Intelligent-instrument-oriented communication signal parameter blind estimation method, device and program product

Abstract

The application discloses a kind of communication signal parameter blind estimation methods, devices and program products for intelligent instrument, involve signal identification field, to with low complexity calculation cost, accurately estimate the carrier frequency offset value of electromagnetic signal, symbol rate and pulse shaping filter type.The application carries out normalization to received signal, carrier frequency offset rough correction, estimates carrier frequency offset and meets rate, carries out the pre-estimation of pulse shaping filter type from the spectrum diagram of rough correction signal, and, for square root raised cosine filter / Raised cosine filter, first sequentially carrier frequency offset correction is carried out to first signal, based on square root raised cosine filter Matching filter, and based on symbol rate construction constellation diagram, finally according to constellation diagram, distinguish whether square root raised cosine filter or Raised cosine filter.The application realizes comprehensive, accurate, reliable estimation to electromagnetic signal parameter with low calculation complexity, realizes the automation and wisdom of electromagnetic signal measurement.

Description

Technical Field

[0001] This invention relates to the field of signal recognition technology, and in particular to a blind estimation method, apparatus and program product for communication signal parameters of intelligent instruments. Background Technology

[0002] Currently, the automation and intelligence levels of electronic measuring instruments are relatively low. When measuring electromagnetic signals, measurement parameters need to be configured based on human experience, making the measurement process complex and the measurement results highly dependent on human experience, making it difficult to achieve standardized and accurate measurements. Against this backdrop, automation and intelligence have become important development directions for electronic measuring instruments. Communication signal parameter estimation is one of the key modules in electronic measuring instruments and the foundation for automatic demodulation and analysis of communication signals.

[0003] Chinese patent document CN115118564A discloses a carrier frequency deviation estimation method and apparatus, which calculates the carrier frequency offset based on a corrected phase difference sequence. Although this method has the advantages of low computational complexity and wide applicability, it is susceptible to adverse factors such as noise, interference, signal fading, and time-varying parameters, resulting in poor stability.

[0004] Chinese patent document CN112702293A proposes a signal parameter analysis method and system based on cyclic spectrum. This method estimates the carrier frequency offset and symbol rate based on the cyclic spectral density function. The accuracy of this method is positively correlated with the number of symbol points. Increasing the number of symbol points can improve the accuracy of signal parameter estimation, but increasing the number of symbol points significantly increases the computational complexity of the cyclic spectral density function, thus increasing the difficulty of technology deployment.

[0005] Chinese patent document CN117596100A proposes a signal parameter estimation method based on a neural network. This method constructs a two-dimensional matrix from the real and imaginary parts of a complex modulated signal, inputs it into a deep neural network, and outputs modulation type and symbol rate parameters. This method uses a deep neural network to uncover the potential relationship between modulation type and symbol rate parameters and the characteristics of the complex modulated signal. However, deep learning heavily relies on massive amounts of high-quality labeled data, which limits its application in practical scenarios.

[0006] Chinese patent document CN114465854A discloses a blind demodulation parameter optimization method, system, and medium based on the quasi-Newton method. This method estimates the carrier frequency and symbol rate of the signal after modulation pattern identification and obtains the optimal parameter estimates based on the demodulation effect feedback from the configured demodulator. Although this method has high accuracy, it requires prior knowledge of the signal's modulation pattern and does not consider cases with large carrier frequency offsets, resulting in poor reliability of signal parameter estimation.

[0007] In summary, existing signal parameter estimation methods have the following shortcomings:

[0008] 1) In traditional methods based on expert experience features, the experience features are designed by experts, so they rely heavily on human experience. This is not only time-consuming and labor-intensive, but also makes it difficult to fully take into account the effects of noise, interference, signal fading, and time-varying parameters in the signal estimation results.

[0009] 2) Most existing methods only consider two signal parameter types: carrier frequency offset and symbol rate, without considering pulse shaping filter parameters. However, pulse shaping filter is one of the key parameters in automatic demodulation analysis, and incorrect pulse shaping filter parameters will seriously affect the signal demodulation quality of electronic measuring instruments.

[0010] 3) In the automatic demodulation analysis of electronic measuring instruments, the estimation accuracy of parameters such as carrier frequency offset and symbol rate is required to be extremely high, often below 10 ppm. Most existing methods improve the frequency resolution by increasing the number of symbol points, thereby improving the accuracy of signal parameter estimation, but this method will significantly increase the computational complexity. Summary of the Invention

[0011] The purpose of this invention is to provide a blind estimation method for communication signal parameters of intelligent instruments, addressing all or part of the problems mentioned above, so as to accurately estimate the carrier frequency offset, symbol rate, and pulse shaping filter type of electromagnetic signals with low computational complexity.

[0012] The technical solution adopted in this invention is as follows:

[0013] A blind estimation method for communication signal parameters of intelligent instruments, comprising:

[0014] S1. Normalize the power of the received signal to obtain a normalized signal;

[0015] S2. Perform coarse carrier frequency offset correction on the normalized signal to obtain the first signal;

[0016] S3. Estimate the carrier frequency offset and symbol rate of the first signal;

[0017] S4. Based on the spectral characteristics of the first signal, the pulse shaping filter type is pre-estimated to determine whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter.

[0018] S5. If the pre-estimation result belongs to a square root raised cosine filter / raised cosine filter, then the carrier frequency of the first signal is corrected based on the estimated carrier frequency offset value; the corrected first signal is matched and filtered using a square root raised cosine filter; a constellation diagram of the filtered first signal is constructed based on the estimated symbol rate; the pulse shaping filter type is estimated from the constellation diagram to determine whether it belongs to a square root raised cosine filter or a raised cosine filter.

[0019] To address the aforementioned problems, the present invention also provides a blind estimation device for communication signal parameters of intelligent instruments, comprising a processor and a storage medium, wherein the storage medium stores computer instructions, and when the processor executes the computer instructions, it can perform the aforementioned blind estimation method for communication signal parameters of intelligent instruments.

[0020] To address the above problems, the present invention also provides another blind estimation device for communication signal parameters of intelligent instruments, comprising:

[0021] The normalization module is used to normalize the power of the received signal and output a normalized signal;

[0022] The carrier frequency offset coarse correction module is used to perform coarse correction of the carrier frequency offset on the normalized signal and output a first signal;

[0023] The carrier frequency offset and symbol rate estimation module is used to estimate the carrier frequency offset and symbol rate of the first signal;

[0024] The pulse shaping filter type identification module is used to pre-estimate the pulse shaping filter type based on the spectral characteristics of the first signal, and determine whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter; and, if the pre-estimated result belongs to a square root raised cosine filter / raised cosine filter, to perform carrier frequency correction on the first signal based on the estimated carrier frequency offset value; to perform matched filtering on the corrected first signal using a square root raised cosine filter; to construct a constellation diagram of the filtered first signal based on the estimated symbol rate; and to estimate whether the pulse shaping filter type belongs to a square root raised cosine filter or a raised cosine filter from the constellation diagram.

[0025] To address the aforementioned problems, the present invention also provides a computer program product, comprising a computer program that, when executed by a processor, can perform the aforementioned blind estimation method for communication signal parameters of intelligent instruments.

[0026] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0027] This application integrates the advantages of prior knowledge in signal processing and the ability of deep learning to fully explore the potential features of signals. Through ingenious reasoning logic, it achieves accurate estimation of the carrier frequency offset, symbol rate, and pulse shaping filter type of electromagnetic signals with lightweight computational cost. This significantly improves the ability of electronic measuring instruments to test and analyze complex signals in complex electromagnetic environments, realizing automated and intelligent electromagnetic signal measurement. Attached Figure Description

[0028] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0029] Figure 1 This is a flowchart of a blind estimation method for communication signal parameters of intelligent instruments proposed in an embodiment of this application;

[0030] Figure 2 This is a data flow diagram of the SK module in an embodiment of this application;

[0031] Figure 3 This is a structural diagram of the self-attention module in an embodiment of this application;

[0032] Figure 4 This is a flowchart illustrating the operation of the separable convolutional layer in an embodiment of this application;

[0033] Figure 5 (1) to (12) in the embodiments of this application are the eighth power spectra of 12 communication signal modulation methods;

[0034] Figure 6 (1) to (12) in the embodiments of this application are differential spectrum diagrams of 12 communication signal modulation methods;

[0035] Figure 7 (1) to (12) are the spectrum diagrams of the signal under the rectangular filter in the embodiments of this application;

[0036] Figure 8 (1) to (12) are the spectrum diagrams of the signal under the square root raised cosine filter in the embodiments of this application;

[0037] Figure 9 (1) to (12) are the spectrum diagrams of the signal under the raised cosine filter in the embodiments of this application;

[0038] Figure 10 (1) to (12) in the embodiments of this application are multi-time constellation diagrams of the signal under the square root raised cosine filter;

[0039] Figure 11 (1) to (12) in the embodiments of this application are multi-time constellation diagrams of the signal under the raised cosine filter;

[0040] Figure 12 This is a graph showing the carrier frequency offset parameter estimation error of the ASK, PSK and QAM communication signal modulation methods in the embodiments of this application;

[0041] Figure 13 This is a graph showing the symbol rate parameter estimation error of ASK, PSK and QAM communication signal modulation methods in the embodiments of this application;

[0042] Figure 14 This is a pulse shaping filter type recognition rate curve of the communication signal in the embodiments of this application. Detailed Implementation

[0043] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0044] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0045] To address the current problems of poor reliability, weak completeness, and high computational complexity in electromagnetic signal parameter estimation, this application proposes a blind estimation method, device, and program product for communication signal parameters of intelligent instruments. The aim is to achieve a more comprehensive, accurate, and reliable estimation of electromagnetic signal parameters with low computational complexity, thereby empowering electronic measuring instruments and realizing automated and intelligent electromagnetic signal measurement.

[0046] like Figure 1 As shown, the blind estimation method for communication signal parameters of intelligent instruments proposed in this application includes the following steps:

[0047] S1. Normalize the power of the received signal to obtain a normalized signal.

[0048] Assuming the received signal is R(n), the power of the received signal R(n) is normalized using equation (1).

[0049]

[0050] in, P f R represents the power value of the received signal R(n), N represents the length of the received signal R(n), n∈[0,N], R a (n) is the normalized signal.

[0051] S2. Perform coarse carrier frequency offset correction on the normalized signal to obtain the first signal.

[0052] In some feasible implementations, the centroid frequency estimation method is used to normalize the signal R. a (n) Perform a coarse estimate of the carrier frequency offset to obtain a coarse estimate of the carrier frequency offset, and then use the coarse estimate of the carrier frequency offset to adjust the normalized signal R. a The carrier frequency of (n) is coarsely corrected to obtain the first signal S. a (n).

[0053] As an optional implementation, step S2 includes the following sub-steps:

[0054] S2-1. Based on the power spectrum and amplitude spectrum of the normalized signal, perform a coarse estimation of the carrier frequency offset value of the normalized signal.

[0055] 1) For the normalized signal R a (n) Perform power spectrum estimation to obtain its power spectrum P(k), calculated using the following formula:

[0056]

[0057] 2) For the normalized signal R a (n) Perform amplitude spectrum estimation to obtain its amplitude spectrum F(k), calculated using the following formula:

[0058]

[0059] 3) Based on the power spectrum P(k) and amplitude spectrum F(k), the centroid frequency method is used to calculate a coarse estimate of the carrier frequency offset f. a The calculation formula is as follows:

[0060]

[0061] S2-2. The carrier frequency of the normalized signal is coarsely corrected using the coarse estimate of the carrier frequency offset.

[0062] The formula for coarsely correcting the carrier frequency of the normalized signal is as follows:

[0063]

[0064] After the above operations, the first signal S can be obtained. a (n).

[0065] S3, Estimating the first signal S a The carrier frequency offset and symbol rate of (n).

[0066] As an optional implementation, step S3 includes the following sub-steps:

[0067] S3-1, Regarding the first signal S a(n) Perform downsampling and extract time-domain features.

[0068] In some feasible implementations, the first signal S a (n) Perform a double downsampling to obtain the downsampled signal S b (n). A feature extraction method based on CCNN-Attention is used to mine the downsampled signal S. b The time-domain characteristics of (n).

[0069] The formula for calculating downsampling is as follows:

[0070] S b (n)=S a (2n-1)(6).

[0071] CCNN-Attention consists of a complex CNN, an SK module, a self-attention module, and a representation module in that order.

[0072] Complex CNNs for mining downsampled signals S b (n) Deeper temporal features: Complex CNN consists of 6 convolutional modules in sequence. Convolutional module-1 and convolutional module-3 include convolutional layers, batch normalization, and ReLU activation function; convolutional module-2 includes convolutional layers, batch normalization, average pooling layer, and ReLU activation function; convolutional module-4 includes convolutional layers, batch normalization, ReLU activation function, and Dropout layer; convolutional modules-5 and convolutional module-6 include convolutional layers and ReLU activation function.

[0073] See appendix Figure 2 The SK module utilizes three convolutional kernels to capture multi-scale features of the signal, adapting to parameter estimation of signals with varying sampling rates in real-world scenarios. This is achieved through three processes: separation, fusion, and selection. The "separation" process uses 3×3, 5×5, and 7×7 convolutional kernels to perform convolution operations on the signal features, yielding feature vectors U1, U2, and U3 respectively. The "fusion" process consists of an addition operation, a global average pooling layer, a fully connected layer-1, a fully connected layer-2, and a Softmax layer in sequence. The addition operation sums feature vectors U1, U2, and U3 to obtain feature vector U; the global average pooling layer performs average pooling on the features of each channel; the fully connected layer-1 is used for feature dimensionality reduction; the fully connected layer-2 is used for multi-branch feature dimensionality enhancement; and Softmax is used for feature normalization. The "selection" process multiplies the multi-branch signal feature vectors by feature vectors U1, U2, and U3 respectively, obtaining feature vectors A1, A2, and A3, and then sums feature vectors A1, A2, and A3 to output feature vector A.

[0074] See appendix Figure 3The self-attention module is used to capture the temporal dependencies of the signal. It consists of a causal convolutional layer, a multi-head self-attention layer, and a feedforward neural network in sequence. The causal convolutional layer encodes the positional information of the input signal features, mining the local temporal dependencies of the signal without destroying the temporal sequence of the signal. The multi-head self-attention layer aims to mine the global temporal dependencies of the signal, using three multi-head self-attention mechanisms. Each self-attention mechanism adopts an attention matrix sparsity method, namely a random self-attention mechanism, a window self-attention mechanism, and a global self-attention mechanism. The feedforward neural network consists of a normalized layer-1, a fully connected layer-1, a normalized layer-2, and a fully connected layer-2 in sequence.

[0075] The representation module consists of a flattening module, an average pooling layer, and a fully connected layer in sequence. It aims to map high-dimensional features to a low-dimensional space, thereby compressing the temporal features of the input signal and effectively representing the temporal information of the signal.

[0076] S3-2. Extract the 8th power spectral features of the first signal, and extract the carrier frequency offset features from the 8th power spectral features.

[0077] In some feasible implementations, the first signal S a (n) Perform an octet spectral transformation to obtain its octet spectral feature P. a (k). And by employing a VGG-based feature extraction method, P is mined. a Carrier frequency offset features in (k).

[0078] The formula for calculating the 8th power spectral transform is as follows:

[0079]

[0080] VGG consists of a CNN and representation modules in sequence. The CNN is used to mine P. a The carrier frequency offset feature in (k) is composed of two convolutional layers -1, -2, -1, -3, -4, -5, and -2 in sequence. The representation module is used for feature dimension compression and feature representation, including a flattening module, an average pooling layer, a fully connected layer, and the ReLU activation function.

[0081] S3-3, Extract the first signal S a The difference spectrum features of (n) are used to extract symbol rate features.

[0082] In some feasible implementations, the first signal S a (n) Perform a differential transform to obtain the differential signal S. r (n), and then applying the Fourier transform method, the difference spectrum feature P is obtained. r (k).

[0083] The formula for calculating the difference transform is as follows:

[0084] S r (n)=|S a (n+1)-S a (n)|·f s (8).

[0085] In the formula, f s Indicates the sampling frequency.

[0086] Difference spectral characteristics P r The calculation process for (k) is as follows:

[0087]

[0088] Furthermore, a VGG-based feature extraction method is employed to mine the differential spectral features P. r The symbol rate characteristics in (k). The structure of VGG has been described in step S3-2.

[0089] S3-4, from (downsampled signal S) b The carrier frequency offset value is precisely estimated by using the time-domain characteristics and carrier frequency offset characteristics of (n).

[0090] S3-5. Estimating the first signal S from time-domain characteristics and symbol rate characteristics. a The symbol rate of (n).

[0091] In some feasible implementations, a frequency offset estimation module is designed to estimate the carrier frequency offset, and a symbol rate estimation module is designed to estimate the symbol rate. The frequency offset estimation module takes time-domain features and carrier frequency offset features as input, and the symbol rate estimation module takes time-domain features and symbol rate features as input.

[0092] In some feasible implementations, the frequency shift estimation module described above consists of a convolutional layer, a fully connected layer-1, a fully connected layer-2, and an activation function Tanh in that order. Similarly, in some feasible implementations, the symbol rate estimation module consists of a convolutional layer, a fully connected layer-2, a fully connected layer-2, and a sigmoid function in that order.

[0093] In the carrier frequency offset estimation task, considering that the carrier frequency offset value of the signal fluctuates within a range centered at 0, the activation function Tanh is used, whose output is between -1 and 1. Therefore, during the training process of the frequency offset estimation module, the input true carrier frequency offset value needs to be normalized. The calculation formula is as follows:

[0094]

[0095] In the formula, fnorm It is the normalized carrier frequency offset value, f real It is the actual carrier frequency offset value, f max It is the largest carrier frequency offset value in the dataset.

[0096] In the symbol rate estimation task, since the sigmoid activation function is used, the output result is between 0 and 1. Therefore, during the training of the symbol rate estimation module, the input true symbol rate value needs to be normalized. The calculation formula is as follows:

[0097]

[0098] In the formula, R norm It is the normalized sign rate value, R real It is the actual symbol rate value, R max and R min These are the maximum and minimum symbol rate values ​​in the dataset, respectively.

[0099] Since both the frequency offset estimation module and the symbol rate estimation module normalize the output results, when performing precise carrier frequency offset estimation and symbol rate estimation, the normalized carrier frequency offset estimate and symbol rate value are inversely normalized to output the precise carrier frequency offset estimate f. b Symbol rate estimate R est .

[0100] S3-6, Coarse estimate of carrier frequency offset f a and carrier frequency offset precise estimate f b Calculate the first signal S a The carrier frequency offset value of (n).

[0101] coarse estimate of carrier frequency offset f a The estimation method has been explained in step S2-1 above. Combined with the precise estimate of the carrier frequency offset f obtained here... b Then the first signal S is obtained. a The carrier frequency offset value f of (n) est The calculation formula is as follows:

[0102] f est =f a +f b (12).

[0103] Using the above method, the first signal S can be estimated through steps S3-1, S3-2, S3-4, and S3-6. a The carrier frequency offset value f of (n) est From steps S3-1, S3-3, and S3-5, the first signal S can be estimated.a The symbol rate of (n).

[0104] S4. Based on the spectral characteristics of the first signal, pre-estimate the type of pulse shaping filter and determine whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter.

[0105] As an optional implementation, step S4 includes the following sub-steps:

[0106] S4-1, Regarding the first signal S a Perform a Fourier transform on (n) to obtain the spectral characteristics S(k). The transform formula is as follows:

[0107]

[0108] S4-2. Using a trained lightweight convolutional neural network (CNN), perform binary classification of pulse shaping filter types based on spectral features S(k), and determine whether the pulse shaping filter type belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter.

[0109] As an optional implementation, the lightweight CNN described above consists of a first feature extraction module and a first classifier module.

[0110] The first feature extraction module is used to perform multi-level feature extraction on the spectral features to uncover subtle features of the spectrogram. The first feature extraction module consists of a two-dimensional convolutional layer-1, bottleneck layers-1 to-7, a two-dimensional convolutional layer-2, an SE module-4, an average pooling layer, and a two-dimensional convolutional layer-3 in sequence. Specifically, the bottleneck layer is constructed in the form of a separable convolutional layer + SE module + residual connection, or a separable convolutional layer + residual connection; bottleneck layer-1 consists of a separable convolutional layer-1, an SE module-1, and a ReLU activation function in sequence; bottleneck layers-2 / 3 consist of a separable convolutional layer-2 / 3 and a ReLU activation function in sequence; bottleneck layers-4 / 6 consist of a separable convolutional layer-4 / 6, an SE module-2 / 3, and an h-swish activation function in sequence; and bottleneck layers-5 / 7 consist of a separable convolutional layer-5 / 7 and an h-swish activation function in sequence.

[0111] The first classifier module is used to fuse the multi-level features extracted by the first feature extraction module and classify the pulse shaping filter type based on the fused features, determining whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter. Specifically, the first classifier module is used to fuse composite signal features to classify the shaping filter, and it consists of a fully connected layer-1, a ReLU activation function, a fully connected layer-2, a fully connected layer-3, and a softmax layer in that order.

[0112] See Figure 4 Separable convolution is achieved through two processes: depthwise convolution and pointwise convolution. Depthwise convolution performs convolution on the input spectrogram without changing the depth, with each convolution kernel applied to one channel of the image. Pointwise convolution is similar to regular convolution, using a 1×1 convolution kernel to perform convolution operations, performing weighted combinations on the depthwise convolution to generate new feature maps.

[0113] The SE module is used to enhance important features and suppress irrelevant features, which is achieved through three processes: compression, activation, and scaling. The "compression" operation uses a global average pooling layer to perform average pooling on the features of each channel; the "activation" operation consists of a fully connected layer-1, a ReLU activation function, a fully connected layer-2, and a Sigmoid activation function in sequence. The fully connected layer-1 compresses the number of channels from C to C / Ratio, and the fully connected layer-2 restores the number of channels from C / Ratio to C, where Ratio is the compression ratio; the scaling operation weights attention to the features of each channel to enhance the network's focus on important features.

[0114] S5. If the pre-estimation result belongs to a square root raised cosine filter / raised cosine filter, then the carrier frequency of the first signal is corrected based on the estimated carrier frequency offset value; the corrected first signal is matched and filtered using a square root raised cosine filter; a constellation diagram of the filtered first signal is constructed based on the estimated symbol rate; the pulse shaping filter type is estimated from the constellation diagram to determine whether it belongs to a square root raised cosine filter or a raised cosine filter.

[0115] Step S5 shall be implemented in the following sub-steps:

[0116] S5-1, Based on the estimated carrier frequency offset value f est For the first signal S a (n) Perform carrier frequency correction.

[0117] For the first signal S a (n) The formula for carrier frequency correction is:

[0118] S c (n)=S a (n)e -j2πnfest (14),

[0119] Among them, S c (n) is the first signal for correction.

[0120] S5-2, The first signal S for correction c (n) Matched filtering is performed using a square root raised cosine filter.

[0121] Use a square root cosine filter to filter S c (n) Perform matched filtering with a filter roll-off factor of 0.5.

[0122] S5-3, Based on the estimated symbol rate R est Construct the multi-time constellation diagram of the filtered first signal. The calculation formula is as follows:

[0123] C k =m(r k (15),

[0124] Where, r k =r(K*(n-1)+1+k),k∈[0,1,...,K-1](16),

[0125] r = S c (2n-1)(17),

[0126]

[0127] In the formula, m(·) represents the constellation mapping function of the signal in the I / Q plane, and r k C represents the nth signal sequence of the decimated signal with a frequency shift of k. k Represents the signal sequence r k The constellation diagram, K is the oversampling factor, and [·] represents the rounding operation.

[0128] Because the symbol rates of the received signals differ, the monitoring receiver uses different demodulation bandwidths (different sampling rates) to acquire the signals, resulting in different oversampling factors K and an indeterminate number of constellation diagrams. Since neural networks cannot directly process constellation diagrams with an indeterminate number, this application employs a completion approach to solve this problem. Specifically, based on the demodulation bandwidth level of the monitoring receiver and the parameters of the communication signals acquired in this application, the maximum oversampling factor K is calculated. max (K calculated) max =12); for K=K max In many cases, the number of constellations does not need to be filled in; for K <K max In this case, the cumulative signal sequence length is used, and the number of constellations is completed by using the multi-time constellation diagram of the next time step, forming constellation diagram features at 12 time steps.

[0129] S5-4. Estimate the pulse shaping filter type from the constellation diagram to determine whether it is a square root raised cosine filter or a raised cosine filter.

[0130] In some feasible implementations, a trained CNN-SE is used to classify pulse shaping filter types from constellation diagrams, determining whether they belong to square root raised cosine filters or raised cosine sequential filters.

[0131] CNN-SE consists of a cascaded multi-feature extraction module, a feature fusion module, and a second classifier module. The multi-feature extraction module is composed of multiple first-feature extraction modules.

[0132] A multi-feature extraction module is used to mine subtle features of the multi-temporal constellation graph. As an optional implementation, the multi-feature extraction module consists of 12 feature extraction sub-modules, each with the same structure as the first feature extraction module of the lightweight CNN in the previous embodiment. After processing by the multi-feature extraction module, the constellation graphs C1, C2, ..., C... 12 Each of the first feature extraction modules transforms the vectors into feature vectors f1, f2, ..., f. 12 Each feature vector has a dimension of 1×L. These feature vectors are then concatenated to form a 12×L matrix f = {f1, f2, ..., f...}. 12}

[0133] The feature fusion module consists of a second SE module, convolutional layer-1, and convolutional layer-2 in sequence. The structure of the second SE module is similar to that of the SE module in the previous embodiment, the difference being the feature dimension. The SE module in step S4-2 is used to process two-dimensional feature maps, while the second SE module here is used to process one-dimensional feature vectors. Specifically, for f = {f1, f2, ..., f...} 12 The second SE module is applied to extract important features from the multi-time constellation graph and suppress irrelevant features.

[0134] The second classifier module is used for shaping filter classification and consists of a fully connected layer-1, a ReLU activation function, a fully connected layer-2, and a softmax layer in that order.

[0135] By using the above methods, it can be finally determined whether the pulse shaping filter is a square root raised cosine filter or a raised cosine filter.

[0136] The effectiveness of this application can be further illustrated by the following simulation experiments:

[0137] I. Simulation Parameter Settings

[0138] In the experiments on blind estimation methods for communication signal parameters, the performance of carrier frequency offset, symbol rate, and pulse shaping filter type estimation algorithms were verified for 12 classic communication signal types, including 4ASK, QPSK, OQPSK, 16QAM, and 64QAM. This application fully considers the application requirements of blind signal parameter estimation in practical communication scenarios. An experimental environment was built using a Rohdeschwarz SMW200A vector signal generator and a Siyi monitoring receiver. The communication signal was transmitted through the signal generator, transmitted via coaxial cable, and then received and acquired by the monitoring receiver. The monitoring receiver requires the selection of an appropriate demodulation bandwidth based on the bandwidth of the received signal. The relationship between demodulation bandwidth and sampling rate is shown in Table 1.

[0139] Table 1. Relationship between demodulation bandwidth and sampling rate

[0140] Demodulation bandwidth Sampling rate 300kHz 1.4MHz 500kHz 2MHz 800kHz 3.5MHz 1000KHz 4MHz 1250KHz 5.6MHz 1500KHz 7MHz 2000KHz 7MHz 5000KHz 14MHz 8000KHz 28MHz 10000KHz 28MHz

[0141] To establish a deep learning model for signal parameter estimation, training and validation datasets need to be constructed. These datasets are generated using a signal generator and receiver, requiring the use of a synchronization line to ensure the transmitting and receiving instruments have the same clock frequency. Simultaneously, the transmitting and receiving ends are connected using a coaxial cable, ignoring factors such as the Doppler effect in the wireless channel; in this case, the carrier frequency offset of the received signal is 0. Further, random frequency offsets are added to the dataset. On one hand, this simulates small-range carrier frequency offsets introduced by factors such as clock frequency deviations between the transmitting and receiving instruments. On the other hand, for the receiver, the transmitting frequency is unknown, and the receiving frequency cannot be accurately aligned with the transmitting frequency, resulting in a large-range carrier frequency offset. To test the performance of the parameter estimation algorithm, a test dataset needs to be constructed. This dataset is also generated using a signal generator and monitoring receiver, without using a synchronization line to preserve the carrier frequency offset caused by clock frequency differences between the transmitting and receiving instruments. The dataset covers different modulation schemes, carrier frequencies, symbol rates, pulse shaping filter types, and pulse shaping filter coefficients. The communication signal parameters are shown in Table 2.

[0142] Table 2 Communication signal parameters

[0143]

[0144] According to the specific implementation scheme of this application, a total of 12 communication signal modulation methods can be estimated, the carrier frequency estimation range is 100MHz to 6GHz, the symbol rate estimation range is 100KHz to 2MHz, and the pulse shaping filter type is identified as square root raised cosine filter, raised cosine filter, and rectangular filter, wherein the coefficient range of the square root raised cosine filter is 0.1 to 0.9. The signal length N = 10000 used for power normalization in step S1; the signal length N = 10000 used for coarse estimation and coarse correction of carrier frequency offset in step S2.

[0145] In step S3, during the simulation of carrier frequency offset and symbol rate estimation for the first signal, firstly, for the signal S with length N = 10000... a (n) Perform a 2x downsampling, then concatenate the real and imaginary parts of the downsampled signal according to the number of channels, and input it into CCNN-Attention to extract the temporal features of the signal. The key layers and parameters of CCNN-Attention are shown in Table 3. Next, for the first signal S with length N = 10000... a (n) An octet spectral transform is performed to obtain an octet spectrum. Further, the signal feature spectrum of size 64×64×3 is input into the VGG to extract the carrier frequency offset features from the signal feature spectrum. The key layers and parameters of the VGG are shown in Table 4. Next, for the first signal S with length N = 10000... a (n) Perform differential transform and Fourier transform to obtain differential spectral features. Further, input the signal feature spectrum map of size 64×64×3 into VGG to extract the symbol rate features from the signal feature spectrum map. Finally, design frequency offset estimation module and symbol rate estimation module respectively, and combine the coarse estimate of carrier frequency offset in step S2 to output the estimated carrier frequency offset and symbol rate of the signal. The frequency offset estimation module integrates the time-domain features extracted by CCNN-Attention and the carrier frequency offset features extracted by VGG, and the symbol rate estimation module integrates the time-domain features extracted by CCNN-Attention and the symbol rate features extracted by VGG. Table 5 shows the key layers and parameters of the frequency offset estimation module and the symbol rate estimation module.

[0146] Table 3 Key layers and parameters of CCNN-Attention

[0147]

[0148]

[0149] Table 4 Key Layers and Parameters of VGG

[0150]

[0151] Table 5 Key Layers and Parameters of the Frequency Offset / Symbol Rate Estimation Module

[0152]

[0153] Since carrier frequency offset estimation and symbol rate estimation are regression tasks, the loss function used for training the deep learning model is the mean squared error (MSE), calculated as follows:

[0154]

[0155] In the formula, m is the number of samples in the dataset. y is the predicted value of the i-th sample. i It is the true value of the i-th sample.

[0156] In the simulation of pulse shaping filter type estimation for the signal in step S4, firstly, for the input signal S... a (n) Perform a Fast Fourier Transform to obtain the signal spectrum. The signal length used is N = 10000. Further, input the spectrum of size 64×64×3 into a lightweight CNN and output the pulse shaping filter pre-identification type. The lightweight CNN structure is shown in Table 6. Next, in step S5, for the square root raised cosine / raised cosine filter, after completing the frequency correction, use a square root raised cosine filter with a roll-off factor of 0.5 for blind matched filtering, and construct a multi-time constellation diagram method based on the symbol rate estimate. The signal length used is N = 10000. Further, input 12 constellation diagrams of size 64×64×3 into a special CNN-SE and output the pulse shaping filter type. The CNN-SE structure is shown in Table 7.

[0157] Since pulse shaping filter type estimation is a classification task, the loss function used for training the deep learning model is cross-entropy, calculated as follows:

[0158]

[0159] In the formula, M is the number of samples in the dataset, and K is the number of categories. In this invention, K = 2. It is the predicted value of the i-th sample in the K-th class. It is the true value of the sample.

[0160] Table 6 Lightweight CNN Architecture

[0161]

[0162]

[0163] Table 7 CNN-SE Structure

[0164]

[0165]

[0166] This application uses ppm, a commonly used unit in instruments, to measure the estimation error of signal parameters such as carrier frequency offset and symbol rate. The formula for calculating the carrier frequency offset estimation error is as follows:

[0167]

[0168] In the formula, f is the actual carrier frequency offset of the signal, and R s It is the symbol rate value of the signal; f est It is the estimated carrier frequency offset of the signal.

[0169] The formula for calculating the symbol rate estimation error is as follows:

[0170]

[0171] In the formula, R est It is the signal symbol rate estimate.

[0172] II. Simulation Experiment Content and Results

[0173] Simulation 1: By Figure 5 Neutron diagrams (1) to (12) show the octet spectrum of communication signals, namely 2ASK, 4ASK, 8ASK, BPSK, QPSK, OQPSK, 8PSK, 16QAM, 32QAM, 64QAM, 128QAM, and 256QAM. It can be seen that all 12 communication signals have high-frequency peaks at the carrier frequency offset points, exhibiting obvious carrier frequency shift characteristics. Therefore, by combining octet spectrum transformation and deep learning methods, the carrier frequency shift parameters of the signals can be effectively estimated.

[0174] Simulation 2: By Figure 6 Neutron diagrams (1) to (12) show the differential spectra of communication signals 2ASK, 4ASK, 8ASK, BPSK, QPSK, OQPSK, 8PSK, 16QAM, 32QAM, 64QAM, 128QAM, and 256QAM, respectively. It can be seen that all 12 communication signals have high-frequency peaks at the symbol rate frequency point, exhibiting obvious symbol rate characteristics. Therefore, combining differential spectral transform and deep learning methods can effectively estimate the symbol rate parameters of the signals.

[0175] Simulation 3: By Figure 7 Neutron diagrams (1) to (12) Figure 8 Neutron diagrams (1) to (12), and Figure 9The neutron diagrams (1) to (12) show the spectrum diagrams of 2ASK, 4ASK, 8ASK, BPSK, QPSK, OQPSK, 8PSK, 16QAM, 32QAM, 64QAM, 128QAM, and 256QAM communication signals under rectangular filters, square root raised cosine filters, and raised cosine filters, respectively. It can be seen that the spectrum diagrams of rectangular filters are significantly different from those of the other two types of pulse shaping filters, and can be accurately distinguished by using deep learning methods.

[0176] Simulation 4: By Figure 10 Neutron diagrams (1) to (12) Figure 11 The neutron diagrams (1) to (12) show the multi-time constellation diagrams of communication signals 2ASK, 4ASK, 8ASK, BPSK, QPSK, OQPSK, 8PSK, 16QAM, 32QAM, 64QAM, 128QAM, and 256QAM under the square root raised cosine filter and the raised cosine filter, respectively. It can be seen that after blind matching, the multi-time constellation diagram under the square root raised cosine filter has obvious signal constellation characteristics, that is, the signal constellation characteristics at the optimal time. However, the raised cosine filter does not have obvious signal constellation characteristics due to filter mismatch. The deep learning method can effectively distinguish them.

[0177] Simulation 5: By Figure 12 As shown, the carrier frequency offset parameter estimation error curves for ASK, PSK, and QAM communication signals are presented. It can be seen that at a signal-to-noise ratio of 25dB, the carrier frequency offset parameter estimation error for ASK, PSK, and QAM modulation schemes is less than 5ppm.

[0178] Simulation 6: By Figure 13 The figure shows the symbol rate parameter estimation error curves for ASK, PSK, and QAM communication signals. It can be seen that at a signal-to-noise ratio of 25dB, the symbol rate parameter estimation errors for ASK, PSK, and QAM modulation schemes are all less than 6ppm.

[0179] Simulation 7: By Figure 14 The figure shows the pulse shaping filter type recognition rate curves for 12 different communication signal modulation methods. It can be seen that the recognition accuracy of the pulse shaping filter type is greater than 90% when the signal-to-noise ratio is 35dB.

[0180] The simulation results above show that:

[0181] (1) Considering that methods based on expert experience and knowledge perform poorly in complex electromagnetic environments, and that deep learning relies heavily on massive amounts of high-quality labeled data, this application proposes a knowledge-data jointly driven blind estimation method for communication signal parameters, which integrates signal processing methods and deep learning technology. On the one hand, deep learning technology fully explores the potential features of the signal, gets rid of the dependence on human experience, and alleviates the influence of undesirable factors such as noise and time-varying parameters. On the other hand, it integrates the prior knowledge of signal processing into the communication signal parameter estimation method, which greatly improves the accuracy of parameter estimation.

[0182] (2) This application fully considers the needs of practical application scenarios and has high engineering application value. On the one hand, the experimental environment is built using a Rohde Schwarz SMW200A vector signal generator and a Siyi monitoring receiver, and all experimental data are collected in this environment. On the other hand, considering the algorithm deployment in practical application scenarios, a large number of lightweight deep neural network modules are used. Even for multi-head self-attention networks, this invention reduces the complexity of matrix operations by using the attention matrix sparsification method.

[0183] Based on the concept of this application, this application also proposes a blind estimation device for communication signal parameters of intelligent instruments, including a processor and a storage medium, wherein the storage medium stores computer instructions, and when the processor executes the computer instructions, it can perform the blind estimation method for communication signal parameters of intelligent instruments described above.

[0184] Furthermore, this application also proposes another blind estimation device for communication signal parameters of intelligent instruments, which includes:

[0185] The normalization module is used to normalize the power of the received signal and output a normalized signal;

[0186] The carrier frequency offset coarse correction module is used to perform coarse correction of the carrier frequency offset on the normalized signal and output a first signal;

[0187] The carrier frequency offset and symbol rate estimation module is used to estimate the carrier frequency offset and symbol rate of the first signal;

[0188] The pulse shaping filter type identification module is used to pre-estimate the pulse shaping filter type based on the spectral characteristics of the first signal, and determine whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter; and, if the pre-estimated result belongs to a square root raised cosine filter / raised cosine filter, to perform carrier frequency correction on the first signal based on the estimated carrier frequency offset value; to perform matched filtering on the corrected first signal using a square root raised cosine filter; to construct a constellation diagram of the filtered first signal based on the estimated symbol rate; and to estimate whether the pulse shaping filter type belongs to a square root raised cosine filter or a raised cosine filter from the constellation diagram.

[0189] The specific configuration information of the operations performed by the normalization module, carrier frequency offset coarse correction module, carrier frequency offset and symbol rate estimation module, and pulse shaping filter type identification module can be referred to the relevant features in steps S1 to S5 of the method embodiment above.

[0190] Furthermore, this application also proposes a computer program product, including a computer program that, when run by a processor, can execute the blind estimation method for communication signal parameters of intelligent instruments described in the preceding embodiments.

[0191] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A blind estimation method for communication signal parameters of intelligent instruments, characterized in that, include: S1. Normalize the power of the received signal to obtain a normalized signal; S2. Perform coarse carrier frequency offset correction on the normalized signal to obtain the first signal; S3. Estimate the carrier frequency offset and symbol rate of the first signal; S4. Based on the spectral characteristics of the first signal, the pulse shaping filter type is pre-estimated to determine whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter. S5. If the pre-estimation result belongs to a square root raised cosine filter / raised cosine filter, then the carrier frequency of the first signal is corrected based on the estimated carrier frequency offset value; the corrected first signal is matched and filtered using a square root raised cosine filter; a constellation diagram of the filtered first signal is constructed based on the estimated symbol rate; the pulse shaping filter type is estimated from the constellation diagram to determine whether it belongs to a square root raised cosine filter or a raised cosine filter.

2. The blind estimation method for communication signal parameters of intelligent instruments as described in claim 1, characterized in that, Step S2 includes: S2-1. Based on the power spectrum and amplitude spectrum of the normalized signal, perform a coarse estimation of the carrier frequency offset value of the normalized signal; S2-2. The carrier frequency of the normalized signal is coarsely corrected using the coarse estimate of the carrier frequency offset.

3. The blind estimation method for communication signal parameters of intelligent instruments as described in claim 2, characterized in that, The method for estimating the carrier frequency offset of the first signal includes: S3-1. Downsample the first signal to obtain a downsampled signal, and extract the time-domain features of the downsampled signal; S3-2. Extract the 8th power spectral features of the first signal, and extract the carrier frequency offset features from the 8th power spectral features; S3-3. Perform a precise estimation of the carrier frequency offset value based on the time-domain features and the carrier frequency offset features; S3-4. Calculate the carrier frequency offset value of the first signal from the coarse estimate of the carrier frequency offset and the fine estimate of the carrier frequency offset.

4. The blind estimation method for communication signal parameters of intelligent instruments as described in claim 3, characterized in that, Methods for estimating the symbol rate of the first signal include: S3-5. Extract the differential spectrum features of the first signal, and extract the symbol rate features from the differential spectrum features; S3-6. Estimate the symbol rate of the first signal based on the time-domain features and the symbol rate features.

5. The blind estimation method for communication signal parameters of intelligent instruments as described in claim 1, characterized in that, Step S4 includes: S4-1. Perform a Fourier transform on the first signal to obtain its spectral characteristics; S4-2. Using a trained lightweight convolutional neural network, perform binary classification of pulse shaping filter types based on the spectral features to determine whether the pulse shaping filter type belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter.

6. The blind estimation method for communication signal parameters of intelligent instruments as described in claim 5, characterized in that, The lightweight convolutional neural network is composed of a first feature extraction module and a first classifier module cascaded together; wherein: The first feature extraction module is used to perform multi-level feature extraction on the spectral features; The first classifier module is used to fuse the multi-level features extracted by the first feature extraction module, and classify the pulse shaping filter type based on the fused features, determining whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter.

7. The blind estimation method for communication signal parameters of intelligent instruments as described in claim 6, characterized in that, In step S5, estimating whether the pulse shaping filter type is a square root raised cosine filter or a raised cosine filter based on the constellation diagram includes: The trained CNN-SE is used to classify pulse shaping filter types based on the constellation diagram, determining whether they belong to square root raised cosine filters or raised cosine filters; where: The CNN-SE is composed of a multi-feature extraction module, a feature fusion module, and a second classifier module cascaded together; the multi-feature extraction module is composed of multiple first feature extraction modules.

8. A blind estimation device for communication signal parameters of intelligent instruments, comprising a processor and a storage medium, wherein the storage medium stores computer instructions, characterized in that, When the processor executes the computer instructions, it can perform the blind estimation method for communication signal parameters of intelligent instruments as described in any one of claims 1-7.

9. A blind estimation device for communication signal parameters of intelligent instruments, characterized in that, include: The normalization module is used to normalize the power of the received signal and output a normalized signal; The carrier frequency offset coarse correction module is used to perform coarse correction of the carrier frequency offset on the normalized signal and output a first signal; The carrier frequency offset and symbol rate estimation module is used to estimate the carrier frequency offset and symbol rate of the first signal; The pulse shaping filter type identification module is used to pre-estimate the pulse shaping filter type based on the spectral characteristics of the first signal, and determine whether it belongs to a rectangular filter or a square root raised cosine filter / raised cosine filter. Furthermore, if the pre-estimation result belongs to a square root raised cosine filter / raised cosine filter, then the first signal is subjected to carrier frequency correction based on the estimated carrier frequency offset value; the corrected first signal is subjected to matched filtering using a square root raised cosine filter; a constellation diagram of the filtered first signal is constructed based on the estimated symbol rate; and the pulse shaping filter type is estimated from the constellation diagram to be either a square root raised cosine filter or a raised cosine filter.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is run by the processor, it can execute the blind estimation method for communication signal parameters of intelligent instruments as described in any one of claims 1-7.

Images