Biological neural network bionic method, system and related equipment

Abstract

The invention discloses a biological neural network bionic method and system and related equipment. The method comprises the steps of digital unit design simulating biological neuron characteristics, network dynamic adjustment based on connection plasticity, network structure generation of task adaptability, optimal allocation of hardware resources, real-time optimization of multi-dimensional performance, enhancement of safety and robustness, cross-platform hardware adaptation and the like. The problems of poor adaptability, insufficient robustness, low resource utilization rate and the like when a current neural network model processes complex tasks are solved. The brain-like computing system is efficient, safe and flexible, can be widely applied to the fields of intelligent sensing, autonomous decision making, man-machine interaction and the like, and provides important support for development of a new-generation artificial intelligence technology. Meanwhile, the invention further provides an Internet of Things device integrated with the bionic neural network and a chip implementation scheme, and the Internet of Things device and the chip implementation scheme have the advantages of being high in energy efficiency ratio, high in adaptability, wide in application range and the like.

Description

Technical Field

[0001] The present invention relates to the fields of brain-like intelligence, artificial intelligence and neural networks, and in particular to a biological neural network bionic method, system and related equipment. Technical Background

[0002] Artificial neural network is a mathematical model that imitates the structure and function of biological nervous system, and has been widely used in pattern recognition, natural language processing and other fields. However, traditional artificial neural network still has many limitations when dealing with complex tasks.

[0003] First, most existing neural network models use static network structures and connection weights, lacking the ability to dynamically adapt to environmental changes and task requirements. This makes it difficult for network models to flexibly handle changing real-world problems. Secondly, the robustness and fault tolerance of current neural networks need to be improved. Faced with abnormal situations such as noise interference, data loss, and hardware failure, the network often finds it difficult to maintain stable performance output. This limits its application in safety-critical areas. Furthermore, the computational efficiency and resource utilization of neural networks need to be optimized. Massive parameter storage and high-frequency data interaction place extremely high demands on hardware performance. However, the current resource allocation and scheduling strategies are difficult to fully utilize hardware performance, resulting in low energy efficiency. In addition, the heterogeneity of different hardware platforms brings inconvenience to the transplantation and deployment of neural networks. Most of the existing network implementations are optimized for specific platforms and lack portability and versatility. This is not conducive to the standardization and promotion of technology.

[0004] In summary, traditional neural networks still have deficiencies in adaptability, robustness, efficiency, etc., and a new brain-inspired computing paradigm is urgently needed to break through the current bottleneck and better support technological progress in the intelligent era. The present invention is proposed in this context in order to solve the above problems. Summary of the invention

[0005] The present invention proposes a bionic method, system and related equipment for biological neural network, aiming to solve the problems of poor adaptability, insufficient robustness and low resource utilization faced by traditional artificial neural networks when processing complex tasks.

[0006] In a first aspect, the present invention provides a biological neural network biomimetic method, comprising the following steps:

[0007] Step S101 designs a digital unit by imitating the excitation, inhibition and threshold characteristics of biological neurons, realizes dynamic adjustment of synaptic weights and activation function adaptation based on activity history, designs an initial connection matrix, and randomly initializes network parameters such as weights, biases, thresholds, and outputs an initial neural network structure; the digital neuron includes an input receiver, a signal integrator, a threshold controller, an output generator, and an adaptive regulator;

[0008] Step S102 is based on the initial neural network structure, and dynamically adjusts the synaptic connection strength according to the neuron activity level through the improved Hebbian learning rule, realizes the activity-based connection generation and elimination mechanism, designs the synchronous activation reinforcement rule, optimizes the network topology, and constructs a plastic network model; the improved Hebbian learning rule is:

[0009]

[0010] Where ΔWab is the connection weight from neuron a to b, η is the connection learning rate, Aa and Ab are the activity levels of neurons a and b respectively, and is the average activity level of neurons a and b;

[0011] Step S103 extracts task features and evaluates complexity C based on the plastic network model and specific task input, designs a network structure adaptability decision mechanism, and generates a task adaptation network through a progressive network reconstruction algorithm;

[0012] Step S104 uses the task to adapt to the network structure and hardware resource information, predict resource requirements, design a dynamic resource allocation mechanism, implement a load balancing algorithm, create a resource utilization efficiency optimization module, and output an optimized resource allocation plan;

[0013] Step S105 designs a multi-dimensional performance indicator collection module based on the optimized resource allocation scheme and runtime performance data, develops a performance data analysis algorithm, creates an automatic network parameter adjustment mechanism, implements a feedback optimization loop, and generates a performance optimization instruction set;

[0014] Step S106 receives the performance optimization instruction set and predefined security policies, develops a network anomaly detection algorithm, designs a critical task redundancy processing mechanism, implements a security policy execution module, creates a system recovery and fault tolerance mechanism, and outputs a security enhanced network model;

[0015] Step S107 designs a hardware abstraction layer based on the security enhanced network model and specific hardware specifications, develops a cross-platform compatibility module, implements a software and hardware interface optimization algorithm, creates a full system performance test and optimization suite, and generates a final optimized system.

[0016] Furthermore, the response process of the digital neuron in step S101 includes:

[0017] a. Signal reception, the neuron receives a set of input signals X = {x1, x2, ..., xn}, each of which is associated with a weight W = {w1, w2, ..., wn};

[0018] b. Signal integration, the neuron calculates the weighted sum and calculates the internal state of the neuron:

[0019] S=∑(xi·wi)+B

[0020] Where S represents the internal state of the neuron, xi represents the i-th signal in a set of signals received by the neuron, Wi represents the weight corresponding to the i-th signal in a set of signals received, and B represents the bias value;

[0021] c. Activate judgment and generate output. Use the activation function to simulate the response characteristics of biological neurons. If f(S) exceeds the activation threshold θ, the neuron generates an output pulse y:

[0022] f(S)=1 / (1+exp(-k·(S-θ)))y=H(f(S)-θ)wherein f(S) is the output value of the neuron activation function, k is the steepness parameter of the function, θ is the activation threshold, and H is a step function; the steepness parameter k and the activation threshold θ can be dynamically adjusted to simulate the plasticity of neurons.

[0023] Furthermore, the activation threshold θ can be dynamically adjusted according to the history of neuron activity. Frequently activated neurons will increase their activation threshold, achieving a "fatigue" effect:

[0024] θ(t+1)=θ(t)+α*(y(t)-yavg)

[0025] Among them, θ(t+1) represents the threshold of neuron i at time t+1, θ(t) represents the threshold of neuron i at time t, α is the neuron adaptation rate, y(t) is the activation frequency of the neuron before time t, and yavg is the average activity level of the neuron activation frequency.

[0026] Furthermore, the connection learning rate η in step S102 may be dynamically adjusted according to the overall activity level of the system to achieve stable learning:

[0027] η=η0 / (1+β*Atotal)

[0028] Where η0 is the initial connection learning rate, β is the decay factor, and Atotal is the total activity level of the entire network;

[0029] The synchronous activation reinforcement rule designed in step S102 is specifically: if the activity correlation of two unconnected neurons (c, d) exceeds a threshold θg, a connection is generated with a probability Pg:

[0030] Pg = sigmoid(Corr(c, d) - θg)

[0031] If the strength of an existing connected neuron (e, f) is lower than the threshold θe, the connection is deleted with probability Pe:

[0032] Pesigmoid(θe-Wef)

[0033] Where Wef represents the weight of neurons e to f, Corr(c, d) is the activity correlation of neurons c and d, and the sigmoid function is used to map the correlation or weight difference to the interval [0,1];

[0034] The process of optimizing the network extension in step S102 is specifically as follows: for each neuron, if the number of its input connections exceeds a preset threshold value Nmax, the system will retain the Nmax connections with the largest weights and eliminate the remaining connections.

[0035] Furthermore, the specific process of designing the network structure adaptability decision-making mechanism in step S103 is as follows:

[0036] a. Determine the number of neurons N that the network should contain based on the task complexity C value: N = N0*(1+tanh(α*(C-C0))), where N0 is the basic number of neurons, C0 is the complexity benchmark value, and α is the scaling factor;

[0037] b. Adjust the overall connection density D of the network: D = D0*(1+β*(C-C0)), where D0 is the basic connection density and β is the adjustment coefficient;

[0038] c. Adjust the modularity of the network J according to the characteristics of the task: J = J0 / (1+γ*C), where J 0 is the basic modularity level, and γ is the adjustment coefficient.

[0039] Furthermore, in step S104, resource demand is predicted, and the process uses a resource demand prediction function Ω: Where Ni is the resource demand vector of neuron i, ρi is the number of input connections of the neuron, σi is the computational complexity of the neuron, is the average activity of neurons;

[0040] The dynamic resource allocation mechanism designed in step S104 is specifically as follows: giving priority to meeting the neuron demand on the critical path; allocating the remaining resources in proportion to the demand while meeting the basic demand; and retaining 10% of the resources as a buffer for dynamic adjustment;

[0041] The load balancing process of the load balancing algorithm in step S104 is specifically as follows: calculate the load index L of each processing unit: L=∑(wi*ui), where ui is the utilization rate of resource type i, and wi is the weight corresponding to resource type i; if Lmax-Lmin>θ, where θ is a preset threshold, load redistribution is triggered; use a simulated annealing algorithm to find the optimal task redistribution solution, and the objective function is: min(∑(Li-Lavg)2), where Lavg is the average load of all processing units, and Li represents the load index of the i-th voluntary type.

[0042] Furthermore, the process of generating the performance optimization instruction set in step S105 includes:

[0043] a. Collect multi-dimensional performance indicators and analyze performance bottlenecks;

[0044] b. Optimize network parameters using a reinforcement learning-based approach, where the process uses a deep Q-net to learn the optimal parameter adjustment strategy: Q(s, a) = r + γ*max(Q(s′, a′)), where s is the current state, a is the parameter adjustment action, r is the performance improvement reward, and γ is the discount factor;

[0045] c. Dynamically select the most suitable optimization algorithm from the algorithm library, predict performance trends and make adjustments in advance.

[0046] Furthermore, the process of outputting the security enhanced network model in step S106 includes: continuously monitoring network behavior to detect anomalies, verifying the integrity of key data and codes, performing redundant processing on key tasks, defining service level agreements to achieve performance adaptation in the event of failures, and integrating privacy protection technology to enhance data security.

[0047] Furthermore, the process of generating the final optimized system in step S107 includes: in-depth analysis of hardware characteristics to guide optimization, quantification and compression of the security enhanced network model to reduce resource requirements, optimization of the calculation graph to improve execution efficiency, efficient scheduling of tasks, and minimization of system energy consumption.

[0048] In a second aspect, the present invention provides a biological neural network bionic system for executing any of the biological neural network bionic methods described above, including: a digital neuron design module, a network connection dynamic adjustment module, a task adaptation network generation module, a resource allocation optimization module, a performance optimization module, a security enhancement module and a hardware adaptation module.

[0049] In a third aspect, the present invention provides an Internet of Things device, comprising a processor and a memory, wherein the processor is used to execute a program stored in the memory, and when the program is running, any biological neural network bionic method provided in the specification of the present invention is executed.

[0050] In a fourth aspect, the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the program is executed by a processor, it implements any one of the biological neural network biomimetic methods provided in the specification of the present invention.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. Significantly improve the generalization and adaptability of the model. By introducing a dynamic and plastic network connection adjustment mechanism, the present invention enables the neural network to adaptively optimize its own structure and parameters according to task requirements and environmental feedback. This online learning capability enables the model to quickly adapt to new task domains, solving the problem of insufficient generalization ability faced by traditional static networks.

[0053] 2. Automatically generate targeted network structures. The task-adaptive network generation module proposed in the present invention can automatically design the optimal network structure according to the task characteristics. Compared with the traditional manually designed network, this method significantly improves the computational efficiency and task performance of the model while reducing the cost of network design. Targeted network customization enables optimal use of computing resources.

[0054] 3. Reduce hardware resource waste and improve system energy efficiency. The present invention adopts predictive resource demand estimation and multi-granularity dynamic scheduling mechanism, which can maximize resource utilization while meeting task performance requirements. Compared with the traditional fixed resource allocation strategy, the dynamic scheduling method of the present invention can significantly reduce the idle waste of computing resources, thereby achieving the same computing tasks at a lower hardware cost.

[0055] 4. Dynamic intelligent optimization of network hyperparameters. Traditional neural networks rely on manual setting of hyperparameters, and the tuning process relies on experience and is time-consuming and laborious. The present invention uses a reinforcement learning algorithm to continuously optimize network parameter configuration based on multi-dimensional real-time monitoring data. This method can quickly search for the optimal parameter combination, so that the system performance is in a dynamically optimal state.

[0056] 5. Enhance the security and robustness of the model. The present invention integrates a variety of privacy protection technologies, fault-tolerant design, and anti-attack detection mechanisms to comprehensively enhance the data security and robustness of the system. Compared with traditional models, the present invention can effectively resist security threats such as malicious data theft, parameter tampering, and the influence of uncertain factors such as node failure and computing power fluctuations.

[0057] 6. Support efficient edge deployment of models. For computing power-constrained environments such as the Internet of Things, the present invention proposes an integrated edge device solution. This solution achieves high compression and acceleration of brain-like models through software and hardware collaborative design, enabling real-time reasoning on resource-constrained terminal devices. Compared with the traditional cloud-based centralized computing paradigm, this edge computing solution significantly reduces communication bandwidth pressure and latency, making intelligent services closer to usage scenarios.

[0058] The above beneficial effects fully prove that the present invention has achieved a comprehensive breakthrough in learning efficiency, generalization ability, resource utilization, security, etc., and represents an important progress in the new generation of high-intelligence computing paradigm. These beneficial effects strongly support the actual deployment of the present invention in complex application scenarios such as intelligent perception, autonomous decision-making, and human-computer interaction, and also provide new ideas for the development of basic theories of artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required to be used in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those skilled in the art, other relevant drawings can be obtained based on these drawings without creative work.

[0060] Figure 1 A workflow diagram of a bionic method of biological neural network;

[0061] Figure 2 A schematic diagram of the structure of a biological neural network bionic system;

[0062] Figure 3 Schematic diagram of the structure of the bionic system of biological neural network running on Internet of Things (IoT) devices. Specific implementation plan

[0063] The present invention proposes a biological neural network bionic method, the working process is as follows Figure 1 As shown, it mainly includes steps S101 to S107.

[0064] The present invention proposes a biological neural network bionic system 200, such as Figure 2 The system shown includes: a digital neuron design module 201, a network connection dynamic adjustment module 202, a task adaptation network generation module 203, a resource allocation optimization module 204, a performance optimization module 205, a security enhancement module 206, and a hardware adaptation module 207.

[0065] The following are the workflow steps of the biological neural network bionics method and the specific process functions of the modules.

[0066] Step S101: Define digital units that mimic the characteristics of biological neurons, including excitation, inhibition, and threshold characteristics, and implement adjustable synaptic weights and activation functions. At the same time, design the initial connection matrix and initialize the network parameters, and finally output the initial neural network structure (INS).

[0067] Exemplarily, the structure of the digital neuron includes the following key components: input receiver: simulates the dendrites of biological neurons, responsible for receiving signals from other neurons. Each input has an associated weight value, indicating the degree of influence of the input on the neuron; signal integrator: corresponds to the cell body of biological neurons, responsible for summarizing all weighted inputs, and implements a nonlinear function to convert input signals into the internal state of the neuron; threshold controller: simulates the action potential threshold of biological neurons, determining whether the neuron is activated; output generator: corresponds to the axon, responsible for generating and sending output signals to other neurons; adaptive regulator: simulates the plasticity of biological neurons, and can adjust the characteristics of neurons according to past activity history.

[0068] Furthermore, the digital neuron performs the following steps, and the working process of the digital neuron is as follows: signal reception is performed, the neuron receives a set of input signals X={x1, x2, ..., xn}, each input signal is associated with a weight W={w1, w2, ..., wn}; after the signal is received, signal integration is performed, the neuron calculates the weighted sum, and adds a bias value B, and the internal state of the neuron can be calculated according to the following formula: S=∑(xi·wi)+B, where S represents the internal state of the neuron, xi represents the i-th signal in a set of signals received by the neuron, wi represents the weight corresponding to the i-th signal in the set of received signals, and B represents the bias value; after completing the signal integration, activation judgment is performed and output is generated, the internal state S of the neuron is converted by a nonlinear activation function f, and an improved logic function (sigmoid) can be used to simulate the response characteristics of biological neurons through the logic function: f(S)=1 / (1+exp(-k·(S-θ))), where f(S) is the output value of the neuron activation function, k is the steepness parameter of the function, and θ is the activation threshold. The two parameters, steepness parameter and activation threshold, can be adjusted dynamically to simulate the plasticity of neurons. If f(S) exceeds a certain activation threshold θ, the neuron will generate an output pulse. The output pulse u can be expressed as: y = H(f(S)-θ), where H is a step function.

[0069] Exemplarily, neurons adjust their parameters based on their activity history. For example, a frequently activated neuron may increase its threshold, achieving a "fatigue" effect: θ(t+1)=θ(t)+α*(y(t)-yavg), where θ(t+1) represents the threshold of neuron i at time y+1, θ(t) represents the threshold of neuron i at time t, α is the neuron adaptation rate, y(t) is the activation frequency of the neuron before time t, and yavg is the average activity level of the neuron activation frequency.

[0070] It should be noted that during the initialization process: the random value range of weight W is initialized to [-0.5, 0.5]; the bias value B is initialized to 0; the steepness parameter k of the activation function is initialized to 1, and the activation threshold θ is initialized to 0; T represents the "fatigue time constant" of the neuron, that is, the number of time steps required for the neuron to recover from the "excited" state to the "normal" state, and its size determines the duration of the neuron's "fatigue" effect, which is initialized to 0.5; the neuron adaptation rate α is initialized to 0.01. The output data of the digital neuron model include: the current state S of the neuron, the output of the activation function f(S), the pulse output y of the neuron, and the current parameter value (W, B, k, θ, T, α).

[0071] Step S102: Based on INS, realize dynamic adjustment of network connections. Develop a dynamic adjustment algorithm for connection strength, realize an activity-based connection generation and elimination mechanism, design synchronous activation reinforcement rules, and create a local connection optimization algorithm to form a plasticity network model (PNM).

[0072] Furthermore, the network connection dynamics simulates the plasticity and adaptability of biological neural networks, receives data from the output of the initial neural network structure INS digital neurons, and dynamically adjusts the structure and connection strength of the network based on these data.

[0073] Exemplarily, data from the output of INS digital neurons is received, including the current state of each neuron, activation threshold, pulse output, etc. The system organizes this data into a state matrix M, where Mij represents the connection state from the i-th neuron to the j-th neuron. For each pair of connected neurons (a, b) in the network, the system calculates their activity correlation and uses a modified Hebbian learning rule to adjust the connection strength: , where ΔWab is the connection weight from neuron a to b, η is the connection learning rate, Aa and Ab are the activity levels of neurons a and b respectively, and is the average activity level of neurons a and b. This rule allows the system to strengthen connections between neurons that are active together while weakening irrelevant connections. The connection learning rate η can be dynamically adjusted according to the overall activity level of the system to achieve stable learning: η = η0 / (1+β*Atotal), where η 0 is the initial connection learning rate, β is the decay factor, and Atotal is the total activity level of the entire network.

[0074] Exemplarily, the system periodically evaluates all possible pairs of neurons to decide whether to establish new connections between them or delete existing connections. This process can use the following rules: If the activity correlation of two unconnected neurons (c, d) exceeds a threshold θg, a connection is generated with probability Pg: Pg = sigmoid(Corr(c, d)-θg); If the strength of an existing connected neuron (e, f) is below a threshold θe, the connection is deleted with probability Pe: Pe = sigmoid(θe-Wef), where Wef represents the weight of neuron e to f, Corr(c, d) is the activity correlation of neurons c and d, and the sigmoid function is used to map the correlation or weight difference to the [0,1] interval.

[0075] Exemplarily, in order to maintain the efficiency and controllability of the network, the system will regularly optimize the local network topology. This process includes: for each neuron, if the number of its input connections exceeds the preset threshold Nmax, the system will retain the Nmax connections with the largest weight and eliminate the remaining connections. The system calculates the local clustering coefficient of each neuron and tries to make it close to the target value Ctarget by adding or deleting connections. This helps maintain the modular structure of the network while maintaining the efficiency of information transmission.

[0076] Exemplarily, after the above steps, the updated network connection matrix M (plastic network model, PNM) is output, where Wij represents the connection weight from neuron i to j. If Wij=0, it means that there is no direct connection between the two neurons.

[0077] Step S103: Receive PNM and specific task input, extract task features and evaluate complexity, design a network structure adaptability decision mechanism based on this information, and finally generate a task adaptation network (TAN) through a progressive network reconstruction algorithm.

[0078] Furthermore, based on the dynamically adjusted network structure, the most suitable neural network configuration for the current task is generated, the output data (connection matrix M) from the network connection dynamic adjustment module and the specific task input are received, and then a targeted optimized network structure is generated.

[0079] Exemplarily, a specific task input time step number T and the current network connection matrix M are received. The task input may be various forms of data, such as images, text, time series, etc. The system first standardizes the task input and converts it into a unified vector representation V. At the same time, an adaptive feature extractor is used to analyze the task input and dynamically adjust according to the complexity of the input. The feature extraction process can be expressed as: F = E (V, θe), where F is the extracted feature vector, E is the feature extraction function, and θe is the parameter set of the extractor. The parameters of the feature extractor are continuously optimized through the back propagation algorithm to adapt to different types of task inputs. Based on the extracted feature vector F, the system evaluates the complexity C of the task. A complexity C scoring function is defined: C = ∑ (wi*fi) + B, where fi is the i-th element of the feature vector F, wi is the corresponding weight, and B is the bias value. The weights and bias values ​​are adjusted according to the system's long-term task processing experience.

[0080] For example, according to the task complexity C, the system decides how to adjust the network structure. This process includes the following steps: the system determines the number of neurons N that the network should contain according to the C value: N = N0*(1+tanh(α*(C-C0))), where N 0 is the basic number of neurons, C 0 is the complexity reference value, α is the scaling factor; the system adjusts the overall connection density D of the network: D = D0 (+β*(C-C0)), where D 0 is the basic connection density, β is the adjustment coefficient; the system adjusts the modularity of the network according to the characteristics of the task: J = J0 / (1+γ*C), where J 0 is the basic modularity, and γ is the adjustment coefficient. Complex tasks tend to use a more integrated network structure.

[0081] Exemplarily, based on the adjusted network parameters (N, D, J), the system selects and combines the most suitable sub-networks from the existing network structure. This process can use an improved genetic algorithm: randomly generate multiple sub-network candidates that meet the N, D, J requirements; use a fast simulation process to evaluate the fitness of each candidate for the current task; retain a portion of the candidates with the highest fitness; generate new candidates by exchanging network fragments (crossover) and randomly adjusting connections (mutation); repeat the above steps until the preset number of iterations or fitness threshold is reached.

[0082] Exemplarily, after the above steps, a task adaptation network structure TAN is output, which includes: a selected neuron set, connection weights between the neurons, and the overall topological structure of the network.

[0083] Step S104: Utilize TAN and hardware resource information to predict resource demand, design a dynamic resource allocation mechanism, implement a load balancing algorithm, and create a resource utilization efficiency optimization module, and finally output an optimized resource allocation scheme (ORAS).

[0084] Furthermore, according to the task adaptation network structure (TAN) and available hardware resources, the allocation of computing resources is optimized to achieve efficient task processing. This module receives the output (TAN) from the task adaptation network generation module and the system hardware resource information, and then generates an optimized resource allocation plan.

[0085] Exemplarily, a task adaptation network structure (TAN) and currently available hardware resource information H are received. The hardware resource information includes the number of CPU cores, memory capacity, number of GPU units, etc. The system initializes a resource allocation matrix R, where Riq represents the number of resource type q allocated to neuron i.

[0086] Exemplarily, the system analyzes the structural characteristics of TAN and estimates the computational resources required for each neuron and connection. This process uses a resource requirement prediction function Ω: Where Ni is the resource requirement vector of neuron i, ρi is the in-degree (number of input connections) of the neuron, σi is the computational complexity of the neuron (determined by the type of activation function), is the average activity of neurons. The Ω function is implemented by a multi-layer perceptron, whose parameters are trained by historical data.

[0087] Exemplarily, based on the results of resource demand analysis, the system performs initial resource allocation. This process can use an improved knapsack algorithm to allocate limited resources to the neurons and connections with the most urgent needs. The allocation principles are as follows: give priority to meeting the needs of neurons on the critical path; allocate the remaining resources in proportion to the demand while meeting the basic needs; retain a certain proportion (such as 10%) of resources as a buffer for dynamic adjustment.

[0088] Exemplarily, the system continuously monitors the load status of each processing unit and makes dynamic adjustments when necessary. The load balancing process is as follows: Calculate the load index L of each processing unit: L = Σ(wi*ui), where ui is the utilization rate of resource type i, and wi is the weight corresponding to resource type i; if Lmax-Lmin>θ (θ is a preset threshold), trigger load redistribution; Use the simulated annealing algorithm to find the optimal task redistribution solution, and the objective function is: min(Σ(Li-Lavg)2), where Lavg is the average load of all processing units, and Li represents the load index of the i-th voluntary type.

[0089] Exemplarily, the system improves overall resource utilization by dynamically adjusting resource allocation strategies. This process includes: tracking the resource utilization efficiency E of each neuron: E = Output / Input, where Output is the output contribution of the neuron and Input is the amount of resources allocated to the neuron; for inefficient neurons, the system will gradually reduce their resource allocation and reallocate the released resources to efficient neurons; regularly perform "resource stress testing" to temporarily reduce the resource allocation of certain neurons and evaluate the impact on overall performance, so as to find the best balance point for resource allocation; while ensuring performance, the system also needs to optimize energy consumption. Energy management strategies include: for temporarily inactive neurons, reduce their operating frequency or put them into sleep mode; when the task load is light, integrate the workload into some processing units and allow other units to enter low power mode; dynamically adjust the balance weight between performance and energy consumption according to the urgency of the task and the energy consumption status.

[0090] Exemplarily, after the above steps, an optimized resource allocation scheme (ORAS) is output, which includes: resource allocation details of each neuron, task allocation scheme of processing units, and dynamic adjustment strategy and threshold setting.

[0091] Step S105: Based on ORAS and runtime performance data, design a multi-dimensional performance indicator collection module, develop a performance data analysis algorithm, create a network parameter automatic adjustment mechanism, and implement a feedback optimization loop to generate a performance-optimized instruction set (POIS).

[0092] Furthermore, it monitors system performance in real time, analyzes performance bottlenecks, and optimizes the overall performance of the system by dynamically adjusting various parameters. It receives the output from the resource allocation optimization module (ORAS) and various performance indicators during system operation, and then generates a series of performance optimization instructions.

[0093] Exemplarily, an optimized resource allocation scheme (ORAS) and real-time system performance data are received. Performance data include, but are not limited to: processing speed, response time, throughput, accuracy, energy consumption, etc. Use the sliding window method to preprocess these data to reduce the impact of noise; continuously collect multi-dimensional performance indicators and perform real-time analysis. This process includes: defining a comprehensive performance indicator P: P = ∑ (wi*pi), where pi is the standardized value of each performance indicator, and wi is the corresponding weight. The weight will be dynamically adjusted according to the characteristics of the current task; use time series analysis methods (such as differential integrated moving average autoregressive model) to predict short-term performance trends; calculate the correlation matrix between each performance indicator to identify potential performance influencing factors.

[0094] Illustratively, based on the performance analysis results, the system uses a multi-level bottleneck identification algorithm to locate performance problems: compare the performance differences between different processing units to identify load imbalance problems; analyze the processing time and resource utilization of each layer of the neural network to find the performance bottleneck layer; check the computing efficiency and communication overhead of individual neurons to locate specific performance obstacles.

[0095] Exemplarily, for the identified bottlenecks, the system adopts a parameter tuning method based on reinforcement learning: define the parameter space Θ, including connection learning rate, batch size, activation function parameters, etc. Use the deep Q network (DQN) to learn the optimal parameter adjustment strategy: Q(s, a) = r + γ*max(Q(s′, a′)), where s is the current state, a is the parameter adjustment action, r is the performance improvement reward, and γ is the discount factor; in each optimization cycle, the system selects the parameter adjustment action that maximizes the Q value; improve the stability and efficiency of learning through techniques such as experience replay and target network.

[0096] For example, the system maintains an algorithm library containing a variety of optimization algorithms for different scenarios. Based on the current task characteristics and system status, the most suitable algorithm is dynamically selected: the multi-armed bandit algorithm is used for preliminary screening to quickly find potential efficient algorithms; small-scale tests are conducted on the screened algorithms to collect performance data; the Bayesian optimization method is used to fine-tune the hyperparameters of the selected algorithm; and the algorithm performance statistics are continuously updated for future selection decisions.

[0097] For example, to predict and prevent potential performance degradation: establish a system performance model and map various factors (such as data characteristics, network structure, resource allocation, etc.) to performance indicators; use this model to conduct "what-if" analysis to evaluate system performance under different conditions; if performance bottlenecks are predicted to occur in the future, the system will adjust resource allocation or network structure in advance to prevent them.

[0098] Exemplarily, after the above steps, a performance optimization instruction set (POIS) is output, which includes: parameter adjustment instructions (such as connection learning rate changes, batch size adjustments, etc.), algorithm switching suggestions, resource reallocation suggestions, network structure adjustment suggestions, early warning information and corresponding preventive measures.

[0099] Step S106: Receive POIS and predefined security policies, develop network anomaly detection algorithms, design critical task redundancy processing mechanisms, implement security policy execution modules, and create system recovery and fault tolerance mechanisms, and finally output a security-enhanced network model (SENM).

[0100] Exemplarily, it ensures that the system has sufficient security and robustness while maintaining high performance. This module receives the output POIS from the performance optimization module and the predefined security policy, and then generates a security-enhanced network model.

[0101] Exemplarily, a performance optimization instruction set (POIS) and a predefined security policy set SECpolicies are received. The security policy includes various security rules, thresholds, and response measures. The system initializes a security status matrix M for recording and tracking the security status of each component.

[0102] Exemplarily, the system continuously monitors the behavior and output of the network to detect potential anomalies and attacks. This process includes: defining a multi-dimensional anomaly indicator vector A: A = [a 1 ,a 2 ,...,an], where ai represents different types of abnormal indicators, such as input distribution shift, output anomaly, resource usage anomaly, etc.; unsupervised learning methods (such as the isolation forest algorithm) are used to detect anomalies. For each indicator ai, its anomaly score ADscore is calculated: Where h(ai) is the path length of ai in the isolation tree, E(h(ai)) is the average path length, and c(N) is the normalization factor of the data volume N; if the ADscore exceeds the preset threshold τi, the corresponding response measures are triggered, such as: temporarily isolating suspicious components, increasing the monitoring frequency, and starting a deep scanning program.

[0103] Exemplarily, to ensure that data is not tampered with during transmission and storage, the system adopts the following mechanism: implement an improved Merkle tree structure for efficiently verifying the integrity of large amounts of data: divide the data into blocks of a fixed size, calculate the hash value of each data block, then combine adjacent hash value pairs and hash them again until the root hash is obtained; regularly calculate and update the Merkle tree root hash of key data, and store it in a secure area; when accessing data, quickly verify data integrity by comparing the currently calculated root hash with the stored root hash.

[0104] Exemplarily, to prevent the core components of the system from being maliciously modified, the following protective measures are implemented: use a hardware root of trust (such as a Trusted Platform Module, TPM) to verify the integrity of the startup code and to verify all components from the bottom layer to the application layer step by step; regularly calculate the hash value of key code segments, and compare the calculation result with the pre-stored whitelist. If a mismatch is found, immediately isolate the suspicious component and issue an alarm; implement a code signing mechanism to ensure that only authorized updates can be applied.

[0105] Exemplarily, in order to improve the robustness and reliability of the system, the following mechanisms are implemented: for critical tasks, they are executed in parallel on multiple independent processing units, and the Byzantine fault-tolerant protocol is used to check the consistency of the results: if the results of more than 2 / 3 of the execution units are consistent, the result is adopted; otherwise, it is marked as a suspicious task and reallocated; at the same time, system status snapshots are created regularly, including network weights, connection structures, etc.; snapshot data is stored incrementally to reduce storage overhead; when a serious anomaly is detected, the system can roll back to the most recent stable state; a service-level agreement (SLA) is defined to clarify performance requirements under different conditions; when resources are limited or some components fail, the system can automatically downgrade to ensure the normal operation of core functions.

[0106] Exemplarily, to protect user data privacy and system sensitive information, the following technologies are adopted: adding calibrated noise during data collection and analysis, and the noise level ε is dynamically adjusted according to the data sensitivity and the desired privacy protection level; implementing a decentralized learning framework so that the model can be trained without direct access to the original data, and using a secure aggregation protocol to ensure that intermediate results do not leak individual information; performing homomorphic encryption on sensitive data, allowing calculations in the encrypted domain, and decrypting the results only when necessary, minimizing the exposure of plaintext data.

[0107] Exemplarily, after the above steps, a security enhanced network model (SENM) is output, which includes: original network structure and parameters, security policy and threshold settings, anomaly detection model, integrity verification data, fault tolerance and recovery mechanism configuration, and privacy protection parameters.

[0108] Step S107: Based on SENM and specific hardware specifications, design a hardware abstraction layer, develop a cross-platform compatibility module, implement a software and hardware interface optimization algorithm, and create a full-system performance test and optimization suite to ultimately generate a final optimized system (FOS).

[0109] Exemplarily, the Security Enhanced Network Model (SENM) is efficiently deployed on various hardware platforms. This module receives the SENM and the specification information of the target hardware, and then generates the final system implementation optimized for the specific hardware.

[0110] Exemplarily, a Security Enhanced Network Model (SENM) and detailed specification information H of the target hardware are received. The hardware information includes processor architecture, memory hierarchy, cache size, instruction set characteristics, etc. The system initializes a hardware abstraction layer HAL to shield the underlying hardware differences.

[0111] Exemplarily, the system deeply analyzes the characteristics of the target hardware to guide the subsequent optimization process: processor analysis: including instruction set characteristics (such as Single Instruction Multiple Data, SIMD), pipeline depth and branch prediction capabilities, cache structure and capacity; then memory system analysis: including memory bandwidth and latency, memory hierarchy, direct memory access (DMA) capabilities; secondly, parallel capability analysis: including multi-core / multi-thread support, general-purpose computing on Graphics Processing Units (GPGPU) computing capabilities, and dedicated AI accelerator characteristics (such as Tensor Processing Unit Tensor Processing Unit, TPU).

[0112] Exemplarily, according to the hardware characteristics, the SENM is appropriately quantized and compressed to reduce the computation and storage requirements: according to the data type supported by the hardware, the floating-point weights are quantized to fixed-point representations, and a non-uniform quantization method such as k-means quantization is used to minimize the accuracy loss, and the weight w is mapped to the nearest quantization center ci: Q(w)=argmini||w-ci||, where argmini represents finding the index i that minimizes ||w-ci||, and ||w-ci|| represents the distance between w and ci; the importance of each weight is evaluated using the importance scoring function I(w), and weights with scores below a threshold τ are removed: if I(w)>τ, w′=w, else 0, where w ′ Represents the new weight, w represents the original weight, I(w) represents the importance scoring function, τ represents the threshold, and the algorithm indicates that if the importance score of weight w is higher than the threshold τ, the original weight is retained; otherwise, the weight is set to 0; a small student network is trained to imitate the behavior of SENM: L = α*Lce(y, y′)+β*Lkd(z, z′), where β represents the weight coefficient for knowledge distillation loss (Lkd), α represents the weight coefficient for cross entropy loss (Lce), y represents the true label, y′ represents the predicted label, Lce is the cross entropy loss, Lkd is the knowledge distillation loss, z and z ′ are the soft labels for the teacher and student networks, respectively.

[0113] Exemplarily, the computational graph of the network is optimized to improve execution efficiency: identify continuous operators that can be merged, such as the deep learning operation combination convolution-batch normalization-ReLU activation (Conv+BN+ReLU), which is constructed by convolution layers, batch normalization layers, and activation function layers, and create fused operators to reduce memory access and computational overhead; rearrange the storage order of weights and activation values ​​according to the memory access pattern of the hardware, implement a data prefetch mechanism, and reduce cache misses; analyze computational dependencies, reorder the computational order to maximize parallelism, and use soft pipelining technology to overlap computation and data transmission.

[0114] Exemplarily, efficient task allocation and scheduling are achieved based on the parallel computing capabilities of the hardware: network computing is decomposed into task units of appropriate granularity, and a task dependency graph G(V,E) is constructed, where V is a set of tasks and E represents a dependency relationship; the earliest start time (EST) and the latest end time (LFT) of each task are calculated, and tasks are arranged in ascending order of (LFT-EST), and computing resources are allocated to each task in turn, giving priority to tasks on the critical path; a work stealing algorithm is implemented to allow idle processing units to "steal" tasks from busy units, and an adaptive threshold mechanism is used to dynamically adjust the stealing strategy according to the system load.

[0115] For example, while ensuring performance, minimize system energy consumption: dynamically adjust processor frequency according to task urgency and computing intensity, implement a predictive model, and adjust the frequency in advance to minimize performance impact; identify idle functional units during the computing process, implement fine-grained power gating, and shut down unnecessary units; implement a data reuse strategy to minimize high-energy external memory accesses, and use near-memory computing technology to delegate part of the computing to the memory system.

[0116] Exemplarily, after the above steps, this module outputs the final optimized system (FOS), which includes: hardware-specific network implementation code, optimized weights and parameters, task scheduling and resource allocation strategies, hardware control instruction settings (such as dynamic voltage and frequency scaling, Dynamic Voltage and Frequency Scaling, DVFS) and performance monitoring and adaptive adjustment mechanisms.

[0117] Furthermore, the bio-neural network bionic system 200 proposed by the present invention is as follows: Figure 2As shown, it is mainly composed of the following modules: digital neuron design module 201, network connection dynamic adjustment module 202, task adaptation network generation module 203, resource allocation optimization module 204, performance optimization module 205, security enhancement module 206, and hardware adaptation module 207. Among them, the digital neuron design module is responsible for defining and initializing digital units that imitate the characteristics of biological neurons; the network connection dynamic adjustment module realizes the plasticity and dynamic adjustment of network connections; the task adaptation network generation module generates an adaptive network structure according to specific task requirements; the resource allocation optimization module optimizes the allocation and utilization of system resources; the performance optimization module monitors and optimizes system performance in real time; the security enhancement module improves the security and robustness of the system; the hardware adaptation module realizes cross-platform compatibility and system optimization.

[0118] Furthermore, the present invention also provides an Internet of Things (IoT) device.

[0119] like Figure 3 As shown, the Internet of Things (IoT) device 300 is mainly composed of a processor 301 and a memory 302, which are connected via a bus 303, which can be an I2C (Inter-Integrated Circuit) bus or the like.

[0120] Specifically, processor 301 undertakes computing and control functions and is the core of the operation of the Internet of Things (IoT) device. Processor 301 can be a low-power microcontroller chip of the Internet of Things, or it can be other general-purpose processors, digital signal processors (Digital Signal Processing, DSP), application specific integrated circuits (Application specific integrated circuit, ASIC), field programmable gate arrays (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor here can be a microprocessor or any conventional processor. Memory 302 can be in various forms such as Flash chip, read-only memory ROM, disk, CD, U disk or mobile hard disk.

[0121] It should be noted that Figure 3 The structure shown in the figure is only a partial structural diagram related to the embodiment of the present invention, and does not limit the Internet of Things (IoT) device to which the solution is applied. The actual Internet of Things (IoT) device may include more or fewer components, or merge some components, or use different component arrangements.

[0122] It is worth noting that the main function of the processor 301 is to execute the computer programs stored in the memory 302. When running these programs, the processor can implement various biological neural network bionic systems proposed in the embodiments of the present invention.

[0123] In a specific embodiment of the present invention, the processor is designed to execute a computer program stored in the memory. When the program is running, the processor will complete the following steps:

[0124] a. Design digital units that mimic the characteristics of biological neurons, including excitation, inhibition, and threshold characteristics, implement adjustable synaptic weights and activation functions, initialize network parameters, and output the initial neural network structure (INS);

[0125] b. Based on INS, realize dynamic adjustment of network connections, including developing a dynamic adjustment algorithm for connection strength, realizing an activity-based connection generation and elimination mechanism, designing synchronous activation reinforcement rules, and creating a local connection optimization algorithm to form a plasticity network model (PNM);

[0126] c. Based on PNM and specific task input, extract task features and evaluate complexity, design a network structure adaptive decision mechanism, and generate a task adaptive network (TAN) through a progressive network reconstruction algorithm;

[0127] d. Use TAN and hardware resource information to predict resource requirements, design a dynamic resource allocation mechanism, implement a load balancing algorithm, create a resource utilization efficiency optimization module, and output an optimized resource allocation solution (ORAS);

[0128] e. Based on ORAS and runtime performance data, design a multi-dimensional performance indicator collection module, develop a performance data analysis algorithm, create an automatic network parameter adjustment mechanism, and implement a feedback optimization loop to generate a performance optimization instruction set (POIS);

[0129] f. Receive POIS and predefined security policies, develop network anomaly detection algorithms, design critical task redundancy processing mechanisms, implement security policy execution modules, and create system recovery and fault tolerance mechanisms to output security enhanced network models (SENMs);

[0130] g. Based on SENM and specific hardware specifications, design the hardware abstraction layer, develop cross-platform compatibility modules, implement software and hardware interface optimization algorithms, and create a full-system performance test and optimization suite to ultimately generate the Final Optimized System (FOS).

[0131] The embodiment of the present invention also provides a computer-readable storage medium, in which one or more programs are stored. These programs can be executed by one or more processors to implement the steps of any biological neural network bionic system described in the embodiment of the present invention. The storage medium can be an internal storage unit of the aforementioned Internet of Things (IoT) device, such as a hard disk or memory. It can also be an external storage device, such as a plug-in hard disk, a smart memory card (SmartMediaCard, SMC), a secure digital (SecureDigitalCard, SD) card, a flash memory card, etc. It should be understood by those skilled in the art that the steps in the above method, the functional modules in the system or device can be implemented by software, firmware, hardware or a combination thereof. In hardware implementation, the division of each functional module does not necessarily correspond to the division of physical components. A physical component may perform multiple functions, or a function may be completed by multiple physical components in collaboration. These components can be software executed by a processor (such as a CPU, a digital signal processor (or microprocessor), or they can be hardware implementations, or integrated circuits (such as Application Specific Integrated Circuit, ASIC; Digital Signal Processor, DSP). These software can be distributed on computer-readable media, including computer storage media (non-temporary) and communication media (temporary). Computer storage media include but are not limited to RAM, ROM, electrically erasable programmable read-only memory (EEPROM), flash memory, CD-ROM, DVD, tape, disk, etc., for storing information. Communication media include any media that transmit computer-readable instructions, data structures, etc. in modulated data signals.

Claims

1. A biological neural network biomimetic method, characterized in that: The steps include: Step S101 designs a digital unit by imitating the excitation, inhibition and threshold characteristics of biological neurons, realizes dynamic adjustment of synaptic weights and activation function adaptation based on activity history, designs an initial connection matrix, and randomly initializes network parameters such as weights, biases, thresholds, and outputs an initial neural network structure; the digital neuron includes an input receiver, a signal integrator, a threshold controller, an output generator, and an adaptive regulator; Step S102 is based on the initial neural network structure, and dynamically adjusts the synaptic connection strength according to the neuron activity level through the improved Hebbian learning rule, realizes the activity-based connection generation and elimination mechanism, designs the synchronous activation reinforcement rule, optimizes the network topology, and constructs a plastic network model; the improved Hebbian learning rule is: Where, ΔW ab is the connection weight from neuron a to b, η is the connection learning rate, A a and A b are the activity levels of neurons a and b, respectively, and is the average activity level of neurons a and b; Step S103 extracts task features and evaluates complexity C based on the plastic network model and specific task input, designs a network structure adaptability decision mechanism, and generates a task adaptation network through a progressive network reconstruction algorithm; Step S104 uses the task to adapt to the network structure and hardware resource information, predict resource requirements, design a dynamic resource allocation mechanism, implement a load balancing algorithm, create a resource utilization efficiency optimization module, and output an optimized resource allocation plan; Step S105 designs a multi-dimensional performance indicator collection module based on the optimized resource allocation scheme and runtime performance data, develops a performance data analysis algorithm, creates an automatic network parameter adjustment mechanism, implements a feedback optimization loop, and generates a performance optimization instruction set; Step S106 receives the performance optimization instruction set and predefined security policies, develops a network anomaly detection algorithm, designs a critical task redundancy processing mechanism, implements a security policy execution module, creates a system recovery and fault tolerance mechanism, and outputs a security enhanced network model; Step S107 designs a hardware abstraction layer based on the security enhanced network model and specific hardware specifications, develops a cross-platform compatibility module, implements a software and hardware interface optimization algorithm, creates a full system performance test and optimization suite, and generates a final optimized system.

2. The biological neural network biomimetic method according to claim 1, characterized in that: The response process of the digital neuron in step S101 includes: a. Signal reception: The neuron receives a set of input signals X = {x1, x2, ..., x n }, each input signal is associated with a weight W = {w1, w2, ..., w n }Associated; b. Signal integration, the neuron calculates the weighted sum and calculates the internal state of the neuron: S=∑(x i ·w i )+B Where S represents the internal state of the neuron, x i represents the i-th signal in a set of signals received by the neuron, W i represents the weight corresponding to the i-th signal in a set of received signals, and B represents the bias value; c. Activate judgment and generate output. Use the activation function to simulate the response characteristics of biological neurons. If f(S) exceeds the activation threshold θ, the neuron generates an output pulse y: f(S)=1 / (1+exp(-k·(S-θ))) y=H(f(S)-θ) Wherein f(S) is the output value of the neuron activation function, k is the steepness parameter of the function, θ is the activation threshold, and H is a step function; the steepness parameter k and the activation threshold θ can be dynamically adjusted to simulate the plasticity of neurons.

3. The biological neural network biomimetic method according to claim 2, characterized in that: The activation threshold θ can be dynamically adjusted according to the history of neuronal activity. Frequently activated neurons will increase their activation threshold, achieving a fatigue effect: θ(t+1)=θ(t)+α*(y(t)-yav g ) Among them, θ(t+1) represents the threshold of neuron i at time t+1, θ(t) represents the threshold of neuron i at time t, α is the neuron adaptation rate, y(t) is the activation frequency of the neuron before time t, and y avg is the average activity level of how often neurons fire.

4. The biological neural network biomimetic method according to claim 1, characterized in that: The connection learning rate η in step S102 can be dynamically adjusted according to the overall activity level of the system to achieve stable learning: η=η0 / (1+β*A total ) Where η0 is the initial connection learning rate, β is the decay factor, and A total is the total activity level of the entire network; The synchronous activation reinforcement rule designed in step S102 is specifically: if the activity correlation of two unconnected neurons (c, d) exceeds a threshold value θ g , then with probability P g Generate a connection: P g =sigmoid(Corr(c,d)-θ g ) If the strength of the existing connected neuron (e, f) is lower than the threshold θ e , then with probability P e Delete the connection: Pe=sigmoid(θ e -W ef ) Among them, W ef represents the weights of neurons e to f, Corr(c, d) is the correlation of the activities of neurons c and d, and the sigmoid function is used to map the correlation or weight difference to the interval [0, 1]; The process of optimizing the network extension in step S102 is specifically as follows: for each neuron, if the number of its input connections exceeds a preset threshold N max , the system will retain the N with the largest weight max connections and eliminate the rest.

5. The biological neural network biomimetic method according to claim 1, characterized in that: The specific process of designing the network structure adaptability decision-making mechanism in step S103 is as follows: a. Determine the number of neurons N that the network should contain based on the task complexity C value: N = N0*(1+tanh(α*(C-C0))), where N0 is the basic number of neurons, C0 is the complexity benchmark value, and α is the scaling factor; b. Adjust the overall connection density D of the network: D = D0(1+β*(C-C0)), where D0 is the basic connection density, β is the adjustment coefficient; c. Adjust the modularity of the network J according to the characteristics of the task: J = J0 / (1+γ*C), where J0 is the basic modularity and γ is the adjustment coefficient.

6. The biological neural network biomimetic method according to claim 1, characterized in that: In step S104, resource demand is predicted, and the process uses a resource demand prediction function Ω: Where N i is the resource demand vector of neuron i, ρ i is the number of input connections of the neuron, σ i is the computational complexity of the neuron, is the average activity of neurons; The dynamic resource allocation mechanism designed in step S104 is specifically as follows: giving priority to meeting the neuron demand on the critical path; allocating the remaining resources in proportion to the demand while meeting the basic demand; and retaining 10% of the resources as a buffer for dynamic adjustment; The load balancing process of the load balancing algorithm in step S104 is specifically as follows: calculate the load index L of each processing unit: L = ∑(w i *u i ), where u i is the utilization rate of resource type i, w i is the weight corresponding to resource type i; if there is L max -L min >θ, where θ is the preset threshold that triggers load redistribution; the simulated annealing algorithm is used to find the optimal task redistribution solution, and the objective function is: min(∑(L i -L avg ) 2 ), where L avg is the average load of all processing units, and Li represents the load indicator of the i-th voluntary type.

7. The biological neural network biomimetic method according to claim 1, characterized in that: The process of generating the performance optimization instruction set in step S105 includes: a. Collect multi-dimensional performance indicators and analyze performance bottlenecks; b. Optimize network parameters using a reinforcement learning-based approach, where the process uses a deep Q-net to learn the optimal parameter adjustment strategy: Q(s, a) = r + γ*max(Q(s′, a′)), where s is the current state, a is the parameter adjustment action, r is the performance improvement reward, and γ is the discount factor; c. Dynamically select the most suitable optimization algorithm from the algorithm library, predict performance trends and make adjustments in advance.

8. The biological neural network biomimetic method according to claim 1, characterized in that: The process of outputting the security enhanced network model in step S106 includes: continuously monitoring network behavior to detect anomalies, verifying the integrity of key data and codes, performing redundancy processing on key tasks, defining a service level agreement to achieve performance adaptation in the event of a failure, and integrating privacy protection technology to enhance data security.

9. The biological neural network biomimetic method according to claim 1, characterized in that: The process of generating the final optimized system in step S107 includes: in-depth analysis of hardware characteristics to guide optimization, quantification and compression of the security enhanced network model to reduce resource requirements, optimization of the calculation graph to improve execution efficiency, efficient scheduling of tasks, and minimization of system energy consumption.

10. A biological neural network bionic system, characterized in that: The bionic method of biological neural network used to execute any one of claims 1-9 comprises: a digital neuron design module, a network connection dynamic adjustment module, a task adaptation network generation module, a resource allocation optimization module, a performance optimization module, a security enhancement module and a hardware adaptation module.

11. An Internet of Things device, comprising a processor and a memory, characterized in that: The processor is used to execute the program stored in the memory, and when the program is running, the biological neural network biomimetic method according to any one of claims 1 to 9 is executed.

12. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the biological neural network biomimetic method according to any one of claims 1 to 9.

Images