A broadband spectrum interference sensing and intelligent parameter estimation method and system

Abstract

The application relates to the field of wireless communication anti-interference technology and discloses a wideband spectrum interference sensing and intelligent parameter estimation method and system.The method comprises the following steps: extracting a training sequence part from a mixed receiving signal, performing channel estimation to obtain a channel response estimation value, and utilizing the channel response estimation value to cancel the mixed receiving signal to obtain a mixed signal of interference and noise; dividing the mixed signal of interference and noise into multiple uniform subchannels through a polyphase filter bank with frequency overlap; performing spectrum sensing on each subchannel through a fast covariance matrix eigenvalue (FCME) algorithm, and identifying and splicing the same interference across adjacent subchannels through a multi-subband joint detection algorithm; extracting a time-frequency diagram and IQ data according to the interference detection result, and pre-processing the time-frequency diagram; and inputting the time-frequency diagram into a YOLOv7 deep neural network to output parameter information of each interference through a target detection method, so that accurate intelligent estimation is realized through the above method.

Description

Technical Field

[0001] This invention relates to the field of wireless communication anti-interference technology, and in particular to a broadband spectrum interference sensing and intelligent parameter estimation method and system. Background Technology

[0002] In modern wireless communication systems, the electromagnetic spectrum environment is becoming increasingly complex and volatile, with various intentional or unintentional interference signals seriously threatening the reliability and stability of communication systems. With the rapid development of broadband communication technology, available spectrum resources are becoming increasingly scarce. Communication signals and interference signals highly overlap in the time and frequency domains, posing a significant challenge to signal processing at the receiver. Common types of interference include single-tone interference, multi-tone interference, broadband noise interference, narrowband interference, and time-varying frequency sweep interference. These interference signals vary in form and their parameters change dynamically, making interference detection and parameter estimation extremely difficult.

[0003] Traditional interference detection methods primarily rely on techniques such as energy detection, cyclostationary feature detection, and matched filtering. Energy detection is widely used due to its simplicity and lack of prior information, but its performance degrades significantly under low signal-to-noise ratio conditions and it struggles to distinguish between communication and interference signals. Cyclostationary feature detection, while utilizing the periodic stationarity of signals, demands high computational resources and exhibits poor adaptability to parameter variations. More importantly, these traditional methods share a fatal flaw in complex parameter estimation: they cannot adaptively handle a variable number of interference targets. Traditional methods typically assume the number of interferences is known or fixed, while deep learning methods often output fixed-dimensional parameter vectors, failing to match the dynamic changes in interference in real-world electromagnetic environments. For example, at one moment there might be only a single sweeping interference, while at the next moment multiple broadband noise interferences and comb-spectrum interferences might simultaneously appear. Existing technologies lack the intelligence to output as many sets of parameters as there are interferences, making it difficult to meet practical application requirements.

[0004] Furthermore, existing interference sensing technologies have limited ability to estimate interference parameters. They can often only detect the presence of interference but struggle to accurately estimate key parameters such as the center frequency, bandwidth, and pitch of the interference, failing to provide sufficient basis for subsequent adaptive avoidance by the communication system. Simultaneously, existing technologies lack effective communication signal cancellation mechanisms. During interference detection, residual communication signals mix with interference signals, severely impacting the accuracy of interference detection and parameter estimation.

[0005] In summary, there is an urgent need for an intelligent system that can accurately sense various types of interference, precisely estimate interference parameters, and adaptively handle a variable number of interference targets in a broadband spectrum environment. This system would support communication systems in implementing effective adaptive avoidance strategies and ensure communication reliability in complex electromagnetic environments. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems of low interference detection accuracy, inaccurate parameter estimation, and inability to effectively support communication avoidance in the prior art, and to provide a broadband spectrum interference sensing and intelligent parameter estimation method and system.

[0007] On the one hand, a broadband spectrum interference sensing and intelligent parameter estimation method is provided, including the following steps: S1: Extract the training sequence portion from the mixed received signal, perform channel estimation on it based on the orthogonal matching pursuit (OMP) algorithm to obtain the channel response estimate, and use the channel response estimate to cancel the mixed received signal to obtain a mixed signal of interference and noise; S2: The mixed signal of interference and noise is divided into multiple uniform sub-channels by a polyphase filter bank with a certain frequency overlap. S3: The Fast Covariance Matrix Eigenvalue (FCME) algorithm is used to perform spectrum sensing on each sub-channel to obtain the set of interference frequency points in each sub-band. The multi-sub-band joint detection algorithm is used to identify and splice the same interference that crosses adjacent sub-channels to obtain a complete interference detection result. S4: Extract the time-frequency diagram and IQ data of the channel where the interference is located based on the interference detection results, and preprocess the time-frequency diagram; S5: Input the preprocessed time-frequency graph into the pre-trained YOLOv7 deep neural network, and adaptively output the parameter information of each disturbance through the target detection method.

[0008] Further, in step S1, obtaining the mixed signal of interference and noise includes: S11: Extract the training sequence portion from the received signal The training sequence segment signal of the receiving end is obtained. Its expression is: ,in, For channel impulse response, It is a noise sequence. It is an interference sequence. This represents the convolution operation; S12: Yes Its frequency domain representation is obtained by performing a Discrete Fourier Transform (DFT): , in, For measured values, For the perception matrix, This is the channel response vector. This is a noise frequency domain representation. To represent the interference in the frequency domain, channel estimation is performed using the OMP algorithm based on the compressed sensing principle. The iterative process of the OMP algorithm includes: Initialize residual Index set Iteration counting Sparsity of input channel response and size are Perception matrix In each iteration, the correlation between the residual and each column of the perception matrix is ​​calculated, and the column index with the highest correlation is found. Update the index set ,in The calculated correlation matrix is ​​represented as: ; Reconstructing using the least squares method exist The corresponding signal response estimate component Simultaneously update the residuals ; make ,exist The system iterates within a certain range and outputs the final channel response estimate. ; S13: Estimate the channel response value Transforming to the frequency domain yields the channel frequency response estimate. And according to the formula Signal cancellation is performed, among which, The frequency domain representation of the received signal. The frequency domain representation of the training sequence is used to determine the signal after cancellation, which mainly includes interference and noise.

[0009] Further, in step S2, dividing the mixed signal of interference and noise into multiple uniform sub-channels includes: S21: The broadband signal after cancellation is divided into several sub-channels with the same nominal bandwidth by a channel uniform division unit. S22: A polyphase filter bank with 50% frequency overlap is used to channelize the signal, so that there is a certain overlap bandwidth between adjacent sub-channels, so as to ensure that the signal in each required communication sub-band can fall completely into the passband of the filter. S23: Reconstruct the signal based on the principle of polyphase decomposition, letting The signal after decimation is The prototype low-pass filter is Its multiphase components are ,in Then the final output signal after channelization of the k-th channel can be expressed as: This process enables signal reconstruction after channelization of the Q-path filter.

[0010] Further, in step S3, the spectrum sensing of each sub-channel using the Fast Covariance Matrix Eigenvalue (FCME) algorithm includes: S31: Based on the power spectrum of each sub-band, estimate the power spectrum of the first sub-band. Power of each sub-band: ,in, This represents the total number of frequency points within the sub-band. For the first The first in the individual belt spectral values ​​at each frequency point; S32: Find the minimum subband power value among all subband powers. This is used as a reference for noise power; S33: Set the desired false alarm probability Calculate the decision threshold factor Thus, the decision threshold for the subband level is obtained as follows: ; S34: For subbands whose power exceeds the subband level threshold, i.e. Further calculate the frequency-level decision threshold: ,in, As a frequency-level decision factor, This indicates that the median power of all frequency points within a subband is calculated. S35: By comparing the frequency point power with the frequency point level threshold point by point, i.e., if If so, the frequency point is determined to be an interference frequency point and added to the interference frequency point set. Simultaneously, the interference detection results of each sub-band are recorded to form an interference sub-band detection result vector. .

[0011] Preferably, in step S3, identifying and splicing the same interference spanning adjacent sub-channels using a multi-sub-band joint detection algorithm includes: Each sub-band is divided into four consecutive parts in the frequency domain; For adjacent th and the The child belt, if the first The last part of the sub-band and the first The energy of the first part of each subband exceeds the preset power threshold and satisfies the energy continuity condition. ,in, For continuous decision threshold, and Indicates the adjacent sub-band number The energy of a portion; if the condition is met, then it is determined that these two portions belong to the same cross-subband interference, and the first portion is... and the Subbands are categorized as the same interference and spliced ​​together.

[0012] Further, in step S4, the preprocessing of the time-frequency graph includes: Extract the corresponding time-frequency map and IQ data from the channel where the detected interference is located, normalize the pixel values ​​of the time-frequency map to the [0,1] interval, and adjust it to the size that meets the input requirements of the YOLOv7 network.

[0013] Further, in step S5, the YOLOv7 deep neural network is a YOLOv7-tiny lightweight network, and the training process of the YOLOv7 deep neural network includes: Extracting the optimal anchor boxes from the training dataset using the K-means clustering algorithm specifically includes: setting the number of clusters. The width of all real boxes and height After normalization, the data is used as input, and clustering is performed by maximizing the Intersection over Union (IoU) to obtain the desired results. Each cluster center is used as an anchor, and the clusters are divided into three levels according to size. In network training, the SimOTA dynamic label allocation strategy is used for positive sample matching. This strategy dynamically determines the number of positive samples to match for each ground truth box by calculating the cost matrix between the predicted box and the ground truth box. The cost takes into account both bounding box loss and classification loss. Calculate the total loss function based on the assigned positive samples. Its expression is: ,in, For the bounding box loss, the Distance IoU loss is used, and its expression is: ,in, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. Center point of the prediction box Center point of the real frame The Euclidean distance between them The length of the diagonal of the smallest bounding rectangle that can simultaneously contain both the predicted bounding box and the ground truth bounding box; For the target confidence loss, the binary cross-entropy (BCE) loss is adopted, and its expression is: ,in, The true label is the IoU value between the predicted bounding box and the true bounding box. The confidence level of the network prediction; For classification loss, the BCE loss is also used for calculation; , , The weighting coefficients are used to balance the three types of losses mentioned above.

[0014] Furthermore, in step S5, the parameter information for each interference output includes at least the interference type, center frequency, bandwidth, pitch number, and power. During output, the bounding box pixel coordinates detected by the YOLOv7 network need to be mapped to physical frequency parameters. The specific mapping method is as follows: , , , ; in, The starting frequency, For the termination frequency, For the center frequency, For bandwidth, and These are the left and right boundary pixel coordinates of the detection box on the time-frequency graph, respectively. This represents the total width of the time-frequency plot. and These are the lower and upper limits of the actual physical frequency range corresponding to the spliced ​​time-frequency diagram, respectively.

[0015] Preferably, when outputting the parameter information for each interference in step S5, estimation is performed based on the average energy value within the detection frame and the background noise level, specifically including: Calculate the average power spectral density of all pixels within the detection bounding box. And calculate the average power spectral density of the background region far from the interference. Then the interference power The estimated value is: ,in For bandwidth.

[0016] On the other hand, the present invention provides a broadband spectrum interference sensing and intelligent parameter estimation system, comprising: The communication signal cancellation module is used to extract the training sequence portion from the mixed received signal, perform channel estimation on it based on the orthogonal matching pursuit (OMP) algorithm to obtain the channel response estimate, and use the channel response estimate to cancel the mixed received signal to obtain a mixed signal of interference and noise. A digital channelization module is used to divide the mixed signal of interference and noise into multiple uniform sub-channels through a polyphase filter bank with a certain frequency overlap. The multi-sub-band joint interference detection module is used to perform spectrum sensing on each sub-channel using the Fast Covariance Matrix Eigenvalue (FCME) algorithm to obtain the set of interference frequency points in each sub-band. It also uses the multi-sub-band joint detection algorithm to identify and splice the same interference that spans adjacent sub-channels to obtain a complete interference detection result. The time-frequency diagram preprocessing module is used to extract the time-frequency diagram and IQ data of the channel where the interference is located based on the interference detection results, and to preprocess the time-frequency diagram. The intelligent parameter estimation module is used to input the preprocessed time-frequency map into the pre-trained YOLOv7 deep neural network and adaptively output the parameter information of each disturbance through the target detection method.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention uses the Orthogonal Matching Pursuit (OMP) algorithm to accurately cancel communication signals, eliminating the interference of communication signals on interference detection and significantly improving the accuracy of interference perception in low signal-to-noise ratio environments. This invention employs a multiphase filter bank with overlapping frequencies for channelization processing, overcoming the signal distortion problem caused by edge attenuation of traditional filters, and ensuring the complete preservation and accurate detection of interference signals across the entire frequency band. This invention designs a multi-subband joint detection algorithm, which effectively identifies and splices cross-subband interference by analyzing the energy continuity of neighboring subbands, thus solving the technical problem that the traditional FCME algorithm cannot completely detect broadband interference. This invention introduces the YOLOv7 deep neural network into interference parameter estimation, transforming the interference perception problem into a target detection problem on a time-frequency graph, and realizing adaptive identification of a variable number of interference targets; This invention enables high-precision output of key parameters through precise mapping of pixel coordinates to physical frequencies, providing a comprehensive and reliable decision-making basis for communication systems to implement anti-interference strategies such as frequency avoidance and power control. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a broadband spectrum interference sensing and intelligent parameter estimation method according to the present invention; Figure 2 This is a schematic diagram illustrating the cancellation result of the orthogonal matching pursuit (OMP) algorithm of the present invention on communication signals; Figure 3 This is a schematic diagram of the channelization and interference energy detection results according to the present invention; Figure 4This is a schematic diagram of a YOLOv7 parameter estimation network and process according to the present invention; Figure 5 This is a schematic diagram of a parameter estimation result according to the present invention; Figure 6 This is a schematic diagram illustrating a parameter estimation reasoning method according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In this invention, we introduce a parameter estimation module using a target detection method. For input signals mixed with interference and noise, the first step in interference detection is to address the mixing of interference and communication signals. Here, in the training sequence input method, a sparse matrix is ​​constructed to cancel the communication signal. To address the real-time and accuracy issues of broadband interference detection, we employ polyphase filtering. However, traditional polyphase filters suffer from severe edge attenuation, leading to signal deviations after passing through sub-bands, making accurate detection difficult. Therefore, we use a PFB filter design with 50% channel overlap to effectively ensure signal integrity within each sub-band. The signal falls within the passband of the filter, thus ensuring the accuracy of interference estimation after PFB. Traditional interference detection uses the FCME algorithm to detect energy, but it cannot accurately locate cross-channel interference. This paper designs a multi-subband joint detection FCME algorithm, dividing each subband into four parts. Energy detection within neighboring subbands is used to determine whether cross-band interference belongs to the same type. Then, the sub-channels containing each type of interference are concatenated and fed into interference identification and parameter estimation. A lightweight Yolov7-tiny network is used to adaptively estimate each type of interference on the time-frequency graph, thus accurately outputting the interference parameters. The specific embodiments of the present invention will be described below with reference to the accompanying drawings and examples.

[0021] Example 1 Please see Figure 1 The technical solution of the broadband spectrum interference sensing and intelligent parameter estimation method provided in this embodiment includes: S1: Extract the training sequence portion from the mixed received signal, perform channel estimation on it based on the orthogonal matching pursuit (OMP) algorithm to obtain the channel response estimate, and use the channel response estimate to cancel the mixed received signal to obtain a mixed signal of interference and noise.

[0022] The extracted training sequence segment signal can be represented as a convolutional combination of the receiver training sequence, the training sequence itself, the noise sequence, the interference sequence, and the channel impulse response. Specifically, obtaining the mixed signal of interference and noise includes: S11: Extract the training sequence portion from the received signal The training sequence segment signal at the receiving end is obtained. Its expression is: ,in, For channel impulse response, It is a noise sequence. It is an interference sequence. This represents the convolution operation; S12: Yes Its frequency domain representation is obtained by performing a Discrete Fourier Transform (DFT): , in, For measured values, For the perception matrix, This is the channel response vector. This is a noise frequency domain representation. To represent the interference in the frequency domain, channel estimation is performed using the OMP algorithm based on the compressed sensing principle. The iterative process of the OMP algorithm includes: Initialize residual Index set Iteration counting Sparsity of input channel response and size are Perception matrix In each iteration, the correlation between the residual and each column of the perception matrix is ​​calculated, and the column index with the highest correlation is found. Update the index set ,in The calculated correlation matrix is ​​represented as: ; Reconstructing using the least squares method exist The corresponding signal response estimate component Simultaneously update the residuals ; make ,exist The system iterates within a certain range and outputs the final channel response estimate. ; S13: Estimate the channel response value After DFT transformation, the communication signal is separated from noise and interference, and the channel frequency response estimate is obtained by transforming it to the frequency domain. And according to the formula Signal cancellation is performed, among which, The frequency domain representation of the received signal. The frequency domain representation of the training sequence, after cancellation, mainly includes interference and noise. For example... Figure 2 As shown in the spectrum diagram after OMP algorithm cancellation, the communication signal (peaks at approximately 180MHz and 300MHz) is almost completely cancelled, while the interference signals (single-tone interference 2 and interference 1) and noise signals are completely preserved. Comparing the two spectrum diagrams, it can be seen that the spectrum before cancellation included the superposition of the communication signal and the interference signal, while after cancellation, only the interference signal and noise background remain. This cancellation process greatly improves the accuracy of subsequent interference detection and parameter estimation, and avoids the influence of the communication signal on interference detection.

[0023] In this embodiment, we calculate the perception matrix. When orthogonality is guaranteed, it is first ensured by an orthogonal index set, which is represented as follows: , Placeholder markers indicate that values ​​at certain specific locations are set to 1 (i.e., the row is selected), while values ​​at other locations are 0, thus yielding the perception matrix. : , in, , For the first part of the matrix The former Subarrays composed of columns This represents the maximum multipath delay.

[0024] Next, the signal division in step S2 is performed. In step S2, the mixed signal of interference and noise is divided into multiple uniform sub-channels, including: S21: The broadband signal after cancellation is divided into several sub-channels with the same nominal bandwidth by a channel uniform division unit. S22: A polyphase filter bank with 50% frequency overlap is used to channelize the signal, so that there is a certain overlap bandwidth between adjacent sub-channels, so as to ensure that the signal in each required communication sub-band can fall completely into the passband of the filter. S23: Reconstruct the signal based on the principle of polyphase decomposition, letting The signal after decimation is The prototype low-pass filter is Its multiphase components are ,in Then the final output signal after channelization of the k-th channel can be expressed as: This process enables signal reconstruction after channelization of the Q-path filter.

[0025] Specifically, in this embodiment, we first divide the broadband spectrum into channels using uniform partitioning units. Each sub-channel has a bandwidth of 40MHz, with 50% frequency overlap between adjacent channels; the spectrum sensing unit ensures that the signal within each desired communication sub-band (within 20MHz) falls within the passband of the filter designed for each sub-band, thereby minimizing filter edge decay and aliasing effects.

[0026] Before applying the Fast Covariance Matrix Eigenvalue (FCME) algorithm to each sub-channel, signal reconstruction is performed first: , Low-pass filter (Suppressing the mirror image caused by signal interpolation), according to polyphase decomposition: , in Therefore, we can conclude that: , make We can obtain: .

[0027] make , Multiple interpolation value ,get: .

[0028] Thus, we obtained Signal reconstruction after channelization by the filter; in, Indicates the first The bandpass filter of each sub-channel at discrete time The system is used to suppress mirroring caused by signal interpolation; Denotes the complex exponential modulation factor, where For the first The center angular frequency of each sub-channel is used to move the prototype filter to that frequency band. This represents the total number of sub-channels (which is also the number of phases in the polyphase decomposition). express The complex variable being transformed Indicates the rotation factor. , Indicate its Power of 1 Indicates the modulated Domain delay factor, exponent This represents a negative power of a specific delay amount. Indicates the first The multiphase components are in Representation of a domain For multi-phase branch index; The output signal reconstructed after channelization ( (domain representation) Indicates the first Each channel output signal Transformation, The first phase is obtained by combining the input signal after decimation with a multiphase structure. Signal in the middle of the road Transformation, Represents the prototype filter After frequency shift Transformation, Sub-channel index ( ), For multi-phase branch index ( ), This represents the power of the rotation factor, used for weighted summation in multiphase structures. This indicates the delay term, which corresponds to the fixed delay at the output of the polyphase filter bank, ensuring alignment of each branch. This represents a newly defined intermediate variable, representing all... of The summation result after rotation factor weighting; This is the interpolation factor, which is set here to be equal to the number of sub-channels. .

[0029] Then, the spectrum sensing in step S3 is performed, which involves using the Fast Covariance Matrix Eigenvalue (FCME) algorithm to perform spectrum sensing for each sub-channel, including: S31: Based on the power spectrum of each sub-band, estimate the power spectrum of the first sub-band. Power of each sub-band: ,in, This represents the total number of frequency points within the sub-band. For the first The first in the individual belt spectral values ​​at each frequency point; S32: Find the minimum subband power value among all subband powers. This is used as a reference for noise power; S33: Set the desired false alarm probability Calculate the decision threshold factor Thus, the decision threshold for the subband level is obtained as follows: ; S34: For subbands whose power exceeds the subband level threshold, i.e. Further calculate the frequency-level decision threshold: ,in, As a frequency-level decision factor, This indicates that the median power of all frequency points within a subband is calculated. S35: By comparing the frequency point power with the frequency point level threshold point by point, i.e., if If so, the frequency point is determined to be an interference frequency point and added to the interference frequency point set. Simultaneously, the interference detection results of each sub-band are recorded to form an interference sub-band detection result vector. .

[0030] Simultaneously, through a multi-sub-band joint detection algorithm, the same interference spanning adjacent sub-channels is identified and spliced ​​together, including: Each sub-band is divided into four consecutive parts in the frequency domain; For adjacent th and the The child belt, if the first The last part of the sub-band and the first The energy of the first part of each subband exceeds the preset power threshold and satisfies the energy continuity condition. ,in, For continuous decision threshold, and Indicates the adjacent sub-band number The energy of a portion; if the condition is met, then it is determined that these two portions belong to the same cross-subband interference, and the first portion is... and the Subbands are categorized as the same interference and spliced ​​together.

[0031] like Figure 3 As shown, the channelized power detection results clearly indicate the location of the interfering sub-channel. In this embodiment, among the 20 sub-channels, sub-channels with abnormal power (such as channels 9 and 15) were successfully detected. The power of these two channels was significantly higher than the noise floor of other channels (approximately -150 dBm), indicating the presence of interference signals. The power of normal channels was generally consistent and low, while the power of interfering channels was significantly higher, providing accurate frequency location information for subsequent interference parameter estimation.

[0032] Specifically, to obtain the complete detected interference signal, and given that the parameters (center frequency, bandwidth, etc.) of the interference signal are unknown, for interference spanning one sub-band but not filling two sub-bands, analyzing only the interference information on a single sub-band will result in distorted information. For interference spanning more than one sub-band and filling more than one sub-band, since the interference detection principle is based on the difference between interference and noise power to distinguish interference, if an interference completely fills the sub-channel, the effectiveness of the interference detection algorithm will be significantly affected. Therefore, a multi-sub-band joint interference detection method is designed here. The specific algorithm design is shown in Table 1 below: Table 1. Interference detection algorithm for multi-subband joint operation This step extracts the sub-channel numbers and location information of all detected interference. For each detected interfering sub-channel, the system records its channel number, center frequency, and other key information. For interference spanning multiple sub-channels, adjacent sub-channels belonging to the same type of interference are classified and concatenated based on the judgment results of the multi-subband joint detection algorithm. For example, if a broadband interference spans sub-channels 9 and 10, the system will concatenate the data from these two sub-channels into a complete interference signal segment.

[0033] Next, the time-frequency graph in step S4 is preprocessed, including: Extract the corresponding time-frequency map and IQ data from the channel where the detected interference is located, normalize the pixel values ​​of the time-frequency map to the [0,1] interval, and adjust it to a size that meets the input requirements of the YOLOv7 network. In this example, we normalize the time-frequency map to the [0,1] interval and adjust its size to 640×640×3.

[0034] The YOLOv7 deep neural network is a lightweight YOLOv7-tiny network, comprising a backbone network, a neck network, and a detection head, used for feature extraction and target detection from the input time-frequency map. A parameter mapping unit maps the detected bounding box pixel coordinates to physical frequency parameters, outputting parameters such as the center frequency, bandwidth, pitch number, and power of the interference. The backbone network uses the efficient aggregation network ELAN module and the spatial pyramid pooling module SPPCSPC for feature extraction; the neck network uses the path aggregation network PANet for multi-scale feature fusion; and the detection head includes three detection layers—small-scale, medium-scale, and large-scale—used to detect interference targets of different sizes.

[0035] In step S5, the training process of the YOLOv7 deep neural network includes: Extracting the optimal anchor boxes from the training dataset using the K-means clustering algorithm specifically includes: setting the number of clusters. The width of all real boxes and height After normalization, the data is used as input, and clustering is performed by maximizing the Intersection over Union (IoU) to obtain the desired results. Each cluster center is used as an anchor, and the clusters are divided into three levels according to size. In network training, the SimOTA dynamic label allocation strategy is used for positive sample matching. This strategy dynamically determines the number of positive samples to match for each ground truth box by calculating the cost matrix between the predicted box and the ground truth box. The cost takes into account both bounding box loss and classification loss. Calculate the total loss function based on the assigned positive samples. Its expression is: ,in, For the bounding box loss, the Distance IoU loss is used, and its expression is: ,in, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. Center point of the prediction box Center point of the real frame The Euclidean distance between them The length of the diagonal of the smallest bounding rectangle that can simultaneously contain both the predicted bounding box and the ground truth bounding box; For the target confidence loss, the binary cross-entropy (BCE) loss is adopted, and its expression is: ,in, The true label is the IoU value between the predicted bounding box and the true bounding box. The confidence level of the network prediction; For classification loss, the BCE loss is also used for calculation; , , The weighting coefficients are used to balance the three types of losses mentioned above.

[0036] Specifically, before training the YOLOv7 network, the optimal anchor boxes need to be extracted from the training dataset as the final detection boxes. Traditional manual anchor point setting often fails to adapt to the layout on the time-frequency map of interfering signals; therefore, K-means clustering is used to extract anchors. First, extract the width and height information of each real bounding box, normalize the size of all boxes to the [0,1] range to eliminate the absolute influence of size, and set the initial number of anchors. (Corresponding to the three detection layers of YOLOv7, with 3 anchors per layer), the anchors obtained from clustering are sorted in order and divided into three levels: small-scale, medium-scale, and large-scale, to correspond to different types of interference. The specific method for extracting anchors is shown in Table 2 below: Table 2. Optimal Anchor Extraction by K-means Clustering The forward inference process involves normalizing the 640*640*3 time-frequency graph to [0,1], converting the data into tensor format, and completing data processing through data augmentation (such as Moasic concatenation). The complete YOLOv7 network used in this invention is as follows: Figure 4 As shown.

[0037] In this embodiment, the backbone network mainly consists of convolutional layers, an ELAN module, and an SPPC / SPC module. The network employs an efficient convergence network (ELAN), which departs from the traditional backpropagation method and instead designs the network structure through gradient propagation paths to shorten the shortest gradient path. When input image information passes through multiple layers, gradient vanishing or over-dilation may occur; therefore, introducing the ELAN module deepens the network and improves its accuracy. In the neck network, a path convergence network (PANet) is still used as the feature fusion part, adding bottom-up path enhancement to the feature pyramid network (FPN) to enhance the entire feature hierarchy using accurate low-level localization signals. In the detection head, standard convolutions replace REPConv to adjust the number of channels, and input processing is performed through three convolutional layers. Compared to YOLOv7, YOLOv7-tiny sacrifices some accuracy but has advantages in speed and lightweight design. The ELAN network outputs three feature maps, each with a shape of [Batch, 3, H, W, 9]. For each cell in the grid and each anchor, it predicts the bounding box center offset, width and height scaling factors, target confidence, and class probability. Then, the prediction results of the three detection layers are combined to perform non-maximum suppression (NMS) to remove duplicate detection boxes, thereby completing the prediction of interfering anchors.

[0038] Furthermore, in the loss function, YOLOv7 uses SimOTA for dynamic label assignment. Unlike traditional static label assignment, SimOTA can adaptively determine the number of positive samples (k) matched for each ground truth box based on the IoU quality between the predicted and ground truth boxes. The specific process includes: positive sample matching: cost matrix calculation, dynamic determination of k value, optimal matching. The loss function comprehensively considers three aspects: IoU, confidence, and the BCE loss for classification. First, positive sample matching is performed, i.e., candidate anchors are determined by comparing the anchor aspect ratio with a threshold. The number of positive samples is increased by expanding the center grid. The IoU between candidate and ground truth boxes and the classification loss are calculated. The number of positive samples (k) matched for each ground truth box is dynamically determined, and the total loss is determined based on these factors. The specific loss calculation method is shown in the total loss function above. The calculation involves forward inference to derive the bounding box, and then using the pixel coordinates and the spliced ​​filename (containing channel frequency information) obtained from multi-subband detection to infer the physical parameters, thereby obtaining the actual location information of the interference and completing the parameter estimation.

[0039] Finally, in step S5, the output parameter information for each interference should include at least the interference type, center frequency, bandwidth, pitch number, and power. During output, the bounding box pixel coordinates detected by the YOLOv7 network need to be mapped to physical frequency parameters. The specific mapping method is as follows: , , , ; in, The starting frequency, For the termination frequency, For the center frequency, For bandwidth, and These are the left and right boundary pixel coordinates of the detection box on the time-frequency graph, respectively. This represents the total width of the time-frequency plot. and These are the lower and upper limits of the actual physical frequency range corresponding to the spliced ​​time-frequency diagram, respectively.

[0040] Furthermore, when outputting the parameter information for each disturbance in step S5, estimation is performed based on the average energy value within the detection frame and the background noise level, specifically including: Calculate the average power spectral density of all pixels within the detection bounding box. And calculate the average power spectral density of the background region far from the interference. Then the interference power The estimated value is: ,in For bandwidth.

[0041] The final complete output of the interference parameter estimation results is shown in Table 3 below: Table 3 Parameter estimation inference results These parameters provide accurate decision-making basis for communication systems to implement anti-interference measures such as frequency avoidance, power control, and frequency hopping strategies. Communication systems can select frequency bands to avoid interference based on the estimated interference center frequency and bandwidth; adjust transmission power or adopt corresponding coding and modulation schemes according to the interference power and type; and predict the sweeping trajectory of frequency-sweeping interference to avoid frequencies in advance.

[0042] like Figure 5 and Figure 6 As shown, the system can accurately identify and estimate parameters of various types of interference. Figure 5 The parameter estimation results for different types of interference are presented, along with the NRMSE of these parameters under different interference-to-signal ratios. The NRMSE of different parameters gradually decreases with increasing JSR, and all NRMSE values ​​are below 0.1. Figure 6 The inference effect of parameter estimation is demonstrated, showing the accurate localization of the detection box on the time-frequency graph, as well as the corresponding category label and confidence score. Experimental results verify the effectiveness and practicality of the method of this invention, enabling high-precision, real-time interference parameter estimation in complex electromagnetic environments.

[0043] This invention combines key technologies such as OMP signal cancellation, multi-subband joint detection, and deep learning parameter estimation to construct a complete broadband spectrum interference sensing and parameter estimation system, effectively solving the interference detection problem in complex electromagnetic environments and providing strong support for intelligent anti-interference of communication systems.

[0044] Based on the above methods, this invention provides a broadband spectrum interference sensing and intelligent parameter estimation system, comprising: The communication signal cancellation module is used to extract the training sequence portion from the mixed received signal, perform channel estimation on it based on the orthogonal matching pursuit (OMP) algorithm to obtain the channel response estimate, and use the channel response estimate to cancel the mixed received signal to obtain a mixed signal of interference and noise. A digital channelization module is used to divide the mixed signal of interference and noise into multiple uniform sub-channels through a polyphase filter bank with a certain frequency overlap. The multi-sub-band joint interference detection module is used to perform spectrum sensing on each sub-channel using the Fast Covariance Matrix Eigenvalue (FCME) algorithm to obtain the set of interference frequency points in each sub-band. It also uses the multi-sub-band joint detection algorithm to identify and splice the same interference that spans adjacent sub-channels to obtain a complete interference detection result. The time-frequency diagram preprocessing module is used to extract the time-frequency diagram and IQ data of the channel where the interference is located based on the interference detection results, and to preprocess the time-frequency diagram. The intelligent parameter estimation module is used to input the preprocessed time-frequency map into the pre-trained YOLOv7 deep neural network and adaptively output the parameter information of each disturbance through the target detection method.

[0045] It should be noted that the steps in the broadband spectrum interference sensing and intelligent parameter estimation method provided in this embodiment can be implemented based on the corresponding modules in the broadband spectrum interference sensing and intelligent parameter estimation system. Those skilled in the art can refer to the technical solution of the system to implement the steps of the method. That is, the embodiments in the system can be understood as preferred examples of implementing the method, and will not be elaborated here.

[0046] Besides implementing the system and its various devices provided by this invention in purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the system and its various devices of this invention appear as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices provided by this invention can be considered as a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0047] Finally, it should be noted that the above description is only a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be pointed out that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

[0048] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A broadband spectrum interference sensing and intelligent parameter estimation method, characterized in that, Includes the following steps: S1: Extract the training sequence portion from the mixed received signal, perform channel estimation on it based on the orthogonal matching pursuit (OMP) algorithm to obtain the channel response estimate, and use the channel response estimate to cancel the mixed received signal to obtain a mixed signal of interference and noise; S2: The mixed signal of interference and noise is divided into multiple uniform sub-channels by a polyphase filter bank with a certain frequency overlap. S3: The Fast Covariance Matrix Eigenvalue (FCME) algorithm is used to perform spectrum sensing on each sub-channel to obtain the set of interference frequency points in each sub-band. The multi-sub-band joint detection algorithm is used to identify and splice the same interference that crosses adjacent sub-channels to obtain a complete interference detection result. S4: Extract the time-frequency diagram and IQ data of the channel where the interference is located based on the interference detection results, and preprocess the time-frequency diagram; S5: Input the preprocessed time-frequency graph into the pre-trained YOLOv7 deep neural network, and adaptively output the parameter information of each disturbance through the target detection method.

2. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 1, characterized in that, In step S1, obtaining the mixed signal of interference and noise includes: S11: Extract the training sequence portion from the received signal The training sequence segment signal at the receiving end is obtained. Its expression is: ,in, For channel impulse response, It is a noise sequence. It is an interference sequence. This represents the convolution operation; S12: Yes Its frequency domain representation is obtained by performing a Discrete Fourier Transform (DFT): ， in, For measured values, For the perception matrix, This is the channel response vector. This is a noise frequency domain representation. To represent the interference in the frequency domain, channel estimation is performed using the OMP algorithm based on the compressed sensing principle. The iterative process of the OMP algorithm includes: Initialize residual Index set Iteration counting Sparsity of input channel response and size are Perception matrix In each iteration, the correlation between the residual and each column of the perception matrix is ​​calculated, and the column index with the highest correlation is found. Update the index set ,in The calculated correlation matrix is ​​represented as: ; Reconstructing using the least squares method exist The corresponding signal response estimate component Simultaneously update the residuals ; make ,exist The system iterates within a certain range and outputs the final channel response estimate. ; S13: Estimate the channel response value Transforming to the frequency domain yields the channel frequency response estimate. And according to the formula Signal cancellation is performed, among which, The frequency domain representation of the received signal. The frequency domain representation of the training sequence is used to determine the signal after cancellation, which mainly includes interference and noise.

3. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 1, characterized in that, In step S2, dividing the mixed signal of interference and noise into multiple uniform sub-channels includes: S21: The broadband signal after cancellation is divided into several sub-channels with the same nominal bandwidth by a channel uniform division unit. S22: A polyphase filter bank with 50% frequency overlap is used to channelize the signal, so that there is a certain overlap bandwidth between adjacent sub-channels, so as to ensure that the signal in each required communication sub-band can fall completely into the passband of the filter. S23: Reconstruct the signal based on the principle of polyphase decomposition, letting The signal after decimation is The prototype low-pass filter is Its multiphase components are ,in Then the final output signal after channelization of the k-th channel can be expressed as: This process enables signal reconstruction after channelization of the Q-path filter.

4. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 1, characterized in that, In step S3, the spectrum sensing of each sub-channel using the Fast Covariance Matrix Eigenvalue (FCME) algorithm includes: S31: Based on the power spectrum of each sub-band, estimate the power spectrum of the first sub-band. Power of each sub-band: ,in, This represents the total number of frequency points within the sub-band. For the first The first in the individual belt spectral values ​​at each frequency point; S32: Find the minimum subband power value among all subband powers. This is used as a reference for noise power; S33: Set the desired false alarm probability Calculate the decision threshold factor Thus, the decision threshold for the subband level is obtained as follows: ; S34: For subbands whose power exceeds the subband level threshold, i.e. Further calculate the frequency-level decision threshold: ,in, As a frequency-level decision factor, This indicates that the median power of all frequency points within a subband is calculated. S35: By comparing the frequency point power with the frequency point level threshold point by point, i.e., if If so, the frequency point is determined to be an interference frequency point and added to the interference frequency point set. Simultaneously, the interference detection results of each sub-band are recorded to form an interference sub-band detection result vector. .

5. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 4, characterized in that, In step S3, the multi-subband joint detection algorithm identifies and splices together the same interference spanning adjacent sub-channels, including: Each sub-band is divided into four consecutive parts in the frequency domain; For adjacent th and the The child belt, if the first The last part of the sub-band and the first The energy of the first part of each subband exceeds the preset power threshold and satisfies the energy continuity condition. ,in, For continuous decision threshold, and Indicates the adjacent sub-band number The energy of a portion; if the condition is met, then it is determined that these two portions belong to the same cross-subband interference, and the first portion is... and the Subbands are categorized as the same interference and spliced ​​together.

6. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 1, characterized in that, In step S4, the preprocessing of the time-frequency graph includes: Extract the corresponding time-frequency map and IQ data from the channel where the detected interference is located, normalize the pixel values ​​of the time-frequency map to the [0,1] interval, and adjust it to the size that meets the input requirements of the YOLOv7 network.

7. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 1, characterized in that, In step S5, the YOLOv7 deep neural network is a YOLOv7-tiny lightweight network, and the training process of the YOLOv7 deep neural network includes: Extracting the optimal anchor boxes from the training dataset using the K-means clustering algorithm specifically includes: setting the number of clusters. The width of all real boxes and height After normalization, the data is used as input, and clustering is performed by maximizing the Intersection over Union (IoU) to obtain the desired results. Each cluster center is used as an anchor, and the clusters are divided into three levels according to size. In network training, the SimOTA dynamic label allocation strategy is used for positive sample matching. This strategy dynamically determines the number of positive samples to match for each ground truth box by calculating the cost matrix between the predicted box and the ground truth box. The cost takes into account both bounding box loss and classification loss. Calculate the total loss function based on the assigned positive samples. Its expression is: ,in, For the bounding box loss, the Distance IoU loss is used, and its expression is: ,in, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. Center point of the prediction box Center point of the real frame The Euclidean distance between them The length of the diagonal of the smallest bounding rectangle that can simultaneously contain both the predicted bounding box and the ground truth bounding box; For the target confidence loss, the binary cross-entropy (BCE) loss is adopted, and its expression is: ,in, The true label is the IoU value between the predicted bounding box and the true bounding box. The confidence level of the network prediction; For classification loss, the BCE loss is also used for calculation; , , The weighting coefficients are used to balance the three types of losses mentioned above.

8. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 1, characterized in that, In step S5, the output parameter information for each interference should include at least the interference type, center frequency, bandwidth, pitch number, and power. During output, the bounding box pixel coordinates detected by the YOLOv7 network need to be mapped to physical frequency parameters. The specific mapping method is as follows: ， ， ， ； in, The starting frequency, For the termination frequency, For the center frequency, For bandwidth, and These are the left and right boundary pixel coordinates of the detection box on the time-frequency graph, respectively. This represents the total width of the time-frequency plot. and These are the lower and upper limits of the actual physical frequency range corresponding to the spliced ​​time-frequency diagram, respectively.

9. The broadband spectrum interference sensing and intelligent parameter estimation method according to claim 1, characterized in that, In step S5, when outputting the parameter information for each interference, estimation is performed based on the average energy value within the detection frame and the background noise level, specifically including: Calculate the average power spectral density of all pixels within the detection bounding box. And calculate the average power spectral density of the background region far from the interference. Then the interference power The estimated value is: ,in For bandwidth.

10. A broadband spectrum interference sensing and intelligent parameter estimation system, characterized in that, include: The communication signal cancellation module is used to extract the training sequence portion from the mixed received signal, perform channel estimation on it based on the orthogonal matching pursuit (OMP) algorithm to obtain the channel response estimate, and use the channel response estimate to cancel the mixed received signal to obtain a mixed signal of interference and noise. A digital channelization module is used to divide the mixed signal of interference and noise into multiple uniform sub-channels through a polyphase filter bank with a certain frequency overlap. The multi-sub-band joint interference detection module is used to perform spectrum sensing on each sub-channel using the Fast Covariance Matrix Eigenvalue (FCME) algorithm to obtain the set of interference frequency points in each sub-band. It also uses the multi-sub-band joint detection algorithm to identify and splice the same interference that spans adjacent sub-channels to obtain a complete interference detection result. The time-frequency diagram preprocessing module is used to extract the time-frequency diagram and IQ data of the channel where the interference is located based on the interference detection results, and to preprocess the time-frequency diagram. The intelligent parameter estimation module is used to input the preprocessed time-frequency map into the pre-trained YOLOv7 deep neural network and adaptively output the parameter information of each disturbance through the target detection method.

Images