A non-backtracking intelligent frequency sweeping method based on a lightweight neural network

Abstract

The application discloses a non-backtracking intelligent frequency sweeping method based on a lightweight neural network and belongs to the technical field of intelligent frequency sweeping. The method specifically comprises the following steps: firstly, full-band is rapidly scanned in starting to extract the mean and standard deviation of the noise power; after the starting is completed, the power values of each frequency point in the window function coverage interval are collected to extract a feature vector, which is input into a pre-trained lightweight neural network to output a confidence level; when the confidence level exceeds a threshold, the frequency step and the resolution bandwidth are dynamically reduced through a non-backtracking frequency sweeping control mechanism; otherwise, the parameters remain unchanged and the noise statistics are updated; finally, the collected power values are subjected to bandwidth power normalization and window function gain correction to obtain final power values and draw a spectrum diagram. The method successfully solves the contradiction between the frequency sweeping efficiency and the hardware lock overhead under the premise of ensuring the spectrum detection accuracy, and provides an efficient and feasible solution for high-bandwidth and high-real-time spectrum analysis tasks.

Description

Technical Field

[0001] This invention belongs to the field of intelligent frequency scanning technology, and particularly relates to a non-backtracking intelligent frequency scanning method based on a lightweight neural network. Background Technology

[0002] Superheterodyne spectrum analyzers have a wide range of applications and play an irreplaceable role in modern wireless communication and electromagnetic environment monitoring. Their applications include, but are not limited to, wireless communication R&D, electromagnetic compatibility testing, and spectrum monitoring. Due to their extremely wide range of applications, especially as they expand into higher frequency bands, the demand for scanning speed and real-time performance has exploded. This makes optimizing the underlying scanning algorithm crucial for enhancing the instrument's competitiveness. The core working principle of a superheterodyne spectrum analyzer involves mixing the radio frequency (RF) signal with a signal generated by a local oscillator (LO) via a mixer, thereby down-converting the signal to a fixed intermediate frequency (IF), and then performing energy detection via an RBW filter.

[0003] Resolution bandwidth (RBW) is a crucial parameter in superheterodyne spectrum analyzers, specifically the 3dB bandwidth of the RBW filter. It defines the minimum frequency difference between two different frequency signals that the instrument can clearly distinguish, directly determining the measurement's frequency resolution and noise floor level. The instrument can only clearly distinguish different spectral lines when the frequency difference between the two signals is greater than or equal to the resolution bandwidth; the smaller the resolution bandwidth, the higher the resolution of the superheterodyne spectrum analyzer.

[0004] The sweep time of a superheterodyne spectrum analyzer is also closely related to its resolution bandwidth, primarily because the physical response of the RBW filter has a time limitation. Therefore, when performing a full-band scan, a superheterodyne spectrum analyzer often faces a trade-off between high resolution and high scan speed. Increasing the RBW for higher scan speed makes it impossible to distinguish subtle, adjacent signals. Decreasing the RBW for higher resolution significantly increases the scan time. Furthermore, using the same small RBW in the noise floor region (no-signal frequency bands) results in a substantial waste of computing power and time.

[0005] To address this issue, traditional methods often employ segmented triggering scanning algorithms based on fixed thresholds, primarily using the following two backtracking implementation logics:

[0006] Instantaneous trigger backtracking: A coarse scan with a large RBW / large step is performed. Once the power at a certain frequency exceeds a fixed threshold, backtracking is immediately initiated and a fine scan with a small RBW is switched until the power falls below the threshold, at which point the coarse scan is switched back. However, when multiple discrete small signals exist in the environment, the local oscillator needs to be frequently interrupted, backtracked, and undergo multiple lock-in times, resulting in a longer total scan time than a normal scan.

[0007] Retrospective regression: First, a large RBW scan of the entire frequency band is performed and the frequency points exceeding the standard are recorded. Then, a fine scan is performed on these frequency points, and finally, the coarse and fine scan data are stitched together in the software. However, due to the time difference, it is very easy to miss burst signals that last only a few milliseconds.

[0008] In addition, both of the above segmented trigger scanning algorithms have the problem that the fixed threshold cannot adapt to changing environmental thermal noise or antenna impedance mismatch, and will frequently trigger invalid fine scans when noise fluctuates, which greatly reduces the optimization effect. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose a non-backtracking intelligent frequency sweeping method based on a lightweight neural network.

[0010] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0011] This invention provides a non-backtracking intelligent frequency sweeping method based on a lightweight neural network, comprising the following steps:

[0012] S1. During the startup process of the superheterodyne spectrum analyzer, the oscillator performs a single full-band scan by stepping at the maximum frequency to match the maximum resolution bandwidth, thereby extracting the background noise power data as quickly as possible and calculating the power mean representing the current ambient noise and the standard deviation used to measure power fluctuations.

[0013] S2. After the superheterodyne spectrum analyzer is started, the oscillator performs signal acquisition by stepping the corresponding resolution bandwidth at a preset frequency and in real time obtains the original power values ​​of each frequency point within the frequency range covered by the window function, which are used to extract the feature vectors input to the lightweight neural network.

[0014] S3. Input the extracted feature vector into the pre-trained lightweight neural network. The network outputs a confidence level that reflects the probability of the signal’s existence through a combination of weight distribution and nonlinear logic. When the confidence level exceeds the preset confidence threshold, the frequency step and resolution bandwidth are dynamically adjusted through a non-backtracking frequency sweep control mechanism based on confidence level prediction, and the scanning continues according to the adjusted frequency step and resolution bandwidth.

[0015] S4. Perform bandwidth power normalization and window function gain correction on the original power values ​​at each frequency point to obtain the final power value at each frequency point, which is used to draw a spectrum diagram and visualize the output.

[0016] Based on the above scheme, each step can be implemented in the following preferred manner.

[0017] Preferably, in S2, the length of the window function is equal to the main lobe bandwidth correction factor of the window function multiplied by the sampling frequency, and then divided by the current resolution bandwidth.

[0018] Preferably, in S2, the feature vector input to the lightweight neural network consists of three feature components: the standardized power peak, the standardized power variance, and the normalized spectral energy entropy.

[0019] Preferably, both the peak power and the power variance are obtained based on the power sequence, which is a sequence of the original power values ​​of all frequency points within the frequency range covered by the current window function.

[0020] As a preferred method, the calculation process of spectral energy entropy is as follows: within the frequency range covered by the window function, the original power values ​​of each frequency point are converted into a probability distribution through normalization, and then the uncertainty measure of the probability distribution is calculated through the definition of information entropy in information theory. Finally, the result of the uncertainty measure is used as the spectral energy entropy.

[0021] Preferably, in S3, the inference process of the lightweight neural network is as follows: multiple neurons perform linear fusion of feature components in parallel, and then the linear result output by each neuron is processed by a nonlinear activation function to obtain the activation state of the neuron. After linearly fusing the activation states of all neurons, the output result is generated. Finally, the output result is compressed by the Sigmoid logistic regression function to obtain the confidence level.

[0022] As a preferred embodiment, in S3, the non-backtracking frequency sweep control mechanism based on confidence prediction is as follows: when the output confidence exceeds the preset confidence threshold, the frequency range covered by the window function is determined as a potential signal area, and the current frequency step and resolution bandwidth are reduced; otherwise, the frequency range covered by the window function is determined as a noise floor area, and the current frequency step and resolution bandwidth remain unchanged. At the same time, the noise floor power data in this frequency range is extracted to dynamically update the power mean and standard deviation of the current environmental noise.

[0023] Preferably, in the bandwidth power normalization of S4, the original power value is added to the correction term to obtain the power value after bandwidth power normalization. The correction term adopts a logarithmic relationship and is determined by the ratio of the reference resolution bandwidth to the current resolution bandwidth.

[0024] Preferably, in the window function gain correction of S4, the length of the window function is divided by the sum of all coefficients of the window function to calculate the amplitude correction factor and obtain the compensation amount in decibels. The magnitude of the compensation amount depends on the amplitude correction factor.

[0025] Preferably, in S4, the final power value is the sum of the power value after bandwidth power normalization and the compensation amount.

[0026] Compared with the prior art, the method of the present invention has the following advantages:

[0027] This invention utilizes a lightweight neural network to achieve high-precision nonlinear identification of signals and random impulse interference, solving the problem of false triggering due to background noise fluctuations in existing technologies, resulting in an extremely low false alarm rate in complex dynamic noise environments. Specifically, this invention constructs a multi-dimensional feature vector based on peak power, power variance, and spectral energy entropy, and uses a lightweight neural network to mine the "entropy reduction" characteristic (ordered energy distribution) and stability characteristics of power variance exhibited by the signal in the frequency domain. It can effectively identify weak signals submerged in sudden, high-power impulse noise (high entropy, high variance), solving the problem of false triggering that is easily generated by the traditional amplitude threshold method in environments with severe background noise fluctuations or complex interference.

[0028] This invention utilizes the leading-edge geometric lead characteristics of a sliding window on the frequency axis to achieve seamless switching of look-ahead parameters across cycles during a unidirectional advance. Specifically, this invention maintains a sliding window of a specific bandwidth during the scanning process; when the leading edge of the sliding window touches the starting skirt of the potential signal before the current local oscillator center frequency, a lightweight neural network performs millisecond-level real-time inference on the multi-dimensional features already collected within the window and outputs the signal region confidence score. When this confidence score is greater than a threshold, the resolution bandwidth and scanning step of the next advance cycle are dynamically adjusted without interrupting the current frequency scanning process or performing a backtracking rescan.

[0029] This invention ensures high-precision physical consistency of scanned data at different resolutions by adjusting the frequency step and resolution bandwidth, normalizing the bandwidth and power of the original power value, and correcting the window function gain. Specifically, this invention configures the corresponding frequency step and resolution bandwidth based on the confidence level of the lightweight neural network output, and designs compensation and correction terms to correct the acquired original power value, thus maintaining the accuracy of data acquired in different precision modes (fast scan and fine scan). Attached Figure Description

[0030] Figure 1 This is a flowchart of the present invention;

[0031] Figure 2 This is a schematic diagram of the inference process of the lightweight neural network of the present invention;

[0032] Figure 3 This is a schematic diagram illustrating real-time parameter switching based on a non-backtracking frequency sweep control mechanism.

[0033] Figure 4 This is a comparison chart of the scanning efficiency of different schemes provided in this embodiment;

[0034] Figure 5 A comparison chart of hardware relock overhead for different schemes provided in this embodiment;

[0035] Figure 6This is a schematic diagram illustrating the verification of the physical consistency of the spectrum after variable parameter scanning for different schemes provided in this embodiment. Detailed Implementation

[0036] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0037] like Figure 1 As shown, in a preferred embodiment of the present invention, the above-mentioned non-backtracking intelligent frequency sweeping method based on a lightweight neural network includes the following steps S1 to S4. The specific implementation process of each step will be described in detail below.

[0038] S1. During the startup process of the superheterodyne spectrum analyzer, the oscillator steps at its maximum frequency. Adapt to maximum resolution bandwidth A single full-band scan is performed to extract the noise floor data most quickly and calculate the power mean representing the current ambient noise. and the standard deviation used to measure power fluctuations .

[0039] It should be noted that the extraction of background noise power data and the calculation of the mean and standard deviation of environmental noise power in S1 of this invention are common practices in the field, and therefore will not be described in detail in this invention.

[0040] S2. After the superheterodyne spectrum analyzer is started, the oscillator steps at a preset frequency. Adapt to the corresponding resolution bandwidth Signal acquisition is performed, and the raw power values ​​of each frequency point within the frequency range covered by the window function are obtained in real time, which are used to extract the feature vectors input to the lightweight neural network.

[0041] It should be noted that in S2 of this invention, the frequency step and resolution bandwidth (RBW) are two closely related but functionally different parameters, which together determine the scanning accuracy and speed. In signal acquisition, a window function is established to obtain the raw power values ​​of all frequency points within that window function. Specifically, in this embodiment, the length of the window function is... In other words, within each window function, it is necessary to obtain... The original power values ​​at each frequency point. The length of the window function. Equal to the main lobe bandwidth correction factor of the window function With sampling frequency Multiply, then divide by the current resolution bandwidth. , represented as:

[0042]

[0043] In terms of physical logic, the above formula for calculating the window function length originates from the resolution bandwidth in the Fourier transform. With effective signal observation duration The reciprocal relationship, that is, for the ideal rectangular window function, satisfies In digital sampling systems, the effective observation duration... Furthermore, the number of sampling points (i.e., the window function length) With sampling frequency Decision, satisfaction Considering that in practical engineering, non-rectangular window functions used to suppress spectral leakage widen the main lobe while suppressing sidelobes, leading to a decrease in resolution, this invention introduces the aforementioned main lobe bandwidth correction factor to dynamically adapt the window function length, thereby ensuring that the digital filter can accurately meet the preset resolution limit at a specific sampling rate. The specific value of the aforementioned main lobe bandwidth correction factor depends on the preset window function type. When using a Kaiser-Bessel window, K is set to 2.23–2.3; when using a Hanning window, K is set to 1.44; when using a Hamming window, K is set to 1.30; and when using a flat-top window, K is set to 3.77.

[0044] It should be noted that in S2 of this invention, the feature vector input to the lightweight neural network It consists of three characteristic components, namely the normalized peak power. Standardized power variance and the normalized spectral energy entropy ,Right now .

[0045] Furthermore, both the power peak and power variance are obtained based on the power sequence, which is a sequence of the original power values ​​of all frequency points within the frequency range covered by the current window function.

[0046] In this embodiment, the significance of signal strength relative to the dynamic noise floor benchmark is characterized by the standardization of peak power; the dispersion of energy distribution is measured by the standardization of power variance. Specifically, the standardization methods for peak power and power variance are as follows:

[0047]

[0048]

[0049] in, and These correspond to the peak power and power variance before standardization, respectively; The standardization coefficient is a preset value, typically between 1.0 and 1.2, under ideal Gaussian white noise. This represents the mathematical expectation of the noise power. However, in actual engineering, due to ADC quantization errors or the nonlinear flatness of the front-end amplifier, the average noise power floor will fluctuate within a small range. Setting it to 1.2 is to avoid the lightweight neural network being interfered with by extremely weak hardware stray interference during the initialization phase.

[0050] Furthermore, the calculation process of spectral energy entropy is as follows: within the frequency range covered by the window function, the original power values ​​of each frequency point are converted into a probability distribution through normalization, and then the uncertainty measure of the probability distribution is calculated through the definition of information entropy in information theory. Finally, the result of the uncertainty measure is used as the spectral energy entropy.

[0051] In this embodiment, spectral energy entropy quantitatively characterizes the degree of concentration or dispersion of spectral energy along the frequency axis: the higher the entropy value, the more uniform the energy distribution; the lower the entropy value, the more concentrated the energy is at a few frequency points. When calculating spectral energy entropy, the normalized probability of each frequency point within the window function is first calculated:

[0052]

[0053] in, For the first window function The normalized probability of the i-th frequency point, i.e., the i-th frequency point The proportion of the original power value at each frequency point to the total power value within the window function is obtained by summing the original power values ​​at all frequency points within the window function, i.e., the corresponding denominator. ; For the first time within this window function The original power values ​​at each frequency point; For the first time within this window function The original power values ​​at each frequency point are then used. Subsequently, the spectral energy entropy within the current window function is calculated using Claude Shannon's information entropy formula. :

[0054]

[0055] Finally, in this embodiment, the spectral energy entropy is normalized according to the following formula. By eliminating the dimensional influence of the window function length, the ordered characteristics of the energy distribution are locked within a certain range. Within the interval:

[0056]

[0057] S3. The extracted feature vector is input into a pre-trained lightweight neural network. The network outputs a confidence level reflecting the probability of the signal's existence through a combination of weight distribution and nonlinear logic. When the confidence level exceeds a preset confidence threshold, the frequency step and resolution bandwidth are dynamically adjusted through a non-backtracking frequency sweep control mechanism based on confidence level prediction, and the scanning continues according to the adjusted frequency step and resolution bandwidth.

[0058] It should be noted that in S3 of this invention, the inference process of the lightweight neural network is as follows: multiple neurons perform linear fusion of feature components in parallel, and then the linear result output by each neuron is processed by a nonlinear activation function to obtain the activation state of the neuron. After linearly fusing the activation states of all neurons, the output result is generated. Finally, the output result is compressed by the Sigmoid logistic regression function to obtain the confidence level.

[0059] In the lightweight neural network of this embodiment S3, such as Figure 2 As shown, each neuron performs a weighted summation of the three feature components of the input to achieve linear fusion, aiming to map the original statistical features to a high-dimensional feature space. For the ... Each neuron outputs a linear result. Represented as:

[0060]

[0061] in, For feature component index; For the first The corresponding neuron in the nth neuron each feature component The weight matrix, which has been pre-trained, It can be the standardized peak power, the standardized power variance, or the normalized spectral energy entropy; For the first The bias term of each neuron is then applied to the linear result using a nonlinear activation function. The process is then performed. The nonlinear activation function here can be the ReLU function, which utilizes its "one-sided inhibition" property as a feature filter, allowing only salient features that conform to a specific pattern combination (such as "high power and low energy entropy") to pass through, thus obtaining the activation state. This effectively eliminates random stray and invalid interference information in the environmental background. To obtain the maximum value, the activation states of all neurons are then aggregated to the output layer neurons. The activation states extracted in the previous step are then weighted and aggregated again to generate the output result. :

[0062]

[0063] in, and For the first The weight matrix and bias terms corresponding to each output layer neuron. The above output results combine the orderliness and significance of the energy distribution within the window, and are the final criterion for distinguishing between "target signal" and "random noise". To obtain standardized discrimination probabilities, this embodiment uses the Sigmoid logistic regression function to... Compress to Within the interval, the confidence level is obtained. :

[0064]

[0065] It should be noted that in S3 of this invention, the non-backtracking frequency sweep control mechanism based on confidence prediction is as follows: when the output confidence exceeds a preset confidence threshold (e.g., 0.85), the frequency range covered by the window function is determined to be a potential signal region, and the current frequency step and resolution bandwidth are reduced; otherwise, the frequency range covered by the window function is determined to be a noise floor region, and the current frequency step and resolution bandwidth remain unchanged. At the same time, the noise floor power data in this frequency range is extracted to dynamically update the power mean and standard deviation of the current environmental noise.

[0066] The following example, using a lightweight neural network with only one neuron, further illustrates the calculation process of the confidence level and the adjustment process of the frequency step and resolution bandwidth. Specifically, this embodiment selects two representative feature vectors. and , where the feature vector Located at the edge of a potential signal, the power increases slightly, the energy tends to be ordered, and the eigenvector... In high-power random noise, although the power is high, the distribution is extremely disordered. Meanwhile, the weights of the aforementioned neurons are set... Bias term This weight distribution reflects the model's preference for signal characteristics characterized by "high power, low variance, and low entropy." Under this setting, the feature vector... linear results With feature vectors linear results They can be represented as:

[0067]

[0068]

[0069] Next, the ReLU activation function is used for feature filtering, leveraging its "one-sided suppression" property to eliminate invalid interference, thereby improving the feature vector. Activation status With feature vectors Activation status They can be represented as:

[0070]

[0071]

[0072] As can be seen, the feature vector The effective features in the feature vector are preserved and passed on, while the feature vectors are... Noise interference is completely suppressed. Then, the output layer uses weights. With bias term By summarizing, the corresponding feature vectors are obtained. Output results With feature vectors Output results :

[0073]

[0074]

[0075] Finally, the output of the two feature vectors is mapped to confidence scores using the Sigmoid function. and :

[0076]

[0077]

[0078] Analysis shows that if the output is a large positive number (i.e., the signal feature is extremely significant), the calculated confidence level approaches 1; if the output is a small negative number (i.e., it conforms to a random noise distribution), the calculated confidence level approaches 0. Therefore, by comparing the confidence level with a preset confidence threshold, intelligent and accurate identification of "potential signal areas" can be achieved. In the example above, the feature vector... The confidence level is relatively high, therefore the frequency range covered by the window function can be identified as the potential signal region. The current frequency sweep step is then switched from faster to slower, and the current resolution bandwidth is simultaneously switched from wider to narrower, thus completing the parameter switching. Relatively speaking, the feature vector... The confidence level is relatively low, so the frequency range covered by the window function can be determined as the noise floor region. At this time, a relatively fast sweep step and a wide resolution bandwidth are still maintained to ensure the scanning efficiency, thereby successfully avoiding false triggering caused by high power noise.

[0079] Figure 3 This further demonstrates the non-backtracking real-time switching process of the two parameters: frequency sweep step and resolution bandwidth. When the frequency range covered by the window function is in the pure noise band, the extracted feature vector exhibits high entropy and high variance characteristics, and the confidence level of the output by the lightweight neural network inference is low. At this time, scanning is performed with a faster frequency sweep step and a wider resolution bandwidth. As the local oscillator (LO) moves forward, the edge of the window function begins to touch the skirt of the signal. Although the peak power has not yet reached the traditional threshold, the presence of the signal increases the orderliness of the energy distribution within the window (the entropy value decreases) and improves the power stability. The lightweight neural network nonlinearly fuses these implicit features, causing the output confidence level to rise smoothly and rapidly. Before the LO reaches the center frequency of the signal, the confidence level exceeds the preset confidence level threshold in advance. At this time, using this advance, the LO locking (phase-locked loop relocking) and adaptive adjustment of the window length are completed in advance without interrupting the current scanning process or performing backtracking rescanning. It is worth noting here that... Figure 3 The rectangular window used is only for illustrating the above phenomenon. In this embodiment, a non-rectangular window is used when performing actual frequency sweep.

[0080] S4. Perform bandwidth power normalization and window function gain correction on the original power values ​​at each frequency point to obtain the final power value at each frequency point, which is used to draw a spectrum diagram and visualize the output.

[0081] It should be noted that in S4 of this invention, because the hardware parameters are switched in real time as needed during the scanning process, the resolution bandwidth, frequency step, and window function length change in real time during the frequency sweep. If the raw power values ​​actually collected at each frequency point are not further processed and directly plotted on the spectrum, the final spectrum will have obvious fragmentation. Therefore, this embodiment reconstructs the data after collecting the raw power values ​​at each frequency point, based on the current resolution bandwidth, frequency step, and window function length, by performing bandwidth power normalization and window function gain correction on the raw power values.

[0082] The following sections will provide a detailed explanation of bandwidth power normalization and window function gain correction.

[0083] 1) Bandwidth power normalization

[0084] Since the thermal noise power spectral density is constant in the frequency domain, the measured noise floor power data is proportional to the current resolution bandwidth. This applies when the oscillator moves from a fast scan mode to a wider resolution bandwidth. Switching to finer mode with narrower resolution bandwidth At this time, the total amount of thermal noise entering the filter decreases, leading to a drop in the measured noise floor level. To ensure the consistency of the noise floor reference across the entire frequency band, this invention introduces a compensation formula:

[0085]

[0086] This formula actually uses the obtained raw power value Adding this to the correction term yields the power value after bandwidth power normalization. This correction term uses a logarithmic relationship, determined by the reference resolution bandwidth. With current resolution bandwidth The ratio is determined by this. Through this operation, the present invention uniformly converts the original power values ​​under different resolution bandwidths to the power level corresponding to the reference resolution bandwidth, eliminating the "gap" phenomenon of spectral noise floor caused by dynamic switching of hardware parameters.

[0087] It's worth noting that the "10" in the correction term comes from the definition of decibels (dB). When comparing two power values, the ratio expressed in decibels is... In this formula, the ratio of the two resolution bandwidths actually represents the power ratio under the two resolution bandwidths, so a coefficient of 10 is naturally used to convert the power ratio into decibel form. In short, this 10 is the standard coefficient for logarithmic power representation in the definition of decibels.

[0088] 2) Window function gain correction

[0089] Because the applied window function is a non-rectangular window, its shape is typically "high in the middle and low at both ends." This means that the signal in the middle of the window is preserved, while the signals at both ends are forcibly compressed to near zero. This leads to a reduction in the total energy collected, resulting in a decrease in the amplitude of the final output. Therefore, an amplitude correction factor is needed. (ACF) is used to compensate for the gain loss caused by windowing, and its calculation formula is as follows:

[0090]

[0091] in, This is the amplitude correction factor; The length of the window function; For the first The window function coefficients at the nth frequency point are used to represent the nth frequency point. The amount of power compression at each frequency point. The calculation method for the window function length is as described above and will not be repeated here.

[0092] During power spectrum reconstruction, the compensation amount The actual result is: It is worth noting that since power is proportional to the square of the amplitude, when converting the amplitude ratio to decibels, the compensation amount will show "20". In other words, the calculation formulas for the amplitude correction factor and compensation amount above require first dividing the length of the window function by the sum of all its coefficients to calculate the amplitude correction factor and obtain the compensation amount in decibels. The magnitude of the compensation amount depends on the amplitude correction factor. Therefore, the final power value is obtained. That is Add compensation amount for window function gain correction Finally, the final power value at each frequency point is used as the vertical axis and the frequency as the horizontal axis to form a spectrum diagram.

[0093] Additionally, it should be noted that in S4 of this invention, when plotting the spectrum, different frequency steps are used for the latent signal region and the noise floor region (the latent signal region corresponds to...). The corresponding noise floor area This results in uneven distribution of sampling points along the frequency axis. In the potential signal region, the interval between adjacent frequency points is... In the background noise region, this interval is expanded to To obtain more refined spectral details, approximate power values ​​for unmeasured frequency points can be calculated using existing power values. This allows for the addition of frequency points to the spectrum without increasing the number of measurements, reducing the interval between adjacent frequency points to a fixed value. In this embodiment, cubic spline interpolation is used for this process. The power envelopes of known frequency points are used to accurately fill in the frequency gaps, resulting in a continuous spectral curve with continuous slopes and curvatures at the junctions of different regions. The implementation of cubic spline interpolation is existing technology and will not be described further.

[0094] To verify the actual effectiveness of the method of this invention, this embodiment compares it with the traditional threshold-triggered backtracking method (i.e., the basic scheme) through computer simulation. The simulation is set with a full-band bandwidth of 1000MHz, containing 6 randomly distributed narrowband signals, and the PLL lock time is set to 50 time units to simulate real hardware overhead. The simulation results are as follows. Figure 4 , Figure 5 and Figure 6 As shown in Table 1, the specific performance indicators are compared.

[0095] Table 1. Comparison Test Results of Frequency Sweep Performance

[0096] Therefore, the method of this invention abandons the traditional backtracking mechanism. Its core lies in introducing a lightweight neural network discrimination engine. By performing nonlinear fusion and real-time probability extrapolation on multi-dimensional features such as spectral energy entropy, power variance, and peak power, the frequency sweeping decision is evolved from the traditional "posterior backtracking based on explicit thresholds" to "spatial forward-looking prediction scheduling based on intelligent confidence assessment". Under the premise of no backtracking and no missed detection of burst signals, an intelligent balance between high resolution and high scanning speed is achieved.

[0097] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A non-backtracking intelligent frequency sweeping method based on a lightweight neural network, characterized in that, Includes the following steps: S1. During the startup process of the superheterodyne spectrum analyzer, the oscillator performs a single full-band scan by stepping at the maximum frequency to match the maximum resolution bandwidth, thereby extracting the background noise power data as quickly as possible and calculating the power mean representing the current ambient noise and the standard deviation used to measure power fluctuations. S2. After the superheterodyne spectrum analyzer is started, the oscillator performs signal acquisition by stepping the corresponding resolution bandwidth at a preset frequency and in real time obtains the original power values ​​of each frequency point within the frequency range covered by the window function, which are used to extract the feature vectors input to the lightweight neural network. S3. Input the extracted feature vector into the pre-trained lightweight neural network. The network outputs a confidence level that reflects the probability of the signal’s existence through a combination of weight distribution and nonlinear logic. When the confidence level exceeds the preset confidence threshold, the frequency step and resolution bandwidth are dynamically adjusted through a non-backtracking frequency sweep control mechanism based on confidence level prediction, and the scanning continues according to the adjusted frequency step and resolution bandwidth. S4. Perform bandwidth power normalization and window function gain correction on the original power values ​​of each frequency point to obtain the final power value of each frequency point, which is used to draw a spectrum diagram and visualize the output. In S3, the non-backtracking frequency sweep control mechanism based on confidence prediction works as follows: when the output confidence exceeds the preset confidence threshold, the frequency range covered by the window function is identified as a potential signal region, and the current frequency step and resolution bandwidth are reduced; otherwise, the frequency range covered by the window function is identified as a noise floor region, and the current frequency step and resolution bandwidth remain unchanged. At the same time, the noise floor power data in this frequency range is extracted to dynamically update the power mean and standard deviation of the current ambient noise.

2. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 1, characterized in that, In S2, the length of the window function is equal to the main lobe bandwidth correction factor of the window function multiplied by the sampling frequency, and then divided by the current resolution bandwidth.

3. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 1, characterized in that, In S2, the feature vector input to the lightweight neural network consists of three feature components: the standardized peak power, the standardized power variance, and the normalized spectral energy entropy.

4. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 3, characterized in that, Both peak power and power variance are obtained based on the power sequence, which is a sequence of the original power values ​​of all frequency points within the frequency range covered by the current window function.

5. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 3, characterized in that, The calculation process of spectral energy entropy is as follows: within the frequency range covered by the window function, the original power values ​​of each frequency point are converted into a probability distribution through normalization. Then, the uncertainty measure of this probability distribution is calculated through the definition of information entropy in information theory. Finally, the result of the uncertainty measure is used as the spectral energy entropy.

6. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 3, characterized in that, In S3, the inference process of the lightweight neural network is as follows: multiple neurons perform linear fusion of feature components in parallel, and then the linear result output by each neuron is processed by a nonlinear activation function to obtain the activation state of the neuron. After linearly fusing the activation states of all neurons, the output result is generated. Finally, the output result is compressed by the Sigmoid logistic regression function to obtain the confidence level.

7. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 1, characterized in that, In the bandwidth power normalization of S4, the original power value is added to the correction term to obtain the power value after bandwidth power normalization. The correction term adopts a logarithmic relationship and is determined by the ratio of the reference resolution bandwidth to the current resolution bandwidth.

8. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 7, characterized in that, In the window function gain correction of S4, the length of the window function is divided by the sum of all the coefficients of the window function to calculate the amplitude correction factor and obtain the compensation amount in decibels. The magnitude of the compensation amount depends on the amplitude correction factor.

9. The non-backtracking intelligent frequency sweeping method based on a lightweight neural network as described in claim 8, characterized in that, In S4, the final power value is the sum of the power value after bandwidth power normalization and the compensation amount.

Images