Secondary modulation signal recognition method and system based on joint model deep learning

Abstract

The application discloses a secondary modulation signal recognition method and system based on a joint model deep learning, relates to electronic countermeasures, wireless communication and signal processing technology, and comprises the following steps: obtaining a sample secondary modulation signal and performing pretreatment; based on signal reconstruction and phase space statistical analysis permutation entropy, space-time characteristic samples are extracted from the sample secondary modulation signal after pretreatment; a lightweight deep learning model is adopted to jointly construct a data local space characteristic perception model and a time sequence backtracking control model, and the extracted space-time characteristic samples are used to train the joint model; space-time characteristics of a target secondary modulation signal are extracted, and a recognition result is output by using the trained joint model. The application can solve the problems of low recognition accuracy, large operation amount, and large number of prior samples required in the field of signal and information processing.

Description

Technical Field

[0001] This application relates to the fields of electronic countermeasures, wireless communication and signal processing technology, and in particular to a method and system for recognizing secondary modulated signals based on joint model deep learning. Background Technology

[0002] Double-modulated signals are generated by performing two independent modulation operations on a signal in a communication system. They aim to achieve multi-dimensional signal optimization through a layered modulation strategy. Essentially, the modulated signal (the output of the first modulation) is used as the baseband signal and then loaded onto another carrier for secondary parameter adjustments (such as amplitude, frequency, or phase), forming a double-modulated waveform. Double-modulated signals are essentially a two-stage modulation operation: an inner modulation followed by an outer modulation. They are widely used in various communication systems and UAV telemetry and control. Double-modulated identification signals reveal the inherent modulation parameters and information encoding mechanisms through the hierarchical structure of the modulated signal, demonstrating key application value in military and civilian communications. Double-modulated identification signals can be used for anti-interception communications. In tactical radios and UAV telemetry links, double modulation hides signal characteristics. The receiving end needs to analyze the spreading code and modulation parameters to achieve reliable demodulation at low signal-to-noise ratios, while making it difficult for adversaries to reconstruct the signal through conventional spectrum analysis. Secondary modulation signals also play a crucial role in satellite communications. For example, satellite television broadcasting (such as DVB-S2X) employs layered secondary modulation signals to adapt to the receiving capabilities of different terminals. Ground stations dynamically switch the inner demodulation mode by identifying the outer QPSK parameters, achieving adaptive transmission. Secondary modulation signals can also be used for network spectrum sharing. Due to their wide range of applications, the identification, analysis, and detection of secondary modulation signals are of paramount importance.

[0003] Traditional signal modulation recognition methods can be broadly categorized into two types: likelihood theory-based and feature-based methods. Likelihood theory-based methods typically achieve optimal recognition accuracy in the Bayesian estimation sense, but they suffer from high computational complexity. Feature-based methods primarily learn representative features from training samples and use the trained model to classify the input signal. Typical feature types used in feature-based methods include instantaneous time-domain features, transform-domain features, and statistical features. Compared to likelihood theory-based methods, feature-based methods usually only yield suboptimal solutions, but they have lower computational complexity and stronger multi-modulation recognition capabilities.

[0004] ZL 202311297799.0 An Automatic Modulation Recognition Method for Composite Modulation Signals in Aerospace Telemetry and Control Links This patent discloses an automatic modulation recognition method for composite modulation signals in aerospace telemetry and control links. This method uses traditional signal processing methods (cyclic spectrum, wavelet transform, etc.) to identify secondary modulation signals, without employing deep learning methods. The recognition accuracy and stability are insufficient, and it requires prior knowledge of the secondary modulation signals.

[0005] ZL 202310624327.5 A method for blind identification of composite modulated signals based on frequency domain data compressed sensing. This patent uses compressed sensing technology combined with shallow neural networks to blindly identify secondary modulated signals, which improves the identification accuracy. However, the design of the neural network is relatively simple. It only uses convolutional layers and does not use a temporal network structure, so it cannot analyze the temporal characteristics of the signal and the identification accuracy is insufficient. Summary of the Invention

[0006] This application provides a method and system for recognizing secondary modulated signals based on joint model deep learning, which solves the problems of low recognition accuracy, large computational load, and large number of prior samples required by secondary modulated signal recognition methods in the field of signal and information processing.

[0007] This application provides a method for identifying secondary modulation signals based on joint model deep learning, including:

[0008] Acquire the sample secondary modulation signal and perform preprocessing;

[0009] For the preprocessed sample modulated signal, spatiotemporal feature samples are extracted based on signal reconstruction and phase space statistical analysis permutation entropy.

[0010] A lightweight deep learning model is used to jointly construct a data local spatial characteristic perception model and a temporal backtracking control model, and the extracted spatiotemporal feature samples are used to train the joint model.

[0011] The spatiotemporal features of the target secondary modulated signal are extracted, and the recognition result is output using the trained joint model. The extracted spatiotemporal features are respectively input into the trained data local spatial characteristic perception model and the time-series backtracking control model. After the two models complete their respective classification calculations, the classification weights of the secondary modulated signal of the two models are optimized based on the optimal association information criterion. The modulation mode recognition results of the secondary modulated signal are then fused according to the classification weights to output the recognition result.

[0012] This application provides a secondary modulation signal recognition system based on joint model deep learning, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the secondary modulation signal recognition method based on joint model deep learning as described above.

[0013] The embodiments of this application adopt a lightweight processing approach, which optimizes the amount of computation and has a high classification and recognition accuracy in complex and ever-changing electromagnetic environments. It can optimize signal classification calculation, while requiring less prior knowledge, having strong robustness, and having lower training overhead than other methods with the same recognition accuracy level, making it highly practical for engineering applications.

[0014] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0016] Figure 1 This is a schematic diagram of the basic process of the secondary modulation signal recognition method based on joint model deep learning in this application embodiment;

[0017] Figure 2 This illustration shows a comparison of the recognition accuracy performance of the method in this application embodiment with other methods. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] The technical problem this application aims to solve is that traditional methods for identifying second-modulated signals in complex electromagnetic environments have low recognition rates, involve numerous feature extraction steps, and have high computational complexity for classification, requiring substantial prior information about signal parameters such as signal code rate and modulation index. Furthermore, some traditional signal identification methods are computationally intensive and have long latency, and this high computational complexity hinders their deployment on hardware-restricted devices. This application provides a second-modulated signal identification method based on joint model deep learning, such as... Figure 1 As shown, it includes the following steps:

[0020] In step S101, the sampled secondary modulated signal is acquired and preprocessed. In some embodiments, the main carrier and sidelobe characteristics of the secondary modulated signal are located by scanning the frequency band using Fast Fourier Transform; and the signal power spectral density is calculated, combined with threshold triggering to lock the effective signal window. Then, the secondary modulated signal is effectively captured based on an analog-to-digital converter (ADC), and the gain of the front-end amplifier is dynamically adjusted to avoid ADC saturation or quantization noise degradation. After the above process, the secondary modulated signal acquisition data is input, providing an initial data source for subsequent steps.

[0021] After signal acquisition and input, the signal is decomposed into corresponding envelopes based on the dynamic envelope multi-level decomposition method, and then reassembled to realize signal data preprocessing.

[0022] In step S102, spatiotemporal feature samples are extracted from the preprocessed sample secondary modulated signal based on signal reconstruction and phase space statistical analysis permutation entropy. In a specific example, after completing the preprocessing of the secondary modulated signal, the obtained signal normalized sequence is decomposed, and then the decomposed sequence is mapped from low dimension to high dimension phase space to realize the phase space reconstruction of the signal. An adaptive scaling factor and noise suppression weight coefficient are introduced to optimize the phase space statistical analysis of the reconstructed signal, thereby realizing the nonlinear feature extraction of the signal.

[0023] In step S103, a lightweight deep learning model is used to jointly construct a data local spatial characteristic perception model and a temporal backtracking control model, and the joint model is trained using the extracted spatiotemporal feature samples. In some embodiments, the jointly constructed data local spatial characteristic perception model and temporal backtracking control model include: a data local spatial characteristic perception module, a temporal backtracking control module, and a joint classification calculation module. The joint model is used to process the extracted temporal and spatial features of the secondary modulation signal using the data local spatial characteristic perception model and the temporal backtracking control model respectively, and then fuses the classification results of the two models according to the joint classification calculation module.

[0024] In step S104, the spatiotemporal features of the target secondary modulation signal are extracted, and the recognition result is output using the trained joint model. The extracted spatiotemporal features are respectively input into the trained data local spatial characteristic perception model and the time-series backtracking control model. After the two models complete their respective classification calculations, the classification weights of the secondary modulation signal of the two models are optimized based on the optimal association information criterion. The modulation mode recognition results of the secondary modulation signal are then fused according to the classification weights to output the recognition result.

[0025] The embodiments of this application adopt a lightweight processing approach, which optimizes the amount of computation and has a high classification and recognition accuracy in complex and ever-changing electromagnetic environments. It can optimize signal classification calculation, while requiring less prior knowledge, having strong robustness, and having lower training overhead than other methods with the same recognition accuracy level, making it highly practical for engineering applications.

[0026] In some embodiments, acquiring and preprocessing the sampled secondary modulation signal further includes:

[0027] The acquired secondary modulation signal is segmented into sampling points, saved as a vector format, and then normalized.

[0028] Next, all extreme points of the analytic signal are calculated, and the extreme envelope v is fitted using a cubic spline function. + (t).

[0029] Then, obtain all the extreme values ​​of the function and fit their effect envelope v. - (t), and calculate the mean m(t) of the two envelopes.

[0030] Subtracting the calculated mean from the normalized quadratic modulated signal yields a new signal. This decomposition process is repeated until a signal output satisfying the conditions is obtained. The result can be expressed as:

[0031] r1(t)=x(t)-k1 n (t)=x(t)-imf1(t)

[0032] Where r1(t) represents the decomposed second-modulated signal, x(t) represents the input sampled data, and imf1(t) represents the first-order eigenmode function of the signal. This is the mean signal. Based on the first decomposition of the signal data, the envelope of the signal can be preliminarily analyzed and processed to achieve noise reduction.

[0033] The decomposition process is repeated for each iteration of the newly generated signal. Decomposition is performed on each new signal to obtain second-order, third-order, ..., n-order components, thus acquiring the noise components of the quadratic modulation signal. This process is repeated for each new signal, where the residual signal r... nWhen (t) becomes a monotonic function and cannot be further decomposed into oscillating components, or becomes a constant, the decomposition terminates. After multi-stage filtering of the signal through the above process, the signal is represented as:

[0034]

[0035] Among them, imf n (t) is the nth order envelope component.

[0036] The sum of the third-order components and the residual terms is selected as the signal preprocessing result.

[0037] X input (t)=imf last (t)+imf last-1 (t)+imf last-2 (t)+r n (t)

[0038] Among them, X input (t) represents the result after preprocessing the output signal, and imf1(t) represents the envelope components of each order r. n (t) represents the nth-order residual term. Based on multi-level component filtering of the secondary modulated signal, the influence of noise components on the signal is eliminated, achieving efficient preprocessing of the input data. The processed result is then input into the next step for feature extraction based on signal reconstruction and phase space statistical analysis.

[0039] After preprocessing the second-modulated signal, the obtained signal normalization sequence is decomposed. The decomposed sequence is then mapped from low-dimensional to high-dimensional phase space to achieve phase space reconstruction. An adaptive scaling factor and noise suppression weight coefficient are introduced to optimize the phase space statistical analysis of the reconstructed signal, thereby achieving nonlinear feature extraction. In some embodiments, extracting spatiotemporal feature samples from the preprocessed sample second-modulated signal based on signal reconstruction and phase space statistical analysis permutation entropy includes:

[0040] The preprocessed sample secondary modulated signal is subjected to phase parameter transformation to construct the phase space. The inverse transformation of the input y-row parameters satisfies:

[0041]

[0042] Where x is the preprocessed signal data, For the phase parameter of the signal, R is the output after phase transformation, and I is the function for operating on the real and imaginary parts of the data.

[0043] Based on the structure of the phase parameter transformation, the phase space is reconstructed from the time series data using the time delay embedding method, enabling data feature analysis in a high-dimensional space, as shown in the following equation:

[0044]

[0045] in, The output is the reconstruction result. The signal data is after phase transformation; N is the length of the preprocessed signal, and s is the scale of data reconstruction. The reconstructed signal vector is obtained based on the scale.

[0046] After obtaining the reconstructed signal, the signal sequence in the phase space is quantized. The quantization interval is set according to the reconstructed signal vector, and the same discrete value is assigned to the reconstructed signal within the same interval to obtain sequence symbols. The frequency of occurrence of each different sequence symbol is then counted. The permutation entropy, a statistical feature of the phase space of the reconstructed signal, is calculated as shown in the following formula, thereby extracting the nonlinear features of the signal:

[0047]

[0048] Where, p l =N l / N represents the proportion of the number of the l-th symbol in the total number of symbols, L is the total number of symbols, and m = N / s is the length of the reconstructed signal sequence.

[0049] The phase space statistical feature vector extracted from the reconstructed signal is used as the extracted spatiotemporal feature sample, which is then output to the subsequent data local spatial characteristic perception model and time-series backtracking control model for secondary modulation signal feature analysis and classification.

[0050] After extracting the features of the second-modulated signal, a lightweight deep learning model is employed, namely, jointly constructing a data local spatial characteristic perception model and a temporal backtracking control model, and training the model for subsequent classification and recognition of the obtained signal features. In some embodiments, training the joint model using the extracted spatiotemporal feature samples includes:

[0051] First, the signal spatial features are compressed using the local sensing computation unit of the data local spatial characteristics sensing model.

[0052] The local sensing computing unit extracts the local vector of signal features, multiplies it with the sensing computing kernel, and performs sliding operations to realize the successive analysis of each local feature of the signal, obtaining the scale-compressed spatial feature vector. The class probability vector is then calculated through an exponential normalization function, which improves the classification and recognition accuracy of the second-modulated signal and reduces the computational complexity.

[0053] The spatial feature vectors obtained from the local spatial characteristic perception model of data are reconstructed by scaling transformation to adapt to the parameter input requirements of the time-series backtracking control model.

[0054] Next, the vectors obtained from the local spatial characteristic perception model are reconstructed by scaling to adapt to the parameter input requirements of the time-series backtracking control model. The scaled vectors are then input into the time-series backtracking control model. This model uses a state-selective update unit to store the feature data involved in the computation and updates the parameter z during each training calculation output. t Control historical state data h t-1 With the status data to be updated The mixing ratio satisfies:

[0055]

[0056] in, For the new state of the output model, h t-1 For historical state data, W h Let x be the scaling function. t The input data for the model is the feature vector obtained from the local spatial characteristics perception model, b h The bias term, tanh, is the hyperbolic tangent function used to activate the model's calculation results. Exponential vector normalization is then applied to transform the results into a probability vector, i.e., the modulation scheme identification vector for the secondary modulation signal. By selectively updating historical and current data, the vanishing gradient problem caused by processing long time series data can be reduced, thus improving the accuracy of secondary modulation signal feature calculation.

[0057] In some embodiments, training the joint model using the extracted spatiotemporal feature samples further includes: after calculating the data local spatial characteristic perception model and the temporal backtracking control model, the outputs of the two models are fused according to the joint classification calculation module, and the loss of the negative expected log probability optimization loss function between the model output and the sample data is calculated, that is, the loss of the negative expected log probability optimization loss function between the model output and the sample data is calculated, and the loss is used as the gradient for parameter iterative update to train the model, as shown in the following formula.

[0058]

[0059] Among them, W h b h For the linear calculation parameters and offset parameters of the model, y t (i) represents the probability that the sample value is i at time t, which is either 0 or 1. Let be the probability that the signal identification and classification result at time t is i, which is the output result of the joint model of data local spatial characteristic perception model and temporal backtracking control, [W]. h ,b h ] is the joint vector matrix composed of the corresponding parameters of the model, and η is the update step size, which can be set to a fixed value.

[0060] The model parameters are iteratively updated until the model converges by calculating the loss function. Then, signal classification and recognition are achieved based on a joint model combining local spatial data perception and temporal backtracking control.

[0061] After completing the construction and training analysis of the joint model in the above steps, based on the optimal association information criterion, the data local spatial characteristic perception model and the time-series backtracking control model are combined to achieve refined mining of data characteristics and improve the accuracy of weight allocation.

[0062] First, the obtained secondary modulation signal feature data is input into the trained local spatial characteristic perception model, along with the corresponding modulation mode identification output. Simultaneously, the feature vector is input into the time-series backtracking control model to calculate the output type identification probability vector and perform normalization. After completing the classification calculations for both models, the classification weights of the secondary modulation signals for both models are optimized based on the optimal association information criterion to reduce classification error, satisfying the following:

[0063]

[0064] Among them, w i Let I be the classification weights for the model, and p be the association information function. i The classification probability is calculated by the data local spatial characteristic perception model and the temporal backtracking control model.

[0065] In some embodiments, the modulation scheme identification results of the secondary modulation signal are fused according to the classification weights to output an identification result that satisfies the following:

[0066] C = w1c1 + w2c2

[0067] Where C is the joint classification vector, w1 and w2 are the weight values ​​of the data local spatial characteristic perception model and the time-series backtracking control model, respectively, and c1 and c2 are the classification result vectors of the data local spatial characteristic perception model and the time-series backtracking control model, respectively. After calculating the joint classification vector, the vector is normalized to calculate the modulation mode recognition probability and sorted. The category corresponding to the highest probability is selected as the final classification and recognition result of the secondary modulation signal modulation mode.

[0068] The method in this application fully considers the characteristics of the second-modulated signal and uses a parameter transformation module to reduce the adverse effects of the channel on the signal data. In terms of recognition accuracy, the method in this application significantly outperforms existing deep learning recognition methods based on CNN network models. The method in this application also has fewer parameters than other existing deep learning-based modulation recognition methods.

[0069] The method in this application can directly use the raw modulation recognition data as input without requiring additional resources for data preprocessing. This method comprehensively considers recognition accuracy and model complexity, achieving a recognition accuracy close to the highest level of current deep learning modulation recognition methods with minimal model parameters.

[0070] This application also proposes implementation examples of a secondary modulation signal recognition method based on joint model deep learning:

[0071] In this example, the secondary modulated communication signal data is first acquired as input, and preprocessed based on dynamic envelope multi-level decomposition to obtain the denoised secondary modulated signal. Then, the signal is reconstructed in phase space based on the phase statistical transformation of the signal data, and the phase space statistical features of the signal are calculated to achieve feature extraction. The obtained feature vector is then input into the joint model of data local spatial characteristic perception and time-series backtracking control. The joint model is trained and analyzed based on sample data to obtain the joint model, which is used to process the features of the secondary modulated signal and perform classification and recognition, and finally outputs the recognition result of the secondary modulated signal.

[0072] The specific implementation process of the efficient identification method for secondary modulation signals based on joint model deep learning:

[0073] The initial conditions were as follows: all models were evaluated on simulated secondary modulation signals including 2FSK-PM, BPSK-FM, BPSK-PM, MSK-FM, OQPSK-FM, OQPSK-PM, QPSK-FM, QPSK-PM, PCM-FM, PCM-PM, and MSK-DSB and PCM-DSB modulation schemes. The dataset was randomly divided into training, validation, and test sets in a 6:2:2 ratio. A GeForce GTX 1080Ti GPU and Keras were used, with TensorFlow as the backend. No expert feature extraction or other preprocessing was performed on the original signals; instead, the original time-series features were directly learned from the high-dimensional data using a real-number network, where the real and imaginary parts of the input signal were respectively input into the model as two dimensions.

[0074] Baseline comparisons are provided using representative State-of-the-art (SOTA) models, including the IC-AMCNET and MCNET models based on CNN networks, the LSTM2 and GRU2 models based on RNN networks, and the MCLDNN model, which currently boasts the highest recognition accuracy.

[0075] Table 1 selects multiple metrics for performance comparison and complexity analysis, including parameter count, training time, testing time, highest recognition accuracy under all signal-to-noise ratios (SNR), and average recognition accuracy under all SNR. The method in this application has the fewest parameters and the shortest training time. Its testing time is only 0.003 seconds / sample slower than the fastest IC-AMCNET model. Its accuracy performance is slightly lower than the MCLDNN model, but comparable to the LSTM2 model. Overall, compared to other methods, the method in this invention has the fewest parameters; even compared to the MCNET model with the fewest parameters, its parameter count is less than 70%. Compared to the high-performance but complex GRU2, LSTM2, and MCLDNN models, the method in this application has a significant advantage in time cost. Compared to the CNN-based IC-ACMNET and MCNET models, the time cost of this method is comparable, but its accuracy is higher.

[0076] Table 1. Performance Comparison between the Invention and Existing Methods

[0077]

[0078] Simulations and experiments were conducted based on the above initial conditions. Figure 2 The paper presents the recognition accuracy of various models as a function of signal-to-noise ratio. The MCLDNN model exhibits the highest recognition accuracy, but the method in this application, with less than one-fifth the parameters of MCLDNN, achieves a recognition accuracy very close to that of MCLDNN. Furthermore, the recognition accuracy of the method in this application is superior to that of the GRU2 model. The IC-AMCNET and MCNET models based on CNN networks show lower recognition accuracy in data simulations compared to the method proposed in this patent, achieving a maximum accuracy of only around 80%.

[0079] In terms of overall performance, the method in this application balances complexity and accuracy, and is an efficient method for identifying secondary modulation signals based on joint model deep learning.

[0080] This application also provides a secondary modulation signal recognition system based on joint model deep learning, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of the secondary modulation signal recognition method based on joint model deep learning as described above.

[0081] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0082] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0083] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0084] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.

Claims

1.A method for secondary modulation signal recognition based on joint model deep learning, characterized in that, The method comprises the following steps: obtaining a sample quadratic modulation signal and pre-processing the same; extracting a spatio-temporal feature sample from the pre-processed sample quadratic modulation signal based on signal reconstruction and phase space statistical analysis permutation entropy; training a joint model using the extracted spatio-temporal feature sample by adopting a lightweight deep learning model to jointly construct a data local space characteristic perception model and a time series backtracking control model; extracting a spatio-temporal feature of a target quadratic modulation signal and outputting a recognition result using the trained joint model, wherein the extracted spatio-temporal feature is respectively input into the data local space characteristic perception model and the time series backtracking control model, after the two models complete their respective classification calculations, the quadratic modulation signal classification weights of the two models are optimized based on an optimal correlation information criterion, and the modulation mode recognition result of the quadratic modulation signal is fused according to the classification weights to output the recognition result; training the joint model using the extracted spatio-temporal feature sample comprises: performing compression processing on the signal space feature through a local perception calculation unit of the data local space characteristic perception model; the local perception calculation unit multiplies the local vector of the signal feature with a perception calculation kernel and performs sliding operation to realize successive analysis of each local feature of the signal, thereby obtaining a scale-compressed space feature vector; performing scale transformation reconstruction on the space feature vector obtained by the data local space characteristic perception model to adapt the parameter input requirement of the time series backtracking control model; The transformed vector is input into a time-series backtracking control model, which uses a state-selective update unit to store the feature data involved in the computation and updates the parameters during each training calculation output. Control historical state data With the status data to be updated The mixing ratio satisfies: wherein, is the new state of the output model, is the historical state data, is the model input data, is the bias term, is the scaling function, tanh is the hyperbolic tangent function, denotes the time at which the state is at. 2.The joint model deep learning based secondary modulation signal recognition method of claim 1, wherein, obtaining a sample quadratic modulation signal and pre-processing the same comprises: scanning a frequency band by fast Fourier transform to locate the main carrier and sidelobe characteristics of the quadratic modulation signal; and calculating the signal power spectral density and triggering the locking of the effective signal window according to the threshold value; obtaining the quadratic modulation signal based on an analog-to-digital converter (ADC). 3.The joint model deep learning based secondary modulation signal recognition method of claim 2, wherein, Further comprising: dividing and normalizing the quadratic modulation signal obtained based on the ADC; calculating the extreme points of the normalized quadratic modulation signal and fitting the extreme envelope to obtain all extreme values; fitting the effect envelope of the extreme value and calculating the mean values of the two envelopes; subtracting the calculated mean values from the normalized quadratic modulation signal to obtain a new signal; performing decomposition on each obtained new signal to obtain second-order components, third-order components, …, n-order components, thereby realizing the acquisition of the noise components of the quadratic modulation signal, wherein the decomposition is terminated in the case that the residual signal becomes a monotonic function and cannot further decompose into an oscillation component or becomes a constant; selecting the sum of the third-order components and the residual term as the signal pre-processing result. 4.The method of claim 1, wherein, The method for extracting a spatio-temporal feature sample from the pre-processed sample quadratic modulation signal based on signal reconstruction and phase space statistical analysis permutation entropy comprises the following steps: performing phase parameter transformation on the pre-processed sample quadratic modulation signal to construct a phase space, which satisfies: wherein, is the pre-processed signal data, is a phase parameter of the signal, is the output of the phase transform, , is a real and imaginary part operation function for the data; reconstructing the phase space from the time series data using the time delay embedding method according to the structure of the phase parameter transformation: wherein, is the output reconstruction result, is the signal data after phase transformation; is the length of the preprocessed signal, is the scale of data reconstruction, represents the output reconstruction result data sequence number, represents the signal data sequence number for phase transformation, and the reconstructed signal vector is obtained according to the scale size ; setting a quantization interval according to the reconstructed signal vector, assigning the same discrete value to the reconstructed signals in the same interval to obtain sequence symbols, and counting the number of occurrences of each different sequence symbol; using the extracted phase space statistical feature vector result in the reconstructed signal as the extracted spatio-temporal feature sample. 5.The joint model deep learning based secondary modulation signal recognition method of claim 4, wherein, The jointly constructed data local space characteristic perception model and time sequence backtracking control model comprise a data local space characteristic perception module, a time sequence backtracking control module and a joint classification calculation module, the joint model is used for processing the time characteristics and the space characteristics of the extracted secondary modulation signal respectively by using the data local space characteristic perception model and the time sequence backtracking control model, and fusing the classification results of the two models according to the joint classification calculation module. 6.The joint model deep learning based secondary modulation signal recognition method of claim 1, wherein, Training the joint model by using the extracted space-time characteristic samples further comprises: After the data local space characteristic perception model and the time sequence backtracking control model are calculated, the outputs of the two models are fused according to the joint classification calculation module, the logarithmic probability of the negative expectation of the model output and the sample data is calculated, and the loss function is optimized to train the model. 7.The joint model deep learning based secondary modulation signal recognition method of claim 6, wherein, The secondary modulation signal classification weights of the two models based on the optimal correlation information criterion optimization satisfy: wherein, is a model classification weight, are weight values of the data local spatial characteristic perception model and the time series backtracking control model, respectively, is a correlation information function, and represent classification probabilities calculated by the data local spatial characteristic perception model and the time series backtracking control model, respectively, wherein and represent any two different data sequence numbers in the model, . 8.The joint model deep learning based secondary modulation signal recognition method of claim 7, wherein, The fusion of the modulation mode recognition results of the secondary modulation signal according to the classification weights to output the recognition results satisfies: wherein, is a joint classification vector, are classification result vectors of the data local space characteristic perception model and the time series backtracking control model, respectively. 9.A system for identifying a secondary modulated signal based on a joint model deep learning, the system comprising: The computer program is stored on the memory and is executed by the processor to realize the steps of the secondary modulation signal recognition method based on the joint model deep learning according to any one of claims 1 to 8.

Images