Intelligent model error localization method for autonomous control systems

Abstract

The application relates to the technical field of artificial intelligence and autonomous decision control, in particular to an intelligent model error positioning method for an autonomous control system. The method registers a forward calculation hook, synchronously collects interaction trajectories and neuron activation tensors, and constructs a heterogeneous fusion dataset. Based on the dataset, structural risk suspiciousness and reward-driven suspiciousness are calculated respectively. Normalization and coverage weighted fusion are performed to generate a comprehensive suspiciousness ranking list, and high-confidence suspicious neurons are intercepted as error positioning results. This method improves the error positioning accuracy, realizes full-form defect capture, enhances the statistical diagnosis robustness, and provides a guarantee for the safe operation of the autonomous control system.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and autonomous decision-making control technology, specifically to an intelligent model error localization method for autonomous control systems. Background Technology

[0002] With the rapid development of artificial intelligence technology, the application of autonomous control systems in transportation, logistics, security, and industrial production has grown exponentially. In these systems, traditional rule-based and finite state machine-based control systems mainly follow the engineering paradigm of "finding patterns—modeling—decision-making." While this paradigm offers strong interpretability and verifiability, it often faces numerous challenges when dealing with highly dynamic, multi-agent interactive complex scenarios, including difficulties in mathematical modeling, insufficient coverage of edge rules, and high costs associated with scenario migration.

[0003] To overcome the limitations of traditional methods, modern autonomous control systems increasingly rely on data-driven deep learning and reinforcement learning techniques. Deep reinforcement learning learns end-to-end policy mappings within the framework of Markov decision processes through continuous interaction between the agent and the environment. Taking autonomous driving scenarios as an example, vehicles need to autonomously complete complex tasks such as lane keeping, automatic following, lane changing and overtaking, and collision avoidance control in traffic environments with multi-source sensor inputs, strong temporal constraints, and multi-participant interactions. Data-driven methods enable systems to develop powerful decision-making capabilities without manually exhaustively enumerating all traffic rules, significantly reducing the cost of manual modeling and demonstrating excellent scenario generalization potential.

[0004] However, while intelligent models bring performance leaps, they also introduce unprecedented security risks and engineering challenges. First, intelligent models typically possess high nonlinearity and a massive parameter dimension, containing tens or even millions of neurons and connection weights. This results in their decision-making mechanisms exhibiting strong black-box characteristics, making reverse engineering using traditional programming logic extremely difficult. Second, these models are often trained on limited datasets or in specific simulation environments. When deployed to real-world physical systems with long-tailed distribution biases, they are highly susceptible to policy biases due to unseen state inputs, potentially leading to serious safety incidents such as vehicle collisions or system malfunctions. Even more challenging is that, because the output of intelligent models is an end-to-end mapping formed by the combined effects of massive numbers of parameters, once an anomaly occurs, developers cannot backtrack errors to specific logical branches or conditional statements as they would with traditional software code.

[0005] Errors in intelligent decision-making models often manifest subtly as abnormal activation peaks in certain deep neurons within specific state subspaces, or as an overemphasis on global action value estimation. This lack of interpretability and coarse-grained localization leads to extremely crude error correction methods for intelligent models in current engineering practices. The commonly used approach in the industry is to spend huge sums of money to recollect scene data and retrain the model globally. This method not only lacks specificity regarding errors but also makes it difficult to guarantee the controllability of the system optimization process. Therefore, in the crucial process of promoting the large-scale and commercial application of autonomous control systems, it is urgent to establish an interpretable and refined internal error localization and diagnosis mechanism for data-driven intelligent models. Summary of the Invention

[0006] The main objective of this application is to provide a method, apparatus, and device for intelligent model error localization in autonomous control systems. It aims to address how to automatically, precisely, and statistically interpretably locate the internal neurons that cause system performance degradation or decision-making errors in deep reinforcement learning policy networks, which possess highly nonlinear and black-box characteristics, within an unlabeled interactive environment lacking a complete testing language. Traditional methods often struggle to attribute macroscopic policy failures to microscopic network parameter defects when dealing with long-term interactive feedback in reinforcement learning models. Since the success or failure of autonomous decision-making tasks such as autonomous driving typically depends on a complete temporal interactive trajectory, single-step actions are often difficult to directly classify as absolutely "correct" or "incorrect," rendering existing supervised learning error localization methods completely ineffective. Furthermore, existing fault localization techniques based on traditional software genealogies heavily rely on discrete "success / failure" binary labels. When dealing with continuous performance degradation caused by cumulative actions in autonomous decision-making tasks, these binary labels lose a significant amount of fine-grained feedback information, failing to accurately characterize the negative correlation between continuous reward decay and the activation patterns of underlying neurons.

[0007] To achieve the above objectives, the first aspect of this application provides an intelligent model error localization method for autonomous control systems, comprising: During the interaction between the autonomous control system and the dynamic physical environment, by registering forward computation hooks, the state-action-reward interaction trajectory of the agent and the neuron activation tensor of the hidden layer of the policy network are collected non-intrusively and synchronously. The neuron activation tensor is aligned with the instantaneous rewards and state transitions fed back from the external environment in time steps to construct a heterogeneous fusion dataset. Based on the heterogeneous fusion dataset, the structural risk suspicion of each neuron is calculated by using the structural risk localization channel and the excitation spectrum statistics of neurons in successful and failed rounds. Based on a heterogeneous fusion dataset, the localization channel is located by continuous reward decay. The reward-driven suspicion of each neuron is calculated based on the degree of negative correlation between the average activation intensity of the neuron in a round and the normalized cumulative reward. A coverage weight factor is constructed based on the global activation frequency of each neuron. The structural risk suspicion and reward-driven suspicion are normalized and weighted by coverage to generate a comprehensive suspicion ranking list. High-confidence suspicious neurons are selected as error localization results to guide the parameter repair of the policy network and improve the operational safety and control efficiency of the autonomous control system.

[0008] Optionally, a heterogeneous fusion dataset is constructed, including: the agent performs multi-round interactions in a dynamic physical environment according to the policy network, records the external state transition quadruples of each round, and calculates the undiscounted cumulative total reward of each round; the entire trajectory is labeled as success or failure using an adaptive median threshold; by registering forward inference hooks in the fully connected or convolutional hidden layers of the policy network, the activation tensors of all underlying neurons are captured in real time at each time step, and the activation tensors are aligned and stored with the state transition quadruples and success / failure labels of the corresponding time steps.

[0009] Optionally, the structural risk suspicion of each neuron is calculated, including: setting an adaptive activation threshold based on the historical activation distribution of a specified neuron in a specified layer of the policy network; determining whether the neuron is activated at a time step using a Boolean function, and aggregating at the round level to statistically determine the frequency characteristics in four dimensions: the number of rounds in which the system is activated and the system is successful, the number of rounds in which the system is activated and the system is unsuccessful, and the number of rounds in which the system is unsuccessful; and substituting the frequency characteristics in four dimensions into the risk assessment mapping function to calculate the structural risk suspicion for the failure and crash event.

[0010] Optionally, the risk assessment mapping function is the Ochiai index, which is the ratio of the number of rounds in which the system is activated and fails to the square root of the sum of the number of rounds in which the system is activated and fails to the number of rounds in which the system is activated and succeeds to the structural risk suspicion of the neuron.

[0011] Optionally, the reward-driven suspicion of each neuron is calculated, including: converting the total round reward into a zero-mean normalized reward variable using batch standard deviation, and dividing historical interactions into high-reward-dominant and low-reward-disadvantaged clusters using quantile thresholds, ignoring rounds in the middle interval; calculating the average activation intensity feature of neurons within a round; calculating the Pearson correlation coefficient between the average activation intensity feature and the round-normalized reward sequence; and extracting strongly negatively correlated components, i.e., components where more intense activation leads to higher system collision risk or lower passage efficiency, and mapping them to non-negative reward-driven suspicion.

[0012] Optionally, the structural risk suspicion and reward-driven suspicion are normalized and weighted by coverage, including: performing minimum-maximum range scaling on the structural risk suspicion vector and the reward-driven suspicion vector respectively, constraining the two scores to a closed interval between zero and one; constructing a structural coverage weight factor using the global activation frequency of each neuron in the full set of test cases; adding the product of the normalized structural risk suspicion and the first weighting coefficient to the product of the normalized reward-driven suspicion and the second weighting coefficient to obtain a dual-channel weighted sum; multiplying the result of adding a smoothing coefficient and the coverage weight factor as the coverage interpolation constraint multiplier; and multiplying the dual-channel weighted sum by the coverage interpolation constraint multiplier to obtain the final comprehensive suspicion.

[0013] Optionally, the autonomous control system is an intelligent connected vehicle autonomous driving system, and the dynamic physical environment is a multi-lane merging traffic environment; the state of the intelligent agent is a multi-lane grid occupancy map perception vector, and the actions include acceleration, deceleration, and lane change control commands; high-confidence suspicious neurons are extracted to identify the internal hidden danger parameters of the strategy network that cause autonomous vehicles to make incorrect avoidance or rear-end collisions in dense merging areas. By performing targeted fine-tuning or pruning repair on the hidden danger parameters, the traffic efficiency and collision avoidance safety of autonomous vehicles can be improved.

[0014] To achieve the above objectives, a second aspect of this application also provides an intelligent model error localization device for autonomous control systems, comprising: The trajectory synchronization acquisition module is used to non-intrusively collect the state-action-reward interaction trajectory of the agent and the neuron activation tensor of the hidden layer of the policy network by registering forward computation hooks during the interaction between the autonomous control system and the dynamic physical environment. The neuron activation tensor is aligned with the immediate rewards and state transitions fed back from the external environment in time steps to construct a heterogeneous fusion dataset. The lineage risk localization module is used to calculate the structural risk suspicion of each neuron based on the heterogeneous fusion dataset and through the structural lineage risk localization channel, according to the excitation spectrum statistics of neurons in successful and failed rounds. The reward decay tracking module is used to locate channels through continuous reward decay tracking based on heterogeneous fused datasets, and calculate the reward-driven suspicion of each neuron based on the degree of negative correlation between the average activation intensity of neurons in a round and the normalized cumulative reward. The fusion ranking and output module is used to construct a coverage weight factor based on the global activation frequency of each neuron, normalize and weight the structural risk suspicion and reward-driven suspicion, generate a comprehensive suspicion ranking list, and extract the suspicious neurons with high confidence as the error localization result to guide the parameter repair of the policy network to improve the operational safety and control efficiency of the autonomous control system.

[0015] Optionally, the phylogenetic risk localization module is also used to set an adaptive activation threshold for a specified neuron in a specified layer of the policy network based on its historical activation distribution; to determine whether the neuron is activated at a time step using a Boolean function, and to aggregate at the round level to count the number of rounds in which the system is activated and successful, the number of rounds in which the system is activated and failed, the number of rounds in which the system is not activated and successful, and the number of rounds in which the system is not activated and failed; and to substitute the frequency features of the four dimensions into the risk assessment mapping function to calculate the structural risk suspicion degree for the failure and collapse event. The reward decay tracking module is also used to convert the total round reward into a zero-mean normalized reward variable using batch standard deviation, and to divide historical interactions into high-reward-dominant and low-reward-disadvantaged clusters using quantile thresholds; calculate the average activation intensity feature of neurons within a round; calculate the Pearson correlation coefficient between the average activation intensity feature and the round-normalized reward sequence; and extract strong negatively correlated components and map them to non-negative reward-driven suspicion. The fusion sorting and output module is also used to perform minimum-maximum range scaling on the structural risk suspicion vector and the reward-driven suspicion vector to a closed interval between zero and one; construct a structural coverage weight factor using the global activation frequency of each neuron in the full test cases; add the product of the normalized structural risk suspicion and the first weighting coefficient to the product of the normalized reward-driven suspicion and the second weighting coefficient to obtain a dual-channel weighted sum; multiply the result of adding a smoothing coefficient and the coverage weight factor as the coverage interpolation constraint multiplier; and multiply the dual-channel weighted sum by the coverage interpolation constraint multiplier to obtain the final comprehensive suspicion.

[0016] To achieve the above objectives, a third aspect of this application also provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the intelligent model error localization method for autonomous control systems provided above is implemented.

[0017] This application discloses an intelligent model error localization method for autonomous control systems, comprising: during the interaction between the autonomous control system and the dynamic physical environment, by registering forward computation hooks, non-intrusive synchronously collecting the state-action-reward interaction trajectory of the agent and the neuron activation tensors of the hidden layer of the policy network, aligning the neuron activation tensors with the immediate rewards and state transitions of the external environment feedback in time steps to construct a heterogeneous fusion dataset; based on the heterogeneous fusion dataset, through the structural spectrum risk localization channel, calculating the structural risk suspicion of each neuron according to the excitation spectrum statistics of neurons in successful and failed rounds; based on the heterogeneous fusion dataset, through the continuous reward decay tracking localization channel, calculating the reward-driven suspicion of each neuron according to the negative correlation between the average activation intensity of neurons in rounds and the normalized cumulative reward; constructing a coverage weight factor based on the global activation frequency of each neuron, normalizing and weighting the structural risk suspicion and reward-driven suspicion, generating a comprehensive suspicion ranking list, and extracting high-confidence suspicious neurons as error localization results to guide the parameter repair of the policy network to improve the operational safety and control efficiency of the autonomous control system.

[0018] This application constructs a heterogeneous fusion dataset through non-intrusive synchronous acquisition, combines a structural spectrum risk localization channel and a continuous reward decay tracking localization channel for dual-line error localization, and introduces a coverage weight factor for fusion, achieving the following technical effects: First, improving the accuracy and interpretability of error localization: Addressing the pain point of difficulty in tracing faults caused by the large parameter matrix of deep reinforcement learning models, this application successfully builds an interpretive bridge from "external physical collision / trajectory boundary violation" directly to "internal specific hidden layer node anomaly," extending the granularity from the coarse network level or hierarchy to the single neuron level, solving the engineering pain point of extremely high costs associated with traditional global blind trial and error troubleshooting. Second, addressing the theoretical bottleneck of the lack of testing language in the field of autonomous control, achieving full-form defect capture: Traditional debugging heavily relies on well-labeled supervised learning labels, while this application, through the continuous reward decay tracking localization channel, breaks away from the coarse judgment logic of black-and-white, enabling not only the localization of "explicit fatal errors" that cause immediate vehicle crashes, but also the keen capture of "implicit chronic errors" that cause excessive braking and slow decline in traffic efficiency, achieving full lifecycle and full-form monitoring of defects in autonomous decision-making models. Third, it enhances the robustness of statistical diagnosis in the face of highly randomized environmental interactions and suppresses spurious correlations: The fusion weighting formula of this application forcibly introduces a coverage frequency adjustment factor, which effectively removes irrelevant normal neurons that are only activated by chance in a few marginal failure rounds and do not actually constitute causal harm. This results in a highly confident diagnostic list with extremely high purity, significantly reducing the misdiagnosis rate and providing a high-confidence safety check-up defense for the model. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall closed-loop verification process of the intelligent model error localization mechanism for autonomous control systems; Figure 2 This is a diagram illustrating the principles of deep reinforcement learning and its interaction with the environment. Figure 3 Schematic diagram of the dual-channel fusion ranking framework for spectrum suspicion and reward suspicion; Figure 4 This is a flowchart of an intelligent model error localization method for autonomous control systems.

[0020] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0022] In existing technologies, both academia and industry have explored error localization and repair for neural networks and deep reinforcement learning models. However, these technologies still exhibit significant drawbacks and limitations when dealing with autonomous control systems: First, existing localization techniques are too coarse-grained and lack interpretability for sequential decision-making scenarios. Traditional spectral software fault localization techniques are primarily geared towards source-level programs with explicit control flow, making it difficult to directly translate them into neural networks with continuous parameter spaces. Although recent works such as DeepFault have introduced the SBFL concept into deep learning, these methods almost entirely focus on supervised learning tasks such as image classification, heavily relying on clear, discrete "right / wrong" data labels. In deep reinforcement learning, agent decision-making is a long-term interactive process, and performance degradation is usually a gradual, cumulative result. Simply using binary labels (success or failure) loses a large amount of fine-grained information about reward fluctuations and policy degradation, failing to accurately characterize the negative correlation between continuous performance degradation and the activation patterns of underlying neurons. Second, deep reinforcement learning systems generally face the problem of missing test predictions. In traditional software testing, test languages ​​are used to determine the correctness of a program's output under specific inputs. However, in complex autonomous decision-making scenarios such as autonomous driving, the environmental state space is vast and continuous, and multi-agent games exist, making it impossible for developers to predefine an absolutely correct optimal action for every possible traffic state. Existing DRL fuzzing tools (such as MDPFuzz and STARLA) can efficiently search for boundary test cases leading to system failure through genetic algorithms or gradient ascent, but their main output is merely a set of high-risk states or a reproduction of failure trajectories. There remains a significant gap between these tools and accurately locating the internal components of the policy network that cause the error (i.e., root cause localization). Third, single-perspective attribution methods lack statistical robustness in complex dynamic environments and are prone to "spurious correlation" misjudgments. Existing network internal localization methods often rely solely on instantaneous gradient backpropagation of erroneous states or only on global activation frequency statistics. In autonomous decision-making scenarios, the random transition probability of the environment and the inherent exploration mechanism of the policy network lead to a large number of accidental trajectory fluctuations. A small number of boundary failure samples can easily cause single-perspective localization algorithms to output extreme scores, thus misclassifying healthy neurons that normally extract useful environmental features as high-risk nodes. Current technologies generally lack joint mathematical constraints on "sufficient neuronal activation coverage" and "continuous reward decay trend," leading to highly unstable diagnostic results under different random seeds. Fourth, there is a lack of a multidimensional alignment mechanism between temporal interaction trajectories and microscopic neuronal representations. Most current diagnostic studies separate the monitoring of external environmental anomalies from the analysis of internal model parameters.In actual engineering debugging, the R&D system lacks a non-intrusive method that can strictly align the Markov trajectory (including state, action, reward, and next state) with the activation vectors of neurons in each hidden layer of the network in time steps without destroying the original inference speed and execution logic. This directly limits the data acquisition of subsequent fine-grained attribution algorithms.

[0023] To address the technical challenges of intelligent models in autonomous control systems, such as black-box characteristics, missing labels, difficulty in attributing continuous performance degradation, and false correlation misjudgments, this invention proposes an error localization method for intelligent models in autonomous control systems. This technical solution creatively integrates the phylogenetic localization concept from traditional software engineering with the continuous reward feedback mechanism of reinforcement learning, and its core content encompasses the following three key technical aspects.

[0024] First, a non-intrusive mechanism for synchronously acquiring state-action-reward trajectories and internal neuron activations is constructed. Without altering the original forward inference logic and parameter space of the deep reinforcement learning model, forward computation hooks are registered to record the agent's multi-round closed-loop interaction sequences in a dynamic environment. At each time step, the output feature tensor of the policy network's hidden layer is precisely extracted and aligned with the immediate rewards and state transitions from the external environment, thereby establishing an isomorphic mapping dictionary dataset between the macroscopic system behavior trajectory and the microscopic parameter response sequence.

[0025] Secondly, a bilinear error localization algorithm at the neuron granularity is proposed based on joint modeling of spectrum and reward patterns. To overcome the limitation of traditional SBFL relying solely on binary labels, this scheme performs risk profiling in two orthogonal dimensions: 1. Structural Spectrum Risk Localization (SBFL Channel): Construct a binary quaternary spectrum of whether neurons are activated using the trajectories of successful and failed rounds. Apply classic risk assessment mapping functions (such as Ochiai, Tarantula, etc.) to calculate the structural risk suspicion degree and screen out structural units that are abnormally active in the rounds of system collapse (such as collision, boundary crossing).

[0026] 2. Continuous Reward Decay Tracking and Localization (RFL Channel): This approach delves into the continuous distribution characteristics of long-term rewards. It extracts the average activation intensity at the round-level, calculates the distributional difference between neuron activation characteristics and normalized cumulative rewards, and directly solves for the Pearson correlation coefficient between the two. Strongly negatively correlated components are extracted to construct reward-driven suspicion, enabling the identification of chronically flawed nodes that, while not causing complete failure, have severely hampered system performance.

[0027] Finally, a comprehensive score fusion based on coverage weighting and a full-order output of suspicious targets are performed. To eliminate spurious interference from a very small number of anomalous test samples, this scheme constructs coverage weights based on the global activation frequency of neurons. A linear weighted fusion model is used to mathematically fuse risk cues from the SBFL and RFL dimensions, outputting a comprehensive list of suspicious neurons with high statistical reliability and performance sensitivity. Finally, the top-M neurons with high confidence are selected as accurate diagnostic conclusions for system defects, providing direct targets for subsequent model repair.

[0028] refer to Figure 1 , Figure 2 , Figure 3 and Figure 4 The first embodiment of this application provides an intelligent model error localization method for autonomous control systems, addressing the technical problem presented in the background art: how to automatically, precisely, and statistically interpretably locate the internal neurons that cause system performance degradation or decision-making errors in deep reinforcement learning policy networks with highly nonlinear and black-box characteristics, within an unlabeled interactive environment lacking a complete testing language. This method can be executed by a processor, which can be located in a terminal or server. The execution process of the intelligent model error localization method for autonomous control systems is as follows: In step S101, during the interaction between the autonomous control system and the dynamic physical environment, the state-action-reward interaction trajectory of the agent and the neuron activation tensor of the hidden layer of the policy network are collected non-intrusively and synchronously by registering forward computation hooks. The neuron activation tensor is aligned with the immediate rewards and state transitions fed back from the external environment in time steps to construct a heterogeneous fusion dataset.

[0029] In one embodiment of this application, a heterogeneous fusion dataset is constructed, including: an agent performing multi-round interactions in a dynamic physical environment according to a policy network, recording the external state transition quadruples for each round, and calculating the undiscounted cumulative total reward for each round; labeling the entire trajectory as a success or failure label using an adaptive median threshold; and capturing the activation tensors of all underlying neurons in real time at each time step by registering forward inference hooks in the fully connected or convolutional hidden layers of the policy network, and aligning and storing the activation tensors with the state transition quadruples and success / failure labels for the corresponding time steps.

[0030] Furthermore, this invention first applies to Markov decision processes (... Within the theoretical framework of [the theory], a rigorous formal definition is given for the autonomous decision-making model and its operating environment. The environmental interaction is defined by a quintuple [structure / mechanism]. Composition. In time step The system is in a state. The deep reinforcement learning policy network inside the intelligent body receives high-dimensional perception vectors (e.g., a flattened multi-lane grid occupancy map) and outputs action value distributions through multi-layer nonlinear transformations. Intelligent agents based on Explore - Use strategies to select actions Distribute to the environment for execution.

[0031] During the data collection phase, the intelligent agent relies on... The strategy executes several rounds of interactions within the environment. The collector records not only the external state transition quadruple for each round, but also... Calculate the total undiscounted cumulative reward for each round. and using an adaptive median threshold The entire trajectory The data is labeled as either successful or failed; more importantly, the collector registers forward inference hooks in the fully connected or convolutional hidden layers of the PyTorch model at each time step. Real-time capture of activation tensors of all underlying neurons Therefore, the system constructs a heterogeneous fusion dataset that includes trajectory structure, environmental rewards, success / failure labels, and deep activation sequences. .

[0032] For example, in the tool's presentation layer interface, the autonomous decision-making model is loaded and its structure is parsed. Typical Highwaymerge-v0 environment parameters are configured, and data acquisition parameters are set to collect state-action-reward trajectories and neuron activation data. Trajectory statistical summaries and disk landing data are output as unified inputs for subsequent modules.

[0033] Step S102: Based on the heterogeneous fusion dataset, the structural risk suspicion of each neuron is calculated through the structural risk localization channel, according to the excitation spectrum statistics of neurons in successful and failed rounds.

[0034] In one embodiment of this application, the calculation of the structural risk suspicion of each neuron includes: setting an adaptive activation threshold for a specified neuron in a specified layer of the policy network based on its historical activation distribution; determining whether the neuron is activated at a time step using a Boolean function, and aggregating at the round level to statistically determine the frequency characteristics in four dimensions: the number of rounds in which the system is activated and the system is successful, the number of rounds in which the system is activated and the system is unsuccessful, and the number of rounds in which the system is unsuccessful; and substituting the frequency characteristics in four dimensions into a risk assessment mapping function to calculate the structural risk suspicion for the failure and crash event.

[0035] Furthermore, drawing on the coverage spectrum concept from traditional software debugging, this solution shifts the logic from "whether code statements are executed" to "whether neurons are significantly activated." For the first... Layer Each neuron has an adaptive threshold set based on its historical activation distribution. Through Boolean functions Determine its time. Whether or not it is activated, and further aggregated at the round level. Thus, the system statistically analyzes the frequency characteristics of each neuron across four key dimensions: (Number of rounds in which the system successfully activated) (Number of rounds in which the system was activated but failed) (Number of rounds inactive but successfully activated by the system) and (Number of rounds inactive and the system failed).

[0036] After obtaining the quaternary statistics, the system calls a series of classic risk assessment mapping functions for quantification. This invention supports the following five types of measurement formulas, and by default prioritizes the Ochiai index, which has excellent robustness: The Ochiai index essentially calculates the cosine similarity between the failure label vector and the neuron activation vector. It exhibits excellent robustness to fluctuations in the size of failure samples.

[0037] ; This formula, by calculating the cosine similarity between the failure label vector and the neuron activation vector, can effectively filter out structural units that are abnormally active during system crashes or collision rounds, forming a spectrum suspicion score as the first line of defense. .

[0038] Tarantula metric: Measures risk from the perspective of conditional failure probability, comparing the difference in activation rates between failed and successful samples.

[0039] ; Jaccard index: Characterizes the overlap between neuron activation events and failed round events from the perspective of the intersection-union ratio of sets.

[0040] ; Ample metric: Directly calculates the absolute difference in activation probabilities under different success and failure states.

[0041] ; Wong's metric: Utilizes the absolute difference in activation counts, with minimal computational overhead.

[0042] ; Through the SBFL pathway, the system can effectively filter out explicit defective nodes that cause catastrophic errors in the model (such as direct collisions or running off the road).

[0043] In an optional embodiment of this application, the risk assessment mapping function is the Ochiai index, which is the ratio of the number of rounds in which the system is activated and the system fails to the square root of the sum of the number of rounds in which the system is activated and the system succeeds to the structural risk suspicion of the neuron.

[0044] For example, the network is automatically segmented into two sets: success and failure, using an adaptive median. Based on a set minimum positive real activation threshold, the tensor is aggregated into Boolean values ​​along the time dimension. All network nodes are traversed, and the frequency of their quadruple coverage in the full trajectory is counted. These quadruples are then substituted into the Ochiai core formula to calculate the high-dimensional cosine similarity for failure / collapse events, generating an initial vector.

[0045] Step S103: Based on the heterogeneous fusion dataset, the localization channel is located by continuous reward decay tracking. The reward-driven suspicion of each neuron is calculated according to the degree of negative correlation between the average activation intensity of the neuron in the round and the normalized cumulative reward.

[0046] In one embodiment of this application, calculating the reward-driven suspicion of each neuron includes: converting the total reward of the round into a normalized reward variable with zero mean using batch standard deviation, and dividing historical interactions into high-reward-dominant clusters and low-reward-disadvantaged clusters using quantile thresholds, ignoring rounds in the middle interval; calculating the average activation intensity feature of the neuron within the round; calculating the Pearson correlation coefficient between the average activation intensity feature and the round-normalized reward sequence; and extracting strongly negatively correlated components, i.e., components where the more intense the activation, the higher the system collision risk or the lower the passage efficiency, and mapping them to non-negative reward-driven suspicion.

[0047] Furthermore, simple binary success or failure labels often mask the subtle degradation trend of cumulative rewards in deep reinforcement learning. To address this, this invention innovatively introduces a suspicion assessment mechanism based on round-based reward patterns. First, the total round reward is converted into a normalized variable with zero mean using the standard deviation. This directly reflects the relative performance of that round compared to the average level.

[0048] ; The historical interactions were then divided into high-reward advantage clusters using quantile thresholds. and low-reward disadvantage clusters , Rounds in the middle range are selectively ignored when modeling the reward pattern distribution to reduce statistical interference caused by label uncertainty.

[0049] In addition, to go beyond simple Boolean activation determination, the system calculates the average activation intensity feature of the neuron within a round. ; Among them, Z i (k) This represents the average activation intensity of the i-th neuron in the k-th round. It is the mean of the neuron's output after non-negative truncation, reflecting the positive activation level of the neuron throughout the round, and can be used to measure the strength of the neuron's positive contribution to the decision.

[0050] i represents the neuron index, used to identify a specific target neuron.

[0051] k represents the episode index, used to identify a specific interaction episode; T represents the total number of time steps within a single round. 1 / T is a normalization coefficient, which is averaged over the total activation amount to eliminate the influence of differences in round length and make the activation intensity comparable across rounds of different lengths.

[0052] The original forward output value of the neuron, where the subscript l indicates the network layer number where the neuron is located (e.g., hidden layers fc1, fc2). The subscript j represents the index number of the neuron in layer l, the superscript (k) corresponds to the kth interaction round, and t represents the time step number within the round, which is the original continuous output value of the neuron after linear calculation.

[0053] This means that the original output of the neuron is truncated using ReLU (taking the non-negative part). If the output value is greater than 0, the original value is retained; if it is less than or equal to 0, it is taken as 0.

[0054] This operation retains only the positive activation of neurons, filters out negative outputs, focuses on the positive activation contribution of neurons, and is consistent with the logic of the commonly used ReLU activation function.

[0055] Subsequently, the system measures the degradation contribution of neurons from two perspectives: one is to calculate the distribution difference in the clusters of good and bad trajectories. Secondly, the Pearson correlation coefficient between the neuronal aggregation activation features and the round-normalized reward sequence is directly calculated. The system extracts strongly negatively correlated components (i.e., the more intense the activation, the higher the system collision risk or the lower the passage efficiency), and maps them to non-negative reward-driven suspiciousness. This diagnostic approach, based on continuous characterization, can keenly identify chronic defects that, while not causing complete failure, have severely hampered system performance.

[0056] For example, the batch standard deviation is calculated and center-normalized for the final cumulative total score of each round. Rounds with extreme ranges are segmented based on dynamic quantile lines. The algebraic mean of the activation intensity of a single neuron within a single round time step is extracted. The linear coupling slope between the normalized reward and the activation mean is rigorously calculated using the Pearson product-moment correlation coefficient method. Positively correlated signals are filtered out, retaining only the negatively correlated components indicating policy regression, generating a continuous degradation penalty vector.

[0057] Step S104: Construct a coverage weight factor based on the global activation frequency of each neuron, normalize and weight the structural risk suspicion and reward-driven suspicion, generate a comprehensive suspicion ranking list, and extract the suspicious neurons with high confidence as the error localization results to guide the parameter repair of the policy network to improve the operational safety and control efficiency of the autonomous control system.

[0058] In one embodiment of this application, the structural risk suspicion and reward-driven suspicion are normalized and weighted by coverage, including: performing minimum-maximum range scaling on the structural risk suspicion vector and the reward-driven suspicion vector respectively, constraining the two scores to a closed interval between zero and one; constructing a structural coverage weight factor using the global activation frequency of each neuron in the full set of test cases; adding the product of the normalized structural risk suspicion and the first weighting coefficient to the product of the normalized reward-driven suspicion and the second weighting coefficient to obtain a dual-channel weighted sum; using the result of adding a smoothing coefficient and the coverage weight factor as the coverage interpolation constraint multiplier; and multiplying the dual-channel weighted sum by the coverage interpolation constraint multiplier to obtain the final comprehensive suspicion.

[0059] Furthermore, since single-channel evaluation is prone to falling into the "spurious correlation" trap of extreme samples when faced with environmental noise, the system introduces a bidirectional complementarity and coverage constraint mechanism in the final stage to output the fusion suspicion level. Spectral Suspicion Level Suspiciousness of rewards The numerical scales are not consistent. To prevent any one channel from absolutely dominating the final ranking on the numerical scale, the system performs an independent normalization mapping: ; right The same scaling operation is also performed to constrain the scores of both paths to a strictly closed interval.

[0060] If a neuron is activated only in a very few rounds of the full test cases, even if its calculated penalty score is extremely high, it severely lacks global statistical confidence. Therefore, the system utilizes its global activation frequency. Constructing a structure to cover weight factors : ; To eliminate spurious interference from a very small number of outlier samples, the system constructs a coverage weight based on the global activation frequency of neurons. And a linear weighted model is used to fuse the conclusions of the two pathways mentioned above: ; Among them, Sus fusion (i) represents the fusion and integration suspicion of the i-th neuron in the final output of the formula. The higher the value, the more likely the neuron is to be a faulty unit that causes performance degradation and erroneous behavior in the decision model. Finally, all neurons are sorted in descending order based on this value, and the Top-K are selected as target candidates for the repair stage.

[0061] Sus spec The normalized spectral suspicion score is calculated based on the SBFL (Spectrum-Based Fault Localization) method. Using rounds as test cases, the activation frequency of neurons in successful / failed rounds is statistically analyzed, and the original suspicion value is obtained using formulas such as Ochiai and Tarantula.

[0062] Sus rew This represents the normalized reward suspicion level, while the original reward suspicion level is calculated using the RFL method. Specifically, the original suspicion value is obtained by analyzing the negative correlation between neuron activation intensity and cumulative reward over rounds (stronger activation and worse reward indicate higher suspicion), using either the distribution difference method or the Pearson correlation coefficient method.

[0063] Similarly, it is normalized to the [0,1] interval using Min-Max to ensure consistency with the scale of spectral suspicion.

[0064] The weighting coefficient for spectral suspicion is a configurable hyperparameter with a value ≥ 0.

[0065] The larger this value, the more the fusion result relies on spectral localization evidence of "co-occurrence activation in failed rounds"; The weighting coefficient for reward suspicion is a configurable hyperparameter with a value ≥ 0.

[0066] The larger this value, the more the fusion result relies on performance degradation-related evidence of "low reward, high activation"; The coverage weight of neuron i is also called the activation sufficiency weight.

[0067] It is calculated as follows: the total number of times neuron i is activated in all rounds, divided by the highest total number of activations of all neurons, and normalized to the interval [0,1].

[0068] Its function is to measure whether the activation statistics of a neuron are sufficient: the more frequent the activation, the more reliable the statistical conclusion; neurons that are activated only in a very few rounds are highly suspicious and are likely to be accidental spurious correlations, and will be given a lower score by this weight.

[0069] The adjustment coefficient for the coverage weight is a configurable hyperparameter with a value range of [0,1].

[0070] In this formula, the first half of the parentheses represents the weighted sum of discrete and continuous evaluations across two channels, achieving positive interactive complementation of information; the second half represents the coverage interpolation constraint multiplier, utilizing... Smoothly magnify the scores of questionable nodes with a solid statistical base.

[0071] Through this rigorous normalization and fusion process, the system ultimately generates a full risk sequence list of global neurons, extracts the top-ranked high-confidence Top-M set (e.g., the top-10 neurons), and generates a structured vulnerability diagnosis list containing layer name identifiers, neuron indexes, and comprehensive scores. At this point, the autonomous decision-making model completes end-to-end tracing from black-box behavioral anomalies to specific network nodes.

[0072] For example, strict Min-Max interval constraints are applied to two orthogonal vectors. The absolute activation frequency of each neuron is extracted to generate smooth weights to offset randomness, and linear weighted interpolation is performed on the dual-path scores. Finally, a fast sorting algorithm is used to sort the scores in descending order, and the system outputs a diagnostic table: the top-ranked neurons are the core internal hidden danger parameters that could induce autonomous vehicles to make incorrect avoidance errors or rear-end collisions in densely flowing areas.

[0073] In an optional embodiment of this application, the autonomous control system is an intelligent connected vehicle autonomous driving system, and the dynamic physical environment is a multi-lane merging traffic environment; the state of the intelligent agent is a multi-lane grid occupancy map perception vector, and the actions include acceleration, deceleration, and lane change control commands; the intercepted high-confidence suspicious neurons are used to calibrate the internal hidden danger parameters of the strategy network that cause autonomous vehicles to make incorrect avoidance or rear-end collisions in dense merging areas, and the traffic efficiency and collision avoidance safety of autonomous vehicles are improved by directional fine-tuning or pruning the hidden danger parameters.

[0074] The following section details the end-to-end unlabeled error localization process in a specific autonomous driving multi-lane merging environment (Highwaymerge-v0) to demonstrate the linkage between various functional modules. However, those skilled in the art should understand that the specific operating environment and configuration thresholds described are for illustrative purposes only and are not intended to limit the scope of the claims.

[0075] Implementation Step 1: Model Deployment and Heterogeneous Trajectory Acquisition.

[0076] In the tool's presentation layer interface, the autonomous decision-making model is loaded and its structure is parsed. Typical Highwaymerge-v0 environment parameters are configured, and data acquisition parameters are set to collect state-action-reward trajectories and neuron activation data. Trajectory statistics summaries and disk landing data are output as unified input for subsequent modules.

[0077] Step 2: Statistical calculation of SBFL structural phylogenetic characteristics.

[0078] The system automatically divides the data into two sets: success and failure, based on an adaptive median. This is done using a predefined activation threshold of a minimal positive real number. Aggregate tensors into Boolean values ​​along the time dimension. Traverse all network nodes and count their respective positions in the full trajectory. Quadruple coverage frequency. Substituting the quadruples into the Ochiai core formula, we calculate the high-dimensional cosine similarity for failure / collapse events, generating an initial vector. .

[0079] Step 3: Calculation of RFL reward decay trend characteristics.

[0080] The final cumulative total score for the round The batch standard deviation is calculated and center normalized. Rounds of range are segmented based on dynamic quantile lines. The algebraic mean of the firing intensity of a single neuron within a single round time step is extracted. The Pearson product-moment correlation coefficient method is used to rigorously calculate the linear coupling slope between the normalized return and the activation mean. Positively correlated signals are filtered out, retaining only the negatively correlated components indicating policy regression, thus generating a continuous degradation penalty vector. .

[0081] Step 4: Range scaling, interpolation fusion, and root cause output.

[0082] Strict Min-Max interval constraints are applied to the two orthogonal vectors. Absolute activation frequencies of each neuron are extracted to generate smooth weights that mitigate randomness. Linear weighted interpolation is performed on the dual-path scores. Finally, the quicksort algorithm is used to sort the scores in descending order, and the system outputs a diagnostic table: the top-ranked neurons are the core internal hidden danger parameters that may cause autonomous vehicles to make incorrect avoidance or rear-end collisions in densely flowing areas.

[0083] To address the aforementioned technical problems, the second embodiment of this application provides an intelligent model error localization device for autonomous control systems, which solves the same technical problems as the method embodiment. The device may include the following modules: a trajectory synchronization acquisition module, a spectral risk localization module, a reward decay tracking module, and a fusion sorting and output module.

[0084] Among them, the trajectory synchronization acquisition module is used to non-intrusively collect the state-action-reward interaction trajectory of the agent and the neuron activation tensor of the hidden layer of the policy network in the process of interaction between the autonomous control system and the dynamic physical environment by registering forward computation hooks. The neuron activation tensor is aligned with the immediate rewards and state transitions fed back by the external environment in time steps to construct a heterogeneous fusion dataset.

[0085] The lineage risk localization module is used to calculate the structural risk suspicion of each neuron based on the heterogeneous fusion dataset and the structural lineage risk localization channel, according to the excitation spectrum statistics of neurons in successful and failed rounds.

[0086] The reward decay tracking module is used to locate channels based on heterogeneous fused datasets through continuous reward decay tracking. It calculates the reward-driven suspicion of each neuron based on the degree of negative correlation between the average activation intensity of neurons in a round and the normalized cumulative reward.

[0087] The fusion ranking and output module is used to construct a coverage weight factor based on the global activation frequency of each neuron, normalize and perform coverage weighted fusion on the structural risk suspicion and reward-driven suspicion, generate a comprehensive suspicion ranking list, and extract suspicious neurons with high confidence as error localization results to guide the parameter repair of the policy network to improve the operational safety and control efficiency of the autonomous control system.

[0088] In one embodiment of this application, the phylogenetic risk localization module is further configured to set an adaptive activation threshold for a specified neuron in a specified layer of the policy network based on its historical activation distribution; determine whether the neuron is activated at a time step using a Boolean function, and aggregate the data at the round level to count the number of rounds in which the system is activated and successful, the number of rounds in which the system is activated and failed, the number of rounds in which the system is not activated and successful, and the number of rounds in which the system is not activated and failed; and substitute the frequency features of the four dimensions into the risk assessment mapping function to calculate the structural risk suspicion degree for the failure and collapse event. The reward decay tracking module is also used to convert the total round reward into a zero-mean normalized reward variable using batch standard deviation, and to divide historical interactions into high-reward-dominant and low-reward-disadvantaged clusters using quantile thresholds; calculate the average activation intensity feature of neurons within a round; calculate the Pearson correlation coefficient between the average activation intensity feature and the round-normalized reward sequence; and extract strong negatively correlated components and map them to non-negative reward-driven suspicion. The fusion sorting and output module is also used to perform minimum-maximum range scaling on the structural risk suspicion vector and the reward-driven suspicion vector to a closed interval between zero and one; construct a structural coverage weight factor using the global activation frequency of each neuron in the full test cases; add the product of the normalized structural risk suspicion and the first weighting coefficient to the product of the normalized reward-driven suspicion and the second weighting coefficient to obtain a dual-channel weighted sum; multiply the result of adding a smoothing coefficient and the coverage weight factor as the coverage interpolation constraint multiplier; and multiply the dual-channel weighted sum by the coverage interpolation constraint multiplier to obtain the final comprehensive suspicion.

[0089] It is understood that this embodiment is a device claim corresponding to the method claim, and therefore this embodiment has all the technical effects of the method claim, which will not be repeated here.

[0090] To address the aforementioned technical problems, the third embodiment of this application provides an electronic device, including a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the intelligent model error localization method for autonomous control systems provided in the previous embodiments.

[0091] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for intelligent model error localization in autonomous control systems, characterized in that, include: During the interaction between the autonomous control system and the dynamic physical environment, by registering forward computation hooks, the state-action-reward interaction trajectory of the agent and the neuron activation tensor of the hidden layer of the policy network are collected non-intrusively and synchronously. The neuron activation tensor is aligned with the immediate rewards and state transitions fed back from the external environment in time steps to construct a heterogeneous fusion dataset. The autonomous control system is an intelligent connected vehicle autonomous driving system, and the dynamic physical environment is a multi-lane merging traffic environment. The state of the agent is a multi-lane grid occupancy map perception vector, and the actions include acceleration, deceleration, and lane change control commands. Based on the heterogeneous fusion dataset, the structural risk suspicion of each neuron is calculated through the structural risk localization channel, according to the excitation spectrum statistics of neurons in successful and failed rounds. Based on the heterogeneous fusion dataset, the reward-driven suspicion of each neuron is calculated by locating the channel through continuous reward decay tracking, according to the degree of negative correlation between the average activation intensity of the neuron in a round and the normalized cumulative reward. A coverage weight factor is constructed based on the global activation frequency of each neuron. The structural risk suspicion and the reward-driven suspicion are normalized and weighted by coverage to generate a comprehensive suspicion ranking list. High-confidence suspicious neurons are extracted as error localization results to guide the parameter repair of the policy network to improve the operational safety and control efficiency of the autonomous control system. The calculation of the structural risk suspicion of each neuron includes: For a specific neuron in a specified layer of the policy network, an adaptive activation threshold is set based on its historical activation distribution. The Boolean function is used to determine whether a neuron is activated at a time step, and the data is aggregated at the round level to statistically analyze the frequency characteristics in four dimensions: the number of rounds in which the system is activated and the system is successful, the number of rounds in which the system is activated and the system is unsuccessful, and the number of rounds in which the system is unsuccessful. Substituting the frequency characteristics of the four dimensions into the risk assessment mapping function, the structural risk suspicion degree for failure and crash events is calculated. The calculation of the reward-driven suspicion of each neuron includes: The total round reward is converted into a normalized reward variable with zero mean using batch standard deviation, and historical interactions are divided into high-reward dominant clusters and low-reward disadvantaged clusters using quantile thresholds, ignoring rounds in the middle range. Calculate the average activation intensity characteristics of the neuron over the round; Calculate the Pearson correlation coefficient between this average activation intensity feature and the round-normalized reward sequence; Extract the strongly negatively correlated components, that is, the components that are activated more intensely, resulting in higher system collision risk or lower passage efficiency, and map them to non-negative reward-driven suspiciousness. The structural risk suspicion and the reward-driven suspicion are normalized and weighted by coverage, including: Minimum-maximum range scaling is applied to both the structural risk suspicion vector and the reward-driven suspicion vector to constrain the scores of both paths to a closed interval between zero and one. Construct a structure coverage weight factor by utilizing the global activation frequency of each neuron in the full set of test cases; The product of the normalized structural risk suspicion degree and the first weighting coefficient is added to the product of the normalized reward-driven suspicion degree and the second weighting coefficient to obtain the dual-channel weighted sum. The result of multiplying the smoothing coefficient and the covering weight factor is used as the covering interpolation constraint multiplier. The final comprehensive suspicion level is obtained by multiplying the dual-channel weighted sum with the coverage interpolation constraint multiplier.

2. The intelligent model error localization method for autonomous control systems as described in claim 1, characterized in that, The construction of the heterogeneous fusion dataset includes: The agent performs multi-round interactions in a dynamic physical environment based on the policy network, records the external state transition quadruple for each round, and calculates the undiscounted cumulative total reward for each round. The entire trajectory is labeled as either success or failure using an adaptive median threshold. By registering forward inference hooks in the fully connected or convolutional hidden layers of the policy network, the activation tensors of all bottom neurons are captured in real time at each time step, and the activation tensors are aligned and stored with the state transition quadruples and success / failure labels of the corresponding time steps.

3. The intelligent model error localization method for autonomous control systems as described in claim 1, characterized in that, The risk assessment mapping function is the Ochiai index, which is the ratio of the number of rounds in which the system is activated and then fails to the square root of the sum of the number of rounds in which the system is activated and then fails to the number of rounds in which the system is activated and then succeeds to the structural risk suspicion of the neuron.

4. The intelligent model error localization method for autonomous control systems as described in claim 1, characterized in that, The extracted high-confidence suspicious neurons are used to identify potential hazards within the strategy network that cause autonomous vehicles to make incorrect avoidance or rear-end collisions in densely trafficked areas. By fine-tuning or pruning these potential hazards, the traffic efficiency and collision avoidance safety of autonomous vehicles can be improved.

5. An intelligent model error localization device for autonomous control systems, characterized in that, include: The trajectory synchronization acquisition module is used to non-intrusively and synchronously acquire the state-action-reward interaction trajectory of the agent and the neuron activation tensors of the hidden layer of the policy network during the interaction between the autonomous control system and the dynamic physical environment by registering forward computation hooks. The neuron activation tensors are aligned with the immediate rewards and state transitions fed back from the external environment in time steps to construct a heterogeneous fusion dataset. The autonomous control system is an intelligent connected vehicle autonomous driving system, and the dynamic physical environment is a multi-lane merging traffic environment. The state of the agent is a multi-lane grid occupancy map perception vector, and the actions include acceleration, deceleration, and lane change control commands. The lineage risk localization module is used to calculate the structural risk suspicion of each neuron based on the heterogeneous fusion dataset and through the structural lineage risk localization channel, according to the excitation spectrum statistics of neurons in successful and failed rounds. The reward decay tracking module is used to locate the channel through continuous reward decay tracking based on the heterogeneous fusion dataset, and calculate the reward-driven suspicion of each neuron according to the degree of negative correlation between the average activation intensity of the neuron in the round and the normalized cumulative reward. The fusion sorting and output module is used to construct a coverage weight factor based on the global activation frequency of each neuron, normalize and perform coverage weighted fusion on the structural risk suspicion degree and the reward-driven suspicion degree, generate a comprehensive suspicion degree ranking list, and extract suspicious neurons with high confidence as error localization results to guide the parameter repair of the policy network to improve the operational safety and control efficiency of the autonomous control system. The phylogenetic risk localization module is also used to set an adaptive activation threshold for a specified neuron in a specified layer of the policy network based on its historical activation distribution; to determine whether the neuron is activated at a time step using a Boolean function, and to aggregate at the round level to count the number of rounds in which the system is activated and successful, the number of rounds in which the system is activated and failed, the number of rounds in which the system is not activated and successful, and the number of rounds in which the system is not activated and failed; and to substitute the frequency features of the four dimensions into the risk assessment mapping function to calculate the structural risk suspicion degree for the failure and collapse event. The reward decay tracking module is also used to convert the total reward of the round into a normalized reward variable with zero mean using batch standard deviation, and to divide historical interactions into high-reward-dominant clusters and low-reward-disadvantaged clusters using quantile thresholds; calculate the average activation intensity feature of the neuron in the round; calculate the Pearson correlation coefficient between the average activation intensity feature and the round normalized reward sequence; and extract strong negatively correlated components and map them to non-negative reward-driven suspicion. The fusion sorting and output module is also used to perform minimum-maximum range scaling on the structural risk suspicion vector and the reward-driven suspicion vector to a closed interval between zero and one; construct a structural coverage weight factor using the global activation frequency of each neuron in the full test cases; add the product of the normalized structural risk suspicion and the first weighting coefficient to the product of the normalized reward-driven suspicion and the second weighting coefficient to obtain a dual-channel weighted sum; multiply the result of adding a smoothing coefficient and the coverage weight factor as the coverage interpolation constraint multiplier; and multiply the dual-channel weighted sum with the coverage interpolation constraint multiplier to obtain the final comprehensive suspicion.

6. An electronic device, characterized in that, It includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, it implements the intelligent model error localization method for autonomous control systems as described in any one of claims 1 to 4.

Images