Interference suppression method and device based on deep learning

Abstract

The invention discloses an interference suppression method and device based on deep learning. According to the method, the pre-trained deep learning model is adopted to execute end-to-end interference suppression, and the limitation that a traditional method depends on a fixed rule can be effectively overcome. According to the method, training is carried out under the condition of a limited training set (such as a fixed signal-to-noise ratio, a multi-signal-to-interference ratio and comb interference combination), so that the model can adaptively learn an end-to-end mapping relation from an interference-containing receiving signal to a clean signal. After training is completed, the model can be deployed offline, shows high scene adaptability and generalization ability in practical application, and can cope with dynamic changes of a signal-to-noise ratio, a signal-to-interference ratio and an interference type, so that more accurate interference suppression is achieved in a complex and changeable interference environment, and damage to useful signals is avoided to the maximum extent.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and more specifically to an interference suppression method and device based on deep learning. Background Technology

[0002] In the field of wireless communication, the electromagnetic environment for transmission is becoming increasingly complex, and signals are susceptible to various interferences during transmission. Particularly in aerial flight scenarios, signals are also vulnerable to malicious interference. This interference can severely impact the performance of communication systems, reducing signal quality and reliability. Traditional interference suppression algorithms have many limitations when facing complex and ever-changing interference environments, and their improvement in anti-interference performance is limited. For example, the setting of the interference detection threshold has a significant impact on the final detection result, potentially leading to missed detections, false detections, or multiple detections. Furthermore, some conventional interference suppression methods, while suppressing interference, also affect the signal to some extent, resulting in performance degradation. Therefore, a new interference suppression method is needed to improve the anti-interference capability of communication systems.

[0003] In recent years, artificial intelligence technologies such as deep reinforcement learning have been widely applied in the field of communications, achieving breakthroughs in areas such as signal recognition, signal demodulation, channel estimation, and end-to-end communication. With its characteristics based on advanced signal processing and intelligent algorithms, it has become a new approach to solving communication system problems. Therefore, incorporating deep learning algorithms into communication systems, leveraging their powerful data feature learning capabilities, can achieve better interference suppression effects.

[0004] In related technologies, under complex and ever-changing wireless communication interference scenarios, traditional interference suppression algorithms rely on fixed processing models and prior knowledge, resulting in inaccurate anti-interference performance and damage to useful signals. Summary of the Invention

[0005] The technical problem this invention aims to solve is that, in complex and ever-changing wireless communication interference scenarios, traditional interference suppression algorithms, relying on fixed processing models and prior knowledge, suffer from inaccurate anti-interference performance and damage to useful signals. The purpose is to provide a deep learning-based interference suppression method and device that solves the problems of inaccurate anti-interference performance and damage to useful signals inherent in traditional interference suppression algorithms that rely on fixed processing models and prior knowledge.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides an interference suppression method based on deep learning, the method comprising:

[0008] Acquire the received signal; wherein the received signal is the original transmitted signal after interference and noise have been superimposed during transmission;

[0009] The received signal is preprocessed to obtain multi-channel frequency domain feature data that meets the input format requirements of the pre-trained interference suppression network model;

[0010] Multi-channel frequency domain feature data is input into a pre-trained interference suppression network model. The interference suppression network model performs forward inference calculations and outputs the corresponding multi-channel frequency domain feature data after interference suppression. The interference suppression network model is pre-trained using training data, which includes sample input data and corresponding sample label data. The sample input data represents the frequency domain feature data of the signal carrying interference, and the sample label data represents the corresponding frequency domain feature data of the signal without interference. The interference suppression network model, through training, learns the mapping relationship from the frequency domain features of the signal with interference to the frequency domain features of the signal without interference.

[0011] The multi-channel frequency domain feature data after interference suppression is post-processed to obtain the time domain signal after interference suppression, which is used for subsequent communication demodulation.

[0012] Further, the step of preprocessing the received signal to obtain multi-channel frequency domain feature data that conforms to the input format requirements of the pre-trained interference suppression network model includes:

[0013] The received signal is subjected to power normalization processing to obtain a normalized received signal;

[0014] The normalized received signal is windowed to obtain the windowed received signal.

[0015] The windowed received signal is processed by Fast Fourier Transform to obtain the corresponding frequency domain representation;

[0016] The frequency domain representation is separated into real and imaginary channels to obtain multi-channel frequency domain feature data.

[0017] Further, the step of performing power normalization processing on the received signal to obtain a normalized received signal includes:

[0018] Calculate the average power characteristic of the received signal over a preset time period;

[0019] The amplitude of the received signal is adjusted based on the average power characteristic to give the adjusted received signal a uniform energy scale, resulting in a normalized received signal.

[0020] Furthermore, the step of windowing the normalized received signal to obtain the windowed received signal includes:

[0021] A Hamming window is applied to the normalized received signal to obtain the windowed received signal.

[0022] Furthermore, the interference suppression network model includes an encoder path and a decoder path; wherein, the encoder path is used to downsample the input feature data to extract multi-scale features; the decoder path is used to upsample the feature data output by the encoder path to restore the signal resolution; wherein, there is a skip connection between the encoder path and the decoder path, which is used to pass the features extracted by the encoder path to the corresponding layer of the decoder path.

[0023] Furthermore, the interference signal included in the signal corresponding to the sample input data is a comb interference signal; the comb interference signal is generated by superimposing multiple independent interference components; wherein, each interference component is generated by band-limited noise signal modulation, and each interference component corresponds to a preset center frequency and a preset phase offset.

[0024] Furthermore, the step of post-processing the multi-channel frequency domain feature data after interference suppression to obtain the time domain signal after interference suppression for subsequent communication demodulation includes:

[0025] The real and imaginary channels of the multi-channel frequency domain feature data after interference suppression are combined to form a complex frequency domain signal;

[0026] The complex frequency domain signal is processed by inverse fast Fourier transform to obtain the time domain signal after interference suppression.

[0027] Secondly, the present invention provides an interference suppression device based on deep learning, the device comprising:

[0028] An acquisition module is used to acquire a received signal; wherein the received signal is the original transmitted signal after interference and noise have been superimposed during transmission.

[0029] The preprocessing module is used to preprocess the received signal to obtain multi-channel frequency domain feature data that meets the input format requirements of the pre-trained interference suppression network model.

[0030] The output module is used to input multi-channel frequency domain feature data into a pre-trained interference suppression network model. The interference suppression network model performs forward inference calculations and outputs the corresponding multi-channel frequency domain feature data after interference suppression. The interference suppression network model is pre-trained using training data, which includes sample input data and corresponding sample label data. The sample input data is the frequency domain feature data of the signal carrying interference, and the sample label data is the corresponding frequency domain feature data of the signal without interference. The interference suppression network model, through training, learns the mapping relationship from the frequency domain features of the signal with interference to the frequency domain features of the signal without interference.

[0031] The post-processing module is used to post-process the multi-channel frequency domain feature data after interference suppression to obtain the time domain signal after interference suppression, which is used for subsequent communication demodulation.

[0032] Thirdly, the present invention provides an electronic device, comprising: a memory, and one or more processors communicatively connected to the memory; the memory stores instructions executable by the one or more processors, the instructions being executed by the one or more processors to cause the one or more processors to implement the method described above.

[0033] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0034] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0035] This invention employs a pre-trained deep learning model to perform end-to-end interference suppression, effectively overcoming the limitations of traditional methods that rely on fixed rules and prior knowledge. The model is trained with limited samples, requiring only a training dataset consisting of a fixed signal-to-noise ratio (SNR), multiple signal-to-interference ratio (SIR), and comb interference. It adaptively learns the mapping relationship between interference signals and useful signals in the frequency domain, exhibiting strong generalization ability. During the inference phase, this method can accurately identify and suppress dynamically changing and diverse interference signals while preserving the integrity of useful signals to the maximum extent. It significantly improves the robustness and anti-interference capability of communication systems in complex interference scenarios, performing excellently under different SNR, SIR, and frequency domain interference environments. Furthermore, it does not overly rely on the distribution of the training set data, effectively reducing the bit error rate caused by inaccurate interference suppression, thereby ensuring the performance and reliability of subsequent communication demodulation stages. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0037] Figure 1 A flowchart illustrating a deep learning-based interference suppression method provided in the embodiments of this specification;

[0038] Figure 2 This is an architecture diagram of a deep learning-based interference suppression method provided in the embodiments of this specification;

[0039] Figure 3 This is a schematic diagram comparing the anti-interference performance under dynamic interference scenarios provided in the embodiments of this specification;

[0040] Figure 4 This is a schematic diagram comparing demodulation performance regression under dynamic interference scenarios provided in the embodiments of this specification;

[0041] Figure 5 This is a schematic diagram of the deep learning-based interference suppression network model provided in the embodiments of this specification;

[0042] Figure 6 This is a schematic diagram of the structure of an interference suppression device based on deep learning, provided in the embodiments of this specification.

[0043] Figure 7 This is a block diagram of a deep learning-based interference suppression device provided in the embodiments of this specification. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0045] In related technologies, reliable signal transmission is a core requirement in wireless communication systems. A typical communication link includes a transmitter, a transmission channel, and a receiver. The transmitter processes the original signal through encoding, modulation, and other methods to convert it into a form suitable for transmission. As the signal propagates in the channel, it is superimposed with additive white Gaussian noise and various intentional or unintentional interference signals. One of the core tasks of the receiver is to recover the original transmitted signal as much as possible from this mixed received signal, with interference suppression being a crucial step.

[0046] In related technologies, interference suppression methods rely on traditional algorithms based on deterministic signal processing models, such as suppression schemes based on energy detection combined with filter banks (band-stop filters) using fixed thresholds. These methods often pre-determine certain a priori characteristics of the interference (e.g., known fixed frequency bands, amplitude thresholds), and filter out signal components that match these pre-determined characteristics as interference. Such methods are effective in scenarios with simple and stable interference patterns.

[0047] However, in the complex electromagnetic environment of modern times, interference often exhibits multiple types, dynamic changes, and a high degree of overlap between its time-frequency characteristics and the useful signal. Related technologies face significant challenges in such scenarios: First, fixed detection thresholds and filter parameters struggle to match the dynamic changes in interference, leading to missed detections of low-power or mismatched interference, while false detections of useful signals with high instantaneous power. Second, filtering based on fixed rules lacks the ability to distinguish subtle differences between interference and the useful signal when suppressing components identified as interference, indiscriminately filtering out all energy within that frequency band or time slot, inevitably damaging the overlapped useful signal. These pain points can be summarized as follows: traditional methods, due to their fundamental limitations of relying on fixed models and prior knowledge, cannot adaptively learn and accurately separate the intrinsic characteristics of complex interference from the useful signal, resulting in inaccurate anti-interference performance and significant signal impairment.

[0048] The reason for this technical problem is that the processing logic of traditional algorithms is deterministic and non-learning. Their performance is limited by manually designed feature extraction rules and suppression criteria, making it unable to discover complex, high-dimensional interference patterns from the data, and lacking the ability to generalize to unknown interference. When the dynamics and complexity of the interference exceed the representation range of the preset model, the performance deteriorates sharply.

[0049] The inventive concept of this invention lies in replacing the interference suppression path of traditional fixed models with a data-driven deep learning model. Unlike deep learning models that rely on massive amounts of data (over-reliance on datasets leads to drastic performance degradation when the input scenario changes), this method trains on a limited training set (e.g., a fixed signal-to-noise ratio, multiple signal-to-interference ratio, and comb interference combination), enabling the model to adaptively learn the end-to-end mapping relationship from the interfered received signal to the clean signal. After training, the model can be deployed offline, demonstrating strong scene adaptability and generalization ability in practical applications. It can cope with dynamic changes in signal-to-noise ratio, signal-to-interference ratio, and interference type, thereby achieving more accurate interference suppression in complex and ever-changing interference environments and minimizing damage to useful signals.

[0050] This embodiment provides a deep learning-based interference suppression method. The execution entity of the method can be a server, such as a cloud computing server, which can be part of a cloud service to receive interference-containing signal data uploaded from terminal devices (e.g., base stations, IoT devices), perform interference suppression processing in the cloud, and then send the clean signal back to the terminal or perform subsequent analysis. The execution entity of the method can be a user terminal device, such as a smartphone, tablet, vehicle communication unit, and IoT terminal. The execution entity of the method can also be a network access device, such as a cellular communication base station, satellite ground station, and wireless access point. It can also be a hardware device in a communication system focused on signal processing, such as an embedded processing module or a digital signal processor. The embedded processing module can be an anti-interference processing board implemented based on FPGA (Field-Programmable Gate Array) or ASIC (Application-Specific Integrated Circuit).

[0051] like Figure 1 As shown, the method may include:

[0052] Step S12: Obtain the received signal; wherein the received signal is the original transmitted signal after interference and noise have been superimposed during transmission.

[0053] In this embodiment, the acquisition action can be represented as an execution entity (communication receiver, base station, terminal device or its internal processor) capturing electromagnetic waves from the physical channel through its hardware interface (e.g., antenna, radio frequency front end, analog-to-digital converter) and converting them into a discrete-time digital sequence that can be processed by a digital system.

[0054] In this embodiment, the original transmitted signal can be represented as a standard signal generated after the information intended to be transmitted by the transmitter has been encoded, modulated, and processed. Its specific form depends on the communication system adopted. For example, it can be a frequency-hopping signal using Turbo coding and QPSK modulation, or a direct-sequence spread spectrum signal using other coding and modulation methods, a single-carrier signal, etc. Interference signals can be represented as unwanted electromagnetic signals originating from outside the system, with the purpose of disrupting normal communication. Specifically, they can be: single-tone interference (a continuous wave of a single frequency), partial-band interference (noise occupying a portion of the signal bandwidth), multi-tone interference (a combination of multiple single-tone interferences), and more complex comb interference (multiple narrowband interferences distributed at intervals in the frequency domain), etc.

[0055] In this embodiment, the noise can be additive white Gaussian noise inherent in the channel, or it can include device thermal noise, etc.

[0056] Step S14: Preprocess the received signal to obtain multi-channel frequency domain feature data that meets the input format requirements of the pre-trained interference suppression network model.

[0057] In this embodiment, the purpose of the preprocessing is to convert the received signal in the time domain into a tensor format suitable for deep learning model processing. Specifically, this may include power normalization, windowing, Fast Fourier Transform (FFT) processing, and multi-channel feature construction. Specifically, power normalization is used to calculate the average power index (e.g., root mean square value) of the received signal over a preset time length (i.e., the network input window length, e.g., 8192 sampling points). Then, the amplitude of all sampling points of the received signal is scaled proportionally based on this average power index. The aim is to eliminate overall power fluctuations in the received signal caused by variations in transmission distance, channel fading, or different interference strengths, ensuring all input samples have a uniform energy scale, thereby accelerating the convergence process of neural network training and improving model stability. Windowing is used to apply a window function (e.g., Hamming window, Hanning window, etc.) to the power-normalized time-domain signal. The length of the window function is the same as the preset time length. The aim is to reduce spectral leakage caused by finite-length truncation of continuous signals, making the spectral profile clearer and the location of interfering frequencies easier to identify in subsequent frequency domain analysis. Fast Fourier Transform (FFT) processing is used to perform an FFT on the windowed time-domain signal, transforming it from the time domain to the frequency domain to obtain a complex sequence. This sequence represents the amplitude and phase information of the signal at different frequency components. Finally, the complex frequency domain sequence obtained from the FFT is separated into two independent data channels: one corresponding to the real part of the complex sequence, and the other corresponding to the imaginary part. The data from these two channels are organized according to a preset dimensional order (e.g., [sequence length, 1, 2]), thus forming the multi-channel frequency domain feature data. This multi-channel can be represented as including at least two channels: one for the real part and one for the imaginary part. This separation method transforms the amplitude and phase information of the complex signal into two-dimensional features that are easy for the network to learn.

[0058] Step S16: Input the multi-channel frequency domain feature data into the pre-trained interference suppression network model. The interference suppression network model performs forward inference calculations and outputs the corresponding multi-channel frequency domain feature data after interference suppression. The interference suppression network model is pre-trained using training data. The training data includes sample input data and corresponding sample label data. The sample input data is the frequency domain feature data of the signal carrying interference, and the sample label data is the corresponding frequency domain feature data of the signal without interference. The interference suppression network model learns the mapping relationship from the frequency domain features of the signal with interference to the frequency domain features of the signal without interference through training.

[0059] In this embodiment, the pre-trained interference suppression network model can be represented as a deep learning model whose parameters have been fixed through offline training optimization. It is used to learn and realize a complex nonlinear mapping from interference-containing frequency domain features to clean frequency domain features.

[0060] In this embodiment, the interference suppression network model can be obtained in advance using training data using a pedestal model. The pedestal model can be a network with an encoder-decoder structure and skip connections (e.g., U-Net and its variants).

[0061] The base model can also be an end-to-end mapping model based on a convolutional neural network, for example, a deep CNN composed of multiple residual blocks or densely connected blocks. The network receives multi-channel frequency domain feature data as input, performs a series of nonlinear transformations, and directly outputs interference-suppressed multi-channel frequency domain feature data.

[0062] In this embodiment, forward inference computation can be represented as taking the multi-channel frequency domain feature data obtained in step S14 as input, and calculating the output of each layer sequentially from front to back according to the inherent inter-layer connection relationship of the network model and the fixed parameters (weights and biases) until the final output layer result is obtained.

[0063] In one possible and specific implementation, the training data may consist of multiple pairs of samples, each sample including sample input data and its corresponding sample label data. The sample label data represents the ideal output that the network expects to learn, i.e., a signal free from external interference, consisting only of the original transmitted signal and the inherent noise of the channel. The sample label data can be generated in the following manner:

[0064] First, the original transmission signal is generated. Specifically, a digital baseband original transmission signal can be generated based on the set communication system parameters (e.g., Turbo coding, QPSK modulation, frequency hopping pattern, etc.).

[0065] Then, Gaussian white noise is superimposed. The original transmitted signal can be added to an additive white Gaussian noise sequence that meets the preset signal-to-noise ratio requirement to form a useful signal containing only noise. It should be noted that this noise is intentionally preserved in order to ensure that the model learns realistic signal characteristics that include channel noise, rather than a theoretically pure signal, thereby enhancing the model's robustness under different signal-to-noise ratio environments.

[0066] Next, the noisy but useful signal undergoes preprocessing operations identical to those performed on the received signal. These operations could include power normalization, windowing, Fast Fourier Transform (FFT), and tag feature construction, which involves separating the complex frequency domain representation into two channels: the real part and the imaginary part, thus obtaining the sample tag data (frequency domain feature form) corresponding to the sample.

[0067] Sample input data can be generated in the following ways:

[0068] First, the interference signal is synthesized. Specifically, to train the model to cope with complex and varied interference, this implementation scheme can use comb interference as the interference source for data preparation. Comb interference can effectively simulate the characteristics of single-tone, multi-tone, and some band interference. More specifically, the comb interference signal can be synthesized as follows: 1. Determine the number of comb teeth N in the sample (e.g., randomly selected between 1 and the maximum number of frequency points); 2. Generate N independent band-limited noise sequences. The bandwidth of each band-limited noise can be controlled by a randomly generated bandwidth factor, which can be randomly selected within a preset range to simulate interference components of different widths. 3. For the i-th band-limited noise, specify a carrier frequency f(i) and an initial phase θ(i). The carrier frequency is randomly selected within the signal bandwidth, and the initial phase is randomly selected within [0, 2π). 4. Modulate (multiply) each band-limited noise with a complex sinusoidal carrier with the corresponding carrier frequency f(i) and initial phase θ(i) to generate sub-interference components. 5. Superimpose all N sub-interference components to form a complete comb interference signal. 6. Adjust the power of the comb-shaped interference signal according to the preset signal-to-interference ratio requirement, and then add it to the above-mentioned noisy useful signal to finally obtain the time-domain simulation of the received signal containing interference.

[0069] Then, preprocessing is also required, that is, preprocessing the simulated received signal containing interference. The steps can be exactly the same as the real-time processing of the received signal (step S14), which can be power normalization, windowing and FFT processing, and input feature construction. Separate the real and imaginary parts of the FFT result to form the sample input data (frequency domain feature form) of the sample.

[0070] In this implementation scheme, the above method can be used to generate tens of thousands of data pairs (sample input data, sample label data) in batches, forming the dataset required for model training.

[0071] In one possible and specific implementation, the interference suppression network model can be trained using supervised learning in the following manner:

[0072] First, the prepared dataset is divided into training and validation sets. Then, a deep neural network with an encoder-decoder structure is initialized. Training parameters are set, including the optimizer, initial learning rate, batch size, and total number of training epochs. Specifically, network training uses the Keras framework in Python 3.7, network parameter initialization uses Xavier, training uses the Adam optimizer, the initial learning rate is set to 0.0006, the minimum learning rate is 0.0000001, the batch size is 65, and the total number of training epochs is 30. The learning rate is halved when the validation set loss no longer decreases after 5 consecutive epochs. Mean squared error (MSE) is used as the loss function, defined as the mean squared error between the frequency domain feature data of the network output and the sample label data. Training is terminated early when the validation set loss no longer decreases after 10 consecutive epochs, and the model parameters with the minimum validation set loss are saved.

[0073] Then, the training loop is executed. Specifically, each training iteration may include: 1. Taking a batch of sample input data from the training set. 2. Inputting the sample input data into the interference suppression network model to be trained; the model performs forward propagation calculations to obtain the predicted output feature data. 3. Calculating the difference between the predicted output and the corresponding batch of sample label data, and quantifying this difference using the mean squared error loss function. This loss function measures the overall deviation between the predicted clean signal features and the true clean signal features by calculating the average of the squares of the differences between the predicted and true values. 4. Calculating the gradient of the loss value with respect to the parameters of each layer of the network model using the backpropagation algorithm. 5. Using the optimizer to update all parameters of the network model based on the gradients to reduce the loss value.

[0074] Through iterative processes, the parameters within the network model are gradually adjusted and optimized. The training objective is to minimize the mean squared error between the predicted output and the true label, thereby internalizing the complex mapping between frequency domain features containing interference and those without. When the loss value on the independent validation set no longer decreases significantly, and its interference suppression performance meets predetermined requirements, iterative training can be stopped, resulting in a pre-trained interference suppression network model.

[0075] Step S18: Post-process the multi-channel frequency domain feature data after interference suppression to obtain the time domain signal after interference suppression, which is used for subsequent communication demodulation.

[0076] In this embodiment, the post-processing can be described as performing an operation that corresponds to but is reversed in direction to the pre-processing in step S14, used to restore the data output by the neural network, which is still in the frequency domain representation, to a time domain waveform that can be directly processed by the communication demodulation module (e.g., demapping, decoder). Specifically, firstly, channel synthesis can be performed. That is, the multi-channel frequency domain feature data output by the neural network is recombined into a complete complex frequency domain representation. For example, when the input is a dual-channel system with real and imaginary parts, the subsequent processing is to recombine the data from the two channels into a complex sequence, where the real part comes from the first channel and the imaginary part comes from the second channel.

[0077] Then, a frequency-time transformation is performed. That is, the synthesized complex frequency domain signal is converted back to the time domain. It is the inverse process of time-frequency transformation. For example, an inverse fast Fourier transform can be performed. This transform can restore the frequency domain complex sequence to the time domain complex sequence.

[0078] Finally, scaling, real part extraction, or overlapping and addition operations can be performed (for example, overlapping framing strategies are also used in preprocessing) to finally obtain the time-domain discrete signal after interference suppression. This signal can then be sent to the subsequent communication demodulation module for processing. This communication demodulation module can be pre-programmed with a series of communication receiver processing flows such as digital demodulation, channel decoding, and source decoding to ultimately recover the transmitted information bits.

[0079] The method provided by this invention effectively overcomes the limitations of traditional methods that rely on fixed rules by employing a pre-trained deep learning model to perform end-to-end interference suppression. This model, through learning from limited data, adaptively grasps the subtle differences and mapping relationships in the frequency domain characteristics of complex interference signals and useful signals. Therefore, during the inference stage, this method can achieve more accurate identification and suppression of dynamically changing and diverse interference types, while preserving the integrity of useful signals to the greatest extent possible. It significantly improves the robustness and anti-interference capability of communication systems in complex interference scenarios, reduces the bit error rate caused by inaccurate interference suppression, and thus ensures the performance and reliability of subsequent communication demodulation stages.

[0080] In some embodiments, the step of preprocessing the received signal to obtain multi-channel frequency domain feature data that conforms to the input format requirements of a pre-trained interference suppression network model includes:

[0081] Step S142: Perform power normalization processing on the received signal to obtain the normalized received signal.

[0082] In this embodiment, the average power characteristic of the received signal segment over a preset time period (which can be matched with the window length and network input dimension) can be calculated. This average power characteristic can be a concentrated representation of the signal energy within that time period; specifically, it can be the root mean square value of the amplitudes of all sampling points in the signal segment, or the average absolute amplitude, or other statistical quantities that can reflect the overall signal strength. Then, the amplitude value of each sampling point in the received signal can be divided by the calculated average power characteristic. This operation is equivalent to proportionally scaling the amplitude of the entire signal segment.

[0083] Understandably, after this processing step, the average power of the normalized received signal is adjusted to a uniform reference level. This significantly reduces the negative impact on the training stability of deep neural networks caused by excessive differences in input signal strength, helping to accelerate model convergence and improve its generalization performance under various signal reception intensities.

[0084] Step S144: Window the normalized received signal to obtain the windowed received signal.

[0085] In this embodiment, a preset window function can be multiplied point-by-point with the normalized received signal obtained in step S142. The window function can be a sequence of the same length as the signal segment, which smoothly decays to zero or near zero at both ends of the segment. Specific window functions can include: Hamming window, Hanning window, and Blackman window, etc. In one specific embodiment, a Hamming window can be used. This operation is performed directly in the time domain, and the output is the windowed received signal. The windowing operation weakens the discontinuity at the signal truncation point, so that the mutual interference (i.e., sidelobe level) between frequency components in the frequency domain (through subsequent FFT) representation of this signal segment can be effectively suppressed.

[0086] Step S146: Perform Fast Fourier Transform on the windowed received signal to obtain the corresponding frequency domain representation.

[0087] In this embodiment, a Fast Fourier Transform (FFT) can be performed on the windowed received signal obtained in step S144. This operation determines the number of transform points (which can be the same as the window length, for example, 8192 points) based on the signal sampling rate and segment length. In specific calculations, to maintain signal energy conservation before and after the transform, the time-domain signal can be preprocessed with a scaling factor (e.g., divided by the square root of the number of transform points) before the FFT. The output of the FFT can be a complex sequence with a length equal to the number of transform points. This complex sequence is the frequency domain representation of the signal, where each complex number corresponds to a frequency component, its magnitude representing the amplitude of that frequency, and its argument representing the phase of that frequency.

[0088] Understandably, this transformation converts a signal from a waveform with time as the independent variable into a spectrum with frequency as the independent variable. Interference signals (e.g., single-tone interference represented by a single spectral line, partial-band interference by a continuous spectral block, and comb interference by multiple discrete spectral lines) can be represented in this format by their frequency position and distribution, while useful signals (e.g., broadband communication signals) appear as relatively flat or wide spectral bands with a specific shape. This representation greatly facilitates the subsequent learning and separation of interference patterns by deep learning models.

[0089] Step S148: Separate the frequency domain representation into real and imaginary channels to obtain multi-channel frequency domain feature data.

[0090] In this embodiment, the complex frequency domain sequence obtained in step S146 can be split into two independent real number sequences. The first sequence can be composed of the real part of each complex number in the original complex number sequence. The second sequence can be composed of the imaginary part of each complex number. Then, the two real number sequences can be organized according to a preset dimensional order. For example, they can be constructed as a three-dimensional tensor with dimensions [L, 1, 2], where L represents the number of frequency points (sequence length), 1 can be a placeholder for the spatial dimension, and 2 represents the number of channels, i.e., the first channel is the real part sequence and the second channel is the imaginary part sequence.

[0091] In this embodiment, the real and imaginary components can be extracted from the complex frequency domain representation and organized into multiple data channels, wherein at least two channels correspond to the real and imaginary components.

[0092] Specifically, the real part values ​​of each complex element in the complex frequency domain sequence can be directly extracted and arranged in their original order to form a real part sequence. Then, the imaginary part values ​​of each complex element are extracted simultaneously to form an imaginary part sequence. These two sequences are then organized into multi-channel data. For example, a two-dimensional tensor of shape [L, 2] can be constructed, where the first column (or the first channel dimension) is the real part sequence and the second column is the imaginary part sequence. Alternatively, to accommodate convolutional networks that require a three-dimensional tensor input, it can be constructed in the form of [L, 1, 2], where the last dimension 2 represents the channel, channel 0 is the real part, and channel 1 is the imaginary part.

[0093] Specifically, normalization can also be performed after separation. First, the original real part sequence and imaginary part sequence are obtained using the method described above. Then, the two real number sequences are subjected to independent numerical normalization. For example, maximum value normalization (dividing each element in the sequence by the absolute maximum value of the sequence) or standardization (subtracting the sequence mean and then dividing by the sequence standard deviation) can be performed on each sequence. After processing, the normalized real part sequence and imaginary part sequence are organized as two independent channels. In some implementations, additional channels derived from the real part and imaginary part channels can be introduced to provide supplementary information while retaining the real part and imaginary part channels. More specifically, the real part sequence and imaginary part sequence are extracted as the two basic channels. At the same time, the amplitude sequence at each frequency point is calculated based on the real part and imaginary part (e.g., the magnitude is calculated). Then, the original real part sequence, the original imaginary part sequence, and the calculated amplitude sequence are organized as three channels. In some implementations, in addition to the basic real and imaginary parts, to provide more symmetrical information, the following channels can be organized: the first channel is the original real part sequence, the second channel is the original imaginary part sequence, the third channel can be the real part sequence of the conjugate of the complex number sequence (which is actually equivalent to the original real part sequence), and the fourth channel is the imaginary part sequence of its conjugate (i.e., the negative of the original imaginary part sequence).

[0094] In some embodiments, the step of performing power normalization processing on the received signal to obtain a normalized received signal includes:

[0095] Step S1422: Calculate the average power characteristic of the received signal within a preset time period.

[0096] In this embodiment, the preset time length can be a basic time unit set for signal processing, which can correspond to the length of a processing frame. This length can be determined by system design parameters, for example, it can be 8192 sampling points, and its specific value can accommodate the typical characteristic period of the signal and interference.

[0097] In this embodiment, the average power representation can be the root mean square (RMS) value. This is achieved by first calculating the squares of the amplitude values ​​at all sampling points of the signal segment, then averaging these squares, and finally taking the square root of the average. The result is the RMS value of the signal, which physically represents the effective amplitude of the signal.

[0098] In this embodiment, the average power characteristic can be the average absolute amplitude, that is, the arithmetic mean of the absolute amplitude values ​​of all sampling points of the signal segment.

[0099] In this embodiment, the average power characteristic can be the mean square value of the signal. The arithmetic mean of the squares of the amplitudes of all sampling points of the signal segment can be calculated; this value directly corresponds to the average energy of the signal segment.

[0100] Step S1424: Adjust the amplitude of the received signal based on the average power so that the adjusted received signal has a uniform energy scale, thereby obtaining a normalized received signal.

[0101] In this embodiment, this step is used to scale the entire signal segment proportionally using the average power representation obtained in the previous step, thereby eliminating the absolute energy difference between different signal segments. Specifically, the amplitude value of each sampling point of the received signal within a preset time length can be divided by the average power representation calculated in step S1422. That is, the same division operation is performed on each sampling point in the signal segment. After this operation, the original sampling point sequence is converted into a new sequence, namely the normalized received signal. It can be understood that through the above operation, the overall energy level of the output signal sequence is adjusted to a uniform benchmark. For example, if the root mean square value is calculated in step S1422, the root mean square value of the normalized signal sequence will become the value one. This ensures that all processed signal segments, regardless of their original intensity, can have a comparable and consistent energy scale when input to the subsequent processing module.

[0102] In some embodiments, the step of windowing the normalized received signal to obtain a windowed received signal includes:

[0103] Step S1442: Apply a Hamming window to the normalized received signal to obtain the windowed received signal.

[0104] In this embodiment, the Hamming window is a cosine-type window function whose time-domain waveform smoothly decays to near zero at both ends. A sequence of Hamming windows of the same length can be generated based on the preset time length (i.e., the length of the signal processing window, for example, 8192 sampling points). Each value in this sequence represents a weighting coefficient at the corresponding sampling point position.

[0105] The generated Hamming window sequence is then multiplied point-by-point with the normalized received signal sequence obtained in step S142. That is, the values ​​at the same position in two sequences of the same length are multiplied to obtain a new sequence. This new sequence obtained after point-by-point multiplication is the windowed received signal. This signal retains the main characteristics of the original signal, but its amplitude at the edges of the time window has been significantly attenuated.

[0106] Understandably, directly truncating a signal introduces spurious frequency components (spectral leakage) in the frequency domain. The Hamming window, by smoothly attenuating signal jumps at the truncation point, can effectively reduce the intensity of these spurious components, making the true spectral profile of the signal (especially the spikes of narrowband interference) more prominent and isolated.

[0107] In some implementations, the interference suppression network model includes an encoder path and a decoder path; wherein the encoder path is used to downsample the input feature data to extract multi-scale features; the decoder path is used to upsample the feature data output by the encoder path to restore signal resolution; wherein there is a skip connection between the encoder path and the decoder path for passing the features extracted by the encoder path to the corresponding layer of the decoder path.

[0108] In this embodiment, the encoder path is used to perform hierarchical feature extraction and condensation on the input multi-channel frequency domain feature data. This path gradually reduces the spatial dimension (i.e., frequency resolution) of the data while increasing the number of feature channels, thereby extracting and condensing multi-scale abstract features of the signal, especially the essential features of interference modes, within a larger receptive field.

[0109] Specifically, the encoder path can be composed of multiple sequentially connected cascaded encoding stages. Each encoding stage may include one or more basic computational units. For example, an encoding stage may consist of two consecutive convolutional layers, each optionally followed by a non-linear activation function (e.g., ReLU or Tanh), and then connected to a downsampling layer. The downsampling layer is used to reduce the spatial size of the feature map, and it can employ convolution operations with a stride greater than 1 (i.e., strided convolution) or max pooling or average pooling operations. As data is passed down the encoder path, the width (corresponding to the number of frequency points) and height of the feature map can be halved at each level, while the number of feature channels can be multiplied at each level. In this way, the encoder path ultimately yields a highly condensed feature representation at the output, with low spatial resolution but rich in high-level semantic information.

[0110] In this embodiment, the decoder path is used to perform hierarchical resolution restoration and detail reconstruction of the features output by the encoder path. This path can upsample and decode the abstract features extracted by the encoder by gradually increasing the spatial dimension of the data and reducing the number of feature channels, so as to reconstruct a frequency domain feature map with the same resolution as the original input and suppressing interference.

[0111] Specifically, the decoder path can also be composed of multiple sequentially connected cascaded decoding stages, and the number of decoding stages typically corresponds to the number of encoding stages. Each decoding stage may include an upsampling operation, followed by one or more convolutional layers. The upsampling layers are used to increase the spatial size of the feature map, and may specifically include: transposed convolution (deconvolution), nearest neighbor interpolation upsampling, or bilinear interpolation upsampling. After upsampling, the resolution of the feature map is improved.

[0112] In this embodiment, the skip connection is a component connecting the encoder path and the decoder path. That is, it establishes a short-circuit connection from a layer in the encoder path to a corresponding layer in the decoder path (of the same depth or similar resolution), used to directly transfer and fuse the feature map extracted by the encoder at that layer, which includes more spatial details and low-frequency information, into the corresponding layer of the decoder path. Specifically, at each decoding stage of the decoder path, before performing convolution operations, the feature map transmitted from the corresponding stage of the encoder path through the skip connection is concatenated with the feature map obtained by upsampling in the previous stage of the decoder path. This concatenation can be performed along the dimension of the feature channel. Through the skip connection, the decoder can simultaneously utilize the shallow features containing rich details from the encoder path and the high-level semantic features that have undergone deep abstraction when reconstructing high-resolution output. This effectively compensates for the details that may be lost during downsampling, especially the subtle boundary features of interference and useful signals in the frequency domain, thereby significantly improving the accuracy of the reconstructed signal, especially the frequency domain contour and phase.

[0113] In some implementations, the interference signal included in the signal corresponding to the sample input data is a comb interference signal; the comb interference signal is generated by superimposing multiple independent interference components; wherein each interference component is generated by band-limited noise signal modulation, and each interference component corresponds to a preset center frequency and a preset phase offset.

[0114] In this embodiment, the comb-shaped interference signal is a single type of interference generated by superimposing multiple independent interference components in the time domain. Its design aims to effectively cover various typical frequency domain interference patterns, such as single-tone, multi-tone, and narrowband noise, using this single interference type. Using this method for training allows the deep learning model to learn more general and robust interference features while significantly reducing the amount of data samples required for training, thereby improving the model's generalization ability when facing unknown or dynamically changing interference scenarios.

[0115] Specifically, the comb-like interference signal can be generated by superimposing N independent interference components, where N is an integer greater than or equal to 1, representing the number of teeth in the comb-like interference. Each interference component can be represented as a band-limited signal modulated by a carrier wave.

[0116] The band-limited noise signal is the baseband portion of each interference component. Specifically, it can be a band-limited noise signal, that is, random noise whose frequency components are confined to a certain bandwidth. It can be generated by passing Gaussian white noise through a bandpass filter, thus obtaining a noise sequence whose energy is mainly concentrated within a preset bandwidth. This band-limited noise is used to simulate the random fluctuation characteristics of the interference signal within a single comb-tooth frequency band. Generating the interference component mainly involves modulating the aforementioned band-limited noise. Specifically, the band-limited noise signal is multiplied point-by-point by a complex sinusoidal carrier. This sinusoidal carrier is uniquely determined by two preset parameters: a preset center frequency and a preset phase offset. More specifically, the frequency value of the preset center frequency determines the center position of the interference component in the final synthesized spectrum. When preparing different samples, different center frequencies can be randomly or according to certain rules for different interference components, so that they are distributed at different positions within the target communication frequency band.

[0117] The preset phase offset determines the initial phase of the sinusoidal carrier at the start time. Setting random and independent phase offsets for different interference components makes the synthesized comb interference signal waveform more random and uncertain, more realistically simulating the characteristics of an unknown interference source. Through the above modulation, the spectrum of the original band-limited noise is shifted to a region centered on the preset center frequency, thus forming a narrowband interference sub-signal centered on that frequency in the frequency domain. Its time-domain waveform is the result of the envelope of the band-limited noise being modulated by the sinusoidal waveform corresponding to the center frequency.

[0118] After generating multiple (e.g., N) independent interference components in parallel, a complete comb-shaped interference signal can be synthesized by superposition (summation). That is, the sampling values ​​of all interference components at the same time are added together to obtain the final composite interference waveform.

[0119] In one specific implementation, when constructing the training dataset, the number of comb teeth (i.e., the number of interference components N), the preset center frequency value of each component, the preset phase offset value, and the relative power of each component's bandwidth (which can be achieved by adjusting the gain of the band-limited noise, corresponding to different signal-to-interference ratio scenarios) can all be flexibly varied in the samples. For example, the center frequency can be randomly selected within the communication bandwidth, and the phase offset can be randomly generated between 0 and 2π zero radians.

[0120] In some embodiments, the step of post-processing the multi-channel frequency domain feature data after interference suppression to obtain the time domain signal after interference suppression for subsequent communication demodulation includes:

[0121] Step S182: Combine the real and imaginary channels of the multi-channel frequency domain feature data after interference suppression to form a complex frequency domain signal.

[0122] In this embodiment, this step is the inverse process corresponding to the channel separation operation (step S148) in the preprocessing process. Its purpose is to recombine the frequency domain information output by the neural network, which is distributed in two independent real number channels, and restore it to a complete complex number representation in order to recover the complete mathematical description of the frequency domain signal.

[0123] In one specific implementation, the multi-channel frequency domain feature data after interference suppression can be a data structure comprising at least two channels, for example, a three-dimensional tensor of dimension [L, 1, 2], where L is the number of frequency points. The first channel (e.g., the channel with index 0) stores the sequence of real part values ​​of the frequency domain signal, and the second channel (e.g., the channel with index 1) stores the sequence of imaginary part values ​​of the frequency domain signal. The merging operation can be performed by constructing a new complex number by taking each value in the real part channel as the real part and the corresponding value in the imaginary part channel as the imaginary part, in a one-to-one correspondence order.

[0124] Step S184: Perform inverse fast Fourier transform on the complex frequency domain signal to obtain the time domain signal after interference suppression.

[0125] In this embodiment, this step may be to perform an inverse transform corresponding to the fast Fourier transform (step S146) in the preprocessing process. The purpose is to convert the frequency domain signal, which has been processed by interference suppression and merged into a complex form, back from the frequency domain to the time domain, thereby generating a time domain waveform that can be directly processed by subsequent modules of the communication receiver chain (e.g., demodulator, decoder).

[0126] In one specific implementation, a deep learning-based interference suppression method is provided. This method can learn online via a computer program and be used offline after training. Besides being directly invoked in application software, it can also be integrated and deployed in data servers, cloud computing servers, mobile terminals, and embedded devices such as FPGAs and DSPs.

[0127] A deep learning-based interference suppression method includes:

[0128] S100: Build an interference suppression network model.

[0129] S101: The deep learning network adopts the UNet network structure, consisting of an encoder and a decoder. The encoder mainly consists of convolutional layers and pooling layers, which are downsampling paths used to extract input features and reduce data resolution, thereby obtaining a larger field of view. The decoder mainly consists of convolutional layers and upsampling layers, restoring the original resolution of the data. Skip connections connect corresponding layers in the encoder and decoder, further reproducing details.

[0130] In one possible and specific implementation, the deep learning-based interference suppression network model includes four pooling convolutional blocks, three fusion convolutional blocks, five convolutional blocks, one upsampling layer, and a connection layer. Specifically, each pooling convolutional block consists of one average pooling layer and two 3×1 convolutional layers; the fusion convolutional block consists of one 3×1 deconvolutional layer, three 3×1 convolutional layers, and one upsampling layer; and each convolutional block consists of one 3×1 convolutional layer, a dropout layer, and a Tanh activation function. The network input dimension is N×1×2, where N is the input signal length and 2 is the number of feature channels (real and imaginary parts). The added dimension accommodates the network input layer dimension.

[0131] In one possible and specific implementation, the input data is first passed through two 1x1 convolutional blocks to increase the number of channels, introducing more non-linear representations. Then, cascaded pooling convolutional blocks are used, doubling the number of channels with each downsampling (feature size decreases, but features become more diverse). After four pooling convolutional blocks, upsampling convolutions are performed and cascaded fusion convolutional blocks are used. At the same time, the outputs of the corresponding pooling convolutional blocks are connected for feature fusion. Finally, the output dimension is transformed back to the input dimension through two convolutional blocks for output.

[0132] like Figure 5 As shown, in one possible and specific implementation, the deep learning-based interference suppression network model may include:

[0133] The input layer receives multi-channel frequency domain feature data of shape [N, 1, 2], where N is the length of the frequency domain sequence and 2 represents the real and imaginary feature channels.

[0134] In the initial feature extraction and transformation layer, the input data first passes through two convolutional layers with a kernel size of 1×1 for preliminary feature enhancement and channel number adjustment to improve the non-linear expressive ability of the model.

[0135] The encoder path (downsampling path) begins with the data passing through three cascaded pooling convolutional blocks. Each pooling convolutional block performs a downsampling operation to reduce the spatial resolution of the feature map, while simultaneously increasing the number of channels through convolution, thereby extracting multi-scale abstract features within a progressively expanding receptive field. The data then continues through more advanced pooling convolutional blocks for deeper feature compression and encoding, resulting in a highly condensed low-level feature representation.

[0136] The bottleneck transition layer (UpSample+Conv) first improves the spatial resolution of the features output from the encoder by upsampling (UpSample), and then adjusts and integrates the features through a convolutional layer (Conv) to provide input for the decoder path.

[0137] The decoder path (upsampling path) consists of three cascaded fused convolutional blocks. Each fused convolutional block can include upsampling operations and convolutional layers. Crucially, each fused convolutional block is concatenated with the feature maps of the corresponding layers in the encoder path via skip connections. This allows the decoder to fuse the detailed information retained in the encoder stage when reconstructing the high-resolution output, accurately restoring the frequency domain structure and contour of the signal.

[0138] The output reconstruction layer, at the end of the decoder, uses two convolutional layers to perform final integration and channel adjustment of the features, transforming the output dimension back to the [N, 1, 2] format consistent with the input, thus obtaining the multi-channel frequency domain feature data after suppressing interference.

[0139] S102: Design the objective function of the interference suppression network model. The network input is the interfered signal, and the desired output is the undisturbed signal. To make the interference-free data generated by the network more closely match the desired output, a mean squared error loss function is used, which is:

[0140] ;

[0141] In the formula, This represents the value of the loss function, where G is the entire interference suppression network model. This value quantifies the gap between the network's current performance and its ideal state, and is the objective to be minimized during model training. N is the total number of data points involved in a single loss calculation. It corresponds to the sum of the dimensions of all samples within a training batch. This represents the actual output or predicted value of the interference suppression network model for the i-th data point. This represents the expected output value or the actual value of the i-th data point corresponding to the network input.

[0142] S200: Received signals with different types and intensities of superimposed interference are transformed through data preprocessing to adapt to the network input, serving as the dataset for deep learning network training. For details, please refer to [link to relevant documentation]. Figure 2 .

[0143] S201: Perform data preprocessing on network input.

[0144] Suppose the signal we receive is:

[0145] ;

[0146] In the formula, This represents the amplitude value of the received signal at the discrete-time index point m. This represents the amplitude value of the original transmitted signal at the discrete-time index m. This represents the amplitude value of the Gaussian white noise at the discrete-time index point m. This represents the amplitude of the interference signal at the discrete-time index m. m is the discrete-time index.

[0147] Comb noise interference was used when constructing the dataset. The mathematical model is as follows:

[0148] ;

[0149] In the formula, The number of comb teeth. This represents band-limited noise, as well as Gaussian white noise. different, The spectrum is confined to a narrower frequency band, so that each interference component i is not a pure tone of a single frequency, but a noise band with width.

[0150] It is a complex exponential function in complex form, representing the frequency as... The initial phase is The complex sinusoidal carrier wave, where It is the imaginary unit. Let be the center frequency of the i-th interference component (i.e., the i-th comb tooth). Each interference component i can be preset with a different frequency. This allows the synthesized comb interference to cover multiple discrete frequency points or narrow bands in the frequency domain. This represents the initial phase offset of the i-th interference component.

[0151] To ensure the network is not affected by interference signals of varying intensities during training and to accelerate the network's learning speed, Perform power normalization. Input length for the network:

[0152] ;

[0153] In the formula, This represents the amplitude value of the received signal at time index m after power normalization. This represents the amplitude value of the original received signal at time index m. This is the network input length, which is also the signal frame length.

[0154] The network input length is limited, necessitating truncation of the received signal. To reduce spectral leakage caused by time-domain truncation, it is necessary to... Add windows, using Hanming windows, with a length of [missing information]. After windowing, the signal is There is no need for overlapping windowing here to compensate for the signal-to-noise ratio loss caused by single-path windowing.

[0155] To address frequency domain interference, the received signal needs to be transformed to the frequency domain and a normalized FFT operation performed.

[0156] ;

[0157] In the formula, This represents the set of multi-channel frequency domain feature data finally obtained from the preprocessing steps. This represents the sequence of the real part of the processed frequency domain signal, which constitutes the first channel of the neural network's input feature data. This represents the imaginary part sequence of the processed frequency domain signal, which constitutes the second channel of the neural network input feature data. For the real part operator, This is the operator for retrieving the imaginary part. For Fast Fourier Transform. This represents the received signal after windowing at the discrete-time index m.

[0158] S202: Perform data preprocessing on network labels.

[0159] The network expects to output the following signal:

[0160] ;

[0161] In the formula, This represents the network's expected output signal at the discrete-time index point m, which is the label signal of the training data. This represents the original transmitted signal at the discrete-time index point m. This represents Gaussian white noise superimposed at the discrete-time index m. Its purpose is to increase the robustness of the network. By allowing the model to learn features that include noise, it can prevent the model from over-cleaning and attempting to eliminate the inherent and necessary background noise in the channel, thus making it more adaptable to different signal-to-noise ratio environments.

[0162] The presence of noise is intended to increase the robustness of the network, and the network simultaneously learns noise characteristics to adapt to different signal-to-noise ratios.

[0163] right Perform maximum value normalization:

[0164] ;

[0165] In the formula, This represents the amplitude value of the tag signal at time index m after maximum value normalization.

[0166] You can obtain network tags:

[0167] ;

[0168] In the formula, This is the final generated multi-channel frequency domain feature data for network labels used in training. The first channel of the tag's frequency domain feature data is composed of the real part of a complex frequency domain signal. The second channel for the tag's frequency domain feature data is composed of the imaginary part of the complex frequency domain signal.

[0169] S203: Construct the network training dataset.

[0170] For the frequency domain interference dataset, comb interference combines the characteristics of tone interference with narrowband and partial band interference. Therefore, comb interference is used in the dataset to effectively balance the network's complexity, training efficiency, and robustness while ensuring that the network can cope with multiple types of interference. The dataset is generated under single signal-to-noise ratio and multiple signal-to-interference ratio conditions. The number of comb teeth and bandwidth of the comb interference are set according to the communication system.

[0171] S204: Training the interference suppression network.

[0172] Prepare a training dataset, using the received signal containing interference as the network input and the signal without interference as the network label, and train the network to obtain an interference suppression network model that can be used offline.

[0173] S300: Build an anti-interference communication system model based on a deep learning network. Replace the traditional interference suppression algorithm module with the trained interference suppression network, and it can be used offline. Pay attention to dimensionality transformation and matching it with the network input dimension.

[0174] S301: The present invention will be further described below with reference to the embodiments. The embodiments adopt a frequency hopping communication system, but the present invention is not limited to a specific communication system.

[0175] S302: The specific embodiments of the present invention are described below.

[0176] The communication system architecture adopts a frequency hopping system, with Turbo coding, QPSK modulation, 51 frequency hopping points, a signal bandwidth of 5MHz, a sampling bandwidth of 160MHz, and a Hamming window length of 8192. The simulation of the embodiment uses MATLAB language.

[0177] The deep learning framework used is Keras in Python 3.7, with an input dimension of 8192×1×2. The training data consists of signals after coded modulation and frequency hopping, followed by the superposition of noise and interference. The training labels are the signals after coded modulation and frequency hopping. The comb interference parameters in the training dataset are as follows: the number of comb teeth is randomly generated between 1 and 51, the location of the interference frequency is randomly generated, and the interference bandwidth factor is randomly generated between 0.1 and 0.5. The signal-to-interference ratio (SIR) of the training dataset is [-100:10:20] dB, and the signal-to-noise ratio (SNR) is fixed at 10 dB.

[0178] S303: Presents test results under the most complex interference scenario. Specifically, in dynamic interference scenarios, the performance of the interference suppression network at the same signal-to-noise ratio (SNR) but different SNRs is tested, and compared with traditional interference suppression algorithms based on overlapping windowing. The test results are as follows: Figure 3 As shown, under dynamic interference with periodic single-tone interference, partial-band interference, comb interference, and multi-tone interference switching in the middle of the data segment, the proposed scheme can resist interference strength significantly better than traditional interference suppression algorithms based on overlapping windowing at the same bit error rate. This is achieved with a signal-to-noise ratio of 11 dB and a bit error rate of [missing information]. At that time, the anti-interference performance can be improved by about 30dB.

[0179] S304: In dynamic interference scenarios, the source bit error rate performance of the interference suppression network is tested under the same signal-to-interference ratio (SIR) but different SNRs. The results are compared with those of a traditional interference suppression algorithm based on overlapping windowing. Figure 4 As shown, under dynamic interference with periodic single-tone interference, partial band interference, comb interference, and multi-tone interference switching in the middle of the data segment, the proposed scheme of this invention exhibits better demodulation backoff performance than traditional algorithms at the same bit error rate. This is achieved with a signal-to-interference ratio of -50dB and a bit error rate of [missing information]. At this time, the demodulation back-off performance can be improved by about 0.6dB. With the increase of interference intensity, the demodulation back-off performance can be further improved.

[0180] like Figure 6 As shown, this embodiment provides an interference suppression device based on deep learning, including: a memory, a processor, and a computer program. The computer program trains a network model in the processor and stores it in the memory. The processor directly calls and executes the trained network model to achieve interference suppression.

[0181] like Figure 7 As shown, this embodiment provides an interference suppression device based on deep learning, the device comprising:

[0182] An acquisition module is used to acquire a received signal; wherein the received signal is the original transmitted signal after interference and noise have been superimposed during transmission.

[0183] The preprocessing module is used to preprocess the received signal to obtain multi-channel frequency domain feature data that meets the input format requirements of the pre-trained interference suppression network model.

[0184] The output module is used to input multi-channel frequency domain feature data into a pre-trained interference suppression network model. The interference suppression network model performs forward inference calculations and outputs the corresponding multi-channel frequency domain feature data after interference suppression. The interference suppression network model is pre-trained using training data, which includes sample input data and corresponding sample label data. The sample input data is the frequency domain feature data of the signal carrying interference, and the sample label data is the corresponding frequency domain feature data of the signal without interference. The interference suppression network model, through training, learns the mapping relationship from the frequency domain features of the signal with interference to the frequency domain features of the signal without interference.

[0185] The post-processing module is used to post-process the multi-channel frequency domain feature data after interference suppression to obtain the time domain signal after interference suppression, which is used for subsequent communication demodulation.

[0186] According to embodiments of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a computer, causes the computer to perform the method described in any of the above embodiments.

[0187] According to embodiments of the present invention, a computer program product comprising instructions is also provided, which, when executed by a computer, cause the computer to perform a method in any of the above embodiments.

[0188] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based interference suppression method, characterized in that, The method includes: Acquire the received signal; wherein the received signal is the original transmitted signal after interference and noise have been superimposed during transmission; The received signal is preprocessed to obtain multi-channel frequency domain feature data that meets the input format requirements of the pre-trained interference suppression network model; Multi-channel frequency domain feature data is input into a pre-trained interference suppression network model. The interference suppression network model performs forward inference calculations and outputs the corresponding multi-channel frequency domain feature data after interference suppression. The interference suppression network model is pre-trained using training data, which includes sample input data and corresponding sample label data. The sample input data represents the frequency domain feature data of the signal carrying interference, and the sample label data represents the corresponding frequency domain feature data of the signal without interference. The interference suppression network model, through training, learns the mapping relationship from the frequency domain features of the signal with interference to the frequency domain features of the signal without interference. The multi-channel frequency domain feature data after interference suppression is post-processed to obtain the time domain signal after interference suppression, which is used for subsequent communication demodulation.

2. The method according to claim 1, characterized in that, The step of preprocessing the received signal to obtain multi-channel frequency domain feature data that conforms to the input format requirements of the pre-trained interference suppression network model includes: The received signal is subjected to power normalization processing to obtain a normalized received signal; The normalized received signal is windowed to obtain the windowed received signal. The windowed received signal is processed by Fast Fourier Transform to obtain the corresponding frequency domain representation; The frequency domain representation is separated into real and imaginary channels to obtain multi-channel frequency domain feature data.

3. The method according to claim 2, characterized in that, The step of performing power normalization processing on the received signal to obtain a normalized received signal includes: Calculate the average power characteristic of the received signal over a preset time period; The amplitude of the received signal is adjusted based on the average power characteristic to give the adjusted received signal a uniform energy scale, resulting in a normalized received signal.

4. The method according to claim 2, characterized in that, The step of windowing the normalized received signal to obtain the windowed received signal includes: A Hamming window is applied to the normalized received signal to obtain the windowed received signal.

5. The method according to claim 1, characterized in that, The interference suppression network model includes an encoder path and a decoder path; wherein, the encoder path is used to downsample the input feature data to extract multi-scale features; the decoder path is used to upsample the feature data output by the encoder path to restore the signal resolution; wherein, there is a skip connection between the encoder path and the decoder path, which is used to pass the features extracted by the encoder path to the corresponding layer of the decoder path.

6. The method according to claim 1, characterized in that, The interference signal included in the signal corresponding to the sample input data is a comb interference signal; the comb interference signal is generated by superimposing multiple independent interference components; wherein, each interference component is generated by band-limited noise signal modulation, and each interference component corresponds to a preset center frequency and a preset phase offset.

7. The method according to claim 6, characterized in that, The step of post-processing the multi-channel frequency domain feature data after interference suppression to obtain the time domain signal after interference suppression for subsequent communication demodulation includes: The real and imaginary channels of the multi-channel frequency domain feature data after interference suppression are combined to form a complex frequency domain signal; The complex frequency domain signal is processed by inverse fast Fourier transform to obtain the time domain signal after interference suppression.

8. An interference suppression device based on deep learning, characterized in that, The device includes: An acquisition module is used to acquire a received signal; wherein the received signal is the original transmitted signal after interference and noise have been superimposed during transmission. The preprocessing module is used to preprocess the received signal to obtain multi-channel frequency domain feature data that meets the input format requirements of the pre-trained interference suppression network model. The output module is used to input multi-channel frequency domain feature data into a pre-trained interference suppression network model. The interference suppression network model performs forward inference calculations and outputs the corresponding multi-channel frequency domain feature data after interference suppression. The interference suppression network model is pre-trained using training data, which includes sample input data and corresponding sample label data. The sample input data is the frequency domain feature data of the signal carrying interference, and the sample label data is the corresponding frequency domain feature data of the signal without interference. The interference suppression network model, through training, learns the mapping relationship from the frequency domain features of the signal with interference to the frequency domain features of the signal without interference. The post-processing module is used to post-process the multi-channel frequency domain feature data after interference suppression to obtain the time domain signal after interference suppression, which is used for subsequent communication demodulation.

9. An electronic device, characterized in that, include: A memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors to cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images