A nonlinear equalization method for single-carrier photonic terahertz communication system based on bidirectional gated cyclic unit

By using bidirectional gated recurrent unit neural network for nonlinear equalization in photon terahertz communication systems, the problem of nonlinear damage in photon terahertz communication is solved, and more efficient signal equalization and lower bit error rate are achieved, improving communication quality.

CN118473874BActive Publication Date: 2024-11-01BEIJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410606113.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-15
Publication Date
2024-11-01
Estimated Expiration
2044-05-15

AI Technical Summary

Technical Problem

In photon terahertz communication systems, nonlinear damage leads to a reduced signal communication quality. The existing nonlinear equalization algorithm has high computational complexity or limited compensation effect, making it difficult to effectively equalize signal.

Method used

Using a nonlinear equalization method of neural network based on bidirectional gating cyclic unit, in the digital signal processing module at the receiving end, signals are processed through the input layer, hidden layer and output layer, and the bidirectional gating cyclic unit and full connection layer are used to perform non-signal signals. Linear equalization, improve signal quality factors and reduce bit error rate.

Benefits of technology

It realizes nonlinear equalization with lower computational complexity, improves signal quality factors, significantly improves signal convergence and bit error rate performance, and is better than traditional nonlinear equalizers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118473874B_ABST
    Figure CN118473874B_ABST
Patent Text Reader

Abstract

The present invention discloses a nonlinear equalization method for a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit, and belongs to the field of photonic terahertz communication. A nonlinear equalization module is added to a digital signal processing module at the receiving end of the photonic terahertz communication system to realize nonlinear equalization, specifically: the input layer extracts features of the real part and the imaginary part of the input signal, and then transmits them to a hidden layer constructed by a bidirectional gated cyclic unit neural network, and the front sequence and the rear sequence of the signal are processed simultaneously by the forward gated cyclic unit and the backward gated cyclic unit respectively. The output values ​​of the bidirectional gated cyclic unit are spliced ​​through a fully connected layer to form a complete gated cyclic unit output sequence. The output layer encodes the information of the fully connected layer, and maps the hidden layer state value to the dimension of the output type through a linear transformation, thereby generating an output sequence for predicting the output information of a 16QAM signal.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of photonic terahertz communication. Aiming at the nonlinear damage problem occurring in the transmission process of photonic terahertz signals, the invention proposes a nonlinear equalization algorithm for performing nonlinear compensation in a digital signal processing part at a receiving end. The invention relates to a nonlinear equalization technology, specifically a nonlinear equalization technology method for a photonic terahertz communication system based on machine learning. Background Art

[0002] With the continuous development of digital society, people's requirements for communication technology are getting higher and higher. As the latest stage of communication technology, 6G technology has the advantages of higher transmission speed, richer spectrum resources, lower latency, more connections and higher security compared to 5G technology. It can support more emerging applications and help promote social and economic development. It can meet people's needs for faster speed, lower latency and wider connections, and provide more comprehensive support for digital life. 6G will use higher frequency radio waves, such as terahertz waves, which can achieve faster speeds and larger bandwidths, which makes 6G have richer spectrum resources.

[0003] Terahertz (THz) waves refer to electromagnetic waves with a frequency range of 0.1 to 10 THz (wavelength of 0.03nm to 3nm). In an era when spectrum resources are becoming increasingly scarce, the development of the THz frequency band is critical. Terahertz wave communication has the advantages of large bandwidth, narrower beam, strong penetration and strong anti-interference ability. Compared with 5G, the transmission rate of photon terahertz can reach 100Gbps-1Tbps, which is much higher than the transmission rate of 5G. This makes photon terahertz have greater advantages in application scenarios that require high-speed transmission. The frequency band resources of terahertz waves are very rich and can support ultra-high-speed wireless communications. In contrast, the frequency band resources of 5G are relatively limited. Terahertz waves have short wavelengths and high energy, can better penetrate obstacles, and have less radiation damage to the human body. Therefore, photon terahertz has greater advantages in application scenarios that require high-security communications.

[0004] Photonic terahertz can play an important role in fields such as materials science, such as diagnosing physical and chemical reaction processes. In addition, terahertz is an internationally recognized cross-cutting frontier technology for the next generation of high-speed wireless communications, which can bring attractive opportunities for technological innovation. In contrast, the application scope of 5G is relatively narrow. The key to the future 6G technology is the "ubiquitous coverage" and "seamless connection" issues, so the use of terahertz waves as the information carrier of the next generation of broadband wireless communications has become a current research hotspot. Among them, nonlinear damage is an important factor that limits the transmission performance of high-order modulation format signals in terahertz communication systems.

[0005] In terahertz communication systems, high-speed signals will suffer a series of linear and nonlinear damages caused by device defects and optical fiber links during transmission, resulting in reduced signal communication quality. Therefore, using digital signal processing technology at the receiving end to equalize and compensate for damaged signals is currently the most common and effective signal recovery method. At present, digital signal processing technology for linear equalization is relatively mature, however, nonlinear equalization algorithms cannot effectively equalize nonlinear damage.

[0006] In order to equalize the nonlinear damage of the signal, researchers have proposed many nonlinear equalization techniques based on digital signal processing, such as digital back propagation algorithm, nonlinear equalization technology based on Volterra series, and perturbation back propagation algorithm at the transmitter. The digital back propagation algorithm uses the nonlinear Schrödinger equation and reversely infers the propagation process of the signal in the optical fiber based on the split-step Fourier algorithm. It can eliminate the influence of the nonlinear effect of the optical fiber to a certain extent and achieve accurate optical fiber nonlinear compensation. However, the computational complexity is high, and this method is based on the ideal scenario in theoretical analysis. In actual terahertz communication systems, the signal will also be affected by non-ideal damage from other aspects such as devices. Nonlinear equalization technology based on Volterra series can adaptively compensate for nonlinear damage when the link parameters are unknown, but the compensation effect is limited, and it is limited by the exponential growth of computational complexity with the series, so it is rarely used in practice. Using the perturbation back propagation algorithm at the transmitter to pre-distort the signal can effectively mitigate nonlinear damage. However, this method requires the addition of a digital signal processing chip at the transmitter, which increases the cost and complexity of the system.

[0007] At present, nonlinear equalization technology based on neural networks is a hot research direction. Neural networks have strong fitting and analysis capabilities for complex nonlinear problems and can complete complex nonlinear mapping relationships. Nonlinear equalization technology based on neural networks introduces neural networks into photonic terahertz communication. According to the received signal data, the neural network can fit the nonlinear mapping function between input and output. The trained neural network can effectively realize the nonlinear equalization of the signal. Artificial neural networks, deep neural networks, convolutional neural networks, and recurrent neural networks have all been proposed and proven to effectively compensate for the nonlinear damage of the signal.

[0008] Recurrent neural network is a recursive neural network, which is an effective way to process serialized data. Based on the advantages of memory, parameter sharing, and Turing completeness, the recurrent neural network with a bidirectional structure can use forward and backward data information when processing the current moment data. Therefore, it has excellent results when learning the nonlinear characteristics of the sequence. Long short-term memory network and gated recurrent unit, as variants of recurrent neural network, are effective ways to reduce the long-distance dependence of learning errors.

[0009] Therefore, applying machine learning technology to the nonlinear equalization problem in the field of photonic terahertz communication breaks through the limitations of traditional nonlinear equalization schemes, has lower computational complexity, higher quality factor and lower bit error rate, and is a research hotspot in the field of nonlinear equalization today. It has very important research significance and broad application prospects. Summary of the invention

[0010] In view of the shortcomings of the prior art in equalizing the nonlinear damage in terahertz communication, the present invention proposes a nonlinear equalization method for a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit, which is suitable for QAM modulation format signals of photonic terahertz communication.

[0011] The nonlinear equalization method of a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit specifically builds a photonic terahertz communication system, wherein a signal reaches a receiving end through a transmitting end, and nonlinear equalization is achieved by adding a nonlinear equalization module to a digital signal processing module at the receiving end. The signal is processed by an input layer, a hidden layer, a fully connected layer, and an output layer in the nonlinear equalization module in sequence, and finally an equalized signal is output.

[0012] The specific process of performing nonlinear equalization of signals in the nonlinear equalization module is as follows:

[0013] Step 1: After the 16QAM signal is preprocessed by the digital signal processing module at the receiving end, it is output to the input layer of the nonlinear equalization module to extract the eigenvalues ​​of the real part and the imaginary part of the 16QAM signal respectively;

[0014] In the input layer, the real and imaginary parts of the 16QAM signal are separated and equalized separately:

[0015] For the real part, the sliding window module is used to generate the eigenvalue of the real part of the current constellation symbol, specifically the real and imaginary parts of the preceding k constellation symbols of the current constellation symbol, the real and imaginary parts of the succeeding k constellation symbols, and the imaginary part of the current signal, totaling 2k+1 eigenvalues.

[0016] Similarly, for the imaginary part, take the real and imaginary parts of the preceding k constellation symbols of the current constellation symbol, the real and imaginary parts of the succeeding k constellation symbols, and the real part of the current signal, totaling 2k+1 eigenvalues.

[0017] Step 2: The extracted eigenvalues ​​of the real part and the imaginary part are output to the hidden layer respectively, and the processed signals are transmitted to the fully connected layer after being processed by the bidirectional gated recurrent unit neural network of the hidden layer.

[0018] The bidirectional gated recurrent unit neural network includes a forward gated recurrent unit and a backward gated recurrent unit;

[0019] For the real and imaginary eigenvalues, the forward gated recurrent unit processes the pre-order sequence, and the backward gated recurrent unit processes the post-order sequence, respectively.

[0020] The gated recurrent unit controls input, memory, and other information through a gating mechanism, including a reset gate that determines how new input information is combined with previous memory and an update gate that defines the current memory information saved from previous memory information.

[0021] The update gate Zt is calculated by the following formula

[0022] Zt=σ(W (z) x t +U z h t-1 )

[0023] where x t is the input vector at time index t, W (z) is the weight matrix. t-1 Save the information of the previous time t-1, which is multiplied by the weight matrix U z The two results are added together and an activation function σ is applied to compress the result between 0 and 1.

[0024] The past information passed to the current unit, or the information of the previous time and the current time passed to the next moment, is determined by the update gate.

[0025] The reset gate determines the past information that needs to be discarded. The reset gate r at time step t is calculated using the following formula t :

[0026] r t =σ(W (r) x t +U r h t-1 )

[0027] h t-1 and x t Multiply it by the weight matrix and then input it into the Sigmoid activation function to get the output value of the reset gate.

[0028] After the reset gate, the new memory information is calculated based on the reset gate information stored in the past, which can be expressed as:

[0029] h′ t =tanh(W (h) x t + rt ⊙U h h t-1 )

[0030] Among them, input x t and previous information h t-1 First, respectively with the weight matrix W (h) and U h Multiply, and then give the reset gate r t and U h h t-1 The Hamdamar product of . Since the reset gate is a vector ranging from 0 to 1, it controls how much the gate opens. Hamdamar determines whether to retain past information. Next, the sum of the Hamdamar results is loaded into the tanh function, which is the hyperbolic tangent activation function. Subsequently, the vector h t The information of the current unit is retained and passed to the next unit. In this process, an update gate is used to manage the data packet h′ in the current memory. t and the information from the previous time h t-1 network transmission path.

[0031] The final output of the gated recurrent unit is expressed as:

[0032] h t =z t ⊙h t-1 +(1-z t )⊙h′ t

[0033] where z t is the output of the update gate. t and h t-1 The Hamdamar product is the data stream transmitted from the previous moment to the current moment. The final output of the gated recurrent unit is used as the training data for the current time.

[0034] Step 3: The fully connected layer concatenates the output signals of the forward gated recurrent unit and the reverse gated recurrent unit to form a complete output sequence.

[0035] Step 4: The output layer encodes the signal output by the fully connected layer and maps the hidden layer state to the dimension of the output type through linear transformation, thereby generating the prediction result of each output sequence for predicting the output information of the 16QAM signal.

[0036] The advantages of the present invention are:

[0037] (1) The present invention proposes a nonlinear equalizer for a terahertz communication system based on a bidirectional gated cyclic unit to compensate for the nonlinear damage of a signal. Compared with traditional equalizers such as a blind equalizer, the present invention has better performance in improving the quality factor Q than traditional nonlinear equalizers, and has lower computational complexity than traditional Volterra equalizers.

[0038] (2) In the first input layer structure of the neural network, the present invention adds a sliding window module to enrich the eigenvalues, enriching the eigenvalues ​​from 2 to 2k+1, thereby improving the prediction accuracy of the neural network.

[0039] (3) The present invention adopts a hidden layer structure of a bidirectional gated recurrent unit in the second hidden layer structure of the neural network. Compared with other neural network structures, the hidden layer structure of the bidirectional gated recurrent unit can process information from the front end and the end of the signal at the same time. Such a structure can not only improve the efficiency of data processing, but also has a simple structure, which is more advantageous for processing ultra-long signal sequences. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of a photonic terahertz communication system constructed for the present invention;

[0041] Figure 2 is a detailed structural diagram of the gated recurrent unit in the present invention;

[0042] Figure 3 Schematic diagram of the photonic terahertz nonlinear equalization method based on a bidirectional gated recurrent unit neural network;

[0043] Figure 4 Comparison of constellation diagrams before and after equalization by Bi-GRU nonlinear equalizer; Figure 4 a is the constellation diagram before equalization, Figure 4 b is the constellation diagram after equalization;

[0044] Figure 5 A comparison chart of various nonlinear equalizers in terms of Q factor improvement performance.

[0045] In the figure:

[0046] ECL: external cavity laser; AWG: arbitrary waveform generator; I / Q MOD: IQ modulator; EDFA: erbium-doped fiber amplifier; PM-OC: polarization-maintaining optical coupler; LNA: low noise amplifier; SSMF: single-mode fiber; PD: photodiode. DETAILED DESCRIPTION

[0047] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0048] The invention discloses a nonlinear equalization method for a single-carrier photon terahertz communication system based on a bidirectional gated cyclic unit.

[0049] The photonic terahertz communication system is constructed by VPI and MATLAB software simulation. Figure 1 As shown, the signal transmission process is:

[0050] like Figure 1 As shown in a, first at the transmitting end, a random sequence generator is used in MATLAB to generate a random 01 sequence offline, and then the 01 sequence is mapped to 16-QAM through constellation mapping, and then serial-to-parallel conversion is performed, and then carrier modulation and digital-to-analog conversion are performed. I / Q MOD modulates ECL1, and ECL2, as a local oscillator light source, generates a free-running optical carrier, which differs from the optical carrier generated by ECL1 by one terahertz frequency. Then the two optical signals are coupled through PM-OC, and then connected to the optical fiber through an optical amplifier. After being transmitted through the optical fiber, they reach the photodetector, and the two signals are beat through the photodetector to generate a photon terahertz signal with a center frequency of 325GHz, realizing photoelectric conversion, and finally transmitted back-to-back to the receiving end.

[0051] The receiving end uses a local oscillator signal source to convert the high-frequency photon terahertz signal into an intermediate frequency signal that can be processed by the receiving end digital signal processing module under the action of a mixer, and then performs subsequent digital signal processing through an offline digital signal processing module (OfflineDSP) and MATLAB for joint simulation. Figure 1 As shown in b, in the DSP process, the signal undergoes IQ orthogonalization, clock recovery, dispersion compensation, polarization equalization, frequency offset estimation, and phase recovery, and is then equalized by a nonlinear equalizer based on Bi-GRU. The final output signal is shown in Figure 1 c. In this embodiment, Bi-GRU is built in Tensorflow2.7, the cross entropy loss function is selected as the loss function, and the Adam optimizer is used to optimize the dual GRU network. The entire dataset is divided into training data (70%) and test data (30%). The maximum training epoch is set to 300. The learning rate is set to 0.001. When the accuracy does not improve for 40 consecutive epochs, the training is terminated early to prevent overfitting.

[0052] Figure 2 Shown is the detailed structure of a gated recurrent unit. A gated recurrent unit neural network (GRU) controls input, memory, and other information through a gating mechanism. The GRU has two gates: a reset gate and an update gate. The reset gate determines how new input information is combined with previous memory. The update gate defines the current memory information that is preserved from previous memory information.

[0053] The update gate Zt is calculated by the following formula

[0054] Zt=σ(W (z) x t +U z h t-1 )

[0055] where x tis the input vector at time index t. It is linearly transformed and multiplied by the weight matrix W (z) .h t-1 Save the information of the previous time t-1, which is multiplied by the weight matrix U z The two results are added together and an activation function σ is applied to squash the result between 0 and 1. The update gate determines how much information from the past should be passed to the current unit in the model, or how much information from the previous time and the current time should be passed to the next moment.

[0056] The reset gate mainly determines how much past information needs to be discarded. The update gate r at time step t is calculated using the following formula t :

[0057] r t =σ(W (r) x t +U r h t-1 )

[0058] The formula is the same as the update gate formula, except for the parameters and weight matrix. As mentioned before, h t-1 and x t Multiply it by the weight matrix and then input it into the Sigmoid activation function to get the output value of the reset gate. After the reset gate, the new memory information is calculated based on the reset gate information stored in the past. This process can be expressed by equation.

[0059] h′ t =tanh(W (h) x t + rt ⊙U h h t-1 )

[0060] The input x t and previous information h t-1 First, respectively with the weight matrix W (h) and U h Then the reset gate r is given t and U h h t-1 The Hamdamar product of . Since the reset gate is a vector ranging from 0 to 1, it controls how much the gate opens. Hamdamar determines whether to retain past information. Next, the sum of the Hamdamar results is loaded into the tanh function, which is the hyperbolic tangent activation function. Subsequently, the vector h t Keep the information of the current cell and pass it to the next cell. In this process, an update gate is used to manage the data packet h′ in the current memory. t And last time h t-1The network transmission path. This formula can be expressed as

[0061] h t =z t ⊙h t-1 +(1-z t )⊙h t

[0062] where z t is the output of the update gate. t and h t-1 The Hamdamar product is the data stream transmitted from the previous moment to the current moment. The final output of the gated recurrent unit is given as the training data at the current time.

[0063] Figure 3 Schematic diagram of the photonic terahertz nonlinear equalization scheme based on a bidirectional gated recurrent unit neural network. The first layer is the input layer. In the input layer, for 16QAM signals, the real and imaginary parts are separated and equalized respectively. For the real part, the real eigenvalue of the current constellation symbol is generated by the sliding window module, specifically the real and imaginary parts of the k constellation symbols before the current constellation symbol, the real and imaginary parts of the k constellation symbols after the current constellation symbol, and the imaginary part of the current signal, a total of 2k+1 eigenvalues. For the imaginary part, similarly, the real and imaginary parts of the k constellation symbols before the current constellation symbol, the real and imaginary parts of the subsequent k constellation symbols, and the real part of the current signal are taken, totaling 2k+1 eigenvalues.

[0064] The second layer is a hidden layer constructed by a bidirectional gated recurrent unit neural network. After the signal processed by the input layer enters the hidden layer, the front sequence and the back sequence of the signal are processed simultaneously by the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent unit respectively.

[0065] The third layer is the fully connected layer. After the output value of the hidden layer of the bidirectional gated recurrent unit neural network enters the fully connected layer, the output of the forward gated recurrent unit and the reverse gated recurrent unit are spliced ​​through the fully connected layer to form a complete gated recurrent unit output sequence. This structure can effectively capture the contextual information of the input sequence.

[0066] The fourth layer is the output layer. The output layer converts the state of the fully connected layer into output, that is, encodes the information of the fully connected layer, performs linear transformation, maps the hidden layer state value to the dimension of the output type, and thus generates an output sequence. The prediction result generates output information that can be used to predict the 16QAM signal.

[0067] The results show

[0068] The results in the present invention are simulated by simulation software Matlab and VPIphotonics. In order to more intuitively reflect the advantages of the present invention, the 16QAM constellation diagram of the signal restored and demodulated by the receiving end without being processed by the Bi-GRU nonlinear equalizer is compared with the 16QAM constellation diagram after being equalized by the nonlinear equalizer.

[0069] like Figure 4 As shown in FIG. 1 , by comparing the 16QAM constellation diagram before and after equalization by the Bi-GRU nonlinear equalizer, it can be seen that the convergence degree of the 16QAM constellation diagram is significantly improved after equalization by the equalizer.

[0070] like Figure 5 As shown in the figure, the comparison of the effects of different nonlinear equalizers on the Q factor shows that compared with the Bi-LSTM, GRU and CMA equalizers, the Bi-GRU has better performance in improving the effect of the Q factor.

[0071] It can be seen from the above experimental comparison that the 16QAM constellation diagram processed by the nonlinear equalizer of the present invention is more concentrated and the Q value is significantly improved.

Claims

1. A nonlinear equalization method for a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit, characterized in that: Build a photonic terahertz communication system, where the signal reaches the receiving end through the transmitting end, and nonlinear equalization is achieved by adding a nonlinear equalization module to the digital signal processing module at the receiving end; The signal is processed by the input layer, hidden layer, fully connected layer and output layer in the nonlinear equalization module, and finally the equalized signal is output. The specific process is as follows: Step 1: After the 16QAM signal is preprocessed by the digital signal processing module at the receiving end, it is output to the input layer of the nonlinear equalization module to extract the eigenvalues ​​of the real part and the imaginary part of the 16QAM signal respectively; Step 2: The extracted eigenvalues ​​of the real part and the imaginary part are output to the hidden layer respectively, processed by the bidirectional gated recurrent unit neural network of the hidden layer, and the processed signals are transmitted to the fully connected layer; The gated recurrent unit includes a reset gate and an update gate. The reset gate determines how the new input information is combined with the previous memory; the update gate defines the current memory information saved from the previous memory information, and determines the past information passed to the current unit through the update gate, or the information of the previous time and the current time passed to the next moment; After the reset gate, the new memory information is calculated based on the reset gate information stored in the past, which can be expressed as: h t =tanh(W (h) x t +r t ⊙U h h t-1 ) Among them, input x t and the information from the previous time h t-1 First, respectively with the weight matrix W (h) and U h Multiply, and then give the reset gate r t and U h h t-1 The reset gate is a vector ranging from 0 to 1 that controls the size of the gate opening; Hamdamar determines whether to retain past information; Next, the sum of the Hamdamar results is loaded into the tanh function, which is the hyperbolic tangent activation function; Then, the vector h t Keep the information of the current unit and pass it to the next unit; in this process, use the update gate to manage the data packet h′ currently in memory t and the information from the previous time h t-1 The network transmission path; The final output of the gated recurrent unit is expressed as: h t =z t ⊙h t-1 +(1-z t )⊙h′ t where z t is the output of the update gate; z t and h t-1 The Hamdamar product is the data stream transmitted from the previous moment to the current moment; Step 3: The fully connected layer concatenates the output signals of the bidirectional gated recurrent unit to form a complete output sequence. Step 4: The output layer encodes the signal output by the fully connected layer and maps the hidden layer state to the dimension of the output type through linear transformation, thereby generating the prediction result of each output sequence for predicting the output information of the 16QAM signal.

2. According to claim 1, a nonlinear equalization method for a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit is characterized in that: In the input layer, the eigenvalues ​​of the real and imaginary parts of the 16QAM signal are extracted respectively, specifically: For the real part, the eigenvalue of the real part of the current constellation symbol is generated through the sliding window module, specifically the real and imaginary parts of the preceding k constellation symbols of the current constellation symbol, the real and imaginary parts of the succeeding k constellation symbols, and the imaginary part of the current signal, totaling 2k+1 eigenvalues; For the imaginary part, take the real and imaginary parts of the preceding k constellation symbols of the current constellation symbol, the real and imaginary parts of the succeeding k constellation symbols, and the real part of the current signal, totaling 2k+1 eigenvalues.

3. The nonlinear equalization method for a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit according to claim 1, characterized in that: The bidirectional gated recurrent unit neural network comprises a forward gated recurrent unit and a backward gated recurrent unit; the pre-order sequence and post-order sequence of the real and imaginary eigenvalues ​​are processed simultaneously by the forward gated recurrent unit and the backward gated recurrent unit respectively.

4. The nonlinear equalization method for a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit according to claim 1, characterized in that: The update gate Zt is calculated by the following formula: Zt=σ(W (z) x t +U z h t-1 ) where x t is the input vector at time index t, W (z) is the weight matrix; h t-1 Save the information of the previous time t-1, which is multiplied by the weight matrix U z ; Add these two results and apply the activation function σ to compress the result between 0 and 1.

5. The nonlinear equalization method of a single-carrier photonic terahertz communication system based on a bidirectional gated cyclic unit according to claim 1, characterized in that: The reset gate r t The calculation formula is: r t =σ(W (r) x t +U r h t-1 ) h t-1 and x t Multiply it by the weight matrix and then input it into the Sigmoid activation function to get the output value of the reset gate.