Low-power-consumption instantaneous frequency measurement system

Abstract

The invention relates to the technical field of low-power-consumption electronics, and discloses a low-power-consumption instantaneous frequency measurement system which comprises a radio-frequency signal acquisition module, a signal time-frequency conversion module, a noise floor coding module and the like. The system generates a feature tensor through time-frequency analysis of a radio frequency signal, realizes dynamic threshold fusion in combination with noise floor sparse coding, extracts a multi-scale transient frequency feature by using a pyramid convolutional neural network, dynamically suppresses noise based on a 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 clock frequency of the processing unit according to the signal strength, and the power management unit adopts multi-voltage domain independent power supply. The system solves the problems of frequency measurement precision and power consumption under complex noise, is suitable for the fields of wireless communication and the like, and has the characteristics of high precision, low power consumption, high spectral resolution, visualization and intuition.

Description

Technical Field

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

[0002] In the fields of wireless communication, radar detection, spectrum monitoring, etc., instantaneous frequency measurement is one of the core technologies for signal analysis, interference identification, and communication protocol analysis. Traditional instantaneous frequency measurement systems generally face two major technical bottlenecks: one is that in a complex electromagnetic environment, noise and interference signals easily lead to a decline in frequency measurement accuracy. Especially in a low signal-to-noise ratio scenario, traditional algorithms are difficult to effectively distinguish the signal and noise characteristics; the other is that in a high-integration hardware architecture, fixed configurations such as the ADC sampling rate and the clock frequency of the digital signal processing unit result in high system power consumption, which cannot meet the stringent low-power requirements of portable devices, wireless sensor networks, etc.

[0003] In the prior art, although the time-frequency analysis method based on the short-time Fourier transform can initially extract the time-frequency characteristics of the signal, it lacks the ability to dynamically model the noise floor. When the environmental noise or the device background noise changes, the signal detection method with a fixed threshold 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 intensity according to the real-time signal-to-noise ratio, resulting in the characteristics of useful signals being possibly over-attenuated or the noise suppression being incomplete. In addition, at the hardware implementation level, the single voltage domain power supply mode and the fixed sampling rate design make the system unable to achieve adaptive power consumption adjustment when processing signals of different intensities. Especially in a weak signal scenario, too high a sampling rate and clock frequency will cause unnecessary energy consumption.

[0004] With the rapid development of Internet of Things 5G communication and intelligent wireless devices, higher requirements are put forward for the instantaneous frequency measurement system: not only high-precision frequency measurement needs to be achieved in a complex noise environment, but also low-power design needs to be taken into account to extend the device battery life. The deficiencies of the prior art in aspects such as noise modeling, dynamic threshold setting, multi-scale feature extraction, and power consumption management make it difficult to simultaneously meet the dual optimization goals of accuracy and power consumption. Therefore, there is an urgent need for a new type of instantaneous frequency measurement system that can fuse the time-frequency characteristics of the signal and the noise floor characteristics, achieve dynamic threshold detection and adaptive noise suppression, and have the ability of dynamic power consumption adjustment. Summary of the Invention

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

[0006] To achieve the above purpose, the present invention provides the following technical solution: A low-power instantaneous frequency measurement system, the system includes: The radio frequency signal acquisition module is used to acquire radio frequency signals and noise floor data in the target frequency band. Among them, the radio frequency signals include time-domain waveforms, signal intensities, and modulation types, and the noise floor data includes environmental noise spectra, device background noise, and interference signal characteristics; The signal time-frequency transformation module is used to perform joint time-frequency analysis on the radio frequency signals to generate a signal time-frequency distribution feature tensor; The noise floor coding module is used to perform sparse coding processing on the noise floor data to generate a noise floor sparse coding feature matrix; The dynamic threshold fusion module is used 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; The transient feature extraction module is used to perform multi-scale feature aggregation on the frequency-domain feature fusion tensor to generate a transient frequency feature vector; The adaptive noise suppression module is used to perform dynamic noise suppression based on the signal-to-noise ratio weight on the transient frequency feature vector 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 feature vector.

[0007] Preferably, the signal time-frequency transformation module includes: The signal segmentation and windowing unit is used to overlap and segment the radio frequency signals and apply Hamming window processing to generate a set of windowed signal segments; The time-frequency analysis unit is used to input the set of windowed signal segments into a short-time Fourier transform network to generate the signal time-frequency distribution feature tensor.

[0008] Preferably, the time-frequency analysis unit includes: The complex spectrum generation sub-unit is used to perform fast Fourier transform on each windowed signal segment to generate a set of complex spectra; The energy spectrum aggregation sub-unit is used to perform logarithmic energy conversion and time dimension stacking on the set of complex spectra to form the signal time-frequency distribution feature tensor.

[0009] Preferably, the noise floor coding module includes: The noise dictionary training unit is used to train and generate an over-complete noise dictionary from historical noise data using the K-SVD algorithm; The sparse representation unit is used to perform sparse decomposition of the current noise floor data on the over-complete noise dictionary using the orthogonal matching pursuit algorithm to generate the noise floor sparse coding feature matrix.

[0010] Preferably, the transient feature extraction module is configured to: after inputting the frequency-domain feature fusion tensor into a pyramid convolutional neural network for spatial dimension compression and channel dimension expansion, generate the transient frequency feature vector through global average pooling.

[0011] Preferably, the adaptive noise suppression module includes: A sub-band signal-to-noise ratio estimation unit, configured to calculate the signal-to-noise ratio at the frequency point level of the transient frequency feature vector to generate a signal-to-noise ratio weight vector; A dynamic filtering unit, configured to construct an adjustable exponential decay function according to the signal-to-noise ratio weight vector to correct the amplitude of the feature components; A phase preservation unit, configured to preserve the original phase information during the frequency-domain feature correction process to generate the corrected transient frequency feature vector.

[0012] Preferably, the dynamic filtering unit is configured to: construct an adaptive filtering response curve with frequency selection characteristics through the non-linear combination of the signal-to-noise ratio weight vector and a preset attenuation coefficient.

[0013] Preferably, the spectral image reconstruction module includes: A frequency point interpolation unit, configured to perform cubic spline interpolation on the corrected transient frequency feature vector to expand the spectral resolution; A color mapping unit, configured to convert the interpolated spectral data into an HSV color space mapping matrix; An image synthesis unit, configured to convert the HSV color space mapping matrix into an RGB format spectral heat map to generate the real-time spectral distribution image.

[0014] Preferably, the system further includes: A power consumption control module, configured to dynamically adjust the ADC sampling rate and the working clock frequency of the digital signal processing unit according to the signal strength; A power management unit, adopting a multi-voltage domain design to independently control the power supply for the RF front end, the digital processor, and the display module respectively.

[0015] Preferably, the power consumption control module includes an environment perception unit, which automatically switches to the low-power operation mode by monitoring the environmental temperature and the device working state in real time.

[0016] Compared with the prior art, the beneficial effects of the present invention are: At the signal processing level, the radio frequency signal is subjected to overlapping segmentation and Hamming window processing through the signal time-frequency transformation module, and the time-frequency distribution feature tensor is generated in combination with the short-time Fourier transform network, realizing the joint analysis of the time-domain and frequency-domain features of the signal. Compared with the traditional single-domain analysis method, it can more accurately capture the time-frequency variation characteristics of transient signals. The noise floor coding module trains an over-complete noise dictionary using the K-SVD algorithm and sparsely decomposes the current noise floor data through the orthogonal matching pursuit algorithm, constructing a dynamic representation model of noise features, enabling the system to adapt to changes in environmental noise, device background noise, and interference signals in real time, and providing an accurate noise reference floor for subsequent dynamic threshold fusion.

[0017] The dynamic threshold fusion module fuses the signal time-frequency distribution features and the noise floor sparse coding features to generate a frequency-domain feature fusion tensor, effectively solving the problem of insufficient adaptability of traditional fixed-threshold detection to complex noise and improving the signal detection reliability in low signal-to-noise ratio scenarios. The transient feature extraction module uses a pyramidal convolutional neural network for multi-scale feature aggregation. Through spatial dimension compression and channel dimension expansion, it enhances the feature expression ability for different frequency components and time scales. Combining with the transient frequency feature vector generated by global average pooling, it realizes the efficient compression of signal features and the retention of key information.

[0018] The adaptive noise suppression module generates a weight vector through frequency-point-level signal-to-noise ratio estimation, driving an adjustable exponential decay function to correct the amplitude of feature components while retaining the original phase information, realizing the dynamic adjustment of the noise suppression intensity, avoiding excessive damage to signal features by fixed filtering parameters, and maximizing the retention of signal integrity while effectively suppressing noise. The spectrum image reconstruction module improves the spectrum resolution through cubic spline interpolation and generates a real-time spectrum heat map by combining the color space mapping from HSV to RGB, providing an intuitive spectrum distribution visualization interface for users and facilitating the rapid positioning of signal features and interference sources.

[0019] In terms of power consumption management, the power consumption control module dynamically adjusts the ADC sampling rate and the working clock frequency of the digital signal processing unit according to the signal strength, avoiding excessive sampling and waste of computing resources in strong signal scenarios; the multi-voltage domain power management unit independently controls the power supply of the radio frequency front end, digital processor, and display module. Combining with the real-time monitoring of temperature and device status by the environment perception unit, it realizes the automatic switching of the low-power operation mode, 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

[0020] Figure 1 It is the working principle diagram of the low-power instantaneous frequency measurement system described in the present invention; Figure 2 It is the working principle diagram of the signal time-frequency transformation module; Figure 3 Schematic diagram of the working principle of the time-frequency analysis unit; Figure 4 Schematic diagram of the working principle of the transient feature extraction module. Specific implementation manners

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0022] Please refer to Figures 1 - 4 , a low-power instantaneous frequency measurement system involved in the present invention, the system includes: a radio frequency signal acquisition module, a signal time-frequency transformation module, a noise floor coding 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: First, the radio frequency signal acquisition module acquires the radio frequency signal and the noise floor data in the target frequency band, where the radio frequency signal includes the time-domain waveform, signal strength, and modulation type, and the noise floor data includes the environmental noise spectrum, device background noise, and interference signal characteristics. Then, the signal time-frequency transformation module performs time-frequency joint analysis on the radio frequency signal to generate a signal time-frequency distribution feature tensor. Next, the noise floor coding module performs sparse coding processing on the noise floor data to generate a noise floor sparse coding feature matrix. After that, the dynamic threshold fusion module fuses the signal time-frequency distribution feature tensor and the noise floor sparse coding feature matrix to generate a frequency-domain feature fusion tensor. Then, the transient feature extraction module 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 the signal-to-noise ratio weight to generate a corrected transient frequency feature vector. Finally, the spectrum image reconstruction module reconstructs the real-time spectrum distribution image of the target signal according to the corrected transient frequency feature vector.

[0023] The present invention will be further described below in conjunction with Embodiments 1 to 5: Embodiment 1: The signal time-frequency transformation module of the system includes a signal segmentation and windowing unit and a time-frequency analysis unit. The signal segmentation and windowing unit overlaps and segments the radio frequency signal. With a preset overlap ratio, the overlap length between two adjacent segments is 1 / 4 of the single segment length, ensuring the continuity of signal processing. After segmentation, a Hamming window is applied to each segment of the signal. The expression of the Hamming window is a specific combination form of cosine functions. By processing with this window function, the spectral leakage problem caused by signal truncation can be reduced, generating 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, which generates a set of complex spectra by performing a fast Fourier transform on each windowed signal segment. The set of complex spectra 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 sub-unit performs a logarithmic energy conversion on the set of complex spectra, converting the complex amplitude into a logarithmic energy value 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.

[0024] When the signal segmentation and windowing unit performs signal segmentation, it first preprocesses the input radio frequency signal, including removing the DC component and smoothing, to improve the signal quality. The preprocessed signal is sent to the segmentation module and segmented according to the preset segmentation length and overlap ratio. The selection of the segmentation length needs to comprehensively consider the characteristics of the signal and processing requirements. Too short a segmentation length will result in reduced frequency resolution, while too long a segmentation length will affect the time resolution. In this embodiment, the segmentation length is adaptively adjusted according to the bandwidth and sampling rate of the signal to ensure that the transient characteristics of the signal can be captured.

[0025] The design of overlapping segmentation is to avoid signal distortion at the segmentation boundary. By overlapping two adjacent segments, the continuity of the signal can be better restored in subsequent processing. In this embodiment, the selection of a 1 / 4 overlap ratio is the result of a trade-off, which can not only ensure the continuity of the signal but also not significantly increase the processing burden. Too small an overlap ratio may result in ineffective compensation for signal distortion at the segmentation boundary, while too large an overlap ratio will increase the computational amount and storage requirements.

[0026] Applying the Hamming window processing is one of the key steps of the signal segmentation and windowing unit. The Hamming window is a commonly used window function, and its expression is: .

[0027] Among them, is the sampling point index, is the length of the window function. The Hamming window is characterized by a moderate main lobe width and a relatively fast decay of side lobes, which can achieve a good balance between suppressing spectral leakage and maintaining frequency resolution. By applying the Hamming window to each segment of the signal, the spectral leakage problem caused by signal truncation can be reduced, making the spectral analysis more accurate.

[0028] When applying the Hamming window, it should be noted that the length of the window function should be the same as the length of the signal segment to ensure that the window function can completely cover the signal segment. At the same time, to avoid the edge effect of the window function, smoothing transitions are usually performed at both ends of the window function. In this embodiment, a symmetric extension method is adopted to symmetrically extend both ends of the window function and then multiply it with the signal segment to reduce the influence of the edge effect.

[0029] After receiving the set of windowed signal segments, the time-frequency analysis unit inputs it into the short-time Fourier transform network. The short-time Fourier transform is a commonly used time-frequency analysis method that obtains the distribution of the signal at different times and frequencies by windowing the signal and then performing the Fourier transform. In this embodiment, the fast Fourier transform algorithm is adopted in the short-time Fourier transform network to improve the calculation efficiency.

[0030] The fast Fourier transform algorithm is an efficient method for calculating the discrete Fourier transform, which reduces the computational complexity from O(N^2) to O(NlogN), where N is the length of the signal. In this embodiment, to further improve the calculation efficiency, the radix-2 fast Fourier transform algorithm is adopted, which requires the signal length to be a power of 2. If the length of the signal segment is not a power of 2, zero-padding processing is required to make its length meet the algorithm requirements.

[0031] By performing the fast Fourier transform on each windowed signal segment, a set of complex spectra is generated. Each element in the set of complex spectra contains the complex amplitude information of the corresponding signal segment at a certain frequency point. The complex amplitude consists of a real part and an imaginary part, which represent the amplitude and phase of the signal at that frequency point respectively. The set of complex spectra can comprehensively reflect the frequency components and phase relationships of the signal, providing rich information for subsequent signal analysis and processing.

[0032] The energy spectrum aggregation subunit performs logarithmic energy conversion and time dimension stacking processing on the set of complex spectra. Logarithmic energy conversion is the process of converting the complex amplitude into a logarithmic energy value, and its calculation formula is: .

[0033] where, is the logarithmic energy value, is the complex amplitude. Through logarithmic energy conversion, the dynamic range of the signal can be expanded, enabling both weak and strong signals to be better represented, and it is also convenient for subsequent processing and analysis.

[0034] Stack the converted energy values 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, and its 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 points 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.

[0035] After forming the signal time-frequency distribution feature tensor, it is also necessary to perform normalization processing on it to eliminate the energy differences between different signal segments. Normalization processing 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 values are within a suitable range.

[0036] The design and implementation of the signal time-frequency transformation module fully consider the time-frequency characteristics and processing requirements of the signal. Through the collaborative work of the signal segmentation windowing unit and the time-frequency analysis unit, it can accurately extract the time-frequency distribution characteristics of the signal, providing a solid foundation for subsequent signal processing and analysis. In practical applications, this module can be adjusted and optimized according to specific signal characteristics and processing requirements to obtain better processing effects.

[0037] Embodiment 2: The noise floor 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, which includes environmental noise spectra, device background noise, and interference signal characteristics in different environments, etc. The K-SVD algorithm iteratively updates the dictionary atoms and sparse representation coefficients to learn an over-complete noise dictionary from the historical noise data that can sparsely represent noise. The number of atoms in the over-complete noise dictionary is more than the signal feature dimension, which 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 floor data on the over-complete noise dictionary. The orthogonal matching pursuit algorithm constructs the sparse representation coefficients by gradually selecting the dictionary atoms most relevant to the current noise data, and finally generates a noise floor sparse coding feature matrix. This matrix represents the current noise floor data in a sparse form, highlighting the main components of the noise, which is convenient for subsequent fusion processing with signal features.

[0038] When the noise dictionary training unit conducts noise dictionary training, it first collects a large amount of historical noise data. These historical noise data come from different environments and scenarios, including indoor, outdoor, different frequency bands, etc., to ensure that various possible noise types and characteristics can be covered. The collected historical noise data needs to be preprocessed, including removing outliers, normalization processing, etc., to improve the quality and consistency of the data.

[0039] 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 alternately updating the dictionary atoms and the 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 continuously until the convergence condition is met.

[0040] In the dictionary update stage, the K-SVD algorithm adopts a greedy strategy, updating only one dictionary atom at a time while keeping the other atoms unchanged. Specifically, the algorithm selects a dictionary atom to be updated, and then calculates the contribution of this atom to all training samples based on the current sparse representation coefficients. By minimizing the reconstruction error, the value of this dictionary atom is updated. This process is repeated until all dictionary atoms have been updated once.

[0041] The design of the overcomplete noise dictionary is the key of the noise dictionary training unit. The number of atoms in the overcomplete noise dictionary is more than the signal feature dimension, which enables the dictionary to represent the complex structure of noise more flexibly. In this embodiment, the number of atoms in the overcomplete noise dictionary is adaptively adjusted according to the characteristics of the historical noise data and the processing requirements. Too many atoms will increase the computational complexity and storage requirements, while too few atoms may not be able to fully represent the characteristics of the noise.

[0042] To improve the quality and generalization ability of the overcomplete noise dictionary, a regularization term is introduced during the training process. The regularization term can constrain the characteristics of the dictionary atoms and avoid the occurrence of overfitting. In this embodiment, the L1 regularization term is adopted, which can promote the sparsity of the sparse representation coefficients, that is, most of the coefficients are zero, and only a few coefficients are non-zero. This sparsity enables the noise basis data to be represented concisely and highlights the main components of the noise.

[0043] After receiving the over-complete noise dictionary and the current noise basis data, the sparse representation unit performs sparse decomposition using the orthogonal matching pursuit algorithm. The orthogonal matching pursuit algorithm is a greedy algorithm that constructs sparse representation coefficients by gradually selecting the dictionary atoms most relevant to the current noise data. Specifically, the algorithm first initializes the residual as the current noise data, and then in 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 continuously until the stopping condition is met.

[0044] When selecting dictionary atoms, the orthogonal matching pursuit algorithm calculates the inner product of 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 most relevant to the current residual, thus enabling a rapid approximation of the sparse representation of the noise data.

[0045] When updating the residual, the orthogonal matching pursuit algorithm uses the Gram - Schmidt orthogonalization method to orthogonalize the selected dictionary atoms and then projects the residual onto the orthogonal complement space. This orthogonalization process ensures that the atoms selected each time are linearly independent, thereby improving the accuracy and stability of the sparse representation.

[0046] The stopping condition of the orthogonal matching pursuit algorithm can be set according to specific requirements. In this embodiment, the stopping condition is to reach a preset number of iterations or the energy of the residual 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.

[0047] Through the orthogonal matching pursuit algorithm, the current noise basis data is sparsely decomposed on the over - complete noise dictionary to generate sparse representation coefficients. These sparse representation coefficients form the noise basis sparse coding feature matrix. This matrix represents the current noise basis data in a sparse form, only retaining the coefficients related to the main components of the noise while ignoring the secondary components and noise. This sparse representation not only reduces the storage space of the data but also highlights the main features of the noise, facilitating subsequent fusion processing with signal features.

[0048] After generating the noise basis sparse coding feature matrix, post - processing is also required, including normalization processing and threshold processing. Normalization processing can eliminate the energy differences between different noise data, improving the stability and accuracy of subsequent processing. Threshold processing can further remove the secondary components in the noise, only retaining the noise features that have an important impact on signal processing.

[0049] The design and implementation of the noise floor 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 the characteristics of noise and represent them in a sparse form. In practical applications, this module can adjust and optimize parameters according to the specific noise environment and processing requirements to obtain better processing effects. For example, in the case of a complex and variable noise environment, the amount of historical noise data collection 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 the computational complexity.

[0050] Embodiment 3: The specific implementation method of the transient feature extraction module is as follows: The frequency-domain feature fusion tensor is input into the pyramid convolutional neural network. The pyramid convolutional neural network contains multiple convolutional layers with different scales, and the convolutional kernel sizes of each convolutional layer are different, which can extract multi-scale spatial features from the frequency-domain feature fusion tensor. Through convolutional operations of different scales, the spatial dimension compression of the feature tensor is realized, reducing the redundant information in the spatial dimension, and at the same time, the channel dimension is expanded to increase the expression ability of the features. After multiple convolutional processes, through the global average pooling operation, the features in the spatial dimension are averaged to generate the transient frequency feature vector. The global average pooling can retain the feature information in the channel dimension while reducing the dimension of the feature vector, facilitating subsequent noise suppression processing.

[0051] As the input of the pyramid convolutional neural network, the frequency-domain feature fusion tensor needs to be preprocessed to meet the input requirements of the network. The preprocessing process includes normalization processing. By adjusting the numerical range of the tensor, it is distributed within the effective interval of network training, improving the convergence speed and stability of the network. The normalization processing can adopt the global normalization method to standardize the mean and variance of each channel of the tensor to ensure the comparability of features in different channels.

[0052] The core of the pyramid convolutional neural network lies in the design of the multi-scale convolutional layer. The network contains multiple convolutional layers, and the convolutional kernel sizes of each layer are different, such as convolutional kernels of different sizes of 3×3, 5×5, 7×7, etc. are set in sequence. Convolutional kernels of different scales can capture spatial features in different ranges: small-size convolutional kernels (such as 3×3) are good at extracting local detail features and are more sensitive to high-frequency information; large-size convolutional kernels (such as 7×7) can cover a larger spatial range and extract global or low-frequency features. Through this multi-scale convolutional operation, the network can capture feature information at different levels from the frequency-domain feature fusion tensor, enhancing the diversity and richness of the features.

[0053] During the convolution operation, an activation function layer and a batch normalization layer are usually connected after each convolution layer. The activation function layer (such as the ReLU function) is used to introduce non-linear transformations, enabling the network to learn more complex feature relationships; the batch normalization layer normalizes the features output by the convolution to reduce internal covariate shift, accelerate network training, and improve generalization ability. The layers form a hierarchical feature extraction structure by stacking, with the output of the previous layer serving as the input of the next layer, gradually abstracting and refining the features.

[0054] Spatial dimension compression is achieved through the stride setting and padding strategy in the convolution operation. For example, in some convolution layers, a convolution operation with a stride of 2 is used to halve the spatial size of the output feature map (such as the size in the time and frequency dimensions), thereby reducing redundant information in the spatial dimension. At the same time, by controlling the number of convolution kernels and channels, channel dimension expansion is performed while compressing the spatial dimension. For example, if the number of channels of the input feature tensor is C, the number of channels increases to 2C after passing through the convolution layer, enabling the network to learn more dimensional feature representations and enhancing the expressive power of the features.

[0055] After multi-layer convolution processing, the spatial dimensions (such as time and frequency dimensions) of the feature tensor have been compressed to a relatively low size, while the channel dimension has increased significantly, forming a high-dimensional feature map. At this time, global average pooling is used to aggregate the features in the spatial dimension. Global average pooling takes the average of all elements in the spatial dimension (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, the length of which is equal to the number of channels, and this vector is the transient frequency feature vector.

[0056] The advantage of global average pooling is that it can retain the feature information in the channel dimension, avoid the overfitting problem that may be brought about by traditional fully connected layers, and at the same time significantly reduce the computational amount. Through the averaging operation, the local features in the spatial dimension are integrated into global features, enabling the generated transient frequency feature vector to comprehensively reflect the multi-scale information in the frequency domain feature fusion tensor.

[0057] During the network training process, the backpropagation algorithm is used to optimize the network parameters. By defining a loss function (such as mean squared error loss or cross-entropy loss, specifically determined according to the subsequent task requirements), the difference between the predicted value and the true value is calculated, and the error is backpropagated to each layer of the network to update the weights and bias parameters of the convolution kernels. Stochastic gradient descent (SGD) and its variant algorithms (such as the Adam optimizer) can be used to adjust the learning rate during the training process to improve the training efficiency and convergence accuracy.

[0058] The structural design of the pyramid convolutional neural network needs to balance the computational complexity and the feature extraction ability. For example, increasing the number of convolutional layers and the size of the convolutional kernels will enhance the richness of feature extraction but also lead to a sharp increase in the computational volume; conversely, reducing the number of layers or using small-sized convolutional kernels may not be able to capture sufficient multi-scale features. Therefore, in practical applications, it is necessary to reasonably design parameters such as the number of network layers, the size of convolutional kernels, and the number of channels according to the specific dimensions of the frequency-domain feature fusion tensor and the computational resource limitations.

[0059] In addition, to avoid overfitting of the network, a regularization method such as a dropout layer can be introduced between convolutional layers. The dropout layer randomly discards the outputs of some neurons with a certain probability during the training process, forcing the network to learn more robust feature representations and improving the generalization ability of the model.

[0060] The transient feature extraction module can effectively extract the spatio-temporal features in the frequency-domain feature fusion tensor through the multi-scale feature aggregation mechanism of the pyramid convolutional neural network, and generate a compact transient frequency feature vector through global average pooling. The design of this module makes full use of the advantages of convolutional neural networks in image processing and feature extraction, combines the multi-scale analysis strategy, and provides high-quality feature representations for subsequent adaptive noise suppression and spectral image reconstruction. In practical applications, the network structure and parameters can be flexibly adjusted according to different signal types and frequency measurement requirements to optimize the feature extraction effect.

[0061] Embodiment 4: The adaptive noise suppression module includes a sub-band signal-to-noise ratio estimation unit, a dynamic filtering unit, and a phase preservation unit. The sub-band signal-to-noise ratio estimation unit calculates the signal-to-noise ratio at the frequency point level for the transient frequency feature vector. For the feature components at each frequency point, it calculates the ratio of the signal energy to the noise energy to generate a signal-to-noise ratio weight vector. Each element in this 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 decay function according to the signal-to-noise ratio weight vector, and determines the parameters of the decay function through the non-linear combination of the signal-to-noise ratio weight vector and a preset decay coefficient, so that the decay function has frequency selection characteristics. For frequency points with a higher signal-to-noise ratio, the decay amplitude of the decay function is smaller, retaining more signal features; for frequency points with a lower signal-to-noise ratio, the decay amplitude is larger, suppressing the noise components. During the process of amplitude correction of the feature components, the phase preservation 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.

[0062] After receiving the transient frequency feature vector, the sub-band signal-to-noise ratio (SNR) estimation unit first divides it into multiple sub-bands. Each sub-band contains several adjacent frequency points, and the division method can be determined according to specific application requirements, such as equal-bandwidth division or adaptive division based on signal characteristics. For the frequency points within each sub-band, the signal energy and noise energy are estimated respectively. The estimation of signal energy can be based on the characteristic amplitude of the frequency point and obtained through statistical analysis of the amplitudes over a period of time. The estimation of noise energy requires using the noise basis sparse coding feature matrix generated by the noise basis coding module, extracting the noise feature information corresponding to the frequency point, and dynamically adjusting it in combination with the characteristics of the current signal.

[0063] When estimating the SNR of each frequency point, the sub-band SNR estimation unit performs time-domain analysis using a sliding window. By sliding the window in the time dimension, the ratio of the signal energy and noise energy within the window is calculated to obtain the SNR estimation values of the frequency point at different times. To improve the estimation accuracy, smoothing filtering technology can also be used to process the SNR estimation values and reduce estimation fluctuations.

[0064] Each element in the generated SNR weight vector corresponds to the SNR weight of a frequency point, and the magnitude of the weight value reflects the quality of the signal at that frequency point. The frequency points with higher SNR correspond to larger weight values, indicating higher signal reliability at these frequency points; the frequency points with lower SNR correspond to smaller weight values, indicating that the signals at these frequency points may be severely contaminated by noise.

[0065] After receiving the SNR weight vector, the dynamic filtering unit constructs an adjustable exponential decay function based on this vector. The parameters of the adjustable exponential decay function are determined by the non-linear combination of the SNR weight vector and the preset decay coefficient. The preset decay coefficient is a set of pre-set parameters used to control the basic shape and characteristics of the decay function. By performing a non-linear combination of the SNR weight vector and the preset decay coefficient, the decay function can have frequency selection characteristics, that is, signals at different frequency points are attenuated to different degrees.

[0066] For the frequency points with higher SNR, the attenuation amplitude of the adjustable exponential decay function is smaller, so that the signal characteristics of these frequency points can be better retained; for the frequency points with lower SNR, the attenuation amplitude is larger, thus effectively suppressing the noise components at these frequency points. This frequency selection characteristic enables the dynamic filtering unit to adaptively adjust the filtering strategy according to the actual situation of the signal and improve the noise suppression effect.

[0067] When constructing the adjustable exponential decay function, the dynamic filtering unit also takes into account the time-varying characteristics of the signal. By dynamically adjusting the decay function in the time dimension, the filtering strategy can better adapt to the changes in the signal. For example, for rapidly changing signals, the adjustment speed of the decay function is correspondingly increased to ensure that the changes in the signal can be tracked in a timely manner; for slowly changing signals, the adjustment speed of the decay function can be appropriately slowed down to reduce unnecessary fluctuations.

[0068] During the process of the dynamic filtering unit correcting the amplitude of the feature components, the phase preservation unit is responsible for retaining the original phase information. Phase information is crucial for the correct reconstruction and understanding of signals, especially when dealing with signals with phase-sensitive characteristics. The phase preservation unit ensures that the phase characteristics of the signal are not affected by tracking the phase changes of the original signal and reapplying the original phase information to the corrected signal after amplitude correction.

[0069] In actual implementation, the phase preservation unit can adopt various methods to retain the phase information. A common method is to extract and store the phase information of the feature components before correcting their amplitudes. After the amplitude correction is completed, the stored phase information is reapplied to the corrected feature components. Another method is to design a special filtering algorithm so that the filtering process itself can maintain the phase characteristics of the signal.

[0070] The three units of the adaptive noise suppression module work together to effectively suppress the noise of the transient frequency feature vector. The subband signal-to-noise ratio estimation unit accurately estimates the signal-to-noise ratio of each frequency point, providing a decision basis for the dynamic filtering unit; the dynamic filtering unit constructs an adjustable exponential decay function based on the signal-to-noise ratio weight vector to adaptively filter the signals at different frequency points; the phase preservation unit ensures that the phase information of the signal is retained during the noise suppression process.

[0071] In practical applications, the performance of the adaptive noise suppression module is also affected by various factors. For example, the subband division method, the accuracy of signal-to-noise ratio estimation, the design of the adjustable exponential decay function, and the selection of the phase preservation method, etc., will all have an important impact on the noise suppression effect. Therefore, in specific implementation, these factors need to be reasonably selected and optimized according to the actual application scenario and signal characteristics.

[0072] To improve the real-time performance and computational efficiency of the module, parallel computing technology can also be used to optimize the adaptive noise suppression module. For example, using the parallel computing capabilities of multi-core processors or GPUs to process the signals of multiple subbands simultaneously to speed up the calculation. In addition, the algorithm can be simplified and approximated to reduce the computational complexity while ensuring the noise suppression effect.

[0073] The adaptive noise suppression module realizes the adaptive noise suppression of the transient frequency feature vector through three key steps: sub-band signal-to-noise ratio estimation, dynamic filtering, and phase preservation. This module can automatically adjust the filtering strategy according to the actual situation of the signal, effectively suppress noise while retaining the important features of the signal, and provides high-quality feature vectors for subsequent spectral image reconstruction. In practical applications, this module can be widely used in fields such as communication, radar, and sonar to improve the quality and reliability of signal processing.

[0074] Embodiment 5: The spectral 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 realizes the expansion of spectral resolution by constructing cubic polynomials between adjacent frequency points, making the reconstructed spectral data denser. The color mapping unit converts the interpolated spectral data into an HSV color space mapping matrix. The HSV color space contains three components: hue, saturation, and value, which can more intuitively represent the energy distribution of the spectral data. In the specific conversion process, the value component is determined according to the energy value of the spectral data, the hue component is determined according to 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 an RGB format spectral heat map. The RGB format is a common image display format that generates an intuitive spectral heat map through the combination of three colors, thereby generating a real-time spectral distribution image of the target signal, facilitating users to intuitively observe the frequency distribution and energy change of the signal.

[0075] When the frequency point interpolation unit performs cubic spline interpolation, it first needs to determine the interpolation nodes and interpolation intervals. Each element in the corrected transient frequency feature vector corresponds to the eigenvalue of a frequency point. These frequency points are usually equally spaced on the frequency axis, but the interval may be large, resulting in low spectral resolution. To improve the spectral resolution, new frequency points need to be inserted between adjacent frequency points to make the sampling points on the frequency axis denser.

[0076] The basic idea of cubic spline interpolation is to divide the entire interpolation interval into several small segments, and use cubic polynomials for interpolation on each small segment. Within each small segment, the cubic polynomial needs to satisfy the conditions that the function values and the first and second derivatives at the endpoints are continuous to ensure the smoothness of the interpolation curve. Through this piecewise interpolation method, more frequency point data can be generated without increasing the amount of original data, improving the spectral resolution.

[0077] When implementing cubic spline interpolation, it is necessary to preprocess the original frequency point data first, including sorting and removing duplicates, to ensure that the frequency points are arranged in the order of frequency and there are no duplicate points. Then, calculate the spacing between adjacent frequency points to determine the interpolation step size. The choice of the interpolation step size needs to be adjusted according to actual requirements. The smaller the step size, the denser the interpolated frequency points, the higher the spectral resolution, but the computational complexity will also increase accordingly.

[0078] After receiving the interpolated spectral data, the color mapping unit converts it into an HSV color space mapping matrix. The HSV color space is a color model oriented to visual perception, where H represents hue, corresponding to the types of colors, such as red, green, blue, etc.; S represents saturation, corresponding to the vividness of the color; V represents value, corresponding to the brightness of the color.

[0079] During the conversion process, first, it is necessary to determine the mapping relationship between the energy value of the spectral data and each component of HSV. The value component V is usually directly related to the energy value of the spectral data. The higher the energy value, the higher the value, and the brighter the corresponding color; the lower the energy value, the lower the value, and the darker the corresponding color. The hue component H can be mapped according to the frequency range. For example, map the low-frequency band to red, the middle-frequency band to green, and the high-frequency band to blue. In this way, signals of different frequencies can be represented by different colors, facilitating users to visually distinguish the high and low frequencies.

[0080] The setting of the saturation component S can be adjusted according to actual requirements. Generally speaking, the saturation can be set to a fixed value or adjusted dynamically according to the size of the energy value. For example, a higher saturation can be set in the area with a higher energy value to make the color more vivid and highlight the strong signal; a lower saturation can be set in the area with a lower energy value to make the color softer and avoid the interference of weak signals on vision.

[0081] The image synthesis unit converts the HSV color space mapping matrix into a spectral heat map in RGB format. The RGB color space is a color model based on the three primary colors. By mixing different proportions of red (R), green (G), and blue (B), various colors can be generated. The conversion from HSV to RGB is a non-linear process, and the corresponding RGB values need to be calculated according to the values of each component of HSV.

[0082] During the specific conversion process, first determine the main color tone range to which the color belongs according to the hue H, and then use different formulas within each range to calculate the values of the RGB components. For example, when the hue H is in the range from red to green, the red component gradually decreases from the maximum value, the green component gradually increases from the minimum value, and the blue component remains at the minimum value; when the hue H is in the range from green to blue, the green component gradually decreases from the maximum value, the blue component gradually increases from the minimum value, and the red component remains at the minimum value, and so on.

[0083] After completing the conversion from HSV to RGB, the image synthesis unit combines the RGB values of all pixels into a two-dimensional matrix, namely the spectral heat map. The abscissa of this heat map represents the frequency, and the ordinate can represent time or other relevant dimensions. The color of each pixel corresponds to the spectral energy value at this frequency point and time point. In this way, users can intuitively observe the energy distribution of the target signal at different frequencies and times, and quickly identify the frequency components, intensity changes, and potential interference signals of the signal.

[0084] To improve the display effect of the spectral heat map, the image synthesis unit can also perform some post-processing operations, such as contrast adjustment, brightness adjustment, edge enhancement, etc. Contrast adjustment can expand the color difference between different energy values and make the spectral distribution clearer; brightness adjustment can overall adjust the brightness of the image to adapt to different display environments; edge enhancement can highlight the edge information in the spectrum and facilitate users to identify the boundaries and mutation points of the signal.

[0085] The design and implementation of the spectral image reconstruction module fully consider the characteristics of human visual perception. It improves the spectral resolution through frequency point interpolation, intuitively represents the spectral energy distribution through color mapping, and generates an easy-to-observe spectral heat map through image synthesis. In practical applications, this module can adjust parameters according to different display devices and user requirements, such as adjusting the interpolation step, color mapping rules, image resolution, etc., to obtain the best display effect.

[0086] In addition, to achieve real-time spectral image reconstruction, it is necessary to ensure that the computational efficiency of each unit meets the real-time requirements. The frequency point interpolation unit can adopt optimized interpolation algorithms and parallel computing technologies to accelerate the interpolation speed; the color mapping unit and the image synthesis unit can utilize hardware acceleration functions, such as the parallel computing ability of the graphics processing unit (GPU), to improve the image processing speed. Through these optimization measures, the spectral image reconstruction module can generate the spectral distribution image of the target signal in real time, providing users with timely signal analysis and monitoring means.

[0087] It should be noted that, in this text, relational terms such as first and second are only used 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 term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0088] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A low-power instantaneous frequency measurement system, characterized in that Including: A radio frequency signal acquisition module, configured to acquire radio frequency signals and noise floor data in a target frequency band, where the radio frequency signals include time-domain waveforms, signal intensities, and modulation types, and the noise floor data includes environmental noise spectra, device background noise, and interference signal characteristics; A signal time-frequency transformation module, configured to perform joint time-frequency analysis on the radio frequency signals to generate a signal time-frequency distribution feature tensor; A noise floor coding module, configured to perform sparse coding processing 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 based on signal-to-noise ratio weights on the transient frequency feature vector to generate a corrected transient frequency feature vector; A spectrum image reconstruction module, configured to reconstruct a real-time spectrum distribution image of a target signal according to the corrected transient frequency feature vector.

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

3. The low-power instantaneous frequency measurement system according to claim 2, wherein The time-frequency analysis unit includes: A complex spectrum generation sub-unit, configured to perform fast Fourier transform on each windowed signal segment to generate a set of complex spectra; An energy spectrum aggregation sub-unit, configured to perform logarithmic energy conversion and time dimension stacking on the set of complex spectra 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 coding module includes: A noise dictionary training unit, configured to train an over-complete noise dictionary using historical noise data by the K-SVD algorithm; A sparse representation unit, configured to perform sparse decomposition of current noise floor data on the over-complete noise dictionary using the orthogonal matching pursuit algorithm to generate the noise floor sparse coding feature matrix.

5. The low-power instantaneous frequency measurement system according to claim 4, characterized in that, The transient feature extraction module is configured to: input the frequency-domain feature fusion tensor into a 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. The low-power instantaneous frequency measurement system according to claim 5, characterized in that, The adaptive noise suppression module includes: A sub-band signal-to-noise ratio estimation unit, configured to calculate the signal-to-noise ratio at the frequency point level of the transient frequency feature vector to generate a signal-to-noise ratio weight vector; A dynamic filtering unit, configured to construct an adjustable exponential decay function according to the signal-to-noise ratio weight vector to correct the amplitude of the feature components; A phase preservation unit, configured to preserve 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 configured to: construct an adaptive filtering response curve with frequency selection characteristics through the non-linear combination of the 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 interpolation unit for performing cubic spline interpolation on the corrected transient frequency feature vector to expand the spectrum resolution; A color mapping unit for converting the interpolated spectrum data into an HSV color space mapping matrix; An image synthesis unit for converting the HSV color space mapping matrix into an RGB-format spectrum heat map to generate the real-time spectrum distribution image.

9. A low-power instantaneous frequency measurement system according to claim 1, characterized in that Further comprising: A power consumption control module for dynamically adjusting the ADC sampling rate and the operating clock frequency of the digital signal processing unit according to the signal strength; A power management unit that uses a multi-voltage domain design to independently control the power supply of the RF front end, the digital processor, and the display module.

10. A low-power instantaneous frequency measurement system according to claim 9, characterized in that, The power consumption control module includes an environment perception unit that automatically switches to a low-power operation mode by real-time monitoring the environmental temperature and the device operating state.

Images