A neural memory-based autonomous driving safety decision method, system and device

Abstract

The application provides a neural memory-based automatic driving safety decision method, system and device, constructs a human-like neuron vehicle model based on biological neuron characteristics, constructs a memory knowledge base based on traffic environment data, trains the human-like neuron vehicle model based on the memory knowledge base, calculates the stability index of the trained human-like neuron vehicle model, calculates the decision difference between the human-like neuron vehicle model and a test driver based on a complex traffic scene if the stability index meets the preset requirements, performs consistency enhancement training on the human-like neuron vehicle model based on the decision difference, and executes automatic driving safety decision based on the human-like neuron vehicle model after consistency enhancement training. By constructing the human-like neuron vehicle model, the human brain memory and cognition are simulated, the humanization processing capability of the system for traffic information is improved, and the decision process is explained and the safety online verification is realized through checking and logical verification.

Description

Technical Field

[0001] This application relates to the field of autonomous driving control technology, specifically to an autonomous driving safety decision-making method, system, and device based on neural memory. Background Technology

[0002] Autonomous driving technology has undergone rapid development from assisted driving to conditional autonomous driving. Currently, commercial and academic research solutions in the field of autonomous driving safety decision-making can be mainly divided into two major technical systems: rule-driven and data-driven. Both types of solutions have inherent technical defects, making it difficult to replicate the anthropomorphic safety decision-making logic of human drivers and unable to adapt to the safety decision-making needs of complex and dynamic traffic scenarios. Among them, rule-driven decision-making methods rely on expert experience and traffic regulations to build a fixed rule base, and realize driving decisions through preset working condition thresholds, avoidance strategies, and control constraints. It has the advantages of logical transparency, high stability, and strong controllability. However, this method has significant limitations. Real road traffic scenarios have extremely high diversity, randomness, and dynamic interactivity. Preset rules can only cover conventional structured standard working conditions and cannot exhaust the long-tail edge working conditions and sudden dangerous scenarios in traffic scenarios. When facing undefined interactive scenarios such as mixed traffic of pedestrians and vehicles, unconventional cutting in, and emergency avoidance, it is prone to problems such as rigid strategies, delayed decisions, and response failures, and its scenario generalization ability is seriously insufficient. Data-driven end-to-end decision-making models based on deep learning and reinforcement learning, which autonomously learn the correlation characteristics between scenarios and decisions using massive amounts of road data, possess stronger adaptive capabilities compared to rule-based methods and represent the mainstream research direction for advanced autonomous driving decision-making. However, these models are typical black-box models, lacking interpretability in their reasoning logic. Their most critical weakness is the absence of neural memory and cognitive mechanisms unique to human drivers. More importantly, existing methods lack simulations of human drivers' memory recall and scenario analogy cognitive mechanisms, failing to make intuitive decisions quickly based on similar historical scenarios. This results in overly aggressive or conservative responses in sudden dangerous situations, making it difficult to meet the comprehensive safety and human-like requirements of highly automated driving.

[0003] Existing data-driven models only output decisions based on real-time scene features and do not draw on human memory to deduce cognitive logic. In situations of sudden danger, unfamiliar scenarios, and highly interactive conditions, they often make decisions that are either too aggressive or too conservative, which deviates significantly from human driving habits. This not only reduces driving comfort and leads to a lack of trust between humans and machines, but also easily induces traffic safety accidents due to insufficient decision-making adaptability. Summary of the Invention

[0004] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method, system, and device for autonomous driving safety decision-making based on neural memory.

[0005] According to one aspect of this application, an autonomous driving safety decision-making method based on neural memory is provided, comprising: constructing an anthropomorphic neuron vehicle-mounted model based on the characteristics of biological neurons; wherein the anthropomorphic neuron vehicle-mounted model includes a neuron-synaptic structure, a multi-node neuron topology, and a neuron network approximation solver; constructing a memory knowledge base based on traffic environment data; wherein the memory knowledge base includes traffic situation stimuli, neural-synaptic reactors, and stimulus result behaviors; training the anthropomorphic neuron vehicle-mounted model based on the memory knowledge base; calculating the stability index of the trained anthropomorphic neuron vehicle-mounted model; wherein the stability index includes a robustness index, a stability index, and a consistency index; if the stability index meets preset requirements, calculating the decision difference between the anthropomorphic neuron vehicle-mounted model and a test driver based on a complex traffic scenario; performing consistency enhancement training on the anthropomorphic neuron vehicle-mounted model based on the decision difference; and executing autonomous driving safety decisions based on the consistency enhancement training of the anthropomorphic neuron vehicle-mounted model.

[0006] In one embodiment, constructing an anthropomorphic neuron vehicle model based on the characteristics of biological neurons includes: constructing the neuron-synapse structure based on the characteristics of biological neurons; wherein, the neuron-synapse structure includes the equivalent capacitance of each neuron, the synaptic weights between neurons, the output pulse sequence and output current of each neuron; constructing a multi-node neuron topology based on the neuron-synapse structure; and constructing an approximate solver for the neuron network based on the multi-node neuron topology and the vehicle dynamics equations.

[0007] In one embodiment, constructing a memory knowledge base based on traffic environment data includes: establishing a traffic situation stimulus source based on traffic environment data; wherein the traffic situation stimulus source includes a standardized traffic situation vector signal; determining a neural-synaptic reactor based on the traffic situation stimulus source; wherein the neural-synaptic reactor includes a standardized neural synaptic identifiable pulse signal sequence; generating the stimulus result behavior based on the traffic situation stimulus source and the neural-synaptic reactor; and constructing a stimulus-behavior bidirectional mapping relationship based on the traffic situation stimulus source and the stimulus result behavior.

[0008] In one embodiment, training the anthropomorphic neuron vehicle-mounted model based on the memory knowledge base includes: determining a training sample set for the anthropomorphic neuron vehicle-mounted model based on the memory knowledge base; and training the anthropomorphic neuron vehicle-mounted model based on the training sample set.

[0009] In one embodiment, the stability index of the anthropomorphic neuron vehicle-mounted model after computational training includes: determining the robustness index based on the degree of influence of external disturbances on the state of the anthropomorphic neuron vehicle-mounted model; calculating the Lyapunov energy function based on the neuron state fluctuation deviation, driving decision output deviation, and scene adaptation deviation of the anthropomorphic neuron vehicle-mounted model; determining the stability index based on the Lyapunov energy function; and calculating the deviation value between the actual output and the benchmark output of the anthropomorphic neuron vehicle-mounted model to obtain the consistency index.

[0010] In one embodiment, calculating the decision difference between the anthropomorphic neuron vehicle-mounted model and the test driver based on a complex traffic scenario includes: generating the output decision of the anthropomorphic neuron vehicle-mounted model based on the complex traffic scenario and the memory knowledge base; and calculating the decision difference between the anthropomorphic neuron vehicle-mounted model and the test driver based on the output decision of the anthropomorphic neuron vehicle-mounted model and the output decision of the test driver.

[0011] In one embodiment, the step of performing consistency enhancement training on the anthropomorphic neuron vehicle model based on the decision difference includes: performing consistency enhancement training on the anthropomorphic neuron vehicle model with the goal of minimizing the decision difference.

[0012] In one embodiment, the autonomous driving safety decision-making method based on neural memory further includes: an online adaptive mechanism based on meta-learning to fine-tune the weight parameters of the anthropomorphic neuron vehicle model.

[0013] According to another aspect of this application, an autonomous driving safety decision-making system based on neural memory is provided, comprising: an on-board model construction module for constructing an anthropomorphic neuron on-board model based on the characteristics of biological neurons; wherein the anthropomorphic neuron on-board model includes a neuron-synaptic structure, a multi-node neuron topology, and a neuron network approximation solver; a knowledge base construction module for constructing a memory knowledge base based on traffic environment data; wherein the memory knowledge base includes traffic situation stimuli, neural-synaptic reactors, and stimulus result behaviors; and an on-board model training module for training the anthropomorphic neuron on-board model based on the memory knowledge base. The system includes: a training module; a stability index calculation module for calculating the stability index of the trained anthropomorphic neuron vehicle model, wherein the stability index includes a robustness index, a stability index, and a consistency index; a decision difference calculation module for calculating the decision difference between the anthropomorphic neuron vehicle model and the test driver based on a complex traffic scenario if the stability index meets preset requirements; a human-machine enhancement training module for performing consistency enhancement training on the anthropomorphic neuron vehicle model based on the decision difference; and an autonomous driving execution module for executing autonomous driving safety decisions based on the consistency enhancement training of the anthropomorphic neuron vehicle model.

[0014] According to another aspect of this application, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; the processor being configured to perform any of the methods described above.

[0015] This application provides a method, system, and device for autonomous driving safety decision-making based on neural memory. Based on the characteristics of biological neurons, a humanoid neuron-based vehicle model is constructed. This humanoid neuron-based vehicle model includes a neuron-synaptic structure, a multi-node neuron topology, and a neural network approximation solver. A memory knowledge base is constructed based on traffic environment data. This memory knowledge base includes traffic situation stimuli, neural-synaptic reactors, and stimulus-result behaviors. The humanoid neuron-based vehicle model is trained based on the memory knowledge base. The stability indices of the trained humanoid neuron-based vehicle model are calculated. These stability indices include robustness, stability, and consistency. If the stability indices meet preset requirements, the model is then used in complex traffic scenarios. The system calculates the decision differences between a humanoid neuron-based in-vehicle model and a test driver; based on these differences, it performs consistency enhancement training on the humanoid neuron-based in-vehicle model; the humanoid neuron-based in-vehicle model trained on consistency enhancement executes autonomous driving safety decisions; by constructing a humanoid neuron-based in-vehicle model and a neural memory knowledge base, and based on neural-synaptic mechanisms and multi-node neuron topology, it achieves brain-like encoding and transmission of in-vehicle perception and decision-making signals, simulating human brain memory and cognitive mechanisms, improving the system's humanoid processing capability of traffic information, and solving the problem of traditional autonomous driving lacking experience accumulation and danger intuition; and through verification and logical validation, it achieves interpretability of the decision-making process and online safety verification, meeting the functional safety and compliance requirements of autonomous driving. Attached Figure Description

[0016] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 This is a flowchart illustrating an exemplary embodiment of the autonomous driving safety decision-making method based on neural memory provided in this application.

[0018] Figure 2 This is a schematic diagram of the structure of an autonomous driving safety decision-making system based on neural memory provided in an exemplary embodiment of this application.

[0019] Figure 3 This is a structural diagram of an electronic device provided in an exemplary embodiment of this application. Detailed Implementation

[0020] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0021] Figure 1 This is a flowchart illustrating an exemplary embodiment of the autonomous driving safety decision-making method based on neural memory provided in this application. Figure 1 As shown, this autonomous driving safety decision-making method based on neural memory includes: Step 110: Based on the characteristics of biological neurons, construct an anthropomorphic neuron vehicle model.

[0022] The anthropomorphic neuron vehicle-mounted model includes a neuron-synapse structure, a multi-node neuron topology, and a neuron network approximation solver.

[0023] Step 120: Construct a memory knowledge base based on traffic environment data.

[0024] The memory knowledge base includes traffic situation stimuli, neural-synaptic reactors, and stimulus-result behaviors.

[0025] Step 130: Train the anthropomorphic neuron vehicle model based on the memory knowledge base.

[0026] Step 140: Calculate the stability index of the trained anthropomorphic neuron vehicle model.

[0027] Among them, stability indicators include robustness indicators, stability indicators, and consistency indicators.

[0028] Step 150: If the stability index meets the preset requirements, calculate the decision difference between the anthropomorphic neuron vehicle model and the test driver based on the complex traffic scenario.

[0029] Step 160: Based on decision differences, perform consistency enhancement training on the anthropomorphic neuron vehicle model.

[0030] Step 170: The anthropomorphic neural vehicle model based on consistency enhancement training executes autonomous driving safety decisions.

[0031] This application provides an autonomous driving safety decision-making method based on neural memory. Based on the characteristics of biological neurons, a humanoid neuron-based vehicle model is constructed. This model includes a neuron-synaptic structure, a multi-node neuron topology, and a neural network approximation solver. A memory knowledge base is constructed based on traffic environment data, comprising traffic situation stimuli, neuron-synaptic reactors, and stimulus-result behaviors. The humanoid neuron-based vehicle model is trained using this knowledge base. The stability indices of the trained model are calculated, including robustness, stability, and consistency. If the stability indices meet preset requirements, calculations are performed based on complex traffic scenarios. The study investigates the decision-making differences between an anthropomorphic neuron in-vehicle model and a test driver. Based on these differences, it conducts consistency enhancement training on the anthropomorphic neuron in-vehicle model. The model, trained with consistency enhancement, then executes autonomous driving safety decisions. By constructing an anthropomorphic neuron in-vehicle model and a neural memory knowledge base, and based on neural-synaptic mechanisms and multi-node neuron topology, it achieves brain-like encoding and transmission of in-vehicle perception and decision-making signals, simulating human brain memory and cognitive mechanisms. This enhances the system's anthropomorphic processing capability of traffic information, addressing the lack of experience accumulation and danger intuition in traditional autonomous driving. Furthermore, through verification and logical validation, it achieves interpretability of the decision-making process and online safety verification, meeting the functional safety and compliance requirements of autonomous driving.

[0032] In one embodiment, the specific implementation of step 110 above may be as follows: based on the characteristics of biological neurons, construct a neuron-synapse structure; wherein, the neuron-synapse structure includes the equivalent capacitance of each neuron, the synaptic weight between neurons, the output pulse sequence and output current of each neuron; based on the neuron-synapse structure, construct a multi-node neuron topology; based on the multi-node neuron topology and vehicle dynamics equations, construct a neuron network approximate solver.

[0033] First, this application mathematically models the electrophysiological characteristics of biological neurons, establishing a neuronal model that includes membrane potential dynamics, ion channel gating, and synaptic transmission. By extracting common neuronal dynamic laws in humans, it focuses on modeling neurons in the prefrontal cortex that control human rational thinking and logical reasoning, while also covering neurons in the motor cortex, visual cortex, superior colliculus, and brainstem that control limb movement, visual processing, head movement, and eye movement. It models the plasticity of synaptic connections between neurons, establishing a dynamic update model for synaptic weights and designing learning rules that simulate the plasticity of biological synapses. This simulates the human brain's experience accumulation and adaptive correction capabilities, enabling the connection weights between neurons to adaptively adjust according to input and output activity patterns, achieving iterative learning of behavioral experience. Furthermore, it integrates multimodal sensory information from vision, hearing, and limb sensation into the neuronal model, establishing a sensory-motor integration loop to form a unified neural integration mechanism for motor commands. A multimodal pulse fusion model is used, transforming information from different modalities into pulse sequences through independent coding layers, converging into an integrated neuronal group, and then forming motor commands through synaptic weight learning, thus forming a neuron-synapse structure.

[0034] Then, topological modeling of the collaborative relationships between neurons is performed to establish a functionally specific neuronal population network. Following a hierarchical architecture of perception, memory, and decision layers, a multi-node neuron connection adjacency matrix is ​​defined to complete the anthropomorphic neural network topology design. A modular small-world network topology is adopted, with feedforward connections between layers and lateral inhibition / excitation connections within layers. Cross-layer shortcut connections are introduced to accelerate information propagation. A propagation model of electrical signals in the neuronal network is established to describe the propagation law of electrical signals in the neuronal population. Considering the transmission delay and attenuation characteristics of synaptic signals, a neural signal temporal attenuation transfer function is constructed to quantify the information transmission loss law between nodes. The signal propagation model of electrical signals in the neuronal network is represented in the form of pulse convolution. The aforementioned topology and signal propagation model in the neuronal network are instantiated on a computationally limited vehicle platform to construct a runnable spiking neural network. The network size and connection density are constrained and optimized to ensure that the model can run efficiently on the embedded platform, forming a multi-node neuronal topology structure that can run in real-time on a vehicle.

[0035] Finally, addressing the uncertainties of vehicle operation caused by road disturbances, sensor noise, and time-varying parameters, this paper abandons the traditional fixed-weight neural network modeling approach and introduces a variational Bayesian deep learning algorithm to model network parameters as random variables. By constructing a dynamic prior distribution and updating the posterior distribution using real-time driving data, network optimization is achieved through divergence constraints, accurately fitting the nonlinear dynamic mapping relationship of the vehicle. This allows for quantifying prediction uncertainties and outputting confidence intervals, effectively improving the robustness and generalization ability of modeling complex operating conditions. Combining the discrete-time dynamic characteristics of the vehicle, a stochastic differential equation-based benchmark model of vehicle dynamics is constructed to accurately characterize the state drift and diffusion characteristics caused by time-varying disturbances. To solve the problem that the true posterior distribution is difficult to solve directly, a variational approximation distribution is introduced to fit the true posterior, minimizing... KL The divergence constraint distribution deviation is used to construct the network optimization objective function as shown in the following equation: ; In the formula, For the learnable hyperparameters of the variational approximation distribution, for KL divergence term, For the weight random variables of the Bayesian neural network, For the drift function network parameters, For the diffusion function parameters, For variational approximation of the posterior distribution, For the initial prior distribution of the parameters, It represents the mathematical expectation of an approximate distribution.

[0036] This paper simplifies large-scale neural networks by using model reduction methods to reduce computational complexity and meet the requirements of vehicle computing power and real-time performance. Addressing the issues of high dimensionality, large number of parameters, and high inference latency in Bayesian dynamics networks, a Krylov subspace orthogonal projection model reduction algorithm is employed. This algorithm compresses the dimensionality of the high-dimensional neuron state space by constructing a Krylov subspace sequence and orthogonalizing it to generate a projection matrix. This maps high-dimensional dynamic parameters and states to a low-dimensional subspace, eliminating redundant parameters and invalid neurons. While fully preserving dynamic characteristics and uncertainty quantification capabilities, the model is lightweight and adaptable to low-computing-power vehicle operation scenarios. Specifically, based on the original high-dimensional Bayesian network state space, a k-order Krylov subspace sequence is constructed. By iteratively sampling the coupling features of vehicle dynamic states and control inputs, a high-dimensional feature vector set is generated. The projection matrix maps the high-dimensional state space to the low-dimensional subspace, completing the order reduction update of the network parameter-dynamic mapping relationship and constructing a low-dimensional lightweight approximate vehicle dynamics model.

[0037] By designing a parallel computing architecture, the real-time performance of the neural network solver is improved. Addressing the issues of fluctuating computing power, high network solution latency, and uneven resource allocation on heterogeneous vehicle platforms, a combined intra-layer and inter-layer parallel solution algorithm is adopted, coupled with dynamic weight adaptive aggregation and latency-tolerant scheduling strategies. Network subtasks are split and distributed to multiple computing cores for parallel computation. Weights are dynamically adjusted based on core load and solution accuracy. The aggregation operator is improved to fuse sub-network outputs, and a latency correction term is introduced to compensate for timing deviations, achieving low-latency, high-fault-tolerant real-time dynamics solution for vehicles. Based on the heterogeneous vehicle computing platform, a strategy combining intra-layer and inter-layer parallelism is employed. The neural network is divided into several sub-networks and distributed to different computing cores for parallel solution. Specifically, the low-dimensional dynamics neural network is hierarchically split, dividing the overall network into multiple structurally independent and functionally complementary parallel sub-networks. Based on the vehicle core load rate and solution accuracy priority, the computational resources of each sub-network are dynamically allocated. The output features of multiple sub-networks are fused, and a latency correction compensation model is constructed to eliminate timing lag errors, ultimately building a complete neural network approximate solver.

[0038] In one embodiment, step 120 can be implemented as follows: based on traffic environment data, establish traffic situation stimulus sources; wherein, the traffic situation stimulus sources include standardized traffic situation vector signals; based on the traffic situation stimulus sources, determine neural-synaptic reactors; wherein, the neural-synaptic reactors include standardized neural synaptic identifiable pulse signal sequences; based on the traffic situation stimulus sources and neural-synaptic reactors, generate stimulus-result behaviors; based on the traffic situation stimulus sources and stimulus-result behaviors, construct a bidirectional stimulus-behavior mapping relationship.

[0039] This application integrates multi-source traffic data from vehicle cameras and radar to abstract perceived data into a structured scene description. It constructs a comprehensive digital data set of traffic scenes encompassing vehicles, pedestrians, roads, and weather, covering various conditions including highways, urban areas, and intersections. This generates traffic situation data and the dynamic coupling behaviors between vehicles within the situation, establishing a traffic situation structure data that can be processed by a memory system. By establishing spatiotemporal relationships between stimulus sources, a graph representation of the traffic situation is constructed. Based on spatiotemporal distance constraints, the correlation and coupling relationships between traffic participants are extracted, and a spatiotemporal topological correlation matrix of the traffic situation is constructed to characterize the dynamic coupling features of the scene, resulting in a dynamic graph sequence of a neural network. The structured scene description is transformed into vector signals that can be used for training and memory retrieval. A graph neural network is used to encode the spatiotemporal correlation graph and integrate it with the scene model to construct a standardized traffic situation signal.

[0040] By employing a key node extraction method, we identify neuron clusters that have a decisive impact on driving decisions. Through feature contribution calculation, we screen the core neural nodes that play a dominant role in risk assessment, driving decisions, and vehicle handling stability control from the full dataset of these clusters, eliminating redundant and ineffective nodes to obtain a set of key neural nodes. A variable-step adaptive temporal sampling method is designed to discretize and truncate standardized traffic situation vector signals. Small-step sampling is used in high-frequency hazardous scenarios to preserve detailed features, while large-step sampling is used in normal operating conditions to reduce computational load, balancing scene detail with onboard computing efficiency, thus achieving lightweight preprocessing of neural input signals. By integrating key node selection, adaptive sampling, and feature normalization modules, a complete neural-synaptic reactor is constructed to clean, transform, and encode traffic situation signals, achieving standardized transformation of external environmental stimuli into recognizable synaptic pulse signals.

[0041] This study employs a temporal semantic coding network to deeply analyze traffic situation feature vector signals. Combined with a prior knowledge base of human driving, it mines potential risks and driving intentions within scenarios, enabling quantitative inference of driver-like behavioral intentions. The coding network captures the spatiotemporal interaction patterns of traffic participants through multi-layer temporal convolution and attention mechanisms, and corrects the inference results of scenario risk levels and driving intentions based on prior knowledge. Based on driving behavior intention feature vectors, a spatiotemporal correlation graph of stimulus sources is integrated to construct an intention-operation mapping matrix. Simultaneously, vehicle dynamics constraints and boundary limitations of standard vehicle handling and stability procedures are introduced to constrain the rationality of behavior generation and vehicle driving stability, ultimately outputting a continuous and executable sequence of human-like driving behaviors. A mapping relationship between behavior and stimuli is established, forming a complete stimulus-response closed loop to quantify the correspondence between different traffic stimulus intensities, scenario hazard levels, and driving operation amplitudes. This results in storable and searchable association rule samples, which can be used to deduce behavior from stimuli or vice versa. The mapping relationship is stored as a stimulus-behavior bidirectional mapping neural network model.

[0042] This study quantifies sample similarity across four dimensions: scene spatiotemporal features, risk level, driving behavior, and stimulus intensity. It combines density clustering algorithms to group and classify massive historical traffic driving samples, aggregating memory samples with similar features, shared scenes, and similar behaviors into unified memory units. This achieves structured and hierarchical organization and management of disordered samples, and the clustering tree is periodically updated with newly added memory samples to improve memory utilization efficiency. The four-dimensional driving sample similarity measurement formula is described below. The specific expression is shown in the following formula: ; in, For any two groups of driving samples, These are respectively the spatiotemporal characteristics of the scene, risk level, driving behavior, and stimulus intensity. These are the weights corresponding to the four dimensions.

[0043] Based on clustered structured memory units, multi-dimensional memory association triggering rules are constructed. Combining the characteristics of neural memory decay, scene temporal association logic, and causal relationships of driving behavior, association rules for similar scene matching, similar risk association, and homologous behavior inference are defined to achieve rapid association and matching between the current scene and historical memory. When similar traffic situations are detected, corresponding historical driving experience is automatically matched. Simulating the biological forgetting law, an exponential decay model is constructed. As shown in the following formula: ; in, The inherent decay coefficient of memory, To access the sparsity coefficient, For memory in Frequency of invocation at any given moment The storage duration for historical memory samples.

[0044] Establish risk attention gain function In high-risk scenarios, the association weight is automatically increased, as shown in the following formula: ; in, Risk attention value, The threshold for triggering risk attention.

[0045] Establish time-series dependency enhancements The enhanced temporal correlation for long-term continuous scenarios is shown in the following equation: ; in, For timing enhancement coefficients, For time series duration, This is the timing decay time constant.

[0046] Combining four-dimensional driving sample similarity measurement, exponential decay model, risk attention gain function, and temporal dependency enhancement term, a multi-dimensional memory association triggering rule is constructed as shown in the following formula: ; in, For memory association matching degree, This is the memory decay coefficient.

[0047] By integrating clustering organization with multi-dimensional memory association triggering rules, a dynamically updatable memory association knowledge base is constructed based on the memory association matching degree. A dynamic incremental update mechanism is designed to collect new traffic scenarios and driving behavior samples during vehicle operation in real time, complete the clustering and storage of new memories, and iterative update of rules, so as to realize the continuous iterative optimization of the knowledge base, allowing the vehicle model to continuously accumulate driving experience and continuously improve human-like decision-making and scenario adaptation capabilities.

[0048] In one embodiment, step 130 can be implemented as follows: determining the training sample set of the anthropomorphic neuron vehicle model based on the memory knowledge base; and training the anthropomorphic neuron vehicle model based on the training sample set.

[0049] This application integrates a memory association knowledge base, standardized traffic situation vector signals, and human-like driving behavior sequences. It performs integrated knowledge-data encoding on prior knowledge, scene vector signal data, and driving behavior sequences, encoding them into training labels and samples suitable for reinforcement learning, generating a standardized sample set required for a driving awareness neural network. The standardized sample set undergoes temporal alignment, outlier removal, data augmentation, and normalization to address the issue of asynchronous timestamps from multiple data sources. Samples are then categorized and batched according to driving scenarios, into training, validation, and test sets. A deep human-like driving awareness network integrating memory association, situational awareness, and awareness deduction modules is constructed. Based on a preprocessed dataset, a multi-objective reinforcement learning algorithm is used for model training. Human driving behavior bias, driving safety, and passenger comfort are used as a joint loss function (weighted summation). The network weights are iteratively optimized to fit the logic of human driving thought processes. Using a dynamic sequence of neural network graphs as input, temporal convolution and spatial attention are used to extract the interaction features of traffic participants. The network is first pre-trained with human driving data, and then the reward signals provided by the memory knowledge base are used for policy optimization.

[0050] Based on the Hidden Markov Model (HMM) temporal recursion algorithm, dynamic state updates of multi-level neural networks are achieved, simulating the physiological mechanism of human brain neurons responding to external scene stimuli in real time. Initialization of the initial membrane potential, synaptic weights, and hidden states of the multi-level neural network is completed. A temporal state iteration equation is established based on neural signal transmission rules. Combined with real-time neural pulse input, the operating states of all neurons in the perception layer, memory layer, and decision layer are updated, realizing the dynamic evolution of driving consciousness with traffic scenes. Deep convolutional neural networks are used to extract multi-scale spatial features. A distributed parallel training algorithm is combined to achieve multi-node neuron collaborative computation. Based on the temporally updated multi-node neuron state sequence, three scales are divided: microscopic local features, mesoscopic traffic flow features, and macroscopic scene features. Multi-layer convolutional kernels of the deep convolutional neural network are used to extract features at different scales. A distributed parallel training architecture is introduced, distributing the multi-node neuron feature extraction task to multiple parallel computing units to simultaneously complete the extraction and preprocessing of features at each level. Based on the topological relationships of neuronal synapses, a multi-node feature fusion weight matrix is ​​constructed. Weighted fusion, splicing, and redundancy removal are performed on the three scales of features to complete the collaborative activation of multi-node neurons, forming a fused driving consciousness feature representation. This invention integrates steady-state visual evoked potential brain-computer interface technology with knowledge model data-driven algorithms, combined with a spatiotemporal attention mechanism, to achieve precise decoding and mapping of abstract driving consciousness features to concrete vehicle driving behavior. First, a spatiotemporal attention mechanism is introduced to redistribute the weights of the fused multidimensional consciousness features, automatically focusing on high-risk, highly correlated core traffic feature regions, weakening invalid background interference, and strengthening the representational ability of key scenario decision features. Then, based on the steady-state visual evoked potential brain-computer interface decoding principle, it simulates the human visual neural response mechanism, transforming the machine consciousness features extracted by the neural network into human-like driving thought signals, achieving feature alignment between machine consciousness and human driving thought. Simultaneously, a knowledge model data-driven dynamic calibration mechanism is built, retrieving standard driving behavior paradigms from the memory association knowledge base to perform deviation calibration, compliance correction, and comfort optimization on the decoded preliminary decision signals, eliminating model decoding errors. Finally, the abstract driving consciousness thought signals are precisely mapped into standardized vehicle control commands, outputting behavioral sequences that conform to human driving habits.

[0051] Based on the states of multi-node neurons, a Lyapunov function is constructed to measure the stability of the driving awareness system. Based on the principle of energy conservation and steady state, a Lyapunov energy function is constructed to characterize the overall operating energy state of the driving awareness system. The system's steady-state trend is determined by the convergence characteristics of the function. When the function continuously converges and tends to a stable value, it proves that the driving awareness system is in a stable state; if the function oscillates and diverges, it is determined that the awareness system has abnormal fluctuations.

[0052] In one embodiment, step 140 can be implemented as follows: a robustness index is determined based on the degree of influence of external disturbances on the state of the anthropomorphic neuron vehicle model; a Lyapunov energy function is calculated based on the neuron state fluctuation deviation, driving decision output deviation, and scene adaptation deviation of the anthropomorphic neuron vehicle model; a stability index is determined based on the Lyapunov energy function; and a consistency index is obtained by calculating the deviation between the actual output and the baseline output of the anthropomorphic neuron vehicle model.

[0053] This study analyzes the impact of external disturbances on driving awareness, calculates awareness stability indices, and evaluates the system's robustness. By substituting disturbance parameters into the Lyapunov stability function, the convergence rate, oscillation amplitude, and error convergence residual under disturbance conditions are calculated. A quantitative evaluation system for driving awareness stability is constructed, and a normalized driving awareness stability score is calculated using three core indicators: convergence speed, fluctuation amplitude, and disturbance resistance threshold. The current driving awareness-generated behavioral decisions are compared with a baseline safe driving mode to verify the consistency and rationality of the decisions. The driving decision behavior sequence, awareness deduction logic, and vehicle control parameters output by the model in real time are compared with the baseline driving mode frame-by-frame and dimension-by-dimensional to quantify behavioral deviation, decision consistency, and scenario adaptation matching. A deviation tolerance threshold is set. When the overall deviation is within the threshold range, the driving awareness state is considered stable and the decisions are compliant; when the deviation exceeds the threshold, the awareness state is considered to be fluctuating, and the deviation dimension, deviation amplitude, and corresponding correction scheme are output simultaneously.

[0054] In one embodiment, step 150 can be implemented as follows: generating the output decision of a humanoid neuron vehicle-mounted model based on a complex traffic scenario and a memory knowledge base; and calculating the decision difference between the humanoid neuron vehicle-mounted model and the test driver based on the output decision of the humanoid neuron vehicle-mounted model and the output decision of the test driver.

[0055] A high-fidelity traffic scene generator is constructed based on real-world road data and generative adversarial networks. It automatically synthesizes diverse test environments covering extreme conditions, edge cases, and common driving scenarios. Standardized traffic situation vectors are retrieved, and the scene sample memory is expanded. Human driver manipulation data, vehicle status, and environmental perception results are collected in the same scenarios. The collected data is then timestamped, cleaned, and stored in a human-machine joint database. The differences between human driver and autonomous driving system decisions are analyzed, and the confidence level of the system's decisions is assessed. Behavioral bias and consciousness stability are incorporated as correction terms in the confidence level calculation. Key scenario features of decision-making discrepancies are identified. Based on the collected human-machine paired dataset, a multi-dimensional quantitative evaluation index for decision-making differences is constructed to quantitatively analyze the differences between machine and human driving decisions. The temporal similarity of human-machine decisions is calculated using the memory matching degree, and the temporal confidence of machine decisions is also calculated to pinpoint the core scenarios and core feature dimensions of human-machine decision-making bias.

[0056] In one embodiment, step 160 can be implemented by performing consistency enhancement training on the anthropomorphic neuron vehicle model with the goal of minimizing decision differences.

[0057] To minimize the discrepancy between human and machine decision-making, this study optimizes autonomous driving decision-making strategies and improves human-machine consistency through a combination of adversarial training and imitation learning. A human-machine consistency enhancement loss function is constructed, employing a hybrid optimization strategy of transfer learning and adversarial training to constrain machine decisions towards high-quality human driving decisions. Using human driving decisions as supervised ground truth, a confidence-weighted mechanism is used to enhance fitting of high-confidence samples and adaptively correct low-confidence samples. Adversarial perturbation training is also introduced to improve the model's decision consistency and robustness in differentiated scenarios. Consistency Enhancement Loss Function As shown in the following formula: ; in, For the sample size, For machine decision-making timing confidence. The overall bias error in human-machine decision-making. To counteract training losses, To counteract the loss weighting coefficient.

[0058] This application can also construct a dual-branch parallel safety decision-making network, including a human-like decision-making branch and a safety constraint correction branch, to achieve deep integration of human-like decision-making and safety constraint boundaries. The human-like decision-making branch outputs basic decisions that conform to human driving habits; the safety constraint correction branch uses safety boundary conditions as prior constraints, identifies the safety risks of the current decision, and outputs dynamic correction amounts. A safety constraint loss function needs to be designed, incorporating boundary exceedance values, risk prediction values, and decision compliance values ​​into a joint optimization objective, forcing the network decision to always remain within a safe range while preserving human-like driving characteristics to the greatest extent. Furthermore, by establishing an online verification mechanism for safety decisions, the network output is checked for safety in real time. When potential dangers are detected, a multi-level failure protection strategy is triggered to ensure that the system can complete real-time verification of decision compliance, safety, and stability in normal scenarios. In high-risk extreme scenarios, failure protection logic is automatically triggered to block abnormal human-like decisions and force a switch to failure protection mode, eliminating safety hazards caused by decision failures.

[0059] This application can also employ gradient-weighted activation mapping and layer-by-layer correlation propagation algorithms to perform refined attribution and source tracing of network decision outputs, locate the core features, key temporal nodes, and network activation units influencing the current driving decision, quantify the contribution weight of each input feature to the final decision, and ultimately generate a global feature attribution heatmap and temporal contribution curve. Understandable decision rules are extracted from the trained neural network and transformed into human-readable semantic expressions to enhance the transparency of human-computer interaction. Through association rule mining and semantic mapping algorithms, based on the positive mapping function from traffic situation feature vector signals to human-like driving behavior sequences, the memory stimulus-behavior mapping relationship and consciousness output are combined with the implicit network numerical decision logic to transform into explicit, human-understandable natural language driving rules. After redundancy removal and conflict resolution, the rules generate natural language text. A quantitative evaluation index for decision interpretability is established, and human drivers' trust in the system is enhanced through a human-computer interaction interface. A comprehensive interpretability rating is completed by combining subjective ratings from human drivers with objective machine indicators, and a dynamic evolution parameter for human-computer trust based on differential equations is established.

[0060] This application can also evaluate the coverage of the training scenario library to real-world traffic scenarios, identify typical scenarios that are not fully covered, and guide the expansion and optimization of the scenario library. By constructing a multi-dimensional feature space for traffic scenarios, it clusters the scenarios in the training set, validation set, and test set, quantitatively evaluating the scenario coverage, scenario redundancy, representativeness of typical scenarios, and the blankness of unknown scenarios in the existing dataset. Through a high-dimensional scenario feature clustering algorithm, it selects core typical working conditions, rare extreme working conditions, and blank unknown working conditions to accurately locate the scenarios where the model has generalization shortcomings. By constructing a multi-dimensional decision generalization ability evaluation system, including three core indicators—"scenario generalization degree, working condition robustness, and decision stability"—it quantifies the difference in decision performance of the model under known and unknown scenarios.

[0061] In one embodiment, the above-mentioned autonomous driving safety decision-making method based on neural memory may further include: an online adaptive mechanism based on meta-learning to fine-tune the weight parameters of the anthropomorphic neuron vehicle model.

[0062] This application can also design an online adaptive mechanism based on meta-learning, enabling the system to quickly adapt to new scenarios and new driver styles, continuously improve its generalization ability, and mine general driving decision-making prior knowledge through a meta-learning two-layer optimization mechanism to quickly adapt to new working conditions; at the same time, an online incremental learning module is built to collect unknown scenario data in real time, complete rapid iterative updates with small samples, and continuously improve the model's generalization ability.

[0063] Figure 2 This is a schematic diagram of the structure of an autonomous driving safety decision-making system based on neural memory, provided in an exemplary embodiment of this application. Figure 2 As shown, the autonomous driving safety decision-making system 20 based on neural memory includes: an on-board model construction module 21, used to construct an anthropomorphic neuron on-board model based on the characteristics of biological neurons; wherein, the anthropomorphic neuron on-board model includes neuron-synaptic structure, multi-node neuron topology structure and neuron network approximation solver; a knowledge base construction module 22, used to construct a memory knowledge base based on traffic environment data; wherein, the memory knowledge base includes traffic situation stimuli, neural-synaptic reactors and stimulus result behaviors; and an on-board model training module 23, used to train the anthropomorphic neuron on-board model based on the memory knowledge base; and a stable... The stability index calculation module 24 is used to calculate the stability index of the trained anthropomorphic neuron vehicle model; the stability index includes robustness index, stability index and consistency index; the decision difference calculation module 25 is used to calculate the decision difference between the anthropomorphic neuron vehicle model and the test driver based on a complex traffic scenario if the stability index meets the preset requirements; the human-machine enhancement training module 26 is used to perform consistency enhancement training on the anthropomorphic neuron vehicle model based on the decision difference; the autonomous driving execution module 27 is used to execute autonomous driving safety decisions based on the anthropomorphic neuron vehicle model trained with consistency enhancement.

[0064] This application provides an autonomous driving safety decision-making system based on neural memory. The vehicle-mounted model construction module 21 constructs a humanoid neuron vehicle-mounted model based on the characteristics of biological neurons; wherein the humanoid neuron vehicle-mounted model includes a neuron-synaptic structure, a multi-node neuron topology, and a neuron network approximation solver; the knowledge base construction module 22 constructs a memory knowledge base based on traffic environment data; wherein the memory knowledge base includes traffic situation stimuli, neural-synaptic reactors, and stimulus-result behaviors; the vehicle-mounted model training module 23 trains the humanoid neuron vehicle-mounted model based on the memory knowledge base; and the stability index calculation module 24 calculates the stability index of the trained humanoid neuron vehicle-mounted model; wherein the stability index includes a robustness index, a stability index, and a consistency index; if the stability index meets preset requirements, decision difference calculation is performed. Module 25 calculates the decision differences between the anthropomorphic neuron vehicle model and the test driver based on complex traffic scenarios; the human-machine enhancement training module 26 performs consistency enhancement training on the anthropomorphic neuron vehicle model based on the decision differences; the autonomous driving execution module 27 executes autonomous driving safety decisions based on the anthropomorphic neuron vehicle model trained with consistency enhancement; by constructing anthropomorphic neuron vehicle models and neural memory knowledge bases, and based on neural-synaptic mechanisms and multi-node neuron topologies, the system realizes brain-like encoding and transmission of vehicle perception and decision signals, simulates human brain memory and cognitive mechanisms, improves the system's anthropomorphic processing capability of traffic information, and solves the problem of traditional autonomous driving lacking experience accumulation and danger intuition; and through verification and logical validation, the system achieves interpretability of the decision-making process and online safety verification, meeting the functional safety and compliance requirements of autonomous driving.

[0065] In one embodiment, the vehicle-mounted model construction module 21 can be further configured to: construct a neuron-synapse structure based on the characteristics of biological neurons; wherein, the neuron-synapse structure includes the equivalent capacitance of each neuron, the synaptic weights between neurons, the output pulse sequence and output current of each neuron; construct a multi-node neuron topology based on the neuron-synapse structure; and construct a neuron network approximate solver based on the multi-node neuron topology and the vehicle dynamics equations.

[0066] In one embodiment, the knowledge base construction module 22 can be further configured to: establish traffic situation stimulus sources based on traffic environment data; wherein the traffic situation stimulus sources include standardized traffic situation vector signals; determine neural-synaptic reactors based on the traffic situation stimulus sources; wherein the neural-synaptic reactors include standardized neural synaptic identifiable pulse signal sequences; generate stimulus result behaviors based on the traffic situation stimulus sources and neural-synaptic reactors; and construct a stimulus-behavior bidirectional mapping relationship based on the traffic situation stimulus sources and stimulus result behaviors.

[0067] In one embodiment, the above-mentioned vehicle-mounted model training module 23 can be further configured to: determine the training sample set of the anthropomorphic neuron vehicle-mounted model based on the memory knowledge base; and train the anthropomorphic neuron vehicle-mounted model based on the training sample set.

[0068] In one embodiment, the stability index calculation module 24 can be further configured to: determine the robustness index based on the degree of influence of external disturbances on the state of the anthropomorphic neuron vehicle model; calculate the Lyapunov energy function based on the neuron state fluctuation deviation, driving decision output deviation, and scene adaptation deviation of the anthropomorphic neuron vehicle model; determine the stability index based on the Lyapunov energy function; and calculate the deviation value between the actual output and the benchmark output of the anthropomorphic neuron vehicle model to obtain the consistency index.

[0069] In one embodiment, the decision difference calculation module 25 can be further configured to: generate the output decision of the anthropomorphic neuron vehicle model based on the complex traffic scenario and the memory knowledge base; and calculate the decision difference between the anthropomorphic neuron vehicle model and the test driver based on the output decision of the anthropomorphic neuron vehicle model and the output decision of the test driver.

[0070] In one embodiment, the human-machine enhancement training module 26 can be further configured to perform consistency enhancement training on the anthropomorphic neuron vehicle model with the goal of minimizing decision differences.

[0071] In one embodiment, the aforementioned autonomous driving safety decision-making system 20 based on neural memory can be further configured to: fine-tune the weight parameters of the anthropomorphic neuron vehicle model based on an online adaptive mechanism of meta-learning.

[0072] Below, for reference Figure 3 This application describes an electronic device according to embodiments thereof. The electronic device may be either or both of a first device and a second device, or a standalone device independent of them, which may communicate with the first device and the second device to receive acquired input signals from them.

[0073] Figure 3 A block diagram of an electronic device according to an embodiment of this application is illustrated.

[0074] like Figure 3 As shown, the electronic device 10 includes one or more processors 11 and memory 12.

[0075] The processor 11 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 10 to perform desired functions.

[0076] The memory 12 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 11 may execute the program instructions to implement the methods of the various embodiments of this application described above and / or other desired functions. Various contents such as input signals, signal components, and noise components may also be stored in the computer-readable storage medium.

[0077] In one example, the electronic device 10 may also include an input device 13 and an output device 14, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0078] When the electronic device is a standalone device, the input device 13 can be a communication network connector for receiving the collected input signals from the first device and the second device.

[0079] In addition, the input device 13 may also include, for example, a keyboard, a mouse, etc.

[0080] The output device 14 can output various information to the outside, including determined distance information, direction information, etc. The output device 14 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0081] Of course, for the sake of simplicity, Figure 3 Only some of the components of the electronic device 10 relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device 10 may include any other suitable components depending on the specific application.

[0082] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of this application described in the "Exemplary Methods" section above.

[0083] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0084] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of this application described in the "Exemplary Methods" section above.

[0085] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0086] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0087] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0088] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0089] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0090] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for autonomous driving safety decision-making based on neural memory, characterized in that, include: Based on the characteristics of biological neurons, an anthropomorphic neuron vehicle-mounted model is constructed; wherein, the anthropomorphic neuron vehicle-mounted model includes a neuron-synapse structure, a multi-node neuron topology structure, and a neuron network approximation solver; Based on traffic environment data, a memory knowledge base is constructed; wherein, the memory knowledge base includes traffic situation stimuli, neural-synaptic reactors, and stimulus-result behaviors; The anthropomorphic neuron vehicle model is trained based on the aforementioned memory knowledge base. Calculate the stability index of the trained anthropomorphic neuron vehicle model; wherein the stability index includes robustness index, stability index and consistency index. If the stability index meets the preset requirements, the decision difference between the anthropomorphic neuron vehicle model and the test driver is calculated based on the complex traffic scenario. Based on the aforementioned decision differences, the anthropomorphic neuron vehicle model undergoes consistency enhancement training. The anthropomorphic neural vehicle model, trained based on consistency enhancement, performs autonomous driving safety decisions.

2. The autonomous driving safety decision-making method based on neural memory according to claim 1, characterized in that, The construction of the anthropomorphic neuron vehicle model based on the characteristics of biological neurons includes: Based on the characteristics of the biological neurons, the neuron-synapse structure is constructed; wherein, the neuron-synapse structure includes the equivalent capacitance of each neuron, the synaptic weights between neurons, the output pulse sequence of each neuron, and the output current; Based on the neuron-synapse structure, a multi-node neuron topology is constructed; Based on the multi-node neuron topology and vehicle dynamics equations, an approximate solver for the neuron network is constructed.

3. The autonomous driving safety decision-making method based on neural memory according to claim 1, characterized in that, The construction of the memory knowledge base based on traffic environment data includes: Based on traffic environment data, the traffic situation stimulus source is established; wherein, the traffic situation stimulus source includes a standardized traffic situation vector signal; Based on the traffic situation stimuli, the neural-synaptic reactor is determined; wherein, the neural-synaptic reactor includes a standardized neural synaptic identifiable pulse signal sequence; Based on the traffic situation stimulus source and the neural-synaptic reactor, the stimulus result behavior is generated; Based on the traffic situation stimulus sources and the stimulus-result behaviors, a two-way stimulus-behavior mapping relationship is constructed.

4. The autonomous driving safety decision-making method based on neural memory according to claim 1, characterized in that, The training of the anthropomorphic neuron vehicle model based on the memory knowledge base includes: Based on the memory knowledge base, the training sample set of the anthropomorphic neuron vehicle model is determined; The anthropomorphic neuron vehicle model is trained based on the training sample set.

5. The autonomous driving safety decision-making method based on neural memory according to claim 1, characterized in that, The stability metrics of the anthropomorphic neuron vehicle model after computational training include: The robustness index is determined based on the degree of influence of external disturbances on the state of the anthropomorphic neuron vehicle model. Based on the neuron state fluctuation deviation, driving decision output deviation, and scene adaptation deviation of the anthropomorphic neuron vehicle model, the Lyapunov energy function is calculated. The stability index is determined based on the Lyapunov energy function. The deviation between the actual output and the baseline output of the anthropomorphic neuron vehicle model is calculated to obtain the consistency index.

6. The autonomous driving safety decision-making method based on neural memory according to claim 1, characterized in that, The calculation of the decision differences between the anthropomorphic neural vehicle model and the test driver based on complex traffic scenarios includes: Based on the complex traffic scenario and the memory knowledge base, the output decision of the anthropomorphic neuron vehicle model is generated. Based on the output decisions of the anthropomorphic neuron vehicle model and the output decisions of the test driver, the decision differences between the anthropomorphic neuron vehicle model and the test driver are calculated.

7. The autonomous driving safety decision-making method based on neural memory according to claim 1, characterized in that, The consistency enhancement training of the anthropomorphic neuron vehicle model based on the decision differences includes: With the goal of minimizing the decision difference, the anthropomorphic neuron vehicle model is subjected to consistency enhancement training.

8. The autonomous driving safety decision-making method based on neural memory according to claim 1, characterized in that, The autonomous driving safety decision-making method based on neural memory also includes: The weight parameters of the anthropomorphic neuron vehicle model are fine-tuned based on an online adaptive mechanism of meta-learning.

9. An autonomous driving safety decision-making system based on neural memory, characterized in that, include: The vehicle-mounted model construction module is used to construct an anthropomorphic neuron vehicle-mounted model based on the characteristics of biological neurons; wherein, the anthropomorphic neuron vehicle-mounted model includes a neuron-synapse structure, a multi-node neuron topology structure, and a neuron network approximation solver; The knowledge base construction module is used to build a memory knowledge base based on traffic environment data; wherein, the memory knowledge base includes traffic situation stimuli, neural-synaptic reactors, and stimulus result behaviors; The vehicle-mounted model training module is used to train the anthropomorphic neuron vehicle-mounted model based on the memory knowledge base. The stability index calculation module is used to calculate the stability index of the trained anthropomorphic neuron vehicle model; wherein, the stability index includes robustness index, stability index and consistency index. The decision difference calculation module is used to calculate the decision difference between the anthropomorphic neuron vehicle model and the test driver based on a complex traffic scenario if the stability index meets the preset requirements. The human-machine enhancement training module is used to perform consistency enhancement training on the anthropomorphic neuron vehicle model based on the decision differences. An autonomous driving execution module is used to execute autonomous driving safety decisions based on the anthropomorphic neuron vehicle model trained with consistency enhancement.

10. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is used to execute the method described in any one of claims 1-8.

Images