Broadband wireless communication signal detection method and system under complex environment
By applying deep learning centripetal offset network in broadband wireless communication signal detection, using multi-scale features and corner point matching algorithm of time spectrum diagram, the problem of missing detection and estimation error in complex environments is solved, and the problem of more Efficient signal detection effect.
Patent Information
- Application Number
- CN202111543399.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-16
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-12-16
AI Technical Summary
In complex environments, the existing broadband wireless communication signal detection methods have problems such as high missed detection probability, large errors in signal start and end time and carrier frequency parameter estimation, especially in the case of time-frequency aliasing, channel fading and non-Gaussian noise. The detection effect is not good.
Using the centripetal offset network in deep learning, the signal detection model is constructed through the multi-scale feature extraction and corner point matching algorithm of the time spectrum diagram, the depth features can be extracted using the hourglass basic network and the cross star, and the depth features can be extracted through centripetal The offset module performs signal target positioning and generates a signal sample database to improve detection accuracy.
It effectively improves the detection effect of broadband wireless communication signals in complex environments, improves the accuracy and efficiency of signal detection, reduces the probability of missed detection, and enhances the ability to estimate the start and end time of the signal.
Smart Images

Figure CN114358054B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wireless communication signal recognition, and in particular relates to a method and system for detecting broadband wireless communication signals in a complex environment. Background Art
[0002] With the rapid development of wireless communication technology in recent years, the wireless communication channel environment has become increasingly complex. For non-cooperative receivers, it is very common for signals to experience time-frequency aliasing, channel fading, non-Gaussian noise, and low signal-to-noise ratio, which brings considerable difficulties to the traditional blind detection technology of broadband wireless communication signals. At present, many researchers have proposed solutions to this problem. Some methods use the energy index of the broadband signal power spectrum as the basis for detection, but this type of method is limited by the accuracy of the signal noise floor estimation, and its performance is significantly reduced in complex electromagnetic environments. Some methods channelize broadband signals and use narrowband signal detection methods to detect signals in each channel after channelization. In the narrowband detection link, the detection effect of this type of method is greatly affected by noise and the selection of thresholds, and in the channelization process, there is an unsatisfactory segmentation and a certain amount of original signal data information is lost.
[0003] In addition, as neural network technology is introduced into the field of wireless communication signal analysis and processing, some methods use neural networks to learn the one-dimensional power spectrum characteristics of broadband communication signals to carry out signal detection. This type of method can distinguish environmental background noise and carrier signals by transforming the signal detection problem into a binary classification problem of the power spectrum interval. The disadvantage is that it does not use the time domain information of the wireless communication signal, so it is impossible to estimate the start and end time of the signal. Other methods use the target detection network model in deep learning to extract and process the time-frequency spectrum of the signal, which can estimate the start and end time of the signal and determine the existence of the signal target. The anchor (AB)-based detection algorithm used in this type of method is not effective enough for the detection of scale-varying signals in real scenes, and often misses the detection under low signal-noise conditions. In addition, the current research work in this field is mainly focused on the ideal channel environment with no time-frequency aliasing, and the detection under complex channel environment conditions such as time-frequency aliasing, channel fading, and non-Gaussian noise is not considered. Summary of the invention
[0004] To this end, in order to solve the problems of high missed detection probability and large estimation errors of signal start and end time, carrier frequency and other parameters in broadband wireless communication signal detection under complex environments, the present invention provides a method and system for broadband wireless communication signal detection under complex environments, which introduces the centripetal offset network in the field of deep learning natural target recognition and processing into the field of signal detection, and can achieve effective detection of broadband wireless communication signals under complex environments, thereby improving the signal detection effect.
[0005] According to the design scheme provided by the present invention, a method for detecting broadband wireless communication signals in a complex environment includes the following contents:
[0006] The collected broadband wireless communication signals are preprocessed to obtain a time-frequency spectrum diagram of uniform size, and the time-frequency spectrum diagram is used to generate a signal sample database for model training;
[0007] A signal detection model based on a centripetal shift network is constructed, and the signal detection model is trained using signal data in a signal sample database, wherein the signal detection model comprises: an hourglass basic network for obtaining multi-scale features by performing up- and down-sampling operations on input data, a feature processing module for obtaining corner prediction results and signal depth features by performing corner pooling and deformable convolution on multi-scale features, and a centripetal shift module for locating corner prediction results and depth features using a corner matching algorithm to output a target corner position;
[0008] The broadband wireless communication signal to be detected is preprocessed, the preprocessed signal data is used as a model input, and the target position of the broadband wireless communication signal to be detected is identified through the trained signal detection model.
[0009] As the broadband wireless communication signal detection method under complex environment of the present invention, further, a standard broadband wireless communication signal for constructing a signal sample database is generated according to the random symbol sequence and the signal modulation mode.
[0010] As a broadband wireless communication signal detection method under a complex environment of the present invention, further, when a standard broadband wireless communication signal is generated, several narrowband signals are randomly selected, and the narrowband signals and the generated standard broadband wireless communication signal are aliased according to a preset aliasing probability; and α-stable noise is added to the standard broadband wireless communication signal.
[0011] As a broadband wireless communication signal detection method under a complex environment of the present invention, further, in signal aliasing, a cosine wave equal to the bandwidth of a narrowband signal is used as a fading channel, and the fading intensity of the fading channel and the termination phase of the fading channel are randomly selected, and Gaussian noise is used to simulate the actual channel environment to generate broadband mixed multi-signal sample data.
[0012] As a broadband wireless communication signal detection method under a complex environment of the present invention, further, a deep feature extraction module in the signal detection model includes: a corner point pooling unit for performing corner point pooling on multi-scale feature data, a corner point prediction unit for performing corner point prediction processing on a branch of a corner point pooling result, and a feature learning unit for performing local feature learning on another branch of the corner point pooling result through a cross-star variable convolution operation, wherein the corner point prediction unit and the feature learning unit perform corresponding corner point prediction processing and local feature learning on the lower right corner and the upper left corner of the feature map respectively.
[0013] As a broadband wireless communication signal detection method under a complex environment of the present invention, further, in the feature learning unit, first, convolution is performed on another branch of the corner pooling result to obtain an offset field; then, variability convolution is performed on the offset field and another branch of the corner pooling result data to extract signal depth features.
[0014] As a broadband wireless communication signal detection method under a complex environment of the present invention, further, in the centripetal shift module, according to the consistency of the center points of the lower right corner and the upper left corner of the same target, the centripetal shift algorithm is used to calculate the center point coordinates through the corner points, and the quantization weights of the corner point pairs in the corner point prediction results are set according to the degree of proximity between the signal target center and the regression center of each corner point, and the corner point pairs are re-scored by the original scores and quantization weights obtained by corner point pooling for each group of corner point pairs, and the corner point pairs as the corner points of the signal target position are determined according to the size of the score.
[0015] Furthermore, the present invention also provides a broadband wireless communication signal detection system in a complex environment, comprising: a sample collection module, a model building module and a signal recognition module, wherein:
[0016] A sample collection module is used to collect broadband wireless communication signals and pre-process the signals to obtain a time-frequency spectrum diagram of uniform size, and use the time-frequency spectrum diagram to generate a signal sample database for model training;
[0017] A model construction module, used to construct a signal detection model based on a centripetal shift network, and train the signal detection model using signal data in a signal sample database, wherein the signal detection model includes: an hourglass basic network for obtaining multi-scale features by performing up- and down-sampling operations on input data, a feature processing module for obtaining corner prediction results and signal depth features by performing corner pooling and deformable convolution on multi-scale features, and a centripetal shift module for locating corner prediction results and depth features using a corner matching algorithm to output a target corner position;
[0018] The signal recognition module is used to preprocess the broadband wireless communication signal to be detected, use the preprocessed signal data as model input, and identify the target position of the broadband wireless communication signal to be detected through the trained signal detection model.
[0019] Beneficial effects of the present invention:
[0020] The present invention uses a centripetal shift network model for detecting broadband wireless communication signals, takes the time-frequency spectrum of the wireless communication signal as a shallow expression of the signal characteristics, extracts and processes the deep features of the input time-frequency spectrum data, and utilizes an hourglass basic network (HN), a cross-star deformable convolution, and a corner point adaptation algorithm based on centripetal shift to achieve effective detection of broadband wireless communication signals in complex environments. Furthermore, according to the characteristics of broadband wireless communication signal detection, the model training loss function is improved, and a multi-scale training strategy is used to improve the training efficiency and generalization performance of the network model, so as to achieve good signal detection effects and facilitate application in actual scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a schematic diagram of the process of the broadband wireless communication signal detection method in a complex environment in the embodiment;
[0022] Figure 2 This is a schematic diagram of the signal sample generation process in the embodiment;
[0023] Figure 3 This is a schematic diagram of the signal detection model structure in the embodiment;
[0024] Figure 4 Schematic diagram of the signal target corner point and its centripetal offset in the embodiment;
[0025] Figure 5 Schematic diagram of the signal corner points corresponding to the center points in the embodiment;
[0026] Figure 6 This is a schematic diagram of the multi-scale training process of the signal detection model in the embodiment;
[0027] Figure 7 This is an example of the wireless communication signal detection process in the embodiment. DETAILED DESCRIPTION
[0028] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.
[0029] The present invention provides a method for detecting broadband wireless communication signals in a complex environment. Figure 1 As shown, it contains the following contents:
[0030] S101, collecting broadband wireless communication signals, and preprocessing the signals to obtain a time-frequency spectrum diagram of uniform size, and using the time-frequency spectrum diagram to generate a signal sample database for model training;
[0031] S102, constructing a signal detection model based on a centripetal shift network, and training the signal detection model using signal data in a signal sample database, wherein the signal detection model includes: an hourglass basic network for acquiring multi-scale features by performing up- and down-sampling operations on input data, a feature processing module for acquiring corner prediction results and signal depth features by performing corner pooling and deformable convolution on the multi-scale features, and a centripetal shift module for locating the corner prediction results and the depth features using a corner matching algorithm to output the target corner position;
[0032] S103 , preprocessing the broadband wireless communication signal to be detected, using the preprocessed signal data as a model input, and identifying the target position of the broadband wireless communication signal to be detected by using the trained signal detection model.
[0033] The hourglass basic network is combined with deformable convolution to perform deep feature extraction on the input signal sample data, and the signal target is accurately located and detected through a corner point matching algorithm based on centripetal shift. This can solve the problems of low positioning accuracy of signal targets under low signal-to-noise ratio conditions, high probability of missed detection when severe time-frequency aliasing occurs in broadband signals, and serious degradation of wireless communication signal target detection effect when color noise exists. The implementation process has clear steps, is easy to structure and apply, and improves the signal detection effect.
[0034] As a method for detecting broadband wireless communication signals in a complex environment in an embodiment of the present invention, further, a standard broadband wireless communication signal for constructing a signal sample database is generated according to a random symbol sequence and a signal modulation method. Further, when generating a standard broadband wireless communication signal, a number of narrowband signals are randomly selected, and the narrowband signal and the generated standard broadband wireless communication signal are aliased according to a preset aliasing probability; and α-stable noise is added to the standard broadband wireless communication signal. Further, in signal aliasing, a cosine wave with the same bandwidth as the narrowband signal is used as a fading channel, and the fading intensity of the fading channel and the termination phase of the fading channel are randomly selected, and Gaussian noise is used to simulate the actual channel environment to generate broadband mixed multi-signal sample data.
[0035] See also Figure 2As shown, according to several selected specific modulation modes and random symbol sequences, corresponding standard wireless communication signals are generated. After the baseband signal is generated, when the frequency is loaded, each broadband signal sample randomly selects several narrowband signals, and randomly aliases the frequency and time with other signals. The degree of aliasing can be randomly selected between 20% and 100%, and the preset probability of sample aliasing can be set to 0.7, that is, 30% of the signal samples are aliased and 70% of the samples are aliased, which is close to the actual application scenario. Subsequently, α-stable noise is added to the broadband signal to further meet the non-Gaussian noise conditions with obvious spike pulse characteristics that are common in actual wireless communication channels. The characteristic function of the added α-stable noise can be expressed as:
[0036] χ(t)=exp{jδt-γ|t| α [1+jβsgn(t)ω(t,α)]} (1)
[0037] in,
[0038]
[0039]
[0040] Where 0<α≤2, -1≤β≤1, γ>0, -∞<δ<∞. In the embodiment of this case, the parameters used can be designed as follows: the characteristic index α is 1.3, the symmetry parameter β is 1, the proportional parameter γ is 0.025, and the position parameter δ is 0.1. In a complex environment, wireless communication signals will fade during transmission and reception. In this case, a cosine wave equal to the bandwidth of each narrowband signal can be used as a fading channel. The fading intensity of the channel is randomly selected between 40% and 100%, and the termination phase of the fading channel is randomly selected between π and 2π. The probability of fading for each broadband signal sample is 80%, which further increases the richness of the sample. Finally, Gaussian noise is used to simulate the actual channel environment, which can be specifically expressed as:
[0041] s awgn (t) = s br (t)+v(t) (4)
[0042] Where s awgn (t) is the sample signal after adding Gaussian noise, s br (t) is a broadband mixed multi-signal sample without adding Gaussian noise.
[0043] As a broadband wireless communication signal detection method under a complex environment in an embodiment of the present invention, further, a deep feature extraction module in the signal detection model includes: a corner point pooling unit for performing corner point pooling on multi-scale feature data, a corner point prediction unit for performing corner point prediction processing on a branch of a corner point pooling result, and a feature learning unit for performing local feature learning on another branch of a corner point pooling result through a cross-star variable convolution operation, wherein the corner point prediction unit and the feature learning unit perform corresponding corner point prediction processing and local feature learning on the lower right corner and the upper left corner of the feature map, respectively. Further, in the feature learning unit, first, convolution is performed on another branch of the corner point pooling result to obtain an offset field; then, the offset field and another branch of the corner point pooling result data are subjected to variable convolution to extract signal depth features. Furthermore, in the centripetal shift module, based on the consistency of the center points of the lower right corner and the upper left corner of the same target, the centripetal shift algorithm is used to calculate the center point coordinates through the corner points, and the quantization weights of the corner point pairs in the corner point prediction results are set according to the degree of proximity between the target signal center and the regression center of each corner point. The corner point pairs are re-scored using the original scores and quantization weights obtained by corner point pooling for each group of corner point pairs, and the corner point pairs that are used as the corner points of the target signal position are determined based on the size of the score.
[0044] like Figure 3 As shown in the figure, considering the structure and parameters of the signal network model, the size of the time-frequency spectrum of the input network can be set to 512 pixels × 512 pixels. When the time-frequency spectrum sample of the signal is input into the network, the hourglass basic network (HN) is first used to perform downsampling and upsampling operations to obtain the feature map of the broadband signal sample, and then sent to the corner prediction and feature adaptation modules of the lower right corner and upper left corner for processing. Since the network model performs the same processing on the lower right corner and the upper left corner, Figure 3 Only the processing module structure of the upper left corner point is shown in the figure. In the actual process, after the signal feature map is corner pooled (CP), several cross-stars (CS) will be left in the local area of the feature map, where the feature extraction is not comprehensive but contains a lot of relevant information in this local area. Therefore, the cross-star deformable convolution (CDC) is used to learn the geometric structure near the cross-star of the signal feature map. Figure 3 As shown in , the cross-star deformable convolution module obtains the offset field (OF) by using three convolutions, and then convolves the obtained offset field with the original feature map to implement the cross-star deformable convolution processing.
[0045] For a certain signal target, the network model will detect several groups of corner point pairs representing the lower right corner and upper left corner positions of the target, such as Figure 4 As shown in the figure, let the lower right corner be (brx, bry) and the upper left corner be (tlx, tly), then the range of the predicted signal target can be expressed as:
[0046] bbox=(tlx,tly,brx,bry) (5)
[0047] The center point coordinates are:
[0048]
[0049] For two corner points belonging to the same target, their corresponding center points should be the same, so the centripetal offset algorithm is used to calculate the corresponding center point coordinates from the corner points, and the proximity of the center points after regression is used to select each pair of corner points. The centripetal offsets of the lower right corner and the upper left corner are defined as:
[0050]
[0051] Based on:
[0052] R central ={(x,y)∣x∈[ctlx,cbrx],y∈[ctly,cbry]} (8)
[0053] Define the central area of each signal target to compare the closeness between the signal target center and the regression center of each corner point, such as Figure 5 As shown. central The four corner positions:
[0054]
[0055] In the formula, the scaling factor is generally taken as 0<μ≤1, and the center point corresponding to the lower right corner point can be obtained by centripetal shift (br ctx ,br cty ) and the center point corresponding to the upper left corner (tl ctx ,tl cty ) position, and then calculate to satisfy:
[0056] (tl ctx ,tl cty )∈R central ∩(br ctx ,br cty )∈R central (10)
[0057] Quantized weights of corner point pairs:
[0058]
[0059] In the formula, the numerator is the absolute value of the product of the horizontal and vertical coordinate differences of the candidate corner point pair regression center point (light-colored area), and the denominator is the area of the fixed area (dark-colored area). If the center points of the two corner point regressions are closer, the calculated weight value will be larger. Subsequently, the original score obtained by corner point pooling for each group of corner point pairs is multiplied by the above weight, and the predicted corner point pairs are re-scored and selected, and the corner point pairs are finally retained as the position corner points of the target signal.
[0060] Signal detection models can use multi-scale training strategies during training, such as Figure 6 As shown in the figure, during the training process, there is a 50% probability that each signal spectrum sample will be subjected to multi-scale transformation processing. The width and height of the sample will be scaled to 0.5 to 2 times the original size. At the same time, the corresponding target label annotation is generated, and then the scale-transformed samples are used for training, so that the network model can process sample images within this size ratio range, and the generalization ability is stronger.
[0061] At the same time, in the detection model training stage, the quantification tool for the training loss is the training loss function. In this embodiment, the improved Smooth L1 loss function is used to calculate the centripetal shift loss (CSL) and guided shift loss (GSL) of signal position positioning:
[0062]
[0063] Where y is the true value representing the signal target position information, is the corresponding predicted value. When in the early stage of training, the gap between the predicted value and the true value Generally, it is relatively large. In this case, the original Smooth L1 loss function is used to make the network model insensitive to abnormal samples of the signal, so that the gradient decreases in the effective direction. , it means that the detection model has reached the late stage of training. In this stage, the logarithmic hyperbolic cosine function is used instead. This function is also quadratically differentiable. Compared with the original Smooth L1 loss function Its curve is smoother and its value is smaller, which is conducive to fine-tuning the network node parameters during back-propagation operations in the later stage of training, and facilitates the detection model to find network node weight parameters with better convergence.
[0064] Based on the above training strategy and the design of the model loss function, the specific process of this solution in practical application can be found in Figure 7As shown in the figure, first, the broadband wireless communication signal sample is preprocessed to obtain a time-frequency spectrum of size 512×512, and a training signal data set is generated. Then, the centripetal shift network model is fully trained using the training signal data set until the detection error loss value of the network model is reduced to below the preset threshold, the training is terminated and the network model is saved. An actual broadband communication signal time-frequency spectrum sample is input, and multiple scale feature data are extracted using the hourglass basic network, followed by corner pooling. The result of corner pooling is output to two branches. The upper branch performs corner prediction processing, and the lower branch learns the local feature information of the feature map, which is realized by the cross-star deformable convolution operation. The operation results of the cross-star deformable convolution and corner prediction are input into the centripetal shift module to carry out corner pair selection. The centripetal shift adaptation algorithm is used to recalculate the weights of each corner pair and perform selection operations to obtain the final detection target detection result and output it, which provides a reliable and effective implementation scheme for broadband wireless communication signal detection in complex environments, while ensuring detection efficiency, it can improve the signal target detection accuracy.
[0065] Furthermore, based on the above method, an embodiment of the present invention also provides a broadband wireless communication signal detection system in a complex environment, comprising: a sample collection module, a model building module and a signal recognition module, wherein:
[0066] A sample collection module is used to collect broadband wireless communication signals and pre-process the signals to obtain a time-frequency spectrum diagram of uniform size, and use the time-frequency spectrum diagram to generate a signal sample database for model training;
[0067] A model construction module, used to construct a signal detection model based on a centripetal shift network, and train the signal detection model using signal data in a signal sample database, wherein the signal detection model includes: an hourglass basic network for obtaining multi-scale features by performing up- and down-sampling operations on input data, a feature processing module for obtaining corner prediction results and signal depth features by performing corner pooling and deformable convolution on multi-scale features, and a centripetal shift module for locating corner prediction results and depth features using a corner matching algorithm to output a target corner position;
[0068] The signal recognition module is used to preprocess the broadband wireless communication signal to be detected, use the preprocessed signal data as the model input, and identify the broadband wireless communication signal target to be detected through the trained signal detection model.
[0069] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0070] Based on the above method and / or system, an embodiment of the present invention also provides a server, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the above method.
[0071] Based on the above method and / or system, an embodiment of the present invention further provides a computer-readable medium having a computer program stored thereon, wherein the above method is implemented when the program is executed by a processor.
[0072] In all examples shown and described herein, any specific values should be interpreted as merely exemplary and not as limiting, and thus other examples of the exemplary embodiments may have different values.
[0073] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0074] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. A method for detecting broadband wireless communication signals in a complex environment, characterized in that: Contains the following: The collected broadband wireless communication signals are preprocessed to obtain a time-frequency spectrum diagram of uniform size, and the time-frequency spectrum diagram is used to generate a signal sample database for model training; A signal detection model based on a centripetal shift network is constructed, and the signal detection model is trained using signal data in a signal sample database, wherein the signal detection model comprises: an hourglass basic network for obtaining multi-scale features by performing up- and down-sampling operations on input data, a feature processing module for obtaining corner prediction results and signal depth features by performing corner pooling and deformable convolution on multi-scale features, and a centripetal shift module for locating the corner prediction results and depth features using a corner matching algorithm to output the target corner position; the feature processing module comprises: a corner pooling unit for performing corner pooling on multi-scale feature data, a corner prediction unit for performing corner prediction processing on a branch of a corner pooling result, and a feature learning unit for performing local feature learning on another branch of a corner pooling result by cross-star deformable convolution operations, and the corner prediction unit and the feature learning unit respectively perform corresponding corner prediction processing and local feature learning on the lower right corner and the upper left corner of the feature map; The broadband wireless communication signal to be detected is preprocessed, the preprocessed signal data is used as a model input, and the target position of the broadband wireless communication signal to be detected is identified through the trained signal detection model.
2. The method for detecting broadband wireless communication signals in a complex environment according to claim 1, characterized in that: A standard broadband wireless communication signal for constructing a signal sample database is generated according to a random symbol sequence and a signal modulation method.
3. The method for detecting broadband wireless communication signals in a complex environment according to claim 2, characterized in that: When generating a standard broadband wireless communication signal, a number of narrowband signals are randomly selected, and the narrowband signals and the generated standard broadband wireless communication signal are aliased according to a preset aliasing probability; and stable noise is added to the standard broadband wireless communication signal.
4. The method for detecting broadband wireless communication signals in a complex environment according to claim 1, characterized in that: In signal aliasing, a cosine wave with the same bandwidth as the narrowband signal is used as the fading channel, and the fading intensity of the fading channel and the termination phase of the fading channel are randomly selected. Gaussian noise is used to simulate the actual channel environment to generate broadband mixed multi-signal sample data.
5. The method for detecting broadband wireless communication signals in a complex environment according to claim 1, characterized in that: In the feature learning unit, first, the other branch of the corner pooling result is convolved to obtain the offset field; then, the offset field and the other branch data of the corner pooling result are variably convolved to extract the signal depth features.
6. The method for detecting broadband wireless communication signals in a complex environment according to claim 1, characterized in that: In the centripetal shift module, the centripetal shift algorithm is used to calculate the center point coordinates through the corner points according to the consistency of the lower right corner and upper left corner center points of the same target. The quantization weights of the corner point pairs in the corner point prediction results are set according to the degree of proximity between the signal target center and the regression center of each corner point. The corner point pairs are re-scored based on the original scores and quantization weights obtained by corner point pooling for each group of corner point pairs, and the corner point pairs as the corner points of the signal target position are determined according to the size of the score.
7. A broadband wireless communication signal detection system in a complex environment, characterized in that: It includes: sample collection module, model building module and signal recognition module, among which, A sample collection module is used to collect broadband wireless communication signals and pre-process the signals to obtain a time-frequency spectrum diagram of uniform size, and use the time-frequency spectrum diagram to generate a signal sample database for model training; A model construction module, for constructing a signal detection model based on a centripetal shift network, and training the signal detection model using signal data in a signal sample database, wherein the signal detection model comprises: an hourglass basic network for obtaining multi-scale features by performing up- and down-sampling operations on input data, a feature processing module for obtaining corner prediction results and signal depth features by performing corner pooling and deformable convolution on multi-scale features, and a centripetal shift module for locating corner prediction results and depth features using a corner matching algorithm to output target corner positions; the feature processing module comprises: a corner pooling unit for performing corner pooling on multi-scale feature data, a corner prediction unit for performing corner prediction processing on a branch of a corner pooling result, and a feature learning unit for performing local feature learning on another branch of a corner pooling result by performing a cross-star deformable convolution operation, and the corner prediction unit and the feature learning unit respectively perform corresponding corner prediction processing and local feature learning on the lower right corner and the upper left corner of the feature map; The signal recognition module is used to preprocess the broadband wireless communication signal to be detected, use the preprocessed signal data as model input, and identify the target position of the broadband wireless communication signal to be detected through the trained signal detection model.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.
