A low-power instantaneous frequency measurement system

Abstract

The present invention relates to the field of low-power electronic technology, and discloses a low-power instantaneous frequency measurement system, which includes a radio frequency signal acquisition module, a signal time-frequency conversion module, a noise base coding module, etc. The system generates a feature tensor by performing time-frequency analysis on the radio frequency signal, realizes dynamic threshold fusion in combination with noise base sparse coding, extracts multi-scale transient frequency features using a pyramid convolutional neural network, dynamically suppresses noise based on the signal-to-noise ratio weight, and reconstructs a real-time spectrum distribution image. The power consumption control module can adjust the ADC sampling rate and the processing unit clock frequency according to the signal strength, and the power management unit adopts multiple voltage domains for independent power supply. The system solves the problems of frequency measurement accuracy and power consumption under complex noise, is applicable to fields such as wireless communications, and has the characteristics of high precision, low power consumption, high spectrum resolution, and intuitive visualization.

Description

Technical Field

[0001] The present invention relates to the field of low-power electronic technology, and in particular to a low-power instantaneous frequency measurement system. Background Art

[0002] In fields such as wireless communications, radar detection, and spectrum monitoring, instantaneous frequency measurement is one of the core technologies for signal analysis, interference identification, and communication protocol parsing. Traditional instantaneous frequency measurement systems generally face two major technical bottlenecks: First, in complex electromagnetic environments, noise and interference signals can easily lead to a decrease in frequency measurement accuracy. This is especially true in low signal-to-noise ratio scenarios, where traditional algorithms struggle to effectively distinguish signal and noise characteristics. Second, in highly integrated hardware architectures, fixed configurations such as the ADC sampling rate and the clock frequency of the digital signal processing unit (DSP) result in high system power consumption, making it impossible to meet the stringent low-power requirements of portable devices and wireless sensor networks.

[0003] In the existing technology, the time-frequency analysis method based on short-time Fourier transform can achieve the preliminary extraction of signal time-frequency characteristics, but lacks the ability to dynamically model the noise floor. When the environmental noise or equipment noise floor changes, the fixed threshold signal detection method is prone to false alarms or missed detections. At the same time, traditional noise suppression algorithms mostly use fixed filtering parameters and cannot dynamically adjust the suppression strength according to the real-time signal-to-noise ratio, resulting in the possible excessive attenuation of useful signal characteristics or incomplete noise suppression. In addition, at the hardware implementation level, the single voltage domain power supply mode and fixed sampling rate design make it impossible for the system to achieve adaptive power consumption adjustment when processing signals of different strengths. Especially in weak signal scenarios, excessively high sampling rates and clock frequencies will cause unnecessary energy loss.

[0004] The rapid development of 5G communications and smart wireless devices in the Internet of Things (IoT) is placing higher demands on instantaneous frequency measurement systems: not only must they achieve high-precision frequency measurement in complex noisy environments, but they must also maintain low power consumption to extend device battery life. Existing technologies lack the ability to simultaneously optimize both accuracy and power consumption due to shortcomings in noise modeling, dynamic threshold setting, multi-scale feature extraction, and power management. Therefore, there is an urgent need for a new instantaneous frequency measurement system that can integrate signal time-frequency characteristics with noise floor characteristics, implement dynamic threshold detection and adaptive noise suppression, and dynamically adjust power consumption. Summary of the Invention

[0005] The object of the present invention is to provide a low-power instantaneous frequency measurement system to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a low-power instantaneous frequency measurement system, the system comprising:

[0007] The RF signal acquisition module is used to obtain the RF signal and noise floor data of the target frequency band, wherein the RF signal includes the time domain waveform, signal strength, and modulation type, and the noise floor data includes the ambient noise spectrum, device noise floor, and interference signal characteristics;

[0008] A signal time-frequency conversion module, configured to perform a time-frequency joint analysis on the radio frequency signal to generate a signal time-frequency distribution feature tensor;

[0009] a noise floor coding module, configured to perform sparse coding on the noise floor data to generate a noise floor sparse coding feature matrix;

[0010] A dynamic threshold fusion module, configured to fuse the signal time-frequency distribution feature tensor and the noise floor sparse coding feature matrix to generate a frequency domain feature fusion tensor;

[0011] A transient feature extraction module, configured to perform multi-scale feature aggregation on the frequency domain feature fusion tensor to generate a transient frequency feature vector;

[0012] an adaptive noise suppression module, configured to perform dynamic noise suppression on the transient frequency feature vector based on a signal-to-noise ratio weight to generate a corrected transient frequency feature vector;

[0013] The spectrum image reconstruction module is used to reconstruct the real-time spectrum distribution image of the target signal according to the corrected transient frequency eigenvector.

[0014] Preferably, the signal time-frequency conversion module includes:

[0015] a signal segmentation and windowing unit, configured to perform overlapping segmentation on the radio frequency signal and apply Hamming window processing to generate a set of windowed signal segments;

[0016] The time-frequency analysis unit is used to input the windowed signal segment set into a short-time Fourier transform network to generate the signal time-frequency distribution feature tensor.

[0017] Preferably, the time-frequency analysis unit includes:

[0018] a complex spectrum generating subunit, configured to perform a fast Fourier transform on each windowed signal segment to generate a complex spectrum set;

[0019] The energy spectrum aggregation subunit is used to perform logarithmic energy conversion and time dimension stacking on the complex spectrum set to form the signal time-frequency distribution feature tensor.

[0020] Preferably, the noise floor encoding module includes:

[0021] A noise dictionary training unit is used to train historical noise data using the K-SVD algorithm to generate an overcomplete noise dictionary;

[0022] The sparse representation unit is used to perform sparse decomposition on the current noise floor data on the overcomplete noise dictionary using an orthogonal matching pursuit algorithm to generate the noise floor sparse coding feature matrix.

[0023] Preferably, the transient feature extraction module is used to: input the frequency domain feature fusion tensor into the pyramid convolutional neural network for spatial dimension compression and channel dimension expansion, and then generate the transient frequency feature vector through global average pooling.

[0024] Preferably, the adaptive noise suppression module includes:

[0025] a subband signal-to-noise ratio estimation unit, configured to perform frequency-level signal-to-noise ratio calculation on the transient frequency feature vector to generate a signal-to-noise ratio weight vector;

[0026] A dynamic filtering unit, configured to construct an adjustable exponential attenuation function according to the signal-to-noise ratio weight vector to perform amplitude correction on the characteristic component;

[0027] The phase holding unit is used to retain the original phase information during the frequency domain feature correction process to generate the corrected transient frequency feature vector.

[0028] Preferably, the dynamic filtering unit is used to construct an adaptive filtering response curve with frequency selection characteristics through a nonlinear combination of a signal-to-noise ratio weight vector and a preset attenuation coefficient.

[0029] Preferably, the spectrum image reconstruction module includes:

[0030] a frequency point interpolation unit, configured to perform cubic spline interpolation on the corrected transient frequency eigenvector to expand spectrum resolution;

[0031] A color mapping unit is used to convert the interpolated spectrum data into an HSV color space mapping matrix;

[0032] An image synthesis unit is used to convert the HSV color space mapping matrix into a spectrum heat map in RGB format to generate the real-time spectrum distribution image.

[0033] Preferably, the system further comprises:

[0034] Power consumption control module, used to dynamically adjust the ADC sampling rate and the working clock frequency of the digital signal processing unit according to the signal strength;

[0035] The power management unit adopts a multi-voltage domain design to independently control the power supply of the RF front end, digital processor, and display module.

[0036] Preferably, the power consumption control module includes an environment sensing unit, which automatically switches to a low power consumption operation mode by monitoring the ambient temperature and the working status of the device in real time.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] At the signal processing level, the signal time-frequency transformation module performs overlapping segmentation and Hamming window processing on the RF signal. Combined with the short-time Fourier transform network, the time-frequency distribution feature tensor is generated, enabling joint analysis of the signal's time and frequency domain characteristics. Compared to traditional single-domain analysis methods, this method can more accurately capture the time-frequency variation characteristics of transient signals. The noise floor encoding module uses the K-SVD algorithm to train a complete noise dictionary and performs sparse decomposition of the current noise floor data using the orthogonal matching pursuit algorithm. This constructs a dynamic representation model of the noise characteristics, enabling the system to adapt in real time to changes in environmental noise, device noise floor, and interference signals, providing an accurate noise reference floor for subsequent dynamic threshold fusion.

[0039] The dynamic threshold fusion module fuses the signal's time-frequency distribution features with the sparsely coded features of the noise floor to generate a frequency-domain feature fusion tensor. This effectively addresses the lack of adaptability of traditional fixed-threshold detection to complex noise, improving signal detection reliability in low signal-to-noise ratio scenarios. The transient feature extraction module utilizes a pyramid-like convolutional neural network for multi-scale feature aggregation. Through spatial dimension compression and channel dimension expansion, it enhances the ability to express features across different frequency components and time scales. Combined with the transient frequency feature vector generated by global average pooling, it achieves efficient compression of signal features while retaining key information.

[0040] The adaptive noise suppression module generates a weight vector by estimating the signal-to-noise ratio at the frequency level. This then drives an adjustable exponential attenuation function to modify the amplitude of characteristic components while preserving the original phase information. This allows for dynamic adjustment of noise suppression strength, avoiding excessive damage to signal characteristics caused by fixed filtering parameters. This effectively suppresses noise while preserving signal integrity to the greatest extent possible. The spectrum image reconstruction module enhances spectral resolution through cubic spline interpolation and generates real-time spectrum heat maps using HSV to RGB color space mapping. This provides users with an intuitive visualization interface for spectral distribution, facilitating the rapid location of signal characteristics and interference sources.

[0041] In terms of power consumption management, the power consumption control module dynamically adjusts the ADC sampling rate and the operating clock frequency of the digital signal processing unit according to the signal strength, avoiding oversampling and waste of computing resources in strong signal scenarios; the multi-voltage domain power management unit independently controls the power supply of the RF front end, digital processor and display module, and combines the environmental perception unit's real-time monitoring of temperature and device status to achieve automatic switching of low-power operation modes, significantly reducing the overall power consumption of the system and extending the battery life of the device in portable scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a working principle diagram of the low-power instantaneous frequency measurement system of the present invention;

[0043] Figure 2 This is the working principle diagram of the signal time-frequency conversion module;

[0044] Figure 3 This is the working principle diagram of the time-frequency analysis unit;

[0045] Figure 4 This is the working principle diagram of the transient feature extraction module. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] See also Figures 1-4 The present invention relates to a low-power instantaneous frequency measurement system, which includes: a radio frequency signal acquisition module, a signal time-frequency conversion module, a noise floor encoding module, a dynamic threshold fusion module, a transient feature extraction module, an adaptive noise suppression module, and a spectrum image reconstruction module. The specific implementation steps are as follows:

[0048] First, the RF signal acquisition module acquires the RF signal and noise floor data of the target frequency band. The RF signal includes the time-domain waveform, signal strength, and modulation type, while the noise floor data includes the ambient noise spectrum, device noise floor, and interference signal characteristics. Next, the signal time-frequency conversion module performs a joint time-frequency analysis on the RF signal to generate a signal time-frequency distribution feature tensor. The noise floor encoding module then performs sparse coding on the noise floor data to generate a noise floor sparse coding feature matrix. The dynamic threshold fusion module then fuses the signal time-frequency distribution feature tensor with the noise floor sparse coding feature matrix to generate a frequency domain feature fusion tensor. The transient feature extraction module then performs multi-scale feature aggregation on the frequency domain feature fusion tensor to generate a transient frequency feature vector. Subsequently, the adaptive noise suppression module performs dynamic noise suppression on the transient frequency feature vector based on signal-to-noise ratio weighting to generate a corrected transient frequency feature vector. Finally, the spectrum image reconstruction module reconstructs a real-time spectrum distribution image of the target signal based on the corrected transient frequency feature vector.

[0049] The present invention will be further described below in conjunction with Examples 1 to 5:

[0050] Example 1: The signal time-frequency conversion module of the system includes a signal segmentation and windowing unit and a time-frequency analysis unit. The signal segmentation and windowing unit divides the radio frequency signal into overlapping segments, using a preset overlap ratio. The overlapping length of two adjacent segments is 1 / 4 of the length of a single segment to ensure the continuity of signal processing. After the segmentation is completed, Hamming window processing is applied to each signal segment. The expression of the Hamming window is a specific cosine function combination. This window function processing can reduce the spectrum leakage problem caused by signal truncation and generate a set of windowed signal segments. The time-frequency analysis unit inputs the set of windowed signal segments into a short-time Fourier transform network. The network generates a set of complex spectra by performing a fast Fourier transform on each windowed signal segment. The complex spectrum set contains the complex amplitude information of each signal segment at different frequency points, reflecting the frequency components and phase relationships of the signal. Subsequently, the energy spectrum aggregation subunit performs logarithmic energy conversion on the complex spectrum set, converting the complex amplitudes into logarithmic energy values to expand the dynamic range and facilitate subsequent processing. At the same time, the converted energy values are stacked in the time dimension to form a signal time-frequency distribution feature tensor. The dimensions of this tensor include time, frequency, and energy, which can intuitively reflect the energy distribution characteristics of the signal in the time-frequency domain.

[0051] When performing signal segmentation, the signal segmentation and windowing unit first preprocesses the input RF signal, including removing the DC component and smoothing to improve signal quality. The preprocessed signal is then fed into the segmentation module and segmented according to a preset segment length and overlap ratio. The selection of segment length requires comprehensive consideration of the signal characteristics and processing requirements. Segments that are too short will result in reduced frequency resolution, while segments that are too long will affect time resolution. In this embodiment, the segment length is adaptively adjusted based on the signal bandwidth and sampling rate to ensure that the transient characteristics of the signal can be captured.

[0052] Overlapping segments is designed to avoid signal distortion at segment boundaries. By overlapping adjacent segments, signal continuity can be better restored during subsequent processing. In this embodiment, the choice of an overlap ratio of 1 / 4 is a balancing act that ensures signal continuity without significantly increasing the processing burden. An overlap ratio that is too small may prevent effective compensation for signal distortion at segment boundaries, while an overlap ratio that is too large may increase computational complexity and storage requirements.

[0053] Applying Hamming window processing is one of the key steps in the signal segmentation and windowing unit. The Hamming window is a commonly used window function, and its expression is:

[0054] .

[0055] in, is the sampling point index, is the length of the window function. The Hamming window is characterized by a moderate mainlobe width and rapid sidelobe decay, achieving a good balance between suppressing spectral leakage and maintaining frequency resolution. By applying a Hamming window to each signal segment, spectral leakage caused by signal truncation can be reduced, resulting in more accurate spectrum analysis.

[0056] When applying a Hamming window, it's important to ensure the length of the window function matches the length of the signal segment to ensure it completely covers the segment. Furthermore, to avoid edge effects, a smooth transition is typically performed at both ends of the window function. In this embodiment, a symmetrical expansion method is employed, symmetrically expanding both ends of the window function before multiplying it with the signal segment to mitigate the effects of edge effects.

[0057] After receiving the windowed signal segments, the time-frequency analysis unit inputs them into a short-time Fourier transform (SFT) network. SFT is a commonly used time-frequency analysis method that performs a windowing process on a signal and then performs a Fourier transform to obtain the signal's distribution over time and frequency. In this embodiment, the SFT network utilizes a fast Fourier transform (FFT) algorithm to improve computational efficiency.

[0058] The Fast Fourier Transform (FFT) algorithm is an efficient method for calculating discrete Fourier transforms, reducing computational complexity from O(N^2) to O(NlogN), where N is the length of the signal. In this embodiment, to further improve computational efficiency, a radix-2 FFT algorithm is employed. This algorithm requires that the signal length be a power of 2. If the length of a signal segment is not a power of 2, zero padding is performed to ensure that the length meets the algorithm's requirements.

[0059] By performing a fast Fourier transform on each windowed signal segment, a complex spectrum set is generated. Each element in the complex spectrum set contains the complex amplitude information of the corresponding signal segment at a specific frequency point. The complex amplitude consists of a real part and an imaginary part, representing the amplitude and phase of the signal at that frequency point, respectively. The complex spectrum set can comprehensively reflect the frequency components and phase relationships of the signal, providing rich information for subsequent signal analysis and processing.

[0060] The energy spectrum aggregation subunit performs logarithmic energy conversion and time dimension stacking on the complex spectrum set. Logarithmic energy conversion is the process of converting complex amplitudes into logarithmic energy values. Its calculation formula is:

[0061] .

[0062] in, is the logarithmic energy value, The logarithmic energy conversion can expand the dynamic range of the signal, so that both weak and strong signals can be better represented, and it also facilitates subsequent processing and analysis.

[0063] The converted energy values are stacked in the time dimension to form a signal time-frequency distribution feature tensor. The signal time-frequency distribution feature tensor is a three-dimensional tensor whose three dimensions represent time, frequency, and energy, respectively. In this embodiment, the time dimension corresponds to the segmentation order of the signal, the frequency dimension corresponds to the frequency point of the fast Fourier transform, and the energy dimension corresponds to the logarithmic energy value. In this way, the signal time-frequency distribution feature tensor can intuitively reflect the energy distribution characteristics of the signal in the time-frequency domain, providing strong support for subsequent signal processing and analysis.

[0064] After forming the signal time-frequency distribution feature tensor, it needs to be normalized to eliminate energy differences between different signal segments. Normalization can improve the stability and accuracy of subsequent processing. In this embodiment, a global normalization method is adopted, that is, the entire signal time-frequency distribution feature tensor is normalized so that its energy value is within an appropriate range.

[0065] The design and implementation of the signal time-frequency conversion module fully considers the signal's time-frequency characteristics and processing requirements. Through the collaborative work of the signal segmentation and windowing unit and the time-frequency analysis unit, it can accurately extract the signal's time-frequency distribution characteristics, providing a solid foundation for subsequent signal processing and analysis. In practical applications, this module can adjust and optimize parameters based on specific signal characteristics and processing requirements to achieve better processing results.

[0066] Example 2: The noise base coding module includes a noise dictionary training unit and a sparse representation unit. The noise dictionary training unit uses the K-SVD algorithm to train historical noise data. The historical noise data includes environmental noise spectra, equipment background noise, and interference signal characteristics under different environments. The K-SVD algorithm learns an over-complete noise dictionary that can sparsely represent noise from historical noise data by iteratively updating dictionary atoms and sparse representation coefficients. The number of atoms in the over-complete noise dictionary is greater than the signal feature dimension, and it can more flexibly represent the complex structure of noise. The sparse representation unit uses the orthogonal matching pursuit algorithm to sparsely decompose the current noise base data on the over-complete noise dictionary. The orthogonal matching pursuit algorithm constructs sparse representation coefficients by gradually selecting the dictionary atoms that are most relevant to the current noise data, and finally generates a noise base sparse coding feature matrix. The matrix represents the current noise base data in a sparse form, highlights the main components of the noise, and facilitates subsequent fusion processing with signal features.

[0067] The noise dictionary training unit first collects a large amount of historical noise data when training the noise dictionary. This historical noise data comes from a variety of environments and scenarios, including indoor and outdoor environments, and different frequency bands, to ensure that it covers all possible noise types and characteristics. The collected historical noise data requires preprocessing, including outlier removal and normalization, to improve data quality and consistency.

[0068] The preprocessed historical noise data is fed into the K-SVD algorithm for training. The K-SVD algorithm is an iterative algorithm that learns the optimal dictionary by alternating between updating dictionary atoms and sparse representation coefficients. In each iteration, the algorithm first fixes the dictionary atoms and solves for the sparse representation coefficients; then, it fixes the sparse representation coefficients and updates the dictionary atoms. This process is repeated until convergence conditions are met.

[0069] During the dictionary update phase, the K-SVD algorithm adopts a greedy strategy, updating only one dictionary atom at a time while keeping all other atoms unchanged. Specifically, the algorithm selects a dictionary atom to be updated and then calculates its contribution to all training samples based on the current sparse representation coefficients. The value of this dictionary atom is then updated by minimizing the reconstruction error. This process is repeated until all dictionary atoms have been updated.

[0070] The design of an overcomplete noise dictionary is key to the noise dictionary training unit. An overcomplete noise dictionary has more atoms than signal feature dimensions, enabling the dictionary to more flexibly represent the complex structure of noise. In this embodiment, the number of atoms in the overcomplete noise dictionary is adaptively adjusted based on the characteristics of historical noise data and processing requirements. Excessive atoms increase computational complexity and storage requirements, while too few atoms may fail to fully represent noise characteristics.

[0071] To improve the quality and generalization of the overcomplete noise dictionary, a regularization term is introduced during training. Regularization constrains the properties of dictionary atoms and prevents overfitting. In this implementation, an L1 regularization term is used, which ensures that the sparse representation coefficients are sparse, meaning that most coefficients are zero, with only a few non-zero coefficients. This sparsity allows for a concise representation of the noise floor data, highlighting the primary components of the noise.

[0072] After receiving the overcomplete noise dictionary and the current noise floor data, the sparse representation unit uses the orthogonal matching pursuit algorithm for sparse decomposition. The orthogonal matching pursuit algorithm is a greedy algorithm that constructs sparse representation coefficients by gradually selecting the dictionary atoms that are most relevant to the current noise data. Specifically, the algorithm first initializes the residual to the current noise data. Then, at each step, it selects the dictionary atom most relevant to the residual and adds it to the sparse representation. Next, the residual is updated to be orthogonal to the space spanned by the selected dictionary atoms. This process is repeated until a stopping condition is met.

[0073] When selecting dictionary atoms, the orthogonal matching pursuit algorithm calculates the inner product between the current residual and each dictionary atom and selects the atom with the largest absolute value of the inner product. This selection strategy ensures that the atom selected each time is the one most relevant to the current residual, allowing it to quickly approximate a sparse representation of noisy data.

[0074] When updating the residual, the orthogonal matching pursuit algorithm uses the Gram-Schmidt orthogonalization method to orthogonalize the selected dictionary atoms and then project the residual into the orthogonal complement space. This orthogonalization process ensures that the atoms selected at each time are linearly independent, thereby improving the accuracy and stability of the sparse representation.

[0075] The stopping condition of the orthogonal matching pursuit algorithm can be set according to specific needs. In this embodiment, the stopping condition is to reach a preset number of iterations or the residual energy is lower than a certain threshold. The preset number of iterations determines the sparsity of the sparse representation, that is, the number of non-zero coefficients. The more iterations, the lower the sparsity, the more accurate the sparse representation, but the higher the computational complexity. The residual energy threshold controls the size of the reconstruction error. The smaller the threshold, the smaller the reconstruction error, but more iterations may be required.

[0076] Using the orthogonal matching pursuit algorithm, the current noise floor data is sparsely decomposed on an overcomplete noise dictionary to generate sparse representation coefficients. These sparse representation coefficients form the noise floor sparse coding feature matrix. This matrix represents the current noise floor data in a sparse form, retaining only coefficients related to the primary noise components while ignoring secondary components and noise. This sparse representation not only reduces data storage space but also highlights the primary noise features, facilitating subsequent fusion with signal features.

[0077] After generating the noise floor sparse coding feature matrix, post-processing is required, including normalization and thresholding. Normalization eliminates energy differences between different noise data, improving the stability and accuracy of subsequent processing. Thresholding further removes minor noise components, retaining only the noise features that are important for signal processing.

[0078] The design and implementation of the noise basis encoding module fully consider the characteristics and processing requirements of noise. Through the collaborative work of the noise dictionary training unit and the sparse representation unit, it can accurately extract noise features and represent them in a sparse form. In practical applications, this module can adjust and optimize parameters based on the specific noise environment and processing requirements to achieve better processing results. For example, in the case of complex and changing noise environments, the amount of historical noise data collected can be increased to improve the quality of the overcomplete noise dictionary. In the case of limited computing resources, the number of dictionary atoms and the number of iterations of the orthogonal matching pursuit algorithm can be appropriately reduced to reduce computational complexity.

[0079] Example 3: The specific implementation method of the transient feature extraction module is: input the frequency domain feature fusion tensor into the pyramid convolutional neural network. The pyramid convolutional neural network contains multiple convolution layers of different scales, and the convolution kernel size of each convolution layer is different, which can extract multi-scale spatial features from the frequency domain feature fusion tensor. Through convolution operations of different scales, the spatial dimension of the feature tensor is compressed, the redundant information of the spatial dimension is reduced, and the channel dimension is expanded at the same time to increase the expressive ability of the feature. After multi-layer convolution processing, the features of the spatial dimension are averaged through the global average pooling operation to generate a transient frequency feature vector. Global average pooling can retain the feature information of the channel dimension while reducing the dimension of the feature vector, which is convenient for subsequent noise suppression processing.

[0080] As the input to the pyramidal convolutional neural network, the frequency-domain feature fusion tensor first requires preprocessing to meet the network's input requirements. This preprocessing includes normalization, which adjusts the tensor's numerical range to fit within the effective training range, improving the network's convergence speed and stability. Global normalization can be used to standardize the mean and variance of each channel in the tensor, ensuring comparability of features across channels.

[0081] The core of the pyramid convolutional neural network lies in the design of multi-scale convolutional layers. The network comprises multiple convolutional layers, each with a different kernel size, such as 3×3, 5×5, and 7×7. Kernels of varying sizes capture spatial features at different scales: small kernels (such as 3×3) excel at extracting local details and are more sensitive to high-frequency information; large kernels (such as 7×7) cover a wider spatial range and extract global or low-frequency features. This multi-scale convolution operation allows the network to capture feature information at different levels from the frequency-domain feature fusion tensor, enhancing feature diversity and richness.

[0082] During convolution, each convolutional layer is typically followed by an activation layer and a batch normalization layer. The activation layer (such as the ReLU function) introduces nonlinear transformations, enabling the network to learn more complex feature relationships. The batch normalization layer normalizes the features of the convolution output, reducing internal covariate shift, accelerating network training, and improving generalization. Layers are stacked to form a hierarchical feature extraction structure, with the output of the previous layer serving as the input to the next layer, gradually abstracting and refining features.

[0083] Spatial dimension compression is achieved through the stride setting and padding strategy in convolution operations. For example, using a stride of 2 in some convolutional layers halves the spatial dimensions of the output feature map (such as the time and frequency dimensions), thereby reducing redundant information in the spatial dimension. Simultaneously, by controlling the number of convolution kernels and channels, the channel dimension is expanded while compressing the spatial dimension. For example, if the input feature tensor has C channels, this number increases to 2C after passing through the convolutional layer. This allows the network to learn feature representations in more dimensions, improving feature expressiveness.

[0084] After multiple layers of convolution, the spatial dimensions of the feature tensor (such as time and frequency) are compressed to a lower size, while the channel dimension increases significantly, forming a high-dimensional feature map. At this point, the spatial features are aggregated through global average pooling. Global average pooling averages all elements in the spatial dimensions (i.e., time and frequency dimensions) of each channel, compressing the features of each channel into a scalar value. After global average pooling, the feature tensor is converted from a three-dimensional structure (time × frequency × channel) to a one-dimensional feature vector with a length equal to the number of channels. This vector is the transient frequency feature vector.

[0085] The advantage of global average pooling is that it preserves channel-dimensional feature information, avoiding the overfitting problem that can occur with traditional fully connected layers while significantly reducing computational complexity. Through the averaging operation, local features in the spatial dimension are integrated into global features, allowing the resulting transient frequency feature vector to comprehensively reflect the multi-scale information in the frequency-domain feature fusion tensor.

[0086] During network training, a backpropagation algorithm is used to optimize network parameters. By defining a loss function (such as mean squared error or cross entropy loss, depending on subsequent task requirements), the difference between the predicted and true values is calculated. This error is then backpropagated through each network layer to update the convolution kernel weights and bias parameters. During training, stochastic gradient descent (SGD) and its variants (such as the Adam optimizer) can be used to adjust the learning rate to improve training efficiency and convergence accuracy.

[0087] The structural design of pyramidal convolutional neural networks requires a balance between computational complexity and feature extraction capabilities. For example, increasing the number of convolutional layers and the size of the convolution kernels improves the richness of feature extraction, but also leads to a surge in computational complexity. Conversely, reducing the number of layers or using smaller convolution kernels may not capture sufficient multi-scale features. Therefore, in practical applications, the number of layers, kernel size, number of channels, and other parameters of the network must be appropriately designed based on the specific dimensions of the frequency domain feature fusion tensor and computational resource constraints.

[0088] Furthermore, to prevent network overfitting, regularization methods, such as dropout layers, can be introduced between convolutional layers. During training, the dropout layer randomly discards the outputs of some neurons with a certain probability, forcing the network to learn more robust feature representations and improving the model's generalization capabilities.

[0089] The transient feature extraction module utilizes a multi-scale feature aggregation mechanism within a pyramid-style convolutional neural network to effectively extract spatiotemporal features from the frequency-domain feature fusion tensor. It then generates a compact transient frequency feature vector through global average pooling. This module's design leverages the strengths of convolutional neural networks in image processing and feature extraction, incorporating a multi-scale analysis strategy to provide high-quality feature representation for subsequent adaptive noise suppression and spectrum image reconstruction. In practical applications, the network structure and parameters can be flexibly adjusted to optimize feature extraction based on varying signal types and frequency measurement requirements.

[0090] Example 4: The adaptive noise suppression module includes a sub-band signal-to-noise ratio estimation unit, a dynamic filtering unit and a phase holding unit. The sub-band signal-to-noise ratio estimation unit performs frequency-level signal-to-noise ratio calculation on the transient frequency feature vector, calculates the ratio of its signal energy to noise energy for the feature component of each frequency point, and generates a signal-to-noise ratio weight vector. Each element in the vector corresponds to the signal-to-noise ratio weight of a frequency point, reflecting the quality of the signal at that frequency point. The dynamic filtering unit constructs an adjustable exponential attenuation function based on the signal-to-noise ratio weight vector, and determines the parameters of the attenuation function through a nonlinear combination of the signal-to-noise ratio weight vector and a preset attenuation coefficient, so that the attenuation function has a frequency selection characteristic. For frequency points with a higher signal-to-noise ratio, the attenuation amplitude of the attenuation function is smaller, retaining more signal features; for frequency points with a lower signal-to-noise ratio, the attenuation amplitude is larger, suppressing the noise component. In the process of amplitude correction of the feature component, the phase holding unit retains the original phase information to ensure that the phase characteristics of the signal are not affected, and finally generates a corrected transient frequency feature vector.

[0091] After receiving the transient frequency feature vector, the subband signal-to-noise ratio estimation unit first divides it into multiple subbands. Each subband contains several adjacent frequency points, and the division method can be determined according to the specific application requirements, such as equal bandwidth division or adaptive division based on signal characteristics. For each frequency point in the subband, its signal energy and noise energy are estimated separately. The signal energy estimation can be based on the characteristic amplitude of the frequency point by performing statistical analysis of the amplitude over a period of time. The estimation of noise energy requires the use of the noise base sparse coding feature matrix generated by the noise base coding module to extract the noise feature information of the corresponding frequency point, and dynamically adjust it based on the characteristics of the current signal.

[0092] To estimate the signal-to-noise ratio (SNR) of each frequency, the subband SNR estimation unit uses a sliding window approach for time-domain analysis. By sliding the window over time, the ratio of signal energy to noise energy at each frequency within the window is calculated, yielding SNR estimates for that frequency at different times. To improve estimation accuracy, smoothing filtering techniques can be used to process the SNR estimates and reduce fluctuations.

[0093] Each element in the generated SNR weight vector corresponds to the SNR weight of a frequency, and the weight value reflects the signal quality at that frequency. Frequencies with higher SNRs have larger weights, indicating higher signal reliability. Frequencies with lower SNRs have smaller weights, indicating that the signal at that frequency may be severely contaminated by noise.

[0094] After receiving the signal-to-noise ratio weight vector, the dynamic filtering unit constructs an adjustable exponential attenuation function based on it. The parameters of the adjustable exponential attenuation function are determined by a nonlinear combination of the signal-to-noise ratio weight vector and preset attenuation coefficients. The preset attenuation coefficients are a set of pre-set parameters that control the basic shape and characteristics of the attenuation function. This nonlinear combination of the signal-to-noise ratio weight vector and the preset attenuation coefficients enables the attenuation function to have frequency-selective properties, effectively attenuating signals of different frequencies to varying degrees.

[0095] For frequencies with high signal-to-noise ratios, the adjustable exponential attenuation function has a smaller attenuation amplitude, effectively preserving the signal characteristics of these frequencies. For frequencies with low signal-to-noise ratios, the attenuation amplitude is larger, effectively suppressing the noise components at these frequencies. This frequency-selective characteristic enables the dynamic filtering unit to adaptively adjust the filtering strategy based on the actual signal conditions, improving noise suppression effectiveness.

[0096] When constructing the adjustable exponential attenuation function, the dynamic filtering unit also takes into account the time-varying nature of the signal. By dynamically adjusting the attenuation function in the time dimension, the filtering strategy can better adapt to signal changes. For example, for rapidly changing signals, the attenuation function's adjustment speed is accelerated to ensure timely tracking of signal changes; for slowly changing signals, the attenuation function's adjustment speed can be appropriately slowed to reduce unnecessary fluctuations.

[0097] The phase hold unit is responsible for preserving the original phase information while the dynamic filtering unit performs amplitude correction on the characteristic components. Phase information is crucial for correct signal reconstruction and understanding, especially when processing phase-sensitive signals. The phase hold unit tracks the phase changes of the original signal and reapplies the original phase information to the corrected signal after amplitude correction, ensuring that the signal's phase characteristics are not affected.

[0098] In practical implementation, the phase preservation unit can employ a variety of methods to preserve phase information. One common approach is to extract and store the phase information of the characteristic component before performing amplitude correction. After the amplitude correction is complete, the stored phase information is reapplied to the corrected characteristic component. Another approach involves designing a specialized filtering algorithm so that the filtering process itself preserves the signal's phase characteristics.

[0099] The three units of the adaptive noise suppression module work together to effectively suppress transient frequency eigenvectors. The subband signal-to-noise ratio estimation unit accurately estimates the signal-to-noise ratio at each frequency point, providing a basis for decision-making by the dynamic filtering unit. The dynamic filtering unit constructs an adjustable exponential attenuation function based on the signal-to-noise ratio weight vector to perform adaptive filtering on signals at different frequencies. The phase preservation unit ensures that the signal's phase information is preserved during the noise suppression process.

[0100] In practical applications, the performance of the adaptive noise suppression module is affected by a variety of factors. For example, the subband division method, the accuracy of the signal-to-noise ratio estimation, the design of the adjustable exponential decay function, and the choice of phase preservation method all have a significant impact on the noise suppression effect. Therefore, during implementation, these factors must be appropriately selected and optimized based on the actual application scenario and signal characteristics.

[0101] To improve the module's real-time performance and computational efficiency, the adaptive noise suppression module can be optimized using parallel computing techniques. For example, the parallel computing capabilities of multi-core processors or GPUs can be leveraged to simultaneously process signals from multiple subbands, accelerating computational speed. Furthermore, the algorithm can be simplified and approximated to reduce computational complexity while maintaining effective noise suppression.

[0102] The adaptive noise suppression module achieves adaptive noise suppression of transient frequency eigenvectors through three key steps: subband signal-to-noise ratio estimation, dynamic filtering, and phase preservation. This module automatically adjusts its filtering strategy based on the actual signal conditions, effectively suppressing noise while preserving key signal features, providing high-quality eigenvectors for subsequent spectrum image reconstruction. In practical applications, this module can be widely used in communications, radar, sonar, and other fields to improve the quality and reliability of signal processing.

[0103] Example 5: The spectrum image reconstruction module includes a frequency point interpolation unit, a color mapping unit, and an image synthesis unit. The frequency point interpolation unit performs cubic spline interpolation on the corrected transient frequency feature vector. Cubic spline interpolation is a piecewise interpolation method that expands the spectrum resolution by constructing a cubic polynomial between adjacent frequency points, making the reconstructed spectrum data denser. The color mapping unit converts the interpolated spectrum data into an HSV color space mapping matrix. The HSV color space contains three components: hue, saturation, and lightness, which can more intuitively represent the energy distribution of the spectrum data. In the specific conversion process, the lightness component is determined based on the energy value of the spectrum data, the hue component is determined based on the frequency range, and the saturation component can be set according to preset rules. The image synthesis unit converts the HSV color space mapping matrix into a spectrum heat map in RGB format. RGB format is a common image display format. The combination of three colors generates an intuitive spectrum heat map, thereby generating a real-time spectrum distribution image of the target signal, allowing users to intuitively observe the frequency distribution and energy changes of the signal.

[0104] When performing cubic spline interpolation, the frequency interpolation unit first needs to determine the interpolation nodes and interpolation interval. Each element in the corrected transient frequency eigenvector corresponds to the eigenvalue of a frequency point. These frequency points are usually evenly spaced on the frequency axis, but the spacing may be large, resulting in low spectral resolution. To improve spectral resolution, new frequency points are inserted between adjacent frequency points to make the sampling points on the frequency axis more dense.

[0105] The basic idea behind cubic spline interpolation is to divide the entire interpolation interval into several small segments, each of which is interpolated using a cubic polynomial. Within each segment, the cubic polynomial must satisfy the continuity conditions of the function value at the endpoints and the first- and second-order derivatives to ensure the smoothness of the interpolated curve. This segmented interpolation method can generate more frequency data without increasing the original data volume, thereby improving the spectral resolution.

[0106] When implementing cubic spline interpolation, the original frequency point data must first be preprocessed, including sorting and deduplication, to ensure that the frequency points are arranged in frequency order and there are no duplicate points. Next, the spacing between adjacent frequency points is calculated to determine the interpolation step size. The choice of interpolation step size needs to be adjusted according to actual needs. A smaller step size results in denser interpolated frequency points and higher spectral resolution, but this also increases the computational effort.

[0107] After receiving the interpolated spectrum data, the color mapping unit converts it into an HSV color space mapping matrix. The HSV color space is a color model for visual perception, where H represents hue, corresponding to the type of color, such as red, green, and blue; S represents saturation, corresponding to the vividness of the color; and V represents value, corresponding to the brightness of the color.

[0108] During the conversion process, the mapping relationship between the energy value of the spectral data and each HSV component must be determined. The lightness component V is typically directly related to the energy value of the spectral data. Higher energy values correspond to higher lightness and brighter colors; lower energy values correspond to lower lightness and darker colors. The hue component H can be mapped based on the frequency range, for example, mapping low frequencies to red, mid-range frequencies to green, and high frequencies to blue. This allows signals of different frequencies to be represented by different colors, making it easier for users to intuitively distinguish between high and low frequencies.

[0109] The saturation component, S, can be adjusted based on actual needs. Generally speaking, saturation can be set to a fixed value or dynamically adjusted based on the energy value. For example, areas with higher energy values can have higher saturation to make colors more vivid and highlight strong signals; areas with lower energy values can have lower saturation to make colors softer and avoid visual interference from weak signals.

[0110] The image synthesis unit converts the HSV color space mapping matrix into a spectrum heatmap in RGB format. The RGB color space is a color model based on the three primary colors. Various colors can be generated by mixing red (R), green (G), and blue (B) in varying proportions. The HSV to RGB conversion is a nonlinear process, requiring the calculation of the corresponding RGB values based on the HSV component values.

[0111] The conversion process first determines the primary hue interval to which the color belongs based on the hue H. Then, different formulas are used within each interval to calculate the values of each RGB component. For example, when the hue H is in the red to green interval, the red component gradually decreases from its maximum value, the green component gradually increases from its minimum value, and the blue component remains at its minimum value. When the hue H is in the green to blue interval, the green component gradually decreases from its maximum value, the blue component gradually increases from its minimum value, and the red component remains at its minimum value, and so on.

[0112] After completing the HSV to RGB conversion, the image synthesis unit combines the RGB values of all pixels into a two-dimensional matrix, known as a spectral heatmap. The abscissa of this heatmap represents frequency, and the ordinate can represent time or other relevant dimensions. The color of each pixel corresponds to the spectral energy value at that frequency and time. This allows users to intuitively observe the energy distribution of the target signal at different frequencies and times, quickly identifying the signal's frequency components, intensity variations, and potential interfering signals.

[0113] To enhance the display quality of the spectrum heat map, the image synthesis unit can also perform post-processing operations such as contrast adjustment, brightness adjustment, and edge enhancement. Contrast adjustment can enhance the color differences between different energy values, making the spectrum distribution clearer; brightness adjustment can adjust the overall brightness of the image to adapt to different display environments; and edge enhancement can highlight edge information in the spectrum, making it easier for users to identify signal boundaries and mutation points.

[0114] The design and implementation of the spectral image reconstruction module fully consider the characteristics of human visual perception. It improves spectral resolution through frequency interpolation, intuitively represents spectral energy distribution through color mapping, and generates easy-to-understand spectral heat maps through image synthesis. In practical applications, this module can adjust parameters such as interpolation step size, color mapping rules, and image resolution to achieve optimal display effects based on different display devices and user needs.

[0115] Furthermore, to achieve real-time spectrum image reconstruction, it's necessary to ensure that the computational efficiency of each unit meets real-time requirements. The frequency interpolation unit can utilize optimized interpolation algorithms and parallel computing techniques to accelerate interpolation. The color mapping unit and image synthesis unit can leverage hardware acceleration, such as the parallel computing capabilities of graphics processing units (GPUs), to increase image processing speed. Through these optimizations, the spectrum image reconstruction module can generate a spectral distribution image of the target signal in real time, providing users with timely signal analysis and monitoring capabilities.

[0116] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0117] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A low-power instantaneous frequency measurement system, characterized in that: include: The RF signal acquisition module is used to obtain the RF signal and noise floor data of the target frequency band, wherein the RF signal includes the time domain waveform, signal strength, and modulation type, and the noise floor data includes the ambient noise spectrum, device noise floor, and interference signal characteristics; A signal time-frequency conversion module, configured to perform a time-frequency joint analysis on the radio frequency signal to generate a signal time-frequency distribution feature tensor; a noise floor coding module, configured to perform sparse coding on the noise floor data to generate a noise floor sparse coding feature matrix; A dynamic threshold fusion module, configured to fuse the signal time-frequency distribution feature tensor and the noise floor sparse coding feature matrix to generate a frequency domain feature fusion tensor; A transient feature extraction module, configured to perform multi-scale feature aggregation on the frequency domain feature fusion tensor to generate a transient frequency feature vector; an adaptive noise suppression module, configured to perform dynamic noise suppression on the transient frequency feature vector based on a signal-to-noise ratio weight to generate a corrected transient frequency feature vector; The spectrum image reconstruction module is used to reconstruct the real-time spectrum distribution image of the target signal according to the corrected transient frequency eigenvector.

2. A low-power instantaneous frequency measurement system according to claim 1, characterized in that: The signal time-frequency conversion module includes: a signal segmentation and windowing unit, configured to perform overlapping segmentation on the radio frequency signal and apply Hamming window processing to generate a set of windowed signal segments; The time-frequency analysis unit is used to input the windowed signal segment set into a short-time Fourier transform network to generate the signal time-frequency distribution feature tensor.

3. A low-power instantaneous frequency measurement system according to claim 2, characterized in that: The time-frequency analysis unit includes: a complex spectrum generating subunit, configured to perform a fast Fourier transform on each windowed signal segment to generate a complex spectrum set; The energy spectrum aggregation subunit is used to perform logarithmic energy conversion and time dimension stacking on the complex spectrum set to form the signal time-frequency distribution feature tensor.

4. A low-power instantaneous frequency measurement system according to claim 3, characterized in that: The noise floor encoding module includes: A noise dictionary training unit is used to train historical noise data using the K-SVD algorithm to generate an overcomplete noise dictionary; The sparse representation unit is used to perform sparse decomposition on the current noise floor data on the overcomplete noise dictionary using an orthogonal matching pursuit algorithm to generate the noise floor sparse coding feature matrix.

5. A low-power instantaneous frequency measurement system according to claim 4, characterized in that: The transient feature extraction module is used to: input the frequency domain feature fusion tensor into the pyramid convolutional neural network for spatial dimension compression and channel dimension expansion, and then generate the transient frequency feature vector through global average pooling.

6. A low-power instantaneous frequency measurement system according to claim 5, characterized in that: The adaptive noise suppression module comprises: a subband signal-to-noise ratio estimation unit, configured to perform frequency-level signal-to-noise ratio calculation on the transient frequency feature vector to generate a signal-to-noise ratio weight vector; A dynamic filtering unit, configured to construct an adjustable exponential attenuation function according to the signal-to-noise ratio weight vector to perform amplitude correction on the characteristic component; The phase holding unit is used to retain the original phase information during the frequency domain feature correction process to generate the corrected transient frequency feature vector.

7. A low-power instantaneous frequency measurement system according to claim 6, characterized in that: The dynamic filtering unit is used to construct an adaptive filtering response curve with frequency selection characteristics through a nonlinear combination of a signal-to-noise ratio weight vector and a preset attenuation coefficient.

8. A low-power instantaneous frequency measurement system according to claim 7, characterized in that: The spectrum image reconstruction module includes: a frequency point interpolation unit, configured to perform cubic spline interpolation on the corrected transient frequency eigenvector to expand spectrum resolution; A color mapping unit is used to convert the interpolated spectrum data into an HSV color space mapping matrix; An image synthesis unit is used to convert the HSV color space mapping matrix into a spectrum heat map in RGB format to generate the real-time spectrum distribution image.

9. The low-power instantaneous frequency measurement system according to claim 1, characterized in that: Further including: Power consumption control module, used to dynamically adjust the ADC sampling rate and the working clock frequency of the digital signal processing unit according to the signal strength; The power management unit adopts a multi-voltage domain design to independently control the power supply of the RF front end, digital processor, and display module.

10. The low-power instantaneous frequency measurement system according to claim 9, characterized in that: The power consumption control module includes an environment sensing unit, which automatically switches to a low power consumption operation mode by monitoring the ambient temperature and the working status of the device in real time.

Images