Pulse neural network-based time series signal classification model construction method and system

Abstract

The application discloses a kind of based on pulse neural network's time sequence signal classification model construction method and system, comprising the following steps: the overall framework of network is built, network is entirely composed of pulse neuron, including input module, liquid state machine module, feedforward classification network module;Setting neuron inside liquid state machine, set appropriate neuron number, set and optimize synapse connection inside liquid state machine, so that liquid state machine reaches critical state;The structure of feedforward classification network is set;Network is trained, and the classification network of time sequence signal is obtained.The model presented in the application utilizes the characteristics of pulse neurons to efficiently represent and process time series information, utilizes LSM and feedforward classification network for feature extraction and pattern recognition, and realizes end-to-end automated time sequence signal classification. It is a model with strong biological interpretability, which can be deployed on a brain-like hardware platform to achieve lightweight, low-energy, and high-speed computing.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and brain-like computing, specifically relating to a method and system for constructing a time-series signal classification model based on a spiking neural network. Background Technology

[0002] With the development of communication technology and the rapid increase in communication data volume, the automated monitoring and identification of information has become a hot topic, especially the identification of time series signals, which has broad application prospects. Currently, methods for identifying time series signals can be broadly divided into two categories: pattern recognition methods based on time-domain and frequency-domain feature extraction, and methods based on deep learning. However, the former relies on manual selection of the feature space, and signal processing and classification are performed separately, resulting in low efficiency and poor real-time performance. The latter utilizes the powerful feature extraction capabilities of deep neural networks, achieving higher classification performance and generalization, but the increased network size and computational load, along with higher accuracy, significantly increase hardware resource consumption and power consumption. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a time-series signal classification method and system based on spiking neural networks. By utilizing the characteristics of spiking neurons, the time-series processing capabilities of liquid state machines, and the pattern recognition capabilities of feedforward classification networks, end-to-end automated time-series signal classification is achieved.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: a method for constructing a time-series signal classification model based on a spiking neural network, comprising the following steps:

[0005] The overall framework for constructing the network model is as follows: the network model consists of spiking neurons, including an input module, a liquid state machine module, and a feedforward classification spiking neural network module;

[0006] Configure the parameters of the spiking neurons inside the liquid state machine module; configure the synaptic connections inside the liquid state machine module to achieve a critical state in which the excitation and inhibition states are balanced; configure the structure of the feedforward classification spiking neural network, which adopts a hierarchical SNN model.

[0007] Optimize the internal weight connection parameters of the network model to obtain a classification network model for time series signals.

[0008] Furthermore, within the overall framework of the network model:

[0009] The input module is used to configure multi-dimensional input of time-series signals of any dimension according to input requirements. The input module is connected to the excitatory neurons of the liquid state machine module according to a set ratio. The input module is only connected to the excitatory neurons of the liquid state machine.

[0010] The liquid state machine module is used to encode time-series signals to obtain pulse sequences and extract time-series features. The liquid state machine module is a recursive spiking neural network, composed of excitatory spiking neurons and inhibitory spiking neurons. The ratio of excitatory to inhibitory spiking neurons is set at M1:M2, where M1>M2. Excitatory spiking neurons have recursive synaptic connections within themselves, and there are bidirectional synaptic connections between excitatory and inhibitory spiking neurons. Inhibitory spiking neurons do not have recursive synaptic connections within themselves. Each neuron in the liquid state machine module is in an active state, and the degree of activation is within a set range.

[0011] The feedforward classification spiking neural network is composed of convolutional layers, downsampling layers, and fully connected layers. The neurons in each layer have the same function and are independent of each other.

[0012] Furthermore, when setting the parameters of the spiking neurons inside the liquid state machine module: spiking neurons include two main types of neurons, excitatory and inhibitory, and the neuron model can be configured with multiple different types of spiking neuron models, and the parameters of the neuron model are multi-configurable;

[0013] Various types of spiking neuron models include the Hodgkin-Huxley neuron model, the leakage cumulative firing neuron model, the Izhikevich neuron model, and the impulse response model.

[0014] Furthermore, a brain-inspired optimization method based on impulse temporal-dependent plasticity rules is employed for the liquid state machine module to achieve synaptic learning, enabling the liquid state machine to reach a critical state where excitation and inhibition states are balanced.

[0015] The critical state in which excitation and inhibition are in equilibrium must satisfy either of the following two criteria: First, the difference between the weighted sum of the number of pulses released by excitatory spiking neurons and their corresponding weights and the weighted sum of the number of pulses released by inhibitory spiking neurons and their corresponding weights is less than a threshold ∈; Second, the difference between the information entropy of the input time series information received by the liquid state machine module and the pulse sequence information output by the liquid state machine module within a certain period of time is less than a threshold ∈.

[0016] Furthermore, the feedforward classification spiking neural network consists of multiple convolutional layers, downsampling layers, and fully connected layers. The neurons in each layer are composed of spiking neurons. The input to the feedforward classification spiking neural network is the pulse sequence output by the excitatory spiking neurons in the liquid state machine module. After the liquid state machine module completes the pulse encoding and temporal feature extraction of the time-series signal, it inputs the pulse sequence into the feedforward classification spiking neural network. The feedforward classification spiking neural network performs spatial feature extraction and automatically completes the classification of the time-series signal.

[0017] Furthermore, training the network and optimizing the weight connection parameters to obtain a classification network model for time-series signals includes: using the pulse time-dependent plasticity rule or an error backpropagation method that replaces the gradient for the feedforward classification spiking neural network;

[0018] The weights of the connections input to the liquid state machine and the connections inside the liquid state machine are frozen. The network is optimized using the pulse time-dependent plasticity rule or the error backpropagation method that replaces the gradient. The weights of the connections at the LSM output and the connections inside the feedforward classification network are optimized to obtain the classification network model of the time-series signal.

[0019] Based on the same technical concept, the present invention provides a system for constructing a time-series signal classification model based on a spiking neural network, including a network structure construction module, a network parameter preset module, and a training module;

[0020] The network structure building module is used to build the overall framework of the network model, which consists of spiking neurons, including an input module, a liquid state machine module, and a feedforward classification spiking neural network module.

[0021] The network parameter preset module is used to set the parameters of the spiking neurons inside the liquid state machine module; set the synaptic connections inside the liquid state machine module so that the liquid state machine module reaches a critical state in which the excitation and inhibition states are balanced; and set the structure of the feedforward classification spiking neural network, which adopts a hierarchical SNN model.

[0022] The training module is used to optimize the internal weight connection parameters of the network model to obtain a classification network model for time series signals.

[0023] Based on the above technical solutions, the present invention also provides a time series signal classification method based on spiking neural networks, comprising the following steps: acquiring a time series signal with n channels and k length, passing it through an end-to-end classification network model to output the classification corresponding to the signal; the end-to-end classification network model is obtained by the time series signal classification model construction method based on spiking neural networks.

[0024] The present invention also provides a computer device, including a processor and a memory. The memory is used to store a computer executable program. The processor reads part or all of the computer executable program from the memory and executes it. When the processor executes part or all of the executable program, it can realize the time-series signal classification model construction method based on spiking neural networks or the time-series signal classification method based on spiking neural networks described in the present invention.

[0025] Simultaneously, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, it can implement the time-series signal classification model construction method based on spiking neural networks or the time-series signal classification method based on spiking neural networks described in this invention.

[0026] Compared with the prior art, the present invention has at least the following beneficial effects:

[0027] The present invention provides a method for constructing a time-series signal classification model based on a spiking neural network. The model uses spiking neurons throughout, which efficiently represent and process time-series information. It designs a critical-state liquid state machine and a feedforward classification network structure to process the signal, train the network, and achieve end-to-end automated time-series signal classification. It is a biologically interpretable model. The spiking neurons, critical-state liquid state machine, and network optimization methods in the model can all find corresponding experimental evidence in neurobiology. Moreover, it can be deployed on brain-like hardware platforms such as DYNAP-SE, BrainScaleS, and Tianji Core to achieve lightweight, low-power, and high-speed computation. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall network architecture in this invention.

[0029] Figure 2 The diagram shows the neuron and synapse model in the spiking neural network of this invention. Among them, (a) is the equivalent circuit diagram of the Hodgkin-Huxley neuron model, (b) is the equivalent circuit diagram of the leakage current cumulative firing neuron model, and (c) is the model diagram of synaptic connections.

[0030] Figure 3 This is a schematic diagram of the liquid state machine module in the network of this invention.

[0031] Figure 4 This is a schematic diagram of the feedforward network optimization method in this invention. Among them, (a) is the image of the step function, derivative function, impulse function and its substitution function in the backpropagation algorithm based on pseudo-gradient, and (b) is the window diagram of the impulse time-dependent plasticity learning rule.

[0032] Figure 5 This is a flowchart of the network operation in this invention. Detailed Implementation

[0033] The problem this invention aims to solve is pattern recognition and classification of time-series signals. A time-series signal with n channels and k length is processed through an end-to-end network model, which outputs the corresponding classification for the signal.

[0034] For example, electromyographic data of corresponding forearm muscle activities are collected under different gestures. The electromyographic signal acquisition device obtains a time-series signal with 8 channels and a length of 400, which is then encoded and input into the network model to output the gesture category corresponding to the current signal.

[0035] The input of a network model can be various time-series signal collection devices, such as wireless signal receivers, event cameras, and bioelectric measurement devices, while the output can be specific monitoring and indicator devices, task execution devices, or more complex computational models for further analysis and processing.

[0036] The network model is constructed entirely using spiking neurons, and consists of three modules arranged in the order of information transmission: an input module, a liquid state machine module, and a feedforward classifier module. The original multi-channel time-series signal is input into the network model, and the predicted class value of the signal is obtained at the output of the feedforward classification module. The overall architecture of the network model is as follows: Figure 1 As shown.

[0037] The network model and its modules are as follows:

[0038] Spiking neuron and synapse models transmit encoded information in the form of pulse sequences and are computational models constructed based on the dynamic characteristics of biological neurons. The firing activity of spiking neurons follows an "all-or-none" rule; whether a pulse is generated depends on whether the received stimulus reaches a threshold.

[0039] Different spiking neuron models have different characteristics and can be selected according to computational requirements. Commonly used models include the Hodgkin-Huxley (HH) neuron model, the Leaky Integrate-and-Fire Neuron (LIF) model, the Izhikevich neuron model, and the spike response model (SRM).

[0040] The HH neuron model uses differential equations to reveal the fundamental laws governing neuronal activity. Various ion channels exist on the neuron membrane, and the channel conductivity is closely related to the membrane potential; the principle can be seen in… Figure 2 (a) Equivalent circuit diagram. In the HH model, part of the input current charges the membrane's capacitance potential, and the other part flows through the various ion channels, which can be described by the following expression:

[0041]

[0042] Where I(t) represents the input, I C (t) and I R (t) represents the current charging the capacitor and the current flowing through the ion channel, respectively.

[0043] LIF model: simplifies the HH model, considering only the leakage current in the circuit. The equivalent circuit is as follows: Figure 2 (b) The firing of a presynaptic neuron generates a current input to the current neuron. Part of this current is used to charge the membrane capacitance, and the other part leaks through the membrane resistance. When the voltage across the membrane capacitance exceeds a fixed firing threshold, the neuron fires a pulse. The principle is explained below. Figure 2 (b) Equivalent circuit diagram. The membrane potential change can be described by the following differential equation:

[0044]

[0045] Where u represents the membrane potential of the neuron; τ m It represents the time constant of its membrane potential decay.

[0046] Izhikevich neuron model: The HH neuron model is relatively complex and difficult to compute quickly, while the LIF model is too simplified and cannot simulate the dynamic behavior of multiple types of neurons. By simplifying the HH neuron model, the Izhikevich neuron model can be obtained, which can be described by second-order differential equations:

[0047]

[0048]

[0049] Synaptic model: such as Figure 2 As shown in (c), the synaptic connection weights of neurons reflect the strength of the influence of the spiking signals generated by the presynaptic neuron on the postsynaptic neuron. Synaptic plasticity refers to the ability of synaptic connection weights to change. In the design of spiking neural networks, synaptic weights are set as learnable parameters, and corresponding rules are used to update the synaptic weights to optimize the network performance and complete the specified task requirements.

[0050] The original signal is converted into a pulse sequence using time encoding and rate encoding methods, and then input into the subsequent liquid state machine module with random weights. The primary encoding methods considered here are time encoding and rate encoding. If the original time-series signal is a pulse sequence, the encoding step can be omitted, and the pulse sequence signal can be directly input into the liquid state machine module.

[0051] A liquid state machine (LSM) is a special type of spiking neural network model that can effectively process input time-series information and output computational results for continuous time series. It is an adaptive system that supports real-time computation. The neurons within an LSM are recursively connected, without a clear hierarchical structure, and the connection weights are randomly determined. The activity of an internal neuron represents the current liquid state, determined by the input information and the previous liquid state, and it possesses memory capabilities.

[0052] The computational power of LSM depends on the separation property SP and the approximation property AP. The separation property SP describes the degree of separation of the internal state trajectories of the system caused by two different sets of input sequences, reflecting the difference in the system's response to different inputs. The approximation property AP describes the recoding capability of the readout mechanism, reflecting the system's ability to transform different internal states to the target output.

[0053] In such a recursive network module, the network module possesses maximum computational and storage capabilities when it reaches a critical state—that is, when the internal neurons achieve an excitation-inhibition equilibrium. There is also some evidence that the human brain also exists in such a critical state. In the network architecture described in this invention, the critical state is achieved by adjusting the number of neurons, internal parameters, connection methods, and connection weights. Therefore, it can be considered that the LSM module maps the input signal from a low-dimensional (channel number) space to a high-dimensional (LSM neuron number) space, achieving highly efficient encoding through a critical state network configuration.

[0054] The working principle of LSM is as follows: Figure 3 As shown, the input sequence is u(t), the target output is y(t), and the current liquid state is x. M (t)=(K M u)(t), using L M This represents the mapping function from the input sequence to the internal state of the liquid. For the output, a memoryless readout mapping f is introduced. M The output of the liquid state machine Choose a suitable mapping f M This ensures that the output of the liquid state machine aligns with the target output, allowing for real-time extraction of the required input information from the current state of the loop. Leveraging the characteristics of high-dimensional dynamic systems, instantaneous states are transformed into stable target outputs. As an example, in this invention, a linear mapping with variable weights is used as f. M And in the subsequent optimization of the feedforward classification network, the connection weights are automatically adjusted so that f M Learn to extract signal category-related features from LSM.

[0055] In the network architecture of this invention, the feedforward classification spiking neural network adopts a hierarchical SNN model. The distribution of neurons exhibits a clear hierarchical characteristic. Neurons within each layer have the same function and are independent of each other, with no synaptic connections within layers. Connections between layers are made as needed, and various combinations of architectures can be used, such as convolutional layers, pooling layers, and fully connected layers. The network directly receives and processes the LSM state to obtain the classification result. The synaptic weights in the feedforward classification spiking neural network are variable and can be optimized using rules similar to backpropagation in deep neural network training or biologically interpretable spiking temporal dependence plasticity rules.

[0056] (1) Backpropagation rules based on pseudo-gradients:

[0057] Backpropagation is a commonly used algorithm in deep neural network training. It defines the error L between the network's target output and the actual output, calculates the gradient, and updates the weights according to the following rules:

[0058]

[0059] The output of a spiking neuron is a discontinuous impulse function. Direct backpropagation will result in either vanishing or exploding gradients. Therefore, a smooth function is used instead of the impulse function, and a backpropagation rule based on pseudo-gradients is executed. The principle is as follows: Figure 4 As shown in (a).

[0060] (2) Spike-Time Dependent Plasticity (STDP) Rule

[0061] Biological experiments have shown that changes in synaptic connection strength are influenced by the impulse firing activity of presynaptic and postsynaptic neurons, and the specific trends and magnitudes of these changes are closely related to the precise timing of the impulse firing. Based on this, the proposed Spike-Time Dependent Plasticity (STDP) satisfies the following form: if the presynaptic neuron fires before the postsynaptic neuron, a causal relationship is considered between the two neurons, and the synaptic weight increases; conversely, the synaptic weight decreases. The specific learning rules are as follows:

[0062]

[0063] Where w ij This represents the synaptic connection weight from neuron j to neuron i. a represents the time difference in impulse firing between the postsynaptic and presynaptic neurons. + a - τ represents the coefficient. + and τ - Representing the time constant, the STDP learning window is as follows: Figure 4 As shown in (b).

[0064] The flowchart of the network operation of this invention is as follows: Figure 5 As shown, the setup and training steps are as follows:

[0065] 1. Construct an overall framework based on spiking neural networks. The spiking neural network is entirely composed of spiking neurons. The overall architecture includes an input module, an LSM module, and a feedforward classification spiking neural network module, forming an end-to-end time series signal classification network model. Synapses from the input to the LSM module are randomly connected with predetermined probabilities. The excitatory neurons of the LSM are directly connected to the lowest layer of the feedforward classification network module. The output of the last layer of the feedforward classification network module is the signal classification prediction result. The LSM is a recursive SNN, and the feedforward classification network is a hierarchical SNN with feedforward connections.

[0066] 2. Configure the neurons within the LSM. Based on the number of input signal channels n and the signal length k, determine the number of neurons within the LSM: m = m1 + m2 (m1 and m2 are the number of excitatory neurons and inhibitory neurons, respectively). Configure the parameters and characteristics of the neurons within the LSM according to task requirements, such as the neuron's time constant, firing threshold, impulse response kernel function, refractory period, etc. The computational power of the LSM can be adjusted by modifying the diversity of neurons.

[0067] 3. Configure synaptic connections within the LSM to reach a critical state. The LSM employs random and sparse synaptic connections, including connections between excitatory and inhibitory neurons, as well as connections within excitatory neurons themselves. Experiment with input signals, observe the activity state of the LSM neurons, and adjust the ratio of the two types of neurons and the probability of each synaptic connection based on the results. Combine this with the STDP learning rules to adjust the connection weights, continuing this process until the LSM network reaches a critical state, where excitatory and inhibitory activities cancel each other out in response to the input signal. The critical state must satisfy either of the following two criteria: first, the difference between the weighted sum of the number of pulses released by excitatory spiking neurons and their corresponding weights and the weighted sum of the number of pulses released by inhibitory spiking neurons and their corresponding weights is less than a threshold ∈; second, the difference between the information entropy of the input time-series information received by the liquid state machine module and the pulse sequence information output by the liquid state machine module within a certain period is less than a threshold ∈. ∈ is set according to the difficulty and accuracy requirements of the classification task.

[0068] 4. Design the structure of the feedforward classification spiking neural network. Based on the task requirements and signal characteristics, design the size and connection method of each layer of the network, which may include convolutional layers, pooling layers, fully connected layers, and residual connections.

[0069] 5. Train the network. Freeze the weights of the connections input to the LSM and the connections within the LSM, and use the error backpropagation method based on alternative gradients or the STDP method to optimize the weights of the connections at the LSM output and the connections within the feedforward classification network; after training, a time-series signal classification model based on a spiking neural network is obtained.

[0070] For the liquid state machine module, a brain-like optimization method based on the pulse temporal dependence plasticity rule is used to achieve synaptic learning. For the feedforward classification spiking neural network, the pulse temporal dependence plasticity rule or the error backpropagation method that replaces the gradient is used for network optimization.

[0071] The present invention can also provide a computer device, including a processor and a memory, wherein the memory is used to store a computer executable program, the processor reads part or all of the computer executable program from the memory and executes it, and when the processor executes part or all of the computer executable program, it can realize the time-series signal classification model construction method based on spiking neural network and the time-series signal classification method based on spiking neural network described in the present invention.

[0072] On the other hand, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the time-series signal classification model construction method based on spiking neural networks and the time-series signal classification method based on spiking neural networks described in the present invention.

[0073] The computer equipment may be a laptop, desktop computer, vehicle-mounted computer, or workstation.

[0074] The processor described in this invention may be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), or an off-the-shelf programmable gate array (FPGA).

[0075] The memory described in this invention can be an internal storage unit of a laptop, desktop computer, vehicle computer, or workstation, such as memory or hard disk; or it can be an external storage unit, such as a portable hard disk or flash memory card.

[0076] Computer-readable storage media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media can include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. Random access memory can include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

Claims

1. A time-series signal classification method based on a spiking neural network, characterized in that, Includes the following steps: The number of channels to obtain is , length is The signal is a time-series signal; the signal is electromyographic data of corresponding forearm muscle activities collected under different gestures, and the time-series signal with 8 channels and a length of 400 is obtained by electromyographic signal acquisition device; the original signal is converted into a pulse sequence through time encoding and rate encoding. The pulse sequence is processed by an end-to-end classification network model to output the gesture category corresponding to the signal; The construction of an end-to-end classification network model includes the following steps: The overall framework of the network model is constructed. The network model consists of spiking neurons, including an input module, a liquid state machine module, and a feedforward classification spiking neural network module. The input module is used to configure multi-dimensional inputs for time-series signals of arbitrary dimensions according to input requirements. The input module is connected to the excitatory neurons of the liquid state machine module according to a set ratio. The input module is only connected to the excitatory neurons of the liquid state machine. The liquid state machine module is used to encode the time-series signals to obtain pulse sequences and extract time-series features. The liquid state machine module is a recursive spiking neural network, composed of excitatory spiking neurons and inhibitory spiking neurons. The excitatory spiking neurons and inhibitory spiking neurons are set in a ratio of M1:M2, where M1>M2. There are recursive synaptic connections within the excitatory spiking neurons, and bidirectional synaptic connections between excitatory and inhibitory spiking neurons. There are no recursive synaptic connections within the inhibitory spiking neurons. Each neuron in the liquid state machine module is in an active state, and the activation level is within a set range. The feedforward classification spiking neural network is composed of convolutional layers, downsampling layers, and fully connected layers. The neurons in each layer have the same function and are independent of each other. The parameters of the spiking neurons within the liquid state machine module are set; the synaptic connections within the liquid state machine module are configured to achieve a critical state where excitation and inhibition are balanced; the structure of the feedforward classification spiking neural network is set, employing a hierarchical SNN model; a brain-like optimization method based on pulse temporal-dependent plasticity rules is used to achieve synaptic learning for the liquid state machine module, ensuring that the liquid state machine reaches a critical state where excitation and inhibition are balanced. The critical state of equilibrium between excitation and inhibition states must satisfy either of the following two criteria: First, the difference between the weighted sum of the number of pulses released by excitatory spiking neurons and their corresponding weights and the weighted sum of the number of pulses released by inhibitory spiking neurons and their corresponding weights is less than a threshold. Secondly, the difference in information entropy between the input time series information received by the liquid state machine module and the pulse sequence information output by the liquid state machine module within a certain period is less than a threshold. ; Optimize the internal weight connection parameters of the network model to obtain a classification network model for time series signals. During optimization, the connection weights input to the liquid state machine and the connection weights inside the liquid state machine are frozen.

2. The time-series signal classification method based on spiking neural networks according to claim 1, characterized in that, When setting the parameters of the spiking neurons inside the liquid state machine module: spiking neurons include two main types of neurons: excitatory and inhibitory, and the neuron model can be configured with multiple different types of spiking neuron models, and the parameters of the neuron model are multi-configurable; Various types of spiking neuron models include the Hodgkin-Huxley neuron model, the leakage cumulative firing neuron model, the Izhikevich neuron model, and the impulse response model.

3. The time-series signal classification method based on spiking neural networks according to claim 1, characterized in that, The feedforward classification spiking neural network consists of multiple convolutional layers, downsampling layers, and fully connected layers. Each layer's neurons are composed of spiking neurons. The input to the feedforward classification spiking neural network is the pulse sequence output by the excitatory spiking neurons in the liquid state machine module. After the liquid state machine module completes pulse encoding and temporal feature extraction of the time-series signal, it inputs the pulse sequence into the feedforward classification spiking neural network. The feedforward classification spiking neural network then performs spatial feature extraction and automatically classifies the time-series signal.

4. The time-series signal classification method based on spiking neural networks according to claim 1, characterized in that, Training the network, optimizing the weight connection parameters, and obtaining a classification network model for time-series signals include: using the pulse time-dependent plasticity rule or an error backpropagation method that replaces the gradient for the feedforward classification spiking neural network; Network optimization is performed using the pulse timing-dependent plasticity rule or the error backpropagation method that replaces the gradient. The connection at the LSM output and the weights of the internal connections of the feedforward classification network are optimized to obtain a classification network model for the time series signal.

5. A time-series signal classification system based on a spiking neural network, characterized in that, The method for classifying time-series signals based on spiking neural networks as described in any one of claims 1 to 4 includes a data acquisition module, a network structure construction module, a network parameter preset module, and a training module. The data acquisition module is used to obtain the number of channels. , length is Time series signals; The network structure building module is used to build the overall framework of the network model, which consists of spiking neurons, including an input module, a liquid state machine module, and a feedforward classification spiking neural network module. The network parameter preset module is used to set the parameters of the spiking neurons inside the liquid state machine module; set the synaptic connections inside the liquid state machine module so that the liquid state machine module reaches a critical state in which the excitation and inhibition states are balanced; and set the structure of the feedforward classification spiking neural network, which adopts a hierarchical SNN model. The training module is used to optimize the internal weight connection parameters of the network model to obtain a classification network model for time series signals.

6. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store a computer-executable program, the processor reading part or all of the computer-executable program from the memory and executing it, and the processor executing part or all of the computer-executable program is able to implement the timing signal classification method based on a spiking neural network as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the time-series signal classification method based on a spiking neural network as described in any one of claims 1 to 4.

Images