Signal-to-noise ratio estimation algorithm based on auto-encoder

Abstract

The invention relates to the technical field of signal processing, in particular to a signal-to-noise ratio estimation algorithm based on an auto-encoder, which comprises the following steps of: firstly, selecting four groups of parameter signals with different code element rates and roll-off factors, each group containing 16 mainstream modulation type sub-signals, segmenting the sub-signals, adding Gaussian white noise to construct a data set, and performing data processing on the data set; dividing into a training set and a verification set according to the ratio of 7: 3, and calculating a noise adding signal and a power spectrum of corresponding noise by adopting a Welch method; secondly, constructing a stacked full-connection auto-encoder, taking a training set noise adding signal power spectrum as input, taking a noise power spectrum as a target, and introducing L2 regularization suppression over-fitting according to set batchsize and epoch training; and finally, inputting a test set noise adding signal into the trained model to obtain the noise average power, and calculating the signal average power and the signal-to-noise ratio by combining the noise adding signal power spectrum. The method can cover multiple signal parameters and modulation scenes, the low signal-to-noise ratio interval precision is better, the process is clear, engineering landing is easy, and the method is suitable for a multi-standard communication system.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, specifically to a signal-to-noise ratio estimation algorithm based on an autoencoder. Background Technology

[0002] In modern communication systems, the signal-to-noise ratio (SNR), as a core parameter for measuring signal quality, directly determines the accuracy of signal demodulation at the receiver, the reliability of data transmission, and the overall performance of the communication system. Whether in wireless, satellite, or wired communication scenarios, accurate SNR estimation is a prerequisite for achieving key functions such as adaptive modulation and demodulation, channel equalization, and power control. Therefore, developing efficient and high-precision SNR estimation algorithms remains an important research direction in the field of communication signal processing.

[0003] Current mainstream signal-to-noise ratio (SNR) estimation algorithms are mainly divided into two categories: non-data-aided and data-aided. Data-aided methods rely on known training sequences or pilot signals, calculating the SNR by comparing the received signal with the known sequence. While these methods can achieve high accuracy under ideal channel conditions, they have significant limitations: inserting training sequences consumes additional channel bandwidth, reducing data transmission efficiency; and in scenarios with rapidly changing channels, the channel characteristics of the training sequence may deviate from those of the actual signal transmission period, leading to a significant decrease in estimation accuracy and making it difficult to adapt to dynamically changing communication environments.

[0004] Non-data-aided methods do not rely on known sequences and directly estimate the signal-to-noise ratio (SNR) based on the characteristics of the received signal itself, which better meets the real-time and bandwidth efficiency requirements of practical communication systems. Common examples include periodogram methods based on power spectrum analysis and moment estimation methods based on higher-order statistics. However, existing non-data-aided methods still have several technical bottlenecks: First, they have poor signal adaptability. Most algorithms are only designed for specific modulation types, such as single QPSK, 16QAM, or fixed symbol rates. When signal parameters such as symbol rate, roll-off factor, or modulation scheme change, the algorithm parameters need to be readjusted or even the model reconstructed, making it impossible to achieve universality across multiple scenarios. Second, the estimation accuracy in low SNR ranges is insufficient. In low SNR environments of 0dB to 10dB, signal characteristics are severely masked by noise. Traditional power spectrum analysis methods are easily affected by noise fluctuations, and higher-order statistics estimation methods are also susceptible to interference. First, the algorithm suffers from several drawbacks. First, insufficient signal sample information leads to increased errors, making it difficult to meet the demands of high-precision communication scenarios. Second, the model's generalization ability is weak; some machine learning-based estimation algorithms are prone to overfitting during training, significantly increasing estimation errors for new signal samples outside the training set, and lacking effective overfitting suppression mechanisms. Third, the engineering implementation is complex; parameter settings for some steps in existing algorithms rely on empirical debugging, such as the selection of window functions for power spectrum calculation and the configuration of hyperparameters for model training. The lack of clear and unified operational standards increases the difficulty of integrating and applying these algorithms in actual communication devices.

[0005] As communication technology develops towards multiple standards, multiple frequency bands, and high dynamics, the shortcomings of existing signal-to-noise ratio (SNR) estimation algorithms in terms of signal adaptation range, low SNR accuracy, generalization ability, and engineering practicality are becoming increasingly prominent. There is an urgent need for an SNR estimation algorithm that can cover multiple signal parameters and modulation types, maintain high accuracy over a wide SNR range, and is easy to implement in engineering, so as to meet the actual needs of complex communication scenarios. This provides the necessary technical background and practical significance for the proposal of this invention. Summary of the Invention

[0006] The purpose of this invention is to provide a signal-to-noise ratio (SNR) estimation algorithm based on an autoencoder, in order to solve the problems mentioned in the background art, such as poor signal adaptability, insufficient estimation accuracy in low SNR intervals, and weak generalization ability of existing SNR estimation algorithms.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A signal-to-noise ratio estimation algorithm based on an autoencoder includes the following steps: Step 1: Signal selection and dataset construction. Four sets of parameter signals are selected. The four sets of parameter signals are signals with a symbol rate of 150kSps and a shaping filter roll-off factor of 0.35, signals with a symbol rate of 150kSps and a shaping filter roll-off factor of 0.65, signals with a symbol rate of 62.5kSps and a shaping filter roll-off factor of 0.35, and signals with a symbol rate of 62.5kSps and a shaping filter roll-off factor of 0.65. Each set of parameter signals contains sub-signals of 16 modulation types. Each modulation type of sub-signal sequence is segmented with a segment length of N=2000 points and a segment number of M=25. The signal-to-noise ratio range is set to 0~30dB, and noise of 16 different signal-to-noise ratios is added to each sub-signal segment in 2dB increments, resulting in 4×25×16×16 segments of noisy signal. The noisy signal is then divided into a training set and a validation set according to a certain ratio, resulting in the original data S1 of the training set noisy signal and the original data S2 of the validation set noisy signal. The training set noise sequence N1 is obtained by subtracting the original noisy signal S1 from the corresponding original sub-signal. The validation set noise sequence N2 is obtained by subtracting the original noisy signal S2 from the corresponding original sub-signal. The power spectrum Ps1 of the training set noise signal S1, the power spectrum Pn1 of the training set noise sequence N1, the power spectrum Ps2 of the validation set noise signal S2, and the power spectrum Pn2 of the validation set noise sequence N2 are calculated using the Welch method. A window function is used in the calculation process, with the window length set to N divided by 10 and the overlap length set to N divided by 20. The dimensions of Ps1 and Pn1 are both 17920×2000, and the dimensions of Ps2 and Pn2 are both 7680×2000. Step 2: Model Training. A stacked autoencoder is constructed, with all layers forming fully connected layers. The encoding dimension is 2000→1000→500→250, and the decoding dimension is 250→500→1000→2000. The power spectrum Ps1 of the noisy training set signal is used as the input data of the stacked autoencoder, and the power spectrum Pn1 of the noise in the training set is used as the target output data. The mean square error between the output data after encoding and decoding the input data and the target output data is calculated. This mean square error is used as the loss function. The batch size is set to 64, and the epoch size to 30. The parameters of the stacked autoencoder are updated using the backpropagation principle of deep learning until the loss function value is minimized, resulting in the trained autoencoder model. The training effect of the trained autoencoder model is verified using the power spectrum Ps2 of the noisy validation set signal and the power spectrum Pn2 of the noise in the validation set. Step 3: Calculate the signal-to-noise ratio (SNR). Input the noisy test set signal into the trained autoencoder model to obtain the model output data. Calculate the mean of the model output data and use this mean as the average power of the noise. Calculate the power spectrum of the noisy signal in the test set according to the formula: ; Calculate the average power of the noisy signal in the test set. ,in The power spectrum of the noisy signal in the test set is given, where N is the number of points in the Fourier transform of the signal; according to the formula... Calculate the average power of the signal in The average power of the signal; according to the formula Calculate the signal-to-noise ratio .

[0008] Preferably, the 16 modulation types of sub-signals in step one are 2ASK signal, 4ASK signal, 8ASK signal, 2FSK signal, 4FSK signal, 8FSK signal, MSK signal, BPSK signal, QPSK signal, OQPSK signal, 4DQPSK signal, 8PSK signal, 16PSK signal, 16QAM signal, 32QAM signal, and 64QAM signal.

[0009] Preferably, in step one, the noisy signal is divided into a training set and a validation set in a 7:3 ratio. The specific calculation process is as follows: the total number of noisy signal segments S total =4×25×16×16=25600 segments; Number of training set segments S train =S total×0.7=17920 segments, corresponding to 17920 segments of raw training set noisy signal data S1; validation set segment number S val =S total ×0.3=7680 segments, corresponding to 7680 segments of the original data S2 of the validation set noise signal; the number of segments of the training set noise sequence N1 is the same as that of S1, the number of segments of the validation set noise sequence N2 is the same as that of S2, and the length of each segment of N1 and N2 is the same as the length of the corresponding noise signal segment, which is 2000 points.

[0010] Preferably, the encoding process of the stacked autoencoder in step two includes three fully connected encoding layers: the first fully connected encoding layer has an input dimension of 2000 and an output dimension of 1000; the second fully connected encoding layer has an input dimension of 1000 and an output dimension of 500; and the third fully connected encoding layer has an input dimension of 500 and an output dimension of 250. The decoding process includes three fully connected decoding layers: the first fully connected decoding layer has an input dimension of 250 and an output dimension of 500; the second fully connected decoding layer has an input dimension of 500 and an output dimension of 1000; and the third fully connected decoding layer has an input dimension of 1000 and an output dimension of 2000.

[0011] Preferably, the window function used in step one is the Hanning window, and the mathematical expression of the Hanning window is: ; Where n = 0, 1, ..., , For the length of the window, and When using the Welch method to calculate the power spectrum, the formula for calculating the power spectrum is: ; in, The signal is divided into segments, with each segment of the noisy signal or noise sequence treated as a separate segment; therefore, K=S. train Or K=S val R is the energy normalization coefficient of the window function. , For the first Given a segment of noisy signal or noise sequence n=0,1,…,1999, FFT(·) is the Fast Fourier Transform operation, and |·| is the modulo operation.

[0012] As a preferred option, step two sets batch_size=64 and epoch=30; the loss function is the mean squared error loss function, and the specific calculation is divided into sample-by-sample loss and batch total loss: (1) Loss per sample for: ; in For power spectrum dimension =2000, For the first The decoded output of the nth sample Dimensional data, For the first Individual sample targets The Dimensional data; (2) Total batch loss: ; in =batch_size=64, i=1,2,...,B; Model parameter updates use stochastic gradient descent, and the weight update formula is:

[0013] The bias update formula is: ; Where η is the learning rate, initially set to 0.001, decaying to 0.8 times the original value every 5 epochs, until training stops when the loss function value is less than the preset threshold of 0.005.

[0014] Preferably, the number of points N in the Fourier transform of the signal in step three is 2000, which is consistent with the segment length of the sub-signal sequence in step one; the power spectrum of the noisy signal in the test set. The calculation formula is: ; in The sampling rate, and With symbol rate satisfy The oversampling rate is 4. =150kSps =600kSps, when =62.5kSps =250kSps; Average power of the noisy signal Calculation formula Average signal power = Signal-to-noise ratio And the calculated The estimation error must be satisfied: ; Standard deviation: ; in, Let j = 1, 2, ..., be the number of samples in the test set. .

[0015] Preferably, in step two, when verifying the training effect of the trained autoencoder model using the validation set, the loss value of the validation set is calculated first. : ; ; Next, calculate the signal-to-noise ratio estimation error of the validation set: ; If within 3 consecutive epochs The increase exceeded 5% and If the value is greater than 1 dB, the model is considered overfitted. In this case, L2 regularization needs to be introduced. The formula for the total loss after regularization is: ; in Let be the regularization coefficient, set to 0.0001. Let L2 norm sum of squares of all weight matrices be the sum of squares of all weight matrices; if Stable within the range of 0.005~0.01 and If the value is ≤1dB, the model training effect is considered acceptable.

[0016] Preferably, the noise added to each sub-signal in step one is Gaussian white noise, and the formula for generating the Gaussian white noise is: ; in (·) is a function that generates standard normally distributed random numbers. The standard deviation of noise; With target signal-to-noise ratio (SNR) true The relationship is given by the formula The derivation yields that: ; in The average power of the original sub-signal, after amplitude normalization, is given by the 2ASK signal. =0.5, for QPSK signal P_s=0.5.

[0017] As a preferred option, when using the Welch method to calculate the power spectrum in step one, the window length is... Overlap length overlap rate Total signal length training set Or the total length verification set of the signal ; Number of segments for: ; in The function is the floor function; frequency resolution. ,when =600kSps =600k / 200=3kHz, when =250kSps =250k / 200=1.25kHz, and the power spectrum Corresponding frequency point ( =0,1,...,1999).

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention excels in signal adaptability, covering multiple signal parameters and modulation scenarios to effectively cope with complex communication environments. It selects four sets of parameter signals with different symbol rates of 150kSps and 62.5kSps and shaping filter roll-off factors of 0.35 and 0.65, while also including sub-signals of 16 common modulation types, covering mainstream modulation methods such as ASK, FSK, PSK, and QAM. Furthermore, the signal-to-noise ratio coverage ranges from 0dB to 30dB, adapting to noise environments of varying intensities. This comprehensive signal coverage capability eliminates the need for individual adjustments for single signal scenarios, allowing direct application to multi-standard, multi-scenario communication systems, reducing development costs for scenario adaptation.

[0019] 2. The model design of this invention is scientifically sound and reasonable, with stable training and strong anti-interference capabilities, ensuring the reliability of the model output. It employs a stacked fully connected autoencoder, using a progressive compression and restoration dimensionality design: encoding dimension 2000→1000→500→250, decoding dimension 250→500→1000→2000, effectively extracting noise features from noisy signals. During training, the mean squared error is used as the loss function, combined with reasonable batch size and epoch settings, and a progressive decay strategy for the learning rate, achieving stable parameter updates. Simultaneously, a validation set is introduced to monitor training performance in real time. When overfitting occurs, L2 regularization is used for correction, preventing the model from overfitting on the training set and losing generalization ability. Ultimately, this ensures that the trained model maintains stable feature extraction and output performance across different datasets.

[0020] 3. The signal-to-noise ratio (SNR) estimation accuracy of this invention is high, meeting the needs of high-precision communication scenarios, especially performing better in the low SNR range. In the power spectrum calculation stage, the algorithm employs the Welch method combined with the Hanning window, effectively reducing the impact of noise fluctuations on power spectrum accuracy. During the SNR solution process, the average noise power is accurately obtained through model output, and the average signal power is calculated by combining it with the power spectrum of the noisy signal. The final SNR estimation error is strictly controlled within 1 dB, with a standard deviation of less than 0.35 dB. For the challenging low SNR scenario, further optimization can be achieved by increasing the number of samples and adjusting training parameters, reducing the estimation error in the 0 dB to 10 dB range to below 0.9 dB, with a standard deviation of less than 0.3 dB. This solves the problem of weak signal characteristics and high estimation difficulty under low SNR conditions, ensuring estimation accuracy in complex noise environments.

[0021] 4. This invention possesses strong practicality and operability, with a clear process and easy engineering implementation. From dataset construction and model training to signal-to-noise ratio (SNR) calculation, each step has clearly defined parameters and operational standards. For example, the data segment length, window length and overlap length for power spectrum calculation, model training parameter settings, and SNR calculation logic all provide specific and executable solutions, eliminating the need for complex experience-based adjustments. Furthermore, the algorithm supports flexible optimization based on actual needs. For instance, in low SNR scenarios, the number of sample segments and training epochs can be increased; in overfitting scenarios, regularization can be introduced. This flexibility allows the algorithm to meet the basic requirements of general scenarios as well as adapt to the customized requirements of special scenarios, reducing the difficulty of engineering implementation and facilitating integration into actual communication equipment or signal processing systems. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0023] Figure 1 This is a flowchart illustrating the signal-to-noise ratio estimation algorithm based on an autoencoder according to the present invention. Figure 2 This is a diagram illustrating the training process of the signal-to-noise ratio estimation autoencoder model in this invention. Figure 3 This describes the process of estimating the noise power spectrum using an autoencoder in this invention. Figure 4 This is a performance index analysis chart of Embodiment 1 of the present invention; Figure 5 This is an analysis diagram of the autoencoder network structure in Embodiment 1 of the present invention; Figure 6 This is a dataset partitioning and analysis diagram from Embodiment 1 of the present invention; Figure 7 This is a comparative analysis chart of the dataset sizes in Embodiment 1 and Embodiment 3 of the present invention; Figure 8 This is a comparative analysis chart of the effects of embodiments one and three of the present invention in the low signal-to-noise ratio range. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: Signal-to-noise ratio estimation based on standard parameters 1. Signal selection and dataset construction First, four sets of parameter signals are selected: a signal with a symbol rate of 150 kSps and a shaping filter roll-off factor of 0.35; a signal with a symbol rate of 150 kSps and a shaping filter roll-off factor of 0.65; a signal with a symbol rate of 62.5 kSps and a shaping filter roll-off factor of 0.35; and a signal with a symbol rate of 62.5 kSps and a shaping filter roll-off factor of 0.65. Each set of parameter signals contains sub-signals with 16 modulation types: 2ASK, 4ASK, 8ASK, 2FSK, 4FSK, 8FSK, MSK, BPSK, QPSK, OQPSK, 4DQPSK, 8PSK, 16PSK, 16QAM, 32QAM, and 64QAM.

[0026] The sub-signal sequence for each modulation type was segmented, with each segment having a length of 2000 points, and each sub-signal sequence was divided into 25 segments. The signal-to-noise ratio was set from 0dB to 30dB, and Gaussian white noise was added to each sub-signal segment at 2dB intervals, resulting in 25600 segments of noisy signals.

[0027] The 25,600 noisy signal segments were divided into a training set and a validation set in a 7:3 ratio. The training set contained 17,920 noisy signal segments, which served as the original data for the training set noisy signal; the validation set contained 7,680 noisy signal segments, which served as the original data for the validation set noisy signal.

[0028] Calculate the noise sequences for the training and validation sets separately: subtract the corresponding original sub-signals from the original noisy data of the training set to obtain the training set noise sequence; subtract the corresponding original sub-signals from the original noisy data of the validation set to obtain the validation set noise sequence. The number of segments in the training set noise sequence is the same as the number of segments in the original noisy data of the training set, and the number of segments in the validation set noise sequence is the same as the number of segments in the original noisy data of the validation set, with each noise sequence segment having a length of 2000 points.

[0029] The Welch method was used to calculate the power spectra of the original noisy training data, the power spectrum of the noisy training sequence, the power spectrum of the original noisy validation data, and the power spectrum of the noisy validation sequence. A Hanning window was used in the calculations, with a window length of 200 points and an overlap length of 100 points. The final power spectra of the noisy training data and the noisy validation data both had dimensions of 17920 x 2000, while the dimensions of the power spectra of the noisy validation data and the noisy validation data were both 7680 x 2000.

[0030] 2. Model Training A stacked autoencoder is constructed, whose network structure consists entirely of fully connected layers. The encoding process includes three fully connected layers, with the dimension decreasing sequentially from 2000 to 1000, then to 500, and finally to 250. The decoding process also includes three fully connected layers, with the dimension increasing sequentially from 250 to 500, then to 1000, and finally to 2000.

[0031] The power spectrum of the noisy training signal is used as the input data of the stacked autoencoder, and the power spectrum of the noise in the training set is used as the target output data. The mean square error between the output data after encoding and decoding the input data and the target output data is used as the loss function. The batch size is set to 64, and the number of training epochs is 30. The autoencoder parameters are updated using stochastic gradient descent. The initial learning rate is set to 0.001, and it is decayed to 0.8 times the original value every 5 training epochs. Training continues until the loss function value is less than 0.005, resulting in the trained autoencoder model.

[0032] The model training effect was verified using a validation set: The validation set loss was calculated by averaging the per-sample loss for each data segment in the validation set. The per-sample loss is calculated by summing the squared differences between the decoded output data for each dimension and the corresponding dimension of the target noise power spectrum, then dividing by the power spectrum dimension of 2000. Simultaneously, the signal-to-noise ratio (SNR) estimation error was calculated, which is the average of the absolute values ​​of the differences between the estimated SNR and the true SNR for each data segment in the validation set. In this training, the validation set loss remained stable between 0.005 and 0.01, and the SNR estimation error did not exceed 1 dB, indicating that the model training effect was satisfactory.

[0033] 3. Solving for the signal-to-noise ratio Prepare a test set of noisy signals. The parameters of the test set of noisy signals are consistent with those of the training set and validation set. It contains the same 4 sets of parameter signals and 16 modulation types. The segment length is 2000 points and the signal-to-noise ratio ranges from 0dB to 30dB.

[0034] The test set with added noise is input into the trained autoencoder model to obtain the model output data. The mean of the output data is calculated and used as the average power of the noise.

[0035] Calculate the power spectrum of the noisy signal in the test set: Perform a Fast Fourier Transform on the noisy signal in the test set, square the result modulo, and then divide it by the product of the number of points of the Fourier Transform and the sampling rate. The sampling rate is determined according to the symbol rate. When the symbol rate is 150 kSps, the sampling rate is 600 kSps, and when the symbol rate is 62.5 kSps, the sampling rate is 250 kSps.

[0036] Calculate the average power of the noisy signal in the test set: sum all the frequency points of the power spectrum of the noisy signal in the test set and divide by the number of points of the Fourier transform of the signal, 2000.

[0037] Calculate the average power of the signal: Subtract the average power of the noise from the average power of the noisy signal.

[0038] Calculate the signal-to-noise ratio (SNR): Divide the average power of the signal by the average power of the noise, take the logarithm to base 10, and then multiply by 10. Testing showed that the SNR estimation error for all test samples was less than 1 dB, and the standard deviation was less than 0.35 dB, meeting the performance requirements. Figures 4-6 As shown.

[0039] Parameter categories Specific content Signal basic parameters Symbol rate: 150 kSps, 62.5 kSps; roll-off factor: 0.35, 0.65; modulation type: 16 types (including 2ASK, QPSK, etc.) Dataset partitioning Total number of noisy signal segments: 25,600; Training set: 17,920 segments; Validation set: 7,680 segments; Segment length: 2,000 points Power spectrum calculation parameters Window function: Hanning window; Window length: 200 points; Overlap length: 100 points; Power spectrum dimension (training set): 17920×2000 Model structure parameters Encoding dimension: 2000→1000→500→250; Decoding dimension: 250→500→1000→2000; Batch size: 64; Epoch: 30 Training and performance metrics Initial learning rate: 0.001 (decays to 0.8x every 5 epochs); validation set loss range: 0.005~0.01; estimation error <1dB; standard deviation <0.35dB. Example 2: Signal-to-noise ratio estimation after model overfitting correction 1. Signal selection and dataset construction The signal selection and dataset construction process in this embodiment is completely consistent with that in Embodiment 1: Four sets of parameter signals are selected, each set containing 16 types of modulation sub-signals; each sub-signal sequence is segmented into 2000 segments, and each segment is further divided into 25 segments; Gaussian white noise is added at 2dB intervals within the 0dB to 30dB signal-to-noise ratio range to obtain 25600 segments of noisy signals; the noisy signals are divided into a training set of 17920 segments and a validation set of 7680 segments at a ratio of 7:3; the noise sequences and corresponding power spectra of the training set and validation set are calculated. A Hanning window with a window length of 200 points and an overlap length of 100 points is used in the calculation process. The final power spectrum dimension of the training set is 17920 multiplied by 2000, and the power spectrum dimension of the validation set is 7680 multiplied by 2000.

[0040] 2. Model Training A stacked autoencoder with the same structure as in Example 1 was constructed, with the encoding process having dimensions of 2000→1000→500→250, and the decoding process having dimensions of 250→500→1000→2000. The batch size was set to 64, the training epochs to 30, and the initial learning rate to 0.001, decreasing to 0.8 times the original rate every 5 training epochs. The mean squared error between the input data and the target output data was used as the loss function.

[0041] After training began, the training set loss value continued to decrease in the first 10 training rounds. However, starting from the 11th training round, the validation set loss value increased by more than 5% in the next 3 consecutive training rounds, and the signal-to-noise ratio estimation error of the validation set was greater than 1dB. Based on this, it was determined that the model was overfitting.

[0042] To correct the overfitting problem, L2 regularization is introduced. The total loss after regularization is the original loss function value plus the product of the regularization coefficient and the sum of squares of the L2 norm of all weight matrices, where the regularization coefficient is set to 0.0001.

[0043] After introducing L2 regularization and continuing training, the increase in the loss value of the validation set was effectively controlled and eventually stabilized between 0.005 and 0.01. The signal-to-noise ratio estimation error of the validation set was reduced to below 1 dB, and the model training was completed and the training effect was satisfactory.

[0044] 3. Solving for the signal-to-noise ratio The parameters of the test set with added noise are consistent with those of the training and validation sets. It includes 4 sets of parameter signals and 16 modulation types, with a segment length of 2000 points and a signal-to-noise ratio range of 0dB to 30dB.

[0045] The test set with added noise is input into the overfitted autoencoder model to obtain the model output data. The mean of the output data is calculated and used as the average power of the noise.

[0046] The power spectrum of the noisy signal in the test set was calculated using the same method as in Example 1: a Fast Fourier Transform was performed on the noisy signal, the modulus was taken, the square was then taken, and the result was divided by the product of the number of Fourier Transform points and the corresponding sampling rate; then the average power of the noisy signal was calculated, which is the sum of all frequency points of the power spectrum and divided by the number of Fourier Transform points, 2000; the average power of the signal was obtained by subtracting the average power of the noise from the average power of the noisy signal; finally, the signal-to-noise ratio was calculated by the ratio of the average power of the signal to the average power of the noise.

[0047] Test results show that the signal-to-noise ratio estimation error of all test samples is less than 1dB, the standard deviation is less than 0.35dB, and the estimation accuracy in the high signal-to-noise ratio range (20dB to 30dB) is improved by about 0.1dB compared with the uncorrected overfitting, which meets the requirements of practical applications.

[0048] Parameter categories Specific content Signal and Dataset Parameters Consistent with Example 1 (4 sets of signal parameters, 16 modulation schemes, 25,600 noisy signal segments, 7:3 training / validation set division). Model basic parameters Consistent with Example 1 (fully connected stacked autoencoder, same encoding / decoding dimension, batch_size=64, epoch=30) Overfitting Overfitting occurred in epochs 11-13; the increase in validation set loss was >5%; the validation set signal-to-noise ratio estimation error was >1dB. Corrective measures Regularization type: L2 regularization; Regularization coefficient: 0.0001; Corrected total loss: Original loss + 0.0001 × sum of squared L2 norms of the weights Performance after correction Validation set loss range: 0.005~0.01; estimation error ≤1dB; accuracy improvement of 0.1dB in the high signal-to-noise ratio range (20~30dB). Example 3: Signal-to-noise ratio (SNR) estimation optimized for low SNR regions 1. Signal selection and dataset construction Four sets of parameter signals, identical to those in Example 1, were selected, each containing 16 sub-signals with different modulation types. Considering that the signal characteristics in the low signal-to-noise ratio (0dB to 10dB) range are weaker and more difficult to estimate, the number of segments for each sub-signal sequence was increased from 25 to 40 during segmentation to increase the number of samples in the low signal-to-noise ratio range and improve the model's learning performance on low signal-to-noise ratio signals.

[0049] The signal-to-noise ratio is set to a range of 0dB to 10dB, with 1dB intervals (more dense than the 2dB intervals in Example 1). Gaussian white noise is added to each sub-signal segment, resulting in a total of 4 x 40 x 16 x 11 segments, or 28,160 segments.

[0050] The 28,160 noisy signal segments were divided into a training set and a validation set in a 7:3 ratio. The training set contained 19,712 noisy signal segments, and the validation set contained 8,448 noisy signal segments.

[0051] Calculate the noise sequences of the training set and validation set: The noise sequence of the training set is obtained by subtracting the original data of the training set noise-added signal from the corresponding original sub-signal, and the noise sequence of the validation set is obtained by subtracting the original data of the validation set noise-added signal from the corresponding original sub-signal. The length of each noise sequence is 2000 points.

[0052] The Welch method was used to calculate the power spectrum, still using the Hanning window with a window length of 200 points and an overlap length of 100 points. The final dimensions of the power spectrum of the training set with added noise and the power spectrum of the training set with noise are 19712 x 2000, and the dimensions of the power spectrum of the validation set with added noise and the power spectrum of the validation set with noise are 8448 x 2000.

[0053] 2. Model Training A stacked autoencoder was constructed, with the network structure consistent with that in Example 1. The encoding process dimension was 2000→1000→500→250, and the decoding process dimension was 250→500→1000→2000.

[0054] To adapt to the learning requirements of low signal-to-noise ratio (SNR) samples, the training parameters were adjusted as follows: the batch size was reduced from 64 to 32, enabling the model to update parameters more precisely when processing low SNR data and reducing the impact of differences in single batches on parameter updates; the number of training epochs was increased from 30 to 40 to ensure that the model fully learns the signal and noise features in the low SNR range; the initial learning rate remained at 0.001, decaying to 0.8 times its original value every 5 training epochs, and the mean squared error loss function was adopted.

[0055] During training, the validation set loss value and signal-to-noise ratio (SNR) estimation error are monitored in real time. When training reaches the 35th epoch, the loss function value drops to 0.0045, the validation set loss value stabilizes between 0.005 and 0.008, and the validation set SNR estimation error stabilizes below 0.8 dB. Training is then stopped, resulting in an autoencoder model optimized for the low SNR range.

[0056] 3. Solving for the signal-to-noise ratio The test set noisy signal focuses on the low signal-to-noise ratio range, and the parameters are consistent with those of the training set: it contains 4 sets of parameter signals, 16 modulation types, a segment length of 2000 points, each sub-signal sequence is divided into 40 segments, and the signal-to-noise ratio ranges from 0dB to 10dB with 1dB intervals.

[0057] The test set with added noise is input into the optimized autoencoder model to obtain the model output data. The mean of the output data is calculated and used as the average power of the noise.

[0058] Calculate the power spectrum of the noisy signal in the test set: Determine the sampling rate based on the symbol rate (sampling rate is 600kSps when symbol rate is 150kSps, and sampling rate is 250kSps when symbol rate is 62.5kSps). Perform a fast Fourier transform on the noisy signal, take the square of the modulus, and then divide it by the product of the number of Fourier transform points 2000 and the sampling rate.

[0059] Calculate the average power of the noisy signal: sum all frequency points of the power spectrum of the noisy signal and divide by the number of Fourier transform points, 2000; calculate the average power of the signal: subtract the average power of the noise from the average power of the noisy signal; finally, calculate the signal-to-noise ratio: multiply the base-10 logarithm of the ratio of the average power of the signal to the average power of the noise by 10.

[0060] Test results show that in the low signal-to-noise ratio (SNR) range of 0 dB to 10 dB, the SNR estimation error for all test samples is less than 0.9 dB, and the standard deviation is less than 0.3 dB. Compared with Example 1, the estimation accuracy in the same range is improved by approximately 0.2 dB. Figure 7 and Figure 8 As shown, it can meet the high-precision estimation requirements in low signal-to-noise ratio scenarios.

[0061] Parameter categories Specific content Low signal-to-noise ratio optimization parameters Signal-to-noise ratio range: 0~10dB (in 1dB steps); Number of sub-signal segments: 40 segments / type; Total number of noisy signal segments: 28160 segments Dataset partitioning Training set: 19712 segments; Validation set: 8448 segments; Segment length: 2000 points; Power spectrum dimension (training set): 19712×2000 Model adjustment parameters Encoding / decoding dimensions: Same as in Example 1; batch_size: 32; epoch: 40; initial learning rate: 0.001 (same decay rule) Training termination conditions Training epochs: 35 epochs; Loss function value: 0.0045; Validation set loss range: 0.005~0.008 Low signal-to-noise ratio performance Estimation error <0.9dB; standard deviation <0.3dB; accuracy improved by 0.2dB compared to the previous example. The advantages of the signal-to-noise ratio estimation algorithm based on autoencoder proposed in this invention are as follows: This invention excels in signal adaptability, covering a wide range of signal parameters and modulation scenarios to effectively handle complex communication environments. It selects four sets of parameter signals with different symbol rates of 150kSps and 62.5kSps and shaping filter roll-off factors of 0.35 and 0.65, and includes sub-signals for 16 common modulation types, covering mainstream modulation methods such as ASK, FSK, PSK, and QAM. Furthermore, the signal-to-noise ratio coverage ranges from 0dB to 30dB, adapting to noise environments of varying intensities. This comprehensive signal coverage capability eliminates the need for individual adjustments for single signal scenarios, allowing direct application to multi-standard, multi-scenario communication systems and reducing development costs for scenario adaptation.

[0062] The model design of this invention is scientifically sound and reasonable, with stable training and strong anti-interference capabilities, ensuring the reliability of the model output. It employs a stacked fully connected autoencoder, using a progressive compression and restoration dimensionality design: encoding dimension 2000→1000→500→250, decoding dimension 250→500→1000→2000, effectively extracting noise features from noisy signals. During training, the mean squared error is used as the loss function, combined with reasonable batch size and epoch settings, and a progressive decay strategy for the learning rate, achieving stable parameter updates. Simultaneously, a validation set is introduced to monitor training performance in real time. When overfitting occurs, L2 regularization is used for correction, preventing the model from overfitting on the training set and losing generalization ability. Ultimately, this ensures that the trained model maintains stable feature extraction and output performance across different datasets.

[0063] This invention offers high signal-to-noise ratio (SNR) estimation accuracy, meeting the demands of high-precision communication scenarios, especially excelling in the low SNR range. The algorithm employs the Welch method combined with the Hanning window in the power spectrum calculation stage, effectively reducing the impact of noise fluctuations on power spectrum accuracy. During SNR calculation, the average noise power is accurately obtained through model output, and the average signal power is calculated by combining this with the power spectrum of the noisy signal. The final SNR estimation error is strictly controlled within 1 dB, with a standard deviation of less than 0.35 dB. For the challenging low SNR scenario, further optimization can be achieved by increasing the number of samples and adjusting training parameters, reducing the estimation error in the 0 dB to 10 dB range to below 0.9 dB, with a standard deviation of less than 0.3 dB. This solves the problem of weak signal characteristics and high estimation difficulty under low SNR conditions, ensuring estimation accuracy in complex noise environments.

[0064] This invention possesses strong practicality and operability, with a clear process and easy engineering implementation. From dataset construction and model training to signal-to-noise ratio (SNR) calculation, each step has clearly defined parameters and operational standards. For example, the data segment length, window length and overlap length for power spectrum calculation, model training parameter settings, and SNR calculation logic are all provided with specific and executable solutions, eliminating the need for complex experience-based adjustments. Furthermore, the algorithm supports flexible optimization based on actual needs. For instance, in low SNR scenarios, the number of sample segments and training epochs can be increased; in overfitting scenarios, regularization can be introduced. This flexibility allows the algorithm to meet the basic requirements of general scenarios as well as adapt to the customized requirements of special scenarios, reducing the difficulty of engineering implementation and facilitating integration into actual communication equipment or signal processing systems.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A signal-to-noise ratio estimation algorithm based on an autoencoder, characterized in that, Includes the following steps: Step 1: Signal selection and dataset construction. Four sets of parameter signals are selected. The four sets of parameter signals are signals with a symbol rate of 150kSps and a shaping filter roll-off factor of 0.35, signals with a symbol rate of 150kSps and a shaping filter roll-off factor of 0.65, signals with a symbol rate of 62.5kSps and a shaping filter roll-off factor of 0.35, and signals with a symbol rate of 62.5kSps and a shaping filter roll-off factor of 0.

65. Each set of parameter signals contains sub-signals of 16 modulation types. Each modulation type of sub-signal sequence is segmented with a segment length of N=2000 points and a segment number of M=25. The signal-to-noise ratio range is set to 0~30dB, and noise of 16 different signal-to-noise ratios is added to each sub-signal segment in 2dB increments, resulting in 4×25×16×16 segments of noisy signal. The noisy signal is then divided into a training set and a validation set according to a certain ratio, resulting in the original data S1 of the training set noisy signal and the original data S2 of the validation set noisy signal. The training set noise sequence N1 is obtained by subtracting the original noisy signal S1 from the corresponding original sub-signal. The validation set noise sequence N2 is obtained by subtracting the original noisy signal S2 from the corresponding original sub-signal. The power spectrum Ps1 of the training set noise signal S1, the power spectrum Pn1 of the training set noise sequence N1, the power spectrum Ps2 of the validation set noise signal S2, and the power spectrum Pn2 of the validation set noise sequence N2 are calculated using the Welch method. A window function is used in the calculation process, with the window length set to N divided by 10 and the overlap length set to N divided by 20. The dimensions of Ps1 and Pn1 are both 17920×2000, and the dimensions of Ps2 and Pn2 are both 7680×2000. Step 2: Model Training. A stacked autoencoder is constructed, with all layers forming fully connected layers. The encoding dimension is 2000→1000→500→250, and the decoding dimension is 250→500→1000→2000. The power spectrum Ps1 of the noisy training set signal is used as the input data of the stacked autoencoder, and the power spectrum Pn1 of the noise in the training set is used as the target output data. The mean square error between the output data after encoding and decoding the input data and the target output data is calculated. This mean square error is used as the loss function. The batch size is set to 64, and the epoch size to 30. The parameters of the stacked autoencoder are updated using the backpropagation principle of deep learning until the loss function value is minimized, resulting in the trained autoencoder model. The training effect of the trained autoencoder model is verified using the power spectrum Ps2 of the noisy validation set signal and the power spectrum Pn2 of the noise in the validation set. Step 3: Calculate the signal-to-noise ratio (SNR). Input the noisy test set signal into the trained autoencoder model to obtain the model output data. Calculate the mean of the model output data and use this mean as the average power of the noise. ; Calculate the power spectrum of the noisy signal in the test set using the formula: ； Calculate the average power of the noisy signal in the test set. ,in The power spectrum of the noisy signal in the test set is given, where N is the number of points in the Fourier transform of the signal; according to the formula... Calculate the average power of the signal in The average power of the signal; according to the formula Calculate the signal-to-noise ratio .

2. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, The 16 modulation types of sub-signals mentioned in step one are 2ASK signal, 4ASK signal, 8ASK signal, 2FSK signal, 4FSK signal, 8FSK signal, MSK signal, BPSK signal, QPSK signal, OQPSK signal, 4DQPSK signal, 8PSK signal, 16PSK signal, 16QAM signal, 32QAM signal, and 64QAM signal.

3. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, In step one, the noisy signal is divided into a training set and a validation set in a 7:3 ratio. The specific calculation process is as follows: the total number of noisy signal segments S total =4×25×16×16=25600 segments; Number of training set segments S train =S total ×0.7=17920 segments, corresponding to 17920 segments of raw training set noisy signal data S1; validation set segment number S val =S total ×0.3=7680 segments, corresponding to 7680 segments of the original data S2 of the validation set noise signal; the number of segments of the training set noise sequence N1 is the same as that of S1, the number of segments of the validation set noise sequence N2 is the same as that of S2, and the length of each segment of N1 and N2 is the same as the length of the corresponding noise signal segment, which is 2000 points.

4. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, The encoding process of the stacked autoencoder described in step two includes three fully connected encoding layers: the first fully connected encoding layer has an input dimension of 2000 and an output dimension of 1000; the second fully connected encoding layer has an input dimension of 1000 and an output dimension of 500; and the third fully connected encoding layer has an input dimension of 500 and an output dimension of 250. The decoding process includes three fully connected decoding layers: the first fully connected decoding layer has an input dimension of 250 and an output dimension of 500; the second fully connected decoding layer has an input dimension of 500 and an output dimension of 1000; and the third fully connected decoding layer has an input dimension of 1000 and an output dimension of 2000.

5. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, The window function used in step one is the Hanning window, and the mathematical expression for the Hanning window is: ； Where n = 0, 1, ..., , For the length of the window, and When using the Welch method to calculate the power spectrum, the formula for calculating the power spectrum is: ； in, The signal is divided into segments, with each segment of the noisy signal or noise sequence treated as a separate segment; therefore, K=S. train Or K=S val R is the energy normalization coefficient of the window function. , For the first Given a segment of noisy signal or noise sequence n=0,1,…,1999, FFT(·) is the Fast Fourier Transform operation, and |·| is the modulo operation.

6. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, In step two, batch_size is set to 64 and epoch to 30; the loss function is the mean squared error loss function, which is specifically calculated as per-sample loss and batch total loss. (1) Loss per sample for: ； in For power spectrum dimension =2000, For the first The decoded output of the nth sample Dimensional data, For the first Individual sample targets The Dimensional data; (2) Total batch loss: ； in =batch_size=64, i=1,2,...,B; Model parameter updates use stochastic gradient descent, and the weight update formula is: ； The bias update formula is: ； Where η is the learning rate, initially set to 0.001, decaying to 0.8 times the original value every 5 epochs, until training stops when the loss function value is less than the preset threshold of 0.

005.

7. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, The number of points N=2000 in the Fourier transform of the signal in step three is consistent with the segment length of the sub-signal sequence in step one; the power spectrum of the noisy signal in the test set. The calculation formula is: ； in The sampling rate, and With symbol rate satisfy The oversampling rate is 4. =150kSps =600kSps, when =62.5kSps =250kSps; Average power of the noisy signal Calculation formula Average signal power = Signal-to-noise ratio And the calculated The estimation error must be satisfied: ； Standard deviation: ； in, Let j = 1, 2, ..., be the number of samples in the test set. .

8. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, In step two, when validating the training performance of the trained autoencoder model using the validation set, the loss value of the validation set is first calculated. : ； ； Next, calculate the signal-to-noise ratio estimation error of the validation set: ； If within 3 consecutive epochs The increase exceeded 5% and If the value is greater than 1 dB, the model is considered overfitted. In this case, L2 regularization needs to be introduced. The formula for the total loss after regularization is: ； in Let be the regularization coefficient, set to 0.0001. Let L2 norm sum of squares of all weight matrices be the sum of squares of all weight matrices; if Stable within the range of 0.005~0.01 and If the value is ≤1dB, the model training effect is considered acceptable.

9. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, The noise added to each sub-signal in step one is Gaussian white noise, and the formula for generating the Gaussian white noise is as follows: ； in (·) is a function that generates standard normally distributed random numbers. The standard deviation of noise; With target signal-to-noise ratio (SNR) true The relationship is given by the formula The derivation yields that: ； in The average power of the original sub-signal, after amplitude normalization, is given by the 2ASK signal. =0.5, for QPSK signal P_s=0.

5.

10. The signal-to-noise ratio estimation algorithm based on an autoencoder according to claim 1, characterized in that, In step one, when using the Welch method to calculate the power spectrum, the window length is... Overlap length overlap rate Total signal length training set Or the total length verification set of the signal ; Number of segments for: ； in The function is the floor function; frequency resolution. ,when =600kSps =600k / 200=3kHz, when =250kSps =250k / 200=1.25kHz, and the power spectrum Corresponding frequency point ( =0,1,...,1999).

Images