A Method and Device for Identifying Modulation Signal Formats Based on Reservoir Computing

Abstract

The present invention provides a method and apparatus for identifying a modulation signal format based on reservoir computing. The method includes: obtaining an input signal S(t) through an input layer of a reservoir; passing the input signal S(t) through a reservoir layer of the reservoir to obtain a state result, where the reservoir layer includes a vertical cavity surface emitting laser (VCSEL) and a delay feedback loop; training an output layer of the reservoir according to the state result to determine the weights and output result of the output layer; using the residual between a predicted value and a target value, and the state result for gradient boosting to perform modulation signal format identification, where the predicted value includes the product of the output result and the weight value. In this embodiment, only the weights of the output layer are trained, reducing the complexity of modulation recognition processing and improving the implementation efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of modulation recognition, and in particular, to a method and device for identifying modulation signal formats based on reservoir computing. Background Art

[0002] In a non-cooperative communication system, modulation recognition is a crucial step between signal detection and information demodulation. Whether in military or civilian communication fields, it is very important to efficiently and accurately identify the modulation methods of various signals.

[0003] In the prior art, modulation recognition methods generally rely on feature extraction. First, significant differential feature parameters are extracted from the signal to be recognized, such as the instantaneous amplitude of the signal, the frequency distribution histogram of the signal, higher-order statistics and higher-order spectra, constellation diagrams and other feature parameters as the classification features of the signal. Then, a suitable classification and recognition method is selected according to the feature parameters of the modulation signal. Finally, different types of modulation format signals are distinguished using these different features. Usually, neural networks are used for classification and recognition, which perform well and can make full use of the characteristics of the feature parameters.

[0004] However, the traditional modulation signal recognition method based on neural networks has problems such as large computational complexity, high resource consumption, and dependence on parallel acceleration computing devices, resulting in high complexity in the processing of modulation recognition and low implementation efficiency. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention provides a method and device for identifying modulation signal formats based on reservoir computing, aiming to reduce the complexity of modulation recognition processing and improve the implementation efficiency.

[0006] The present invention is implemented through the following technical solutions:

[0007] On the one hand, an embodiment of the present invention provides a method for identifying modulation signal formats based on reservoir computing, including:

[0008] Obtain an input signal S(t) through the input layer of the reservoir, where the input signal S(t) includes a signal determined according to the signal modulation format;

[0009] Pass the input signal S(t) through the reservoir layer of the reservoir to obtain a state result, where the reservoir layer includes a vertical cavity surface emitting laser (VCSEL) and a delay feedback loop;

[0010] Train the output layer of the reservoir according to the state result to determine the weights and output results of the output layer;

[0011] Gradient boosting is performed using the residual between the predicted value and the target value, as well as the state result, for modulation signal format recognition. The predicted value includes the value obtained by multiplying the output result and the weight value.

[0012] Further, training the output layer of the reservoir according to the state result to determine the weights and output result of the output layer includes:

[0013] Determine the training target value;

[0014] Determine the training weights according to the training target value, the VCSEL state matrix, and the ridge coefficient;

[0015] Multiply the result obtained by passing the input signal S(t) through the reservoir layer by the training weights, and determine the obtained result as the output result.

[0016] Further, the gradient boosting using the residual between the predicted value and the target value, as well as the state result, for modulation signal format recognition includes:

[0017] Subtract the first predicted value from the target value in the output result to determine the first residual. The output result includes at least the first predicted value, the second predicted value, the third predicted value, and the fourth predicted value;

[0018] Train the first residual according to the state result, add the training result to the predicted value, and then compare it with the target value again to obtain the second residual;

[0019] Use the second predicted value in the output result as the predicted value, perform the operation of subtracting the predicted value from the target value to determine the first residual, and perform the training of the first residual according to the output result, add the training result to the first predicted value, and then compare it with the target value again to obtain the second residual. Repeat this three times, and the output result is the result of modulation signal format recognition.

[0020] Further, before obtaining the input signal S(t), it further includes:

[0021] Obtain a modulation signal, and the signal modulation format includes at least: 2ASK, 2FSK, 2PSK;

[0022] Perform Hilbert transform on the modulation signal to obtain the complex signal information of the signal;

[0023] Pass the complex signal information through feature extraction in sequence to obtain an input sequence. The feature extraction includes at least: the maximum value γ of the spectral density of the zero-centered normalized instantaneous amplitude max and the standard deviation σ of the non-weak signal segment instantaneous phase non-linear component of the zero-centereddp The standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa ;

[0024] Preprocess the input sequence in the input layer of the input sequence to obtain the input signal S(t).

[0025] Furthermore, the preprocessing includes multiplying the input sequence by a binary random mask signal with a sampling period T.

[0026] On the other hand, an embodiment of the present invention provides a modulation signal format recognition device based on reservoir computing, including:

[0027] An acquisition module, configured to acquire an input signal S(t) through the input layer of the reservoir, where the input signal S(t) includes a signal determined according to a signal modulation format;

[0028] A processing module, configured to pass the input signal S(t) through the reservoir layer of the reservoir to obtain a state result, where the reservoir layer includes a vertical cavity surface emitting laser (VCSEL) and a delay feedback loop;

[0029] A training module, configured to train the output layer of the reservoir according to the state result to determine the weights and output results of the output layer;

[0030] An identification module, configured to perform gradient boosting using the residual between the predicted value and the target value, and the state result to perform modulation signal format identification, where the predicted value includes the value obtained by multiplying the output result and the weight value.

[0031] Furthermore, the training module is specifically configured to determine a training target value; determine training weights according to the training target value, the VCSEL state matrix, and the ridge coefficient; multiply the result obtained by passing the input signal S(t) through the reservoir layer by the training weights, and determine the obtained result as the output result.

[0032] Furthermore, the identification module is configured to subtract the first predicted value from the target value in the output result to determine a first residual, where the output result includes at least a first predicted value, a second predicted value, a third predicted value, and a fourth predicted value; train the first residual according to the state result, add the training result to the predicted value, and then compare it with the target value again to obtain a second residual; use the second predicted value in the output result as the predicted value, perform the operation of subtracting the predicted value from the target value to determine a first residual, and perform the training of the first residual according to the output result, add the training result to the first predicted value, and then compare it with the target value again to obtain a second residual. After repeating three times, the output result is the result of modulation signal format identification.

[0033] Further, the obtaining module is further configured to obtain a modulation signal, and the signal modulation format at least includes: 2ASK, 2FSK, 2PSK; perform Hilbert transform on the modulation signal to obtain complex signal information of the signal; sequentially perform feature extraction on the complex signal information to obtain an input sequence, and the feature extraction at least includes: the maximum value γ of the spectral density of the zero-centered normalized instantaneous amplitude max , the standard deviation σ of the non-linear component of the instantaneous phase of the zero-centered non-weak signal segment dp , the standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa ; preprocess the input sequence in the input layer to obtain the input signal S(t).

[0034] Further, the obtaining module is further configured to multiply the input sequence by a binary random mask signal through a sampling period T.

[0035] Compared with the prior art, the present invention has the following beneficial technical effects:

[0036] The modulation signal format recognition method based on reservoir computing provided by the embodiment of the present invention obtains an input signal S(t) through the input layer of the reservoir; passes the input signal S(t) through the reservoir layer of the reservoir to obtain a state result, and the reservoir layer includes: a vertical cavity surface emitting laser VCSEL and a delay feedback loop; trains the output layer of the reservoir according to the state result to determine the weights and output results of the output layer; uses the residual between the predicted value and the target value, and the state result to perform gradient boosting to perform modulation signal format recognition, and the predicted value includes the value obtained by multiplying the output result and the weight value.

[0037] In this embodiment, only the weights of the output layer are trained, reducing the complexity of modulation recognition processing and improving the implementation efficiency. Description of the Drawings

[0038] Figure 1 is a schematic flowchart of a modulation signal format recognition method based on reservoir computing according to an embodiment of the present invention;

[0039] Figure 2 is a feature sequence extracted from 2ASK, 2PSK, and 2FSK signals according to an embodiment of the present invention;

[0040] Figure 3 is a schematic diagram of a reservoir based on VCSEL according to an embodiment of the present invention (the line in front of VCSEL in the figure should be a double arrow);

[0041] Figure 4Schematic diagram of the feature sequence and the corresponding target value sequence generated in an embodiment of the present invention;

[0042] Figure 5 Schematic diagram of the structure of a modulation signal format recognition device based on reservoir computing in an embodiment of the present invention. Detailed implementation manners

[0043] The present invention will be further described in detail below with reference to specific embodiments, which are explanations of the present invention rather than limitations.

[0044] As Figure 1 shown, an embodiment of the present invention provides a modulation signal format recognition method based on reservoir computing, including:

[0045] Step 101: Obtain an input signal S(t) through the input layer of the reservoir.

[0046] Specifically, the reservoir mainly includes three layers: an input layer, a reservoir layer, and an output layer. Among them, the reservoir layer includes a large number of sparsely connected non-linear nodes. The input signal first passes through the input layer and then enters each node in the reservoir layer. The input connection weights between the input layer and the reservoir layer and the connection weights inside the reservoir layer are randomly generated before the start of training and remain fixed during the training process. Therefore, reservoir computing only needs to train the output weights between the output layers, and the computational complexity of the whole process is greatly reduced, and good training results are obtained. This reservoir is built based on a VCSEL-based photonic reservoir.

[0047] The input signal S(t) in this embodiment includes signals determined according to the signal modulation format. For example, a modulation signal is obtained, and the modulation signal at least includes: 2ASK, 2FSK, 2PSK; the modulation signal is subjected to Hilbert transform to obtain complex signal information; the complex signal information is sequentially subjected to feature extraction to obtain an input sequence, and the feature extraction at least includes: the maximum value γ of the spectral density of the zero-centered normalized instantaneous amplitude max , the standard deviation σ of the non-linear component of the zero-centered instantaneous phase in the non-weak signal segment dp , the standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa ; the input sequence is preprocessed in the input layer to obtain the input signal S(t).

[0048] Step 102: Pass the input signal S(t) through the reservoir layer of the reservoir to obtain a state result.

[0049] The reservoir layer in this embodiment includes: a vertical cavity surface emitting laser (VCSEL) and a delay feedback loop;

[0050] Specifically, the reservoir layer consists of a VCSEL and a delay feedback loop. The input signal S(t) is injected into the VCSEL with delay feedback. Due to the existence of the delay loop, the input signal generates a non-linear transient under the condition of having generated an input response. The implementation process of this transient simulation can be simulated by the Spin-Flip model.

[0051] Step 103: Train the output layer of the reservoir according to the state result to determine the weights and output result of the output layer.

[0052] Specifically, determine the training target value; determine the training weights according to the training target value, the VCSEL state matrix, and the ridge coefficient; determine the calculated loss according to the training weights, the training sample target value, and the set parameter λ.

[0053] Step 104: Use the residual between the predicted value and the target value, and the state result to perform gradient boosting for modulation signal format recognition.

[0054] In this embodiment, the predicted value includes the value obtained by multiplying the output result and the weight value.

[0055] Specifically, subtract the first predicted value in the output result from the target value to determine the first residual. The output result includes at least the first predicted value, the second predicted value, the third predicted value, and the fourth predicted value; train the first residual according to the state result, add the training result to the predicted value, and then compare it with the target value again to obtain the second residual; use the second predicted value in the output result as the predicted value, perform the operation of subtracting the predicted value from the target value to determine the first residual, and perform the training of the first residual according to the output result, add the training result to the first predicted value, and then compare it with the target value again to obtain the second residual. Repeat this three times, and the output result is the result of modulation signal format recognition.

[0056] In this embodiment, the input signal S(t) is obtained through the input layer of the reservoir; the input signal S(t) passes through the reservoir layer of the reservoir to obtain the state result. The reservoir layer includes a vertical cavity surface emitting laser (VCSEL) and a delay feedback loop; the output layer of the reservoir is trained according to the state result to determine the weights and output result of the output layer; the residual between the predicted value and the target value, and the state result are used to perform gradient boosting for modulation signal format recognition. The predicted value includes the value obtained by multiplying the output result and the weight value. In this embodiment, only the weights of the output layer are trained, reducing the complexity of modulation recognition processing and improving the implementation efficiency.

[0057] An embodiment of the present invention is a method for identifying modulation signal formats based on reservoir computing (RC), which identifies three conventional digital modulation signals (2ASK, 2FSK, 2PSK). Reservoir computing is a special recurrent neural network (RNN). As a novel information processing method, it only needs to train the weights of the output layer, which can greatly reduce the training cost and computing cost. At present, the nonlinear characteristics of a single vertical cavity surface emitting laser (VCSEL) can replace a large number of previous nonlinear nodes to complete the mapping from low dimension to high dimension of the input signal, and the VCSEL has many advantages such as small footprint, easy integration, and low power consumption.

[0058] Further, on the basis of the above-mentioned embodiment of the present invention, before step 101, it further includes:

[0059] Step 1001: Obtain a modulation signal, where the modulation signal at least includes: 2ASK, 2FSK, 2PSK;

[0060] Specifically, first generate a modulation signal, set the carrier frequency f c = 1e4Hz, and the sampling rate f s = 1e5Hz modulation signal. There are mainly three types of this modulation signal: 2ASK, 2FSK, 2PSK; then, perform feature extraction. The features to be extracted are: the maximum value γ max of the spectral density of the zero-centered normalized instantaneous amplitude, the standard deviation σ dp of the instantaneous phase non-linear component in the zero-centered non-weak signal segment, and the standard deviation σ aa of the absolute value of the zero-centered normalized instantaneous amplitude.

[0061] Step 1002: Pass the modulation signal through the Hilbert transform to obtain the complex signal information of the signal;

[0062] Step 1003: Pass the complex signal information through feature extraction in sequence to obtain an input sequence.

[0063] As Figure 2 shown, the embodiment of the present invention provides a feature sequence composed of 2ASK, 2PSK, and 2FSK signals. The feature extraction at least includes: the maximum value γ max of the spectral density of the zero-centered normalized instantaneous amplitude, the standard deviation σ dp of the instantaneous phase non-linear component in the zero-centered non-weak signal segment, and the standard deviation σ aa of the absolute value of the zero-centered normalized instantaneous amplitude;

[0064] For example, the maximum value γ max of the spectral density of the zero-centered normalized instantaneous amplitude = max|FFT[a cn(i)] 2 | / N s , where and m a is the instantaneous frequency of the signal, a cn (i) is the average value of the instantaneous amplitudes of the i signals, N s is the number of sampling points of the digital signal, and FFT(·) represents the Fourier transform.

[0065] Standard deviation of the instantaneous phase non-linear component of the zero-centered non-weak signal segment:

[0066] where

[0067] φ NL [n] represents the non-linear component in the instantaneous phase, in, φ NL [n] is calculated by subtracting the linear weight from the expanded phase φ[n]; f c is the carrier frequency, f s is the sampling frequency;

[0068] where the expanded phase φ[n] is obtained by adding the correction sequence C[n] to the signal instantaneous phase sequence , that is:

[0069] where the calculation formula of the correction sequence C[n] is:

[0070]

[0071] Standard deviation of the absolute value of the zero-centered normalized instantaneous amplitude:

[0072]

[0073] Next, an input sequence is formed in turn according to the order of the above three features, and the input sequence can be subjected to maximum-minimum normalization processing.

[0074] Step 1004. Preprocess the input sequence in the input layer to obtain the input signal S(t).

[0075] In this embodiment, the preprocessing includes: multiplying the input sequence after maximum-minimum normalization processing by the sampling period T and the binary random mask signal.

[0076] As Figure 3 shown. The experimental setup of the time-delay RC system scheme based on a single VCSEL is as Figure 3As shown in the figure. The training process can be completed through the following steps. First, a commercial 1550nm VCSEL without isolation (SEOUL VIOSYS) is regarded as a non-linear node. The VCSEL is driven by a low-noise laser diode controller (ILX-Lightwave, LDC-3724C). Second, the external light generated by a tunable laser (TL) with isolation is sent into a Mach-Zehnder modulator (MZM), and then modulated with a shielded input signal S(t). The shielded input signal is generated by an arbitrary waveform generator (AWG, 70002A). Third, the modulated optical signal passes through fiber couplers FC2 and FC1, through a polarization controller (PC2), an EDFA (erbium-doped fiber amplifier) and a variable optical attenuator (VOA) and is injected into the VCSEL. Then, the output of the VCSEL is first connected to 10 / 90 fiber couplers (FC1) and FC2, and then, after passing through a polarization controller (PC1) and a fiber feedback mirror with a reflectivity of 0.8 (mirror), it is fed back to the VCSEL through the 90% port of FC2. This path is the delay feedback loop. A power meter (PM) is connected to the 10% output of FC1 to measure the optical power injected into the VCSEL. Finally, in order to measure the output of the RC system, first, the output light is detected by a photodetector (PD, HP 11982A), and finally, it is displayed on an oscilloscope (OSC, Agilent DSOV334A). In addition, the output light of the VCSEL is sent to a spectrum analyzer (OSA, Q8384).

[0077] Furthermore, on the basis of the above embodiments, step 103 can be specifically implemented in the following manner:

[0078] Step 113: Determine the training target value;

[0079] Specifically, obtain a modulation signal. There are mainly 3 types of modulation formats for this signal: 2ASK, 2FSK, and 2PSK. Then, perform One-Hot encoding on the modulation signal to determine the training target value, where the label [1, 0, 0] represents the 2ASK signal, the label [0, 1, 0] represents the 2FSK signal, and the label [0, 0, 1] represents the 2PSK signal.

[0080] Obtain the result after the feature sequence input undergoes non-linear high-dimensional mapping through the VCSEL on the oscilloscope, and then use this data to train the output after One-Hot encoding. The schematic diagram of its input and output is as Figure 4 shown.

[0081] Step 123: Train the reservoir with the extracted feature training dataset, and train with the target value defined in step 113 as the target to determine the weight w;

[0082] The method for solving the weight w is ridge regression:

[0083] w = (X T X + λI) -1 X T y

[0084] where λ is the ridge coefficient, I is the identity matrix, X represents the state matrix generated after the VCSEL laser is stimulated by injection, y represents the target value, and T represents the transpose of the matrix.

[0085] Step 133: Calculate the loss function:

[0086] The loss function is used to evaluate the degree of difference between the predicted value and the true value of the model. When the difference between the predicted value and the actual value is larger, the value of the loss function, which is the sum of the total sample errors, is larger.

[0087] Loss function:

[0088] where is called the L2 regularization term, m is the training sample parameter, h θ (x i ) is the training sample target value, y i is the predicted value, n is the number of features, θ is the weight w value mentioned in step 123, and the parameter λ is the set parameter, that is, the regularization coefficient or penalty coefficient, which needs to be continuously adjusted.

[0089] Step 134: Save the weights of the current training cycle, multiply the output of the test set after passing through the reservoir by the training weights to obtain the result Y pre and then compare it with the test target Y target :

[0090]

[0091] Specifically, the comparison result corresponds to the recognition accuracy after this training (the predicted value refers to the actual calculated result value). If the predicted value is equal to the target value, it proves that the recognition is successful.

[0092] Furthermore, on the basis of the above embodiments, the gradient is boosted using the residual between the predicted value and the target value, and the state result to perform modulation signal format recognition. It can be understood as using the residual between the predicted value and the target value and the high-dimensional mapping sequence obtained after VCSEL delay for gradient boosting (GBDT).

[0093] Specific steps:

[0094] 4.1: Obtain the residual E between the first predicted value and the target value for the test set 1 = Y pre - Y target .

[0095] 4.2: Retrain the residual value using the previously generated high-dimensional mapping sequence. After obtaining the training result, add it to the first predicted value and then compare it with the target value again:

[0096]

[0097] 4.3: Then calculate the residual value after step 4.1 again and continuously repeat step 4.2:

[0098]

[0099] 4.3: In this experimental scheme, we repeated the operation 4 times.

[0100] This embodiment has the advantages of fast calculation speed and high calculation efficiency. This patent conducts high-dimensional mapping and calculation of information based on the non-linear characteristics of VCSEL. Compared with the large number of weight iteration inference calculations in traditional artificial neural network algorithms during the calculation process, the calculation principle of ridge regression algorithm is relatively simple, so it has a faster calculation speed. And using the hardware VCSEL as a carrier realizes a recognition effect similar to other neural networks, and VCSEL has the characteristics of low power consumption and small occupied area, which provides a basis and feasibility for future large-scale hardware integration and application.

[0101] As Figure 5 shown, an embodiment of the present invention provides a modulation signal format recognition device based on reservoir computing, including: an acquisition module 51, a processing module 52, a training module 53, and a recognition module 54, wherein,

[0102] The acquisition module 51 is used to obtain the input signal S(t) through the input layer of the reservoir, and the input signal S(t) includes a signal determined according to the signal modulation format;

[0103] The processing module 52 is used to pass the input signal S(t) through the reservoir layer of the reservoir to obtain a state result, and the reservoir layer includes: a vertical cavity surface emitting laser VCSEL and a delay feedback loop;

[0104] The training module 53 is used to train the output layer of the reservoir according to the state result to determine the weights and output results of the output layer;

[0105] The recognition module 54 is used to perform gradient boosting using the residual between the predicted value and the target value, and the state result to perform modulation signal format recognition, and the predicted value includes the value obtained by multiplying the output result and the weight value.

[0106] The working principle and technical effects of a modulation signal format recognition device based on reservoir computing provided by an embodiment of the present invention are similar to those of the above method, and will not be elaborated here.

[0107] Based on the above embodiment, the training module 53 is specifically configured to determine a training target value; determine a training weight according to the training target value, the VCSEL state matrix, and the ridge coefficient; multiply the result obtained by multiplying the input signal S(t) by the training weight after passing through the reservoir layer, and determine the obtained result as the output result.

[0108] Further, based on the above embodiment, the recognition module 54 is configured to subtract the first predicted value in the output result from the target value to determine a first residual. The output result at least includes a first predicted value, a second predicted value, a third predicted value, and a fourth predicted value; train the first residual according to the state result, add the training result to the predicted value, and then compare it with the target value again to obtain a second residual; use the second predicted value in the output result as the predicted value, perform the operation of subtracting the predicted value from the target value to determine a first residual, and perform the training of the first residual according to the output result, add the training result to the first predicted value, and then compare it with the target value again to obtain a second residual. After repeating three times, the output result is the result of modulation signal format recognition.

[0109] Further, based on the above embodiment, the acquisition module 51 is further configured to acquire a modulation signal, and the signal modulation format at least includes: 2ASK, 2FSK, 2PSK; obtain complex signal information of the signal through Hilbert transform; sequentially perform feature extraction on the complex signal information to obtain an input sequence. The feature extraction at least includes: the maximum value γ of the spectral density of the zero-centered normalized instantaneous amplitude max , the standard deviation σ of the non-weak signal segment instantaneous phase non-linear component of the zero-centered dp , the standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa ; preprocess the input sequence in the input layer to obtain the input signal S(t).

[0110] Further, based on the above embodiment, the acquisition module 51 is further configured to multiply the input sequence by a binary random mask signal through a sampling period T.

[0111] The working principle and technical effects of a modulation signal format recognition device based on reservoir computing provided by an embodiment of the present invention are similar to those of the above method, and will not be elaborated here.

Claims

1. A method for identifying modulation signal formats based on reservoir computing, characterized in that: Comprising: An input layer through a reservoir to obtain an input signal S(t), where the input signal S(t) includes a signal determined according to a signal modulation format; Pass the input signal S(t) through the reservoir layer of the reservoir to obtain a state result, where the reservoir layer includes: a vertical cavity surface emitting laser (VCSEL) and a delay feedback loop; According to the state result, train the output layer of the reservoir to determine the weights and output results of the output layer; Subtract the first predicted value in the output result from the target value to determine the first residual, where the output result at least includes the first predicted value, the second predicted value, the third predicted value, and the fourth predicted value; According to the state result, train the first residual, add the training result to the predicted value, and then compare it with the target value again to obtain the second residual; Use the second predicted value in the output result as the predicted value, perform the operation of subtracting the predicted value from the target value to determine the first residual, and perform the training of the first residual according to the state result, add the training result to the first predicted value, and then compare it with the target value again to obtain the second residual. After repeating three times, the output result is the result of modulating signal format recognition; the predicted value includes the value obtained by multiplying the output result by the weight value.

2. The method according to claim 1, wherein: The training of the output layer of the reservoir according to the state result to determine the weights and output results of the output layer includes: Determine the training target value; According to the training target value, the VCSEL state matrix, and the ridge coefficient, determine the training weights; Multiply the result obtained by passing the input signal S(t) through the reservoir layer by the training weights to determine the output result.

3. The method according to claim 1 or 2, characterized in that: Before obtaining the input signal S(t), it further includes: Obtain a modulation signal, where the signal modulation format at least includes: 2ASK, 2FSK, 2PSK; Perform a Hilbert transform on the modulation signal to obtain the complex signal information of the signal; The complex signal information is sequentially subjected to feature extraction to obtain an input sequence, and the feature extraction at least includes: the maximum value γ of the spectral density of the zero-centered normalized instantaneous amplitude max , the standard deviation σ of the non-weak signal segment instantaneous phase non-linear component of the zero center dp , the standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa ; Preprocess the input sequence in the input layer to obtain the input signal S(t).

4. The method according to claim 3, wherein: The preprocessing includes: multiplying the input sequence by a binary random mask signal through a sampling period T.

5. A modulation signal format recognition device based on reservoir computing, characterized in that: Comprising: An acquisition module for obtaining an input signal S(t) through the input layer of the reservoir, where the input signal S(t) includes a signal determined according to a signal modulation format; A processing module for passing the input signal S(t) through the reservoir layer of the reservoir to obtain a state result, where the reservoir layer includes: a vertical cavity surface emitting laser (VCSEL) and a delay feedback loop; A training module for training the output layer of the reservoir according to the state result to determine the weights and output results of the output layer; The recognition module is used to subtract the first predicted value from the target value in the output result to determine the first residual. The output result at least includes the first predicted value, the second predicted value, the third predicted value, and the fourth predicted value. According to the state result, train the first residual, add the training result to the predicted value, and then compare it with the target value again to obtain the second residual. Use the second predicted value in the output result as the predicted value, perform the operation of subtracting the predicted value from the target value to determine the first residual, and perform the training of the first residual according to the state result, add the training result to the first predicted value, and then compare it with the target value again to obtain the second residual. After repeating three times, the output result is the result of modulation signal format recognition. The predicted value includes the value obtained by multiplying the output result by the weight.

6. The device according to claim 5, characterized in that: The training module is specifically used to determine the training target value; determine the training weight according to the training target value, the VCSEL state matrix, and the ridge coefficient; multiply the result obtained by passing the input signal S(t) through the reservoir layer by the training weight, and determine the obtained result as the output result.

7. The device according to claim 5 or 6, characterized in that: The obtaining module is further configured to obtain a modulation signal, where the signal modulation format at least includes: 2ASK, 2FSK, 2PSK; perform a Hilbert transform on the modulation signal to obtain complex signal information of the signal; sequentially perform feature extraction on the complex signal information to obtain an input sequence, where the feature extraction at least includes: the maximum value γ of the spectral density of the zero-centered normalized instantaneous amplitude max , the standard deviation σ of the non-linear component of the instantaneous phase in the zero-centered non-weak signal segment dp , the standard deviation σ of the absolute value of the zero-centered normalized instantaneous amplitude aa ; preprocess the input sequence in the input layer to obtain the input signal S(t).

8. The device according to claim 7, wherein: The acquisition module is further used to multiply the input sequence by the binary random mask signal through the sampling period T.

Images