Time-delay reservoir computing method based on dispersion orthogonal polarization light feedback vcsel

Abstract

The application discloses a time delay reservoir computing method based on dispersion orthogonal polarization light feedback VCSEL, and the time delay RC system comprises an adjustable laser, a Mach-Zehnder modulator, a VCSEL, an optical isolator, an optical coupler, a polarization controller, a variable optical attenuator, a chirped fiber Bragg grating and a photodetector; the signal of an input layer is injected into the X polarization component of the VCSEL of a reservoir layer after being modulated by the Mach-Zehnder modulator; the emission light of the VCSEL is divided into two paths after passing through an optical circulator, one path is reflected back to the VCSEL by the chirped fiber Bragg grating after being rotated by the polarization controller, thereby constructing the reservoir layer of the time delay RC system; and the other path is transmitted to an output layer and converted into an electrical signal output by the photodetector. The application can further inhibit the time delay characteristics, thereby further inhibiting the generation of harmful resonance, and has stronger nonlinear characteristics and better computing performance.

Description

Technical Field

[0001] This invention belongs to the field of time-delayed storage pool calculation systems, and particularly relates to a time-delayed storage pool calculation method based on dispersive orthogonal polarization feedback VCSEL. Background Technology

[0002] Recurrent Transformation (RC), as an emerging information processing method, is a further development based on Recurrent Neural Networks (RNNs). It was proposed to overcome the limitations of traditional neural network models. While early neural network models possessed good nonlinear mapping capabilities, FNNs, due to their fixed hierarchical structure and unidirectional information propagation, struggled to effectively handle complex tasks with time dependencies. To address this, RNNs introduced recurrent structures, establishing temporal connections between the hidden layers to remember and utilize historical information. However, RNNs are prone to problems such as vanishing or exploding gradients during training, and their parameter optimization process is complex. These issues make training long sequences extremely difficult, limiting their widespread adoption in practical applications. Against this backdrop, RC, as a concise and efficient information processing method, emerged. Its core idea is to fix and randomly initialize the weights of the recurrent network portion of a traditional RNN, training only the output layer, thus simplifying the training process.

[0003] In recent years, the structure of Reliable RC (Reliable RC) has moved beyond spatial structures composed of multiple nonlinear nodes, evolving into time-delayed structures consisting of a single nonlinear node and a delayed feedback loop. Traditional RC reservoir structures are spatial interconnections of numerous nonlinear physical nodes. Therefore, physical hardware implementation requires controlling the internal states of many nodes, inter-node connections, and the acquisition of node outputs. This limits the construction of large-scale RC networks with an exponentially increasing number of nodes. In contrast, a time-delayed RC system consists of a single nonlinear node with a time-delayed feedback loop. Virtual nodes in the feedback loop replace a large number of nonlinear physical nodes in traditional RC. Therefore, time-delayed RC hardware implementation is simpler, leading to the research on time-delayed RC systems that construct reservoirs using various nonlinear time-delay systems. In a time-delayed RC system, the total time delay feedback time is... At the same interval in the feedback loop Set N virtual nodes, that is The input layer of the delay RC is masked. Assign a corresponding input weight value to each virtual node. (Time mask) time span equal And by It consists of a uniformly distributed random value ranging from -1 to 1. The time-delayed RC reservoir layer consists of at least one nonlinear node and one feedback loop. In the output layer, at time intervals... The sampled output electrical signal is used as the virtual node state. Then, a relevant algorithm is used to train and optimize the output weights. Finally, the RC system output is obtained by linearly summing the virtual node states and the output weights.

[0004] With the expansion of delay-based RC applications and the diversification of requirements, designing RC systems with stronger computational performance remains a worthwhile challenge. Typically, the nonlinearity of nonlinear nodes in an RC system directly affects its computational performance; systems operating under higher nonlinear states often possess stronger computational capabilities. Currently... Under normal settings, a resonance can occur, weakening the nonlinearity of the nodes. The root cause of this resonance is the introduction of a time-delay feedback loop. The output of the reservoir carries certain time-delay characteristics, resulting in a weak periodicity in its impulse response. This periodicity matches the periodicity of the masked input signal. This resonance severely impairs the computational power of time-delay RC. Therefore, exploring how to suppress resonance to improve the performance of time-delay RC is of great significance.

[0005] Among existing optical delay RC schemes that enhance system computational performance by suppressing resonance, the closest to this invention is the delay RC technique based on orthogonally polarized feedback VCSELs. This scheme uses VCSELs as nonlinear nodes and utilizes the polarization dynamics characteristics of VCSELs to suppress, to some extent, the nonlinearity of the feedback VCSEL. and The resonance induced by matching enhances the nonlinearity of VCSELs and improves the computational performance of RC. The apparatus for this method is as follows: Figure 2 As shown.

[0006] Figure 2 The scheme shown mainly consists of a tunable laser, a Mach-Zehnder modulator, a VCSEL, an optical isolator, an optical coupler, a polarization controller, a variable optical attenuator, and a photodetector.

[0007] The implementation steps of this scheme are as follows: In the input layer, the input data is multiplied by the mask signal to obtain the input signal, which is then modulated by a Mach-Zehnder modulator. The modulated signal is injected into the X-polarization component of the VCSEL. The light emitted from the VCSEL passes through an optical circulator, an optical coupler, a variable optical attenuator, and a polarization controller before being fed back to the VCSEL, thus constructing the reservoir layer of the time-delay RC system. During the feedback process, the polarization controller performs orthogonal polarization rotation on the light emitted from the VCSEL, i.e., the X-polarization component of the light emitted from the VCSEL is fed back to the Y-polarization component of the VCSEL, and the Y-polarization component is fed back to its X-polarization component. The output optical signal of the VCSEL is sent to the output layer through an optical circulator and an optical coupler, and then converted into an output electrical signal by a photodetector. The output electrical signal is sampled at equal time intervals and linearly weighted and summed to obtain the RC output result.

[0008] Existing techniques utilize the polarization dynamics of VCSELs to perform orthogonal polarization rotation on the emitted light in the feedback loop, suppressing the time delay characteristics of the output light and reducing the periodicity of the reservoir impulse response. This mitigates harmful resonances caused by the matching of different periodic characteristics, resulting in higher computational performance. However, in this approach, the VCSEL output light undergoes two feedback cycles: the first feedback rotates the polarization component to its orthogonal polarization direction, while the second feedback rotates it back to its original direction, generating a strong time delay. Therefore, this approach has a limited upper limit for time delay suppression and cannot completely eliminate resonance; the computational performance of the RC system can be further improved.

[0009] Terminology Explanation:

[0010] Feedforward Neural Network (FNN): A neural network model containing multiple hidden layers. Each neuron receives input from the neurons in the previous layer, sums the inputs in a weighted manner, passes the sum through a non-linear activation function, and then passes the result to the neurons in the next layer.

[0011] Recurrent Neural Network (RNN): A type of neural network model used to process sequential data. Its recurrent structure allows it to capture temporal dependencies in the data.

[0012] Reservoir Computing (RC) is a type of recurrent neural network. Its internal weights are randomly fixed, transforming input data into a high-dimensional dynamic sequence. Only the network's output needs to be trained, effectively capturing temporal information from the input sequence.

[0013] Vertical-cavity surface-emitting laser (VCSEL): A special type of laser that can generate two mutually orthogonal polarization components, namely the X-polarization component and the Y-polarization component.

[0014] Orthogonal polarization feedback: The X-polarization component of the VCSEL output is fed back to the Y-polarization component, and the Y-polarization component is fed back to the X-polarization component.

[0015] Polarization dynamics: In VCSEL, it refers to the interaction and dynamic changes between two orthogonal polarization components.

[0016] Chirped Fiber Bragg Grating (CFBG): A fiber grating device with a non-uniform periodic structure.

[0017] Time delay feedback time: The time required for a signal to propagate and cycle once within the photoelectric feedback loop.

[0018] Delay characteristics: The autocorrelation peak value of the autocorrelation function of the laser output time series at the loop delay. Summary of the Invention

[0019] To address the aforementioned problems, this invention provides a method for calculating the time delay reservoir based on a dispersive orthogonal polarization feedback VCSEL.

[0020] The present invention discloses a time delay reservoir calculation method based on a dispersive orthogonal polarization feedback VCSEL. The time delay RC system used includes a tunable laser, a Mach-Zehnder modulator, a VCSEL, an optical isolator, an optical coupler, a polarization controller, a variable optical attenuator, a chirped fiber Bragg grating, and a photodetector.

[0021] In the input layer, the input data is multiplied by the mask signal to obtain the input signal. The input signal is then modulated by a Mach-Zehnder modulator and injected into the X-polarization component of the VCSEL in the reservoir layer.

[0022] The emitted light from the VCSEL is split into two paths after passing through an optical circulator. One path undergoes orthogonal polarization rotation by a polarization controller and is then reflected back to the VCSEL by a chirped fiber Bragg grating, forming the reservoir layer of the time-delay RC system. The other path is transmitted to the output layer and converted into an electrical signal output by a photodetector. The output electrical signal is sampled at equal time intervals and then linearly weighted and summed to obtain the system output result.

[0023] Compared with the traditional time-delay RC scheme based on orthogonally polarized feedback VCSELs, this invention utilizes the dispersion effect of chirped fiber Bragg gratings in the time-delay loop of the reservoir layer to further suppress time-delay characteristics. Its specific theoretical model rate equation is described as follows:

[0024]

[0025] in, and The amplitudes of the slowly varying complex electric field, representing the X-polarization and Y-polarization components, are respectively. and These represent the difference between the total number of charge carriers and the numerical difference between charge carriers with opposite spin directions, respectively. For field attenuation rate, Linewidth enhancement factor; represent The attenuation rate, The spin-flip rate, Represents linear dichroism. Represents linear birefringence; It is the normalized injection current; the third term in equations (1) and (2) is the time-delay feedback term, where Represents the intensity of feedback. Representing the feedback phase, the impulse response of a chirped fiber Bragg grating is: , The feedback delay time represents the feedback loop. and These represent the angular frequencies of the X-polarization component and the Y-polarization component, respectively. , At the speed of light, The center wavelength of the VCSEL; the last term of equations (1) and (2) , , representing Lang Zhiwan noise.

[0026] The fourth term of equation (1) describes the input layer, which multiplies the input data with the mask signal to obtain a preprocessed input signal. This preprocessed signal is then modulated by a Mach-Zehnder modulator and injected into the X-polarization component of the VCSEL. The injection intensity of the external laser. To determine the injected optical angular frequency, since the input signal is injected into the X-polarization component, we use... This indicates frequency detuning between the injected light and the laser; It is the preprocessed input signal after mask modulation, where This is the mask scaling factor. For mask signal, The input data is sampled and held; the mask signal is periodic over a delay loop time and at virtual node intervals. The above is constant; the input signal after masking. Transient dynamics can be excited in the output layer at time intervals. Sampled output electrical signal As a virtual node state Then, the ridge regression algorithm is used to train and optimize the output weights. Finally, during the testing process, the following was used: Obtain the output of the RC system.

[0027] The beneficial technical effects of this invention are as follows:

[0028] This invention, building upon the traditional time-delay RC system based on orthogonally polarized feedback VCSELs, utilizes the dispersion effect of CFBG to propose a time-delay RC calculation system based on dispersive orthogonally polarized feedback VCSELs. Compared to traditional schemes, this time-delay RC system can further suppress time-delay characteristics, thereby further suppressing the generation of harmful resonances, exhibiting stronger nonlinear characteristics and better computational performance. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the time delay RC of the VCSEL based on dispersive orthogonal polarization feedback in this invention.

[0030] Figure 2 This is a schematic diagram of the time delay RC based on orthogonally polarized light feedback VCSEL. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0032] This invention discloses a method for calculating the time delay reservoir based on a dispersive orthogonal polarization feedback VCSEL, employing a time delay RC system such as... Figure 1 As shown, the system includes a tunable laser, a Mach-Zehnder modulator, a VCSEL, an optical isolator, an optical coupler, a polarization controller, a variable optical attenuator, a chirped fiber Bragg grating, and a photodetector. The signal from the input layer is modulated by the Mach-Zehnder modulator and injected into the X-polarization component of the VCSEL in the reservoir layer. The emitted light from the VCSEL passes through the optical coupler; one path undergoes orthogonal polarization rotation and is reflected back to the VCSEL by the chirped fiber Bragg grating, while the other path is transmitted to the output layer and converted into an electrical signal by the photodetector.

[0033] This invention, based on the traditional time-delay RC system using orthogonal polarization feedback from VCSELs, further suppresses the time-delay characteristics by utilizing the dispersion effect of CFBGs, thereby further suppressing the generation of harmful resonances, enhancing the nonlinear characteristics of the system, and significantly improving the computational performance of RC. The specific implementation process is as follows: In the input layer, the input data is multiplied by a mask signal to obtain the input signal. The input signal is modulated by a Mach-Zehnder modulator and injected into the X-polarization component of the VCSEL in the reservoir layer. The emitted light from the VCSEL is split into two paths after passing through an optical circulator. One path undergoes orthogonal polarization rotation by a polarization controller and is reflected back to the VCSEL by a chirped fiber Bragg grating, constructing the reservoir layer of the time-delay RC system. The other path is transmitted to the output layer and converted into an electrical signal output by a photodetector. The output electrical signal is sampled at equal time intervals and then linearly weighted and summed to obtain the system output result.

[0034] Compared with the traditional time-delay RC scheme based on orthogonally polarized feedback VCSELs, this invention utilizes the dispersion effect of chirped fiber Bragg gratings in the time-delay loop of the reservoir layer to further suppress time-delay characteristics. Its specific theoretical model rate equation is described as follows:

[0035]

[0036] in, and The amplitudes of the slowly varying complex electric field, representing the X-polarization and Y-polarization components, are respectively. and These represent the difference between the total number of charge carriers and the numerical difference between charge carriers with opposite spin directions, respectively. For field attenuation rate, Linewidth enhancement factor; represent The attenuation rate, The spin-flip rate, Represents linear dichroism. Represents linear birefringence; It is the normalized injection current; the third term in equations (1) and (2) is the time-delay feedback term, where Represents the intensity of feedback. Representing the feedback phase, the impulse response of a chirped fiber Bragg grating is: , The feedback delay time represents the feedback loop. and These represent the angular frequencies of the X-polarization component and the Y-polarization component, respectively. , At the speed of light, The center wavelength of the VCSEL; the last term of equations (1) and (2) , , representing Langevin noise.

[0037] The fourth term of equation (1) describes the input layer, which multiplies the input data with the mask signal to obtain a preprocessed input signal. This preprocessed signal is then modulated by a Mach-Zehnder modulator and injected into the X-polarization component of the VCSEL. The injection intensity of the external laser. To determine the injected optical angular frequency, since the input signal is injected into the X-polarization component, we use... This indicates frequency detuning between the injected light and the laser; It is the preprocessed input signal after mask modulation, where This is the mask scaling factor. For mask signal, The input data is sampled and held; the mask signal is periodic over a delay loop time and at virtual node intervals. The above is constant; the input signal after masking. Transient dynamics can be excited in the output layer at time intervals. Sampled output electrical signal As a virtual node state Then, the ridge regression algorithm is used to train and optimize the output weights. Finally, during the testing process, the following was used: Obtain the output of the RC system.

[0038] Traditional time-delay RC systems suffer from resonance problems caused by the matching of different periodic characteristics, which impairs the system's computational performance. Existing schemes utilize the polarization dynamics of VCSELs to suppress this harmful resonance, but the suppression upper limit is relatively low, and resonance still has a strong impact on the system's computational performance.

[0039] This invention simply adds a chirped fiber Bragg grating, an inexpensive passive optical device, to the feedback loop to reflect the light emitted by the laser. This modification, while maintaining the simple structure of the traditional VCSEL-based time-delay RC system, further suppresses the time delay characteristics, thereby further suppressing resonance that impairs system computational performance and significantly improving the computational performance of the RC system.

Claims

1. A time-delay reservoir computing method based on dispersion-orthogonal polarization light feedback VCSEL, characterized in that, The time-delay RC system employed includes a tunable laser, a Mach-Zehnder modulator, a VCSEL, an optical isolator, an optical coupler, a polarization controller, a variable optical attenuator, a chirped fiber Bragg grating, and a photodetector. In the input layer, the input data is multiplied by the mask signal to obtain the input signal. The input signal is then modulated by a Mach-Zehnder modulator and injected into the X-polarization component of the VCSEL in the reservoir layer. The emitted light from the VCSEL is split into two paths after passing through an optical circulator. One path undergoes orthogonal polarization rotation by a polarization controller and is then reflected back to the VCSEL by a chirped fiber Bragg grating, forming the reservoir layer of the time-delay RC system. The other path is transmitted to the output layer and converted into an electrical signal output by a photodetector. The output electrical signal is sampled at equal time intervals and then linearly weighted and summed to obtain the system output result. In the time delay loop of the reservoir layer, the dispersion effect of the chirped fiber Bragg grating is utilized to further suppress the time delay characteristics. The specific theoretical model rate equation is described as follows: ； in, and The amplitudes of the slowly varying complex electric field, representing the X-polarization and Y-polarization components, are respectively. and These represent the difference between the total number of charge carriers and the numerical difference between charge carriers with opposite spin directions, respectively. For field attenuation rate, Linewidth enhancement factor; represent The attenuation rate, The spin-flip rate, Represents linear dichroism. Represents linear birefringence; It is the normalized injection current; the third term in equations (1) and (2) is the time-delay feedback term, where Represents the intensity of feedback. Representing the feedback phase, the impulse response of a chirped fiber Bragg grating is: , The feedback delay time represents the feedback loop. and These represent the angular frequencies of the X-polarization component and the Y-polarization component, respectively. , At the speed of light, The center wavelength of the VCSEL; the last term of equations (1) and (2) , This represents Langevin noise; The fourth term of equation (1) describes the input layer, which multiplies the input data with the mask signal to obtain a preprocessed input signal. This preprocessed signal is then modulated by a Mach-Zehnder modulator and injected into the X-polarization component of the VCSEL. The injection intensity of the external laser. To determine the injected optical angular frequency, since the input signal is injected into the X-polarization component, we use... This indicates frequency detuning between the injected light and the laser; It is the preprocessed input signal after mask modulation, where This is the mask scaling factor. For mask signal, The input data is sampled and held; the mask signal is periodic over a delay loop time and at virtual node intervals. The above is constant; the input signal after masking. Transient dynamics can be excited in the output layer at time intervals. Sampled output electrical signal As a virtual node state Then, the ridge regression algorithm is used to train and optimize the output weights. Finally, during the testing process, the following was used: Obtain the output of the RC system.

Images