A method for artificial intelligence-assisted mode locking and pulse classification based on speckle patterns

Abstract

The present invention discloses a method for artificial intelligence-assisted mode locking and pulse classification based on speckle patterns. In an NPR mode-locked laser, the speckle pattern obtained by a CCD is input into a trained neural network for judgment to achieve automatic mode locking, and it is possible to judge whether it is a single pulse or multiple pulses based on the trained neural network. The present invention can achieve automatic mode locking and pulse type recognition by using the intensity and appearance of speckle patterns in industry and academia. This method completely replaces traditional spectrometers and autocorrelators for judging whether mode locking and pulse types are present. The overall network structure is simple and easy to operate, while reducing the cost of laser measurement equipment.

Description

Technical Field

[0001] The present invention relates to a method for artificial intelligence-assisted mode locking and pulse classification based on speckle patterns, and specifically to achieving the purpose of automatic mode locking according to the brightness of the speckle pattern, and classifying the speckle patterns of single pulse, multi-pulse and non-mode locking according to a CNN network, and belongs to the field of lasers. Background Art

[0002] Ultrashort optical pulses have a wide range of applications in modern science, including high-resolution atomic clocks, optical frequency measurement, precise distance measurement, signal processing, and astronomy. Mode-locking technology is one of the main methods for generating picosecond and femtosecond optical pulses and has developed rapidly. People have been striving for optical pulses with narrower pulse widths and higher peak powers. Among these technologies, passive mode-locking of NPR based on polarization control and Kerr nonlinearity has become the preferred method for researchers to realize passive mode-locked lasers because of its simple structure and superior performance. This method mainly uses Kerr nonlinearity and intensity-related nonlinear phase shift in optical fibers to make the polarization state change with the change of light intensity, thereby achieving an effective artificial saturable absorption effect by adjusting the polarization state in the cavity. In addition, through flexible polarization control, NPR mode-locked lasers can generate a variety of pulse states, including soliton molecules, multi-soliton modes, harmonic mode-locking, soliton explosions, and soliton rain.

[0003] Due to the small solution space of different polarization states, it becomes difficult to manually control the mode-locked laser to reach the desired state. In addition, mode-locked lasers have problems such as multi-pulse and Q-switched pulse (QS) instability, which further affects the stability of NPR lasers. Often in practical applications, NPR mode-locked lasers require long-term stability, but environmental factors may cause polarization state drift, which further increases the difficulty of maintaining the desired state of the laser. In order to solve these problems, researchers have proposed self-starting mode-locked lasers and have achieved certain results. However, since the polarization drift caused by environmental interference is continuous, this technology cannot completely solve the problems of NPR mode-locked lasers. Therefore, automatic mode locking (AML) technology has become a new research field for ultrafast lasers and has received widespread attention and research. Traditional stabilization technology is different from AML technology. AML technology attempts to use various algorithms and automated control of certain parameters in the cavity to solve the dilemma of NPR mode-locked lasers. In the early stage of AML technology, the most commonly used algorithm is the traversal algorithm. In AML technology, the traversal algorithm is a commonly used search strategy. The algorithm directly traverses the entire controllable parameter space and combines the judgment criteria to achieve the target pulse state. Common discrimination criteria include pulse counting and detection of certain special parameters. Although the search efficiency of AML lasers based on traversal algorithms is low in a huge parameter space, resulting in unsatisfactory time-consuming performance, the fundamental frequency mode-locked pulse state can still be automatically searched. In addition, some AML lasers adopt a real-time implementation method, using a microcontroller unit or a field programmable gate array as a computing center, and using circuits such as analog-to-digital converters or voltage comparators to obtain feedback data. Compared with offline implementation methods, AML lasers based on real-time implementation methods have huge advantages in time performance. However, these studies use large equipment (including oscilloscopes, spectrometers, and spectrometers) to obtain feedback data, which is very time-consuming and expensive. The present invention uses the intensity and appearance of the speckle pattern to achieve automatic mode locking and pulse type recognition. The overall network structure is simple and easy to operate, and it can replace traditional spectrometers and autocorrelators to determine whether the mode is locked and the pulse type, reducing the cost of laser measurement equipment. The specific method is to first adjust QWP1, HWP1, and QWP2 in sequence to minimize the speckle brightness, and then use the speckle pattern at this time through a trained neural network to determine whether it is mode locked at this time. If the mode is not locked at this time, the second step is required. Select any one of the three wave plates and fine-tune it within 10°. Input the speckle pattern generated during fine-tuning into the trained neural network for judgment until mode locking is achieved. Summary of the invention

[0004] The purpose of the present invention is to provide a method for artificial intelligence-assisted mode locking and pulse classification based on speckle patterns. The method achieves mode locking by obtaining a speckle pattern through frosted glass, and can classify pulse types using a simple one-layer convolution and pooling CNN network structure. The method has low cost, high reliability, simple algorithm and high recognition efficiency.

[0005] The technical method adopted by the present invention to achieve its inventive purpose is an artificial intelligence-assisted mode locking and pulse classification method based on speckle pattern, which includes a nonlinear polarization rotation (NPR) laser device, a speckle generating device, a method for judging the mode locking state, speckle collection of various pulse states, and the use of convolutional neural networks to classify various pulse types.

[0006] The NPR device described in the present invention includes a semiconductor pump source (Pump), a wavelength division multiplexer (WDM), a single-mode optical fiber (2.5m), a Yb-doped optical fiber (0.3m), two optical fiber collimators (Collimator), two quarter-wave plates (QWP), two half-wave plates (HWP), two polarization beam splitters (PBS), a grating (Grating), a Faraday magneto-optical rotator (FR), a 10:90 beam splitter (Coupler), etc., which are used to generate mode-locked pulses.

[0007] The scattering medium and CCD of the speckle generating device of the present invention are used to generate and detect speckles.

[0008] The speckle collection process of various pulse states described in the present invention is to generate three pulse states (single pulse, multiple pulses and unlocked state) by fine-tuning the wave plate left and right. This process requires the use of an autocorrelation instrument to determine single or multiple pulses.

[0009] The process of using a convolutional neural network to classify various pulse types in the present invention is to obtain a data volume of 6000 images by rotating and translating 2000 speckle patterns of three states (single pulse, multi-pulse and unlocked state), and then train 4000 images using a CNN network, test 1000 images using a CNN network, and verify the remaining 1000 images using a CNN network.

[0010] The method for artificial intelligence-assisted mode-locking state judgment of the present invention refers to using CCD to detect the intensity and appearance of speckle after the laser passes through the scattering medium while adjusting the angles of HWP1, QWP1 and QWP2. The specific method is to first adjust QWP1, HWP1 and QWP2 in sequence to minimize the speckle brightness, and use the trained neural network to determine whether the mode is locked at this time. If the mode is not locked at this time, the second step is required to select any one of the three wave plates for fine-tuning within 10°, and input the speckle pattern generated during fine-tuning into the trained neural network for judgment until the mode is locked, and then perform the third step, input the speckle pattern during mode locking into the trained neural network again to determine whether it is a single pulse, if so, the mode locking operation stops; if not, the fourth step is required to reduce the power and further transmit the speckle pattern obtained by the CCD to the network structure under the power for judgment, until a single pulse is finally output, and the mode locking operation is completely completed.

[0011] The principle of the NPR laser described in the present invention is that light is modulated into linear polarized light after passing through the Faraday magneto-optical rotator, and the linear polarized light becomes elliptically polarized light after passing through QWP2 (elliptically polarized light is composed of S light and P light that are perpendicular to each other and have different intensities). When the two linear polarized lights that make up the elliptically polarized light propagate in the optical fiber, different nonlinear phase shifts will be accumulated due to different intensities, which eventually causes the polarization state to rotate and the rotation at the pulse peak is the largest. Therefore, HWP and QWP1 are adjusted to make the light transmittance at the peak large, and the intensity of the two wings is absorbed, which eventually leads to the narrowing of the light pulse. This process is equivalent to the effect of saturable absorption. The spectra of the corresponding output lasers under different saturable and absorption states are different, resulting in different speckle appearances or intensities. Compared with HWP and QWP1, QWP2 has a more significant effect on the modulation depth because they play different roles at different positions. In the experiment, the effect of nonlinear polarization rotation can be enhanced or weakened by adjusting the angle of QWP2.

[0012] The method for judging the mode-locking state of the present invention is to start adjusting the wave plates in sequence when the ratio of the output power to the Pump power is above 15%, so that the speckle brightness corresponding to each wave plate is the lowest, and then fix QWP1 and QWP2 and rotate HWP1 to record the state of the pulse. We found that there are 4 mode-locking areas within 360°, and the angles differ by about 90°. Then, fix QWP1 and rotate QWP2 again to find that there are also mode-locking areas at an angle of 180°. Then, repeatedly rotate HWP1, and we again find that the mode can still be locked under the 4 mode-locking areas with basically the same angle as before.

[0013] In the present invention, HWP1, QWP1 and QWP2 are installed on an electric rotating platform in the NPR laser, and then the speckle and convolutional neural network are used to automatically rotate to achieve mode locking so as to achieve the purpose of automatic mode locking, and automatically judge the pulse type (single pulse, multi-pulse and non-mode-locked state). This method completely replaces the spectrometer and autocorrelator to judge whether the mode is locked and the pulse type, and greatly improves the efficiency of single pulse realization.

[0014] The present invention utilizes CNN network to classify various pulse types. Excluding the input and output layers, there are three layers of network. We compress the input image to a 128 pixel × 128 pixel image, and then a convolution layer is used. The size of the convolution kernel is 5*5. After the convolution layer, the image becomes a 124 pixel × 124 pixel convolution feature map containing 6 pixels. After the pooling layer, the image becomes a 62 pixel × 62 pixel pooling feature map. The pooling feature map is then flattened to 1 pixel × 23064 pixels, and finally passed to a fully connected layer to reach the classification output.

[0015] The present invention is further described in detail below through specific implementation modes and accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a diagram of the experimental device of the present invention;

[0017] Figure 2 The polarization state diagram of the pulse at different positions in the cavity of the device of the present invention;

[0018] Figure 3 The rotating HWP1 and QWP2 in the present invention present a periodic locking region;

[0019] Figure 4 The corresponding spectrum, autocorrelation and speckle pattern of the laser in the typical single pulse, multi-pulse and unlocked state of the present invention;

[0020] Figure 5 It is a flow chart of the mode locking judgment of the present invention.

[0021] Figure 6 This is the CNN network structure diagram of the present invention. DETAILED DESCRIPTION

[0022] like Figure 1 The figure shows the NPR experimental setup. The fiber section in the experiment consists of 2.5 meters of 1μm band single-mode fiber (Corning HI1060) and 0.3 meters of ytterbium-doped gain fiber (LEIKKI1200-4 / 125). The gain fiber has an absorption coefficient of 1200dB / m at 976nm and a second-order dispersion coefficient of β2 = 26.2fs 2 / mm, β2 of single-mode fiber = 24.7fs2 / mm. The spatial optical path is about 0.5 meters, and the net dispersion is about 0.07ps 2 , which belongs to full positive dispersion. The pump wavelength of the semiconductor pump source used is 976nm, and the pump light is coupled into the laser cavity through a wavelength division multiplexer. The fiber collimator transmits the light in the fiber to the spatial optical path. A 10:90 beam splitter is used to divide the optical path into two paths, of which 90% of the light enters the spatial optical path. From left to right in the spatial optical path, there are a quarter wave plate, a half wave plate and a polarization beam splitter prism, which realizes mode-locked output by controlling the polarization evolution of the pulse in the cavity. Next is the Faraday rotator, the half wave plate and the longitudinally placed polarization beam splitter prism. Based on the irreversibility of the polarization rotation of the Faraday rotator, this part controls the polarization of the light and plays the role of a spatial isolator, making the laser run unidirectionally and protecting the pump source. Then there is the spectral filtering structure composed of a 600 linesmm blazed grating combined with a fiber collimator. The last quarter wave plate is used to convert linear polarized light into elliptically polarized light. The polarization beam splitter prism (PBS1) also serves as the output port of the entire laser. The remaining 10% of the light is connected to a collimator placed in front of the scattering medium, and connected to a CCD at the rear of the scattering medium for generating and detecting speckles.

[0023] When the pump power reaches a certain value, the polarization state diagrams of the pulses at different positions in the ring cavity are as follows: Figure 2 As shown, the light is modulated into linear polarized light after passing through ISO, and the linear polarized light becomes elliptically polarized light after passing through QWP2 (elliptically polarized light is composed of S light and P light that are perpendicular to each other and have different intensities). When the two linear polarized lights that make up the elliptically polarized light propagate in the optical fiber, different nonlinear phase shifts will be accumulated due to different intensities, which will eventually cause the polarization state to rotate, and the rotation at the pulse peak is the largest. Therefore, QWP2 and HWP1 are adjusted to make the light transmittance at the peak large, while the intensity of the two wings is absorbed, which eventually leads to the narrowing of the light pulse. This process is equivalent to the effect of saturable absorption.

[0024] The 2000 speckle patterns of the three states (single pulse, multi-pulse and unlocked state) were collected by rotation and translation to obtain the final 6000 images, and then 4000 images were used for CNN ( Figure 6 The network was trained with CNN (as shown in Figure 1), 1,000 images were tested with CNN network, and the remaining 1,000 images were verified with CNN network. The accuracy of the final verification set was above 99%. This shows that the convolutional neural network (CNN) can be used to classify single solitons and multiple solitons, completely replacing the spectrometer and autocorrelator in judgment.

[0025] The method of using artificial intelligence to assist in judging the mode-locking state is to use CCD to detect the intensity and appearance of the speckle after the laser passes through the scattering medium while adjusting the angles of HWP1, QWP1 and QWP2. The specific method is to first adjust QWP1, HWP1, and QWP2 in sequence to minimize the speckle brightness, and then use the speckle pattern at this time to determine whether the mode is locked at this time through a trained neural network. If the mode is not locked at this time, the second step is required to select any of the three wave plates for fine-tuning within 10°, and input the speckle pattern generated during fine-tuning into the trained neural network for judgment until mode locking is achieved ( Figure 5 shown).

Claims

1. A method for artificial intelligence-assisted mode locking and pulse classification based on speckle patterns, characterized in that: The method includes a nonlinear polarization rotation (NPR) laser device, a speckle generating device, speckle collection of various pulse states, a method for classifying various pulse types using a convolutional neural network, and a method for judging a mode-locking state; The NPR device includes a semiconductor pump source (Pump), a wavelength division multiplexer (WDM), a single-mode optical fiber (2.5m), a Yb-doped optical fiber (0.3m), two optical fiber collimators (Collimator), two quarter-wave plates (QWP), two half-wave plates (HWP), two polarization beam splitters (PBS), a grating (Grating), a Faraday magneto-optical rotator (FR), a 10:90 beam splitter (Coupler), etc., which are used to generate mode-locked pulses; The speckle generating device includes a scattering medium and a CCD, which is used to generate and detect speckles; The speckle collection process of various pulse states is to generate three pulse states (single pulse, multiple pulses and unlocked state) by fine-tuning the wave plate left and right. This process requires the use of an autocorrelation instrument to determine whether it is a single or multiple pulse; The process of using the convolutional neural network to classify various pulse types is to obtain 6000 images of data by rotating and translating 2000 speckle images of three states (single pulse, multi-pulse and unlocked state), and then train 4000 images by using the CNN network, test 1000 images by using the CNN network, and verify the remaining 1000 images by using the CNN network; The method for artificial intelligence-assisted mode-locking state judgment refers to using CCD to detect the speckle intensity and appearance after the laser passes through the scattering medium while adjusting the angles of HWP1, QWP1 and QWP2. The specific method is to first adjust QWP1, HWP1 and QWP2 in sequence to minimize the speckle brightness, and use the trained neural network to judge whether the speckle pattern at this time is mode-locked at this time. If the mode is not locked at this time, the second step is required to select any one of the three wave plates for fine-tuning within 10°, and input the speckle pattern generated during the fine-tuning into the trained neural network for judgment until the mode is locked, and then perform the third step, input the speckle pattern during mode locking into the trained neural network again to judge whether it is a single pulse, if so, the mode-locking operation stops; if not, the fourth step is required to reduce the power and further transmit the speckle pattern obtained by the CCD to the network structure under the power for judgment, until a single pulse is finally output and the mode-locking operation is completely ended.

2. The method of artificial intelligence-assisted mode locking and pulse classification based on speckle pattern according to claim 1, characterized in that: The principle of the NPR laser is that light is modulated into linearly polarized light after passing through a Faraday magneto-optical rotator, and the linearly polarized light becomes elliptically polarized light after passing through QWP2 (elliptically polarized light is composed of mutually perpendicular S light and P light with different intensities). When the two linearly polarized lights constituting the elliptically polarized light propagate in the optical fiber, different nonlinear phase shifts will be accumulated due to different intensities, which eventually causes the polarization state to rotate and the rotation at the pulse peak is the largest. Therefore, HWP1 and QWP1 are adjusted to make the light transmittance at the peak large, while the intensity of the two wings is absorbed, which eventually leads to the narrowing of the light pulse. This process is equivalent to the effect of saturable absorption. The spectra of the corresponding output lasers under different saturable and absorption states are different, resulting in different speckle appearances or intensities. Compared with HWP1 and QWP1, QWP2 has a more significant effect on the modulation depth. This is because they play different roles at different positions. In the experiment, the effect of nonlinear polarization rotation can be enhanced or weakened by adjusting the angle of QWP2.

3. The method of artificial intelligence-assisted mode locking and pulse classification based on speckle pattern according to claim 1, characterized in that: The method for judging the mode-locking state is to optimize the laser loss, and start adjusting the wave plates in sequence only when the ratio of the output power to the pump power of the semiconductor laser is above 15%, so that the speckle brightness corresponding to each wave plate is the lowest, and then fix QWP1 and QWP2, rotate HWP1 to record the state of the pulse, and there are 4 mode-locking areas within 360°, and the angles differ by about 90°, and then fix QWP1 and rotate QWP2 again to find that there are also mode-locking areas at an angle of 180°, and then repeat the rotation of HWP1, and we find again that mode-locking can still be achieved in the 4 fast mode-locking areas with basically the same angle as the previous one.

4. The method of artificial intelligence-assisted mode locking and pulse classification based on speckle pattern according to claim 1, characterized in that: The CNN network is used to classify various pulse types. Excluding the input and output layers, there are three layers of the network. We compress the input image to a size of 128 pixels × 128 pixels, and then the convolution layer is used. The size of the convolution kernel is 5*5. After the convolution layer, the image becomes a convolution feature map with 6 sizes of 124 pixels × 124 pixels. After the pooling layer, the image becomes a pooling feature map with 6 sizes of 62 pixels × 62 pixels. The pooling feature map is then flattened to 1 pixel × 23064 pixels, and finally passed to a fully connected layer to reach the classification output.

5. The method of artificial intelligence-assisted mode locking and pulse classification based on speckle pattern according to claim 4, characterized in that: In the NPR laser, HWP1, QWP1 and QWP2 are installed on an electric rotating platform, and then the speckle and convolutional neural network are used to automatically rotate to achieve the purpose of automatic mode locking, and automatically determine the pulse type (single pulse, multiple pulses and unlocked state). This method completely replaces the spectrometer and autocorrelator to determine whether the mode is locked and the pulse type, reduces the cost of laser measurement equipment, and can achieve mode locking in about 2 minutes.

6. A method for artificial intelligence-assisted mode locking and pulse classification based on speckle pattern as described in any one of claims 1 to 5 is applied in a mode-locked laser. Automatic mode locking can be quickly achieved by changing the brightness of the speckle pattern, and to a certain extent, it completely replaces the use of spectrometers and autocorrelation instruments to determine whether the mode is locked and the pulse type.

Images