High speed optical reservoir computing system for time series signal processing and test method

Abstract

This invention discloses a high-speed optical reservoir computing system and testing method for time-series signal processing, belonging to the field of optical reservoir computing technology. The method includes the following steps: filtering and denoising the dataset using a computer, signal normalization, feature enhancement, time window extraction, data segmentation, and data balancing; importing the processed data into an arbitrary waveform generator; the signal from the arbitrary waveform generator is led out by wires, passes through a T-bias, and is input into the saturable absorption region through a GSG probe; the pulse signal is then amplified through the gain region; the output surface of the gain region is connected to a photodetector via fiber optic coupling, converting the optical signal into an electrical signal which is then imported into an oscilloscope; and the output weight matrix between the output vector and the corresponding label vectors of the input vector is fitted using linear regression methods such as ridge regression. This invention enables ultra-high-speed processing of large-volume time-series signals, such as electrocardiogram signals, significantly reducing the energy consumption of existing processing systems and increasing processing speed.

Description

Technical Field

[0001] This invention relates to the field of optical reservoir computing technology, and more specifically to a high-speed optical reservoir computing system and testing method for time-series signal processing. Background Technology

[0002] In neural networks, the training process involves calculating and minimizing the target loss function by finding the difference between the ideal and actual outputs to optimize network parameters and achieve network convergence. When the network finally performs a prediction task, it achieves the expected results. Artificial neural network training takes various forms, including reinforcement learning, unsupervised learning, and supervised learning. The choice of training method depends on the model and the problem to be solved. In practical applications using supervised learning, collecting or generating a sufficiently large "good" dataset containing known answers is just as important as designing or implementing the neural network itself. Supervised learning requires a labeled dataset (x, t), where x is the example network input and t is the associated known label, providing the true answer. During parameter optimization, backpropagation and gradient descent are used to minimize the network's cost function. The main idea is that after the ANN (Artificial Neural Network) completes the forward propagation process, the error between the network's estimate and the actual value is calculated, and this error is backpropagated from the output layer to the hidden layers, and so on, until it reaches the input layer. During backpropagation, the values ​​of various parameters are adjusted based on the error, and this process is iterated until the network converges.

[0003] For photophysical systems, once training is complete, the entire structure can perform optical signal calculations at the speed of light without requiring additional energy input. In other words, the trained network is passive and can perform optical signal calculations with minimal power consumption.

[0004] In optical neural networks, the output layer uses a photodetector or oscilloscope to detect the output light intensity and calculate the difference between the output and target light intensities. The loss function is defined by the mean squared error, and the amplitude and phase of the entire network are updated using backpropagation and stochastic gradient descent, all performed on a computer. However, implementing backpropagation and performing gradient descent using analog hardware is generally much more difficult than implementing digital hardware. The computer needs an expectation function to simulate the physical process. But finding a physical system that behaves like a specific mathematical function is extremely difficult. Furthermore, the backpropagation algorithm requires calculating the derivative of a nonlinear activation function. For an optical computing system, the activation function itself can be realized through a series of nonlinear optical phenomena, but optically implementing the backpropagation and its derivative is extremely challenging.

[0005] Therefore, the promising optical neural networks for industrial applications should currently utilize only optical hardware for inference and make new predictions on unseen data, while training is entirely done using digital simulation, typically on a computer. The advantage of this approach is that optimal weights can be found using backpropagation or fitting, and then the pre-trained model weights are mapped to the physical system to perform inference. However, such hardware systems that rely on physical phenomena for computation and determine optimal weights through backpropagation will inevitably exhibit defects or distortions, introducing systematic errors. For example, the two output ports of a Mach-Zehnder interferometer (MZI) struggle to achieve an ideal 50:50 beam split, which significantly impacts offline training schemes because the optimal weights found using a noiseless digital system may not be the optimal weights for an error-prone analog system. This problem becomes even more complex in deep networks with many layers: small differences between the physical and analog systems accumulate and amplify with increasing network layers, ultimately severely affecting model performance. A solution to the reality gap problem is to train the neural network using the analog hardware itself, rather than relying on a digital model. However, with current technology, this requires a large number of optical feedback devices.

[0006] In summary, under the current industrial context, there are problems such as the excessive amount of input signal data to be processed, the slow data processing speed of electronic architecture, and the high cost of existing optical chips (requiring a large number of optical feedback devices).

[0007] Based on this, the present invention designs a high-speed optical reservoir computing system and testing method for time-series signal processing to solve the above problems. Summary of the Invention

[0008] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a high-speed optical reservoir computing system and testing method for time-series signal processing.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A test method for a high-speed optical reservoir computing system for time-series signal processing includes the following steps: Step (1): Process the dataset using a computer as follows: 1.1 Filtering and noise reduction; 1.2 Signal normalization; 1.3. Perform feature enhancement on each peak; 1.4. For each heartbeat, a time window is extracted centered on the R peak; 1.5. During data segmentation, the AAMI classification category corresponding to each heartbeat is extracted based on the annotation information, and sample-label pairs are constructed: the input is the heartbeat waveform vector, and the output is the corresponding heartbeat category vector, thereby establishing a one-to-one mapping relationship between the heartbeat waveform vector and the heartbeat category vector. 1.6 Then, perform data balancing on the processed dataset; 1.7 Then, divide the dataset into training and test sets for verification; 1.8. Import the processed data into an arbitrary waveform generator; Step (2): The signal in the arbitrary waveform generator is led out by wires, biased by T, and input into the saturable absorption region through the GSG probe. The pulse signal is then amplified through the gain region. The light output surface of the gain region is connected to the photodetector through fiber optic coupling, and the optical signal is converted into an electrical signal and introduced into the oscilloscope. Step (3): Input the oscilloscope signal into the computer, and fit the output weight matrix between the heartbeat waveform after nonlinear processing by the quantum dot mode-locked laser and the heartbeat category vector corresponding to the heartbeat waveform input into the quantum dot mode-locked laser after computer preprocessing by using linear regression methods such as ridge regression.

[0010] Furthermore, step 1.1 specifically involves removing high-frequency noise, power frequency interference, and baseline drift through filtering.

[0011] Furthermore, step 1.2 specifically involves: using z-score standardization. =x- By adjusting the mean μ and standard deviation σ, the normalized value is made to be within the voltage range of any waveform generator.

[0012] Furthermore, in step 1.5, the input is waveform data of a single heartbeat, represented as a one-dimensional vector of fixed length.

[0013] Furthermore, in step 1.5, the output is the corresponding heartbeat category, represented by the following vector: (0,1 / 2 / 3 / 4) or (0,0,0,0,1).

[0014] Furthermore, the ridge regression formula is: =( y in, Let X be the feature matrix and y be the target variable, representing the optimal weight vector. It is a hyperparameter that controls the intensity of L2 regularization, and I is the identity matrix.

[0015] To better achieve the objectives of this invention, the present invention also provides a high-speed optical reservoir computing system for time-series signal processing, comprising an arbitrary waveform generator, a mode-locked laser, a photodetector, and an oscilloscope; the arbitrary waveform generator is electrically connected to the saturable absorption region of the mode-locked laser, the gain region of the mode-locked laser and the photodetector are connected via optical fiber, and the photodetector is electrically connected to the oscilloscope.

[0016] Furthermore, the sampling rate of the arbitrary waveform generator is matched with the repetition frequency of the mode-locked laser; the mode-locked laser is a GaAs / InAs quantum dot mode-locked laser; and the photodetector is an InGaAs photodetector.

[0017] Furthermore, the physical processes of the passive chip on the mode-locked laser are described by the Haus master equation: ; Where A is the complex amplitude of the pulse, z is the propagation distance, T is the time, α is the linear loss coefficient, β2 is the group velocity dispersion coefficient, γ is the nonlinear coefficient, and i is the imaginary unit.

[0018] Furthermore, the passive chip on the mode-locked laser is packaged on a high-speed copper heat sink, and an external gold block is connected to the gain region of the mode-locked laser via gold wires. A forward bias is applied to the gain region, and a reverse bias is applied to the saturable absorption region.

[0019] Compared with the prior art, the advantages of this invention are as follows: this invention only requires a mode-locked laser, an arbitrary waveform generator, a photodetector and an oscilloscope, which greatly reduces the testing cost and simplifies the testing device and testing method; Compared to other methods that rely on optical feedback optical paths, this invention utilizes the characteristics of the saturable absorption region of a mode-locked laser, and has the advantages of simplicity and high speed. Compared to other neural network structures, this invention uses a reservoir computing structure, which greatly simplifies parameter training and is energy-efficient and highly effective.

[0020] In summary, this invention enables ultra-high-speed processing of large-volume time-series signals, such as electrocardiogram signals, which can greatly reduce the energy consumption of existing processing systems and improve processing speed. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0022] Figure 1 This is a schematic diagram of the high-speed optical reservoir computing system for time-series signal processing according to the present invention; Figure 2 This is a diagram illustrating the test results.

[0023] Figure 3 This is an epitaxial structure for a GaAs / InAs quantum dot mode-locked laser. Detailed Implementation

[0024] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Example 1: In some embodiments, please refer to the accompanying drawings. Figures 1-2 A high-speed optical reservoir computing system for time-series signal processing includes an arbitrary waveform generator, a mode-locked laser, a photodetector, and an oscilloscope. The arbitrary waveform generator (AWG(SDG6052X)) is electrically connected to the saturable absorption region (SA) of the mode-locked laser, the gain region (GAIN) of the mode-locked laser and the photodetector are connected via optical fiber, and the photodetector is electrically connected to the oscilloscope.

[0026] The sampling rate of the arbitrary waveform generator is matched with the repetition frequency of the mode-locked laser; the mode-locked laser is a GaAs / InAs quantum dot mode-locked laser (its epitaxial structure is shown in...). Figure 3 The cavity length is 2000μm and the repetition frequency is 20GHz; the photodetector is an InGaAs photodetector (UPD series, DC-25GHz, spectral range 170-2600nm).

[0027] The physical processes of a mode-locked laser on a passive chip can be described by the Haus master equation: ; Where A is the complex amplitude of the pulse, z is the propagation distance, T is the time, α is the linear loss coefficient, β² is the group velocity dispersion coefficient, γ is the nonlinear coefficient, and i is the imaginary unit. This equation considers the effects of linear loss, group velocity dispersion, and intensity-dependent nonlinear effects on ultrashort pulses.

[0028] To meet the needs of ultra-high-speed computing or communication, the passive chip on the mode-locked laser is packaged on a copper heat sink, and an external gold block is connected in the gain region (GAIN) of the mode-locked laser through gold wires to facilitate the application of forward current through the gold block and the application of reverse bias voltage in the saturable absorption region.

[0029] The specific testing method includes the following steps: Step (1): Download the MIT-BIH dataset from the PhysioNet website and process the dataset using a computer as follows: 1.1 Filtering and noise reduction: High-frequency noise (>40HZ), power frequency interference (>40HZ), and baseline drift (<0.5HZ) are removed by filtering.

[0030] 1.2 Signal Normalization: z-score normalization is adopted. =x- By adjusting the mean μ and standard deviation σ, the normalized value is made to be within the voltage range of the arbitrary waveform generator (AWG).

[0031] 1.3 Feature Enhancement: Perform feature enhancement on each peak.

[0032] 1.4. Extracting a time window: Centered on the R peak for each heartbeat, with a sampling rate of 360Hz, extract a certain time window (e.g., 0.25s before + 0.45s after → 252 sample points).

[0033] 1.5. Next, perform data segmentation, extract the AAMI classification categories (N, S, V, F, Q) corresponding to each heartbeat based on the annotation file (ATR), and construct sample-label pairs: The input is waveform data of a single heartbeat, represented as a one-dimensional vector of fixed length; The output is the corresponding heartbeat category, which can be represented by two vectors: (0,1 / 2 / 3 / 4) or (0,0,0,0,1), thereby establishing a one-to-one mapping relationship between the heartbeat waveform vector and the heartbeat category vector; 1.6 Then, the processed dataset is balanced. Since the MIT-BIH dataset is unbalanced in terms of heart rate (N class is the most, while V and F classes are less), oversampling (random copying) or undersampling (appropriately reducing N class) is performed.

[0034] 1.7 Then, the training set and the test set are divided, and 5-fold cross-validation is used.

[0035] 1.8. Import the processed data into an arbitrary waveform generator; Step (2): The signal in the arbitrary waveform generator is led out by wires, biased by T, and input into the saturable absorption region through the GSG probe. The pulse signal is then amplified through the gain region. The light output surface of the gain region is connected to the photodetector through fiber optic coupling, and the optical signal is converted into an electrical signal and introduced into the oscilloscope. Step (3): Copy the oscilloscope signal into an Excel file and input it into the computer. Use linear regression methods such as ridge regression to fit the output weight matrix between the output vector (the heartbeat waveform after nonlinear processing by the quantum dot mode-locked laser) and the corresponding label vector (heartbeat category vector) of the input vector (the heartbeat waveform input into the quantum dot mode-locked laser after computer preprocessing).

[0036] The core idea of ​​ridge regression is to find an optimal weight vector that minimizes the loss function (e.g., choosing mean squared error), while adding a regularization term (L2 norm) to penalize excessively large weights and prevent overfitting. The ridge regression formula is:

[0037] =( y in, Let X be the feature matrix and y be the target variable, representing the optimal weight vector. It is a hyperparameter that controls the intensity of L2 regularization, and I is the identity matrix.

[0038] This invention uses a quantum dot mode-locked laser as the reservoir and utilizes the physical characteristics of the mode-locked laser to replace the hidden layers of a traditional recurrent neural network (RNN). It uses an arbitrary waveform generator as the input layer and a photodetector and an oscilloscope as the output layer to build a high-speed optical reservoir computing system for time-series signal processing.

[0039] Compared to other electronic or physical implementations of reservoir computing, quantum dot mode-locked lasers exhibit fast nonlinear response, rapidly capturing electrically pumped carriers into discrete energy levels; they also possess lower phase noise and carrier fluctuations, with less noise impact; they demonstrate good high-temperature stability, and the quantum dot bandgap effect further stabilizes their dynamic characteristics; they can output multiple modes and inherently possess a high-dimensional time-domain structure, making them suitable for time-delay mapping (RC). This means that quantum dot mode-locked lasers are a natural high-speed analog nonlinear node array, faster and more energy-efficient than other methods.

[0040] Example 2: The operation process for electrocardiogram (ECG) recognition includes the following steps: Step (1) Download the MIT-BIH dataset from the PhysioNet website and process the dataset using a computer as follows: 1.1 First, noise reduction is performed by filtering to remove high-frequency noise (>40Hz), power frequency interference (>40Hz), and baseline drift (<0.5Hz).

[0041] 1.2 Then, perform signal normalization using z-score standardization: =x- By adjusting μ and σ, the normalized value is made to be within the voltage range of the arbitrary waveform generator (AWG).

[0042] 1.3 Then, feature enhancement is performed on each peak.

[0043] 1.4. Then perform R-peak detection and slicing: For each heartbeat, take the R-peak as the center, sample at a rate of 360Hz, and extract a certain time window (e.g., 0.25s before + 0.45s after → 252 sample points).

[0044] 1.5. The preprocessed ECG signals are segmented, and the AAMI classification category (N, S, V, F, Q) corresponding to each heartbeat is extracted based on the annotation file (ATR), and sample-label pairs are constructed. The input is waveform data of a single heartbeat, represented as a one-dimensional vector of fixed length; The output is the corresponding heartbeat category, which can be represented by two vectors: (0,1 / 2 / 3 / 4) or (0,0,0,0,1), thereby establishing a one-to-one mapping relationship between the heartbeat waveform vector and the heartbeat category vector; 1.6 Then, the processed dataset is balanced. Since the MIT-BIH heart rate is unbalanced (N class is the most, V and F classes are less), oversampling (random copying) or undersampling (appropriately reducing N class) is performed.

[0045] 1.7 Then, divide the training / test sets and use 5-fold cross-validation.

[0046] 1.8. Convert the processed data into a CSV file, copy it to a USB drive, and input it into an arbitrary waveform generator (AWG).

[0047] Step (2): The electrocardiogram signal is drawn from the arbitrary waveform generator by the wire, and a T bias is added to the electrical signal to convert it into a radio frequency signal. The signal is then driven into the saturable absorption region by the GSG probe and a reverse bias is applied. Then, the probe is used to apply a forward current to the gain region through the gold block and gold wire. The current magnitude depends on the threshold current. Here, a current of 1.2A is applied to the gain region. Then, the optical fiber is coupled to the light output point of the ridge of the gain region. The optical fiber is moved back and forth by controlling the computer software. When the power meter reading is between 0 and -10dB, it can be considered aligned. The signal passes through an InGaAs 1300nm band photodetector and a 10 GHz bandwidth oscilloscope to obtain the output waveform and sample 500 points to show the high-dimensional state.

[0048] Step (3): The heartbeat waveform output from the oscilloscope after processing by the quantum dot mode-locked laser is linked with the heartbeat category vector of the corresponding input vector (the heartbeat waveform that will be input into the quantum dot mode-locked laser after computer preprocessing) through the output matrix. The output matrix is ​​obtained by ridge regression fitting.

[0049] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A test method for a high-speed optical reservoir computing system for time-series signal processing, characterized by comprising the following steps: Step (1): Process the dataset using a computer as follows: 1.1 Filtering and noise reduction; 1.2 Signal normalization; 1.

3. Perform feature enhancement on each peak; 1.

4. For each heartbeat, a time window is extracted centered on the R peak; 1.

5. During data segmentation, the AAMI classification category corresponding to each heartbeat is extracted based on the annotation information, and sample-label pairs are constructed: the input is the heartbeat waveform vector, and the output is the corresponding heartbeat category vector, thereby establishing a one-to-one mapping relationship between the heartbeat waveform vector and the heartbeat category vector. 1.6 Then, perform data balancing on the processed dataset; 1.7 Then, divide the dataset into training and test sets for verification; 1.

8. Import the processed data into an arbitrary waveform generator; Step (2): The signal in the arbitrary waveform generator is led out by wires, biased by T, and input into the saturable absorption region through the GSG probe. The pulse signal is then amplified through the gain region. The light output surface of the gain region is connected to the photodetector through fiber optic coupling, and the optical signal is converted into an electrical signal and introduced into the oscilloscope. Step (3): Input the oscilloscope signal into the computer, and fit the output weight matrix between the heartbeat waveform after nonlinear processing by the quantum dot mode-locked laser and the heartbeat category vector corresponding to the heartbeat waveform input into the quantum dot mode-locked laser after computer preprocessing by using linear regression methods such as ridge regression.

2. The test method according to claim 1, characterized in that, Step 1.1 specifically involves removing high-frequency noise, power frequency interference, and baseline drift through filtering.

3. The test method according to claim 1, characterized in that, Step 1.2 specifically involves: using z-score standardization. =x- By adjusting the mean μ and standard deviation σ, the normalized value is made to be within the voltage range of any waveform generator.

4. The test method according to claim 1, characterized in that, In step 1.5, the input is waveform data of a single heartbeat, represented as a one-dimensional vector of fixed length.

5. The test method according to claim 1, characterized in that, In step 1.5, the output is the corresponding heartbeat category, represented by the following vector: (0,1 / 2 / 3 / 4) or (0,0,0,0,1).

6. The test method according to claim 1, characterized in that, The ridge regression formula is: =（ y in, Let X be the feature matrix and y be the target variable, representing the optimal weight vector. It is a hyperparameter that controls the intensity of L2 regularization, and I is the identity matrix.

7. A high-speed optical reservoir computing system for time-series signal processing, used in the test method according to any one of claims 1 to 6, characterized in that, It includes an arbitrary waveform generator, a mode-locked laser, a photodetector, and an oscilloscope; the arbitrary waveform generator is electrically connected to the saturable absorption region of the mode-locked laser, the gain region of the mode-locked laser and the photodetector are connected via optical fiber, and the photodetector is electrically connected to the oscilloscope.

8. The high-speed optical reservoir computing system for time-series signal processing according to claim 7, characterized in that, The sampling rate of the arbitrary waveform generator is matched with the repetition frequency of the mode-locked laser; the mode-locked laser is a GaAs / InAs quantum dot mode-locked laser; and the photodetector is an InGaAs photodetector.

9. The high-speed optical reservoir computing system for time-series signal processing according to claim 7, characterized in that, The physical processes of a mode-locked laser on a passive chip are described by the Haus master equation: ; Where A is the complex amplitude of the pulse, z is the propagation distance, T is the time, α is the linear loss coefficient, β2 is the group velocity dispersion coefficient, γ is the nonlinear coefficient, and i is the imaginary unit.

10. The high-speed optical reservoir computing system for time-series signal processing according to claim 8, characterized in that, The mode-locked laser is on-chip passive chip packaged on a high-speed copper heat sink, and an external gold block is connected in the gain region of the mode-locked laser through gold wires. A forward bias is applied in the gain region and a reverse bias is applied in the saturable absorption region.

Images