Reservoir system based on multiple random nanowire networks
By using a concatenation of multiple random nanowire networks and graph theory methods, combined with the ridge regression algorithm, a high-dimensional vector processing system is constructed. This solves the problem of low prediction accuracy of single-layer nanowire networks and achieves richer nonlinear behavior and higher time series prediction accuracy.
Patent Information
- Application Number
- CN202310715854.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-16
- Publication Date
- 2025-11-28
- Estimated Expiration
- 2043-06-16
AI Technical Summary
Existing single-layer nanowire networks cannot generate rich nonlinear behavior when predicting time series, resulting in low prediction accuracy.
Design a reservoir system based on multiple random nanowire networks, including an input layer, a random nanowire network layer, a training layer, and an output layer. By connecting multiple nanowire networks in series and using graph theory methods, the conductivity variation characteristics of memristor components are utilized, and a ridge regression algorithm is combined to generate high-dimensional vectors for data processing, thereby realizing a single-input-multiple-output reservoir network structure.
This improves the nonlinear behavior complexity and predictive ability of nanowire networks, enhancing the accuracy and fidelity of time series prediction.
Smart Images
Figure CN116757244B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of time series prediction, in particular to a reserve pool system based on a multiple random nanowire network. BACKGROUND
[0002] With the progress of science and technology, devices with memristive characteristics have realized artificial synapses, including resistive random access memory, and currently there are also modules that use memristor components to simulate the functions of neurons and synapses. In a large-scale cross array, a memristive device is organized to realize artificial neural networks and neuromorphic computing. However, the hardware architecture of a regular array of memristor components is very different from the inherent structural complexity of the human brain, and this method cannot simulate the topological structure and emergency behavior of biological neural networks. Therefore, another hardware neural network based on memristive synaptic devices has attracted widespread attention, and the nanowire network developed in recent years is a unique representative.
[0003] Traditional single-layer nanowire networks implement reserve pool computing, such as inputting voltage data, mapping input voltage data to different node voltages in the network, and analyzing the changes in different node voltages in the network in the time domain to process time series information. As the input voltage changes, the node voltage in the nanowire network changes constantly, and the change in the voltage of each node in the network depends on the historical input voltage. When the nanowire network is stimulated by an external voltage, the voltage distribution of different nodes in the network has differences, and the trend of data change can be predicted by analyzing the changes in different node voltages over time. The node voltage state value of each node at each time is read for subsequent data analysis and processing, and finally all the read node voltage state values are input to the network output layer for processing. However, single-layer nanowire networks cannot produce rich nonlinear behavior, and the nanowire network is insufficient in predicting time series and has low prediction accuracy.
[0004] The disadvantages of the prior art are that it cannot produce rich nonlinear behavior, and it is insufficient in predicting time series and has low prediction accuracy. SUMMARY
[0005] The main purpose of the present application is to provide a reserve pool system based on a multiple random nanowire network composed of memristor components, to provide more rich nonlinear behavior, and to improve the ability of nanowire network reserve pool computing in predicting time series.
[0006] To achieve the above purpose, the technical scheme of the present application is as follows:
[0007] A reservoir system based on a multi-random nanowire network includes an input layer, a random nanowire network layer, a training layer, and an output layer. The key feature is that the input layer collects vector data sets with time-series attributes, and these vector data sets are normalized and transformed into a voltage input matrix W. in The random nanowire network layer is provided with N layers of nanowire network, which are connected in series.
[0008] The input layer output is connected to the first nanowire network layer, and the input layer transmits the voltage input matrix W one by one. in A voltage is applied to the first nanowire network layer, which stimulates the input nodes of the first nanowire network layer and randomly generates output nodes in the first nanowire network layer. The number of output nodes of the first nanowire network layer is proportional to the number of input nodes of the first nanowire network layer. After being stimulated by the voltage, the first nanowire network layer forms a conductive path between the input nodes and the output nodes. The first nanowire network layer completes data transmission and generates vector X1.
[0009] The ratio of the number of output nodes to the number of input nodes of the first nanowire network layer is a positive integer, and the larger the ratio, the better the storage effect and the higher the fidelity of the nanowire network.
[0010] Within the Kth nanowire network layer, there is a one-to-one correspondence between input and output nodes, where K ≥ 2; in adjacent nanowire network layers, there is a one-to-one correspondence between the output node of the previous layer and the input node of the next layer; the first nanowire network layer transmits the vector X1 to the training layer and the second nanowire network layer, and the second nanowire network layer generates vector X2 after obtaining vector X1; the second nanowire network layer transmits the vector X2 to the training layer and the third nanowire network layer, and the third nanowire network layer generates vector X3 after obtaining vector X2, until the Nth nanowire network layer obtains vector X... N-1 The generated vector X N The Nth nanowire network layer transmits the vector X. N The training layer is given training vectors X1, X2, ..., X... N A high-dimensional vector X is then generated; the output of the training layer is connected to the input of the output layer, the training layer transmits vector X to the output layer, and the output layer outputs the weight matrix W. out ;
[0011] The high-dimensional vector X = vector X1 + vector X2 + ... + vector X N The vectors X1, X2, ... and vector X N The dimensions are equal.
[0012] The output vector X of the training layer is a stack of data vectors X1, X2, and X3, not a vector sum, thus increasing the dimension.
[0013] The input data in the random nanowire network layer is mapped into a voltage vector in part of the nodes in the first layer nanowire network layer. The voltage vector of the node in the previous layer is used as the input of the next layer by taking advantage of the characteristics of the nanowire network allowing the spatiotemporal processing of multiple input signals, and the voltage vector of each layer is trained by a training layer to generate a high-dimensional vector X, and finally the high-dimensional vector X is input into the output layer for processing, thereby designing a reserve pool network structure with single-input-multiple-output and multiple-input-multiple-output.
[0014] As a preferred: the dynamic behavior of the random nanowire network layer is realized by graph theory, specifically as follows:
[0015] When the random nanowire network layer is stimulated by voltage, a conductive path is formed on the two outer layers wrapped with insulating polyvinylpyrrolidone silver nanowires, and the current on the conductive path is:
[0016] I(t)=[G low(1-g(t)) +G hight(g(t)) ]V(t)
[0017] Wherein, G low , G hight represent the minimum and maximum conductance of the memristive edge, g(t) represents the conductance of the memristive edge, and V(t) represents the stimulation voltage;
[0018] The change of the conductance of the memristive edge is represented by a voltage-controlled potential-inhibition rate balance equation:
[0019] dg(t) / dt=ω a (1-g(t))-ω b g(t)
[0020] Wherein, ω a , ω b represent the inhibition coefficient and the enhancement coefficient controlling the change of the conductance;
[0021]
[0022]
[0023]
[0024] Wherein, represents the historical memory state of the network;
[0025] represents the process of the network changing constantly under voltage stimulation.
[0026] The random nanowire network is represented as a weighted directed graph by using the method of graph theory, each intersection in the nanowire network represents a nanowire node, and the connection line between the nanowire nodes represents the connection between the nanowire nodes, which simplifies the calculation and network complexity and is more suitable for the internal connection of the reserve pool calculation.
[0027] As preferred: the input layer collects vector data sets with time series attributes, including but not limited to: voltage, current, temperature, humidity, displacement vector sets changing with time series, and the vector data is converted into a voltage input matrix W after normalization in , the normalization is realized by difference standardization, and the specific process is as follows:
[0028]
[0029] Wherein, V is the current data, V min is the minimum value of the data, V max is the maximum value of the data, and V' is the normalized data.
[0030] The data normalization uniformly converts the vector data with time series attributes into pulse data in the interval of 0-1, reduces the gap between the data, and accelerates the training process of the network.
[0031] As preferred: the output weight matrix W out is obtained by ridge regression algorithm, and the specific calculation is as follows:
[0032] W out =Y target X T (XX T +γI) -1
[0033] Where Y target is the prediction target, X is the high-dimensional vector generated by the training layer, I is the unit matrix, the symbol "T" represents matrix transposition, and the symbol "-1" represents matrix inversion; γ is the regularization coefficient, and the value is γ=10 -8 .
[0034] As preferred: the output end of the output layer is also connected with data prediction, and the data prediction output Y' is calculated as follows:
[0035] Y′=W out X.
[0036] As preferred: the number of nanowire network layers increases by one layer for every increase of 10 input nodes in the first nanowire network layer, and the number of nanowire network layers is ≥2.
[0037] The beneficial effects of this invention are: it provides richer nonlinear behavior for the reservoir computing system, improves the connectivity of network nodes and the randomness of node distribution, enhances the complexity of the network, and at the same time improves the ability of the reservoir computing based on nanowire networks to predict time series, thereby improving the actual prediction effect, enhancing the storage effect of the reservoir system and improving the fidelity. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the structure of the present invention;
[0039] Figure 2 This is a diagram of a single-layer nanowire structure.
[0040] Figure 3 This is a network structure diagram of adjacent nanowires. Detailed Implementation
[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0042] like Figure 1 As shown, a reservoir system based on a multi-random nanowire network includes an input layer, a random nanowire network layer, a training layer, and an output layer. The input layer collects one data point with time-series attributes within one time slice, and the vector data is normalized and converted into a voltage input matrix W. in Voltage input matrix W in The dimension is 1; the random nanowire network is provided with 2 layers of nanowire network, which are connected in series.
[0043] The input layer output is connected to the first nanowire network layer, and the input layer transmits the voltage input matrix W one by one. in A voltage is applied to the first nanowire network layer, which stimulates the input nodes of the first nanowire network layer and randomly generates output nodes in the first nanowire network layer. The number of output nodes of the first nanowire network layer is proportional to the number of input nodes of the first nanowire network layer. In this embodiment, 15 output nodes are randomly generated. After being stimulated by the voltage, the first nanowire network layer forms a conductive path between the input nodes and the output nodes. The first nanowire network layer completes data transmission and generates a 15-dimensional vector X1.
[0044] like Figure 3As shown, in the two adjacent nanowire network layers, the output nodes of the upper layer and the input nodes of the lower layer correspond one-to-one, and the input nodes and the output nodes inside the second nanowire network layer correspond one-to-one; the first nanowire network layer transmits the vector X1 to the training layer and the second nanowire network layer, and the second nanowire network layer generates a 15-dimensional vector X2 after obtaining the vector X1; the second nanowire network layer transmits the vector X2 to the output layer, and the output layer generates a 30-dimensional vector X after training the vector X1 and the vector X2; the output end of the training layer is connected to the input end of the output layer, the training layer transmits the vector X to the output layer, and the output layer outputs the weight matrix W out
[0045] The 30-dimensional vector X is a data stack of the vector X1 and the vector X2, the dimension of the vector X = the dimension of the vector X1 + the dimension of the vector X2, and the dimensions of the vector X1 and the vector X2 are equal.
[0046] The input data in the random nanowire network layer is mapped into a voltage vector in part of the nodes in the first layer nanowire network layer. By using the characteristics of the nanowire network that allows the spatiotemporal processing of multiple input signals, the voltage vector of the node of the upper layer is used as the input of the next layer, the voltage vector of each layer is trained by the training layer to generate a high-dimensional vector X, and finally the high-dimensional vector X is input to the output layer for processing, thereby designing a reserve pool network structure with single-input-multiple-output and multiple-input-multiple-output.
[0047] As shown in Figure 2 , the dynamic behavior between the random nanowire network layers is realized by using graph theory, and the specific implementation is as follows:
[0048] When the random nanowire network layer is stimulated by voltage, a conductive path is formed on the two outer layers of the insulating polyvinylpyrrolidone silver nanowire, and the current on the conductive path is:
[0049] I(t)=[G low(1-g(t)) +G hight(g(t)) ]V(t)
[0050] Where G low , G hight represent the minimum and maximum conductance of the memristive edge, g(t) represents the conductance of the memristive edge, and V(t) represents the stimulation voltage.
[0051] The change of the conductance of the memristive edge is represented by a voltage-controlled potential-inhibition rate balance equation:
[0052] dg(t) / dt=ω a (1-g(t))-ω b g(t)
[0053] Where ω a , ω b The inhibition coefficient and the enhancement coefficient represent the control of the conductance change;
[0054]
[0055]
[0056]
[0057] wherein, The historical memory state of the network is represented;
[0058] The process of the network changing under voltage stimulation is represented.
[0059] By using the method of graph theory, the random nanowire network is represented as a weighted directed graph, each intersection in the nanowire network represents a nanowire node, and the connection line between the nodes of the adjacent two layers of nanowire network represents the connection between the nanowire nodes, which simplifies the calculation and network complexity, and is more suitable for the internal connection of the reserve pool calculation.
[0060] The input layer collects vector data sets with time series attributes, including but not limited to: voltage, current, temperature, humidity, displacement vector sets changing with time series, and the vector data is converted into a voltage input matrix W in after normalization, and the normalization is realized by difference standardization, which is as follows:
[0061]
[0062] wherein, V is the current data, V min is the minimum value of the data, V max is the maximum value of the data, and V' is the normalized data.
[0063] Data normalization uniformly converts vector data with time series attributes to the interval 0-1, reduces the gap between data, and accelerates the training process of the network.
[0064] The output weight matrix W out is obtained by the ridge regression algorithm, and the specific calculation is as follows:
[0065] W out =Y target C T (XX T +γI) -1
[0066] wherein Y target is the prediction target, X is the vector generated by the training layer, I is the unit matrix, the symbol "T" represents matrix transposition, and the symbol "-1" represents matrix inversion; γ is the regularization coefficient, and the value is γ=10 -8 .
[0067] The prediction result Y' is calculated as follows:
[0068] Y' = W out X.
[0069] The number of nanowire network layers is increased by one layer for every 10 input nodes in the first nanowire network layer, and the number of nanowire network layers is ≥ 2.
[0070] In this embodiment, the voltage input matrix W in The dimension = 1, only one voltage stimulates the input nodes of the first nanowire network, and the first nanowire network generates one input node; the number of input nodes ≤ 10, and the number of nanowire network layers is not less than two, so two layers of nanowire networks are selected in this embodiment.
Claims
1. A reserve pool system based on multiple random nanowire network, comprising an input layer, a random nanowire network layer, a training layer and an output layer, characterized in that: The input layer collects a vector data set with time series attribute, and the vector data set is converted into a voltage input matrix W through normalization processing in The random nanowire network layer is provided with N layers of nanowire networks which are connected in series. The input layer output is connected to the first nanowire network layer, and the input layer transmits the voltage input matrix W one by one. in A voltage is applied to the first nanowire network layer, which stimulates the input nodes of the first nanowire network layer and randomly generates output nodes in the first nanowire network layer. The number of output nodes of the first nanowire network layer is proportional to the number of input nodes of the first nanowire network layer. After being stimulated by the voltage, the first nanowire network layer forms a conductive path between the input nodes and the output nodes. The first nanowire network layer completes data transmission and generates vector X1. Within the Kth nanowire network layer, there is a one-to-one correspondence between input and output nodes, where K ≥ 2; in adjacent nanowire network layers, there is a one-to-one correspondence between the output node of the previous layer and the input node of the next layer; the first nanowire network layer transmits the vector X1 to the training layer and the second nanowire network layer, and the second nanowire network layer generates vector X2 after obtaining vector X1; the second nanowire network layer transmits the vector X2 to the training layer and the third nanowire network layer, and the third nanowire network layer generates vector X3 after obtaining vector X2, until the Nth nanowire network layer obtains vector X... N-1 The generated vector X N The Nth nanowire network layer transmits the vector X. N The training layer is given training vectors X1, X2, ..., X... N A high-dimensional vector X is then generated; the output of the training layer is connected to the input of the output layer, the training layer transmits vector X to the output layer, and the output layer outputs the weight matrix W. out ; The high-dimensional vector X = vector X1 + vector X2 +... + vector X N , the vector X1, vector X2,... and vector X N have equal dimensions.
2. The reservoir system based on multiple random nanowire network of claim 1, wherein: The dynamic behavior of the random nanowire network layer is realized by graph theory, and is specifically as follows: When the random nanowire network layer is stimulated by a voltage, a conductive path is formed on two outer layers wrapped with insulating polyvinylpyrrolidone silver nanowires, and the current on the conductive path is: I(t) = [G low(1-g(t)) +G hight(g(t)) ] V(t) where G low , G hight denote the minimum and maximum conductance of the memristive edge, g(t) denotes the memristive edge conductance, and V(t) denotes the stimulus voltage. The change of the edge conductance of the memristor is represented by a voltage-controlled potential-inhibition rate balance equation: dg(t) / dt = ω a (1 - g(t)) - ω b g(t) where ω a ,ω b denote the inhibition and enhancement coefficients of the control conductance variation; wherein, represents the historical memory state of the network; represents the process of the network changing constantly under voltage stimulus.
3. The reservoir system based on multiple random nanowire network of claim 1, wherein: The input layer collects vector data sets with time series attributes, including but not limited to: voltage, current, temperature, humidity, displacement vector sets that change over time, and the vector data is converted into a voltage input matrix W after normalization in Normalization is achieved by difference standardization, as follows: where V is the current data, V min is the data minimum, V max is the data maximum, and V' is the normalized data.
4. The reservoir system based on multiple random nanowire network of claim 1, wherein: The output weight matrix W out Obtained by the ridge regression algorithm, specifically calculated as follows: W out = Y target X T (XX T + γI) -1 where Y target is the prediction target, X is the high-dimensional vector generated by the training layer, I is the identity matrix, the symbol "T" represents matrix transposition, the symbol "-1" represents matrix inversion; γ is the regularization coefficient, and the value is γ = 10 -8 .
5. The reservoir system based on multiple random nanowire network of claim 1, wherein: The output layer output end is also connected with a data prediction, and the data prediction outputs Y', and is calculated as follows: Y' = W out X.
6. The reservoir system based on multiple random nanowire network of claim 1, wherein: With the increase of 10 input nodes in the first nanowire network layer, the number of nanowire network layers increases by one, and the number of nanowire network layers is greater than or equal to 2.