Processor temperature calculation method based on spiking neural network
Through the processor temperature calculation method based on pulsed neural network, the problem of processor temperature management in the prior art relying on high computing capabilities is solved, and high-efficiency temperature estimation in embedded systems is realized, which reduces GPU dependence and model training costs. Improved convergence speed.
Patent Information
- Application Number
- CN202410812467.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-21
- Publication Date
- 2024-11-29
AI Technical Summary
Existing processor temperature estimation and management methods rely on artificial neural networks and require high-frequency access to memory, resulting in strong computing power and inability to deploy in embedded systems and ineffective management of processor temperature.
Using the processor temperature calculation method based on pulsed neural network, by collecting the relevant hardware information and temperature data of the CPU, establishing densely connected feedforward artificial neural networks and neural network models that simulate the dynamic behavior of biological neurons, and using membrane potential aggregation algorithm to estimate Processor temperature reduces dependence on GPU.
It realizes efficient operation on GPU devices, reduces model training costs, increases convergence speed by about 50%, and establishes a local learning mechanism through biology-related knowledge, which is more in line with the way the human brain processes information.
Smart Images

Figure CN119047289A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of processor performance optimization, and in particular to a processor temperature calculation method based on a pulse neural network. Background Art
[0002] With the rapid development of computer science and technology, the performance of processors is constantly improving. However, the temperature management of processors has become an important issue. Overheating of the processor will lead to reduced system performance and may even damage the hardware. Therefore, accurate estimation and effective management of the processor temperature are of great significance to ensure the stable operation of the system and extend the service life of the hardware.
[0003] Currently, artificial neural networks (ANNs) are often used to estimate and manage processor temperature, but there are some problems. Kaushik Roy et al. proposed that ANNs need to access memory frequently, which requires powerful computing power. Relying on GPU cannot be deployed in embedded systems
[0004] Therefore, it is urgent to improve this shortcoming. The present invention studies and improves the existing structure and shortcomings, and provides a processor temperature calculation method based on a pulse neural network. Summary of the invention
[0005] The purpose of the present invention is to provide a processor temperature calculation method based on a pulse neural network to solve the problems raised in the above background technology.
[0006] To achieve the above object, the present invention provides the following technical solution: a processor temperature calculation method based on a pulse neural network, comprising the following specific steps:
[0007] S1: Collect data to form a data set. Use the CPU data collection framework to collect CPU hardware information and temperature data when the CPU is periodically in full load and idle state.
[0008] S2: Establish a densely connected feedforward artificial neural network, which is ANN. Specifically: From the input space To output space function, given by the nonlinear function h (l) The combination of definitions makes: Here, x represents the dimension d x The input vector is Represents dimension d y The output vector of the nonlinear function h (l)It is called a layer and defines a l-fold cross validation, mapping the input vector to the output vector. The first layer h (0) Defined as the input layer, the last layer h (nL) As the output layer,
[0009]
[0010] The layer h between the input layer and the output layer (l) It is called the hidden layer and is defined as:
[0011]
[0012] in is the lth layer h (l) The nth neural unit, n u represents the total number of neural units in each layer, is the lth layer h (l) The weight vector of the nth neural unit in (l-1) is the output of the previous layer, where the bias value is absorbed;
[0013] S21:φ (l) :R→R is a nonlinear activation function that takes into account historically relevant input and output data and, The formula for the hidden layer is:
[0014]
[0015] Since time is discrete, each discrete time slice of the input vector of the previous layer is processed in turn:
[0016] S3: Establish a neural network model that simulates the dynamic behavior of biological neurons. The spiking neural network is SNN. SNN is a history-related ANN that introduces memory effects through a biologically inspired process. Therefore, the activation function φ in S21 (l) It can be expressed as:
[0017] and in, is the membrane potential of the nth neuron at time t, represents the membrane threshold, is the membrane potential decay rate, They are the standard ANN weights multiplied by the previous layer at the current time step;
[0018] S31: The membrane potential does not reach the membrane threshold or remains silent through time evolution (φ spk =0), if the membrane potential exceeds the membrane threshold, the SNN activates the neural unit and outputs a discrete pulse (φspk =1), these pulses are called spikes, in S3, It is called the reset mechanism. Once a spike is emitted, the membrane potential is reset. The membrane threshold and membrane potential decay rate are optimized during training, so that the optimized parameters of SNN are:
[0019] S32: A spike is a single event expressed by a Dirac delta function or a discrete pulse in a continuous or discrete setting. A group of spikes over time is called a spike train, i = [i t ,t=0,…,n t ] For this purpose, spiking neurons are trained with spikes at time intervals consisting of either a spike (1) or zero input (0), and the membrane potential The incoming spike i t regulation, in the absence of input spikes, due to the membrane decay rate The membrane voltage decays over time;
[0020] S33: Use the backpropagation algorithm from a standard artificial neural network, use a proxy gradient to replace the activation during backpropagation, and use a fast sigmoid function:
[0021]
[0022] where k is the parameterized slope, and for some input x, the proxy φ surr (x) is continuously differentiable and preserves the gradient dynamics of the network, using φ' surr Used for back-propagation training;
[0023] S4: Design the network structure: first normalize the real data between 0 and 1, and then directly input the normalized data into the first input layer of the network. The input layer will select the appropriate encoding method based on the real data. Since the membrane potential directly outputs a continuous value, the decoding part directly uses the membrane potential value of the output layer as a real number to represent the calculated temperature value.
[0024] Furthermore, the data collected in S1 also includes CPU-related hardware counts and four core temperatures at different fan speeds and different CPU frequencies, and the four core temperatures are Core 0, Core 1, Core 2, and Core 3 respectively. The data set was collected under the Ubuntu 20.04.5LTS system, and the hardware device for collecting the data set was a CPU model i5-6500.
[0025] Furthermore, the CPU data collection framework is used for data collection using custom script execution, which follows the specified execution order outlined in the framework. The process is as follows:
[0026] Step a: Generate a PTM configuration file and set the corresponding change strategy for the running status of each core;
[0027] Step b: Adjust the operating frequency of the multi-core to match the frequency f supported by the CPU;
[0028] Step c: Start two processes at the same time. Process 1 executes the data collection program, and process 2 starts McFTP to manage the scheduled operation or sleep state.
[0029] Furthermore, the weight in S21 Shared across all time steps.
[0030] Furthermore, the SNN in S3 is a history-related ANN, where Also known as the LIF neuron model.
[0031] Furthermore, the input peak in S31 Triggering a change in membrane potential when sufficient excitation exceeds the threshold U thr The neuron emits an output spike when
[0032] Furthermore, the S4 designed network structure for the real data x with a time step of t t , directly input it into the input layer in the network structure In Chinese, that is: The hidden layer For network transmission, and transmit the data of the hidden layer to the next layer, the last output layer The output of the decoder layer is designed It outputs the membrane potential of the output layer directly as a real number, namely:
[0033]
[0034] The sum of the membrane potential of each time step of the output layer is taken as the final required temperature real number y t ,Right now: Where n0 represents the number of times a set of input data is simulated in the network.
[0035] Furthermore, in S4, the real data x t After normalization preprocessing, the data is directly transmitted to the input layer in a constant current manner, and then reaches the output layer, i.e. the decoding layer, after passing through the hidden layer. Then the required real number y is obtained by summing up the membrane potentials of all simulated decoding layers t .
[0036] The present invention provides a processor temperature calculation method based on a pulse neural network, which has the following beneficial effects: the processor temperature calculation method based on a pulse neural network estimates the temperature of the processor at the current moment based on the membrane potential aggregation (AMP) algorithm of the SNN, and can efficiently run the model even on a device without a GPU, thereby reducing the cost of training the model;
[0037] Compared with the traditional temperature estimation algorithm based on membrane potential, the convergence speed is improved by about 50%;
[0038] Providing a program for collecting CPU-related information reduces the difficulty of data collection and can establish a local learning mechanism. The information required for learning does not require a specific transmission path, but only relies on the information of the previous and subsequent neurons themselves (such as membrane potential, firing rate, etc.), which increases biology-related knowledge and is more in line with the way the human brain processes information. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 A schematic diagram of a dense feedforward artificial neural network of the present invention;
[0040] Figure 2 Schematic diagram of the calculation of SNN forward propagation in the present invention;
[0041] Figure 3 It is a schematic diagram of the calculation of the SNN back propagation of the present invention;
[0042] Figure 4 Schematic diagram of the regression model structure of the spike network of the present invention. DETAILED DESCRIPTION
[0043] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0044] like Figure 1-Figure 4 As shown, a processor temperature calculation method based on a pulse neural network includes the following specific steps:
[0045] S1: Collect data to form a data set. Use the CPU data collection framework to collect CPU-related hardware information and temperature data when the CPU is periodically in full load and idle states.
[0046] S2: Establish a densely connected feedforward artificial neural network, which is ANN. Specifically: From the input space To output space function, given by the nonlinear function h (l) The combination of definitions makes: Here, x represents the dimension d x The input vector is Represents dimension d y The output vector of the nonlinear function h (l) It is called a layer and defines a l-fold cross validation, mapping the input vector to the output vector. The first layer h (0) Defined as the input layer, the last layer h (nL) As the output layer,
[0047]
[0048] The layer h between the input layer and the output layer (l) It is called the hidden layer and is defined as:
[0049]
[0050] in is the lth layer h (l) The nth neural unit, n u represents the total number of neural units in each layer, is the lth layer h (l) The weight vector of the nth neural unit in (l-1) is the output of the previous layer, where the bias value is absorbed;
[0051] S21:φ (l) :R→R is a nonlinear activation function that takes into account historically relevant input and output data and, The formula for the hidden layer is:
[0052]
[0053] Since time is discrete, each discrete time slice of the input vector of the previous layer is processed in turn:
[0054] S3: Establish a neural network model that simulates the dynamic behavior of biological neurons. The spiking neural network is SNN. SNN is a history-related ANN that introduces memory effects through a biologically inspired process. Therefore, the activation function φ in S21 (l) It can be expressed as:
[0055]
[0056] in, is the membrane potential of the nth neuron at time t, represents the membrane threshold, is the membrane potential decay rate, They are the standard ANN weights multiplied by the previous layer at the current time step;
[0057] S31: The membrane potential does not reach the membrane threshold or remains silent through time evolution (φ spk =0), if the membrane potential exceeds the membrane threshold, the SNN activates the neural unit and outputs a discrete pulse (φ spk =1), these pulses are called spikes, in S3, It is called the reset mechanism. Once a spike is emitted, the membrane potential is reset. The membrane threshold and membrane potential decay rate are optimized during training, so that the optimized parameters of SNN are:
[0058] S32: A spike is a single event expressed by a Dirac delta function or a discrete pulse in a continuous or discrete setting. A group of spikes over time is called a spike train, i = [i t ,t=0,…,n t ] For this purpose, spiking neurons are trained with spikes at time intervals consisting of either a spike (1) or zero input (0), and the membrane potential The incoming spike i t regulation, in the absence of input spikes, due to the membrane decay rate The membrane voltage decays over time;
[0059] S33: Use the backpropagation algorithm from a standard artificial neural network, use a proxy gradient to replace the activation during backpropagation, and use a fast sigmoid function:
[0060]
[0061] where k is the parameterized slope, and for some input x, the proxy φ surr (x) is continuously differentiable and preserves the gradient dynamics of the network, using φ' surr Used for back-propagation training;
[0062] S4: Design the network structure: first normalize the real data between 0 and 1, and then directly input the normalized data into the first input layer of the network. The input layer will select the appropriate encoding method based on the real data. Since the membrane potential directly outputs a continuous value, the decoding part directly uses the membrane potential value of the output layer as a real number to represent the calculated temperature value.
[0063] The data collected in S1 also includes CPU-related hardware counts and four core temperatures at different fan speeds and CPU frequencies. The four core temperatures are Core 0, Core 1, Core 2, and Core 3. The data set was collected under the Ubuntu 20.04.5LTS system, and the hardware device for collecting the data set is the CPU model i5-6500.
[0064] The CPU data collection framework is used for data collection using custom script execution, which follows the specified execution order outlined in the framework. The process is as follows:
[0065] Step a: Generate a PTM configuration file and set the corresponding change strategy for the running status of each core;
[0066] Step b: Adjust the operating frequency of the multi-core to match the frequency f supported by the CPU;
[0067] Step c: Start two processes at the same time. Process 1 executes the data collection program, and process 2 starts McFTP to manage the scheduled operation or sleep state.
[0068] S21 Medium Weight Shared across all time steps.
[0069] The SNN in S3 is a history-related ANN, in which, Also known as the LIF neuron model.
[0070] Input spike in S31 Triggering a change in membrane potential when sufficient excitation exceeds the threshold U th r The neuron emits an output spike when
[0071] S4 designs the network structure for the real data x with time step t t , directly input it into the input layer in the network structure In Chinese, that is: The hidden layer For network transmission, and transmit the data of the hidden layer to the next layer, the last output layer The output of the decoder layer is designed It outputs the membrane potential of the output layer directly as a real number, namely:
[0072]
[0073] The sum of the membrane potential of each time step of the output layer is taken as the final required temperature real number y t ,Right now: Where n0 represents the number of times a set of input data is simulated in the network.
[0074] Furthermore, in S4, the real data x t After normalization preprocessing, the data is directly transmitted to the input layer in a constant current manner, and then reaches the output layer, i.e. the decoding layer, after passing through the hidden layer. Then the required real number y is obtained by summing up the membrane potentials of all simulated decoding layers t .
[0075] In summary, the working principle of the processor temperature calculation method based on pulse neural network is:
[0076] S1: Collect data to form a data set. Use the CPU data collection framework to collect CPU-related hardware information and temperature data when the CPU is periodically in full load and idle states. The collected data also includes CPU-related hardware counts and four core temperatures at different fan speeds and CPU frequencies. The four core temperatures are Core 0, Core 1, Core 2, and Core 3. The data set was collected under the Ubuntu 20.04.5LTS system. The hardware device for collecting the data set is the CPU model i5-6500, and a fan speed controller, a chassis fan speed regulator model STW-5036, is also installed. In addition, you need to install the necessary software in the Ubuntu system. Performance Application Programming Interface (PAPI) is a library that provides performance counters on various platforms. During a specified run, performance counters provide accurate low-level information about processor behavior, including simple indicators such as total cycle counts, cache misses, and executed instructions, as well as more advanced information (such as total FLOPS and warp occupancy), etc., which can view the relationship between software performance and hardware events in the computer system in real time. Use PAPI to view low-level performance hardware counts in the computer system. lm_sensors (Linux monitoring sensors) is a free open source software tool for Linux that can be used to monitor temperature, voltage, and fans. It is used to monitor the temperature of each core of the CPU. Multi-core Fast Thermal Prototyping (McFTP) uses physical processors to perform real tasks or benchmarks to fully load the physical processors. The purpose is to make the CPU periodically fully loaded and idle; use the fan speed controller to set a fixed fan speed, and the CPU frequency can be changed every T time. The CPU is in a periodic change of working for t1 time and resting for t2 time, and data is collected every t3 time.
[0077] The CPU data collection framework is used for data collection using custom script execution, which follows the specified execution order outlined in the framework. The process is as follows:
[0078] Step a: Generate a PTM configuration file and set the corresponding change strategy for the running status of each core;
[0079] Step b: Adjust the operating frequency of the multi-core to match the frequency f supported by the CPU;
[0080] Step c: Start two processes at the same time: process 1 executes the data collection program, and process 2 starts McFTP to manage the scheduled operation or sleep state;
[0081] S2: Establish a densely connected feedforward artificial neural network, which is ANN. Specifically: From the input space To output space function, given by the nonlinear function h (l) The combination of definitions makes: Here, x represents the dimension d x The input vector is Represents dimension d y The output vector of the nonlinear function h (l) It is called a layer and defines a l-fold cross validation, mapping the input vector to the output vector. The first layer h (0) Defined as the input layer, the last layer h (nL) As the output layer,
[0082]
[0083] The layer h between the input layer and the output layer (l) It is called the hidden layer and is defined as:
[0084]
[0085] in is the lth layer h (l) The nth neural unit, n u represents the total number of neural units in each layer, is the lth layer h (l) The weight vector of the nth neural unit in (l-1) is the output of the previous layer, where the bias value is absorbed;
[0086] S21:φ (l) :R→R is a nonlinear activation function that takes into account historically relevant input and output data and, The formula for the hidden layer is:
[0087]
[0088] Since time is discrete, each discrete time slice of the input vector of the previous layer is processed in turn: Weight Shared across all time steps;
[0089] S3: Establish a neural network model that simulates the dynamic behavior of biological neurons. The spiking neural network is SNN. SNN is a history-related ANN that introduces memory effects through a biologically inspired process. Therefore, the activation function φ in S21 (l) It can be expressed as:
[0090] and in, is the membrane potential of the nth neuron at time t, represents the membrane threshold, is the membrane potential decay rate, are the standard ANN weights multiplied by the previous layer of the current time step, and SNN is a history-dependent ANN, where Also known as the LIF neuron model.
[0091] S31: The membrane potential does not reach the membrane threshold or remains silent through time evolution (φ spk =0), if the membrane potential exceeds the membrane threshold, the SNN activates the neural unit and outputs a discrete pulse (φ spk =1), these pulses are called spikes, in S3, It is called the reset mechanism. Once a spike is emitted, the membrane potential is reset. The membrane threshold and membrane potential decay rate are optimized during training, so that the optimized parameters of SNN are: Input spike Triggering a change in membrane potential when sufficient excitation exceeds the threshold U th r The neuron emits an output spike when
[0092] S32: A spike is a single event expressed by a Dirac delta function or a discrete pulse in a continuous or discrete setting. A group of spikes over time is called a spike train, i = i t ,t=0,…,n t ] For this purpose, spiking neurons are trained with spikes at time intervals consisting of either a spike (1) or zero input (0), and the membrane potential The incoming spike i t regulation, in the absence of input spikes, due to the membrane decay rate The membrane voltage decays over time;
[0093] S33: Use the backpropagation algorithm from a standard artificial neural network, use a proxy gradient to replace the activation during backpropagation, and use a fast sigmoid function:
[0094]
[0095] where k is the parameterized slope, and for some input x, the proxy φ surr (x) is continuously differentiable and preserves the gradient dynamics of the network, using φ' surr Used for back-propagation training;
[0096] S4: Design network structure: First, normalize the real data between 0 and 1, and then directly input the normalized data into the first input layer of the network. The input layer will select the appropriate encoding method according to the real data. Since the membrane potential directly outputs a continuous value, the decoding part directly uses the membrane potential value of the output layer as a real number to represent the calculated temperature value. In the design network structure, for the real data x with a time step of t t , directly input it into the input layer in the network structure In Chinese, that is: The hidden layer For network transmission, and transmit the data of the hidden layer to the next layer, the last output layer The output of the decoder layer is designed It outputs the membrane potential of the output layer directly as a real number, namely:
[0097]
[0098] The sum of the membrane potential of each time step of the output layer is taken as the final required temperature real number y t ,Right now: Where n0 represents the number of times a set of input data is simulated in the network.
[0099] In S4, the real data x t After normalization preprocessing, the data is directly transmitted to the input layer in a constant current manner, and then reaches the output layer, i.e. the decoding layer, after passing through the hidden layer. Then the required real number y is obtained by summing up the membrane potentials of all simulated decoding layers t ;
[0100] Data collection was performed on a computer located in a controlled environment with an air-conditioning system that simulated different environmental conditions;
[0101] Before starting the data collection process, we must thoroughly evaluate and terminate unnecessary processes, excluding essential system processes that are required for the computer to run. This precaution ensures that extraneous process factors do not interfere with the integrity of the data collection process. After completing the data collection process, it is recommended that the computer be left idle for a period of 1 hour. For example, the CPU frequency may be alternating at intervals, with the CPU running for a period of time and then entering a resting state at t2, and data collected every t3 during a period T. This approach ensures that the state of the computer before subsequent iterations of the initialization number are very similar to the state during the initial run. As a result, the data process that may affect the continuous collection is minimized.
[0102] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.
Claims
1. A processor temperature calculation method based on pulse neural network, characterized in that: The specific steps include: S1: Collect data to form a data set. Use the CPU data collection framework to collect CPU hardware information and temperature data when the CPU is periodically in full load and idle state. S2: Establish a densely connected feedforward artificial neural network, which is ANN. Specifically: From the input space To output space function, given by the nonlinear function h (l) The combination of definitions makes: here, Represents dimension d x The input vector is Represents dimension d y The output vector of the nonlinear function h (l) It is called a layer and defines a l-fold cross validation, mapping the input vector to the output vector. The first layer h (0) Defined as the input layer, the last layer h (nL) As the output layer, The layer h between the input layer and the output layer (l) It is called the hidden layer and is defined as: in is the lth layer h (l) The nth neural unit, n u represents the total number of neural units in each layer, is the lth layer h (l) The weight vector of the nth neural unit in (l-1) is the output of the previous layer, where the bias value is absorbed; S21:φ (l) :R→R is a nonlinear activation function that takes into account historically relevant input and output data and, The formula for the hidden layer is: Since time is discrete, each discrete time slice of the input vector of the previous layer is processed in turn: S3: Establish a neural network model that simulates the dynamic behavior of biological neurons. The spiking neural network is SNN. SNN is a history-related ANN that introduces memory effects through a biologically inspired process. Therefore, the activation function φ in S21 (l) It can be expressed as: and in, is the membrane potential of the nth neuron at time t, represents the membrane threshold, is the membrane potential decay rate, They are the standard ANN weights multiplied by the previous layer at the current time step; S31: The membrane potential does not reach the membrane threshold or remains silent through time evolution (φ spk =0), if the membrane potential exceeds the membrane threshold, the SNN activates the neural unit and outputs a discrete pulse (φ spk =1), these pulses are called spikes, in S3, It is called the reset mechanism. Once a spike is emitted, the membrane potential is reset. The membrane threshold and membrane potential decay rate are optimized during training, so that the optimized parameters of SNN are: S32: A spike is a single event expressed by a Dirac delta function or a discrete pulse in a continuous or discrete setting. A group of spikes over time is called a spike train, i = [i t ,t=0,…,n t ] For this purpose, spiking neurons are trained with spikes at time intervals consisting of either a spike (1) or zero input (0), and the membrane potential The incoming spike i t regulation, in the absence of input spikes, due to the membrane decay rate The membrane voltage decays over time; S33: Use the backpropagation algorithm from a standard artificial neural network, use a proxy gradient to replace the activation during backpropagation, and use a fast sigmoid function: where k is the parameterized slope, and for some input x, the proxy φ surr (x) is continuously differentiable and preserves the gradient dynamics of the network, using φ' surr Used for back-propagation training; S4: Design the network structure: first normalize the real data between 0 and 1, and then directly input the normalized data into the first input layer of the network. The input layer will select the appropriate encoding method based on the real data. Since the membrane potential directly outputs a continuous value, the decoding part directly uses the membrane potential value of the output layer as a real number to represent the calculated temperature value.
2. The method for calculating processor temperature based on pulse neural network according to claim 1, characterized in that: The data collected in S1 also includes CPU-related hardware counts and four core temperatures at different fan speeds and CPU frequencies, and the four core temperatures are Core 0, Core 1, Core 2, and Core 3. The data set was collected under the Ubuntu 20.04.5 LTS system, and the hardware device for collecting the data set was the CPU model i5-6500.
3. The method for calculating processor temperature based on pulse neural network according to claim 2, characterized in that: The CPU data collection framework is used for data collection using custom script execution, which follows the specified execution order outlined in the framework. The process is as follows: Step a: Generate a PTM configuration file and set the corresponding change strategy for the running status of each core; Step b: Adjust the operating frequency of the multi-core to match the frequency f supported by the CPU; Step c: Start two processes at the same time. Process 1 executes the data collection program, and process 2 starts McFTP to manage the scheduled operation or sleep state.
4. The method for calculating processor temperature based on pulse neural network according to claim 1, characterized in that: The weight of S21 Shared across all time steps.
5. The method for calculating processor temperature based on pulse neural network according to claim 1, characterized in that: The SNN in S3 is a history-related ANN, where Also known as the LIF neuron model.
6. The method for calculating processor temperature based on pulse neural network according to claim 1, characterized in that: The S31 input spike Triggering a change in membrane potential when sufficient excitation exceeds the threshold The neuron emits an output spike when .
7. The method for calculating processor temperature based on pulse neural network according to claim 1, characterized in that: The S4 designs the network structure for the real data x with time step t t , directly input it into the input layer in the network structure In Chinese, that is: The hidden layer For network transmission, and transmit the data of the hidden layer to the next layer, the last output layer The output of the decoder layer is designed It outputs the membrane potential of the output layer directly as a real number, namely: The sum of the membrane potential of each time step of the output layer is taken as the final required temperature real number y t ,Right now: Where n0 represents the number of times a set of input data is simulated in the network.
8. The method for calculating processor temperature based on pulse neural network according to claim 7, characterized in that: In S4, the real data After normalization preprocessing, it is directly transmitted to the input layer in a constant current manner, and reaches the output layer, i.e. the decoding layer, after passing through the hidden layer. , and then accumulate the membrane potentials of all simulated decoding layers to get the required real number .
