A radio communication interference identification method and network based on FPGA system

Abstract

This invention discloses a radio communication interference identification method and network based on an FPGA system, mainly addressing the problems of low identification accuracy and high resource consumption in current technologies. The solution involves: establishing a hole depth separable convolutional network; acquiring target sample signals containing different interference types, and constructing a four-channel multi-physical domain feature tensor including time-domain envelope features, frequency-domain amplitude spectrum features, square transform spectrum features, and second-order cyclic autocorrelation spectrum features; training the hole depth separable convolutional network using the four-channel multi-physical domain feature tensor of the target sample signals; acquiring the original low-clock-rate I / Q data of the signal under test and performing cross-clock domain processing and overlapping sliding frame division to calculate the four-channel multi-physical domain feature tensor of the signal under test, inputting it into the trained hole depth separable convolutional network, and outputting the identification result of the interference signal modulation type. This invention offers high identification accuracy, low resource consumption, and can be deployed on resource-constrained hardware platforms for radio communication.

Description

Technical Field

[0001] This invention belongs to the field of interference identification technology, and specifically relates to a radio communication interference identification method and network, which can be used in radio communication. Background Technology

[0002] In real-world radio communication scenarios, tactical radios are highly vulnerable to intense "targeted suppression jamming" from the enemy. To ensure the radio's survivability against jamming, a real-time automatic jamming modulation type identification mechanism must be introduced. However, existing jamming identification technologies exhibit significant limitations. On the one hand, traditional jamming identification methods rely excessively on manually extracted single features in the time or frequency domains. Targeted suppression jamming has the physical characteristics of highly concentrated energy and precise overlap with the communication signal frequency band. Especially when the jamming signal exhibits non-stationary modulation characteristics or is accompanied by dynamic channel multipath fading, the communication signal and the jamming become deeply coupled in both the time and frequency domains. At this point, traditional single-dimensional shallow features, such as simple time-domain I / Q or frequency-domain spectra, are difficult to effectively decouple, leading to blurred decision boundaries for the classifier and a significant drop in identification accuracy. On the other hand, although deep convolutional neural networks (CNNs) possess powerful advantages in nonlinear feature extraction, conventional network architectures are difficult to deploy directly on radio underlying hardware platforms with limited computing power and high real-time requirements, such as FPGAs, due to their large computational overhead, complex exponential operations, and high storage resource requirements. Simply compressing the model size by pruning channels can easily lead to the loss of spectral details and degradation of deep features.

[0003] Patent application CN201910772412.X discloses a satellite navigation interference signal identification device and method based on a convolutional neural network. The implementation steps are as follows: converting a typical one-dimensional satellite navigation interference signal into a two-dimensional image; inputting the two-dimensional image into a pre-trained convolutional neural network; and then using the convolutional neural network to automatically extract features from the signal and complete classification and identification. The convolutional neural network in this method contains three convolutional layers: 32 x 3x3 convolutional layers, 16 x 3x3 convolutional layers, and 16 x 3x3 convolutional layers, respectively. Each convolutional layer is followed by a normalization layer, and each normalization layer is followed by a pooling layer. Finally, a fully connected layer is added as the output layer, and the fully connected layer uses Softmax as the activation function. The exponential and division operations involved in Softmax increase hardware computing resource consumption, and the complex network architecture leads to a sharp increase in the number of model parameters and computational complexity with increasing network depth. Therefore, this method is difficult to deploy on resource-constrained platforms such as FPGAs.

[0004] Therefore, how to achieve an interference modulation type identification network that balances recognition accuracy and lightweight hardware deployment has become a technical bottleneck that urgently needs to be overcome in the current field. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing a radio communication interference identification method and network based on an FPGA system, which improves the accuracy of interference identification while reducing the consumption of hardware resources, thus achieving a balance between identification accuracy and lightweight hardware deployment.

[0006] To achieve the above objectives, the technical approach of this invention is as follows: A four-channel feature tensor is constructed by extracting "temporal envelope, frequency domain amplitude, squared transform spectrum, and second-order cyclic autocorrelation spectrum," utilizing the signal's features in different dimensions to improve the accuracy of interference identification; By introducing depthwise separable convolution and dilated convolution, the receptive field is expanded while reducing resource consumption, and a lightweight inverse residual structure is used to reduce on-chip cache consumption of high-dimensional features; Simultaneously, the squeeze-excitation attention mechanism and inverse residual structure are combined to enhance feature representation capabilities and improve the feature transfer efficiency of deep networks; By using the Hard-Sigmoid activation function instead of the more resource-intensive Sigmoid function, and replacing the Softmax operation with a multi-level pipelined binary comparison tree in the classification decision unit, computational resource consumption is reduced.

[0007] Based on the above ideas, the technical solution of the present invention includes:

[0008] 1. A radio communication interference identification method based on an FPGA system, characterized in that it includes:

[0009] (1) Use signal source, interference source and USRP hardware to obtain target sample signals containing different types of interference, and construct a four-channel multi-physical domain feature tensor including time domain envelope features, frequency domain amplitude spectrum features, square transform spectrum features and second-order cyclic autocorrelation spectrum features based on the target sample signals;

[0010] (2) The hole depth separable convolutional network is trained using the four-channel multi-physics domain feature tensor of the target sample signal;

[0011] (3) Obtain the original low clock rate I / Q data of the signal under test and perform cross-clock domain processing to obtain high clock rate I / Q data;

[0012] (4) Overlapping sliding frames are performed on the high clock rate I / Q data to obtain the framed data, and then parallel computation and timing isolation are performed on it to obtain the four-channel multi-physical domain feature tensor of the signal under test.

[0013] (5) Input the four-channel multi-physical domain feature tensor of the signal to be tested into the trained hole depth separable convolutional network and output the identification result of the interference signal modulation type.

[0014] Furthermore, the construction of the four-channel multi-physics domain feature tensor in (1), which includes time-domain envelope features, frequency-domain amplitude spectrum features, square-transform spectrum features, and second-order cyclic autocorrelation spectrum features, includes the following implementation:

[0015] (1a) Acquire target sample signal real part and the virtual part Calculate the temporal envelope features ;

[0016] (1b) For the target sample signal Perform a Fast Fourier Transform and take the modulus to obtain the frequency domain amplitude spectrum characteristics. ;

[0017] (1c) For the target sample signal Perform a squaring operation, then perform a Fast Fourier Transform on the squared signal and take the modulus to obtain the spectral characteristics of the squared transform. ;

[0018] (1d) Construct the target sample signal At a specific time delay The conjugate sequence below and compare it with the target sample signal Multiplying yields a conjugate multiplication sequence. A Fast Fourier Transform is then performed on this conjugate product sequence, and the modulus is taken to obtain the second-order cyclic autocorrelation spectral characteristics. ;

[0019] (1e) The four feature sequences generated above are standardized to eliminate the dimensional differences of each physical domain feature, and then spliced ​​in the channel dimension to finally generate the four-channel multi-physical domain feature tensor.

[0020] Furthermore, in step (4), parallel computation and temporal isolation are performed on the framed data to obtain the four-channel multi-physical domain feature tensor of the signal under test. Its implementation includes:

[0021] (4a) The data after framing is synchronously input to four parallel computing channels to extract time-domain envelope features, frequency-domain amplitude spectrum features, square transform spectrum features and second-order cyclic autocorrelation spectrum features respectively. Independent FIFO buffers are configured at the output end of each channel to isolate and latch the calculated feature data in time.

[0022] (4b) The ready flags of the four independent FIFO buffers are monitored in real time by the main control state machine. When all four ready flags are valid, a unified read enable signal is triggered so that the four features are aligned and output at the same clock edge. The normalized arithmetic operation of the four features is completed, and finally the four-channel multi-physical domain feature tensor of the signal under test after normalization is generated.

[0023] 2. A depth-separable convolutional network, characterized in that it comprises:

[0024] The backbone feature extraction unit, including the input layer, convolutional layer, and pooling layer, is used to perform initial spatial dimensionality reduction on the input multi-channel physical domain feature tensor to generate shallow basic features.

[0025] The channel attention unit, which includes a global average pooling layer, two fully connected layers, a ReLU activation function, and a Hard-Sigmoid activation function, is used to calculate channel weights based on the shallow basic features output by the backbone feature extraction unit to complete the nonlinear recalibration of features.

[0026] The multi-scale dilated depth separable convolutional unit contains multiple asymmetric parallel feature extraction branches, which are used to capture multi-dimensional scale features of the data in parallel based on recalibrated features and generate multi-scale fusion features.

[0027] Lightweight inverted residual unit, containing 2 1 convolutional layer The convolutional layers and channel attention units are used to receive multi-scale fused features. Through the inverse network topology structure of "first expand, then convolution, then compress", and embedding the channel attention mechanism in the shortcut connection of low-dimensional features, deep abstract features are finally generated.

[0028] The classification decision unit contains one global average pooling layer and one fully connected layer. It performs spatial compression and linear mapping based on deep abstract features to generate confidence score vectors for interference categories, outputs the category index corresponding to the maximum score, and obtains the discrimination result of the interference type.

[0029] Furthermore, in the multi-scale dilated depth separable convolutional unit, the method of utilizing multiple asymmetric parallel feature extraction branches to capture multi-dimensional scale features of the data in parallel based on recalibrated features includes:

[0030] The first asymmetric parallel feature extraction branch, through Convolutional kernels perform linear combination and compression of the channel dimension of the input nonlinear recalibrated features, preserving the most original feature information of the signal with minimal computational cost;

[0031] The second asymmetric parallel feature extraction branch first utilizes... Convolution performs channel transformation on the input nonlinear recalibrated features to obtain channel-transformed features, which are then passed through a system with a dilation rate of 2. The convolutional layers perform spatial filtering on the channel transform features to capture the local micro-texture of the signal, and finally... Convolutional layers achieve feature fusion using spatially filtered features;

[0032] The third asymmetric parallel feature extraction branch first utilizes... Convolution performs channel transformation on the input nonlinear recalibrated features to obtain channel-transformed features, which are then passed through a filter with a dilation rate of 5. The convolutional layers perform spatial filtering on the channel transformation features to expand the equivalent receptive field enhancement module's ability to model long-term dependent features. Finally, the... Convolutional layers achieve feature fusion using spatially filtered features;

[0033] The fourth asymmetric parallel feature extraction branch is... Max pooling cascade The convolutional layer structure utilizes max pooling to filter the nonlinear recalibrated features of the input, enhancing the module's ability to extract small frequency or time shift features. This results in pooled features, which are then further processed using... The convolutional layer adjusts the dimensions of the pooling features to generate the output features of that branch.

[0034] Furthermore, in the lightweight inverse residual unit, the reverse network topology structure of "first expansion, then convolution, then compression" is used to process the multi-scale fused features through a process of first... Convolutional layers are mapped to a high-dimensional space, and then utilized... Spatial correlation is extracted by group convolution, and finally by... Convolutional layers reproject features back into a low-dimensional space.

[0035] Compared with the prior art, the present invention has the following advantages:

[0036] Firstly, this invention avoids the sigmoid activation function and softmax function in traditional networks, which involve a large amount of exponential operations and floating-point division. The implementation of these functions is extremely resource-intensive in FPGAs. Therefore, this network replaces the sigmoid activation function with a hard-sigmoid activation function. Furthermore, in the classification decision stage, a multi-level pipelined binary comparison tree is directly used to retrieve the maximum confidence index, thus avoiding the complex calculations of softmax. This significantly reduces the logic operation resources required for implementing traditional sigmoid and softmax functions in FPGAs.

[0037] Secondly, by constructing a "four-channel feature tensor" containing multi-domain physical features in the network input, this invention can fully utilize the features of the signal in different dimensions to identify interference types by extracting the time-domain envelope, frequency-domain amplitude, square transform spectrum and second-order cyclic autocorrelation spectrum in parallel, which greatly improves the identification accuracy.

[0038] Third, by introducing depthwise separable convolution and dilated convolution into the network, this invention can expand the receptive field of the network while reducing resource consumption; at the same time, by introducing a lightweight inverse residual unit that integrates squeeze-excited attention units into the network, it can enhance the deep feature representation capability, alleviate the performance degradation problem in the training process of deep networks, and realize the replacement of simple stacked convolutional layers with structural optimization, effectively reducing the overall parameter scale of the model. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the implementation of the radio communication interference identification method based on an FPGA system according to the present invention.

[0040] Figure 2 This is a diagram showing the overall structure of the hole depth separable convolutional network in this invention.

[0041] Figure 3 This is a graph showing the accuracy of interference identification under different interference-to-signal ratios according to the present invention.

[0042] Figure 4 This is the interference identification confusion matrix diagram of the present invention under the condition of 10dB interference-to-signal ratio; Detailed Implementation

[0043] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should all fall within the protection scope of the present invention.

[0044] Example 1: A radio communication interference identification method based on an FPGA system

[0045] Reference Figure 1 The implementation steps of this example include the following:

[0046] Step 1: Obtain the target sample signal.

[0047] The system uses a signal source, an interference source, and USRP hardware to collect target sample signals containing different types of interference. ,in, The discrete time-domain sampling point indexes, signal types, and parameters are shown in Table 1.

[0048] Table 1. Target Sample Signal Types and Parameters

[0049]

[0050] Step 2, based on the target sample signal Construct a four-channel multi-physics domain feature tensor.

[0051] (2.1) Obtain the target sample signal real part and the virtual part Calculate the temporal envelope features :

[0052] ,

[0053] in, Input feature length to the network;

[0054] This envelope feature It can preserve the instantaneous energy fluctuation information of the signal in the time domain. Compared with the separate IQ components, the overall envelope can more intuitively reflect the instantaneous energy and amplitude change law of the signal. For amplitude modulation interference such as AM and ASK or pulse interference with a specific duty cycle, the time domain envelope is an important criterion.

[0055] (2.2) For the target sample signal Execution length is The fast Fourier transform is performed, and the frequency domain amplitude spectrum characteristics are obtained by taking the modulus. :

[0056] ;

[0057] in, This serves as an index for discrete frequency domain sampling points; spectral characteristics are the core basis for distinguishing between narrowband and broadband interference. This feature is introduced to provide the network with crucial information such as the signal's carrier frequency location, bandwidth, and number of spectral peaks.

[0058] (2.3) For the target sample signal Perform squaring and fast Fourier transform operations sequentially, and then take the modulus to generate the spectral features of the square transform. :

[0059] ;

[0060] This step addresses the characteristic of digital modulation interference such as BPSK, which suppresses the carrier signal's lack of carrier frequency components. By introducing nonlinear square operations, the original phase balance of the signal is disrupted, causing discrete spectral lines to appear at twice the carrier frequency. This compensates for the blind spots of linear spectrum analysis and enhances the network's sensitivity to digital modulation interference.

[0061] (2.4) Constructing the target sample signal At a specific time delay The following time delay sequence The product is multiplied by the target sample signal to obtain a conjugate product sequence. A fast Fourier transform is performed on the conjugate product sequence and the modulus is taken to generate a second-order cyclic autocorrelation spectrum feature. :

[0062] ;

[0063] Based on the statistical difference that AWGN does not possess cyclic features, this second-order cyclic autocorrelation spectrum feature can still retain significant correlation peaks under strong noise, which can effectively improve the network's noise robustness.

[0064] (2.5) The four feature sequences generated above—time-domain envelope feature, frequency-domain amplitude spectrum feature, square transform spectrum feature, and second-order cyclic autocorrelation spectrum feature—are standardized to eliminate the dimensional differences between the features in each physical domain, thus obtaining the features of the target sample signal after standardization. :

[0065] ;

[0066] in This represents the mean value of the feature as a whole, calculated from the set of features. This represents the standard deviation corresponding to that feature channel;

[0067] (2.6) The four standardized feature vectors are concatenated along the channel dimension to form a vector with dimension . The target sample signal four-channel multi-physics domain feature tensor.

[0068] Step 3, network training.

[0069] (3.1) Match the four-channel multi-physical domain feature tensor of the target sample with its corresponding real label of the interference type, and divide the matched feature tensor into training set and validation set in a ratio of 6:4;

[0070] (3.2) Input the training set into the depth-separable convolutional network with dilatancy established in Example 1, and output the prediction confidence score of each interference category;

[0071] (3.3) Calculate the loss value between the predicted score and the true label based on the cross-entropy loss function. Perform error backpropagation with the goal of minimizing this loss value, and calculate the error gradient of the parameters of each layer of the network. Here, the expression of the cross-entropy loss function is:

[0072] ;

[0073] in, There are a total of 10 categories representing interference signals. Representative sample in the Real labels in each category This represents the confidence score of the prediction given by the network model;

[0074] (3.4) Based on the above error gradient, use the Adam optimizer to update the convolutional kernel weights, fully connected layer weights and the mapping parameters of the Hard-Sigmoid activation function in the network, and after each training round, input the validation set into the network to calculate its accuracy.

[0075] (3.5) Repeat steps (3.2) to (3.4) until the network’s recognition accuracy on the validation set reaches convergence, then stop training to obtain a well-trained depth-separable convolutional network.

[0076] Step 4, cross-clock domain processing.

[0077] (4.1) In the actual communication environment, the original low clock rate I / Q data of the signal under test is acquired, and the original low clock rate I / Q data is continuously written into the asynchronous FIFO according to the low-speed sampling clock of the front end;

[0078] (4.2) The internal empty flag status of the asynchronous FIFO is detected in real time through the FPGA underlying control logic. When the empty flag is determined to change from valid to invalid, the read instruction under the high-speed processing clock domain of the backend is immediately triggered to obtain high clock rate I / Q data.

[0079] Step 5, overlapping sliding frames.

[0080] (5.1) Write the above high-frequency clock rate I / Q data into the dual-port RAM, set the read / write base address state machine, and set the frame length to [value missing]. Frame shift length is ;

[0081] (5.2) When the dual-port RAM first writes data to reach When the high-speed burst read frame state is triggered, a single frame signal is output. After the single frame signal is output, the system enters the waiting frame shift state and maintains the RAM overwrite operation.

[0082] (5.3) When the newly stored data in RAM reaches At that time, the pre-read base address pointer is updated and burst read is triggered again, thereby realizing overlapping sliding frames for continuous data.

[0083] Step 6: Obtain the four-channel multi-physics domain feature tensor of the signal under test.

[0084] (6.1) After the data is segmented, it is downsampled and then the downsampled data is synchronously input into four parallel feature calculation channels to calculate the time domain envelope, frequency domain amplitude, square transform spectrum and second-order cyclic autocorrelation spectrum features in parallel.

[0085] (6.2) Configure an independent FIFO buffer at the output end of each channel to isolate and latch the features that have been calculated first, and synchronously generate a ready flag bit for the data of that channel to complete the calculation.

[0086] (6.3) Use a state machine to determine whether the current channel's flag is valid:

[0087] If the flag changes from 0 to 1, it is determined to be valid. When the state machine determines that the flags of the four channels are all valid, it triggers a unified read enable signal, so that the four characteristic data are aligned and output on the same clock edge, and executes step (6.4).

[0088] Otherwise, enter a waiting state;

[0089] (6.4) Standardize the time domain envelope, frequency domain amplitude, square transform spectrum and second-order cyclic autocorrelation spectrum feature data according to the calculation formula in (2.5) and output the four-channel multi-physical domain feature tensor of the signal under test after standardization.

[0090] Step 7, interference category determination.

[0091] (7.1) The four-channel multi-physical domain feature tensor of the signal to be tested after standardization is input into the trained dilated depth separable convolutional network, and deep abstract features are extracted and fused step by step. After the deep abstract features are mapped by global average pooling and fully connected layers, a fixed-point confidence score vector is generated.

[0092] (7.2) Input the confidence score vector into the multi-level pipeline binary comparison tree, compare it level by level and output the maximum confidence score and its corresponding interference category index;

[0093] (7.3) Compare the maximum confidence score with the preset effective recognition decision threshold:

[0094] If the maximum confidence score is greater than or equal to the decision threshold, it is determined that there is a valid disturbance, and the system outputs the disturbance category index to the external control terminal.

[0095] If the maximum confidence score is less than the decision threshold, the communication environment is determined to be undisturbed or to be subject to unknown interference, and the system does not output the interference category.

[0096] It should be noted that the step numbers in this embodiment and the claims are only for the purpose of clearly describing the implementation scheme of the present invention and facilitating understanding, and their sequence order is not limited.

[0097] Example 2: A hole depth separable convolutional network.

[0098] Reference Figure 2This example includes a backbone feature extraction unit, a channel attention unit, a multi-scale dilated depth separable convolution unit, a lightweight inverse residual unit, and a classification decision unit, wherein:

[0099] The backbone feature extraction unit contains two... 1 convolutional layer The convolutional layer and two The max-pooling layer receives the input multi-channel physical domain feature tensor data and performs initial spatial dimensionality reduction and shallow basic feature extraction. The input feature tensor is first resized to fit the one-dimensional convolution processing format; subsequently, the data sequentially passes through a... Convolutional layer, one Max pooling layer and one The convolutional layer performs preliminary extraction on the input feature tensor to obtain downsampled features; these downsampled features are then input into two parallel asymmetric branches, one of which uses... Max pooling layers extract local extrema features; another branch uses... Convolutional layers extract locally smooth semantic features; the outputs of these two branches are concatenated to generate shallow basic features containing rich spatiotemporal information.

[0100] The channel attention unit is divided into a main branch and a weight calculation bypass. The main branch is the backbone feature extraction unit. The weight calculation bypass includes one global average pooling layer, two fully connected layers, a ReLU activation function, and a Hard-Sigmoid activation function. It is used to receive the shallow basic features output by the backbone feature extraction unit and adjust the weights of each channel to complete the nonlinear recalibration of the features. Its bypass first compresses the two-dimensional features into a one-dimensional channel descriptor through a global average pooling layer; then, it sequentially passes through a first fully connected layer, a ReLU activation function, and a second fully connected layer to learn the non-linear dependencies between channels; the Hard-Sigmoid activation function maps the features to channel weight vectors with values ​​between 0 and 1; this channel weight vector is multiplied element-wise with the features in the main branch through a multiplier to obtain a weighted feature tensor, which is then non-linearly activated by the ReLU activation function to output the final recalibrated features; the Hard-Sigmoid activation function is a linear fit to the traditional Sigmoid activation function, replacing the Sigmoid activation function which includes exponential operations and floating-point division, thus avoiding the large amount of logic operation resources required for its implementation in FPGA. The Sigmoid activation function calculation formula is as follows:

[0101] ;

[0102] The formula for calculating the Hard-Sigmoid activation function is:

[0103] ;

[0104] The multi-scale, depth-separable convolutional unit consists of four asymmetric parallel feature extraction branches, used to receive recalibrated features and capture multi-dimensional scale features in parallel. Its first asymmetric parallel feature extraction branch contains only one... The convolutional layer performs channel dimensionality reduction on the input nonlinear recalibrated features, preserving the original feature information of the signal; the second asymmetric parallel feature extraction branch uses a depthwise separable convolutional architecture, sequentially passing through... Convolutional layers perform channel transformation and hole ratio adjustment. of Convolutional layers are used for spatial filtering to capture local micro-textures of the signal, and then... The convolutional layer achieves feature fusion using spatial filtering features; the third asymmetric parallel feature extraction branch, with a structure similar to the second asymmetric parallel feature extraction branch, sequentially includes... Convolutional layer Group convolutional layers and Convolutional layers, in which The dilatation rate of the convolutional layer is set to 5. By configuring a larger dilatation rate, the equivalent receptive field can be expanded without increasing the number of model parameters, thereby enhancing the modeling ability of this unit for long-term dependent features of broadband interference; the fourth asymmetric parallel feature extraction branch is composed of a... Max pooling layer and one The system consists of convolutional layers that utilize max pooling to filter prominent features, enhancing the extraction of subtle frequency / temporal shift features. The convolutional layer adjusts the dimensions of the pooling features; the features output from the above four branches are concatenated, batch normalized, and ReLU activated to eliminate the dimensional differences of the features extracted by different branches and generate multi-scale fused features.

[0105] The lightweight inverse residual unit adopts a four-stage pipeline structure of "dimensionality increase-convolution-projection-weighting", which consists of 2 1 convolutional layer The system consists of convolutional layers and squeeze-excited attention units, used to receive the aforementioned multi-scale fused features and generate deep abstract features. The first stage utilizes... Pointwise convolution maps the input features to a high-dimensional space, reorganizing and decoupling the original features in the high-dimensional channels, providing a more sufficient feature foundation for subsequent spatial feature extraction; the second stage performs... Depthwise convolution decouples spatial and channel correlations through grouped convolution. Furthermore, because grouped convolution eliminates the dense inter-channel multiplication and addition operations of standard convolution, it reduces the computational scale by an order of magnitude compared to standard convolution of the same size, improving hardware computational efficiency. The third stage utilizes... Convolution projects features into a low-dimensional space to obtain low-dimensional features. To address the lack of inter-channel information interaction in deep convolution, this layer fuses independent spatial features into combined features with a global receptive field through linear combination across channels, thereby reducing the computational burden of subsequent modules. In the fourth stage, squeeze-excited attention units are embedded after the low-dimensional features to generate weights, which are then multiplied and weighted with the low-dimensional features and activated by ReLU. Finally, they are added element-wise with the original low-dimensional features to obtain deep abstract features.

[0106] The classification decision unit, consisting of one global average pooling layer and one fully connected layer, is used to output the final interference identification result. Located at the end of the network, this unit receives the deep abstract features output by the lightweight inverse residual unit, compresses them into a one-dimensional feature vector through the global average pooling layer, and then linearly maps this feature vector into fixed-point confidence score vectors for various interference modulation types through the fully connected layer. Finally, the maximum confidence score and its corresponding category index are taken as the final interference identification output.

[0107] It should be noted that the appendix Figure 2 The network layer parameters annotated in the documentation are uniformly formatted as number of output channels / stance. For example, 32 / 2 indicates that the network layer has 32 output channels and a stance of 2. Furthermore, if annotated with "..." "1" means that the hole rate parameter configured for this group of convolutions is set to 1. "2" indicates that the hole ratio parameter configured for this group of convolutions is set to 2. "5" indicates that the hole rate parameter configured for this group of convolutions is set to 5.

[0108] The above units are implemented in an FPGA to form a hole-depth separable convolutional network.

[0109] The effectiveness of this invention can be further illustrated by the following experimental results:

[0110] I. Experimental Parameters

[0111] The communication signal type, interference type, interference bandwidth, and input signal-to-interference-plus-noise ratio range are set as shown in Table 2.

[0112] Table 2 Experimental Parameter Table

[0113]

[0114] II. Experiment Content

[0115] Simulation Experiment 1: Under the above simulation conditions, this invention was used to identify nine typical interference signals (CW, AM, FM, ASK, BPSK, QPSK, MSK, 2FSK, and NoiFM) under different interference-to-signal ratios, and their respective identification accuracy curves were obtained, as shown below. Figure 3 As shown.

[0116] Depend on Figure 3 It can be seen that in the low interference-to-signal ratio (JSR < 0 dB) range, the recognition performance of the network model varies greatly and is generally low due to the influence of communication signal masking. When the JSR increases to the range of 5 dB to 10 dB, various interference features gradually become prominent and the accuracy gradually increases. When the JSR is greater than 10 dB, the recognition accuracy of most interference signals, except for AM interference, is stable at over 95%, which proves that the network model of this invention has excellent feature extraction ability and classification robustness.

[0117] Simulation Experiment 2: Under the above simulation conditions, this invention was used to identify nine typical interference signals (CW, AM, FM, ASK, BPSK, QPSK, MSK, 2FSK, and NoiFM) at a signal-to-interference ratio (SNR) of 10 dB. The confusion matrix of the identification results was obtained, as shown below. Figure 4 As shown.

[0118] Depend on Figure 4 It can be seen that when the JSR is 10dB, the recognition accuracy of most interference types reaches over 95%, and the recognition rates of BPSK and QPSK, which are difficult to distinguish under low JSR, are also improved to 97% and 96%, respectively. This is because in the input squared spectral features, the phase features of BPSK are transformed into discrete spectral line features, while QPSK still has no obvious spectral peaks after squared. This difference in features enhances the separability of the two types of signals under high JSR conditions. This demonstrates that the multi-domain feature collaborative sensing method introduced in this invention has advantages in feature decoupling and high-precision recognition that traditional single-feature networks do not possess when dealing with complex and highly similar interference patterns.

Claims

1. A method for identifying radio communication jamming based on FPGA system, characterized in that, include: (1) Use signal source, interference source and USRP hardware to obtain target sample signals containing different types of interference, and construct a four-channel multi-physical domain feature tensor including time domain envelope features, frequency domain amplitude spectrum features, square transform spectrum features and second-order cyclic autocorrelation spectrum features based on the target sample signals; (2) The hole depth separable convolutional network is trained using the four-channel multi-physics domain feature tensor of the target sample signal; (3) Obtain the original low clock rate I / Q data of the signal under test and perform cross-clock domain processing to obtain high clock rate I / Q data; (4) Overlapping sliding frames are performed on the high clock rate I / Q data to obtain the framed data, and then parallel computation and timing isolation are performed on it to obtain the four-channel multi-physical domain feature tensor of the signal under test. (5) Input the four-channel multi-physical domain feature tensor of the signal to be tested into the trained hole depth separable convolutional network and output the identification result of the interference signal modulation type.

2. The method of claim 1, wherein, The construction of a four-channel multi-physics domain feature tensor in step (1), which includes time-domain envelope features, frequency-domain amplitude spectrum features, square-transform spectrum features, and second-order cyclic autocorrelation spectrum features, is implemented as follows: (1a) obtaining a target sample signal the real part and the imaginary part of the complex number , calculating a time-domain envelope feature ； in, For discrete time-domain sampling point indexes, Input feature length to the network; (1b) For the target sample signal Perform a Fast Fourier Transform and take the modulus to obtain the frequency domain amplitude spectrum characteristics. : ； in, For discrete frequency domain sampling point indices; (1c) For the target sample signal Perform a squaring operation, then perform a Fast Fourier Transform on the squared signal and take the modulus to obtain the spectral characteristics of the squared transform. : ； (1d) Construct the target sample signal At a specific time delay The conjugate sequence below and compare it with the target sample signal Multiplying yields a conjugate multiplication sequence. A Fast Fourier Transform is then performed on this conjugate product sequence, and the modulus is taken to obtain the second-order cyclic autocorrelation spectral characteristics. : ； (1e) The four feature sequences generated above are standardized to eliminate the dimensional differences of each physical domain feature, and then spliced ​​in the channel dimension to finally generate the four-channel multi-physical domain feature tensor.

3. The method of claim 1, wherein, Step (2) involves training a depth-separable convolutional network using the four-channel multi-physics domain feature tensor of the target sample signal, including: (2a) The constructed four-channel multi-physical domain feature tensor and its corresponding real labels of interference types are matched and divided into training set and validation set in a ratio of 6:4; (2b) Input the feature tensors in the training set into the dilated depth separable convolutional network and output the prediction confidence score of each interference category; (2c) Calculate the loss function value between the predicted confidence score and the true label based on the cross-entropy loss function, and use the optimizer to update the mapping parameters of the convolutional kernel weights, fully connected layer weights and Hard-Sigmoid activation function in the network using the Adam algorithm; (2d) After the current training round ends, the feature tensors in the validation set are input into the network after updating the parameters for forward inference, the prediction results of the validation set data are output, and the network is compared with the real labels to calculate the recognition accuracy of the current network on the validation set. (2e) Repeat the above steps (2b) to (2d) until the network's recognition accuracy on the validation set converges, thus completing the training of the network.

4. The method of claim 1, wherein: In step (3), the original low clock rate I / Q signal is processed across clock domains by continuously writing the original low clock rate I / Q digital signal into the asynchronous FIFO according to the front-end low-speed sampling clock, and detecting the internal empty flag state of the asynchronous FIFO in real time. When the empty flag state changes from valid to invalid, the read instruction under the back-end high-speed clock domain is immediately triggered to complete the data conversion from the low-speed clock domain to the high-speed clock domain.

5. The method of claim 1, wherein: In step (4), overlapping sliding frames are performed on the high clock rate I / Q data to obtain the framed data. The implementation includes: (4a) write the high frequency clock rate I / Q data into the dual-port RAM, set the read-write base address state machine, set the frame length as , and the frame shift length as ; (4b) When the dual-port RAM first writes data to reach When the high-speed burst read frame state is triggered, a single frame signal is output. After the single frame signal is output, the system enters the waiting frame shift state and maintains the RAM overwrite operation. (4c) When the newly stored data in RAM reaches At that time, the pre-read base address pointer is updated and burst read is triggered again, thereby realizing overlapping sliding frame division of continuous data and obtaining the framed data.

6. A depth-separable convolutional network, characterized in that, include: The backbone feature extraction unit, consisting of 3 convolutional layers and 2 max pooling layers, is used to perform initial spatial dimensionality reduction on the input multi-channel physical domain feature tensor to generate shallow basic features. The channel attention unit, which includes a global average pooling layer, two fully connected layers, a ReLU activation function, and a Hard-Sigmoid activation function, is used to calculate channel weights based on the shallow basic features output by the backbone feature extraction unit to complete the nonlinear recalibration of features. The multi-scale dilated depth separable convolutional unit contains multiple asymmetric parallel feature extraction branches, which are used to capture multi-dimensional scale features of the data in parallel based on recalibrated features and generate multi-scale fusion features. Lightweight inverted residual unit, containing 2 1 convolutional layer The convolutional layers and channel attention units are used to receive multi-scale fused features. Through the inverse network topology of "first expand, then convolution, then compress", and embedding the channel attention mechanism in the shortcut connection of low-dimensional features, deep abstract features are finally generated. The classification decision unit contains one global average pooling layer and one fully connected layer. It performs spatial compression and linear mapping based on deep abstract features to generate confidence score vectors for interference categories, outputs the category index corresponding to the maximum score, and obtains the discrimination result of the interference type.

7. The network according to claim 6, characterized in that, In the backbone feature extraction unit, the initial spatial dimensionality reduction of the input multi-channel physical domain feature tensor to generate shallow basic features is implemented as follows: First, the input feature tensor data is resized to fit the one-dimensional convolution format, and then sequentially processed... Convolutional layers Maximum pooling layer and Convolutional layers perform continuous feature downsampling to achieve dimensionality reduction of features; After the aforementioned dimensionality reduction, the feature flow is divided into two parallel branches in the network, one branch employing... The max pooling layer extracts local extremum features, while the other branch uses... The convolutional layer extracts smooth local features, and the outputs of the two branches are spliced ​​together to generate shallow basic features.

8. The network according to claim 6, characterized in that, In the channel attention unit, the nonlinear recalibration of the completed features is implemented by including: First, the shallow basic features are sequentially processed through global average pooling, a fully connected layer, a ReLU activation function, and another fully connected layer to learn the non-linear dependencies between features and generate a one-dimensional channel description vector. Then, the Hard-Sigmoid activation function is obtained by linearly fitting the existing Sigmoid activation function: ；； in, The variable representing the input; the Sigmoid function is ; Next, the channel description vector is mapped to a channel weight vector between 0 and 1 using the Hard-Sigmoid activation function, and the channel weight vector is multiplied element-wise with the shallow basic features through a multiplier to obtain the weighted feature tensor. Finally, the weighted feature tensor is input into the ReLU activation function for nonlinear activation, and the final nonlinear recalibrated features are output.

9. The network according to claim 6, characterized in that, In a multi-scale, depth-separable convolutional unit with dilatancy, the method of utilizing multiple asymmetric parallel feature extraction branches to capture multi-dimensional scale features of the data in parallel based on recalibrated features includes: The first asymmetric parallel feature extraction branch, through Convolutional kernels perform linear combination and compression of the channel dimension of the input nonlinear recalibrated features, preserving the most original feature information of the signal with minimal computational cost; The second asymmetric parallel feature extraction branch first utilizes... Convolution performs channel transformation on the input nonlinear recalibrated features to obtain channel-transformed features, which are then passed through a system with a dilation rate of 2. Spatial filtering is performed on the channel transform features of the convolutional layers to capture the local micro-texture of the signal, and finally... Convolutional layers achieve feature fusion using spatially filtered features; The third asymmetric parallel feature extraction branch first utilizes... Convolution performs channel transformation on the input nonlinear recalibrated features to obtain channel-transformed features, which are then passed through a filter with a dilation rate of 5. The convolutional layers perform spatial filtering on the channel transformation features to expand the equivalent receptive field enhancement module's ability to model long-term dependent features. Finally, the... Convolutional layers achieve feature fusion using spatially filtered features; The fourth asymmetric parallel feature extraction branch is... Max pooling cascade The convolutional layer structure utilizes max pooling to filter the nonlinear recalibrated features of the input, enhancing the module's ability to extract small frequency or time shift features. This results in pooled features, which are then further processed using... The convolutional layer adjusts the dimensions of the pooling features to generate the output features of that branch.

10. The network according to claim 6, characterized in that: In the lightweight inverse residual unit, the inverse network topology structure of "expansion, convolution, and compression" is used to process the multi-scale fused features through... Convolutional layers are mapped to a high-dimensional space, and then utilized... Spatial correlation is extracted by group convolution, and finally by... Convolutional layers reproject features back into a low-dimensional space; In the classification decision unit, the step of generating a confidence score vector for interference categories by performing spatial compression and linear mapping based on deep abstract features includes: The deep abstract features are spatially compressed by a global average pooling layer, flattening the features into a one-dimensional feature vector. The feature vector is then linearly mapped to the confidence score vector corresponding to each type of interference through a fully connected layer. The maximum value in the confidence score vector is extracted, and its corresponding index category is used as the final interference category identification result.

Images