, a first dense module, a fifth Dropout layer, a third flattening layer, and a first dense layer with an activation function of SoftMax; the second task branch structure includes a fifth convolution module, a sixth Dropout layer, a second dense module, a seventh Dropout layer, a fourth flattening layer, and a second dense layer with an activation function of SoftMax; wherein the fourth convolution module and the fifth convolution module are both composed of a convolution layer, a batch normalization layer, and a nonlinear activation function ReLu; that is, the network structures for extracting the two types of task features in the multi-output layer of this embodiment are basically the same; the multi-task common features output by the shared layer are respectively transmitted to the first task branch structure and the second task branch structure, and the two task branch structures respectively extract features belonging to their respective tasks from the multi-task common features based on their corresponding convolution modules and dense modules, so as to perform classification and identification according to their respective divided tasks.

[0084] During specific implementation, the first task branch structure obtains the multi-task common features, and extracts the first target features of the corresponding signal type recognition task from the multi-task common features through the fourth convolution module and the first dense module, so that the first dense layer outputs the first recognition result of the signal type based on the first target feature; the second task branch structure obtains the multi-task common features, and extracts the second target features of the corresponding modulation mode recognition task from the multi-task common features through the fifth convolution module and the second dense module, so that the second dense layer outputs the second recognition result of the modulation mode based on the second target feature. Among them, the dense layer of the dense module fully connects the input and output of the layer and extracts higher-dimensional features. For different recognition tasks, the number of recognition classifications of the dense layer with the activation function of SoftMax at the end of the neural network is determined by the number of label types of the corresponding task; the output results of the multi-output layer of this embodiment include two categories, namely, the recognition results of the signal type and the modulation mode.

[0085] As a multi-output layer that performs specific task recognition and classification in the MTL model, it can map the shared parameters from the shared layer to multiple specific subtasks (specifically two subtasks in this embodiment, and the two corresponding classifiers can be represented as O1 and O2); the multi-output layer is responsible for further improving the feature extraction process of each subtask and making its output quantity consistent with the number of respective label types. The feature extraction mechanism of the MTL model effectively combines and processes time domain and frequency domain data, and generates a representation vector through multiple input layers and shared layers. The representation vector is used by two task classifiers to predict the labels of known signal types and modulation types. The output type of each subtask determines the output size of the Softmax function in the fully connected output layer; according to the above-mentioned data set design, the signal types of this embodiment specifically include 5 types and the modulation methods include 6 types.

[0086] The MTL model repeatedly adjusts the neural network parameters based on the comparison between the output of multiple output layers and the known labels; this approach of customizing the model for specific subtasks highlights the flexibility and adaptability of the MTL framework, ensuring that each task can adopt a task-specific optimization strategy while benefiting from the common learning and feature extraction process of the shared layers.

[0087] See also Figure 2 In an optional embodiment, the training of a multi-task learning neural network model based on the satellite communication protocol and the modulation data set includes:

[0088] S201: Divide the satellite communication protocol and modulation data set into a training set, a validation set, and a test set according to a preset ratio.

[0089] S202: Construct a multi-task learning neural network model, and iteratively train the multi-task learning neural network model using the training set. Specifically, the two classifiers of the multi-output layer are trained using a multi-task classification loss function. This loss function is crucial to improving the model's ability to accurately classify different wireless signal categories and modulation types. The loss function penalizes the misclassifications generated during the network training process, thereby guiding the training process to achieve higher accuracy. During the training process, for each sample in the MTL dataset, the corresponding classification probability vector is output through the SoftMax function. The probability vectors correspond to the classification tasks of the signal type and modulation mode, respectively, and then the classification results of each task are determined.

[0090] S203: Adjusting the multi-task learning neural network model and determining model parameters based on the validation set. During training, the model can be periodically evaluated on the validation set. Based on the evaluation results on the validation set, the model hyperparameters are adjusted to find the optimal model configuration.

[0091] S204: Input the test set into the multi-task learning neural network model to obtain corresponding output data to evaluate model performance. Specifically, the model is used to classify and identify the data in the test set, outputting corresponding results. The output results are then compared with the true labels of the test set, and the model's evaluation metrics are evaluated based on the comparison results. These evaluation metrics may include accuracy, precision, recall, F1 value, mean square error, confusion matrix, etc.

[0092] After the multi-task learning neural network model is trained, the model can be used to analyze the baseband signal received by the radio frequency system to identify its corresponding signal type and its modulation method; this is specifically achieved through the following step S103.

[0093] S103: Preprocess the baseband signal received by the radio frequency system to obtain a time domain complex signal to be measured and a frequency domain characteristic signal to be measured, and transmit them as input data to the pre-trained multi-task learning neural network model to obtain an identification result of the model output signal type and modulation mode.

[0094] Specifically, preprocessing the baseband signal received by the RF system to obtain the time-domain complex signal to be measured and the frequency-domain characteristic signal to be measured includes: extracting the time-domain complex signal to be measured based on the baseband signal received by the RF system; and calculating the power spectral density of the time-domain complex signal to be measured using a Welch overlapping piecewise average estimator to obtain the corresponding frequency-domain characteristic signal to be measured. That is, the IQ complex signal received by the baseband is naturally treated as a time-domain signal; and the PSD data reflecting the frequency-domain characteristics is obtained based on the IQ complex signal using a Welch overlapping piecewise average estimator. After the IQ complex signal directly received at the baseband and the PSD data obtained by processing are input as two input data into a pre-trained multi-task learning neural network model, so that the multi-input layers of the multi-task learning neural network model extract their own features and then transmit the extracted feature parameters to the shared layer of the multi-task learning neural network, which fuses the multi-input features and further extracts the multi-task common features according to the constraints of multiple task outputs; then the multi-task common features are transmitted to the multi-output layer, and finally the multi-output layer realizes the recognition of signal type and modulation mode and the output of recognition results.

[0095] This embodiment, by implementing highly accurate signal type and modulation identification methods, can help satellites used in integrated space-ground networks process wireless signals of multiple communication protocols on the same baseband hardware, enhancing the software-definable capabilities of satellite systems. A satellite-borne communication terminal with digital signal processing capabilities can identify the signal type and modulation method of satellite communication signals received via a radio frequency link, achieving intelligent multi-task recognition. Signal type and modulation method identification through neural networks eliminates the need for specialized signal processing algorithms and provides the ability to update neural network models, effectively enhancing the iterative update capabilities of satellite signal processing.

[0096] According to the description of the above specific implementation methods, this embodiment carried out corresponding simulation verification tests.

[0097] The MTL model is trained using the constructed satellite communication protocol and modulation dataset, and corresponding network model parameters are obtained through iteration.

[0098] Each signal corresponding to each data point in the dataset is generated using random numbers. The simulation dataset includes five signal types and six modulation schemes, as well as two channel types: AWGN and LMS, totaling 58,880 data points. The training, validation, and test sets are split in a 6:2:2 ratio.

[0099] The neural network architecture used in the simulation experiments was built using Python 3.10 and TensorFlow v2.14. The model used Adam as the optimizer, with a learning rate of 1e-4, β1 and β2 set to 0.9 and 0.999, respectively. The loss function was the cross-entropy function.

[0100] The statistical analysis method for MTL model test results is as follows: Based on the dataset generation method, a test dataset is generated to test the performance of the trained MTL model. First, the classification accuracy based on signal type and modulation method changes is analyzed. The test dataset is divided into 2dB steps of signal-to-noise ratio. The data corresponding to each signal-to-noise ratio (SNR) is input into the trained model and the corresponding recognition accuracy is obtained. A more accurate method to describe recognition accuracy is the confusion matrix, which describes the specific discrimination results for each category. In other words, the confusion matrix can display the quantitative relationship between the actual labels and predicted values ​​of the test set in matrix form; compared with a single classification accuracy value, the confusion matrix can more accurately describe the specific categories of classification errors.

[0101] See also Figure 5a-5b , the signal type recognition rate and modulation mode recognition rate both reached over 90% when the signal-to-noise ratio was greater than 5dB. In the AWGN channel scenario with a signal-to-noise ratio of -5dB, the signal type recognition rate even reached 93%. In modulation mode recognition, the recognition accuracy of signals affected only by the AWGN channel reached 95% faster than that of signals affected by the LMS channel, and signal type recognition also showed a similar trend. This also indicates that the fading effect of the LMS channel is more severe. Despite this, the MTL model still demonstrated excellent feature fitting capabilities; when the signal-to-noise ratio was higher than 8dB, the communication protocol classification accuracy in the AWGN and LMS channel scenarios was close to 100%.

[0102] Under the condition of SNR=4dB, the confusion matrix of multiple tasks is as follows Figure 6a to Figure 6d As shown, the signal type accuracy is over 99%, while the modulation mode recognition rate is relatively low, ranging from 90% to 95%. This performance loss is primarily due to the recognition of 16APSK and 16QAM modulation modes. Analysis shows that when the signal-to-noise ratio is not ideal, the dispersion of the constellation diagram can make it difficult to distinguish between an ideal circle and a square, resulting in modulation mode identification errors. This also reflects that the higher the modulation order, the higher the signal-to-noise ratio required for demodulation. Simulation classification results demonstrate that the proposed MTL model is capable of accurately identifying multiple tasks simultaneously. In the future scenario of integrated communication protocols in space-ground networks, the MTL model is expected to help quickly determine the communication protocol and modulation mode, further optimize synchronization and demodulation strategies, and achieve more efficient satellite communications.

[0103] In summary, the present invention provides a multi-task learning neural network model for satellite baseband processing of radio signals of different communication protocols, which can simultaneously identify signal types and modulation methods. A single network model can be reused to identify multiple signal types and modulation methods, thereby improving the efficiency of signal processing. The multi-task learning neural network model of this embodiment completes the identification of different communication protocols and modulation methods for scenarios containing satellite communication channel characteristics, and conducts targeted training and testing. For scenarios where satellite baseband processes radio signals of different communication protocols, this embodiment also provides a strategy for constructing a data set that can take into account the impact of the satellite-to-ground channel on the wireless signal while considering the signal type and modulation method.

[0104] It should be noted that the present application can be implemented in a combination of software and / or software and hardware, for example, can be implemented using an application specific integrated circuit (ASIC), a general purpose computer or any other similar hardware device. In one embodiment, the software program of the present application can be executed by a processor to implement the above steps or functions. Similarly, the software program of the present application (including relevant data structures) can be stored in a computer-readable recording medium, for example, a RAM memory, a magnetic or optical drive or a floppy disk and similar devices. In addition, some steps or functions of the present application can be implemented using hardware, for example, as a circuit that cooperates with a processor to perform each step or function.

[0105] The method according to the present invention can be implemented on a computer as a computer-implemented method, or implemented in dedicated hardware, or implemented in a combination of the two. The executable code for the method according to the present invention, or parts thereof, can be stored on a computer program product. Examples of computer program products include memory devices, optical storage devices, integrated circuits, servers, online software, etc. Preferably, the computer program product includes non-transitory program code components stored on a computer-readable medium so that when the program product is executed on a computer, the method according to the present invention is executed.

[0106] In a preferred embodiment, the computer program comprises computer program code means adapted to perform all the steps of the method according to the invention when the computer program is run on a computer.Preferably, the computer program is embodied on a computer readable medium.

[0107] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A wireless signal modulation mode and signal type identification method based on multi-task learning, characterized in that: include: Constructing a satellite communication protocol and modulation data set; wherein the satellite communication protocol and modulation data set includes a plurality of composite sample data and label information for marking the composite sample data, each composite sample data includes time domain complex data and power spectrum density frequency domain data, and the label information includes signal type and modulation mode; Based on the satellite communication protocol and modulation data set, a multi-task learning neural network model is trained; the multi-task learning neural network model includes multiple input layers, shared layers, and multiple output layers; the multiple input layers are used to receive multiple input data and extract corresponding feature parameters, the shared layers are used to fuse the feature parameters and extract multi-task common features based on multi-task output constraints, and the multiple output layers are used to output multi-task recognition results based on the multi-task common features; The baseband signal received by the radio frequency system is preprocessed to obtain the time domain complex signal and the frequency domain characteristic signal to be measured, and are transmitted as input data to the pre-trained multi-task learning neural network model to obtain the recognition results of the model output signal type and modulation method.

2. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 1 is characterized in that: The power spectrum density frequency domain data in the composite sample data is calculated based on the corresponding time domain complex data using a Welch overlapping piecewise average estimator; The preprocessing of the baseband signal received by the radio frequency system to obtain the time domain complex signal to be measured and the frequency domain characteristic signal to be measured includes: Extracting the time-domain complex signal to be measured based on the baseband signal received by the radio frequency system; The power spectrum density of the time-domain complex signal to be measured is calculated by using a Welch overlapping segmented average estimator to obtain a corresponding frequency-domain characteristic signal to be measured.

3. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 1 is characterized in that: The multi-input layer includes a first group of neuron structures and a second group of neuron structures, the first group of neuron structures is used to extract features of the time domain complex signal to be measured, and the second group of neuron structures is used to extract features of the frequency domain characteristic signal to be measured; The first group of neuron structures includes a first convolution module, a first dropout layer, a first reconstruction layer, a first maximum pooling layer and a first flattening layer; The second group of neuron structures includes a second convolution module, a second Dropout layer, a second maximum pooling layer and a second flattening layer.

4. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 3 is characterized in that: The first convolution module and the second convolution module are both composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu; The characteristic size of the time-domain complex signal to be measured after being reshaped by the first reconstruction layer and the first maximum pooling layer is consistent with the characteristic size of the frequency-domain characteristic signal to be measured after being reshaped by the second maximum pooling layer.

5. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 1 is characterized in that: The shared layers include: A fusion layer, configured to fuse the feature parameters output by the multiple input layers; The second reconstruction layer is used to reconstruct the dimension of the fused feature data; A third convolution module is used to extract the fusion features of the time domain complex signal and the frequency domain feature signal according to the reconstructed feature data; the third convolution module is composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu; The third Dropout layer is used to randomly disable some features of the fusion feature; The third maximum pooling layer is used to reduce the dimension of the fusion feature to output the multi-task common feature.

6. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 1 is characterized in that: The multi-output layer includes a first task branch structure for classifying and identifying signal types and a second task branch structure for classifying and identifying modulation modes; The first task branch structure includes a fourth convolution module, a fourth Dropout layer, a first dense module, a fifth Dropout layer, a third flattening layer, and a first dense layer with a SoftMax activation function; The second task branch structure includes a fifth convolution module, a sixth Dropout layer, a second dense module, a seventh Dropout layer, a fourth flattening layer, and a second dense layer with a SoftMax activation function; Among them, the fourth convolution module and the fifth convolution module are both composed of a convolution layer, a batch normalization layer and a nonlinear activation function ReLu.

7. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 6 is characterized in that: The first task branch structure obtains the multi-task common features, and extracts a first target feature corresponding to the signal type recognition task from the multi-task common features through the fourth convolution module and the first dense module, so that the first dense layer outputs a first recognition result of the signal type based on the first target feature; The second task branch structure obtains the common features of the multiple tasks, and extracts the second target features of the corresponding modulation mode identification task from the common features of the multiple tasks through the fifth convolution module and the second dense module, so that the second dense layer outputs the second recognition result of the modulation mode based on the second target feature.

8. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 1, characterized in that: The training of the multi-task learning neural network model based on the satellite communication protocol and the modulation data set includes: Dividing the satellite communication protocol and modulation data set into a training set, a validation set, and a test set according to a preset ratio; Constructing a multi-task learning neural network model, and iteratively training the multi-task learning neural network model using the training set; Performing model adjustment and model parameter determination on the multi-task learning neural network model according to the verification set; The test set is input into the multi-task learning neural network model to obtain corresponding output data to evaluate the model performance.

9. The wireless signal modulation mode and signal type identification method based on multi-task learning according to any one of claims 1 to 8, characterized in that: The constructing of satellite communication protocol and modulation data set includes: Select several communication protocols; Based on the modulation and coding mechanisms of the selected communication protocols, generate a plurality of time-domain complex signals, and mark the time-domain complex signals with label information, wherein the label information includes a signal type and a modulation method corresponding to the generation of the time-domain complex signals; Outputting the time domain complex signal through at least one preset channel and signal-to-noise ratio; Performing power spectral density calculation on the time-domain complex signal using a Welch overlapping segmented average estimator to obtain a corresponding frequency-domain characteristic signal; The time domain complex signal, the frequency domain characteristic signal and the corresponding label information are constructed to generate the composite sample data, and stored in a preset satellite communication protocol and modulation data set.

10. The wireless signal modulation mode and signal type identification method based on multi-task learning according to claim 9, characterized in that: The at least one channel includes an AWGN channel and an LMS channel; The communication protocols include S2 protocol, S2X protocol, RCS2 protocol, CCSDS TM protocol and CCSDS TC protocol; The modulation and coding mechanisms of the communication protocol include BPSK, PI / 2BPSK, QPSK, 8PSK, 16APSK and 16QAM.

Images