A method for extracting frequency-hopping signal parameters based on a lightweight target detection model
By optimizing the lightweight MobileNet-SSD model and the joint pruning algorithm, the problems of slow frequency hopping signal detection speed and insufficient anti-interference capability due to limited hardware resources are solved, achieving fast and accurate extraction of frequency hopping signal parameters and reducing storage costs.
Patent Information
- Application Number
- CN202410446616.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-15
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-04-15
AI Technical Summary
Existing deep learning-based frequency hopping signal detection and parameter estimation schemes are slow inference speeds under limited hardware resources and struggle to accurately extract frequency hopping signal parameters in real communication scenarios containing interference and noise.
A lightweight MobileNet-SSD model is used in conjunction with a joint pruning algorithm and an online quantization training method to generate time-frequency maps of various types of interference and noise signals. Through multi-scale target detection, a lightweight single-stage model is trained to reduce the number of parameters and improve anti-interference capability.
It achieves rapid extraction of frequency hopping signals on devices with limited hardware resources, enhances the model's anti-interference capability, maintains high detection accuracy, and reduces storage requirements, making it suitable for deployment on mobile or embedded devices.
Smart Images

Figure CN118332401B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to frequency hopping signal detection technology, specifically to a method for extracting frequency hopping signal parameters based on a lightweight target detection model. Background Technology
[0002] Frequency hopping (FHH) is a modulation technique characterized by frequent frequency changes in communication. By transmitting data at different frequencies, FHH effectively resists interference and eavesdropping, improving the security and stability of communication systems. It is widely used in military communications, radar systems, and wireless sensor networks, providing an efficient and reliable solution for radio communication. FHH enhances the ability of communication systems to resist interference and eavesdropping through frequency switching, thus holding an important position and widespread application in modern communications.
[0003] Blind reception and parameter extraction of frequency-hopping signals have important applications in communication and electronic warfare. Most existing solutions extract parameters from frequency-hopping signals based on time-frequency analysis. Since the time-frequency graph of a signal can be viewed as an image, using image processing algorithms to analyze the time-frequency graph of frequency-hopping signals is an effective strategy and method. Thanks to the rapid development of deep learning, image processing tasks based on deep learning models, such as image classification, object detection, and image segmentation, have made rapid progress. In the field of object detection, early representative deep models used deep convolutional neural networks to process images end-to-end, eliminating the need for manually designed feature extractors and achieving good detection performance. Currently, object detection models are generally divided into single-stage and two-stage detectors. Single-stage detectors directly regress the location of bounding boxes and the object category in a single network, resulting in faster inference speeds. Two-stage detectors, on the other hand, first generate candidate regions and then perform classification and location regression. With the increasing demand for deep learning models in mobile devices and embedded systems, lightweight object detection models have also attracted much attention. These models reduce computational complexity and memory usage while maintaining high detection accuracy through the design of lightweight network structures and optimized algorithms.
[0004] Existing deep learning-based frequency hopping signal detection and parameter estimation schemes mostly use models with a large number of parameters. While these models can achieve high detection accuracy, their large parameter count leads to slow inference speeds in scenarios with limited hardware resources. Because the frequency hopping speed of signals currently used in communication is very fast, and the hardware resources of receiver digital signal processors are limited, these large-parameter models struggle to achieve rapid extraction of frequency hopping signal parameters under limited computing resources. Furthermore, current deep learning algorithms for estimating frequency hopping signal parameters are limited to parameter estimation under interference-free and noise-free conditions. However, in real-world communication scenarios, received signals typically contain multiple signal types, requiring algorithms capable of extracting frequency hopping signal parameters from interference signals. Summary of the Invention
[0005] To address the problems of existing technologies, this invention proposes a method for extracting frequency-hopping signal parameters based on a lightweight target detection model. This method generates time-frequency maps containing various types of interference signals, noise signals, and frequency-hopping signals, and trains a lightweight MobileNet-SSD (Mobilenet-SSD) model. This model employs depthwise separable convolution, significantly reducing the number of model parameters. Furthermore, by combining a joint pruning algorithm and an online quantization training method, the model is further compressed. This significantly reduces the number of model parameters and improves detection speed while maintaining almost no decrease in frequency-hopping signal detection accuracy, and also enhances the detection's anti-interference capability.
[0006] The frequency hopping signal parameter extraction method based on a lightweight target detection model of the present invention includes the following steps:
[0007] 1) Preparing the dataset:
[0008] Based on the prior spectral distribution of the frequency hopping signal in the communication reception scenario, as well as the categories of interference and noise signals, multiple sets of training signals are generated; each set of training signals includes interference signals, noise signals, and a set frequency hopping signal.
[0009] Perform time-frequency analysis on each generated training signal to obtain the corresponding time-frequency map; on each time-frequency map, label the regions where the time-frequency patterns of various types of signals are located using rectangular bounding boxes and record the coordinates and corresponding signal category labels; save the time-frequency maps and the bounding boxes, coordinates and corresponding signal category labels contained in each time-frequency map to form a dataset;
[0010] 2) Enhance the training set:
[0011] The dataset is divided into three parts according to a set ratio: training set, validation set, and test set, in order to train and validate the performance of the model.
[0012] The training set is augmented by operations including translation, noise injection, and brightness and contrast adjustment, thereby improving the generalization ability and noise resistance of the training set.
[0013] 3) Iterative training of the model:
[0014] A convolutional neural network is constructed using the MobileNet-SSD model. The input is a time-frequency graph, and the output is a time-frequency graph containing bounding boxes, coordinates, and corresponding signal category labels.
[0015] When detecting targets, only interference signals and frequency hopping signals of each category are considered, and noise signals are not detected. The number of signal categories for target detection is the number of interference signal categories plus 1. The optimizer, learning rate, batch size, and objective function are set.
[0016] The MobileNet-SSD model is iteratively trained using a training set, with internal parameters adjusted through multiple iterations to minimize the loss function. During training, a validation set is used to verify the effectiveness of each iteration and check the performance of the MobileNet-SSD model. Training ends when the expected detection accuracy is achieved and convergence is achieved. After training, a test set is used to test the model, resulting in a well-trained MobileNet-SSD model. The trained MobileNet-SSD model is then saved.
[0017] 4) Compression model based on joint pruning algorithm:
[0018] The joint pruning algorithm uses L2 norm pruning, batch normalized layer parameter pruning, and similarity pruning to evaluate the importance of each convolutional kernel in each convolutional layer of the convolutional neural network, and obtains the convolutional kernel loss-pruning percentage curve for each convolutional layer under each pruning algorithm.
[0019] The convolutional kernel loss-pruning percentage curves obtained by the L2 norm pruning algorithm, the batch normalized layer parameter pruning algorithm, and the similarity pruning algorithm are compared. The pruning algorithm corresponding to the lowest curve, that is, the pruning algorithm corresponding to the curve with the lowest average value of the ordinate, is selected as the optimal pruning strategy for this layer. The optimal pruning strategy is used to prune the convolutional layer of this layer.
[0020] A joint pruning algorithm is used to prune each convolutional layer of the convolutional neural network, thereby compressing the model and obtaining a compressed quantized model.
[0021] 5) Online quantization training model:
[0022] In the compressed quantization model, the output value and weights of each convolutional layer are quantized, and the output value of each convolutional layer is dequantized. The compressed quantization model is further trained using the dataset from step 2) to restore the original accuracy. Furthermore, while ensuring the model accuracy, the weights of the compressed quantization model are converted from floating-point numbers to integers to further reduce the hardware overhead of the model and improve the inference speed, thus obtaining a well-trained compressed quantization model.
[0023] 6) Detect frequency-hopping targets and extract parameters:
[0024] The trained compression quantization model is used to infer the extracted sample signals and detect the time-frequency patterns of frequency hopping signals and various types of interference signals contained in each time-frequency plot.
[0025] Extract the midpoint coordinates from the time-frequency pattern of the frequency-hopping signal;
[0026] After obtaining the midpoint of each frequency hopping signal, the abscissa of the midpoint of the frequency hopping signal is differentially divided to obtain a differential sequence. The average value of the differential sequence is used to obtain the frequency hopping period. The center frequency of the frequency hopping signal is obtained according to the ordinate of the midpoint of each frequency hopping signal. The center time point in the original time-frequency diagram is obtained according to the abscissa of the midpoint of each frequency hopping signal. The 3dB spectral width of the center time point is obtained on the spectral distribution curve corresponding to the center time point, which is the frequency hopping bandwidth.
[0027] In step 1), the time-frequency analysis employs one of the following: short-time Fourier transform, wavelet transform, or short-time fractional Fourier transform. Interference signals include multiple categories: linear sweep frequency signals, nonlinear sweep frequency signals, and fixed-frequency interference signals; the signal category labels are linear sweep frequency signals, nonlinear sweep frequency signals, fixed-frequency interference signals, noise signals of various categories, and frequency-hopping signals. Noise signals are not detected later; the noise signal categories include Gaussian noise and Rayleigh noise. The dataset contains 6000–8000 time-frequency plots.
[0028] In step 2), the ratio of the dataset to the training set, validation set, and test set is 6-8:1-3:1.
[0029] Translation: Randomly translate the time and frequency axes of the time-frequency plot to simulate the time and frequency offsets of various signals. Noise injection: Add random noise to the time-frequency plot to simulate interference in a real communication environment. Brightness and contrast adjustment: Change the brightness and contrast of the time-frequency plot to simulate different signal reception conditions.
[0030] In step 3), the objective function is obtained by weighting the class loss and localization loss, using an adaptive inertia estimation (Adam) optimizer with a learning rate of 0.01–0.001 and a batch size of 16–64. During training, a validation set is used to verify the effect of each iteration. Training ends when the expected detection accuracy of 70%–80% is achieved.
[0031] In step 4), the L2 norm pruning algorithm evaluates convolutional kernels based on the L2 norm of their weights; kernels with larger L2 norms are more important. The batch normalized layer parameter pruning algorithm evaluates convolutional kernels based on the γ parameter of the corresponding batch normalized layer; kernels with larger γ parameters are more important. The similarity pruning algorithm assesses the similarity between convolutional kernels based on their distance from the central kernel; kernels with lower similarity are more important. In each pruning algorithm, the importance of the convolutional kernels is ranked, and the lowest-ranked kernels are removed in order of ranking. The convolutional kernel loss-pruning percentage curve is obtained according to the removal order.
[0032] In step 5), the weights of the compressed quantization model are converted from 16, 32, or 64-bit floating-point numbers to 8-bit integers.
[0033] In step 6), the sampling sequence index of the center frequency point is obtained based on the ordinate of the midpoint of each frequency hopping signal, and the sampling sequence index is multiplied by the sampling frequency to obtain the center frequency point of the frequency hopping signal.
[0034] Advantages of this invention:
[0035] This invention proposes an efficient and low-resource-consumption solution for frequency-hopping signal detection and parameter estimation by employing a lightweight MobileNet-SSDv2 model combined with a joint pruning algorithm and an online quantization training method. Compared with traditional deep learning models and algorithms, the advantages of this invention are mainly reflected in the following aspects:
[0036] (1) Improve the inference speed and real-time performance of the model: By designing and optimizing the MobileNet-SSD model in a lightweight manner, the number of parameters and computational complexity of the model are significantly reduced, enabling the model to run efficiently on devices with limited hardware resources and meet the real-time processing requirements of rapid frequency hopping signal conversion.
[0037] (2) Enhanced the model’s anti-interference ability: This method introduces various types of interference signals and noise signals in the dataset preparation stage, thereby enhancing the robustness of the model and its adaptability to complex communication environments; in addition, by performing joint pruning and online quantization training on the model, the generalization ability of the model is further optimized, enabling it to extract frequency hopping signal parameters from interference signals more accurately.
[0038] (3) Reduced model size and lowered storage and deployment costs: The joint pruning algorithm greatly reduces the model's storage requirements by removing unimportant parameters and features and converting model weights from floating-point numbers to integers through online quantization training. This makes the model of this invention more suitable for deployment on mobile or embedded devices with limited memory and storage space.
[0039] (4) Maintaining high detection accuracy: Although the model has been compressed and quantized, through carefully designed training strategies and algorithm optimization, the model of this invention can still maintain high detection accuracy close to the original model, effectively realizing accurate detection and parameter extraction of frequency hopping signals. Attached Figure Description
[0040] Figure 1 This is a flowchart of an embodiment of the frequency hopping signal parameter extraction method based on a lightweight target detection model of the present invention;
[0041] Figure 2 This is a time-frequency diagram obtained from an embodiment of the frequency hopping signal parameter extraction method based on a lightweight target detection model according to the present invention;
[0042] Figure 3 This is a flowchart of a joint pruning algorithm in an embodiment of the frequency hopping signal parameter extraction method based on a lightweight target detection model of the present invention;
[0043] Figure 4 This is a structural diagram of a convolutional kernel layer after the insertion of quantization and dequantization nodes, according to an embodiment of the frequency hopping signal parameter extraction method based on a lightweight target detection model of the present invention.
[0044] Figure 5 This is a video image of a frequency hopping signal extracted according to an embodiment of the frequency hopping signal parameter extraction method based on a lightweight target detection model according to the present invention. Detailed Implementation
[0045] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0046] like Figure 1 As shown, the frequency hopping signal parameter extraction method based on a lightweight target detection model in this embodiment includes the following steps:
[0047] 1) Preparing the dataset:
[0048] Based on the prior spectral distribution of frequency-hopping signals in the communication reception scenario, as well as the categories of interference signals and noise signals, multiple sets of training signals are generated. Each set of training signals includes interference signals, noise signals, and a set frequency-hopping signal. In this embodiment, the interference signals are classified as fixed-frequency interference signals and linearly swept-frequency interference signals, and the noise signals are classified as Gaussian noise. The center frequency distribution of the frequency-hopping signals is set to an integer uniform random distribution of [3, 20] Hz, with a total of eight frequency-hopping patterns. The frequency-hopping period is a uniform random distribution of [30, 60] ms. The upper limit frequency of the linearly induced frequency interference signal is set to a uniform random distribution of [1.6, 1.7] kHz, and the lower limit frequency is set to a uniform random distribution of [0.9, 1.1] kHz. The frequency point of the fixed-frequency interference signal is set to a uniform random distribution of [0.9, 1] kHz, and the sampling rate is 4 kHz. 8000 sets of training signals are randomly generated.
[0049] Time-frequency analysis was performed on each of the generated training signals to obtain 8000 time-frequency maps. On each time-frequency map, the regions where the time-frequency patterns of various types of signals were located were labeled with rectangular bounding boxes, and the coordinates and corresponding signal category labels were recorded. The time-frequency maps, as well as the bounding boxes, coordinates and corresponding signal category labels contained in each time-frequency map, were saved to form a dataset containing 8000 time-frequency maps. Figure 2 The image shows time-frequency plots of a set of training signals after time-frequency analysis, with each category of signal labeled. Figure 2 The shorter horizontal lines represent the time-frequency pattern of frequency hopping signals, the slanted lines represent the time-frequency pattern of linear frequency sweep interference signals, the long horizontal lines represent the time-frequency pattern of fixed-frequency interference signals, and the thin boxes represent labels.
[0050] 2) Enhance the training set:
[0051] The dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio to train and validate the model's performance.
[0052] Enhancement operations are performed on the training set, including translation, noise injection, and brightness and contrast adjustment, to improve the generalization ability and noise resistance of the training set. Among them, translation: randomly translates the time axis and frequency axis of the time-frequency plot to simulate the offset of various signals in time and frequency; noise injection: adds random noise to the time-frequency plot to simulate interference in the actual communication environment; brightness and contrast adjustment: changes the brightness and contrast of the time-frequency plot to simulate different signal reception conditions.
[0053] 3) Iterative training of the model:
[0054] A convolutional neural network is constructed using the MobileNet-SSD model. The input is a time-frequency graph, and the output is a time-frequency graph containing bounding boxes, coordinates, and corresponding signal category labels.
[0055] During target detection, only interference signals and frequency-hopping signals of each category are considered, while noise signals are not detected. The number of signal categories for target detection is the number of interference signal categories plus 1. In this embodiment, the categories of interference signals detected are fixed-frequency interference signals and linear frequency-sweeping interference signals, with a total of 2 interference signal categories. The number of signal categories for target detection is 2 plus 1, i.e., the number of signal categories for target detection is 3. The signals detected by the target are frequency-hopping signals, fixed-frequency interference signals, and linear frequency-sweeping interference signals. The optimizer, learning rate, batch size, and objective function are set as follows: the Adam optimizer is used, the initial learning rate is set to 0.0001, the batch size is set to 64, and the maximum number of epochs is set to 200.
[0056] The MobileNet-SSD model is iteratively trained using the training set, with internal parameters continuously adjusted through multiple iterations to minimize the loss function. Every 10 epochs during training, the model's loss and average accuracy are validated using the validation set to ensure no overfitting and to check the MobileNet-SSD model's performance, thereby monitoring overfitting and adjusting the learning rate. Training ends when convergence is achieved. After convergence, the model is tested using the test set to obtain the trained MobileNet-SSD model. The trained MobileNet-SSD model is then saved.
[0057] 4) Compression model based on joint pruning algorithm:
[0058] The joint pruning algorithm employs L2 norm pruning, batch normalized layer parameter pruning, and similarity pruning to evaluate the importance of each convolutional kernel in each convolutional layer of the convolutional neural network, respectively, and calculates the importance of each kernel under various conditions.
[0059] The kernel loss versus pruning percentage curve for each convolutional layer under the pruning algorithm, as shown below. Figure 3 As shown:
[0060] The L2 norm is first used to evaluate the k-th convolutional layer. The formula for the L2 norm of each convolutional kernel is as follows:
[0061]
[0062] Where W represents the convolution kernel tensor, w i represents the i-th weight of each convolutional kernel, and N represents the total number of parameters of each convolutional kernel, i = 1, ..., N; sort the convolutional kernels in descending order according to the size of the L2 norm, and successively subtract convolutional kernels from 10% to 90%, test their corresponding loss values on the validation set, and plot the loss-pruning percentage curve.
[0063] Similarity evaluation was performed using the γ parameter and similarity parameters respectively, in a similar manner. The γ parameter is the parameter of the batch normalization layer.
[0064] The formula is as follows:
[0065]
[0066]
[0067] Where, x i μ represents the activation value of the input. B and These represent the batch mean and batch variance, respectively; it can be seen that the γ parameter represents the scaling of the batch normalization layer on the input activation values. The larger the value of γ, the more important the activation values of that layer are.
[0068] Similarity evaluation is based on the Euclidean distance of convolutional kernels. First, the distance from each convolutional kernel to other convolutional kernels is calculated. These distances are then summed, and the convolutional kernel with the smallest distance is selected as the center kernel. Then, similarity is evaluated based on the distance from each convolutional kernel to the center kernel, with similar convolutional kernels being removed first.
[0069] The convolutional kernel loss-pruning percentage curves obtained by the L2 norm pruning algorithm, the batch normalized layer parameter pruning algorithm, and the similarity pruning algorithm are compared. The pruning algorithm corresponding to the lowest curve, that is, the pruning algorithm corresponding to the curve with the lowest average value of the ordinate, is selected as the optimal pruning strategy for this layer. The optimal pruning strategy is used to prune the convolutional layer of this layer.
[0070] A joint pruning algorithm is used to prune each convolutional layer of the convolutional neural network, thereby compressing the model.
[0071] The compressed quantization model is obtained;
[0072] 5) Online quantization training model:
[0073] In the compressed quantization model, the output value and weights of each convolutional layer are quantized, and the output value of each convolutional layer is dequantized: For example... Figure 4As shown, the output of the previous layer is the activation value. The activation value output from the previous layer is quantized, and the weights, which exist in floating-point form, are also quantized and input into the current convolutional layer for depthwise separable convolution. The output of the convolutional layer is then input into a batch normalization layer for batch normalization, and the output value is dequantized to obtain the activation value of the current layer. The dataset from step 2) is used to further train the compressed quantization model, thereby restoring the original accuracy. Furthermore, while maintaining model accuracy, the weights of the compressed quantization model are converted from 16, 32, or 64-bit floating-point numbers to 8-bit integers to further reduce the hardware overhead of the model and improve the model's inference speed, resulting in a well-trained compressed quantization model. The operation of the quantization node can be expressed as the following formula:
[0074]
[0075] Where Q(x) is the quantized value, x is the original value to be quantized (which can be a weight or activation), S is a scaling factor used to map the original value to the quantization range, and this parameter is a learnable parameter during training; Z is the zero point.
[0076] Used to ensure that the zero value after quantization corresponds to the original zero value, especially when quantizing to an asymmetric range;
[0077] The operation of the dequantization node can be expressed as the following formula:
[0078] x approx = (Q(x)-Z)×S
[0079] Online quantization training methods introduce quantization and dequantization operations during the training process, enabling the model to learn the optimal parameters under quantization constraints.
[0080] 6) Frequency-hopping target detection and parameter extraction:
[0081] The trained compression quantization model is used to infer the extracted sample signals and detect the time-frequency patterns of frequency hopping signals and various types of interference signals contained in each time-frequency plot.
[0082] Extract the midpoint coordinates from the time-frequency pattern of the frequency-hopping signal;
[0083] After obtaining the midpoints of each frequency-hopping signal, the x-coordinates of the midpoints are differentially analyzed to obtain a differential sequence. The averaging of the differential sequence yields the frequency-hopping period. The center frequency of each frequency-hopping signal is obtained based on the y-coordinates of its midpoints. The sampling sequence index of the center frequency is obtained from the y-coordinates of its midpoints. Multiplying the sampling sequence index by the sampling frequency gives the center frequency of the frequency-hopping signal. The 3dB spectral width at the center time point is calculated on the spectral distribution curve corresponding to the center time point, thus obtaining the frequency-hopping bandwidth. The background is removed from the image, and the time-frequency graph containing only the frequency-hopping signals is extracted as shown below. Figure 5 As shown.
[0084] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.
Claims
1. A method for extracting frequency-hopping signal parameters based on a lightweight target detection model, characterized in that, The frequency hopping signal parameter extraction method includes the following steps: 1) Preparing the dataset: Based on the prior spectral distribution of the frequency hopping signal in the communication reception scenario, as well as the categories of interference and noise signals, multiple sets of training signals are generated; each set of training signals includes interference signals, noise signals, and a set frequency hopping signal. Perform time-frequency analysis on each generated training signal to obtain the corresponding time-frequency map; on each time-frequency map, label the regions where the time-frequency patterns of various types of signals are located using rectangular bounding boxes and record the coordinates and corresponding signal category labels; save the time-frequency maps and the bounding boxes, coordinates and corresponding signal category labels contained in each time-frequency map to form a dataset; 2) Enhance the training set: The dataset is divided into three parts according to a set ratio: training set, validation set, and test set, in order to train and validate the performance of the model. The training set is augmented by operations including translation, noise injection, and brightness and contrast adjustment, thereby improving the generalization ability and noise resistance of the training set. 3) Iterative training of the model: A convolutional neural network is established, which adopts the MobileNet-SSD model for single-stage multi-scale object detection. The input is a time-frequency map, and the output is a time-frequency map containing bounding boxes, coordinates, and corresponding signal category labels. When detecting targets, only interference signals and frequency hopping signals of each category are considered, and noise signals are not detected. The number of signal categories for target detection is the number of interference signal categories plus 1. The optimizer, learning rate, batch size, and objective function are set. The MobileNet-SSD model is iteratively trained using a training set, with internal parameters adjusted through multiple iterations to minimize the loss function. During training, a validation set is used to verify the effectiveness of each iteration and check the performance of the MobileNet-SSD model. Training ends when the expected detection accuracy is achieved and convergence is achieved. After training, a test set is used to test the model, resulting in a well-trained MobileNet-SSD model. The trained MobileNet-SSD model is then saved. 4) Compression model based on joint pruning algorithm: The joint pruning algorithm uses L2 norm pruning, batch normalized layer parameter pruning, and similarity pruning to evaluate the importance of each convolutional kernel in each convolutional layer of the convolutional neural network, and obtains the convolutional kernel loss-pruning percentage curve for each convolutional layer under each pruning algorithm. The convolutional kernel loss-pruning percentage curves obtained by the L2 norm pruning algorithm, the batch normalized layer parameter pruning algorithm, and the similarity pruning algorithm are compared. The pruning algorithm corresponding to the lowest curve, that is, the pruning algorithm corresponding to the curve with the lowest average value of the ordinate, is selected as the optimal pruning strategy for this layer. The optimal pruning strategy is used to prune the convolutional layer of this layer. A joint pruning algorithm is used to prune each convolutional layer of the convolutional neural network, thereby compressing the model. The compressed quantization model is obtained; 5) Online quantization training model: In the compressed quantization model, the output value and weights of each convolutional layer are quantized, and the output value of each convolutional layer is dequantized. The compressed quantization model is further trained using the dataset from step 2) to restore the original accuracy. Furthermore, while ensuring the model accuracy, the weights of the compressed quantization model are converted from floating-point numbers to integers to further reduce the hardware overhead of the model and improve the inference speed, thus obtaining a well-trained compressed quantization model. 6) Detect frequency-hopping targets and extract parameters: The trained compression quantization model is used to infer the extracted sample signals and detect the time-frequency patterns of frequency hopping signals and various types of interference signals contained in each time-frequency plot. Extract the midpoint coordinates from the time-frequency pattern of the frequency-hopping signal; After obtaining the midpoint of each frequency hopping signal, the abscissa of the midpoint of the frequency hopping signal is differentially divided to obtain a differential sequence. The average value of the differential sequence is used to obtain the frequency hopping period. The center frequency of the frequency hopping signal is obtained according to the ordinate of the midpoint of each frequency hopping signal. The center time point in the original time-frequency diagram is obtained according to the abscissa of the midpoint of each frequency hopping signal. The 3dB spectral width of the center time point is obtained on the spectral distribution curve corresponding to the center time point, which is the frequency hopping bandwidth.
2. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 1), the time-frequency analysis employs one of the following: short-time Fourier transform, wavelet transform, and short-time fractional Fourier transform.
3. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 1), the signal category labels are linear sweep frequency signal, nonlinear sweep frequency signal, fixed frequency interference signal, noise signal of each category, and frequency hopping signal.
4. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 1), the number of time-frequency images in the dataset is 6,000 to 8,000.
5. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 2), the ratio of the dataset to the training set, validation set, and test set is 6-8:1-3:
1.
6. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 2), translation: random translation on the time and frequency axes of the time-frequency plot to simulate the time and frequency shifts of various signals; noise injection: Towards Random noise is added to the time-frequency plot to simulate interference in a real communication environment: brightness and contrast adjustments are made. The brightness and contrast of the time-frequency plot are used to simulate different signal reception conditions.
7. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 3), the objective function is obtained by weighting the class loss and the localization loss, and an adaptive inertia estimation optimizer is used, with a learning rate of 0.01 to 0.001 and a batch size of 16 to 64.
8. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 4), the L2 norm pruning algorithm evaluates convolutional kernels based on the L2 norm of their weights; kernels with larger L2 norms are more important. The batch normalized layer parameter pruning algorithm evaluates convolutional kernels based on the γ parameter of the corresponding batch normalized layer; kernels with larger γ parameters are more important. The similarity pruning algorithm assesses the similarity between convolutional kernels based on their distance from the central kernel; kernels with lower similarity are more important. In each pruning algorithm, the importance of the convolutional kernels is ranked, and the lowest-ranked kernels are removed in order of ranking. The convolutional kernel loss-pruning percentage curve is obtained according to the removal order.
9. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 5), the weights of the compressed quantization model are converted from 16, 32, or 64-bit floating-point numbers to 8-bit integers.
10. The frequency hopping signal parameter extraction method as described in claim 1, characterized in that, In step 6), the sampling sequence index of the center frequency point is obtained based on the ordinate of the midpoint of each frequency hopping signal, and the sampling sequence index is multiplied by the sampling frequency to obtain the center frequency point of the frequency hopping signal.
Citation Information
Patent Citations
Frequency hopping signal detection and parameter estimation method based on deep neural network
CN111832462A
Remote sensing SAR target detection method based on combination of network pruning and parameter quantification
CN114170512A
