Detection signal analysis method for 5G communication temporary grounding

Abstract

The invention relates to the field of computer systems based on specific calculation models, in particular to a 5G communication temporary grounding-oriented detection signal analysis method, which comprises the following steps of: S1, injecting a detection signal; s2, signal acquisition and preprocessing; s3, constructing a training data set; s4, training data set balance; s5, training a signal classification model; s6, classifying; an existing analysis method has the problem of low analysis accuracy, and the method is high in analysis accuracy.

Description

Technical Field

[0001] The present invention relates to the field of computer systems based on a specific computing model, and particularly to a method for analyzing detection signals for temporary grounding in 5G communication. Background Art

[0002] The fifth-generation mobile communication technology (referred to as 5G communication) is a new generation of broadband mobile communication technology with the characteristics of high speed, low latency, and large connection. To ensure the coverage rate of 5G communication, emergency communication vehicles or temporary base stations will be deployed in some hot spots or blank areas. Both emergency communication vehicles and temporary base stations need to be temporarily grounded and ensure the safe operation of the communication system. It is also necessary to perform grounding detection on the temporary grounding without affecting normal communication. The currently commonly used method is to inject a detection signal into the grounding system, collect the detection signal and the feedback signal fed back by the grounding system, and the signal classification model determines the state of the grounding system according to the detection signal and the feedback signal. The states include two classifications: "normal signal" or "abnormal signal".

[0003] The detection signals include signals for testing the impedance characteristics of the grounding system at a specific frequency, pulse signals for locating breaks, corrosion, or poor contact points of the grounding conductor, sweep signals for evaluating broadband matching, and modulation signals for verifying the dynamic response ability of the grounding system under complex modulation signals.

[0004] However, there are still some problems with the existing methods. First, when training the signal classification model, the number of signal data of different categories in the training dataset is different, resulting in majority-class signal data and minority-class signal data. The imbalance between the two will cause the trained signal classification model to be biased towards the majority-class signal data, ignoring abnormal situations, and the classification result is inaccurate, resulting in limited analysis accuracy, which is very dangerous in practical applications. Second, the existing methods mainly use traditional deep convolutional neural networks to train the signal classification model, and problems such as gradient disappearance, gradient explosion, and training oscillation are likely to occur during the training process, resulting in unstable performance of the signal classification model and inaccurate classification results, resulting in limited analysis accuracy.

[0005] Therefore, the existing method for analyzing detection signals for temporary grounding in 5G communication has the problem of relatively low analysis accuracy. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a method for analyzing detection signals for temporary grounding in 5G communication with relatively high analysis accuracy.

[0007] To solve the above technical problem, the method for analyzing detection signals for temporary grounding in 5G communication provided by the present invention includes:

[0008] S1. Detection signal injection;

[0009] S2. Signal acquisition and preprocessing;

[0010] Use a dual-channel synchronous sampling system for signal acquisition. The dual-channel synchronous sampling system includes channel A for collecting detection signals and channel B for collecting feedback signals;

[0011] S3. Training dataset construction;

[0012] Use the collected detection signals and feedback signals as samples to construct a training dataset;

[0013] S4. Training dataset balancing;

[0014] Compare the sample numbers of different category samples in the training dataset. If the sample numbers of different category samples vary greatly, use the SMOTE algorithm based on information priority to perform sample interpolation on the minority class samples to generate balanced samples until the sample numbers of different category samples are balanced, obtaining a balanced training dataset;

[0015] The SMOTE algorithm based on information priority means that an information priority strategy is adopted in the process of generating balanced samples. The information priority strategy means that by measuring the local density and feature information priority of the minority class samples, more suitable samples for expansion are dynamically determined;

[0016] S5. Signal classification model training;

[0017] Input the balanced training dataset and use a deep convolutional neural network combined with the gradient direction resonance suppression method to train the signal classification model, obtaining a trained signal classification model;

[0018] The deep convolutional neural network combined with the gradient direction resonance suppression method means that during the training process, the weight of the gradient direction is dynamically adjusted to control the fluctuation range of the gradient;

[0019] S6. Classification;

[0020] Input the newly collected detection signals and feedback signals into the trained signal classification model to obtain the states of the newly collected detection signals and feedback signals.

[0021] As a further improvement of the present invention: The dynamic power control strategy in S1 is: when the base station load ≤ 50%, the detection signal power is fixed at -20 dBm; when the load > 50%, the detection signal power is increased linearly.

[0022] As a further improvement of the present invention: The situation that the sample numbers of different category samples in S4 vary greatly means:

[0023]

[0024] The balance of the sample numbers of the different categories of samples means that:

[0025]

[0026] As a further improvement of the present invention: The implementation of sample interpolation for the minority class samples by using the SMOTE algorithm based on information priority in S4 to generate balanced samples includes:

[0027] S401. Measure the importance and sparsity of each minority class sample in the feature space to obtain the feature information priority;

[0028] S402. Generate balanced samples and perform feature matching. Achieve the consistency between the balanced samples and the original sample distribution through the mutual entropy constraint mechanism and retain the feature information with high priority during interpolation. The calculation formula is:

[0029]

[0030] In the formula, is the balanced sample, X c is the minority class sample, f is a positive integer, m c is the total number of sample features, is the weight of the f-th feature, λ f is the dynamic interpolation factor of the f-th feature, X d,f is the minority class sample X c 's neighbor sample X d 's value on the f-th feature, X c,f is the minority class sample X c 's value on the f-th feature;

[0031] S403. Calculate the entropy of the newly generated samples by minimizing the differences between the balanced samples and the original samples in terms of the entropy metric and the noise penalty term;

[0032] S404. Perform feature weighting processing on the generated balanced samples. The calculation formula for the feature weight is:

[0033]

[0034] In the formula, is the weight of the f-th feature, j is a positive integer, X j is the j-th minority class sample, D c is the set of minority class samples, ρ c (X j ) is the local density of the sample X j , f(X j ) is the sample Xj The specific value on feature f;

[0035] S405. Post-process and denoise the generated balanced samples, and screen out samples with too large a distribution difference from the existing signal data through the global similarity metric;

[0036] S406. Repeat S401 - S405 until the number of generated balanced samples reaches the preset threshold.

[0037] Preferably, the dynamic interpolation factor λ of the f-th feature in S402 f has the following calculation formula:

[0038]

[0039] In the formula, P c (X c ) is the feature information priority of the minority class sample X c , X d,f is the value of the neighbor sample X c of the minority class sample X d on the f-th feature, X c,f is the value of the minority class sample X c on the f-th feature, f is a positive integer, and m c is the total number of sample features.

[0040] As a further improvement of the present invention: The training of the signal classification model using the deep convolutional neural network combined with the gradient direction resonance suppression method in S5 includes:

[0041] S501. Define the structure of the deep convolutional neural network;

[0042] S502. Initialize the parameters of the deep convolutional neural network;

[0043] S503. Adopt the gradient direction resonance suppression method, control the fluctuation range of the gradient by dynamically adjusting the weight of the gradient direction, and the current loss gradient has the following calculation formula:

[0044]

[0045] In the formula, η is the learning rate, is the loss gradient of the previous iteration, ρ is the gradient adjustment factor, and θ u is the angle between the current gradient and the previous gradient;

[0046] S504. Use the parameter update inertia suppression method to suppress the inertia of the weight change in each training iteration considering the previous parameter update amount;

[0047] S505. Adopt a multi-loss function optimization strategy. By optimizing multiple loss functions, ensure that the model can constrain the classification results from multiple perspectives, comprehensively considering classification accuracy, regularization effect, and output stability;

[0048] S506. Repeat and iterate S503 - S505 until the preset iteration stop condition is met.

[0049] Preferably, the included angle θ u between the current gradient and the previous gradient in S503

[0050]

[0051] In the formula, θ u is the included angle between the current gradient and the previous gradient, is the current loss gradient, is the loss gradient of the previous iteration, is the loss gradient of the k-th iteration, is the loss gradient of the (k - 1)-th iteration.

[0052] The beneficial effects of the present invention are as follows: The analysis accuracy rate of the detection signal analysis method for 5G communication temporary grounding provided by the present invention is relatively high.

[0053] Firstly, before training the signal classification model, this method first adopts the SMOTE algorithm based on information priority to generate balanced samples through feature weighting and distribution entropy constraint, ensuring high-quality synthesis of minority class samples and maintaining the balance of the training data set, so that the signal classification model will not ignore abnormal situations, making the classification results more accurate, thereby improving the analysis accuracy rate of this method;

[0054] Secondly, this method uses a deep convolutional neural network combined with the gradient direction resonance suppression method to train the signal classification model, solving problems such as gradient disappearance and gradient explosion that are prone to occur in traditional neural networks when processing complex signal data, ensuring a stable and efficient training process, making the finally obtained signal classification model more stable, making the finally obtained classification results relatively accurate, and improving the analysis accuracy rate of this method. Brief Description of the Drawings

[0055] Figure 1 is the principle block diagram of the present invention;

[0056] Figure 2 is the classification performance comparison diagram of different data augmentation methods in a noise environment;

[0057] Figure 3 is the difference comparison diagram of different data augmentation methods in feature weight allocation;

[0058] Figure 4 A comparison chart of the influence of different neighborhood sizes on the quality of sample generation;

[0059] Figure 5 A comparison chart of the consistency in distribution between the generated samples and the real samples;

[0060] Figure 6 A comparison chart for comparing the training errors. Specific implementation manners

[0061] The following further elaborates on the specific implementation manners of the present invention in conjunction with the accompanying drawings.

[0062] As Figure 1 shown, the detection signal analysis method for temporary grounding for 5G communication provided by the present invention includes:

[0063] S1. Detection signal injection;

[0064] Using orthogonal frequency division multiplexing (OFDM) technology to generate a baseband signal, superimposing a pulse sequence with pseudo-random phase modulation in each orthogonal frequency division multiplexing symbol so that the detection signal is completely isolated from the normal communication signal in the time domain and frequency domain, inserting a periodic Zadoff-Chu sequence as a synchronization header in the time domain, and predefining 3 groups of characteristic subcarriers (numbered 100, 200, 300) in the frequency domain, with each group embedding a 127-bit Gold code sequence to uniquely identify the detection signal and enhance the signal discrimination in a multi-user scenario;

[0065] Selecting quadrature phase shift keying (QPSK) modulation to modulate the baseband signal, and shifting the spectrum of the baseband signal to a specific radio frequency band to obtain a modulated signal.

[0066] Using a dynamic power control strategy, monitoring the transmission power of the base station main link in real time, dynamically adjusting the power of the modulated signal through closed-loop control, injecting the modulated signal as a detection signal into the grounding loop using a high-isolation directional coupler, connecting a high-precision current sensor in series in the grounding loop, collecting the loop current waveform in real time, and injecting the detection signal during the guard interval period of each frame based on the base station communication frame structure;

[0067] Among them, the number of subcarriers of the baseband signal is fixed at 512, covering 3.5 GHz (Sub-6 GHz band) and 28 GHz (millimeter wave band) defined by the 5G standard, and the subcarrier spacing strictly matches 15 kHz or 30 kHz specified by the 5G protocol;

[0068] The pulse width of the pulse sequence with pseudo-random phase modulation can be dynamically configured (10ns to 1μs);

[0069] The sequence length of the Zadoff-Chu sequence is fixed at 64, and the root index is dynamically optimized according to the base station environment (typical value is 29) to resist multipath fading and noise interference;

[0070] The dynamic power control strategy is as follows: when the base station load ≤ 50%, the detection signal power is fixed at -20dBm; when the load > 50%, the detection signal power is increased linearly, and the highest value that can be increased to is 10dBm, ensuring that the signal-to-noise ratio (SNR) of the feedback signal is always ≥ 20dB;

[0071] The isolation of the high-isolation directional coupler is ≥ 40dB, and the coupling degree is set to -30dB, ensuring that the loss of the main communication link is controlled within 0.1dB;

[0072] When injecting the detection signal, the injection period can be configured as 1ms to 100ms (default 10ms) to avoid conflicts with uplink and downlink data transmission.

[0073] S2. Signal acquisition and preprocessing;

[0074] Use a dual-channel synchronous sampling system for signal acquisition. The dual-channel synchronous sampling system includes channel A for collecting the detection signal and channel B for collecting the feedback signal. The dual-channel hardware-level synchronization is achieved through FPGA, the signals of channel A and channel B are aligned and compensated for data, noise suppression and signal enhancement are performed on the original collected signals, and feature reference extraction is performed on the denoised detection signal;

[0075] Both channel A and channel B use 12-bit high-speed ADCs (such as AD9625), the sampling rate is ≥ 5GS / s, and the bandwidth covers DC to 6GHz.

[0076] The data alignment and compensation for the signals of channel A and channel B is to align the signals of channel A and channel B based on the cross-correlation peak of the Zadoff-Chu synchronization header, use a digital delay line to compensate for the time delay difference (Δt), and the alignment residual is controlled within ±0.1ns.

[0077] The noise suppression and signal enhancement for the original collected signals is to perform adaptive filtering processing using a two-stage filtering method. First, the power frequency interference and broadband noise are suppressed by an adaptive Kalman filter, and then the high-frequency transient noise is eliminated using a wavelet threshold denoising algorithm (Symlet4 wavelet basis, 5-layer decomposition).

[0078] The feature reference extraction for the denoised detection signal is to extract the initial impedance value of the aligned signal as the reference, the calculation range is 0.1Ω to 100Ω, and the resolution is 0.01Ω.

[0079] S3. Training dataset construction;

[0080] Use the collected detection signals and feedback signals as samples to construct a training dataset; the detection signals and feedback signals are collectively referred to as signal data.

[0081] Use the signal parameter data as the reference information for comparison. The signal parameter data includes:

[0082] Carrier technology and subcarrier information: number of subcarriers (512), frequency range (3.5 GHz and 28 GHz), subcarrier spacing (15 kHz or 30 kHz), etc.;

[0083] Signal modulation: pulse sequence with superimposed pseudo-random phase modulation, pulse width (10 ns to 1 μs), duty cycle (0.1% to 5%), etc.;

[0084] Synchronization header: Zadoff-Chu sequence, length 64, root index (e.g., 29), etc.;

[0085] Characteristic subcarriers: subcarriers numbered 100, 200, 300, embedded with 127-bit Gold code sequence, etc.;

[0086] Power control strategy: when the base station load is less than or equal to 50%, the detection signal power is fixed at -20 dBm; when the load is greater than 50%, the power is increased proportionally, with a maximum of 10 dBm, etc.;

[0087] Isolation and coupling: high-isolation directional coupler (isolation ≥ 40 dB, coupling -30 dB), etc.

[0088] The sources of the detection signals and feedback signals include:

[0089] Dual-channel synchronous sampling system: data from channel A and channel B;

[0090] Sampling parameters: e.g., sampling rate (≥ 5 GS / s), ADC bit depth (12 bits), bandwidth coverage (DC ~ 6 GHz);

[0091] Synchronization accuracy: e.g., FPGA synchronization error less than 1 ps, time-domain alignment accuracy;

[0092] Noise suppression and signal enhancement: e.g., processed using Kalman filter and wavelet threshold denoising algorithm;

[0093] Signal alignment and compensation: use Zadoff-Chu synchronization header to align signals and perform digital delay line compensation.

[0094] The training dataset has discrete attributes. Each sample includes multiple numerical features or information. Samples with time-domain, frequency-domain, and power feature attributes are retained by means of feature engineering and statistical analysis techniques, completing the extraction of the training dataset attributes;

[0095] The attributes of the training dataset include:

[0096] ID attribute: The unique identifier of the sample;

[0097] Carrier number attribute: The number of subcarriers used in the system (e.g., 512);

[0098] Frequency band attribute: The frequency band used (e.g., 3.5 GHz or 28 GHz);

[0099] Subcarrier spacing attribute: The spacing between subcarriers (e.g., 15 kHz or 30 kHz);

[0100] Pulse width attribute: The width of the pulse sequence (e.g., 100 ns, 200 ns, etc.);

[0101] Duty cycle attribute: The duty cycle of the pulse sequence (e.g., 0.1% to 5%);

[0102] Synchronization header root index attribute: The root index of the Zadoff-Chu sequence used for synchronization (e.g., 29 or 30);

[0103] Power control strategy attribute: The power control strategy of the base station, which may include fixed power (e.g., -20 dBm) or dynamic power (e.g., 5 dBm, 10 dBm);

[0104] Synchronization error attribute: The value of the dual-channel synchronization error (e.g., 0.3 ps, 1 ps, etc.);

[0105] Signal alignment residual attribute: The residual of the aligned signal (e.g., 0.05 ns, 0.1 ns, etc.);

[0106] Noise suppression effect attribute: The effectiveness of noise suppression (e.g., "effective", "ineffective");

[0107] Initial impedance value (Ω) attribute: The initial impedance value of the signal, with the unit of Ω (e.g., 12.34 Ω, 15.67 Ω, etc.).

[0108] S4. Balancing the training dataset;

[0109] Compare the number of samples of different classes in the training dataset. If the number of samples of different classes varies greatly, the SMOTE algorithm based on information priority is used to interpolate the samples of the minority class to generate balanced samples until the number of samples of different classes is balanced, obtaining a balanced training dataset;

[0110] The significant difference in the number of samples of different categories means that:

[0111]

[0112] The balance in the number of samples of different categories means that:

[0113]

[0114] For example, if the number of majority-class samples is 10 times that of minority-class samples, balanced sample generation is performed using the SMOTE algorithm based on information priority. When the number of majority-class samples is 2 times that of minority-class samples, the two are considered to achieve a balance in quantity;

[0115] The SMOTE algorithm based on information priority means that an information priority strategy is adopted in the process of generating balanced samples. The information priority strategy refers to dynamically determining the samples that are more suitable for expansion by measuring the local density and feature information priority of minority-class samples, thereby reducing the over-interpolation of noise samples. To calculate the distance between samples more accurately, a weighted metric based on feature weights is used, and weights are dynamically assigned based on the importance of features in the local neighborhood, making the expansion of minority-class samples more consistent with the true feature distribution. At the same time, distribution entropy constraints are used to calculate the feature information priority of samples to quantify the mixing degree of minority-class samples in the local area and ensure the rationality of the expansion strategy.

[0116] S401. Considering the diversity of signal data, features such as the number of carriers, frequency band attributes, subcarrier spacing, etc. are crucial for signal characteristics. Measure the importance and sparsity of each minority-class sample in the feature space to obtain the feature information priority, solve the problem of which minority-class samples in the signal data are more worthy of priority expansion, and reduce the risk of over-interpolation on noise samples;

[0117] The formula for calculating the local density of minority-class samples is:

[0118]

[0119] In the formula, ρ c (X c ) is the local density of the minority-class sample X c , X d is the neighboring sample of the minority-class sample, N k (X c ) is the k nearest neighbors of the minority-class sample X c , k is a positive integer, d w (X c , X d ) is the weighted distance metric based on feature weights, measuring the distance between the minority-class sample X c and the neighbor sample Xd The distance between them, X c is the minority class sample; σ c is the Gaussian kernel width of smoothness, which is used to adjust the smoothness of local density calculation and is set to 0.1.

[0120] To more accurately measure the differences between features, through weight adjustment and dynamic calculation of feature importance, better grasp the differences of these features among samples. For example, when calculating the local density of a sample, the weighted distance metric between features can accurately reflect the contribution of different features to the similarity between samples, so that important features such as specific frequency band attributes or pulse widths can be more retained during balanced sample generation. The calculation formula of the weighted distance metric based on feature weights is:

[0121]

[0122] In the formula, d w (X c , X d ) is the weighted distance metric based on feature weights, f is a positive integer, m c is the total number of sample features; w f is the weight of the f-th feature, which represents the importance of this feature in the distance metric; X c,f is the value of the minority class sample X c on the f-th feature, X d,f is the value of the neighboring sample X d of the minority class sample on the f-th feature;

[0123] To dynamically obtain the size of feature weights, it is allocated by the importance of features in the local neighborhood. The calculation formula is:

[0124]

[0125] In the formula, X d is the neighboring sample of the minority class sample, N k (X c ) are the k nearest neighbors of the minority class sample X c , k is a positive integer, f is a positive integer, m c is the total number of sample features, X d,f is the value of the neighboring sample X d of the minority class sample on the f-th feature, |||| is the L2 norm;

[0126] Using distribution entropy constraint to calculate the feature information priority of each sample. The calculation formula of the feature information priority P c of the minority class sample X c (X c ) is:

[0127]

[0128] Where ρ c (X c ) is the local density of the minority class sample X c . X j is the j-th minority class sample, j is a positive integer, D c is the set of all minority class samples, ρ c (X j ) is the local density of the sample X j ; λ is the entropy weighting coefficient, set to 0.1; H(X c ) is the distribution entropy of the sample X c , which is used to measure the mixing degree of the signal data in the local area;

[0129] The calculation formula for the local distribution entropy H(X c ) of the sample X c is as follows:

[0130]

[0131] Where X j is the j-th minority class sample, D c is the set of all minority class samples, ρ c (X j ) is the local density of the sample X j , X i is the i-th minority class sample; i is a positive integer, ρ c (X i ) is the local density of the sample X i ;

[0132] S402. Generate and match balanced samples, and achieve the consistency between the balanced samples and the original sample distribution through the cross-entropy constraint mechanism, and focus on retaining high-priority feature information during interpolation, so as to solve the problem that new samples may deviate from the true distribution under high-dimensional signal data. The calculation formula for the balanced samples is:

[0133]

[0134] Where is the balanced sample, X c is the minority class sample, f is a positive integer, m c is the total number of sample features, is the weight of the f-th feature, λ f is the dynamic interpolation factor of the f-th feature, X d,f is the value of the neighbor sample X c of the minority class sample X d on the f-th feature, X c,fFor the minority class sample X c The value on the f-th feature;

[0135] To determine the dynamic interpolation factor for each feature, it is calculated based on the local differences between features and the sample information priority. For example, for features related to the sync head root index and power control strategy, the dynamic allocation of feature priorities based on signals can provide a more reasonable structure for sample generation, avoiding some pseudo-samples that do not conform to the true distribution. At the same time, for the dual-channel sync error and signal alignment residual, the calculation of feature priorities helps to better handle these specific errors during the expansion process and avoid error amplification, expressed as:

[0136]

[0137] In the formula, P c (X c ) is the feature information priority of the minority class sample X c , X d,f is the minority class sample X c , X d is the value of the neighbor sample X c,f of the minority class sample X c on the f-th feature, X c is the value of the minority class sample X

[0138] S403. To make the balanced samples closer to the characteristics of the original samples in terms of information distribution, the calculation of the entropy of the newly generated samples is achieved by minimizing the differences between the balanced samples and the original samples in terms of entropy metrics and noise penalty terms; ensuring that the newly generated signal samples do not introduce too much noise. Especially for high-frequency signals, features such as the duty cycle of pulse sequences and sync errors have a greater impact on signal noise. During the generation process, by minimizing the entropy difference between the balanced samples and the original samples and imposing a noise penalty, the quality of the signal data can be effectively controlled, thus ensuring the authenticity of the expanded data, expressed as:

[0139]

[0140] In the formula, is the entropy of the balanced sample , H(X c ) is the entropy of the minority class sample X c ; γ c is the noise tolerance coefficient, set to 0.1; is the noise suppression factor, used to measure the noise that the new balanced sample may introduce, following a normal distribution with a mean of 0 and a variance of 0.01; the symbol "→min" indicates minimizing this expression to constrain the balanced sample to be more in line with the original distribution;

[0141] This process corrects the synthetic points that do not conform to the distribution towards the original distribution through iterative optimization.

[0142] When generating the above balanced samples, in order to take into account the local and global feature distributions, it can also be further improved through a multi-scale collaborative optimization mechanism. By defining loss functions at both the local scale and the global scale simultaneously, combining local interpolation with global distribution constraints, the balanced samples can avoid global distribution imbalance while retaining the local neighborhood structure. The calculation formula is:

[0143]

[0144] In the formula, X c is the minority class sample; λ is the interpolation factor, comprehensively considering the local and global feature distributions; f is a positive integer, m is the number of features, w f is the weight of the f-th feature, X d,f is the neighbor sample X c of the minority class sample X d at the f-th feature, X c,f is the value of the minority class sample X c at the f-th feature;

[0145] By defining the sum of the global distribution constraint loss L global and the local interpolation loss L local as the final objective function L final , and adopting a dynamic gradient update strategy for iterative optimization, the balanced samples can maintain a reasonable distribution at both the global and local scales, thereby reducing the phenomenon of local over-aggregation or distribution breakage in the high-dimensional space.

[0146] S404. Perform feature weighting processing on the generated balanced samples to highlight more important features and suppress redundant features in the high-dimensional feature space, thereby improving the contribution of the augmented samples to subsequent training. The calculation formula for the feature weight is:

[0147]

[0148] In the formula, is the weight of the f-th feature, j is a positive integer, X j is the j-th minority class sample, D c is the set of minority class samples, ρ c (X j ) is the local density of the sample X j , f(X j ) is the specific value of the sample X j on the feature f;

[0149] S405. Post-process and denoise the generated balanced samples to ensure the authenticity and effectiveness of the final augmented set. By means of global similarity measurement, samples with too large a distribution difference from the existing signal data are screened out;

[0150] Avoid the phenomenon of outliers deviating in the high-dimensional space. The calculation formula is:

[0151]

[0152] In the formula, is the balanced sample and the similarity measurement between the balanced sample and the j-th minority class sample, is the balanced sample, X j is the j-th minority class sample; σ sim is the standard deviation of similarity calculation, set to 0.1.

[0153] If the similarity measurement exceeds a certain threshold, it means that the balanced sample may have too large a difference from the existing distribution and needs to be excluded, so as to retain the balanced sample that can better reflect the real distribution.

[0154] S406. Repeat S401 - S405 until the number of generated balanced samples reaches the preset threshold.

[0155] As Figure 2 shown, to verify the advantages of the SMOTE data augmentation method based on information priority (abbreviated as IP-SMOTE) in a noisy environment, the F1-score performances of traditional SMOTE, Borderline-SMOTE, ADASYN, and IP-SMOTE at different noise levels (0.05, 0.1, 0.15, 0.2) were experimentally compared. It can be seen from the experimental results that as the noise level increases, the F1-score of the IP-SMOTE method always remains high, significantly better than other methods. IP-SMOTE shows high classification accuracy at all noise levels, proving its robustness and stability in a noisy environment. Compared with other methods, IP-SMOTE can better handle noise, reduce the over-interpolation of noise samples, and thus improve the overall performance of the signal classification model.

[0156] As Figure 3As shown in the figure, to compare the differences in feature weight allocation of different data augmentation methods, the experiment used heatmaps to show the feature weight allocation of traditional SMOTE and IP-SMOTE when dealing with 5 features (number of carriers, frequency band attributes, subcarrier spacing, pulse width, and synchronization error). The experimental results show that IP-SMOTE can dynamically adjust the weights according to the importance and sparsity of each feature, making the retention of key signal features more accurate, thus improving the quality of sample generation. In contrast, traditional SMOTE distributes weights more evenly and lacks focused attention on high-priority features. The experimental data show that IP-SMOTE can more effectively retain the features most useful for the classification task, thereby improving the quality of balanced samples and the effectiveness of subsequent model training.

[0157] As Figure 4 shown, to investigate the impact of different neighborhood sizes (k values) on the quality of sample generation, especially the performance of IP-SMOTE under adjusted neighborhood sizes, the experiment compared the generation quality scores of SMOTE, ADASYN, and IP-SMOTE at different k values (from 3 to 15). The experimental data show that the generation quality performance of IP-SMOTE is the most stable at different k values. The generation quality score gradually increases with the increase of the k value, and its fluctuation range is small. In contrast, the quality scores of SMOTE and ADASYN show a certain decline or instability when the k value is large. The experimental results show that IP-SMOTE can better adapt to different neighborhood sizes, thus ensuring the quality stability of balanced samples and further demonstrating its superiority in the sample generation process.

[0158] As Figure 5 shown, to compare the consistency in distribution between balanced samples and real samples and prove that IP-SMOTE can better retain the real distribution of data when balancing samples, the scatter plots show the distribution of real samples, traditional SMOTE balanced samples, and IP-SMOTE balanced samples. The experimental results show that the samples generated by IP-SMOTE are the closest to the distribution of real samples, while the samples generated by traditional SMOTE deviate far from the real distribution, resulting in a low similarity between samples. The experiment proves that IP-SMOTE can more accurately reflect the distribution of real data when balancing samples, avoiding the phenomenon that the generated pseudo-samples deviate from the real distribution, and ensuring the authenticity and effectiveness of the generated data.

[0159] S5. Signal classification model training;

[0160] Input the balanced training dataset and train the signal classification model using a deep convolutional neural network combined with the gradient direction resonance suppression method to obtain the trained signal classification model;

[0161] The structure of the deep convolutional neural network is shown in Table 1:

[0162] Table 1

[0163]

[0164] The deep convolutional neural network combined with the gradient direction resonance suppression method refers to dynamically adjusting the weights of the gradient direction during the training process to control the fluctuation range of the gradient;

[0165] Training a signal classification model using a deep convolutional neural network combined with the gradient direction resonance suppression method includes:

[0166] S501. Define the structure of the deep convolutional neural network; the hierarchical structure of the deep convolutional neural network includes 10 hidden layers, the number of neurons between layers is preset artificially, for example, all are 100, and the number of neurons is adaptively adjusted by using the dropout method;

[0167] S502. Initialize the parameters of the deep convolutional neural network; the parameters of the deep convolutional neural network are initialized randomly and the initialized parameters follow a normal distribution with a mean of 0 and a variance of the identity matrix. The parameters of the deep convolutional neural network include weight parameters and bias parameters, and the update methods of both are the same during the training process. Taking the weight parameters as a reference, let the initial weight of the deep convolutional neural network be

[0168] S503. During the training process, the resonance effect of the gradient direction may lead to unstable updates of network parameters. To effectively control this phenomenon, the gradient direction resonance suppression method is adopted. By dynamically adjusting the weights of the gradient direction, the fluctuation range of the gradient is controlled, making the parameter updates during the training process more stable and avoiding the gradient instability problem in traditional deep convolutional neural networks. For example, for the synchronization error and signal alignment residual of the signal, when there are errors in the time alignment or frequency alignment of the signal, it is easy to cause unstable updates of the parameters of the deep convolutional neural network. The gradient direction resonance suppression makes the training process remain stable even when there are problems with signal alignment by controlling the gradient fluctuation range, avoiding gradient instability caused by synchronization errors, ensuring that the model can effectively learn signal features. When the parameters are updated based on the gradient descent method, the current loss gradient The calculation formula is:

[0169]

[0170] In the formula, η is the learning rate, is the loss gradient of the previous iteration, ρ is the gradient adjustment factor, and ρ is set to 0.2; cos() is the cosine function, which measures the resonance effect of the gradient direction here; θ u is the angle between the current gradient and the previous gradient;

[0171] The angle θ between the current gradient and the previous gradient u The calculation formula is as follows:

[0172]

[0173] In the formula, θ u is the angle between the current gradient and the previous gradient, is the current loss gradient, is the loss gradient of the previous iteration, is the loss gradient of the k-th iteration, is the loss gradient of the (k - 1)-th iteration, where k is a positive integer;

[0174] S504. To avoid over-updating and unstable oscillations during training, an inertial suppression method for parameter updates is used. Considering the previous parameter update amount, inertial suppression is performed on the weight changes in each training iteration; to ensure the smoothness of network parameter updates, for signals with relatively small or large pulse widths and duty cycles, traditional neural networks may have difficulty processing them because the changes in these signals in the time domain are relatively significant and prone to cause oscillations and instability in network training. The inertial suppression method can effectively avoid over-updating. Especially when the time-domain characteristics of the signal (such as the width and duty cycle of the pulse sequence) cause large fluctuations, by considering the previous update amount to smooth the current weight update, this is particularly effective for signals with significant periodic changes (such as pulse signals), which can ensure the smoothness of network updates and avoid oscillations during training. The parameter update method of the deep convolutional neural network is expressed as:

[0175]

[0176] In the formula, W u-new is the weight after this update, and W u-last is the weight after the previous update; η is the learning rate, set to 0.01; is the current loss gradient; λ is the inertia factor, set to 0.3; ΔW u is the amount of weight change in the previous update;

[0177] The amount of weight change ΔW in the previous update u is calculated based on the weighted average of the previous two gradient changes to control the inertial effect, and is expressed as:

[0178]

[0179] In the formula, α u is the first-step size adjustment factor, and α u is set to 0.3; is the loss gradient of the previous iteration; βu is the second step size adjustment factor, set to 0.5; ΔW u-last is the weight change amount updated the time before last;

[0180] S505. To improve the classification performance, a multi-loss function optimization strategy is adopted. By optimizing multiple loss functions, it is ensured that the model can constrain the classification results from multiple perspectives, comprehensively considering the classification accuracy, regularization effect, and output stability. Taking the frequency band attribute as an example, there may be significant differences in signals at different frequency bands. The cross-entropy loss can optimize the accuracy of class prediction, the KL divergence loss can handle the probability distribution differences between frequency bands, and the mean square error can improve the stability of signal output. Through the joint optimization of these three loss functions, the model can better adapt to signal data with different frequency band and frequency characteristics, improving the classification accuracy and stability. The loss function L u of the deep convolutional neural network is calculated as follows:

[0181] L u = L CE (y a , y b ) + L RKL (y a , y b ) + L MSE (y a , y b ),

[0182] wherein, L CE (y a , y b ) is the cross-entropy loss, L RKL (y a , y b ) is the relative KL divergence loss, L MSE (y a , y b ) is the mean square error loss, y a is the true label of the signal data sample, "normal signal" corresponds to the label "0", and "abnormal signal" corresponds to the label "1"; y b is the predicted label of the deep convolutional neural network for the signal data sample;

[0183] S506. Repeat and iterate S503 to S505 until the preset stop iteration condition is satisfied; the preset stop iteration condition is to reach the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0184] Such as Figure 6As shown, to verify the effectiveness of the deep convolutional neural network combined with the gradient direction resonance suppression method, the error changes and classification accuracy of the deep convolutional neural network combined with the gradient direction resonance suppression method and conventional techniques (such as traditional neural networks and other optimization techniques) during the training process are analyzed. The advantages of the deep convolutional neural network combined with the gradient direction resonance suppression method are verified by comparing the training effects of different techniques. By comparing the training errors of the three techniques, it can be seen that as the number of training iterations increases, the error gradually decreases. The error of the deep convolutional neural network combined with the gradient direction resonance suppression method decreases more smoothly and is less affected by noise, showing the stability of its training process. In addition, as the training progresses, the classification accuracy also continuously improves. The classification accuracy of the deep convolutional neural network combined with the gradient direction resonance suppression method rises rapidly and reaches a relatively high final accuracy, indicating that the deep convolutional neural network combined with the gradient direction resonance suppression method effectively improves the performance of the model.

[0185] S6. Classification;

[0186] The newly acquired detection signal and the feedback signal are input into the trained signal classification model. The deep convolutional neural network will process the input features layer by layer. Through the calculations of the hidden layers, the high-level features of the signal are gradually extracted and combined. The output layer of the deep convolutional neural network will classify the input signal to obtain the states of the newly acquired detection signal and the feedback signal, and the states include normal signals and abnormal signals.

[0187] According to the classification task set during training, the output layer will generate the probabilities of two classes (normal signal or abnormal signal) for each input signal, including:

[0188] Probability of class 0 (normal signal): P 正常信号

[0189] Probability of class 1 (abnormal signal): P 异常信号

[0190] By setting a threshold (such as 0.5) to determine the classification result. For example, if P(abnormal signal) > 0.5, it is determined as an abnormal signal; otherwise, it is determined as a normal signal, which is expressed as:

[0191]

[0192] The final classification result is the label (normal signal or abnormal signal) of the input signal. If it is a normal signal, "normal signal" is output; if it is an abnormal signal, "abnormal signal" is output.

Claims

1. A detection signal analysis method for temporary grounding of 5G communication, characterized in that: include: S1. Detection signal injection; S2. Signal acquisition and preprocessing; A dual-channel synchronous sampling system is used for signal acquisition, wherein the dual-channel synchronous sampling system includes a channel A for acquiring a detection signal and a channel B for acquiring a feedback signal; S3. Construction of training dataset; The collected detection signals and feedback signals are used as samples to construct a training data set; S4. Balanced training dataset; Compare the sample numbers of samples of different categories in the training data set. If the sample numbers of samples of different categories differ greatly, use the SMOTE algorithm based on information priority to perform sample interpolation on minority class samples to generate balanced samples until the sample numbers of samples of different categories are balanced, thus obtaining a balanced training data set. The SMOTE algorithm based on information priority refers to the use of information priority strategy in the process of generating balanced samples. The information priority strategy refers to dynamically determining samples that are more suitable for expansion by measuring the local density and feature information priority of minority class samples; S5. Signal classification model training; The balanced training data set is input to train the signal classification model using a deep convolutional neural network combined with a gradient directional resonance suppression method to obtain a trained signal classification model; The deep convolutional neural network combined with the gradient direction resonance suppression method refers to controlling the fluctuation range of the gradient by dynamically adjusting the weight of the gradient direction during the training process; S6. Classification; The newly collected detection signal and feedback signal are input into the trained signal classification model to obtain the status of the newly collected detection signal and feedback signal.

2. The detection signal analysis method for temporary grounding of 5G communication according to claim 1 is characterized in that: The dynamic power control strategy in S1 is: when the base station load is ≤50%, the detection signal power is fixed at -20dBm; when the load is >50%, the detection signal power is increased in a linear proportion.

3. The detection signal analysis method for temporary grounding of 5G communication according to claim 1 is characterized in that: The large difference in the number of samples of different categories described in S4 refers to: The sample size balance of different categories of samples refers to:

4. The detection signal analysis method for temporary grounding of 5G communication according to claim 1 is characterized in that: In S4, the SMOTE algorithm based on information priority is used to perform sample interpolation on minority class samples to achieve generation of balanced samples, including: S401. Measure the importance and sparsity of each minority class sample in the feature space to obtain the feature information priority; S402. Generate and match the balanced samples, achieve the consistency of the balanced samples with the original samples through the mutual entropy constraint mechanism, and retain the feature information with high priority during interpolation. The calculation formula is: In the formula, For a balanced sample, X c is a minority class sample, f is a positive integer, m c is the total number of sample features, is the weight of the f-th feature, λ f is the dynamic interpolation factor of the fth feature, X d,f For the minority class sample X c Neighbor sample X d The value of the f-th feature, X c,f For the minority class sample X c The value of the f-th feature; S403. Calculate the entropy of the newly generated sample by minimizing the difference between the balanced sample and the original sample in the entropy index and the noise penalty term; S404. Perform feature weighting processing on the generated balanced samples. The calculation formula of feature weight is: In the formula, is the weight of the fth feature, j is a positive integer, X j is the jth minority class sample, D c is the minority class sample set, ρ c (X j ) is the sample X j The local density, f(X j ) is the sample X j The specific value of feature f; S405. Post-process and denoise the generated balanced samples, and filter out samples that differ too much from the existing signal data distribution by means of global similarity measurement; S406. Repeat S401 to S405 until the number of generated balanced samples reaches a preset threshold.

5. The detection signal analysis method for temporary grounding of 5G communication according to claim 4 is characterized in that: The dynamic interpolation factor λ of the f-th feature in S402 f The calculation formula is: Where P c (X c ) is the minority class sample X c The feature information priority of X d,f For the minority class sample X c Neighbor sample X d The value of the f-th feature, X c,f For the minority class sample X c The value of the fth feature, f is a positive integer, m c is the total number of sample features.

6. The detection signal analysis method for temporary grounding of 5G communication according to claim 1 is characterized in that: The deep convolutional neural network training signal classification model described in S5 using the gradient direction resonance suppression method includes: S501. Define the structure of a deep convolutional neural network; S502. Initialize the parameters of the deep convolutional neural network; S503. Adopt the gradient direction resonance suppression method, dynamically adjust the weight of the gradient direction, control the fluctuation range of the gradient, and the current loss gradient The calculation formula is: In the formula, η is the learning rate, is the loss gradient of the previous iteration, ρ is the gradient adjustment factor, θ u is the angle between the current gradient and the previous gradient; S504. Using the parameter update inertia suppression method, taking into account the last parameter update amount, the weight change in each training iteration is inertially suppressed; S505. Adopt a multiple loss function optimization strategy to ensure that the model can constrain the classification results from multiple angles by optimizing multiple loss functions, and comprehensively consider classification accuracy, regularization effect and output stability; S506. Repeat iterations S503 to S505 until a preset stop iteration condition is met.

7. The detection signal analysis method for temporary grounding of 5G communication according to claim 6 is characterized in that: The angle θ between the current gradient and the previous gradient in S503 u The calculation formula is: In the formula, θ u is the angle between the current gradient and the previous gradient, is the current loss gradient, is the loss gradient of the previous iteration, is the loss gradient of the kth iteration, is the loss gradient of the k-1th iteration.

Images