Nonlinear dynamic reservoir network time sequence prediction system and implementation method

Abstract

The embodiment of the invention provides a nonlinear dynamic reservoir network time sequence prediction system and an implementation method, and the system comprises a microcontroller which is internally provided with a pulse width modulation (PWM) module and is used for receiving an input signal, carrying out the preprocessing of the input signal, obtaining a preprocessed input signal, and transmitting the preprocessed input signal to an output end of the microcontroller; performing PWM coding on the processed input signal through the PWM module to obtain a PWM coded signal; a plurality of linear resistance-capacitance (R-C) networks wherein each R-C network has a different resistance value and capacitance value; the storage pool layer is used for realizing a storage pool layer with distributed nonlinear characteristics and diversified time dynamic characteristics, all R-C networks in the storage pool are connected in parallel, and the PWM coded signals are refitted to obtain final output signals; and the linear regression network is used for acquiring the final output signal and executing a classification and prediction task of the time sequence data.

Description

Technical Field

[0001] This document relates to the field of computational model technology, and in particular to a nonlinear dynamic reservoir network time series prediction system and implementation method. Background Art

[0002] At the forefront of current research in artificial intelligence and machine learning, Deep Neural Networks (DNNs) have been widely used in multiple disciplines such as computer vision, speech recognition, and image processing. However, with the diversification of computing needs, especially in resource-constrained edge computing environments, the demand for lightweight design and efficient reasoning capabilities of models is becoming increasingly urgent. In this context, Reservoir Computing (RC), as a neural network architecture with economical parameters and simplified training process, is gradually becoming a promising alternative to traditional DNNs architecture. Based on the optimization of Recurrent Neural Networks (RNNs), the RC model effectively processes time series data, greatly reduces the complexity of training, and shows excellent compatibility at the hardware implementation level.

[0003] As an innovative neuromorphic computing paradigm, Physical Reservoir Computing (PRC) makes full use of the dynamic physical properties of electronic devices to achieve highly parallelized signal processing and energy efficiency optimization. Based on the theory of nonlinear dynamical systems, PRC performs complex nonlinear transformations and information encoding operations on input signals through recursively interconnected physical networks. Its notable features are high computational efficiency, simple structure, and support for full analog computing architecture. In the field of brain-inspired computing, the study of PRC systems is becoming a key issue, and its application prospects in high-tech fields such as artificial intelligence, edge computing, and nonlinear dynamical systems have attracted much attention. Compared with traditional digital computing methods, PRC systems provide significant improvements in computing performance and reductions in energy consumption at the theoretical level, providing a new scientific basis and research path for the future development of computing technology.

[0004] For the use of analog circuits such as rotating neuron architecture, and replacing the usual structure of multiple connected nodes by a dynamic system composed of nonlinear nodes with delayed feedback. However, analog circuits are easily affected by factors such as temperature, power supply fluctuations, and component aging, which may lead to unstable system performance and reduced reliability. The accuracy is usually lower than that of digital circuits, which may limit the calculation accuracy of PRC.

[0005] Using analog memristor arrays as the output layer, using their nonlinearity and memory to simulate the behavior of neurons and synapses in the human brain, in order to reduce overall power consumption and hardware overhead, and to achieve a highly integrated memristor storage system. However, this RC system usually has an unadjustable time scale and a relatively fixed nonlinear transfer function, which limits its information processing capabilities. In addition, memristor devices are expensive and are still in the laboratory stage and cannot be commercialized.

[0006] Optoelectronic circuits that precisely control the generation, modulation, transmission, and detection of light waves and construct complex optical paths using optical components to achieve dynamic interference and diffraction of light waves. Such high-precision optical instruments require precise optical components and highly pure materials, as well as special processing technology, and require the construction of complex optical paths, which are prone to errors and have high R&D and manufacturing costs. Optical signals may encounter loss problems such as scattering, diffraction, and absorption during transmission, affecting the accuracy of calculations.

[0007] Using small quantum computers, such as percolation networks (PNNs) of nanoparticles composed of metal nanoparticles deposited on atomically smooth insulating substrates, nanoparticles are new self-assembled nanoscale systems with scale-free and small-world properties; and neuromorphic devices based on memristive nanowire networks (NWNs), which use synaptic-like changes in conductance at nanowire-nanowire intersections to exhibit resistive memory switching in response to electrical input. However, most of these ideas are in the stage of theory and simulation, and have not yet been used to perform actual computing tasks, and are still a long way from practical application.

[0008] In summary, existing PRC systems usually rely on nonlinear semiconductor devices or emerging nanodevices, which are susceptible to challenges such as temperature instability and device variability, i.e., random cycle-to-cycle variations and limited durability. In addition, task mismatch and limited device characteristics reduce the performance and efficiency of PRC systems in various computing tasks. Although random device-to-device (D2D) variations can enhance the diversity of reservoir states, they hinder reproducibility and complicate large-scale inference applications. Summary of the invention

[0009] The purpose of the present invention is to provide a nonlinear dynamic reservoir network time series prediction system and implementation method, aiming to solve the above problems in the prior art.

[0010] The present invention provides a nonlinear dynamic reservoir network time series prediction system, comprising: a microcontroller, a plurality of linear resistor-capacitor RC networks connected to the microcontroller, and a linear regression network:

[0011] A microcontroller having a built-in pulse width modulation (PWM) module, for receiving an input signal, preprocessing the input signal to obtain a preprocessed input signal, and performing PWM encoding on the processed input signal through the pulse width modulation (PWM) module to obtain a PWM encoded signal;

[0012] A plurality of linear resistor-capacitor RC networks, wherein each RC network has a different resistance value and capacitance value; used to realize a reservoir layer with distributed nonlinear characteristics and diversified time dynamic characteristics, each RC network in the reservoir is connected in parallel, and the PWM encoded signal is refitted to obtain a final output signal;

[0013] The linear regression network is used to obtain the final output signal and perform classification and prediction tasks of time series data.

[0014] The present invention provides a method for realizing time series prediction of a nonlinear dynamic reservoir network, which is used in the above-mentioned nonlinear dynamic reservoir network time series prediction system. The method specifically comprises:

[0015] Receiving an input signal through a microcontroller, preprocessing the input signal to obtain a preprocessed input signal, and performing PWM encoding on the processed input signal through the pulse width modulation (PWM) module to obtain a PWM encoded signal;

[0016] A reservoir layer having distributed nonlinear characteristics and diversified time dynamic characteristics is realized by multiple linear resistor-capacitor RC networks, and the PWM coded signal is refitted to obtain a final output signal;

[0017] The final output signal is obtained through a linear regression network to perform classification and prediction tasks of time series data.

[0018] An embodiment of the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the above-mentioned method for implementing nonlinear dynamic reservoir network time series prediction when executed by the processor.

[0019] An embodiment of the present invention further provides a computer-readable storage medium, on which a program for implementing information transmission is stored, and when the program is executed by a processor, the steps of the above-mentioned method for implementing nonlinear dynamic reservoir network time series prediction are implemented.

[0020] By adopting the embodiment of the present invention, a PRC system with advanced performance in multiple computing tasks is realized by optimizing the parameter configuration of a linear resistor-capacitor (RC) network, introducing nonlinear characteristics, simplifying the training process, and performing performance evaluation and system robustness testing. The embodiment of the present invention not only reduces the number of parameter nodes and simplifies the training process, but also constructs a grouping reservoir with highly nonlinear characteristics by accurately adjusting the RC network parameters, thereby showing excellent performance and environmental adaptability in multiple fields such as arrhythmia classification and time series prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0022] Figure 1 is a schematic diagram of a nonlinear dynamic reservoir network time series prediction system according to an embodiment of the present invention;

[0023] Figure 2 A physical schematic diagram of a nonlinear dynamic reservoir network time series prediction system according to an embodiment of the present invention;

[0024] Figure 3 Schematic diagram of comprehensive analysis of nonlinear and time dynamic characteristics of the RC combined storage tank system according to an embodiment of the present invention;

[0025] Figure 4 Schematic diagram of frequency adaptability and prediction results of grouped RC circuit storage pools according to an embodiment of the present invention;

[0026] Figure 5 is a schematic diagram of the inter-tank transformation and temperature sensitivity response of the grouped RC circuit tank system of an embodiment of the present invention;

[0027] Figure 6 is a flow chart of a method for implementing nonlinear dynamic reservoir network time series prediction according to an embodiment of the present invention;

[0028] Figure 7 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0030] System Example

[0031] According to an embodiment of the present invention, a nonlinear dynamic reservoir network time series prediction system is provided. Figure 1 Schematic diagram of a nonlinear dynamic reservoir network time series prediction system according to an embodiment of the present invention. Figure 1 As shown, the nonlinear dynamic reservoir network time series prediction system according to the embodiment of the present invention specifically comprises: a microcontroller 10, a plurality of linear resistor-capacitor RC networks 12 connected to the microcontroller, and a linear regression network 14:

[0032] The microcontroller 10 has a built-in pulse width modulation (PWM) module 100, which is used to receive an input signal, pre-process the input signal to obtain a pre-processed input signal, and perform PWM encoding on the processed input signal through the pulse width modulation (PWM) module 100 to obtain a PWM encoded signal; the microcontroller 10 is specifically used to: use mask processing on the input signal, multiply the input signal with a mask matrix composed of 1 and -1, and generate N virtual nodes. The pulse width modulation (PWM) module performs PWM encoding on the processed input signal within the range of input amplitude 0-T, while maintaining a constant pulse voltage, to obtain a PWM encoded signal, and the output voltage V at the last moment of T is out(T) As shown in Formula 1:

[0033]

[0034] Among them, t is the input width after encoding, τ is determined by hardware, τ is the time unit, and its calculation method is: τ = R*C, R is the size of the resistor, and C is the size of the capacitor.

[0035] A plurality of linear resistor-capacitor RC networks 12, wherein each RC network has different resistance values ​​and capacitance values; used to realize a reservoir layer with distributed nonlinear characteristics and diversified time dynamic characteristics, each RC network in the storage pool is connected in parallel, and the PWM encoded signal is refitted to obtain a final output signal; the plurality of linear resistor-capacitor RC networks 12 are 8 parallel RC networks, each RC network has a fixed resistance of 40kΩ, and capacitance values ​​of 10nF, 25nF, 40nF, 55nF, 70nF, 85nF, 100nF and 115nF, respectively, and the 8 parallel RC networks are specifically used to output final output signals of 8*N reservoir states, wherein N is the amount of data after mask processing during preprocessing.

[0036] The linear regression network 14 is used to obtain the final output signal and perform the classification and prediction tasks of time series data. The linear regression network is specifically used for:

[0037] The final output signal is obtained, and the classification and prediction tasks of the time series data are performed according to Formula 2:

[0038] W out =Y target X T (XX T ) + Formula 2;

[0039] Among them, W out Represents the weight value trained by linear regression, Y target represents the target value of the corresponding prediction task, X T represents the transposed matrix obtained by transposing the matrix output after passing through the RC network, X represents the matrix obtained after passing through the RC network, T represents the transposition of the matrix, and + represents taking the pseudo-inverse of the matrix.

[0040] Specifically, in the embodiment of the present invention, the RC model (Reservoir Computing) generally follows a three-layer architecture: an input layer for receiving and preprocessing data, a reservoir layer where input signals trigger nonlinear dynamics, and an output layer that recombines signals to produce final outputs. Since the input layer is directly connected to the reservoir layer, only weights need to be trained, which can greatly reduce the training weight parameters and calculations. In general physical reservoir computing devices, it is usually necessary to have nonlinear and memory effects. In this product, a resistor-capacitor series method is used, according to the capacitor charging formula:

[0041] V out(T) =V in (1-e -t / τ )

[0042] Voltage and time present an exponential function relationship. The voltage at the previous time point will affect the voltage at the next time point, so it satisfies the memory characteristic. And because τ=RC, when the resistance value R or the capacitance C changes, the voltage-time curve is different, and each circuit has different time characteristics.

[0043] like Figure 2 As shown, the system of the embodiment of the present invention is mainly composed of a microcontroller system board (STM32 microcontroller [MCU] development board, blue part) as the input layer and eight RC circuits installed in parallel on a custom test board. For the 8 parallel RC circuits, the resistance is fixed at 40kΩ, and the capacitance values ​​are 10nF, 25nF, 40nF, 55nF, 70nF, 85nF, 100nF and 115nF respectively. Eight output pins are also provided on the test board for sampling and measuring the capacitor voltage for subsequent linear regression calculation. The eight resistor-capacitor series circuits have different parameters, thereby increasing the diversified time dynamic response.

[0044] Since RC circuits are often regarded as linear systems, the embodiment of the present invention introduces pulse width modulation coding (PWM) at the input layer to solve this limitation and introduce necessary nonlinearity into the RC circuit. Among them, the input signal comes from different tasks. For example, in the arrhythmia detection task, the input signal is the heartbeat data of different patients in different time periods. When preprocessing the input signal, mask processing is used to multiply the input signal with a mask matrix composed of 1 and -1 to generate N virtual nodes. After preprocessing, the input signal is converted into a PWM signal through the PWM module inside the microcontroller. The characteristic of this signal is that the magnitude of the input voltage is mapped to the input time t, so the larger the voltage, the longer the input voltage time. For example, when the input signal amplitude is encoded as 0~3.3V, it will be mapped to 0~T after PWM. Specifically, the input amplitude is encoded in the range of 0-T, while maintaining a constant pulse voltage of 3.3V (MCU power supply voltage). In this case, the output voltage at the last moment of T can be expressed as:

[0045]

[0046] Among them, t is the input width after encoding. In this way, each input signal will generate 8 (i.e. 8 RC circuits) × N (N is the amount of data after mask processing) reservoir states, which are then fed into the linear regression network for calculation, thereby performing time series classification or prediction tasks.

[0047] The linear regression that follows is a simple calculation used to calculate the weight value in the prediction equation. The formula is as follows: out =Ytarget X T (XX T ) + In this way, the weight value can be obtained.

[0048] In order to solve the above technical problems, the present invention relates to an efficient computing model, specifically a physical reservoir computing (PRC) system, which aims to optimize the computing architecture of miniaturization, high efficiency, and fast reasoning, and adapt to resource-constrained environments. The PRC system of the embodiment of the present invention significantly reduces the number of parameter nodes and simplifies the training process. The following is a detailed technical implementation plan:

[0049] Reservoir network composition: The PRC system of the embodiment of the present invention is composed of eight unique linear resistor-capacitor (RC) networks, each network having different resistance and capacitance parameter values, realizing a device-to-device (D2D) multi-performance combination.

[0050] Introduction of nonlinear characteristics: The built-in pulse width modulation (PWM) module of the microcontroller is used to perform PWM encoding on the input signal, thereby introducing nonlinear characteristics and enhancing the computing power of the system.

[0051] Training process optimization: By inputting the PWM encoded signal into the linear regression network to perform the classification and prediction tasks of time series data, the present invention uses a simplified regression operation to replace the traditional back propagation algorithm, which greatly shortens the training cycle and reduces resource consumption.

[0052] Parameter tuning and characteristic configuration: By precisely adjusting the resistance and capacitance parameters in the RC network, the distributed nonlinearity and diversified temporal dynamic characteristics of the reservoir are achieved, and a grouped reservoir with highly nonlinear characteristics is constructed.

[0053] Performance Evaluation: Our PRC system achieves state-of-the-art performance on multiple computational tasks, including but not limited to arrhythmia classification, NARMA2 time series prediction, Mackey-Glass time series analysis, and Henon map prediction, with minimal reservoir size.

[0054] System robustness test:

[0055] In order to demonstrate the robustness of the present invention, the following test procedure was implemented:

[0056] Device preparation: Two RC circuit devices with the same parameters but different manufacturing processes were prepared and marked as device A and device B respectively.

[0057] Weight training and migration: Under room temperature (RT), linear regression training is performed on device A to obtain reservoir weights.

[0058] Performance verification: The weights obtained from device A training are applied to device B, and performance tests are performed at RT and a high temperature environment of 85°C.

[0059] The test results show that device B can achieve accurate predicted performance under different environmental conditions, confirming the high robustness and environmental adaptability of this PRC system.

[0060] The above technical solution of the embodiment of the present invention is further described below with reference to the example diagram.

[0061] 1. Arrhythmia detection task

[0062] The types of heart diseases are divided into four categories (A, L, V and N) according to the standards provided by the American Association of Medical Instrumentation (AAMI), and are trained and tested using the MIT-BIH arrhythmia database established by the Massachusetts Institute of Technology. The MIT-BIH dataset contains electrocardiogram (ECG) records from 48 different subjects. First, 750 heartbeat data segments are selected from each of the four categories and randomly connected to form a long ECG waveform sequence of 3000 heartbeat segments. Then, the data is resampled at a frequency of 80Hz (input data interval = 12.5ms) and normalized. The preprocessing process is as follows: Figure 3 As shown in a and b.

[0063] For this classification task, the present invention uses the D2D (device-to-device) method (8 RC circuits are connected in parallel to form a reservoir layer), and uses virtual nodes, the number of nodes is set to 5, and the mask data interval is set to T = 2.5ms. Then the reservoir state of 40 nodes is applied to the output layer (a 41×4 network), and then the classified data is generated at each time step. When the output of the corresponding neuron in the readout layer reaches the maximum value within the set time window, we consider the classification to be correct, such as Figure 3 The recognition results of the target signal and output signal of the four types of labels are shown in Figure 3 As shown in the confusion matrix de in , d is the output result of 8 identical RC circuits connected in parallel, the time constant is set to 2.2ms, the nonlinear coefficient is set to 1.8, and each RC sub-reservoir has a different mask sequence; on this basis, the present invention further proposes to connect 8 RC circuits with different parameters in parallel, such as Figure 3The output result can be seen from e in the figure. When 8 RC circuits with different parameter values ​​are used, the output signal can fit the target signal more effectively and produce a peak when the corresponding heartbeat occurs. Finally, the embodiment of the present invention achieved a 94% accuracy rate in classifying these four types of arrhythmias. In contrast, when the input signal is directly input into the linear regression network, the recognition accuracy rate is only 25%, while when 8 identical sub-reservoirs are used, the final accuracy rate is less than 67%. The recognition accuracy comparison results are shown in Figure 2. Figure 3 As shown in f.

[0064] These results emphasize the feasibility and effectiveness of the grouped pooling proposed in the embodiments of the present invention, which has a wide range of nonlinear and temporal characteristics and can process complex time series signals with high accuracy.

[0065] 2. Versatility and applicability of grouped RC circuits

[0066] Real-world applications, such as predicting motion trajectories, often require the ability to handle a wide range of speed conditions, which in turn requires the PRC system to be highly adaptable. To simulate such a scenario, a modified Henon map prediction task was implemented, in which the signal frequency varied from 10 Hz to 50 Hz, and each segment with a different frequency consisted of 100 time steps. This variation is intended to test the ability of the reservoir to manage different temporal dynamics. Reservoirs with fixed temporal dynamics are generally not suitable for such tasks. For example, in the experiment, a reservoir consisting of 8 parallel, identical sub-reservoirs, each with a t of 400 μs and a different mask sequence, can predict a 50 Hz signal well. However, as Figure 4 As shown in Figures a and d, this setup performs poorly for a 10 Hz signal. This illustrates the limitations of using a reservoir with a narrow range of time constants, which is effective at high frequencies but ineffective at slower dynamics. In contrast, another configuration using parallel reservoirs with a longer t (4.6 ms) is able to accurately predict the 10 Hz signal, as shown in Figure 3. Figure 4 However, it cannot effectively process 50Hz signals, which again highlights the limitations of a single device in processing dynamic data.

[0067] The reservoir optimized by the embodiments of the present invention provides a wide range of time constants, making it possible to process multiple time scales simultaneously. Figure 4 As shown in Figures c and f, this configuration allows the system to accurately predict 10Hz and 50Hz signals within the same framework without any feedback control. Moreover, this broad adaptability can be further enhanced by configuring the t-distribution to span a wider range, thus providing an efficient solution to complex multi-frequency tasks.

[0068] 3. Robust performance of PRC system

[0069] In order to verify the robustness of the embodiment of the present invention, two RC reservoir devices were fabricated, labeled as circuit A and circuit B, and they have the same parameters. Figure 5 As shown in a in Figure 1. In order to simulate the maximum C2C ​​(copy-to-copy) variability, circuit A is assembled using machine soldering, while circuit B is assembled using manual soldering technology. Therefore, although the same parameters are set, the different processes also lead to system parameter differences. Here, circuit A is used as a reference, and linear regression is used to train and read out the layer weights at room temperature (RT). The trained weights are then applied to circuit B, and circuit B is tested at RT and 85°C respectively.

[0070] Figure 5 Figures bd show the prediction performance of circuit B for three time series tasks (NARMA2, Mackey-Glass, and Henon map) under RT. It can be seen that the predicted output (blue curve) is closely consistent with the actual situation (gray curve), indicating that when the weights trained by circuit A act on circuit B, circuit B can effectively capture complex temporal dynamics. Figure 5 Figure 4 shows the performance of Circuit B performing the same task at 85°C. The predicted output (red curve) still matches the actual situation well, indicating that Circuit B maintains stable and accurate performance even at high temperatures. Figure 5 h in Figure 1 compares the relative error of the embodiment of the present invention in three prediction tasks under C2C (copy-to-copy), 80°C and the combination of the above two conditions, with the memristor and ReLU network. It can be seen that the embodiment of the present invention achieves the lowest relative error in three tasks and various working environments. This shows that the method of the embodiment of the present invention ensures the stable performance of the present invention under different tasks and temperature changes, and verifies the robustness of the present invention.

[0071] In summary, the physical reservoir system (PRC) adopted in the embodiment of the present invention introduces the nonlinear characteristics necessary for the reservoir network by utilizing pulse width modulation coding (PWM), so that the reservoir can better capture complex time dynamic information, such as chaotic sequences. By modulating the pulse width, a richer dynamic response is generated, thereby enhancing the performance of classification and prediction. Secondly, by directly using analog PWM, the need for a digital-to-analog converter (ADC) can be eliminated, avoiding problems such as high power consumption, high latency, and affected system performance, achieving low power consumption and low cost, and having more commercial advantages. In addition, the embodiment of the present invention can flexibly adapt to multiple time scales and adopt a wider range of nonlinearities and time constants, thereby improving the performance and versatility of the RC system. By covering a wider range, the reservoir can effectively handle a variety of tasks without modulation for the task. The embodiment of the present invention uses a digital readout layer to create a hybrid system, in which Kirchhoff's laws are allowed to be directly applied because the output is already in voltage form, without the need for resistors to convert current into voltage. In addition, the device is robust to parameter changes and temperature changes, so the full analog circuit of the PRC system of the embodiment of the present invention will become an ideal candidate for large-scale, low-power, high-speed computing edge computing applications in actual scenarios.

[0072] Method Embodiment

[0073] According to an embodiment of the present invention, a method for implementing nonlinear dynamic reservoir network time series prediction is provided, which is used in the above-mentioned nonlinear dynamic reservoir network time series prediction system. Figure 6 is a flow chart of a method for implementing nonlinear dynamic reservoir network time series prediction according to an embodiment of the present invention. Figure 6 As shown, the embodiments according to the present invention specifically include:

[0074] Step S601, receiving an input signal through a microcontroller, preprocessing the input signal to obtain a preprocessed input signal, and performing PWM encoding on the processed input signal through the pulse width modulation (PWM) module to obtain a PWM encoded signal; specifically comprising:

[0075] Applying mask processing to the input signal, multiplying the input signal by a mask matrix consisting of 1 and -1 to generate N virtual nodes;

[0076] The pulse width modulation (PWM) module performs PWM encoding on the processed input signal within the range of input amplitude 0-T, while maintaining a constant pulse voltage to obtain a PWM encoding signal. The output voltage V at the last moment of T out(T) As shown in Formula 1:

[0077]

[0078] Among them, t is the input width after encoding, τ is determined by hardware, τ is the time unit, and its calculation method is: τ = R*C, R is the size of the resistor, and C is the size of the capacitor.

[0079] Step S602, realizing a reservoir layer with distributed nonlinear characteristics and diversified time dynamic characteristics through multiple linear resistor-capacitor RC networks, refitting the PWM encoded signal to obtain a final output signal; specifically comprising: outputting a final output signal of 8*N reservoir states, wherein N is the amount of data after mask processing during preprocessing;

[0080] Step S603, obtaining the final output signal through a linear regression network, and performing the classification and prediction task of time series data. Specifically including:

[0081] The final output signal is obtained, and the classification and prediction tasks of the time series data are performed according to Formula 2:

[0082] W out =Y target X T (XX T ) + Formula 2;

[0083] Among them, W out Represents the weight value trained by linear regression, Y target represents the target value of the corresponding prediction task, X T represents the transposed matrix obtained by transposing the matrix output after passing through the RC network, X represents the matrix obtained after passing through the RC network, T represents the transposition of the matrix, and + represents taking the pseudo-inverse of the matrix.

[0084] The embodiment of the present invention is a method embodiment corresponding to the above-mentioned system embodiment. The specific processing of each step can be understood by referring to the description of the method embodiment, which will not be repeated here.

[0085] Device Example 1

[0086] An embodiment of the present invention provides an electronic device, such as Figure 7 As shown, it includes: a memory 70, a processor 72, and a computer program stored in the memory 70 and executable on the processor 72. When the computer program is executed by the processor 72, the steps described in the method embodiment are implemented.

[0087] Device Example 2

[0088] An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by the processor 72, the steps described in the method embodiment are implemented.

[0089] The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk or optical disk, etc.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A nonlinear dynamic reservoir network time series prediction system, characterized in that: include: A microcontroller, a plurality of linear resistor-capacitor RC networks connected to the microcontroller, and a linear regression network: A microcontroller having a built-in pulse width modulation (PWM) module, for receiving an input signal, preprocessing the input signal to obtain a preprocessed input signal, and performing PWM encoding on the processed input signal through the pulse width modulation (PWM) module to obtain a PWM encoded signal; A plurality of linear resistor-capacitor RC networks, wherein each RC network has a different resistance value and capacitance value; used to realize a reservoir layer with distributed nonlinear characteristics and diversified time dynamic characteristics, each RC network in the reservoir is connected in parallel, and the PWM encoded signal is refitted to obtain a final output signal; The linear regression network is used to obtain the final output signal and perform classification and prediction tasks of time series data.

2. The method according to claim 1, characterized in that: The microcontroller is specifically used for: applying mask processing to the input signal, multiplying the input signal with a mask matrix consisting of 1 and -1, and generating N virtual nodes.

3. The method according to claim 1, characterized in that The microcontroller is specifically used to: perform PWM encoding on the processed input signal within the range of input amplitude 0-T through the pulse width modulation PWM module, while maintaining a constant pulse voltage to obtain a PWM encoding signal, and the output voltage V at the last moment of T out(T) As shown in Formula 1: Among them, t is the input width after encoding, τ is expressed as a time unit, τ=R*C, R is the size of the resistor, and C is the size of the capacitor.

4. The method according to claim 1, characterized in that: The multiple linear resistor-capacitor RC networks are 8 parallel RC networks, the resistance of each RC network is fixed at 40kΩ, and the capacitance values ​​are 10nF, 25nF, 40nF, 55nF, 70nF, 85nF, 100nF and 115nF respectively. The 8 parallel RC networks are specifically used to output final output signals of 8*N reservoir states, wherein N is the amount of data after mask processing during preprocessing.

5. The method according to claim 1, characterized in that The linear regression network is specifically used for: The final output signal is obtained, and the classification and prediction tasks of the time series data are performed according to Formula 2: W out = Y target X T (XX T ) + Formula 2; Among them, W out Represents the weight value trained by linear regression, Y target represents the target value of the corresponding prediction task, X T represents the transposed matrix obtained by transposing the matrix output after passing through the RC network, X represents the matrix obtained after passing through the RC network, T represents the transposition operation of the matrix, and + represents taking the pseudo-inverse of the matrix.

6. A method for realizing time series prediction of nonlinear dynamic reservoir network, characterized in that: A nonlinear dynamic reservoir network time series prediction system for use in any one of claims 1 to 5, wherein the method specifically comprises: Receiving an input signal through a microcontroller, preprocessing the input signal to obtain a preprocessed input signal, and performing PWM encoding on the processed input signal through the pulse width modulation (PWM) module to obtain a PWM encoded signal; A reservoir layer having distributed nonlinear characteristics and diversified time dynamic characteristics is realized by multiple linear resistor-capacitor RC networks, and the PWM coded signal is refitted to obtain a final output signal; The final output signal is obtained through a linear regression network to perform classification and prediction tasks of time series data.

7. The method according to claim 6, characterized in that Receiving an input signal through a microcontroller, preprocessing the input signal to obtain a preprocessed input signal, and performing PWM encoding on the processed input signal through the pulse width modulation (PWM) module to obtain a PWM encoded signal specifically includes: Applying mask processing to the input signal, multiplying the input signal by a mask matrix consisting of 1 and -1 to generate N virtual nodes; The pulse width modulation (PWM) module performs PWM encoding on the processed input signal within the range of input amplitude 0-T, while maintaining a constant pulse voltage to obtain a PWM encoding signal. The output voltage V at the last moment of T out(T) As shown in Formula 1: t is the input width after encoding, τ is expressed as a time unit, τ=R*C, R is the size of the resistor, and C is the size of the capacitor.

8. The method according to claim 6, characterized in that Refit the PWM coded signal to obtain the final output signal specifically includes: The final output signal of 8*N reservoir states is output, where N is the amount of data after mask processing during preprocessing; The final output signal is obtained through a linear regression network, and the classification and prediction tasks of time series data are specifically included: Obtain the final output signal and perform the classification and prediction tasks of time series data according to formula 2: W out = Y target X T (XX T ) + Formula 2; Among them, W out Represents the weight value trained by linear regression, Y target represents the target value of the corresponding prediction task, X T represents the transposed matrix obtained by transposing the matrix output after passing through the RC network, X represents the matrix obtained after passing through the RC network, T represents the transposition of the matrix, and + represents taking the pseudo-inverse of the matrix.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the method for realizing time series prediction of a nonlinear dynamic reservoir network as described in any one of claims 6 to 8 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an information transmission implementation program, and when the program is executed by the processor, the steps of the nonlinear dynamic reservoir network time series prediction implementation method according to any one of claims 6 to 8 are implemented.

Images