A semiconductor laser-based deep pulse reservoir pool computing system and method

Abstract

This invention discloses a deep pulse reservoir computing system and method based on semiconductor lasers, relating to the fields of optical computing and reservoir computing technology. The system includes an input layer, a reservoir layer, and an output layer connected in sequence. The input layer encodes the chaotic time series to be predicted into pulsed optical excitation signals. The reservoir layer consists of n cascaded semiconductor lasers, with adjacent lasers bidirectionally coupled through delayed light injection. The delay time is a positive integer multiple of the pulse interval, forming a nonlinear dynamic network with time memory capabilities. The output layer performs feature extraction, spatiotemporal fusion, and weighted readout of the output responses of each semiconductor laser to complete time series prediction. This invention combines low power consumption, strong memory capabilities, and high prediction accuracy, effectively addressing the core shortcomings of traditional pulse reservoir computing and is applicable to the field of photonic neuromorphic computing.

Description

Technical Field

[0001] This invention relates to the field of optical computing and reservoir computing technology, and in particular to a deep pulse reservoir computing system and method based on semiconductor lasers. Background Technology

[0002] With the rapid development of artificial intelligence and time series prediction technology, recurrent neural networks (RNNs) have shown great application potential in the field of complex dynamic time series data processing due to their excellent nonlinear mapping capabilities. However, their training process has inherent defects such as high computational complexity, gradient vanishing or gradient exploding, which greatly limit their hardware implementation and engineering applications.

[0003] Optical reservoir computing, as a simplified recurrent neural network model, has emerged. By fixing the internal weights of the input and reservoir layers and performing linear training only on the output layer, it significantly reduces training costs and hardware implementation difficulties. It excels at handling nonlinear temporal prediction problems and has become one of the core research directions in neuromorphic computing. Currently, optical reservoir computing is mainly divided into two categories: continuous light injection type and pulsed type. Traditional continuous optical reservoirs rely on the nonlinear dynamics of lasers to construct a high-dimensional state space. Although the performance is stable, the laser must continuously operate at a high bias, resulting in high system energy consumption, making it difficult to adapt to low-power edge intelligence scenarios.

[0004] To address the energy consumption challenge, pulsed reservoir computing schemes have been proposed. These schemes significantly reduce average system power consumption by encoding time-series signals into sparse pulse sequences. However, they suffer from a severe lack of memory capacity: the sparse distribution of pulse signals over time makes it difficult for a single laser to effectively maintain long-range correlations of historical input information. This results in significantly limited prediction accuracy when processing complex chaotic time-series data with long-term dependencies. Existing solutions have consistently failed to achieve a balance between low power consumption, strong memory capacity, and high prediction accuracy, becoming a core bottleneck restricting the large-scale application of pulsed optical reservoir computing technology. Summary of the Invention

[0005] To address this, embodiments of the present invention provide a deep pulse reservoir computing system and method based on semiconductor lasers, which solves the technical problems in the prior art where pulsed optical reservoir computing systems cannot simultaneously achieve low power consumption and strong memory capabilities, and suffer from insufficient memory capacity, limited state space dimension, and inadequate utilization of dynamic response features, thereby resulting in limited prediction accuracy for long-range dependencies of complex chaotic time series.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a depth pulse reservoir computing system based on a semiconductor laser, comprising an input layer, a reservoir layer, and an output layer connected sequentially by signals: The input layer includes a driving laser, an arbitrary waveform generator, and a modulator. The arbitrary waveform generator is used to generate a masked analog signal. The analog signal is multiplied by the chaotic time series to be predicted to obtain an input electrical signal. The modulator is used to modulate the input electrical signal onto a continuous optical carrier output by the driving laser to generate a pulsed optical excitation signal. The storage pool layer is composed of It consists of cascaded semiconductor lasers, in which The optical input terminal of the first semiconductor laser is connected to the optical output terminal of the modulator, and is used to receive the pulsed light excitation signal and generate self-pulse output under optical injection drive; adjacent semiconductor lasers are bidirectionally coupled through a delayed optical injection link, i.e., the first... The output light from the first semiconductor laser is delayed and then injected into the second... The first semiconductor laser, the... The output light of the semiconductor laser is injected into the first... One semiconductor laser, of which The delay time of the delayed light injection link is a positive integer multiple of the interval time of the output pulse of the first semiconductor laser, so that the reservoir layer forms a nonlinear dynamic response network with time memory capability. The output layer includes... Each semiconductor laser is configured in a one-to-one correspondence. Individual optical isolators, The system includes a photodetector, an adder, and a programmable gate array. The input of each optical isolator is connected to the optical output of the corresponding semiconductor laser to ensure unidirectional optical signal transmission. The input of each photodetector is connected to the output of the corresponding optical isolator to convert the output optical signal of the corresponding semiconductor laser into an electrical response signal. The input of the adder is connected to the output of all photodetectors to sequentially concatenate multiple feature vectors extracted from each electrical response signal to generate a fused feature vector. The programmable gate array is connected to the output of the adder to train and generate an output weight matrix based on the fused feature vector, and to perform prediction calculations for chaotic time series.

[0007] Preferably, the pulsed light excitation signal generated by the input layer is injected only into the first semiconductor laser in the reservoir layer, and the remaining semiconductor lasers receive the light signal only through the mutual coupling link between adjacent semiconductor lasers.

[0008] Preferably, the front end of the pulsed light excitation signal generated by the input layer is provided with n preheating pulse sequences for bringing the semiconductor laser in the reservoir layer into a stable dynamic working state.

[0009] Preferably, the semiconductor laser is a laser capable of generating self-pulse output, specifically a common semiconductor laser that generates self-pulse output through external light injection, or a semiconductor laser with an integrated saturable absorber, the output pulse signal of which is used to characterize the state information of the reservoir.

[0010] Preferably, the light injection between adjacent semiconductor lasers is achieved through fiber coupling or free space optical coupling, and the semiconductor lasers at each level are bidirectionally coupled.

[0011] Preferably, the delay time of the delayed light injection link is equal to a positive integer multiple of the interval time of the output pulse of the first semiconductor laser, so that the dynamic response of each semiconductor laser forms a multi-level delay correlation in the time dimension, thereby enhancing the system's ability to retain historical input information.

[0012] Preferably, in the reservoir layer, the output optical signal of the first semiconductor laser simultaneously includes the current pulse excitation signal injected from the input layer, as well as the multi-level historical pulse signals formed by delayed light injection from the subsequent intercoupled semiconductor lasers, thus forming a pulse response network with long-range time memory capability within the reservoir layer.

[0013] Preferably, the features extracted by the adder from each electrical response signal include at least one of the following: mean difference, peak difference, standard deviation, energy integral, dynamic range, and rising edge percentage. The mean difference is the difference between the mean response within the pulse segment and the mean response within the gap segment; the peak difference is the difference between the maximum response within the pulse segment and the minimum response within the gap segment; the standard deviation is the standard deviation of the response sequence within the pulse segment; the energy integral is the integral value of the response within the corresponding time window within the pulse segment; the dynamic range is the difference between the maximum and minimum response values ​​within a single pulse period; and the rising edge percentage is the ratio of the pulse rising edge duration to the pulse period.

[0014] Preferably, the adder will The feature vectors corresponding to each semiconductor laser are concatenated end-to-end according to the cascaded numbering order of the semiconductor lasers to generate a dimensionally expanded fused feature vector, achieving deep fusion of features in the time and spatial dimensions.

[0015] This invention also provides a method for calculating a depth pulse reservoir based on a semiconductor laser, implemented using the aforementioned semiconductor laser-based depth pulse reservoir calculation system, comprising the following steps: S1: The input layer performs normalization, masking, and pulse coding on the chaotic time series to be predicted to generate a pulsed light excitation signal with a preheating sequence. S2: The pulsed light excitation signal is injected into the first semiconductor laser in the reservoir layer to generate a self-pulse output. Through the bidirectional delay mutual coupling link between adjacent semiconductor lasers, each level of semiconductor laser generates a nonlinear dynamic response with historical information association, forming a pulse response network with time memory capability. S3: The output optical signals of each semiconductor laser are converted into electrical response signals by the photodetector of the output layer. Multidimensional feature extraction is performed on each electrical response signal to obtain multi-path feature vectors. Then, the multi-path feature vectors are sequentially concatenated by the adder to generate a spatiotemporal fusion feature vector. S4: Based on the fused feature vector of the training set, a linear readout algorithm is used to train and generate the output weight matrix through a programmable gated array. The fused feature vector of the test phase is multiplied with the output weight matrix to obtain the prediction result of the chaotic time series.

[0016] As can be seen from the above technical solutions, this invention application has the following beneficial effects: First, this invention adopts the core technical route of pulse-type encoding and pulse output. By encoding the timing information into a sparse pulse sequence, the semiconductor laser represents the state of the reservoir in the form of pulses with a low duty cycle. Compared with the traditional continuous light injection type optical reservoir, it significantly reduces the average operating power consumption of the semiconductor laser, while fully preserving the input information and nonlinear mapping capability. It solves the industry pain points of high energy consumption and difficulty in adapting to low-power edge intelligence scenarios in traditional optical reservoir systems, and provides a low-power implementation solution for the hardware implementation of photonic neuromorphic computing.

[0017] Secondly, by constructing a bidirectional delay-coupled architecture between adjacent semiconductor lasers, the present invention precisely matches the inter-coupling delay time to a positive integer multiple of the pulse interval, so that the output of the preceding semiconductor laser simultaneously includes the current input pulse signal and the multi-level historical pulse signals injected by the subsequent semiconductor laser after delay. Historical information can be cyclically retained and transmitted in the coupling network. From the physical architecture level, this invention solves the core defect of traditional pulse-type reservoirs that cannot effectively maintain long-range correlation of historical inputs due to the sparse temporal distribution of pulse signals, and significantly enhances the system's ability to process long-term time-dependent sequential tasks.

[0018] Third, this invention expands the nonlinear state space dimension of the system by constructing a deep reservoir structure through cascaded multiple semiconductor lasers. Simultaneously, by extracting multidimensional dynamic features from the response of each semiconductor laser and concatenating these features in cascaded order, it achieves deep fusion of temporal and spatial features, fully exploiting the nonlinear dynamic information throughout the entire pulse response process. This solves the problems of insufficient utilization of dynamic features and limited state space dimension in existing schemes. Experimental verification shows that the normalized mean square error (NMSE) of this scheme can reach [value missing]. Magnitude, coefficient of determination With a prediction accuracy of over 0.99, the accuracy is improved by more than 80% compared to a single semiconductor laser configuration, significantly enhancing the prediction accuracy and system generalization ability for complex chaotic time series. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Referring to the drawings will make the features and advantages of the present invention clearer. The drawings are illustrative and should not be construed as limiting the present invention in any way. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic diagram of the structure of a depth pulse reservoir computing system based on a semiconductor laser provided by the present invention; Figure 2 This is a schematic diagram of the semiconductor laser mutual coupling structure in this invention; Figure 3 This is a schematic diagram of the feature extraction and splicing process in this invention; Figure 4 This is a schematic diagram of the programmable gate array structure in this invention; Figure 5 This is the output timing diagram of the mutually coupled semiconductor laser in this invention, where (a) is the timing diagram of the externally injected signal, (b) is the output timing diagram of the first semiconductor laser, and (c) is the output timing diagram of the second semiconductor laser. Figure 6 This is a graph showing the prediction results of the deep pulse reservoir calculation system based on semiconductor lasers for chaotic time series in this invention; wherein, (a) is the distribution graph of the true value of the chaotic time series, (b) is the distribution graph of the predicted value of the chaotic time series by this system, and (c) is the distribution graph of the prediction error. Figure 7 This is a flowchart of a method for calculating a deep pulse reservoir based on a semiconductor laser, provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1This embodiment provides a deep pulse reservoir computing system based on a semiconductor laser, comprising an input layer, a reservoir layer, and an output layer connected in sequence. The input layer is used to perform mask modulation and pulse coding on the input chaotic time series to generate a pulsed light excitation signal, and a preheating sequence is added to the front end of the effective input signal; the reservoir layer consists of… A series of mutually coupled semiconductor lasers are cascaded together to form ( The first semiconductor laser generates pulse output under optical injection drive. Adjacent lasers are bidirectionally coupled through delayed optical injection to form a nonlinear dynamic network with time memory capability. The output layer is used to collect the output response of each semiconductor laser, perform multidimensional feature extraction and vector concatenation, and then complete training and prediction through a linear readout algorithm to output the prediction result of the chaotic time series.

[0022] This invention adopts a technical approach of "pulse-coded input—nonlinear response of mutually coupled semiconductor lasers—multidimensional feature fusion—linear readout prediction." It achieves low-power operation through pulse-type encoding, addresses the insufficient memory capacity of traditional pulse reservoirs through a bidirectional delay-coupling structure, and realizes deep fusion of spatiotemporal features through deep cascading of multiple lasers and feature stitching, ultimately achieving high-precision prediction of complex chaotic time series. The following provides a detailed description of each layer's structure and working principle: I. Input Layer Please see Figure 1 The internal structure of the input layer includes a driving laser, an arbitrary waveform generator, an electrical amplifier, and a modulator.

[0023] The driving laser is used to generate a continuous optical carrier as the carrier light source for the input optical signal, providing a stable optical carrier basis for subsequent pulse modulation.

[0024] The arbitrary waveform generator is used to generate a masked analog signal, which is multiplied by the chaotic time series to be predicted to generate an input electrical signal, thereby realizing the masking mapping and precoding of the input time series information.

[0025] The input terminal of the electrical amplifier is connected to the output terminal of the arbitrary waveform generator to amplify the power of the input electrical signal, ensuring that the modulator can operate at the optimal modulation depth and guaranteeing the stability and accuracy of the modulation effect.

[0026] The modulator's electrical input is connected to the output of the electrical amplifier, and its optical input is connected to the optical output of the driving laser. This modulator is used to modulate the amplified input electrical signal onto the continuous optical carrier output by the driving laser, generating a pulsed optical excitation signal. The signal is then injected into the reservoir layer. In this embodiment, the modulator is preferably a Mach-Zehnder intensity modulator. In other alternative embodiments, a phase modulator can also be used, both of which can achieve pulse modulation of the optical carrier.

[0027] Furthermore, this embodiment adds a feature to the front end of the effective input signal. A preheating pulse sequence is used to ensure the semiconductor laser in the reservoir layer enters a stable dynamic operating state before formally receiving the input signal, reducing the interference of the system's initial transient response on the calculation results. In the specific parameter settings, the pulse width is set to 0.5 ns, and the pulse period... The injection amplitude is set to 5ns, the preheating sequence length is 40ns, and the injection amplitude is normalized to 20. This injection method has the characteristics of low duty cycle and low average power consumption. It can not only completely preserve the numerical information of the input sequence, but also form a good match with the dynamic response characteristics of the subsequent semiconductor laser.

[0028] II. Storage Pool Layer Please see Figure 1 The internal structure of the middle storage pool layer and Figure 2 The storage pool layer is composed of It consists of a semiconductor laser array unit composed of cascaded mutually coupled semiconductor lasers, wherein This embodiment uses a two-stage semiconductor laser as an example, including a first semiconductor laser and a second semiconductor laser. In other embodiments, it can be extended to three stages, four stages, or even more, depending on the complexity of the prediction task. A cascaded, interconnected architecture is used to form a deeply interconnected reservoir, thereby obtaining a higher-dimensional state space and longer-term memory capabilities.

[0029] The optical input terminal of the first semiconductor laser is connected to the optical output terminal of the modulator in the input layer, and is used to receive the pulsed light excitation signal from the input layer. The two semiconductor lasers are bidirectionally coupled via a delayed optical injection link, which is implemented through a delay fiber. The delay time is set to the pulse interval of the first semiconductor laser. The positive integer multiple. In this embodiment, the coupling delay from the first semiconductor laser to the second semiconductor laser is set. The coupling delay from the second semiconductor laser to the first semiconductor laser Each pulse cycle creates a multi-level delay relationship in the dynamic response of each semiconductor laser, thereby enhancing the system's ability to retain historical input information.

[0030] In this embodiment, both the first and second semiconductor lasers are spin-vertical-cavity surface-emitting lasers (spin-VCSELs), and each laser has a right-hand circularly polarized light field. Left-handed circularly polarized light field Total number of carriers and spin carrier difference Such lasers exhibit significant polarization coupling and carrier dynamics characteristics, enabling them to not only respond nonlinearly to the current input but also retain some influence from historical inputs, fully meeting the core requirements of reservoir computing for nonlinear mapping and memory capabilities. In other alternative implementations, other types of semiconductor lasers, such as edge-emitting lasers and conventional vertical-cavity surface-emitting lasers, can also be used, as long as they can generate self-pulse outputs under optical injection.

[0031] To clearly illustrate the working mechanism of the mutually coupled laser in this embodiment, the complete dynamic rate equation is established as follows: (1) Evolution equation of light field The evolution of the right-hand and left-hand optical fields of the first semiconductor laser can be expressed as follows: ; .

[0032] The right-hand and left-hand optical field evolution of the second semiconductor laser can be expressed as follows: ; .

[0033] (2) Carrier evolution equation The evolution of the total carrier number and spin carrier difference in the first semiconductor laser can be expressed as follows: ; ; ; .

[0034] Of the above formulas, , and , These represent the right-handed and left-handed circularly polarized light fields of the first and second semiconductor lasers, respectively. , and , These represent the total number of charge carriers and the spin-carrier difference for the first and second semiconductor lasers, respectively. The light field attenuation rate, Linewidth enhancement factor It is the linear dichroic property. For birefringence, Carrier decay rate, It is the spin reversal rate; , These represent the pump intensities of the first and second semiconductor lasers, respectively. , These are the pump ellipticities of the first and second semiconductor lasers, respectively. For external injection strength, Inject an optical field into the pulses generated by the input layer. The externally injected light is detuned to the angular frequency of the first semiconductor laser; and The injection intensities are respectively from the second semiconductor laser to the first semiconductor laser and from the first semiconductor laser to the second semiconductor laser; and These are the delay times from the second semiconductor laser to the first semiconductor laser and from the first semiconductor laser to the second semiconductor laser, respectively. The angular frequency detuning between the two semiconductor lasers and These are the angular frequencies of the first and second semiconductor lasers, respectively.

[0035] The above equations together constitute a complete dynamic model describing the bidirectional mutually coupled semiconductor laser system in this embodiment. Regarding the selection of output quantities, this embodiment preferably uses the right-hand rotation intensity of the first semiconductor laser as the first output response and the right-hand rotation intensity of the second semiconductor laser as the second output response, expressed as: .

[0036] After removing the quiescent period corresponding to the preheating sequence, an effective response sequence is obtained for subsequent feature extraction. This selection method can retain both the direct response information of the first semiconductor laser to external input and the dynamic extension information formed by the second semiconductor laser under mutual coupling and delayed injection. In other optional embodiments, the intensity of left-hand circularly polarized light, the polarization difference signal, or the total light intensity can also be selected as the readout quantity of the reservoir state, all of which can achieve the technical effect of the present invention.

[0037] In the aforementioned bidirectional mutually coupled structure, the output signal of the first semiconductor laser includes not only the current pulse excitation signal injected from the input layer, but also historical pulse signals generated by delayed optical injection from the subsequent semiconductor laser. This historical information can be cyclically retained within the coupling network, thereby forming a pulse response network with long-range time memory capability in the reservoir layer. Optical injection coupling between adjacent semiconductor lasers can be achieved through fiber optic coupling or free-space optical coupling, and all laser stages are bidirectionally mutually coupled.

[0038] III. Output Layer Please see Figure 1 The internal structure of the output layer and Figure 3 The output layer includes and Each semiconductor laser is configured in a one-to-one correspondence. Individual optical isolators, The photodetector also includes an adder, a programmable gate array, and a time window partitioning module, a virtual node generation module, and a feature calculation module integrated into the system.

[0039] Each optical isolator is connected to the output of a corresponding semiconductor laser to maintain unidirectional light transmission and prevent reflected light from being injected back into the semiconductor laser, thus affecting its operational stability. Each optical isolator's output is connected to a photodetector, which converts the output optical signal of the corresponding semiconductor laser into an electrical response signal. In this embodiment, the first semiconductor laser outputs an electrical signal of: The second semiconductor laser outputs an electrical signal of .

[0040] The input of the time window division module is connected to the output of all photodetectors, and is used to divide the time window according to the pulse period of the input layer. The electrical response signals output by each semiconductor laser are segmented; each pulse period corresponds to a virtual node, and each response segment is further divided into a pulse segment (width 0.5ns) and a gap segment (width 4.5ns), providing a unified time window basis for subsequent feature calculations.

[0041] The virtual node generation module is connected to the time window division module and is used to map multiple pulse periods corresponding to each original sample to the response states of multiple virtual nodes according to the mask expansion structure. In this embodiment, each original sample corresponds to... There are 10 virtual nodes, each containing a complete pulse cycle response.

[0042] The feature calculation module is connected to the virtual node generation module and is used to extract multi-dimensional dynamic features from the response of each virtual node. For each semiconductor laser, the feature calculation module extracts six types of features in parallel. Internally, it integrates a mean calculation unit, a peak detection unit, a standard deviation calculation unit, an energy integration unit, a dynamic range calculation unit, and a rise time percentage calculation unit. The functions and feature calculation methods of each unit are as follows: 1. Mean Calculation Unit: Calculates the difference between the mean response within the pulse segment and the mean response within the gap segment corresponding to each virtual node, using this difference as the mean difference feature. ,Right now This is used to reflect the average response intensity under pulse excitation; 2. Peak Detection Unit: Calculates the difference between the maximum response value within the pulse segment and the minimum response value within the gap segment corresponding to each virtual node, using this difference as the peak difference feature. ,Right now , used to characterize the range of response amplitude variation; 3. Standard Deviation Calculation Unit: Calculates the standard deviation of the response sequence within the pulse segment corresponding to each virtual node, as a fluctuation characteristic. ,Right now , used to measure the degree of response fluctuation; 4. Energy Integration Unit: Calculates the integral value of the response within the pulse segment corresponding to each virtual node over the time window, which serves as the energy integration feature. ,Right now In the numerical implementation, the trapezoidal integral method is used to characterize the energy of the impulse response; 5. Dynamic Range Calculation Unit: Calculates the difference between the maximum and minimum response values ​​for each virtual node throughout the entire pulse period, using this difference as the dynamic range feature. ,Right now This is used to reflect the signal span within a complete cycle; 6. Rising Edge Proportion Calculation Unit: Calculates the proportion of the pulse rising edge time to the pulse period for each virtual node, as the rising edge proportion feature. It is used to characterize the shape and trend of pulse changes.

[0043] The above six characteristics comprehensively characterize the nonlinear dynamic behavior of semiconductor lasers under pulsed excitation from multiple perspectives, including mean, peak value, fluctuation, energy, range, and trend. Each semiconductor laser generates a corresponding six-dimensional feature vector. In other alternative implementations, other types of features such as frequency domain features, slope features, polarization difference features, and peak-to-peak value features can be added as needed, all of which fall within the scope of protection of this invention.

[0044] The input of the adder is connected to the output of the feature calculation module, and is used to concatenate the feature vectors corresponding to all semiconductor lasers according to the semiconductor laser numbering order to generate a dimension-expanded fused feature vector. In this embodiment, the feature vector of the first semiconductor laser Second semiconductor laser eigenvector The splicing method is as follows: .

[0045] This splicing method fully preserves the independence and sequence information of the characteristics of each semiconductor laser, and achieves deep integration of the characteristics in the time dimension (virtual nodes corresponding to different pulse periods of the same semiconductor laser) and the spatial dimension (different semiconductor lasers).

[0046] The programmable gate array is connected to the output of the adder and is used to determine the fused feature vector. Perform model training and time series prediction. Please refer to [link / reference]. Figure 4 The programmable gated array includes a weight calculation module and an output module.

[0047] The weight calculation module is connected to the adder and is used to receive the fused feature matrix and target prediction labels from the training phase, and calculates the output weight matrix using a linear readout algorithm. In this embodiment, for the total state matrix... The first 3000 samples are used as the training feature matrix. The corresponding target output is the next time step value of the input sequence. To enhance the model's fitting ability, a bias term of 1 is added after the training feature matrix to obtain the augmented matrix. In this embodiment, the ridge regression algorithm with regularization is preferably used to solve the output weight matrix. The calculation formula is: ; in, The regularization coefficient is set to [value] in this embodiment. Used to suppress overfitting The matrix is ​​the identity matrix. In other alternative implementations, other linear readout algorithms such as least squares regression and LASSO regression can also be used, all of which can achieve the technical effects of this invention.

[0048] The output module is connected to the weight calculation module and is used to output the test feature matrix during the testing phase. With the output weight matrix after training Multiplying them yields the prediction result for the chaotic time series. The calculation formula is: .

[0049] The above method enables prediction of the next moment or multiple subsequent moments in a chaotic time series. During the prediction phase, there is no need to retrain the internal parameters of the reservoir layer; simply inputting the test features into the pre-trained linear readout layer yields rapid prediction results, significantly reducing computational complexity and processing time.

[0050] IV. Calculation Method for Deep Pulse Reserve Pool Based on the Above System Please see Figure 7 This embodiment also provides a method for calculating a deep pulse reservoir based on a semiconductor laser, implemented using the above system, and includes the following steps: Step S1: Normalize, mask, and pulse-encode the chaotic time series through the input layer, and add a signal to the signal front end. A preheating pulse sequence is used to generate a pulsed light excitation signal. Specifically, the input sequence is mapped to... On each virtual node, the symbol corresponding to each virtual node is converted into a combination signal of pulse segment (0.5ns) and gap segment (4.5ns) in one pulse period. The pulse amplitude is proportional to the mask value, and the injected amplitude is normalized to 20.

[0051] Step S2: Apply pulse excitation signal A first semiconductor laser is injected into the reservoir layer, causing it to generate a self-pulse response under light injection. Simultaneously, through a bidirectional mutual coupling structure, the output light signal of the first semiconductor laser is injected into a second semiconductor laser, and the output light signal of the second semiconductor laser is delayed. The first semiconductor laser is then injected to form a deeply coupled pulse response network with time memory capability. The delay time is set to the pulse interval time of the first semiconductor laser. It is a positive integer multiple of the value, and in this embodiment it is 1.

[0052] The system dynamics are described by the aforementioned rate equations, and the fourth-order Runge-Kutta method is used for numerical solution, with a time step of [missing information]. The total simulation time is determined based on the length of the input sequence.

[0053] Step S3: Collect the output optical signals of each stage of the semiconductor laser through the photodetector array of the output layer and convert them into electrical response signals. and After removing the 40ns silence period corresponding to the front-end preheating sequence, the modules are divided according to the pulse period by time windows. The response segments are divided, and virtual node states are constructed through the virtual node generation module. Each original sample corresponds to 5 virtual nodes.

[0054] The feature calculation module extracts six-dimensional features from the response of each virtual node, including mean difference, peak difference, standard deviation, energy integral, dynamic range, and rise time percentage, to generate the first semiconductor laser feature vector. Second semiconductor laser eigenvector The feature vectors of the two semiconductor lasers are concatenated end-to-end according to their cascaded numbering order using an adder to obtain a fused feature vector. .

[0055] Step S4: Train the fused feature vector using ridge regression with a programmable gated array, taking the first 3000 samples as the training set and the last 1000 samples as the test set, and applying a regularization coefficient. Calculate the output weight matrix Multiply the fused feature vector from the testing phase with the output weight matrix to generate the chaotic time series prediction value for the corresponding time step.

[0056] V. System Timing Response Characteristics Please see Figure 5 This embodiment provides an output timing diagram of a mutually coupled semiconductor laser, wherein... Figure 5 (a) is a timing diagram of the externally injected signals generated by the input layer. Figure 5 (b) is a schematic diagram of the output timing of the first semiconductor laser. Figure 5 (c) is a schematic diagram of the output timing of the second semiconductor laser.

[0057] from Figure 5 As can be seen in (a), the externally injected signal is in the form of periodic pulses. Each pulse period contains a distinct pulse segment (0.5ns) and gap segment (4.5ns) structure. The front end contains a 40ns preheating sequence, which enables the semiconductor laser to enter a stable working state before it officially receives the input.

[0058] from Figure 5 As can be seen in (b), when an external pulse signal is injected, the output light intensity of the first semiconductor laser produces a significant transient spike response at the corresponding time position. This response is basically synchronized with the input pulse in time, directly reflecting the fast response characteristics of the semiconductor laser to the current input.

[0059] from Figure 5As can be seen in (c), the output intensity of the second semiconductor laser also exhibits a significant pulse response, but its response time is delayed by one pulse period (5ns) compared to the first semiconductor laser. This directly reflects the historical information injection effect brought about by the bidirectional mutual coupling structure. This delay characteristic allows the output of the second semiconductor laser to fully contain the information of the first semiconductor laser at historical moments, thereby realizing the system's retention and transmission of historical input information at the physical level and significantly enhancing the system's time memory capability.

[0060] VI. Analysis of Prediction Results Please see Figure 6 In this embodiment, the predictive performance of the system is tested using the classic Mackey-Glass chaotic time series. The test parameters are set as follows: the total sequence length is 4000 points, and the test length is 1000 points. The input sequence is masked and mapped to 5 virtual nodes, each virtual node corresponding to one pulse period. The reservoir layer uses two mutually coupled spin-VCSEL lasers. Each laser extracts 6-dimensional features and splices them together, resulting in a total feature dimension of 60 dimensions.

[0061] Figure 6 (a) shows the distribution of the true values ​​of the chaotic time series, illustrating the nonlinear chaotic waveform characteristics of the original time series; Figure 6 (b) shows the distribution of predicted values ​​for chaotic time series by this system. It can be clearly seen from the figure that the predicted curves closely match the actual curves. The normalized mean square error of the system's predictions is calculated to be... Coefficient of determination This indicates that the predicted value has a very high correlation with the actual value; Figure 6 Figure (c) shows the prediction error distribution. The error fluctuates slightly around the zero line, with the maximum error not exceeding 0.1, which further verifies the high-precision prediction capability of this system.

[0062] To further verify the effectiveness of the bidirectional mutually coupled structure and feature splicing technology of the present invention, a control experiment was set up in this embodiment: using only the features of the first semiconductor laser for prediction, the NMSE was approximately Using only the characteristics of the second semiconductor laser for prediction, the NMSE is approximately After feature splicing using dual semiconductor lasers, the NMSE was reduced to Compared to a single semiconductor laser configuration, the prediction accuracy is improved by more than 80%. Experimental results fully demonstrate that the bidirectional mutually coupled structure and feature vector splicing scheme of this invention can more accurately characterize the spatiotemporal evolution of chaotic time series and significantly improve the prediction accuracy of complex time series signals.

[0063] It should be noted that the semiconductor laser in this application is not limited to the spin-VCSEL in this embodiment. In other embodiments, other types of semiconductor lasers such as edge-emitting lasers and ordinary vertical-cavity surface-emitting lasers can also be used, as long as they can generate self-pulse output. At the same time, the output readout is not limited to the right-hand circularly polarized light intensity. The left-hand circularly polarized light intensity, polarization difference signal, or total light intensity can also be selected as the readout of the reservoir state.

[0064] Furthermore, the feature extraction methods in this application are not limited to the six types of features—mean difference, peak difference, standard deviation, energy integral, dynamic range, and rising edge percentage—used in this embodiment. Without departing from the technical concept of this application, frequency domain features, slope features, polarization difference features, peak-to-peak value features, etc., can also be added. The linear readout algorithm for the output layer is not limited to ridge regression; other linear readout methods such as least squares regression and LASSO regression can also be used. All solutions adopting the overall technical route of this invention—"pulse-coded input—mutually coupled semiconductor laser nonlinear response—multi-dimensional feature fusion—linear readout prediction"—should fall within the protection scope of this application.

[0065] The number of semiconductor lasers in the reservoir layer is not limited to the two-stage configuration in this embodiment; it can be expanded to three, four, or more stages depending on the complexity of the prediction task. A cascaded, interconnected structure of semiconductor lasers; as the number of semiconductor laser stages increases, the system can obtain a higher-dimensional state space and a longer time memory capability, while the complexity of parameter adjustment also increases accordingly, and can be flexibly configured according to actual application scenarios.

[0066] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0067] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0068] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A semiconductor laser based deep pulse reservoir computing system, characterized in that, It includes an input layer, a reservoir layer, and an output layer connected in sequence; The input layer includes a driving laser, an arbitrary waveform generator, and a modulator. The arbitrary waveform generator is used to generate a masked analog signal. The analog signal is multiplied by the chaotic time series to be predicted to obtain an input electrical signal. The modulator is used to modulate the input electrical signal onto a continuous optical carrier output by the driving laser to generate a pulsed optical excitation signal. The reserve pool layer is composed of a cascade of semiconductor lasers, wherein the light input end of the first semiconductor laser is connected with the light output end of the modulator, for receiving the pulsed light excitation signal and generating a self-pulsed output under light injection driving. Bidirectional mutual coupling between two adjacent semiconductor lasers is achieved through a delayed light injection link, i.e., the first... The output light from the first semiconductor laser is delayed and then injected into the second... The first semiconductor laser, the... The output light of the semiconductor laser is injected into the first... One semiconductor laser, of which The delay time of the delayed light injection link is a positive integer multiple of the interval time of the output pulses of the first semiconductor laser, enabling the reservoir layer to form a nonlinear dynamic response network with time memory capability. The pulsed light excitation signal generated by the input layer is injected only into the first semiconductor laser in the reservoir layer, while the other semiconductor lasers receive the light signal only through the mutual coupling links between adjacent semiconductor lasers. The pulsed light excitation signal generated by the input layer is injected into the first semiconductor laser in the reservoir layer, causing it to generate a self-pulse output. Through the bidirectional delayed mutual coupling links between adjacent semiconductor lasers, each level of semiconductor laser generates a nonlinear dynamic response with historical information association, forming a pulse response network with time memory capability. The output layer includes and Each semiconductor laser is configured in a one-to-one correspondence. Individual optical isolators, The system includes a photodetector, an adder, and a programmable gate array. The input of each optical isolator is connected to the optical output of the corresponding semiconductor laser to ensure unidirectional optical signal transmission. The input of each photodetector is connected to the output of the corresponding optical isolator to convert the output optical signal of the corresponding semiconductor laser into an electrical response signal. The input of the adder is connected to the output of all photodetectors to sequentially concatenate multiple feature vectors extracted from each electrical response signal to generate a fused feature vector. The programmable gate array is connected to the output of the adder to train and generate an output weight matrix based on the fused feature vector, and to perform prediction calculations for chaotic time series.

2. The deep pulse reservoir calculation system based on a semiconductor laser according to claim 1, characterized in that, The front end of the pulsed light excitation signal generated by the input layer is provided with A preheating pulse sequence used to bring the semiconductor laser in the reservoir layer into a stable dynamic operating state.

3. The deep pulse reservoir calculation system based on a semiconductor laser according to claim 1, characterized in that, The semiconductor laser is a laser capable of generating self-pulse output, specifically a common semiconductor laser that generates self-pulse output through external light injection, or a semiconductor laser with an integrated saturable absorber. The output pulse signal is used to characterize the state information of the reservoir.

4. The deep pulse reservoir calculation system based on a semiconductor laser according to claim 1, characterized in that, Optical injection between adjacent semiconductor lasers is achieved through fiber coupling or free-space optical coupling, and all semiconductor lasers at each level are bidirectionally coupled.

5. The deep pulse reservoir calculation system based on a semiconductor laser according to claim 1, characterized in that, The delay time of the delayed light injection link is equal to a positive integer multiple of the interval time of the output pulse of the first semiconductor laser, so that the dynamic response of each semiconductor laser forms a multi-level delay correlation in the time dimension, thereby enhancing the system's ability to retain historical input information.

6. The deep pulse reservoir calculation system based on a semiconductor laser according to claim 1, characterized in that, In the reservoir layer, the output optical signal of the first semiconductor laser simultaneously includes the current pulse excitation signal injected from the input layer, as well as the multi-level historical pulse signals formed by delayed light injection from the subsequent mutually coupled semiconductor lasers, forming a pulse response network with long-range time memory capability within the reservoir layer.

7. The deep pulse reservoir calculation system based on a semiconductor laser according to claim 1, characterized in that, The features extracted by the adder from each electrical response signal include at least one of the following: mean difference, peak difference, standard deviation, energy integral, dynamic range, and rising edge percentage. The mean difference is the difference between the mean response within the pulse segment and the mean response within the gap segment; the peak difference is the difference between the maximum response within the pulse segment and the minimum response within the gap segment. The standard deviation is the standard deviation of the response sequence within the pulse segment; the energy integral is the integral value of the response within the corresponding time window within the pulse segment; the dynamic range is the difference between the maximum and minimum values ​​of the response within a single pulse period; and the rising edge percentage is the ratio of the pulse rising edge duration to the pulse period.

8. The deep pulse reservoir calculation system based on a semiconductor laser according to claim 1, characterized in that, The adder will The feature vectors corresponding to each semiconductor laser are concatenated end-to-end according to the cascaded numbering order of the semiconductor lasers to generate a dimensionally expanded fused feature vector, achieving deep fusion of features in the time and spatial dimensions.

9. A method for calculating a deep pulse reservoir based on a semiconductor laser, characterized in that, The implementation of the deep pulse reservoir computing system based on semiconductor lasers according to any one of claims 1 to 8 includes the following steps: S1: The input layer performs normalization, masking, and pulse coding on the chaotic time series to be predicted to generate a pulsed light excitation signal with a preheating sequence. S2: The pulsed light excitation signal is injected into the first semiconductor laser in the reservoir layer to generate a self-pulse output. Through the bidirectional delay mutual coupling link between adjacent semiconductor lasers, each level of semiconductor laser generates a nonlinear dynamic response with historical information association, forming a pulse response network with time memory capability. S3: The output optical signals of each semiconductor laser are converted into electrical response signals by the photodetector of the output layer. Multidimensional feature extraction is performed on each electrical response signal to obtain multi-feature vectors. Then, the multi-feature vectors are sequentially concatenated by the adder to generate a spatiotemporal fusion feature vector. S4: Based on the fused feature vector of the training set, a linear readout algorithm is used to train and generate the output weight matrix through a programmable gated array. The fused feature vector of the test phase is multiplied with the output weight matrix to obtain the prediction result of the chaotic time series.

Images