Noise shaping method for passive optical network

Abstract

The invention discloses a noise shaping method for a passive optical network, and belongs to the field of optical fiber communication. The implementation method of the invention is as follows: a tap coefficient of a noise shaping filter is generated through an encoder of an automatic encoder and is applied to noise shaping processing before signal transmission. At a sending end, a randomly generated symbol sequence is converted into a bit sequence, the bit sequence is mapped to a corresponding constellation point, and a modulation signal is obtained after modulation; the modulation signal redistributes quantization noise through a noise shaping filter, and then the modulation signal is sent to a channel; and at a receiving end, after a received signal is demodulated, the received signal is restored into a log-likelihood ratio form through a decoder of the automatic encoder and is used for calculating a loss function and updating a neural network weight. And continuously updating the generated noise shaping filter in the iterative training process of the automatic encoder, and training to obtain a noise shaping filter suitable for a given transmission channel. And realizing noise shaping of the passive optical network according to the noise shaping filtering. The method has the advantages of low calculation complexity, low processing delay and high channel adaptability.

Description

Technical Field

[0001] The invention relates to a noise shaping method for a passive optical network and belongs to the field of optical fiber communication. Background Art

[0002] With the rapid development of emerging services such as social media, cloud computing, and high-definition television, the bandwidth demand of access networks has increased dramatically, which has put forward higher requirements on the capacity and bandwidth of passive optical networks (PONs). In order to support the high capacity of the network, high-resolution digital-to-analog converters (DACs) are widely used in passive optical networks because they can effectively reduce the quantization noise generated during the quantization process. However, compared with low-resolution DACs, high-resolution DACs significantly increase the cost and power consumption of the system, which brings a huge burden to cost-sensitive passive optical networks. In order to solve the quantization problem and promote the use of low-resolution DACs, quantization noise suppression technology is used to redistribute quantization noise between signal bands and unused bands to reduce the impact of quantization noise. Quantization noise suppression technologies include Delta-sigma modulators (DSMs), digital resolution enhancers (DREs), and noise shaping (NS) technology, among which noise shaping technology distributes most of the quantization noise outside the signal band through a feedback high-pass filter, thereby reducing the in-band quantization noise.

[0003] At present, most quantization noise suppression technologies can effectively suppress in-band quantization noise and promote the use of low-resolution DACs, but these technologies usually have high computational complexity and processing delay, and cannot dynamically adjust the residual quantization noise in the band according to the channel environment, which is not ideal for cost-sensitive passive optical networks. Therefore, a method with low complexity that can automatically adjust the residual quantization noise according to the channel environment is urgently needed. Summary of the invention

[0004] Aiming at the problem of significantly increased system cost and power consumption when passive optical networks use high-resolution DAC to reduce quantization noise, the main purpose of the present invention is to provide a noise shaping method for passive optical networks, which performs noise shaping on signals based on an automatic encoder and dynamically adjusts the residual quantization noise in the signal frequency band according to the channel frequency response to reduce the impact of quantization noise on the signal. The present invention has the advantages of low computational complexity and processing delay and strong channel adaptability.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] The present invention discloses a noise shaping method for a passive optical network, which generates tap coefficients of a noise shaping filter through an encoder of an automatic encoder and is applied to noise shaping processing before signal transmission. At the transmitting end, a randomly generated symbol sequence is converted into a bit sequence and mapped to a corresponding constellation point, and a modulated signal is obtained after modulation; the modulated signal redistributes the quantization noise through a noise shaping filter and is then sent to the channel; at the receiving end, the received signal is demodulated and restored to the form of a log-likelihood ratio through a decoder of an automatic encoder, which is used to calculate a loss function and update a neural network weight. The generated noise shaping filter is continuously updated during the iterative training process of the automatic encoder, and a noise shaping filter suitable for a given transmission channel is obtained through training. Noise shaping of a passive optical network is achieved based on noise shaping filtering.

[0007] The present invention discloses a noise shaping method for a passive optical network, comprising the following steps:

[0008] Step 1: Build an autoencoder, which consists of two neural networks, called encoder and decoder;

[0009] The encoder is a four-layer neural network: the input signal of the first layer is a fixed one-dimensional vector t, the second and third layers are hidden layers with 2*ntap nodes, where ntap is the number of taps of the noise shaping filter; the fourth layer is the output layer, and the number of nodes is ntap; Linear connections are used between encoder layers, and ReLU is used as the activation function; the output of the fourth layer node is normalized to obtain the tap coefficient h of the noise shaping filter, and the first value of h is fixed to 1;

[0010] The decoder is a four-layer neural network: the first layer consists of two nodes, and the input of these two nodes is the demodulated received signal; the second and third layers are both hidden layers with 2m nodes, where m is the modulation order; the fourth layer is the output layer, and the number of nodes is log2 m; Linear connections are also used between the layers of the decoder; the activation function of the second and third layers is the ReLU function, and the activation function of the fourth layer is the Sigmoid function;

[0011] Step 2: The encoder generates tap coefficients of the noise shaping filter;

[0012] A fixed one-dimensional vector t is input into the encoder, and after passing through the neural network, the tap coefficient h of the noise shaping filter is obtained, and h is used for noise shaping processing before signal transmission;

[0013] Step 3: Generate a symbol sequence randomly, convert it into a modulated signal, perform noise shaping, and obtain a shaped signal;

[0014] The randomly generated symbol sequence s is converted into a bit sequence b, and the modulated signal x is obtained through constellation point mapping and modulation. The modulated signal x is passed through a quantizer and a noise shaping filter to obtain a shaped signal y. The modulated signal x and the shaped signal y satisfy equation (1):

[0015] Y(e jω )=X(e jω )+(1+G(e jω ))N(e jω ) (1)

[0016] Among them, X(e jω )、Y(e jω ) are the discrete Fourier transforms of the modulation signal x and the shaping signal y, respectively, N(e jω ) represents the quantization noise, 1+G(e jω ) is the discrete Fourier transform of the noise shaping filter, satisfying equation (2):

[0017]

[0018] Where h is the tap coefficient of the noise shaping filter; the quantization noise is redistributed across the entire frequency band through the noise shaping filter;

[0019] Step 4: Transmit the shaped signal y obtained in step 2 through the channel to obtain a received signal y';

[0020] Step 5: Demodulate the received signal y', and then the decoder restores the signal to the form of log-likelihood ratio to obtain the decoder output b';

[0021] Step 6: Calculate the loss function and back-propagate to update the neural network weights;

[0022] The decoder output b' obtained in step 5 is compared with the bit sequence b in step 3, and the bit cross entropy with regularization term is used as the loss function Loss of the neural network for gradient backpropagation to update the neural network weights, as shown in formula (3):

[0023]

[0024] Among them, L BCE is the bit cross entropy loss function, D KL is the Kullback-Leibler (KL) divergence between two posterior probabilities, which is used to measure the difference between two distributions of the same variable; is the true posterior distribution, is the mapping defined for the demodulator; I(Y,Y') is the generalized mutual information GMI of the channel, Y and Y' are the input and output before and after the channel respectively; H(s) is the entropy, which is a fixed constant here; λ is a hyperparameter that controls the size of the regularization term, is a regularization term used to control the residual quantization noise within the signal frequency band, H(e jω ) is the discrete Fourier transform of the noise shaping filter tap coefficient h, Ω s is the cutoff frequency of the signal band;

[0025] Step 7: Train the autoencoder;

[0026] Gradient back propagation, repeating steps 2 to 6, and training the neural network constructed in step 1 according to the Loss obtained in step 5; in the process of continuous iterative training, the Loss of the neural network gradually converges, and the bit mutual information tends to the maximum. When the preset number of iterations is reached, the autoencoder training is completed, and a trained autoencoder is obtained;

[0027] Step 8: According to the autoencoder trained in step 7, the tap coefficient h of the noise shaping filter output by the encoder is obtained to obtain the optimal noise shaping filter under a given channel environment. The obtained noise shaping filter is used for the corresponding channel to reduce the residual quantization noise in the signal frequency band and realize noise shaping of the passive optical network.

[0028] Beneficial effects:

[0029] 1. A noise shaping method for a passive optical network disclosed in the present invention adopts a noise shaping filter to reduce in-band quantization noise, which has lower computational complexity and processing delay compared with traditional quantization noise suppression technology.

[0030] 2. The present invention discloses a noise shaping method for a passive optical network, which uses an automatic encoder to generate tap coefficients of a noise shaping filter, thereby obtaining the optimal noise shaping filter under current channel conditions and maximizing the reduction of in-band residual quantization noise after channel transmission.

[0031] 3. The noise shaping method for a passive optical network disclosed by the present invention adopts a modular approach. The channel model in the automatic encoder is replaced according to different channel conditions. Only the channel parameters need to be changed to achieve the optimal noise shaping filter debugging for different channel environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of a noise shaping method for a passive optical network disclosed in the present invention;

[0033] Figure 2 Schematic diagram of the structure of the automatic encoder constructed in the embodiment;

[0034] Figure 3 It is a system block diagram of the noise shaping processing part in the embodiment;

[0035] Figure 4 1 is the frequency response of the noise shaping filter generated under different transmission distances and different transmission powers in the embodiment; wherein (a)-(c) are the frequency responses of the noise shaping filter generated under the channel environment with a transmission distance of 10 km and transmission powers of -2dBm, 0dBm, and 2dBm respectively; (d)-(f) are the frequency responses of the noise shaping filter generated under the channel environment with a transmission distance of 20 km and transmission powers of -2dBm, 0dBm, and 2dBm respectively;

[0036] Figure 5 This is a comparison chart of bit error rates of two noise shaping schemes at a transmission distance of 10 km and different transmission powers in the embodiment;

[0037] Figure 6 This is a comparison chart of bit error rates of two noise shaping schemes at a transmission distance of 20 km and different transmission powers in the embodiment. DETAILED DESCRIPTION

[0038] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. The technical problems solved by the technical solution of the present invention and the beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate the understanding of the present invention and do not have any limiting effect on it.

[0039] This embodiment is a process of generating an optical signal in a passive optical network transmission link with a transmission distance of 10km and 20km. The modulation format of the transmitted optical signal is 32-order carrier-free phase amplitude modulation (CAP32), and the transmission power is -3 to 2dBm. A noise shaping method for a passive optical network disclosed in this embodiment is used to perform noise shaping on the modulated signal, such as Figure 1 As shown, the specific implementation steps are as follows:

[0040] Step 1: Build an autoencoder;

[0041] In the embodiment, Figure 2 As shown in the figure, the autoencoder consists of an encoder and a decoder. The encoder is a four-layer neural network: the input signal of the first layer is a fixed one-dimensional vector t, the second and third layers are hidden layers with 2*11 nodes, where 11 is the number of taps of the noise shaping filter; the fourth layer is the output layer, with 11 nodes; Linear connections are used between encoder layers, and ReLU is used as the activation function; the output of the fourth layer node is normalized to obtain the tap coefficient h of the noise shaping filter, and the first value of h is fixed to 1;

[0042] The decoder is a four-layer neural network: the first layer consists of two nodes, and the input of these two nodes is the demodulated received signal; the second and third layers are hidden layers with 64 nodes; the fourth layer is the output layer, with 5 nodes; Linear connections are also used between the layers of the decoder; the activation function of the second and third layers is the ReLU function, and the activation function of the fourth layer is the Sigmoid function;

[0043] Step 2: The encoder generates tap coefficients of the noise shaping filter;

[0044] A fixed one-dimensional vector t is input into the encoder, and after passing through the neural network, the tap coefficient h of the noise shaping filter is obtained, and h is used for noise shaping processing before signal transmission;

[0045] Step 3: Generate a random symbol sequence, convert it into a modulated signal, and perform noise shaping;

[0046] The numpy.rand library in Python is used to randomly generate a symbol sequence s, which is converted into a bit sequence b. The modulated signal x is then obtained through 32QAM constellation point mapping and carrier-free phase amplitude (CAP) modulation. The modulated signal x is passed through a quantizer and a noise shaping filter to obtain a shaped signal y. The modulated signal x and the shaped signal y satisfy equation (1):

[0047] Y(e jω )=X(e jω )+(1+G(e jω ))N(e jω ) (4)

[0048] Among them, X(e jω )、Y(e jω ) are the discrete Fourier transforms of the modulation signal x and the shaping signal y, respectively, N(e jω ) represents the quantization noise, 1+G(e jω ) is the discrete Fourier transform of the noise shaping filter, satisfying equation (2):

[0049]

[0050] Where h is the tap coefficient of the noise shaping filter. The noise shaping filter can redistribute the quantization noise across the frequency band.

[0051] Step 4: Transmit the shaped signal y obtained in step 2 through the channel to obtain a received signal y'; in the embodiment, a generative adversarial network (CGAN) is used to model the channel and to simulate the real channel to transmit the shaped signal y to ensure that the gradient can be back-propagated.

[0052] Step 5: Perform CAP demodulation on the received signal y', and then the decoder restores the signal to the form of log-likelihood ratio to obtain the decoder output b';

[0053] Step 6: Calculate the loss function and back-propagate to update the neural network weights;

[0054] The decoder output b' obtained in step 5 is compared with the bit sequence b in step 3, and the bit cross entropy with regularization term is used as the loss function Loss of the neural network for gradient backpropagation to update the neural network weights, as shown in formula (3):

[0055]

[0056] Among them, L BCE is the bit cross entropy loss function, D KL is the Kullback-Leibler (KL) divergence between the two posterior probabilities; is the true posterior distribution, is the mapping defined for the demodulator; I(Y,Y') is the generalized mutual information (GMI) of the channel, Y and Y' are the input and output before and after the channel respectively; H(s) is the source entropy; λ is a hyperparameter that controls the size of the regularization term, is the regularization term, H(e jω ) is the discrete Fourier transform of the noise shaping filter tap coefficient h, Ω s is the cutoff frequency of the signal band.

[0057] Step 7: Train the autoencoder;

[0058] Gradient back propagation, repeating steps 2 to 6, and training the neural network constructed in step 1 according to the Loss obtained in step 5; in the process of continuous iterative training, the Loss of the neural network gradually converges, and the bit mutual information tends to the maximum. When the preset number of iterations reaches 2000, the autoencoder training is completed;

[0059] Step 8: According to the automatic encoder trained in step 7, the tap coefficient h of the noise shaping filter output by the encoder is obtained to obtain the optimal noise shaping filter under the given channel environment, and its frequency response is as follows: Figure 4 shown. Figure 4 (a)-(c) are the frequency responses of the noise shaping filter generated by training under the channel environment with a transmission distance of 10 km and a transmission power of -2dBm, 0dBm, and 2dBm respectively; Figure 4 (d)-(f) are the frequency responses of the noise shaping filter generated under the channel environment with a transmission distance of 20 km and a transmission power of -2 dBm, 0 dBm, and 2 dBm respectively;

[0060] In the embodiment, Figure 5 and Figure 6 As shown, using the passive optical network channel noise shaping method, the noise shaping scheme obtained has achieved significant improvement in the bit error rate at a transmission distance of 10km and 20km and a transmission power of -3 to 2dBm, which is reduced by up to 45% compared with traditional noise shaping technology, effectively reducing the impact of in-band residual quantization noise on the signal.

[0061] The specific description above further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A noise shaping method for a passive optical network, characterized in that: The following steps are included: Step 1: construct an autoencoder; the autoencoder consists of two neural networks, namely an encoder and a decoder; Step 2: The encoder generates tap coefficients of the noise shaping filter; A fixed one-dimensional vector t is input into the encoder, and after passing through the neural network, the tap coefficient h of the noise shaping filter is obtained, and h is used for noise shaping processing before signal transmission; Step 3: randomly generate a symbol sequence, convert it into a modulated signal, perform noise shaping, and obtain a shaped signal; The randomly generated symbol sequence s is converted into a bit sequence b, and the modulated signal x is obtained through constellation point mapping and modulation. The modulated signal x is passed through a quantizer and a noise shaping filter to obtain a shaped signal y. The modulated signal x and the shaped signal y satisfy equation (1): Y(e jω )=X(e jω )+(1+G(e jω ))N(e jω ) (1) Among them, X(e jω )、Y(e jω ) are the discrete Fourier transforms of the modulation signal x and the shaping signal y, respectively, N(e jω ) represents the quantization noise, 1+G(e jω ) is the discrete Fourier transform of the noise shaping filter, satisfying equation (2): Where h is the tap coefficient of the noise shaping filter; the quantization noise is redistributed across the entire frequency band through the noise shaping filter; Step 4: Transmit the shaped signal y obtained in step 2 through the channel to obtain a received signal y'; Step 5: demodulate the received signal y', and then the decoder restores the signal to the form of log-likelihood ratio to obtain the decoder output b'; Step 6: Calculate the loss function and back-propagate to update the neural network weights; The decoder output b' obtained in step 5 is compared with the bit sequence b in step 3, and the bit cross entropy with regularization term is used as the loss function Loss of the neural network for gradient backpropagation to update the neural network weights, as shown in formula (3): Among them, L BCE is the bit cross entropy loss function, D KL is the Kullback-Leibler (KL) divergence between two posterior probabilities, which is used to measure the difference between two distributions of the same variable; is the true posterior distribution, is the mapping defined for the demodulator; I(Y,Y') is the generalized mutual information GMI of the channel, Y and Y' are the input and output before and after the channel respectively; H(s) is the entropy, which is a fixed constant; λ is a hyperparameter that controls the size of the regularization term, is a regularization term used to control the residual quantization noise within the signal frequency band, H(e jω ) is the discrete Fourier transform of the noise shaping filter tap coefficient h, Ω s is the cutoff frequency of the signal band; Step 7: Train the autoencoder. Gradient back propagation, repeating steps 2 to 6, and training the neural network constructed in step 1 according to the Loss obtained in step 5; in the process of continuous iterative training, the Loss of the neural network gradually converges, and the bit mutual information tends to the maximum. When the preset number of iterations is reached, the autoencoder training is completed, and a trained autoencoder is obtained; Step 8: According to the autoencoder trained in step 7, the tap coefficient h of the noise shaping filter output by the encoder is obtained to obtain the optimal noise shaping filter under a given channel environment. The obtained noise shaping filter is used for the corresponding channel to reduce the residual quantization noise in the signal frequency band and realize noise shaping of the passive optical network.

2. A noise shaping method for a passive optical network as claimed in claim 1, characterized in that: The encoder described in step 1 is a four-layer neural network: the input signal of the first layer is a fixed one-dimensional vector t, the second and third layers are hidden layers with 2*ntap nodes, where ntap is the number of taps of the noise shaping filter; the fourth layer is the output layer, and the number of nodes is ntap; Linear connections are used between encoder layers, and ReLU is used as the activation function; the output of the fourth layer node is normalized to obtain the tap coefficient h of the noise shaping filter, and the first value of h is fixed to 1.

3. A noise shaping method for a passive optical network as claimed in claim 1, characterized in that: The decoder described in step 1 is a four-layer neural network: the first layer is composed of two nodes, and the input of the two nodes is the demodulated received signal; the second and third layers are hidden layers of 2m nodes, where m is the modulation order; the fourth layer is the output layer, and the number of nodes is log2 m; Linear connections are also used between the layers of the decoder; the activation function of the second and third layers uses the ReLU function, and the activation function of the fourth layer uses the Sigmoid function.

Images