A radio signal recognition method and system based on machine learning

Abstract

The present invention discloses a radio signal identification method and system based on machine learning, including: obtaining a monitoring signal of a radio spectrum; calculating the time difference of signal arrival and the angle of arrival of the signal, inputting them into a neural network model, and outputting interference identification and positioning results of the monitoring signal; dividing the monitoring signal into signal fragments of a single interference source, and reconstructing the radio spectrum, and inputting them into the neural network model again, and outputting prediction results about the signal type. Combining the time difference of signal arrival, the angle of arrival of the signal, the identification of interference signals, and the precise positioning of the interference source, it is possible to more effectively filter out interference, reconstruct the radio spectrum, and use the labeled spectrum data to train the machine learning model to identify and classify different radio signals, thereby improving the accuracy of radio signal identification while also enhancing the system's adaptability to complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field of radio signal recognition, and in particular to a radio signal recognition method and system based on machine learning. Background Art

[0002] The TACAN system is a radio navigation system used for aircraft navigation. It is mainly used to measure the distance and azimuth between the aircraft and the TACAN station on the ground. The principle of TACAN system angle measurement is based on the propagation time and phase difference of radio signals. The TACAN system relies on high-precision radio signals to measure the distance and azimuth between the aircraft and the TACAN station on the ground, and requires accurate identification of specific signal patterns.

[0003] In the existing technology, the TACAN system faces challenges in radio signal recognition, especially in the location and identification of interference signals. Although the TACAN system is designed to provide precise navigation by measuring the distance and azimuth between the aircraft and the TACAN station on the ground, the presence of multiple radio interference sources, such as other aviation communications and ground radio transmissions, will introduce spectral anomalies and background noise, reduce the signal-to-noise ratio, and thus affect the clarity and recognition of the signal. These interferences will not only lead to a decrease in positioning accuracy, but may also cause communication interruptions and difficulties in signal recognition, increasing flight safety risks. Existing technologies have limitations in identifying and locating interference sources, especially in complex electromagnetic environments, where it is difficult to accurately measure the time difference and arrival angle of signals arriving at different stations, which limits the ability of the TACAN system to adapt to complex environments and improve signal processing efficiency.

[0004] Therefore, the existing technology has obvious deficiencies in the precise positioning and effective management of interference signals. More advanced methods are needed to improve the performance and reliability of the TACAN system in modern aviation navigation. It is necessary to provide a radio signal recognition method based on machine learning to solve the above problems. Summary of the invention

[0005] The present invention overcomes the deficiencies of the prior art and provides a radio signal recognition method and system based on machine learning.

[0006] To achieve the above object, the technical solution adopted by the present invention is: a radio signal recognition method based on machine learning, comprising the following steps:

[0007] Acquisition of monitoring signals of the radio spectrum;

[0008] Pre-process the monitoring signal and calculate the time difference of signal arrival and the angle of arrival of the signal;

[0009] The monitoring signal is input into the neural network model, the time difference of signal arrival and the angle of arrival of the signal are input, and the interference identification and positioning results of the monitoring signal are output;

[0010] Based on the interference identification and location results, the monitoring signal is divided into signal segments of a single interference source;

[0011] Filter signal fragments to remove interference and reconstruct the radio spectrum;

[0012] The reconstructed radio spectrum is fed back into the neural network model, which outputs a prediction about the signal type.

[0013] In a preferred embodiment of the present invention, the method for obtaining the monitoring signal of the radio spectrum is:

[0014] The radio signal between the aircraft and the ground TACAN station is acquired by adopting multi-site signal acquisition and synchronous acquisition in the TACAN system; wherein, multi-site signal acquisition is: by deploying several receivers and antenna arrays geographically.

[0015] In a preferred embodiment of the present invention, the method for calculating the time difference of signal arrival and the angle of arrival of the signal includes:

[0016] Get the signal timestamps from different receivers and calculate the time difference of signal arrival;

[0017] The phase of the antennas in the antenna array is adjusted to form a beam pointing in a specific direction and the angle of arrival of the signal is measured.

[0018] In a preferred embodiment of the present invention, the signal arrival time difference and the signal arrival angle of the monitoring signal are encoded into a heat map, and combined with the corresponding time-frequency map to form a multi-channel input tensor, which is input into the convolutional neural network model; the convolutional neural network model recognizes and classifies different interference signals to distinguish different interference sources;

[0019] Among them, the time-frequency diagram is: the corresponding monitoring signal is subjected to time-frequency analysis to obtain the time-frequency representation of the monitoring signal and converted into a time-frequency diagram.

[0020] In a preferred embodiment of the present invention, the time difference of arrival of signals between different receivers is calculated by maximum likelihood estimation.

[0021] In a preferred embodiment of the present invention, the method for dividing the monitoring signal into signal segments of a single interference source is a time axis marking method.

[0022] In a preferred embodiment of the present invention, the method for filtering interference from signal segments is: applying a filter to each segmented signal segment to remove signal components that do not belong to the interference source, wherein the filter is a bandpass filter or a Wiener filter.

[0023] In a preferred embodiment of the present invention, the method for reconstructing a radio spectrum comprises:

[0024] Perform fast Fourier transform on the signal fragments after filtering out interference, and estimate the frequency and amplitude of the signal by finding the peak with the largest amplitude, thereby reconstructing the radio spectrum;

[0025] The reconstructed radio spectrum is corrected to compensate for errors or distortions introduced during the filtering and reconstruction process.

[0026] In a preferred embodiment of the present invention, the method for preprocessing the monitoring signal includes:

[0027] Convert the received monitoring signal into a digital signal;

[0028] Filter out the noise and signals in non-target frequency bands in the digital signal, and only retain the signals within the working frequency range of the TACAN system;

[0029] Adjust the digital signal amplitude to make it in the best dynamic range.

[0030] The present invention adopts a radio signal recognition system based on machine learning, based on the above-mentioned radio signal recognition method based on machine learning, comprising:

[0031] Signal acquisition and synchronization module, used to obtain monitoring signals of the radio spectrum;

[0032] A signal processing module, used for preprocessing the acquired monitoring signal of the radio spectrum and calculating the time difference of signal arrival and the angle of arrival of the signal;

[0033] Neural network model, used to identify and locate interference of monitoring signals, and to identify and reconstruct radio spectrum, and obtain the characteristics of signal frequency, bandwidth, and signal category;

[0034] A signal segmentation module, used to segment the signal into signal segments of a single interference source according to the interference identification result;

[0035] The signal filtering and reconstruction module is used to filter out interference from signal fragments and reconstruct the radio spectrum.

[0036] The present invention solves the defects existing in the background technology and has the following beneficial effects:

[0037] The present invention provides a radio signal recognition method based on machine learning. By combining the time difference of signal arrival, the arrival angle of the signal, interference signal recognition and precise positioning of the interference source, it is possible to more effectively filter out interference, reconstruct the radio spectrum, and use labeled spectrum data to train a machine learning model to recognize and classify different radio signals. This improves the accuracy of radio signal recognition while also enhancing the adaptability of the TACAN system to complex environments.

[0038] The present invention uses the means of multi-site synchronous acquisition and signal feature extraction, deploys high-sensitivity receivers and antenna arrays at multiple geographically dispersed sites, and uses high-precision GPS clocks or other synchronization devices to ensure that the timestamps of all receivers are accurately synchronized to the nanosecond level. This technical means enables the system to accurately measure the time difference and arrival angle of signals arriving at different sites, providing key time and space information for subsequent signal processing and interference identification; it not only improves the accuracy of interference identification, enhances the signal source positioning capability, but also improves the quality of signal reconstruction. Compared with the prior art, it significantly reduces the false alarm and missed alarm rates, improves the efficiency of signal processing, and enhances the adaptability and robustness of the TACAN system in complex electromagnetic environments, providing a more reliable radio signal identification solution.

[0039] In the present invention, the signal arrival time difference positioning can determine the distance of the signal source in space, and does not rely on the power information of the signal. The signal arrival angle positioning uses the incident angle information when the signal arrives at the receiver to determine the position of the interference source, which is not affected by multipath propagation and non-line-of-sight propagation, and the positioning accuracy is relatively high. By combining the signal arrival time difference positioning and the arrival angle positioning, not only the positioning accuracy is improved, but also the adaptability to complex electromagnetic environments is enhanced. For example, in an environment with severe multipath effects, the signal arrival time difference positioning can provide more stable positioning information, while the signal arrival angle positioning can assist in verifying and correcting the signal arrival angle results through the time difference. The combination of the two technologies can reduce the possible error sources of a single technology, such as clock synchronization errors or antenna array calibration errors, thereby improving the robustness of the overall system. In this way, radio signals can be more accurately identified and classified, providing more reliable data support for radio spectrum management.

[0040] The present invention relates to a method for interference identification, positioning and signal reconstruction, which aims to improve the clarity of radio signals and the reliability of communication. The method uses signal processing technology and machine learning algorithms to identify and locate interference sources of mixed signals, including time-frequency analysis, feature extraction and classification of signals to accurately determine the source and nature of the signals. By accurately identifying the interference sources, these interferences can be effectively managed and controlled to reduce or eliminate their negative impact on normal communications.

[0041] The present invention relates to segmenting a signal into signal segments of a single interference source so that each interference source can be processed in a targeted manner. The segmentation process is based on the characteristics and behavior of the signal, such as frequency, bandwidth and modulation mode, and uses a machine learning model to distinguish different signal sources. The segmented signal segments can be applied with specific filters, such as bandpass filters or Wiener filters, to remove signal components that do not belong to the interference source, thereby purifying the signal and reducing or eliminating the impact of interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 is a flow chart of a radio signal recognition method based on machine learning according to a preferred embodiment of the present invention;

[0044] Figure 2 It is a block diagram of a radio signal recognition system based on machine learning according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0047] Exemplary methods:

[0048] Figure 1 The figure illustrates a flow chart of a radio signal recognition method based on machine learning according to an embodiment of the present application.

[0049] A radio signal recognition method based on machine learning according to an embodiment of the present application includes the following steps:

[0050] Step S1, obtaining a monitoring signal of a radio spectrum;

[0051] Step S2, pre-processing the monitoring signal, and calculating the time difference of signal arrival and the angle of arrival of the signal;

[0052] Step S3, input the monitoring signal into the neural network model, input the time difference of signal arrival and the angle of arrival of the signal, and output the interference identification and positioning results of the monitoring signal;

[0053] Step S4: according to the interference identification and positioning results, the monitoring signal is divided into signal segments of a single interference source;

[0054] Step S5, filtering out interference from the signal fragments and reconstructing the radio spectrum;

[0055] Step S6: input the reconstructed radio spectrum into the neural network model again, and output the prediction result about the signal type.

[0056] Below, each step will be described in detail.

[0057] In step S1, by setting up multi-site signal acquisition in the TACAN system, and acquiring the radio signal between the aircraft and the ground TACAN station in a synchronous acquisition manner, that is, the monitoring signal of the radio spectrum.

[0058] Multi-site signal acquisition is to deploy high-sensitivity receivers and antenna arrays at multiple geographically dispersed sites to synchronously acquire radio signals between TACAN beacons and ground TACAN stations. Synchronous acquisition is to use high-precision GPS clocks or other synchronization devices (such as atomic clocks or Network Time Protocol NTS servers) to ensure that the timestamps of all receivers are accurately synchronized to the nanosecond level. Multi-site acquisition can determine the location of the signal source by measuring the time difference and arrival angle of the signal at different sites, and deploying multiple receivers in a geographical area can enhance signal coverage; synchronous acquisition ensures that all sites work under the same time reference and accurately measures the time difference of the signal arriving at different sites.

[0059] In step S2, the monitoring signal is preprocessed, and the time difference of signal arrival and the angle of arrival of the signal are calculated, which specifically includes:

[0060] preprocessing the acquired monitoring signal of the radio spectrum;

[0061] Get the signal timestamps from different receivers and calculate the time difference of signal arrival;

[0062] The phase of the antennas in the antenna array is adjusted to form a beam pointing in a specific direction and the angle of arrival of the signal is measured.

[0063] Further, monitoring signal preprocessing includes:

[0064] Converting the received monitoring signal into a digital signal can be achieved through an analog-to-digital converter (ADC);

[0065] Filtering out noise and signals in non-target frequency bands in the digital signal and retaining only signals within the operating frequency range of the TACAN system can be done by passing the digital signal through a bandpass filter, wherein the passband range of the bandpass filter covers the operating frequency of the TACAN signal and suppresses frequency components outside the passband;

[0066] To adjust the digital signal amplitude to the optimal dynamic range, the automatic gain control (AGC) system can monitor the power level of the digital signal and dynamically adjust the gain of the receiver to ensure that the signal maintains a stable dynamic range under changing signal strength.

[0067] For example, the 1000MHz signal from the TACAN beacon is sampled and digitized at a rate of 2.56 million times per second; the bandpass filter is designed to cover the range of 960MHz to 1215MHz to filter out interference outside the TACAN signal operating frequency and retain key signal components. The automatic gain control (AGC) system dynamically adjusts the receiver gain to cope with changes in signal strength, ensuring that the signal always remains within the optimal dynamic range, preventing overload or distortion, and providing clear and accurate digital signals for further analysis and processing of the signal.

[0068] Furthermore, the time difference of signal arrival between different receivers is calculated by maximum likelihood estimation. The maximum likelihood estimation method maximizes the probability of observed data by finding the optimal model parameters (the distance from the signal source to the receiver) and estimates the signal arrival time difference using the optimal parameters.

[0069] For example, suppose the signal propagation speed is , the arrival time of the signal at receiver A and receiver B is and , the distance from the signal source to receiver A and receiver B is and , the two receivers observe the signal and ; Build a signal arrival time model: , ;in, is the time when the signal source transmits the signal; establish the time difference model: ;in, is the time difference between the signal reaching receiver A and receiver B; the likelihood function is expressed as: ; Where p represents the probability density function; In order to make and The probability is the largest, and the gradient descent algorithm is used to adjust and The value of to maximize the likelihood function L, and get and The estimated value is and , and then calculate the time difference of signal arrival .

[0070] Further, antenna arrays and beamforming techniques are used to calculate the angle of arrival of the signal. An array consisting of multiple antenna elements is deployed, and by adjusting the phase of the signal received by each antenna element, a beam pointing in a specific direction can be formed. This is usually achieved by changing the delay in the signal path so that the signals from a specific direction are spatially coherently superimposed, while the signals from other directions cancel each other out. The direction in which the beam is pointing is measured, which can be achieved by finding the direction with the greatest signal strength after beamforming.

[0071] For example, consider a linear array consisting of N antenna elements, with the antenna elements evenly distributed on a straight line, the distance between adjacent antenna elements is d, and the signal source is located at an angle θ (relative to the normal of the array). Signal phase difference: For the nth antenna element in the array, the angle θ at which the signal arrives will result in a phase difference between adjacent antenna elements ,in, is the wavelength of the signal; to form a beam pointing at an angle θ, a weight can be assigned to each antenna element , , where j is the imaginary unit and n is the index of the antenna element; the signal received by each antenna element is multiplied by its corresponding weight and then summed to form a beam pointing in a specific direction; by finding the angle θ that maximizes the beam signal strength, the signal arrival angle can be estimated ,in, is the signal received by the nth antenna element.

[0072] In step S3, the monitoring signal is input into the neural network model, the time difference of signal arrival and the angle of arrival of the signal are input, and the interference identification and positioning results of the monitoring signal are output.

[0073] The monitoring signal is input into the convolutional neural network model, which is used to process the superimposed mixed signal of wireless signals and interference waveforms in TACAN. The convolutional neural network model is used to identify and classify different interference signals in the superimposed mixed signal and distinguish different interference sources.

[0074] Since the time difference and arrival angle of the monitoring signal are scalar data, they describe the physical properties of signal propagation, rather than two-dimensional or three-dimensional structured data such as images; in order to identify and locate the interference source of the radio signal, the convolutional neural network (CNN) structure provided in this embodiment can process the time-frequency graph, the time difference heat map and the arrival angle heat map. Here, these scalar data can be processed by CNN together with the time-frequency representation of the signal through feature fusion and multi-modal input.

[0075] The following is a detailed convolutional neural network (CNN) structure. The CNN structure is designed to extract and compress features in the input data and classify interference sources with high accuracy. Specifically, the network consists of the following main parts:

[0076] Input layer: receives a multi-channel tensor , each channel represents the time-frequency diagram, time difference heat map and arrival angle heat map respectively; where H is the height of the frequency axis, W is the width of the time axis, the first channel is the time-frequency diagram, which represents the frequency and time characteristics of the signal; the second channel is the time difference heat map, which represents the time difference of the signal arriving at different stations; the third channel is the arrival angle heat map, which represents the angle information of the signal arrival.

[0077] Convolutional layer and pooling layer: The network gradually extracts features at different levels through multiple convolutional layers and pooling layers; each convolutional layer uses a small 3x3 filter to extract local features and introduces nonlinearity through the ReLU activation function; the pooling layer uses the maximum pooling operation to reduce the dimension and retain the most important features, reducing the amount of calculation and the risk of overfitting. Specifically, the network contains three convolutional layers and corresponding pooling layers: the first convolutional layer has 32 filters, the second convolutional layer has 64 filters, and the third convolutional layer has 128 filters; each convolution layer is followed by a 2x2 maximum pooling layer with a stride of 2 to further compress the size of the feature map.

[0078] Flattening layer: After multiple layers of convolution and pooling operations, the network flattens the multi-dimensional feature map into a one-dimensional vector so that it can be input into the fully connected layer for further feature extraction and classification.

[0079] Fully connected layers: The two fully connected layers contain 256 and 128 nodes respectively. Each node uses the ReLU activation function, and a Dropout layer (dropout rate is 0.5) is added to prevent overfitting; these fully connected layers are used to extract high-level features and map them to the final classification task.

[0080] Output layer: The output layer uses the Softmax activation function to map features to the probability distribution of each category, indicating the possibility that the input signal belongs to different interference sources. The number of nodes in the output layer depends on the number of categories of the classification task. For example, if there are 5 different interference sources, the output layer has 5 nodes.

[0081] Auxiliary input layer: This layer receives the scalar data of time difference and arrival angle and concatenates it with the feature map of the main path; specifically, the time difference values ​​of all site pairs are combined into a vector T, and the arrival angles of all antennas are combined into a vector A. The two vectors are concatenated into a new vector ; and broadcast to each spatial position of the feature map after the last convolutional layer to form a new feature map In this way, the fully connected layer can process both spatiotemporal features and physical property information at the same time, further improving the expressive power of the model.

[0082] The loss function is cross entropy loss, which is used for classification problems; the optimizer is Adam optimizer.

[0083] Through the above structure, CNN can not only process complex spatiotemporal features, but also effectively integrate physical attribute information such as time difference and arrival angle, thereby realizing high-precision interference source identification and location of radio signals. This design makes the model more adaptable and robust in complex electromagnetic environments, and is suitable for radio signal recognition tasks in practical applications.

[0084] The following are the steps to simulate interference signals using simulation tools, build a dataset, and train a convolutional neural network (CNN) model to classify different interference source categories.

[0085] Step 1: Determine different simulated interference types, such as narrowband interference, broadband interference, pulse interference, etc.; use simulation tools (such as MATLAB) to set up the simulation environment, including signal parameters (frequency, bandwidth, pulse width, etc.); generate simulated signals of various interference types in the simulation tool; compare the simulated signals with the real measured data, and assign a category label to each simulated signal and monitoring signal sample for supervised learning;

[0086] Step 2: Perform time-frequency analysis on each simulation signal or monitoring signal, such as short-time Fourier transform (STFT) or Wigner-Ville distribution, to obtain the time-frequency representation of the signal; convert the time-frequency data into a time-frequency graph.

[0087] Step 3: Encode the time difference and arrival angle of the monitoring signal of the simulation signal or the monitoring signal into a heat map, and combine it with the corresponding time-frequency map to form a number of multi-channel input tensors, that is, to form a data set; divide the data set into a training set, a validation set, and a test set;

[0088] Step 4: Design the architecture of CNN, including input layer, convolution layer, pooling layer, flattening layer, fully connected layer, output layer and auxiliary input layer; use the training set data in step 3 to train the CNN model, adjust the network weights through optimization algorithms such as back propagation and gradient descent, and the model will output the classification probability or label for each time point or time period, indicating the type of interference source that the signal is most likely to belong to in that time period.

[0089] For example, when using a convolutional neural network (CNN) model, after calculating the time difference and arrival angle of the monitoring signal after the above signal arrives, the time difference and arrival angle of the monitoring signal and the monitoring signal are input into the above pre-trained CNN model. The model will encode the time difference and arrival angle of the monitoring signal into a heat map and combine it with the corresponding time-frequency map to form a multi-channel input tensor. The method of encoding the time difference and arrival angle of the monitoring signal into a heat map includes: using an interpolation method to extend the time difference value to the time axis of the entire time-frequency graph to generate a time difference heat map; mapping the arrival angle corresponding to each time point into the heat map to generate an arrival angle heat map;

[0090] The multi-channel input tensor is input into the CNN model. The network gradually extracts features through the convolution layer and the pooling layer, and performs high-level feature fusion in the fully connected layer. After being processed by the fully connected layer and the output layer, the network finally outputs the model. The model will output the classification probability or label at each time point or time period, indicating the type of interference source that the signal is most likely to belong to during the time period. The number of nodes in the output layer depends on the number of categories of the classification task. For example, if there are 5 different interference sources, the output layer has 5 nodes. The output value of each node represents the probability of the interference source. In addition, the network outputs the location information of the interference source and predicts the coordinates of the interference source by regression. Specifically, additional nodes can be added to the output layer to predict the values ​​of the time difference and the angle of arrival, so as to achieve accurate positioning of the interference source.

[0091] In step S4, the monitoring signal is segmented into signal segments of a single interference source according to the interference identification and positioning results.

[0092] When multiple interference sources exist at the same time, they may affect each other, making it difficult to effectively remove the overall interference. By decomposing the mixed signal into separate signal fragments from different interference sources, the most suitable technology can be applied to purify each signal fragment separately, thereby improving the interference processing accuracy of the entire system.

[0093] The method of segmenting the signal segments into single interference sources in the present invention is the time axis marking method. According to the classification probability output by the convolutional neural network model, a threshold is set to determine the classification of the signal. According to the classification results, the start and end positions of each interference source are marked on the time axis, which can be achieved by finding continuous time periods with a probability exceeding the threshold, and using peak detection to determine the boundaries of the signal segments. For example, if the model predicts that the probability that a signal in a certain time period belongs to a specific interference source is greater than 95%, the signal in that time period is considered to belong to that interference source.

[0094] For example, using the time axis marking method, the signal fragments belonging to a single interference source can be segmented from the mixed signal, including: the classification probability of the convolutional neural network for a specific interference source at time t ,in, is the output of the convolutional neural network at time t; if , then t belongs to the interference signal segment, where is the set threshold (such as 0.95). Let T be the total duration of the signal, is the signal segment of the ith interference source, is the starting time of the signal segment of the i-th interference source, is the end time of the signal segment of the i-th interference source.

[0095] Initially, and ; For each time point t, if and , then update ;like and , then update ;make is the probability difference of adjacent time points; find The local maximum of the signal segment is used to determine the boundary of the signal segment; for each interference source i, the signal segment is extracted ,in, is the signal value at time t.

[0096] In some embodiments, if the signals of different interference sources overlap in time, this can be achieved by using independent component analysis ICA or a two-channel recurrent neural network method.

[0097] The following dual-channel recurrent neural network (DPRNN) method is used as an example to deal with the problem of temporal overlap of signals from different interference sources. Specifically, the long sequence signal is divided into shorter blocks, each block contains K consecutive time lengths, and the overlapping part length is P; the dual-channel recurrent neural network (DPRNN) model is applied to each block, alternating local and global modeling; the processed blocks are overlapped and added to restore the continuous signal sequence.

[0098] Permutation invariance training, optimizing model parameters to minimize the estimated mask Mask with ideal ratio The mean square error between . , where MSE stands for mean square error; ,in, is the spectrum of a single signal, and Y is the spectrum of all signals. sum.

[0099] In step S5, the signal fragments are filtered to remove interference and the radio spectrum is reconstructed.

[0100] The method of filtering interference from signal segments is to apply a specific filter to each segmented signal segment to remove signal components that do not belong to the interference source. The filter here can be a bandpass filter or a Wiener filter. Among them, the Wiener filter is an adaptive filter that minimizes the mean square error and is suitable for situations where the statistical characteristics of the signal and interference are known. Through the application of the filter, the signal components that do not belong to the interference source are effectively removed from each segmented signal segment, thereby purifying the signal and reducing or eliminating the impact of interference.

[0101] The following takes the Wiener filter as an example.

[0102] Frequency response of the filter ,in, is the autopower spectral density of the signal, is the cross-power spectral density, is the variance of the interference; after applying the above filter, we get the signal fragment after filtering out the interference , ,in, represents the Wiener filter function, is the original signal fragment.

[0103] The method for reconstructing the radio spectrum in the present invention includes: performing fast Fourier transform on the signal segments after filtering out interference, and estimating the frequency and amplitude of the signal by finding the peak with the largest amplitude, thereby reconstructing the signal and the radio spectrum; and correcting the reconstructed spectrum to compensate for errors or distortions that may occur during the filtering and reconstruction process.

[0104] Specifically, a fast Fourier transform (FFT) is performed on the signal segment after interference is filtered out to obtain its spectrum representation. ,in, It is a frequency domain signal; find the peak with the largest amplitude in the spectrum, and the frequency corresponding to the peak is the estimated frequency of the signal ; Estimate the signal amplitude based on the amplitude of the frequency peak ; Using the estimated parameters and , reconstruct the signal according to the formula of the original signal ,in, is the sampling frequency.

[0105] For the phase difference method, a section of the signal can be sampled and the sequence can be processed separately. Dot and FFT analysis of the points is performed and the spectrum is corrected using its phase difference. ,in and They are Dot and The FFT result of the point, Represents the phase angle of a complex number. Use full-phase FFT to suppress spectrum leakage. According to the phase difference and the results of full-phase FFT, the spectrum is corrected to compensate for errors or distortions that may occur during filtering and reconstruction. The spectrum correction formula is: ,in, is the spectrum before correction, is the phase difference.

[0106] By combining the spectrum analysis capability of Fourier transform and the spectrum correction capability of phase difference method, the radio spectrum is reconstructed and corrected to compensate for errors or distortions that may occur during filtering and reconstruction.

[0107] In step S6, the reconstructed spectrum data is provided as input to the trained CNN model. The CNN model analyzes the input reconstructed spectrum based on the previously learned knowledge and outputs a prediction result about the signal type.

[0108] The convolution layer of CNN scans the input data through a sliding window to capture local features (such as frequency peaks, spectrum shapes, etc.), while the pooling layer further compresses these features; the fully connected layer maps the extracted high-level features to specific classification tasks, and finally gives the prediction results through the output layer. According to the needs, additional nodes can be added to the output layer to predict frequency, bandwidth, modulation mode or signal strength, etc.

[0109] Example systems:

[0110] Figure 2 The figure illustrates a block diagram of a radio signal recognition system based on machine learning according to an embodiment of the present application.

[0111] like Figure 2 As shown, the radio signal recognition system based on machine learning according to an embodiment of the present application includes:

[0112] Signal acquisition and synchronization module, used to obtain monitoring signals of the radio spectrum;

[0113] A signal processing module, used for preprocessing the acquired monitoring signal of the radio spectrum and calculating the time difference of signal arrival and the angle of arrival of the signal;

[0114] Neural network model, used to identify and locate interference of monitoring signals, and to identify and reconstruct radio spectrum, and obtain the characteristics of signal frequency, bandwidth, and signal category;

[0115] A signal segmentation module, used to segment the signal into signal segments of a single interference source according to the interference identification result;

[0116] The signal filtering and reconstruction module is used to filter out interference from signal fragments and reconstruct the radio spectrum.

[0117] In one example, the signal acquisition and synchronization module includes: a multi-site receiver deployment unit and a time synchronization unit. The multi-site receiver deployment unit deploys high-sensitivity receivers and antenna arrays at multiple geographically dispersed sites to synchronously acquire radio signals between the aircraft and the ground TACAN station; the time synchronization unit is a high-precision GPS clock or other synchronization equipment (such as an atomic clock or a Network Time Protocol NTS server) to ensure that the timestamps of all receivers are accurately synchronized to the nanosecond level.

[0118] In one example, the signal processing module includes: an analog-to-digital converter, a bandpass filter, an automatic gain control system and a calculation unit; the analog-to-digital converter is used to convert the received monitoring signal into a digital signal, the bandpass filter is used to filter out noise and signals of non-target frequency bands in the digital signal, and only retain the signals within the operating frequency range of the TACAN system; the automatic gain control (AGC) system is used to dynamically adjust the gain of the receiver to cope with changes in signal strength and ensure that the signal always remains within the optimal dynamic range; the calculation unit is used to calculate the time difference of signal arrival and the arrival angle of the signal.

[0119] In one example, the neural network model is a convolutional neural network (CNN) model, which is used to process the superimposed mixed signals of wireless signals and interference waveforms in TACAN. The CNN model is used to identify and classify different interference signals in the superimposed mixed signals to distinguish different interference sources. The trained convolutional neural network (CNN) model is used to analyze the reconstructed radio spectrum to obtain the characteristics of the signal's frequency, bandwidth, and signal category.

[0120] In one example, the signal segmentation module is used to set a threshold to determine the classification of the signal based on the classification probability output by the CNN model, and mark the start and end positions of each interference source on the time axis to segment the signal into signal segments of a single interference source.

[0121] In one example, the signal filtering and reconstruction module includes: a Wiener filter and a signal recombining and correcting unit; the Wiener filter is used to apply a specific filter to each segmented signal segment to remove signal components that do not belong to the interference source; the signal recombining and correcting unit is used to recombine the signal segments after filtering out the interference, reconstruct the radio spectrum, and perform corrections to compensate for possible errors or distortions.

[0122] Those skilled in the art will appreciate that other details of the radio signal identification system based on machine learning according to the embodiment of the present application are the same as the corresponding details previously described in the radio signal identification method based on machine learning according to the embodiment of the present application, and will not be repeated here to avoid repetition.

[0123] The above is based on the ideal embodiment of the present invention. Through the above description, relevant personnel can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the content in the specification, and the technical scope must be determined according to the scope of the claims.

Claims

1. A radio signal recognition method based on machine learning, characterized in that: The following steps are involved: Acquisition of monitoring signals of the radio spectrum; Pre-process the monitoring signal and calculate the time difference of signal arrival and the angle of arrival of the signal; The monitoring signal is input into the neural network model, the time difference of signal arrival and the angle of arrival of the signal are input, and the interference identification and positioning results of the monitoring signal are output; According to the interference identification and location results, the monitoring signal is divided into signal segments of a single interference source using the time axis marking method; The signal segments are filtered to remove interference, and a radio spectrum is reconstructed; wherein the method for filtering the signal segments to remove interference is: applying a filter to each segmented signal segment to remove signal components that do not belong to the interference source, wherein the filter is a bandpass filter or a Wiener filter; The reconstructed radio spectrum is fed back into the neural network model, which outputs a prediction about the signal type. The method for reconstructing the radio spectrum includes: performing fast Fourier transform on the signal fragments after filtering out interference, estimating the frequency and amplitude of the signal by finding the peak with the largest amplitude, and then reconstructing the radio spectrum; correcting the reconstructed radio spectrum to compensate for errors or distortions generated in the filtering and reconstruction process; The signal arrival time difference and the signal arrival angle of the monitoring signal are encoded as a heat map and combined with the corresponding time-frequency map to form a multi-channel input tensor, which is input into the convolutional neural network model. The convolutional neural network model is used to identify and classify different interference signals and distinguish different interference sources. The time-frequency map is as follows: the corresponding monitoring signal is subjected to time-frequency analysis to obtain the time-frequency representation of the monitoring signal and convert it into a time-frequency map.

2. The method for radio signal recognition based on machine learning according to claim 1, characterized in that: The method for obtaining the monitoring signal of the radio spectrum is: The radio signal between the aircraft and the ground TACAN station is acquired by adopting multi-site signal acquisition and synchronous acquisition in the TACAN system; wherein, multi-site signal acquisition is: by deploying several receivers and antenna arrays geographically.

3. The method for radio signal recognition based on machine learning according to claim 2, characterized in that: The method for calculating the time difference of signal arrival and the angle of arrival of the signal includes: Get the signal timestamps from different receivers and calculate the time difference of signal arrival; The phase of the antennas in the antenna array is adjusted to form a beam pointing in a specific direction and the angle of arrival of the signal is measured.

4. The method for radio signal recognition based on machine learning according to claim 3, characterized in that: The time difference of signal arrival between different receivers is calculated by maximum likelihood estimation.

5. The method for radio signal recognition based on machine learning according to claim 1, characterized in that: The method for preprocessing the monitoring signal comprises: Convert the received monitoring signal into a digital signal; Filter out the noise and signals in non-target frequency bands in the digital signal, and only retain the signals within the working frequency range of the TACAN system; Adjust the digital signal amplitude to make it in the best dynamic range.

6. A radio signal recognition system based on machine learning, based on the recognition method according to any one of claims 1 to 5, characterized in that: include: Signal acquisition and synchronization module, used to obtain monitoring signals of the radio spectrum; A signal processing module, used for preprocessing the acquired monitoring signal of the radio spectrum and calculating the time difference of signal arrival and the angle of arrival of the signal; Neural network model, used to identify and locate interference of monitoring signals, and to identify and reconstruct radio spectrum, and obtain the characteristics of signal frequency, bandwidth, and signal category; A signal segmentation module, used to segment the signal into signal segments of a single interference source according to the interference identification result; The signal filtering and reconstruction module is used to filter out interference from signal fragments and reconstruct the radio spectrum.

Images