CN113517932B - Optical convolution signal processing system and method based on cascade modulator - Google Patents

Abstract

The invention provides an optical convolution signal processing system and method based on a cascade modulator, which can utilize the cascade Mach-Zehnder modulator to finish multiplication operation among signals in an optical domain, realize time shift operation by means of a dimmable time delay line among the modulators, realize photoelectric conversion by utilizing a photoelectric detector, and finally accumulate signal sampling results to obtain convolution results. Compared with a convolution computing system based on a microelectronic circuit, the method has the advantages of high computing speed, low power consumption and the like; compared with a convolution calculation system based on an integrated optical sub-line, the method can comprise convolution calculation between signals with larger data volume; compared with a convolution computing system based on the wavelength division multiplexing technology, the invention has simple structure and strong operability. The invention utilizes the advantage of ultra-high bandwidth of optical signal processing, reduces the workload of a digital signal processing part, and has the advantages of high speed, large bandwidth and low delay.

Description

A system and method for optical convolution signal processing based on cascaded modulators

technical field

The invention belongs to the technical field of optical signal processing and optical computing, and relates to an optical convolution signal processing system and method based on cascaded modulators.

Background technique

Convolution computation is one of the most important operations in the field of signal processing. In speech recognition, image processing, geological exploration, ultrasonic diagnosis and other technical fields, convolution computing has a very wide range of applications. However, in most applications, the convolution calculation is performed after sampling the signal and using digital signal processing techniques. At present, as the feature size of microelectronic integrated circuits has been very close to the physical limit, its performance cannot continue to grow according to Moore's Law. Therefore, the computing system using electrons as the information transmission and processing carrier has many bottlenecks in terms of operation speed and power consumption.

Optical computing is an optical computing technology that uses the physical properties of light to process large-capacity information. Considering that the physical properties such as parallelism and high speed of light can be utilized, people try to use photonic devices to construct high-performance signal processing systems that can break through the performance bottleneck of traditional microelectronics technology. In recent years, many applied studies on optical computing have been reported. Among them, the most eye-catching research results include integrated Mach-Zehnder interferometers (Y.Shen, et al, "Deeplearning with coherent nanophotonic circuits" Nature Photonics, vol. 11 (2017): 441-446) and microrings Resonators (J. Feldmann, et al, "All-optical spiking neurosynaptic networks with self-learning capabilities" Nature, 569(2019): 208-214) and other photonic device arrays constructed vector-matrix multiplication optical computing structures. Although, it is also feasible to perform optical convolution calculations based on such integrated photonic devices. However, due to the shortcomings of low integration and large feature size of photonic integrated devices, when the above-mentioned hardware system is used to build a convolution computing framework, the amount of data for convolution computing will be limited by the scale of its hardware system. In addition, another method to realize optical convolution calculation based on non-integrated scheme (X. Xu, et al, "11TOPS photonicconvolutional accelerator for optical neural networks." Nature, 586(2021):44-51) is based on wavelength division According to the principle of multiplexing, the same signal is loaded on lasers of different wavelengths, and a wavelength-dependent fixed time delay is generated by using a dispersive fiber. The above optical convolution calculation scheme has the advantages of fast operation speed and large data processing capacity. However, the complexity of this scheme is relatively high, the system is relatively large, and the implementation cost is relatively expensive. Therefore, how to use new technical means to complete the convolution calculation of a large amount of data and make the convolution calculation system simpler and more compact has become an urgent problem to be solved in the field of optical computing and optical signal processing.

Therefore, there is an urgent need for an optical convolution computing system and a corresponding computing method, which can overcome the problem of a large amount of computing convolution and reduce the limitation of the scale of the hardware system.

SUMMARY OF THE INVENTION

In view of this, the present invention provides an optical convolution signal processing system and method based on cascaded modulators, which can complete the multiplication operation between signals in the optical domain by means of the Mach-Zehnder modulators in the cascade structure, and The time shift operation is realized by means of the adjustable optical time delay line between the modulators, the photoelectric conversion is realized by the photodetector, and finally the convolution result is obtained by accumulating the signal sampling results.

For achieving the above object, technical scheme of the present invention is:

An optical convolution signal processing system based on cascaded modulators, comprising an arbitrary waveform generator, a laser, a first-order modulator, a second-order modulator, a first polarization controller, a second polarization controller, a dimmable light Time delay line, photodetector PD and oscilloscope.

The arbitrary waveform generator is a dual-channel waveform generator, one channel is unidirectionally connected to the first-level modulator, and the other channel is unidirectionally connected to the second-level modulator.

The first-stage modulator is unidirectionally connected to the dimmable time delay line, the dimmable time-delay line is unidirectionally connected to the second-stage modulator, the second-stage modulator is unidirectionally connected to the photodetector PD, and the photodetector PD is unidirectionally connected One-way connection with the oscilloscope; one-way connection between the laser and the first-level modulator; the first polarization controller is between the laser and the first-level modulator, and the second polarization controller is between the adjustable optical delay line and the second-level modulation between the devices.

The first-order modulator and the second-order modulator are identical, and the first polarization controller and the second polarization controller are identical.

An optical convolution signal processing method based on cascaded modulators adopts the following specific steps to process the optical convolution signal:

S1. The laser emits laser light, and the first polarization controller and the second polarization controller are used to adjust the polarization direction of the laser, so that the first-order modulator and the second-order modulator are respectively adjusted to the orthogonal bias point, and then through any The waveform generator generates the first signal f(t) and the second signal g(-t) at the same time, and the first signal and the second signal are respectively sent to the second-stage modulator and the first-stage modulator; t is time.

S2. The second signal g(-t) is modulated into an optical signal through the first-stage modulator, and the adjustable delay line performs a time-domain translation operation on the second signal g(-t), and the time-translated second signal g(ΔT-t) is transmitted to the second-stage modulator for processing; ΔT is the delay size of the adjustable light delay line.

S3. The second-stage modulator processes the received first signal f(t), and generates a product result of the first signal f(t) and the time-domain shifted second signal g(ΔT-t). The resulting optical signal is input to the photodetector PD.

S4. In the photodetector PD, the optical signal with the product result is converted into a radio frequency signal, the radio frequency signal is sent to the oscilloscope, and the time domain waveform data of the radio frequency signal is collected by the oscilloscope, and the time domain waveform data is integrated and processed After multiplying by the known coefficients, the convolution calculation result is obtained.

Further, the first signal f(t) and the second signal g(-t) are two independent signals of arbitrary frequency and amplitude, and do not interfere with each other.

Further, the time-domain waveform data is integrated and then multiplied by a known coefficient to obtain a convolution calculation result. The specific method is as follows:

Among them, v _RF1 is the RF signal output by the first-stage modulator; c _RF2 is the RF signal output by the second-stage modulator; V _π represents the half-wave voltage of the first-stage modulator and the second-stage modulator; T represents the sampling duration; V _t0 (t) represents the voltage of the initial RF signal collected by the oscilloscope when v _RF1 and v _RF2 have not met in the time domain; V _out (t) is the voltage of the RF signal collected by the oscilloscope; G is the The gain of the photodetector PD; α is the optical path loss between the laser and the photodetector PD; _Im is the light intensity of the laser generated by the laser; * is the convolution symbol.

Beneficial effects:

1. The present invention provides an optical convolution signal processing system based on cascaded modulators. Compared with the convolution computing system based on microelectronic circuits, it has the advantages of fast operation speed and low power consumption; The convolution calculation system of the line, the present invention can realize the convolution calculation between signals containing a large amount of data; compared with the convolution calculation system based on the wavelength division multiplexing technology, the system used in the present invention has a simple structure and is operable Strong sex.

2. The present invention provides an optical convolution signal processing method based on cascaded modulators. Using the above optical convolution signal processing system, the part of the product operation in the convolution calculation is completed in the optical domain, and finally only the oscilloscope needs to be connected. The result of integral calculation can be obtained by simple processing such as summation of the collected data. In principle, the multiplication part in the optical domain is only affected by the operating bandwidth of the system. Therefore, the advantage of ultra-high bandwidth in optical signal processing can be used to complete the multiplication operation that is complex for digital circuits in convolution calculations. The system reduces the workload of the digital signal processing part, and has the advantages of high speed, large bandwidth and low delay compared with the traditional electronic convolution calculation.

Description of drawings

Figure 1 is a structural diagram of an optical convolution signal processing system based on cascaded MZMs;

Figure 2 is a comparison diagram of the experimental results and theoretical results based on the Gaussian pulse verification experiment.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

In order to make the implementation and process of the example of the present invention clearer, the implementation of the example of the present invention will be described in more detail below with reference to the accompanying drawings.

As shown in FIG. 1, the present invention provides an optical convolution signal processing system, which includes: an arbitrary waveform generator ①, a laser ②, a first-order modulator ③, a second-level modulator ④, a first polarization control ⑤, second polarization controller ⑥, dimmable time delay line ⑦, photodetector PD ⑧ and oscilloscope ⑨.

The first-level modulator ③ is unidirectionally connected to the dimmable delay line ⑦, the dimmable delay line ⑦ is unidirectionally connected to the second-level modulator ④, and the second-level modulator ④ is unidirectionally connected to the photodetector PD⑧ , the photodetector PD⑧ is unidirectionally connected with the oscilloscope ⑨; the laser ② is unidirectionally connected with the first-stage modulator ③; the first polarization controller ⑤ is between the laser ② and the first-stage modulator ③, and the second polarization controller ⑥ Between the dimmable time delay line ⑦ and the second-order modulator ④. In the embodiment of the present invention, the first-level modulator ③ and the second-level modulator ④ are completely identical, and the first polarization controller ⑤ and the second polarization controller ⑥ are completely identical.

The method of the present invention can be summarized as: generating a first signal f(t) and a second signal g(-t) which has been inverted in time domain by a dual-channel arbitrary waveform generator ①. Among them, g(-t) is loaded into the optical signal via the first-stage modulator ③, and then the signal g(-t) is shifted in the time domain using the adjustable optical delay line ⑦. Next, the signal f(t) is loaded via the second-stage modulator ④ to generate the result of multiplying the two signals. Finally, the optical signal with the product of the signal g(ΔT-t) and the signal f(t) is converted into an electrical signal by the photodetector PD⑧. After data acquisition is performed by the oscilloscope⑨, the integral calculation operation is completed by digital signal processing. .

The principle of the present invention is as follows:

The normalized light intensity transfer function of a modulator (MZM) can be expressed as:

代表由直流偏置电压所带来的调制器上下两臂之间的相位差。由此可知，对于两个级联形式的调制器和加在这两个调制器之间的一个可调光时延线组成的系统，其光强的输入输出关系就可以表示为：where _vRF represents the _RF signal loaded by the modulator, Vπ represents the half-wave voltage of the modulator,

Represents the phase difference between the upper and lower arms of the modulator caused by the DC bias voltage. It can be seen from this that for a system composed of two cascaded modulators and a dimmable time delay line added between the two modulators, the input-output relationship of the light intensity can be expressed as:

和

均为

时，两级调制器的输入输出关系可以表示为：Wherein I _in represents the light intensity input to the cascaded modulator system, and I _out represents the output light intensity of the system after being modulated by the two-stage modulator. ΔT is the delay size of the dimmable time delay line ⑦. Furthermore, when both the former and latter modulators work at the quadrature transmission point, that is, when

and

both

When , the input-output relationship of the two-stage modulator can be expressed as:

When the loaded radio frequency signal v _RF is smaller than the half-wave voltage V _π of the MZM, the above formula can also be approximated as:

By analyzing the results of the above formula, it can be known that the function of the output light intensity of the second-stage modulator consists of four parts: the first constant term, the second and third modulated original signals and the third Four terms The product term produced by the overlapping portion of the two signals in the time domain.

It is easy to prove that when the delay size T _d of the delay line is adjusted, that is, when the first signal loaded is shifted in the time domain, the integral values of the first three terms of the function over the entire definition domain are unchanged. In the above formula, there is only the fourth term that changes with the delay size T _d , that is, the product term generated by the overlapping parts of the two signals. According to the definition of convolution calculation: in the process of translating the signal v _RF1 , whenever the signal v _RF1 is translated to a new position, that is, when T _d takes a new value, it must be phased with the signal v _RF2 . Multiply, and then integrate these product results to obtain the convolution calculation result. In order to obtain the integration result of the product term of the two signals in the above formula, it is only necessary to integrate the output light intensity function I _out (t) and subtract the light intensity function I _T0 (t) when the two signals have not met in the time domain. Integral result. According to the above derivation, the result s(ΔT) of the signal v _RF1 *v _RF2 can be expressed as:

When the insertion loss of the two modulators and the gain of the photodetector need to be considered, according to the previous derivation, the convolution calculation result of the two signals can be expressed as:

Among them, v _RF1 is the RF signal output by the first-stage modulator ③; v _RF2 is the RF signal output by the second-stage modulator ④; V _π represents the half-wave voltage of the first-stage modulator ③ and the second-stage modulator ④ ; T represents the sampling duration; V _t0 (t) represents the voltage of the initial RF signal collected by the oscilloscope ⑨ when v _RF1 and v _RF2 have not met in the time domain; V _out (t) is the RF signal collected by the oscilloscope ⑨ The voltage of the signal; G is the gain of the photodetector PD⑧; α is the optical path loss between the laser ② and the photodetector PD⑧; _Im is the light intensity of the laser generated by the laser ②; t is the time; * is the convolution symbol.

Describe the process of using the optical convolution signal processing system of the present invention to implement the specific signal convolution calculation below in conjunction with Fig. 1:

S1. The laser ② emits laser light, and the first polarization controller ⑤ and the second polarization controller ⑥ are used to adjust the polarization direction of the laser, so that the first-level modulator ③ and the second-level modulator ④ are respectively adjusted to the orthogonal bias After the point, the first signal f(τ) and the second signal g(-t) are simultaneously generated by the arbitrary waveform generator ①, and the first signal and the second signal are respectively sent to the second-level modulator④ and the first-level modulator ③. In the embodiment of the present invention, the first signal f(t) and the second signal g(-t) are two independent signals of arbitrary frequency and amplitude, do not interfere with each other, and both have a period of 1 MHz.

S2. The second signal g(-t) is modulated into an optical signal through the first-stage modulator ③, and the adjustable delay line ⑦ performs a time domain translation operation on the second signal g(-t), The second signal g(ΔT-t) is transmitted to the second stage modulator ④ for processing. Among them, ΔT is the delay size of the dimmable time delay line ⑦.

In this step, by adjusting the optical time delay line, the two signals are made to overlap and then separate. Each time the delay of the optical time delay line is adjusted, use an oscilloscope to collect the time domain waveform data output by the system. The peak position of the second signal g(-t) is taken as the time reference point, that is, time t=0. Record the waveform data output by the system under different time delays.

S3. The second-stage modulator ④ processes the received first signal f(t) to generate a product result of the first signal f(t) and the time-domain shifted second signal g(ΔT-t), The optical signal of the product result is input to the photodetector PD⑧.

S4. In the photodetector PD⑧, the optical signal with the product result is converted into a radio frequency signal, the radio frequency signal is sent to the oscilloscope⑨, and the time domain waveform data of the radio frequency signal is collected by the oscilloscope⑨, and the time domain waveform data is processed. Multiply by the known coefficients after the integral processing to obtain the convolution calculation result. Figure 2 shows the comparison between the experimental results and theoretical results of convolution calculation based on the above method. The time domain waveform data is integrated and then multiplied by the known coefficients to obtain the convolution calculation result. The specific method is as follows:

To sum up, the above are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (3)

1. an optical convolution signal processing method based on cascaded modulator, is characterized in that, adopts following concrete steps to carry out the processing of optical convolution signal: The method is aimed at an optical convolution signal processing system based on cascaded modulators, including an arbitrary waveform generator ①, a laser ②, a first-order modulator ③, a second-level modulator ④, a first polarization controller ⑤, a second Polarization controller ⑥, dimmable time delay line ⑦, photodetector PD ⑧ and oscilloscope ⑨; wherein, the first-level modulator ③ and the second-level modulator ④ are exactly the same, and the first polarization controller ⑤ is exactly the same as the second polarization controller ⑥; The arbitrary waveform generator ① is a dual-channel waveform generator, one channel is unidirectionally connected to the first-level modulator ③, and the other channel is unidirectionally connected to the second-level modulator ④; The first-stage modulator ③ is unidirectionally connected to the dimmable time delay line ⑦, the dimmable time delay line ⑦ is unidirectionally connected to the second-stage modulator ④, and the second-stage modulation The device ④ is unidirectionally connected with the photodetector PD⑧, and the photodetector PD⑧ is unidirectionally connected with the oscilloscope ⑨; the laser ② is unidirectionally connected with the first-order modulator ③; the first polarization The controller ⑤ is between the laser ② and the first-stage modulator ③, and the second polarization controller ⑥ is between the adjustable optical delay line ⑦ and the second-stage modulator ④; S1. The laser ② emits laser light, and the first polarization controller ⑤ and the second polarization controller ⑥ are used to adjust the polarization direction of the laser, so that the first-level modulator ③ and the second-level modulator ④ After adjusting to the quadrature bias point respectively, the first signal f(t) and the second signal g(-t) are simultaneously generated by the arbitrary waveform generator ①, and the first signal and the second signal are respectively sent to the second signal. stage modulator ④ and first stage modulator ③; t is time; S2. The second signal g(-t) is modulated into an optical signal by the first-stage modulator ③, and the adjustable optical delay line ⑦ performs a time domain translation operation on the second signal g(-t), The second signal g(ΔT-t) after time domain translation is transmitted to the second-stage modulator ④ for processing; ΔT is the delay size of the adjustable light delay line ⑦; S3. The second-stage modulator ④ processes the received first signal f(t) to generate a product result of the first signal f(t) and the time-domain shifted second signal g(ΔT-t), The optical signal of the product result is input to the photodetector PD⑧; S4. In the photodetector PD⑧, convert the optical signal with the product result into a radio frequency signal, send the radio frequency signal to the oscilloscope ⑨, and collect the time when the radio frequency signal is collected by the oscilloscope ⑨ domain waveform data, the time domain waveform data is integrated and then multiplied by a known coefficient to obtain a convolution calculation result. 2. The method according to claim 1, wherein the first signal f(t) and the second signal g(-t) are two independent signals of arbitrary frequency and amplitude, and do not interfere with each other. 3. The method according to claim 2, wherein the time-domain waveform data is integrated and multiplied by a known coefficient to obtain a convolution calculation result, and the specific method is:

Among them, v _RF1 is the radio frequency signal of g(t); v _RF2 is the radio frequency signal of f(t); V _π represents the half-wave voltage of the first-level modulator ③ and the second-level modulator ④; T represents the sampling duration; V _t0 (t) represents the voltage of the initial radio frequency signal collected by the oscilloscope ⑨ when v _RF1 and v _RF2 have not met in the time domain; V _out (t) is the radio frequency collected by the oscilloscope ⑨ The voltage of the signal; G is the gain of the photodetector PD⑧; α is the optical path loss between the laser ② and the photodetector PD⑧; I _m is the light intensity of the laser generated by the laser ②; * is Convolution symbol.

Images