KR-102961973-B1 - Optical receiving apparatus and method having complex distortion compensation function based on extended instruction set for neural network operations

KR102961973B1KR 102961973 B1KR102961973 B1KR 102961973B1KR-102961973-B1

Abstract

The present invention relates to an optical receiver device and method equipped with a complex distortion compensation function based on extended instructions for neural network computation, wherein complex non-linear distortions occurring in the semiconductor optical amplifier (SOA), photodiode (PD), and trans-impedance amplifier (TIA) of the optical receiver in 50Gbps Passive Optical Network (PON) communication are compensated based on a neural network, and high-speed neural network computations can be performed directly at the hardware level by applying an extended instruction set for neural network computation to a RISC-V processor. By integrally compensating for complex non-linear distortions and inter-symbol interference (ISI) occurring in the SOA, PD, and TIA based on a neural network, the invention has the effect of performing more accurate distortion correction with less computational load compared to the existing Volterra series non-linear equalizer (VNLE), and applies an extended instruction set for neural network computation to a RISC-V processor. By applying the Extended ISA, high-speed neural network operations can be performed directly at the hardware level, thereby enabling the processing of major operations within a single cycle by directly driving the NPU hardware through extended instructions instead of a general software loop.

Inventors

서인식
박성훈
백준현

Assignees

(주)자람테크놀로지

Dates

Publication Date: 20260508
Application Date: 20260119

Claims (10)

SOA (Semiconductor Optical Amplifier) that amplifies the received optical signal; A PD (Photo Detector) that converts the optical signal amplified in the above SOA into an electrical signal; A TIA (Trans-Impedance Amplifier) that converts and amplifies the current signal output from the above PD into a voltage signal; An Analog-to-Digital Converter (ADC) that converts an analog voltage signal output from the above TIA into a digital signal; and It includes a distortion compensation unit that compensates for complex distortions generated in the SOA, PD, and TIA for the digital signal output from the above ADC, and The above distortion compensation unit is, An extended RISC-V processor in which an extended instruction set for neural network operations is added to a basic RISC-V instruction set, identifies the extended instructions, executes the neural network operations corresponding to the extended instructions in a hardware configuration, and reflects the results in a RISC-V pipeline stage; An NPU compensation unit (Neural Processing Unit Compensation Unit) that configures a neural network model using the above-mentioned extended RISC-V processor and performs training and distortion compensation operations for the model; and An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network computation, characterized by including an FEC (Forward Error Correction) tuning unit that performs FEC restoration on a compensation signal output from the above NPU compensation unit and adjusts the operation parameters of the above SOA, PD, and TIA based on the FEC restoration result.
In claim 1, The neural network model configured in the above NPU compensation unit is one or more of neural network models including RNN (Recurrent Neural Network), GRU (Gated Recurrent Unit), LSTM (Long Short-Term Memory), ESN (Echo State Network), Bi-RNN, Bi-LSTM, 1D-CNN, Transformer, TCN (Temporal Convolutional Network), LNN (Linear Neural Network), or variant models based thereon, and An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network operations, characterized in that the extended instruction set configured in the extended RISC-V processor is for performing neural network operations used in one or more neural network models configured in the NPU compensation unit.
In claim 2, An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network computation, characterized in that the extended instruction set of the RISC-V processor, the hardware configuration for executing the extended instruction set, and the neural network model of the NPU compensation unit can be changed by external control.
In claim 1, The above extended instruction set is, A matrix multiplication-accumulation instruction that performs the multiplication and accumulation operations of an input vector and a weight matrix in a single clock cycle; and Includes an activation function instruction that performs the activation function operation of a neural network in a single clock cycle, The above activation function command is, An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network operations, characterized by supporting at least one of a hyperbolic tangent (tanh) activation function, a sigmoid activation function, and a ReLU (Rectified Linear Unit) activation function.
In claim 1, The above extended RISC-V processor is It includes an NPU ISA section that includes extension instructions for neural network operations and hardware configuration for executing extension instructions, and An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network computation, characterized by determining whether a fetched instruction is a basic RISC-V instruction or an extended instruction, transmitting the instruction determined to be an extended instruction to the NPU ISA unit, and injecting the computation result of the NPU ISA unit into the pipeline stage of the extended RISC-V processor.
In claim 1, The above FEC tuning unit is, An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network computation, characterized by sequentially adjusting the gain of the SOA, the bias voltage of the PD, and the operating current of the TIA according to the FEC recovery result of the received signal.
In claim 1, The neural network model of the above NPU compensation unit is, An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network computation, characterized by being pre-learned for complex distortions occurring in SOA, PD, and TIA using test optical signals at the factory after manufacturing of the optical receiver.
In claim 4, An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network computation, characterized by increasing computational efficiency by using a preset lookup table when the input value is within a preset range and directly outputting ±1 by utilizing the characteristic that the tanh function converges to ±1 when the input value is outside the preset range.
A step in which an optical receiving device amplifies a received optical signal in a semiconductor optical amplifier (SOA), converts it into an electrical signal in a photodiode (PD), amplifies it in a preamplifier (TIA), and then converts it into a digital signal in an analog-to-digital converter (ADC); A distortion compensation unit of an optical receiver comprises the step of configuring an extended RISC-V processor that performs hardware operations for neural network operations when an extended instruction for neural network operations is patched, and configuring one or more neural network models that learn the distortion of a received signal using the extended RISC-V processor and compensate for the distortion of the received signal with the learned content; A step of configuring a learning model for compensating for complex distortion caused by the pattern effect of the SOA and the non-linear characteristics of the PD and TIA by training the one or more neural network models through a learning optical signal; The above distortion compensation unit compensates for the complex distortion of the digital signal output from the ADC based on one selected from the above one or more neural network models in an actual communication environment and checks the FEC restoration state; and An optical receiver equipped with a complex distortion compensation function based on extended instructions for neural network computation, comprising the step of fine-tuning at least one of the gain of the SOA, the bias voltage of the PD, and the control current of the TIA when the above FEC restoration state is below a reference.
In claim 9, The neural network model configured in the distortion compensation unit above is one or more of neural network models including RNN (Recurrent Neural Network), GRU (Gated Recurrent Unit), LSTM (Long Short-Term Memory), ESN (Echo State Network), Bi-RNN, Bi-LSTM, 1D-CNN, Transformer, TCN (Temporal Convolutional Network), LNN (Linear Neural Network), or variant models based thereon, and An optical receiver equipped with a complex distortion compensation function based on an extension instruction for neural network operations, characterized in that the extension instruction configured in the above-mentioned extension RISC-V processor is for performing neural network operations used in one or more neural network models configured in the above-mentioned distortion compensation unit.

Description

Optical receiving apparatus and method having complex distortion compensation function based on extended instruction set for neural network operations The present invention relates to an optical receiving device and method equipped with a complex distortion compensation function based on extended instructions for neural network operations. More specifically, the invention relates to an optical receiving device and method equipped with a complex distortion compensation function based on extended instructions for neural network operations, which compensates for complex non-linear distortion occurring in the semiconductor optical amplifier (SOA), photodiode (PD), and trans-impedance amplifier (TIA) of the optical receiving unit in 50Gbps Passive Optical Network (PON) communication based on a neural network, and enables high-speed neural network operations to be performed directly at the hardware level by applying an extended instruction set for neural network operations to a RISC-V processor. With the advancement of optical communication technology, 50G-PON is currently attracting attention as a next-generation Passive Optical Network (PON) technology. 50G-PON is a technology capable of responding to the rapidly increasing demand for data traffic by providing more than five times the bandwidth compared to existing 10G-EPON or XGS-PON. In a typical Passive Optical Network (PON) configuration, a single Optical Line Terminal (OLT) installed at the telephone exchange and multiple subscribers' Optical Network Terminals (ONTs) or Optical Network Units (ONUs) form a point-to-multipoint network structure via optical splitters (remote nodes), which are passive optical branching devices. Since the ONT and ONU have essentially the same configuration, they will be referred to as ONTs below. In this PON configuration, an OLT equipped with an optical transceiver that mutually converts electrical and optical signals is connected to multiple subscriber ONTs through an optical splitter, and each ONT is also configured with an optical transceiver. Through this configuration, high-speed communication services can be provided to multiple subscriber ONTs. Downstream signals (DS) transmitted from the OLT to the ONT are sent to all ONTs, and since the ONTs receive all downstream signals and use the information they need to receive, continuous signal transmission is possible. Upstream signals (US) transmitted from the ONT to the OLT are performed during the upstream signal transmission interval allocated to each ONT; therefore, in PON communication, the OLT proceeds with the procedure of registering the ONTs and controlling the communication. In particular, for uplink signals transmitted from the ONT to the OLT, the distances between the multiple ONTs and OLTs are irregular, and deviations occur between the internal clocks of the ONTs; therefore, the OLT must recover the clock from the received signal whenever it receives an uplink signal in burst mode from an individual ONT. In other words, compared to downlink transmission, uplink transmission is relatively more complex and difficult, and is significantly affected by communication quality. In the case of such 50Gbps PON communication, it is specified that a standard single-mode optical line (SMF: Single Mode Fiber, e.g., ITU-T G.652 standard optical cable) is used without applying any separate modification or compensation configuration for dispersion, even though the received signal must be transmitted at a high speed of 50Gbps using NRZ (Non-Return-to-Zero) modulation. In particular, considering that it is difficult to recover uplink signals received in burst mode for uplink signal transmission, it is stipulated that a communication speed of 25 Gbps be supported and that a wavelength band of 1300 ± 10 nm be used, which is less affected by dispersion. Therefore, for downlink signal transmission that must support a transmission speed of 50 Gbps, a wavelength of 1342 ± 2 nm is used, but there is a limitation in that the transmission quality is lowered and the transmission distance is drastically reduced due to the significant impact of dispersion. For example, if an ITU-T G.652 standard optical cable is used as a single-mode optical fiber (SMF), and the dispersion at a wavelength of 1300 nm is 2 ps/nm·km, it means that a delay of 2 ps occurs due to dispersion when transmitting 1 km at that wavelength. The allowable error between pulses of an Electro Absorptive Modulation (EAM) laser signal, which provides a commonly used transmission speed of 25 Gbps, is approximately 128 ps, and in this case, transmission up to 64 km is possible. If the transmission speed is set to 50 Gbps, the allowable error between signal pulses is approximately 32 ps, so in this case, the transmission distance is reduced to about 16 km. Meanwhile, deviations in the received signal BER values occur due to various factors such as dispersion, communication speed, signal pattern, wavelength, attenuation, optical sys