An adaptive continuous learning method for a highway incident handling model agent

Abstract

The application provides a highway event handling model intelligent agent adaptive continuous learning method, through highway handling case diversity gain evaluation, blind area filling weighting coefficient calculation, combining diversity gain, blind area filling weighting coefficient and handling effect basic score to obtain diversity perception value. And evaluate the prediction uncertainty of the model to the handling case, combine the diversity perception value and the prediction uncertainty to obtain the active learning value score, select the handling case whose active learning value score is greater than the preset score. Based on the expected return of the model, the model training is triggered. In the model training process, the hierarchical forgetting and historical sample playback mechanism are used to guide the model training. Through the case diversity perception, model uncertainty driving, expected return optimization and hierarchical forgetting control mechanism, the adaptive continuous optimization and capability improvement of the highway event handling model intelligent agent are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic management technology, and more specifically, to an adaptive continuous learning method for an intelligent agent in a highway incident handling model. Background Technology

[0002] Highway incident handling is a core aspect of traffic management, encompassing the entire lifecycle management of various incidents such as abnormal parking, traffic accidents, littering, and congestion. With the development of deep learning technology, visual recognition-based intelligent event detection systems are gradually being applied to highway scenarios. However, the highway operating environment is significantly dynamic and complex: changes in weather conditions (sunny days, rainy days, heavy fog), fluctuations in lighting conditions (daytime, nighttime, tunnels), and differences in traffic flow composition (weekdays, holidays) are all intertwined, leading to significant fluctuations in model performance during actual operation.

[0003] A more prominent problem is the difficulty in handling long-tail events. While low-frequency, high-risk events such as hazardous chemical leaks, chain collisions, and tunnel accidents have a low probability of occurrence, they are extremely dangerous. Traditional static response methods are severely inadequate in identifying such events due to a scarcity of training samples. At the same time, the system generates a large amount of real-world case data during daily operation, but this valuable experience data is often not fully utilized due to a lack of effective knowledge solidification mechanisms.

[0004] Existing technologies mainly include two methods: one is a periodic processing method based on manual annotation, and the other is a dynamically triggered incremental processing method based on performance monitoring. However, both methods lack a global and systematic evaluation mechanism for the learning value of cases. The first method relies entirely on human experience and judgment, while the second method only uses simple quantitative statistics or performance thresholds. Neither of them can make scientific decisions from the perspective of global optimization, such as what knowledge the model most needs to learn, which cases contribute the most to the incremental knowledge system, and when to train to obtain the greatest benefit.

[0005] These shortcomings prevent existing technologies from achieving an optimal balance in several aspects: first, they cannot strike a balance between case quality and case diversity, potentially leading to repeated learning of similar cases while ignoring knowledge gaps; second, they cannot strike a balance between learning new knowledge and retaining old knowledge, easily resulting in catastrophic forgetting; and third, they cannot strike a balance between training frequency and training gains, potentially leading to overtraining or undertraining. These problems restrict the effective exercise of the agent's continuous learning and adaptive optimization capabilities. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive continuous learning method for a highway incident handling model agent, so as to achieve adaptive continuous optimization and capability improvement of the highway incident handling model agent.

[0007] In a first aspect, the present invention provides an adaptive continuous learning method for a highway incident handling model agent, the method comprising:

[0008] Extract highway incident handling cases from the highway incident process logs;

[0009] Diversity gain is obtained based on the similarity between the treatment case and existing cases, and a blind spot filling weighting coefficient is obtained by cluster analysis of existing cases. The perceived value of diversity is obtained by combining the diversity gain, the blind spot filling weighting coefficient and the basic score of treatment effect.

[0010] Evaluate the prediction results of the disposal case executed by the disposal model agent, and determine the prediction uncertainty of the disposal model agent for the disposal case based on the prediction results;

[0011] The active learning value score is obtained by combining the perceived value of diversity and the uncertainty of prediction. Cases with active learning value scores greater than the preset score are added to the sample library.

[0012] Evaluate the benefit-cost ratio of model training based on newly added samples in the sample library. If the benefit-cost ratio is greater than a dynamic threshold, then start model training.

[0013] During model training, a total loss function is constructed that includes the training loss of new samples, the hierarchical EWC constraint loss, and the replay loss of historical samples. The model is trained under the guidance of the total loss function until the preset requirements are met.

[0014] The beneficial effects of the embodiments of the present invention are as follows:

[0015] This invention provides an adaptive continuous learning method for a highway incident handling model agent. It extracts incident handling cases from highway incident process logs, obtains diversity gain based on the similarity between the handled cases and existing cases, and obtains blind spot filling weighting coefficients through cluster analysis of existing cases. The method combines diversity gain, blind spot filling weighting coefficients, and a basic incident handling effect score to obtain a diversity perceived value. The prediction results of the handling case processing by the model agent are evaluated, and the prediction uncertainty of the model agent for the handling case is determined based on the prediction results. An active learning value score is obtained by combining the diversity perceived value and prediction uncertainty, and handling cases with an active learning value score greater than a preset score are added to the sample library. The benefit-cost ratio of model training based on newly added samples in the sample library is evaluated. If the benefit-cost ratio is greater than a dynamic threshold, model training is initiated. During model training, a total loss function is constructed, including training loss for newly added samples, hierarchical EWC constraint loss, and replay loss for historical samples. The model is trained under the guidance of the total loss function until preset requirements are met.

[0016] This solution achieves adaptive continuous optimization and capability enhancement of the intelligent agent of the highway incident handling model through case diversity perception, model uncertainty-driven approach, expected return optimization, and hierarchical forgetting control mechanism. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an adaptive continuous learning method for a highway incident handling model agent provided in an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of the overall logic of the continuous learning method provided in the embodiments of the present invention;

[0020] Figure 3 This is a schematic diagram of the overall logic of diversity perception assessment provided in an embodiment of the present invention;

[0021] Figure 4 A schematic diagram of the overall logic of the model uncertainty-driven active learning process provided in the embodiments of the present invention;

[0022] Figure 5 A schematic diagram illustrating the overall logic of training timing optimization decision based on expected returns, provided in an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the overall logic of hierarchical forgetting control provided in an embodiment of the present invention;

[0024] Figure 7 This is a schematic diagram illustrating the overall logic of progressive grayscale release provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0026] Currently, the main methods for handling highway incidents include periodic processing methods based on manual annotation and dynamic triggering incremental processing methods based on performance monitoring.

[0027] Among these methods, the periodic processing approach based on manual annotation is the mainstream technical route for optimizing the current highway incident handling system model. The basic process is as follows: During system operation, a large number of incident handling records are generated. Maintenance personnel periodically manually select representative typical cases from these records, and then perform detailed manual annotation on these selected cases. This includes annotating the specific type of event (e.g., abnormal parking, spilled debris, traffic accidents), the precise location coordinates of the event, the severity level of the event, and the scope of its impact. The complete annotation of a single case typically requires 2-3 hours of manual time.

[0028] When the labeled data accumulates to a certain amount (e.g., 500 new samples) or reaches a preset fixed time period (e.g., once a month or once a quarter), the system will trigger the model retraining process. The retraining process uses all historical labeled data and newly labeled data to completely retrain the model, which typically takes several hours to several days. After training is complete, the new version of the model is deployed to the production environment to replace the old version and continue running.

[0029] Dynamic triggering incremental processing based on performance monitoring is an improvement that has emerged in recent years, introducing a dynamic triggering mechanism on top of the first method. This method adds a model performance monitoring module to track key performance indicators (KPIs) during online model execution in real time, such as recognition accuracy, event recall, and false positive rate. When the system detects that a performance indicator has decreased by more than a certain threshold (e.g., a 5% drop in accuracy), it immediately triggers the incremental training process without waiting for a fixed training cycle.

[0030] Meanwhile, this method retains the sample-based triggering mechanism, automatically triggering training when the number of new training samples reaches a certain threshold (e.g., 100 new samples). In terms of training method, this method adopts an incremental training mode, using only the newly added samples to incrementally update the model, unlike the first method which uses all historical data for full retraining, thus significantly reducing training time costs.

[0031] However, both of the above methods have many drawbacks. Specifically, the first method has drawbacks such as extremely low knowledge accumulation efficiency, inability to respond to model performance degradation in a timely manner, and excessively long accumulation period for long-tail event samples.

[0032] Specifically, because case selection and data annotation rely entirely on manual operation, processing efficiency is severely limited. Actual operational data shows that in an application scenario handling approximately 2,500 real-world events, due to limitations in manpower and time investment, fewer than 100 cases were ultimately able to complete manual selection and annotation, resulting in a knowledge retention rate of only about 4%. This means that over 96% of real-world handling experience failed to be converted into training data for model learning due to manpower constraints. Particularly for long-tail events (such as hazardous chemical leaks and chain collisions), due to their extremely low frequency of occurrence, timely manual processing and annotation are even more difficult, causing the model's ability to identify these high-risk events to remain at a consistently low level.

[0033] Because of the fixed-period triggering mechanism (e.g., monthly or quarterly), when significant changes in the traffic environment (such as seasonal weather changes, holiday traffic pattern changes, or the emergence of new event types) cause a noticeable decline in model performance, the system cannot promptly detect this performance degradation, nor can it immediately initiate the model optimization process. Real-world testing data shows that when encountering new and complex scenarios such as rainy nights, the model's recognition accuracy can drop sharply from 87% in normal scenarios to 58%. However, with a monthly triggering cycle, the system needs to wait several weeks to complete the next retraining, during which time the system will continue to operate at low performance, severely impacting the timeliness and accuracy of event handling.

[0034] For long-tail events that occur very infrequently but pose a significant threat, manual annotation faces the fundamental challenge of an excessively long sample accumulation period. For example, statistics from a certain province's highway system show that only about three such incidents occur annually, and approximately seven chain-reaction collisions occur annually. At the current rate of manual annotation, it would take several years to accumulate a sufficient number of training samples (typically dozens of samples are needed for effective training). This slow sample accumulation rate makes it extremely difficult to improve the model's ability to identify long-tail events, which are precisely the high-risk scenarios that require the intelligent system to respond quickly and identify accurately.

[0035] In addition, the second method mentioned above has drawbacks such as lack of a quantitative evaluation mechanism for case value, neglect of sample diversity and complementarity, lack of optimization strategies for training timing, and prominent catastrophic forgetting problems.

[0036] Specifically, while this approach can automatically trigger training, it lacks a systematic quantitative evaluation method for the core question of "which cases are worth learning from, and which cases have higher learning value." The system often simply counts the number of new samples or continues to rely on human experience for case selection, failing to scientifically determine the actual contribution of each case to improving the model's capabilities. This results in inconsistent quality of training samples, with some low-value cases being included in the training set while some high-value cases may be overlooked.

[0037] The triggering mechanism of this scheme only focuses on a single dimension of sample quantity or performance indicators, completely disregarding the relationship between new samples and existing training data. This may lead to the system repeatedly learning from a large number of similar cases (e.g., multiple similar events on the same road segment under the same weather conditions), wasting training resources, while failing to effectively cover true knowledge blind spots (e.g., special scenarios never encountered before). The lack of consideration for sample diversity results in structural defects in the model's knowledge system, making it difficult to comprehensively improve generalization ability.

[0038] This scheme employs a simple threshold-triggered mechanism, initiating training immediately upon meeting the trigger condition (performance degradation or sufficient sample size), without considering whether it is the optimal training time. This can lead to two disadvantages: first, training is initiated when training gains are low (e.g., the quality and quantity of new samples are low), wasting valuable computational resources; second, training is delayed when it should be started as soon as possible (e.g., performance has already significantly degraded) because the trigger threshold has not been reached, resulting in persistently low system performance. The lack of scientific decision-making regarding training timing leads to less than ideal utilization of training resources.

[0039] During incremental training, the learning of new tasks alters the model's parameter weights, easily leading to significant performance loss on older tasks—a phenomenon known as "catastrophic forgetting." While this approach employs incremental training, it lacks an effective anti-forgetting mechanism. Particularly concerning is the recognition of low-frequency events, which are easily overwritten and forgotten after learning new high-frequency events, causing significant fluctuations in model performance and insufficient stability. In practical applications, it has been observed that after several incremental training iterations, the model's accuracy in recognizing early-learned long-tail events can decrease by 10%-15%, severely impacting system reliability.

[0040] Based on the above research findings, this invention provides an adaptive continuous learning method for an intelligent agent in a highway incident handling model. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating the adaptive continuous learning method for a highway incident handling model agent provided in an embodiment of the present invention. This adaptive continuous learning method for the highway incident handling model agent can be executed by an adaptive continuous learning device for the highway incident handling model agent. This device can be implemented in software and / or hardware and can be configured in an electronic device, such as a computer device, server, etc., for example, a server used for traffic management monitoring. The detailed steps of the adaptive continuous learning method for the highway incident handling model agent are described below.

[0041] S11, Extract highway incident handling cases from the highway incident process log;

[0042] S12, based on the similarity between the handled cases and existing cases, the diversity gain is obtained, and the blind spot filling weighting coefficient is obtained by cluster analysis of existing cases. The diversity perceived value is obtained by combining the diversity gain, the blind spot filling weighting coefficient and the basic score of the handling effect.

[0043] S13, Evaluate the prediction results of the disposal case executed by the disposal model agent, and determine the prediction uncertainty of the disposal model agent for the disposal case based on the prediction results;

[0044] S14, combine the perceived value of diversity and the uncertainty of prediction to obtain the active learning value score, and add the treatment cases with the active learning value score greater than the preset score to the sample library.

[0045] S15, evaluate the benefit-cost ratio of model training based on newly added samples in the sample library. If the benefit-cost ratio is greater than the dynamic threshold, start model training.

[0046] S16. During model training, a total loss function is constructed that includes the training loss of new samples, the hierarchical EWC constraint loss, and the replay loss of historical samples. The model is trained under the guidance of the total loss function until the preset requirements are met.

[0047] The adaptive continuous learning method for the highway incident handling model agent provided in this embodiment achieves the goal of continuous learning and self-improvement of the handling model agent from real-world operational data by establishing a value assessment mechanism that perceives case diversity, an active learning mechanism driven by model uncertainty, a training scheduling mechanism optimized based on expected returns, and a hierarchical forgetting control mechanism. Here, the handling model agent can be understood as a computer software system with processing capabilities, driven by a large model.

[0048] The specific implementation methods of each of the above steps will be explained in detail below.

[0049] Please see Figure 2 This is a schematic diagram illustrating the overall logic of the continuous learning method provided by this invention. This invention establishes a value assessment mechanism for case diversity perception; please refer to [the relevant documentation / reference]. Figure 3 This mechanism not only assesses the intrinsic quality of individual cases, but more importantly, it evaluates the incremental contribution of the case to existing knowledge and its value in filling knowledge gaps from the perspective of the global knowledge system.

[0050] First, incident handling cases are automatically extracted from the highway incident handling process logs to obtain case information, including key time nodes and operation log information. Key time node information includes the time when the incident was first detected by the monitoring system. The time when the monitor confirms the event The time to begin allocating resources for disposal Time for incident handling and traffic restoration The operation log information includes: the type and quantity of resources allocated (such as tow trucks, ambulances, etc.), the scope of the event's impact (length of congestion, number of lanes occupied), and the specific sequence of handling procedures.

[0051] Based on the information extracted from the cases handled, the system evaluates the basic effectiveness score from five dimensions. Specifically, the basic effectiveness score can be obtained in the following ways:

[0052] The actual response time is obtained based on the detection and handling time of each case. A response timeliness score is then calculated by combining the actual response time, standard response time, and maximum response time. The actual handling time of each case is obtained, along with the standard handling time of historical cases. A handling efficiency score is then calculated based on the actual and standard handling times. The total number of resources actually allocated for each case is calculated, and the optimal resource allocation from the expert knowledge base is obtained. A resource rationality score is then calculated based on the total number of resources and the optimal resource allocation. The congestion length and duration caused by each case are obtained. An impact control score is then calculated based on the congestion length, duration, and set threshold parameters. The actual handling process of each case is matched with a standard processing flow template. A process compliance score is then calculated based on the matching results. Finally, a basic handling effectiveness score is calculated by combining the response timeliness score, handling efficiency score, resource rationality score, impact control score, and process compliance score.

[0053] In this embodiment, the basic score for treatment effectiveness is obtained by evaluating the following five dimensions:

[0054] 1. Response Timeliness Score :

[0055] This score measures the system's response time to events. Response time is defined as... This refers to the time interval from when an event is detected to when action begins. Set a standard response time. Minutes, maximum acceptable response time Minutes. The formula for calculating the timeliness of response is:

[0056]

[0057] A score of 1.0 is awarded when the response time is less than or equal to the standard time; a score of 0 is awarded when the response time exceeds the maximum acceptable time; and the score decreases linearly when the response time is in between.

[0058] 2. Processing efficiency score :

[0059] This rating measures the efficiency of incident handling. The actual handling time is defined as... Establish standard handling times for each type of event based on historical statistical data. The formula for calculating the disposal efficiency score is as follows:

[0060]

[0061] When the actual processing time is less than or equal to the standard processing time, a full score of 1.0 is awarded; when the actual processing time exceeds the standard processing time, the score is the ratio of the standard processing time to the actual processing time, reflecting the efficiency loss.

[0062] 3. Resource rationality score :

[0063] This rating measures the rationality of resource allocation. It also includes the total quantity of various resources actually allocated. Based on the event type and severity, query the expert knowledge base for the optimal resource allocation quantity under this situation. The formula for calculating the resource rationality score is as follows:

[0064]

[0065] A perfect score of 1.0 is awarded when the actual allocation quantity perfectly matches the optimal quantity; the greater the allocation deviation, the lower the score. This scoring system avoids both ineffective handling due to insufficient resources and waste caused by excessive resource allocation.

[0066] 4. Impact on control scores :

[0067] This score measures the effectiveness of traffic control measures in addressing the impact of an event. It extracts the length of congestion caused by the event. (Unit: meters) and congestion duration (Unit: minutes), set threshold parameters Mihe Minutes. The impact control score uses an exponential decay method, as shown below:

[0068]

[0069] When the congestion length and duration are both short, the score is close to 1.0; when the impact range and duration are both large, the score decreases exponentially. This design can significantly distinguish cases of major impact.

[0070] 5. Process Standardization Scoring :

[0071] This rating measures the standardization of the handling process. The system maintains a standard operating procedure template library, defining standard handling step sequences for different event types. The actual executed steps are matched against the standard procedures, and the number of steps conforming to the standard procedures is counted. Total number of steps The formula for calculating the process standardization score is as follows:

[0072]

[0073] This rating reflects the professionalism and standardization of the handling process, and helps to select high-quality cases that are well-organized and highly reusable.

[0074] Combining the scores for timeliness of response, efficiency of handling, rationality of resources, impact control, and process standardization, the scores from these five dimensions are weighted and summed to obtain the basic score for the handling effect of the case. The calculation formula is as follows:

[0075]

[0076] Wherein, the weight coefficients satisfy Based on the actual needs of highway incident handling and expert experience, the weights are set as follows: (Timeliness of response) (Processing efficiency) (Resource rationality) (Impact on control) (Process Specifications). Among them, impact control has the highest weight, reflecting the core objective of highway management: "rapidly restoring traffic and reducing social impact."

[0077] Building upon the basic scoring, this invention introduces the calculation of case diversity gain. Diversity gain assesses the degree of difference between new cases and the existing case library, avoiding the system from repeatedly learning a large number of similar cases and instead prioritizing cases that can supplement new knowledge. Specifically, the diversity gain is obtained based on the similarity between the disposal case and existing cases. This step can be implemented in the following way:

[0078] The case information of the disposed case is encoded into a feature vector; the cosine similarity between the feature vector of the disposed case and the feature vector of the existing case is calculated; the existing case most similar to the disposed case is determined based on the cosine similarity; the diversity gain is calculated based on the cosine similarity between the disposed case and the most similar existing case, wherein the diversity gain is negatively correlated with the cosine similarity.

[0079] To calculate the similarity between cases, the key features of the cases first need to be encoded into a unified vector representation. The key features of the cases include three aspects: event type (e.g., 8 categories such as abnormal parking, littering, traffic accidents, etc.), scene features (weather conditions, lighting conditions, road features, etc.), and handling process (sequence of operation steps, resource allocation methods, etc.).

[0080] The case information is encoded into a feature vector as shown below using a pre-trained multimodal encoder (e.g., the CLIP model):

[0081]

[0082] in, The embedding function maps text descriptions and visual features to a unified high-dimensional vector space (e.g., 512-dimensional or 768-dimensional), so that semantically similar cases are closer in the vector space, while cases with large semantic differences are farther apart.

[0083] For newly generated cases Calculate its eigenvector With all existing cases in the case library eigenvectors Cosine similarity between them:

[0084]

[0085] in, Represents the vector dot product. This represents the L2 norm of the vectors. The cosine similarity ranges from -1 to 1, where a value closer to 1 indicates greater similarity between the two cases, and a value closer to -1 indicates less similarity.

[0086] The diversity gain of a case is defined as the degree of difference between that case and the most similar case in the case library:

[0087]

[0088] The formula means: first, find the existing case in the case library that is most similar to the newly generated disposal case (i.e., the maximum cosine similarity); then subtract this maximum cosine similarity from 1 to obtain the diversity gain. A value close to 1 indicates that the newly generated case has a very low similarity to all existing cases in the case library, meaning that the case is entirely new and unique, possessing high diversity value; if... A value close to 0 indicates that there are already existing cases in the case library that are very similar to the case being handled, and the marginal benefit of learning from the case again is low.

[0089] In addition to assessing the differences between individual cases and existing cases, this invention also identifies "knowledge blind spots" in the case space from the perspective of the global structure of the knowledge system. These are areas with low sample density and insufficient model coverage. Newly generated disposal cases falling into these areas are given additional weights to guide the system to prioritize filling knowledge blind spots.

[0090] Specifically, the blind spot filling weighting coefficient is obtained by performing cluster analysis on existing cases. This step can be achieved in the following way:

[0091] Cluster analysis is performed on existing cases to divide them into multiple clusters, and the cluster density of each cluster is calculated; the target cluster to which the treatment cases are assigned is determined, and the cluster density of the target cluster is obtained; the blind spot filling weighting coefficient of the treatment cases is set according to the cluster density of the target cluster, wherein the blind spot filling weighting coefficient is negatively correlated with the cluster density.

[0092] First, cluster analysis of the case space is performed. The feature vectors of all existing cases in the case library are used as input, and the K-means clustering algorithm is used to divide the case space into multiple clusters:

[0093]

[0094] in, The total number of cases in the case library. This represents the number of clusters. The selection adopts a dynamic strategy: setting ,in Total number of event types (e.g.) ,but This setup allows for further differentiation of different sub-scenes within each event type, enabling a more granular division of the knowledge space.

[0095] Then, cluster density is calculated for each cluster. Calculate its sample density, that is, cluster density. Cluster density is defined as the ratio of the number of samples in a cluster to the cluster volume.

[0096]

[0097] in, This indicates the number of cases contained in a cluster. This represents the volume of a cluster in the feature space. The cluster volume is estimated using the within-cluster sample variance.

[0098]

[0099] in, The dimension of the feature vector. For samples within a cluster at the 1st The standard deviation across all dimensions. High-density clusters indicate that the region has sufficient knowledge coverage and abundant samples; low-density clusters indicate that the region has insufficient knowledge coverage and knowledge blind spots.

[0100] Finally, a weighted calculation for filling coverage blind spots is performed, where newly generated disposal cases are assigned to a specific cluster. At any given time (by calculating the distance between the treatment case and the center of each cluster, selecting the nearest cluster), a blind spot filling weighting coefficient is assigned to the newly generated treatment case based on the cluster density of that cluster:

[0101]

[0102] in, To adjust the parameters, this formula uses an exponential decay form, where the cluster density... At a lower level (knowledge gap). A value close to 1 is assigned a higher weight; this applies when the cluster density is high (in areas with sufficient knowledge). Smaller weighting reduces the overall weight. This design guides the system to prioritize areas with insufficient knowledge coverage, achieving a balanced development of the knowledge system.

[0103] After obtaining the baseline score for treatment effectiveness, diversity gain, and blind spot filling weighted coefficient through the above methods, the perceived value of diversity is obtained by combining the baseline score for treatment effectiveness, diversity gain, and blind spot filling weighted coefficient. Specifically, this can be achieved through the following methods:

[0104] The preliminary perceived value of diversity is obtained by combining diversity gain, blind spot filling weighting coefficient and basic score of treatment effect; when the treatment case belongs to long tail event, the occurrence frequency of long tail event is obtained, and the perceived value of diversity is calculated based on the occurrence frequency, average frequency and preliminary perceived value of diversity.

[0105] By combining the baseline score of the handling effect, the diversity gain, and the weighted coefficient of the blind spot filling, the preliminary perceived value of the diversity of the handling cases is obtained:

[0106]

[0107] The design of this formula reflects the following considerations: basic score of treatment effectiveness. As a multiplicative factor, it ensures that only high-quality cases can obtain high scores; the weight allocation in parentheses reflects the evaluation concept of "quality as the foundation (0.5), diversity as the focus (0.35), and filling blind spots as a supplement (0.15)".

[0108] For long-tail events (event types that occur very infrequently), additional scarcity weighting is required. Define the occurrence frequency of the event type. The average frequency represents the proportion of this type of event to the total number of events. ( (Total number of event types). When If a newly generated case is classified as a long-tail event, and it is also a long-tail event, then scarcity weighting is applied to obtain the final perceived value of diversity:

[0109]

[0110] Among them, the weighting coefficients of the logarithmic terms can be dynamically adjusted according to the scarcity of the events. For example, if the frequency of a certain type of long-tail event is only... The average frequency is (8 types of events), then the weighting coefficient is approximately This significantly elevates the learning priority of this type of case. The outermost layer The function is used to limit the final score to no more than 1.0.

[0111] Ultimately, when the perceived value of the diversity of the cases handled exceeds a certain threshold, for example... If the case is deemed high-value, it will be marked as a high-value case and enter the subsequent proactive learning process; otherwise, it will be judged as a regular case.

[0112] It should be noted that in application scenarios with limited computing resources or extremely high real-time requirements, a simplified diversity assessment mechanism can be adopted: omitting the knowledge blind spot identification step of incremental clustering analysis, and only retaining the diversity gain calculation based on cosine similarity. Specifically, for each new case, the similarity between it and existing cases in the case library is calculated, and the diversity gain is directly applied. As a diversity score, cluster analysis and blind spot filling weighting are no longer performed.

[0113] The advantages of this alternative approach are a reduction in computational complexity of approximately 40% and better real-time performance, making it suitable for scenarios with stringent response speed requirements. The disadvantage is a slight decrease in the effectiveness of diversity optimization due to the lack of awareness of the global structure of the knowledge space, potentially overlooking some low-density areas (knowledge blind spots). It is expected that under this approach, the improvement in scenario coverage of the case library may decrease from 30% to around 20%, but it will still be significantly better than existing technologies.

[0114] Building upon the above, the core innovation of this invention also includes the introduction of model uncertainty quantification technology, actively selecting training samples from the perspective of "what the model needs to learn." Traditional methods passively receive all cases, while this method, by evaluating the model's prediction uncertainty for each case, prioritizes learning from cases where the model is "uncertain," thereby significantly improving sample utilization efficiency. The model uncertainty-driven active learning process is as follows: Figure 4 As shown.

[0115] In this embodiment, the prediction results of the disposal case executed by the disposal model agent are evaluated, and the prediction uncertainty of the disposal model agent for the disposal case is determined based on the prediction results. Specifically, this step can be implemented in the following way:

[0116] The case information of the disposal case is input into the existing model, and multiple prediction results are output; the variance of the multiple prediction results is calculated to obtain the uncertainty measure index; the uncertainty measure index is normalized to a uniform range; based on the normalized uncertainty measure index and a segmented weighting strategy, the prediction uncertainty is obtained.

[0117] First, uncertainty estimation based on Monte Carlo Dropout is performed. Model uncertainty reflects the model's confidence in its predictions. For cases where the model's predictions are highly certain (the model has already "mastered" this type of knowledge), the marginal benefit of further learning is low; for cases where the model's predictions are uncertain (the model lacks confidence in the scenario), the marginal benefit of learning is high. This invention uses the Monte Carlo Dropout method to estimate the model's prediction uncertainty.

[0118] For cases requiring assessment, input information. Including video frame sequences and environmental context, while keeping the model's Dropout layer active, [the process is performed]. Next (usually set) Forward propagation. During each forward propagation, the Dropout layer randomly drops some neuron connections, forming slightly different subnetworks. This inference can be approximated as sampling from the posterior distribution of the model to obtain... One prediction result:

[0119]

[0120] in, Indicates the first The output obtained from the inference is usually a probability distribution vector (such as the probability distribution of 8 types of events).

[0121] Then, the prediction variance, as a measure of uncertainty, is specifically calculated... The variance of each prediction result is used as a measure of model uncertainty.

[0122]

[0123] in, for The average of the prediction results, Let L2 norm be the squared value of the vector. The larger the prediction variance, the greater the difference in prediction results among different sub-networks for this case, and the higher the uncertainty of the model; the smaller the prediction variance, the more consistent the prediction results among different sub-networks, and the lower the uncertainty of the model.

[0124] To facilitate subsequent processing, the uncertainty metric is normalized to a uniform range, such as the [0, 1] interval:

[0125]

[0126] in, and These are the minimum and maximum values ​​of the uncertainty quantification index obtained statistically on the validation set. The closer the normalized uncertainty quantification index is to 1, the more uncertain the model is; the closer it is to 0, the more certain the model is.

[0127] Based on the model's assessment of the uncertainty level of the handled cases, a segmented weighting strategy is employed to obtain the predicted uncertainty. To adjust the learning value of the case studies:

[0128]

[0129] The design philosophy of this strategy is: for cases where the model is highly uncertain ( ), assigning a weighting factor of 2 to give it the highest learning priority; for cases with moderate uncertainty in the model ( ), and assign a weighting coefficient of 1.5; for cases where the model is relatively well-defined ( This maintains a standard 1.0 multiplier. This segmented weighting guides training resources toward the areas where the model most needs improvement.

[0130] Furthermore, this embodiment introduces an additional weighting mechanism for prediction errors. If the case being dealt with has a manually verified ground truth label, the model's prediction can be further evaluated to determine its accuracy, and additional weighting can be applied to cases with incorrect predictions.

[0131]

[0132] in, This represents the category with the highest predicted probability by the model. This represents the true label. Cases with incorrect predictions obviously require more model learning and correction, so they are given a weighting factor of 2.

[0133] Finally, by weighting and synthesizing the perceived value of diversity, predictive uncertainty, and predictive error, the final active learning value score of the handled cases is obtained:

[0134]

[0135] This scoring system comprehensively considers the intrinsic quality of the handled cases, their contribution to diversity, their value in filling knowledge gaps, and the model's learning needs, achieving multi-objective optimization of "high quality + high diversity + what the model needs most." When the active learning value score is greater than the preset score, for example... At that time (this threshold is adjusted according to actual needs), the handled cases are included in the high-value case pool, i.e., the sample library, and are given priority for model training.

[0136] It should be noted that for scenarios with extremely high real-time requirements, the Monte Carlo Dropout method requires multiple forward propagations (e.g., 10 times), resulting in significant computational overhead. A lightweight uncertainty estimation method can be adopted: using the entropy value of the softmax output probability distribution obtained from a single forward propagation as a surrogate index for uncertainty.

[0137] The specific formula is as follows:

[0138]

[0139] in, For the model to predict the first The probability of a class The entropy value represents the total number of categories. A larger entropy value indicates that the model's prediction is less certain (the probabilities of each category are relatively even); a smaller entropy value indicates that the model's prediction is more certain (the probability of one category is dominant).

[0140] The advantage of this alternative is a computational speed increase of approximately 10 times, requiring only a single forward propagation to complete uncertainty estimation. The disadvantage is a slight decrease in the accuracy of uncertainty estimation because a single inference cannot capture the randomness of the model. It is expected that under this approach, the sample efficiency improvement brought by active learning may decrease from 42% to around 30%, but it is still superior to existing passive learning methods.

[0141] In this embodiment, for the high-value disposal cases selected, the system needs to automatically generate standardized samples that can be used for model training and perform quality checks to ensure the usability of the samples.

[0142] First, the system automatically extracts multimodal information from the entire event handling log, extracting three types of key information:

[0143] Environmental information: extract weather conditions (sunny, rainy, foggy, etc.) from meteorological sensor data, extract time period information (daytime, nighttime) from timestamps, and extract traffic density levels (low, medium, high) from vehicle detector data.

[0144] Event characteristics: Extract the event type (8 standard classifications), the precise location of the event (station number, lane number), and the severity level of the event (minor, moderate, severe) from the event confirmation record.

[0145] Handling process: Extract the event identification method (automatic detection or manual reporting), the monitor's confirmation result, the type and quantity of resources allocated, and the specific sequence of handling steps from the operation log.

[0146] Then, the automatic construction of the model input is performed, which mainly includes the following three steps:

[0147] Video frame sequence extraction: Extract a 60-second video segment (30 seconds before and 30 seconds after) before and after the event, sampling at a rate of 2 frames per second, resulting in 120 frames. Adjust all frames to the input size required by the model (e.g., ...). (pixels), and then normalize them (pixel values ​​are scaled to the [0,1] range).

[0148] Sensor data extraction: Extract sensor data synchronized with video time, including traffic flow, speed, and occupancy data recorded by vehicle detectors, and temperature, humidity, and visibility data recorded by meteorological sensors.

[0149] Context encoding: Encodes the timestamp (accurate to the second) and location information (station number, lane number) of the event and converts them into a numerical vector form acceptable to the model.

[0150] Ultimately, the model's input is a multimodal data packet: In addition, automatic generation of training labels is performed, including the following three types of labels:

[0151] Event Category Label: Based on the confirmed event type, an 8-dimensional one-hot encoded vector is generated. For example, if the event type is "traffic accident" (assuming it corresponds to category 3), the label would be... .

[0152] Location coordinate labels: The event's location on the highway, including its station number and lane position, is mapped to pixel coordinates in the video frame using a camera view transformation matrix, generating bounding box annotations (format: ).

[0153] Severity level labels: Based on the scope of the event's impact (congestion length, number of lanes) and duration, the severity is divided into three levels (minor, moderate, severe), generating corresponding category labels.

[0154] Finally, an automatic quality check is performed on the generated training samples. The system performs three automatic quality checks:

[0155] Video integrity check: Check the video frame sequence for missing frames, severe occlusion, or blur (by calculating image sharpness metrics such as gradient magnitude).

[0156] Label consistency check: A pre-trained object detection model is used to perform preliminary identification of the video to determine whether the automatically generated labels match the video content. For example, if the label is "abnormal parking," but the detection model does not find any stationary vehicles in the video, there may be a label error.

[0157] Time validity check: Verify the logical consistency of each timestamp (e.g.) This avoids time chaos caused by log recording errors.

[0158] The pass rate for quality checks is approximately 93% in practical applications. For samples that fail the check (approximately 7%), the system marks them and places them in a manual review queue for manual verification and correction by operations and maintenance personnel to ensure sample quality. In this way, high-value disposal cases added to the sample library can be processed into standardized samples suitable for model training.

[0159] Building upon the above, the core innovation of this invention also includes the establishment of a training benefit prediction model. By predicting the expected performance improvement of this training session and weighing it against training costs, the optimal training time is scientifically determined, avoiding blind training with low returns and missing out on high-return opportunities. The training timing decision-making process based on expected returns is as follows: Figure 5 As shown.

[0160] Specifically, in this embodiment, the benefit-cost ratio of model training based on newly added samples in the sample library is evaluated. If the benefit-cost ratio is greater than a dynamic threshold, model training is initiated (i.e., training of the model agent is performed). The benefit-cost ratio can be obtained in the following ways:

[0161] The current state features are obtained based on the newly added samples in the sample library and the training effect of model training based on the newly added samples. The predicted performance improvement value is obtained based on the current state features. The computation time cost, performance risk cost and system load cost of model training are estimated. The total cost is obtained by combining the computation time cost, performance risk cost and system load cost. The benefit-cost ratio is calculated based on the predicted performance improvement value and the total cost.

[0162] In this embodiment, a meta-learning model for profit prediction is trained. The task of this model is to predict: given the current system state and the situation of new samples, assuming that incremental training is performed at this time, how much performance improvement is expected to be obtained.

[0163] The input features of a meta-learning model include four aspects:

[0164] Current model performance A performance vector is composed of multiple performance metrics, including the model's accuracy, recall, and F1 score on the validation set.

[0165] New sample quality For newly added high-value samples, calculate their active learning value score. The statistical characteristics, including the mean, maximum, minimum, and standard deviation, reflect the overall quality level of the new samples.

[0166] New sample diversity : Calculate the diversity gain of new samples The statistical characteristics reflect the degree of complementarity between the new samples and existing knowledge.

[0167] Historical training effects Record historical data from the last 5 incremental training iterations, including performance before training, performance after training, performance improvement, and training time, to form a historical feature vector.

[0168] The above four features are concatenated to form the input feature vector of the meta-learning model. .

[0169] The meta-learning model adopts the XGBoost regression model (an ensemble learning algorithm based on gradient boosting decision trees), which has good expressive and generalization capabilities and can effectively handle nonlinear relationships.

[0170] The training data for the meta-learning model comes from the system's historical training records. Each time the system completes an incremental training iteration, it records the state features from before training. and actual performance improvement after training As a training sample As the system runs for a period of time, such training samples are continuously accumulated. These historical data are then used periodically (e.g., after every 5 incremental training iterations) to retrain the meta-learning model, enabling it to gradually learn the mapping relationship between "state → reward".

[0171] When the system needs to determine whether to start training, it extracts the current state features. The input is then fed into the trained meta-learning model to obtain the expected improvement in prediction performance:

[0172]

[0173] This predicted performance improvement value indicates how many percentage points the model's F1 score on the validation set is expected to improve if incremental training is performed at this point.

[0174] Furthermore, this invention provides a multi-dimensional cost estimation model for training, encompassing not only explicit computational resource consumption but also implicit costs such as system load and performance risks.

[0175] Calculation time cost Estimate the GPU time (in hours) required to complete this incremental training based on the number of new samples, model size, and hardware configuration. The estimation formula is as follows:

[0176]

[0177] in, To increase the number of samples, The number of parameters to be trained (LoRA adaptation layer parameters). These are the coefficients obtained by fitting based on historical training data.

[0178] Performance risk cost Assess the risk of performance degradation in the new model. Based on historical training records, calculate the failure rate where model performance actually declines after training:

[0179]

[0180] in, This represents the number of times performance has degraded in the last 10 training sessions. Multiply the failure rate by 10 to obtain the risk cost (the unit of measurement should be consistent with the time cost).

[0181] System load cost Assess the additional load on the system during canary releases. Canary releases require running both the old and new versions of the model simultaneously, increasing system resource consumption. Estimate the load increment to be 20% of the normal load, lasting approximately 48 hours, and translate this into load costs.

[0182] The total cost is calculated as follows:

[0183]

[0184] in, , , The relative weights of the three cost categories are given. This weighting reflects the emphasis on performance risk (highest weight), because a failed training session not only wastes computational resources but may also affect the stable operation of the system.

[0185] Based on this, a benefit-cost ratio decision and dynamic threshold adjustment are implemented. The benefit-cost ratio calculation formula is as follows:

[0186]

[0187] BCR (Benefit-Cost Ratio) measures the performance improvement gained per unit cost. A higher BCR indicates a higher cost-effectiveness of training; a lower BCR indicates a lower cost-effectiveness of training.

[0188] Set a dynamically adjustable threshold. Only when Incremental training only begins when the time is right. Otherwise, samples continue to accumulate, waiting for a more suitable training opportunity.

[0189] The dynamic threshold can be adaptively adjusted based on the recent training results:

[0190]

[0191] in: This is the baseline threshold (meaning that at least a 5% improvement in F1 score must be achieved per unit of cost). For the most recent The number of training sessions that yielded poor results (defined as training sessions where the actual benefit was more than 50% lower than the expected benefit). This is the adjustment coefficient.

[0192] The logic behind this dynamic adjustment strategy is as follows: if recent training results are frequently below expectations (high failure rate), then the dynamic threshold is increased to start training more cautiously and avoid frequent ineffective training; if recent training results are in line with expectations (low failure rate), then the dynamic threshold is kept low to encourage more active learning.

[0193] Based on the above mechanism, it can be determined when to start model training. In this embodiment, if model training is started, during the model training process, that is, during the training process of the model agent, a total loss function is constructed that includes the training loss of new samples, the hierarchical EWC constraint loss, and the replay loss of historical samples. The model is trained under the guidance of the total loss function until the preset requirements are met.

[0194] The core innovation of this invention also includes the establishment of a hierarchical forgetting control mechanism. Unlike traditional methods that employ a uniform protection strategy for all old knowledge, this method strengthens the protection of key foundational capabilities based on the importance level of the task, while allowing for moderate forgetting of lower-priority capabilities, thus achieving more refined knowledge management. The hierarchical forgetting control mechanism is as follows: Figure 6 As shown.

[0195] In this embodiment, the handling of highway incidents involves multiple levels of tasks. During model training, the hierarchical EWC constraint loss is constructed in the following way:

[0196] The system constructs a mapping relationship between multiple parameters of the disposal model agent and multiple hierarchical tasks, and assigns each parameter to the corresponding hierarchical task based on the mapping relationship; sets different constraint strength coefficients according to the hierarchical task to which each parameter belongs; calculates the contribution of each parameter to each hierarchical task in historical cases; and constructs a hierarchical EWC constraint loss based on the contribution of each parameter and the constraint strength coefficient.

[0197] In this embodiment, the multiple levels of tasks in handling cases are divided into four levels of tasks according to the level of abstraction from the bottom to the top:

[0198] Layer L0 (Basic Visual Feature Extraction): The most basic capabilities, including the extraction of fundamental visual features such as edge detection, texture recognition, and color recognition. These capabilities are the foundation of all higher-level tasks and must not be overlooked.

[0199] Layer L1 (General Object Detection): Based on basic features, it identifies general objects in the video, such as vehicles, pedestrians, and foreign objects. This is a prerequisite for event recognition and is of great importance.

[0200] Layer 2 (Common Event Recognition): Identifies high-frequency common events such as abnormal parking and minor rear-end collisions. These events account for more than 80% of the total events and are the core function of the system, requiring key protection.

[0201] Layer L3 (Long-tail event recognition): Identifies low-frequency long-tail events such as hazardous chemical leaks and chain collisions. Although important, it can tolerate a larger range of parameter adjustments due to the scarcity of samples.

[0202] In order to apply differentiated protection to tasks at different levels, it is necessary to establish a mapping relationship between model parameters and tasks at different levels, that is, to identify which level of task each parameter contributes the most to.

[0203] For each parameter among the multiple parameters of the processing model agent, such as each parameter of the LoRA adaptation layer. In the datasets of tasks at each level ( Calculate its gradient contribution on ).

[0204]

[0205] in, For the first Loss function for hierarchical tasks (cross-entropy loss). The gradient of the loss function with respect to the parameters. Take the absolute value. This represents the expected value (average value) on the dataset.

[0206] The meaning of this formula is: the larger the absolute value of the gradient, the greater the influence of the parameter on the task at that level, and the higher its contribution; when the gradient is close to 0, it means that the parameter has almost no influence on the task at that level.

[0207] Then, the parameter allocation strategy is executed for each parameter. Calculate its contribution to tasks at each level. Assign parameters to the task at the highest level of contribution:

[0208]

[0209] in, Indicates parameters The task level to which it belongs. In this way, a mapping table is established from all LoRA adaptation layer parameters to the task level.

[0210] Elastic Weight Consolidation (EWC) is a classic method for preventing forgetting, the core idea of ​​which is to impose constraints on changes in important parameters. This invention introduces a hierarchical constraint mechanism based on standard EWC.

[0211] First, Fisher information is calculated for each parameter. Calculate its Fisher information content to measure the importance of this parameter to older tasks (historical cases):

[0212]

[0213] in, For datasets of old tasks (historical cases), The predicted probability distribution of the model, The expected value of the squared gradient is the Fisher information content. The larger the Fisher information content, the greater the contribution of this parameter to the performance of old tasks, and the less it should be changed.

[0214] Next, set the hierarchical constraint strength, and set different EWC constraint strengths according to the hierarchical task to which the parameter belongs:

[0215]

[0216] The lower the level (the more basic), the stronger the constraint and the greater the penalty for parameter changes; the higher the level, the weaker the constraint and the more room for parameter adjustment.

[0217] Based on this, the hierarchical EWC constraint loss is constructed as follows:

[0218]

[0219] in, These are the parameter values ​​after training the old task was completed. This represents the constraint strength of the task at which the parameter belongs. This loss term penalizes the degree to which the parameter deviates from its old value; the strength of the penalty depends on the importance of the parameter (…). ) and hierarchical importance ( (To be decided jointly)

[0220] It should be noted that if gradient analysis of parameter-task mapping cannot be performed in real time (this analysis requires forward-backward propagation on task datasets at each level, which has a large computational cost), a predefined fixed weight scheme can be used: roughly divide the task levels according to the depth of the network layers. For example, the first 25% of parameters of the LoRA adaptation layer can be assigned to the L0 layer (basic features), 25%-50% to the L1 layer (general detection), 50%-75% to the L2 layer (regular events), and 75%-100% to the L3 layer (long-tail events). Then, the hierarchical EWC constraint can be applied directly.

[0221] The advantage of this alternative is that it avoids the computational overhead of dynamic gradient analysis and is simple to implement. The disadvantage is that the parameter-task mapping accuracy is not as good as the main approach, and some parameters may be assigned to the wrong levels. It is expected that under this approach, the forgetting rate control effect may slightly increase from 3% to 4%-5%, but it is still significantly better than the 10%-15% of existing technologies.

[0222] In addition to EWC constraints, this invention also introduces a selective experience replay mechanism, which further enhances the preservation of old knowledge by mixing new and historical samples during training. In this embodiment, the replay loss of historical samples is constructed during model training in the following way:

[0223] Evaluate the performance changes of each task level during model training; calculate the replay sampling weights of each task level based on the importance weights and performance changes of each task level; sample from historical samples according to the replay sampling weights of each task level, and construct the replay loss of historical samples based on the sampled historical samples.

[0224] First, hierarchical forgetting risk monitoring is performed, and the performance changes of each level of task are evaluated periodically on the validation set during incremental training:

[0225]

[0226] in, For the model before training, in the first Performance (such as accuracy) across different levels of tasks. This represents the model's performance on this level of task during the current training process. If This indicates that the task at this level is experiencing performance degradation and is at risk of being forgotten.

[0227] Then, the hierarchical replay weight calculation is performed, and the replay sampling weight of each hierarchical task is calculated based on the importance of the hierarchical task and the risk of forgetting:

[0228]

[0229] in, The importance weight of the hierarchy is set to , , , . The performance degradation is converted to a positive value (0 if the performance improves or remains unchanged). The greater the performance degradation, the greater the weight.

[0230] Finally, a sample replay strategy is implemented. In each training batch, samples are taken from historical samples according to the replay sampling weights, and then the new samples and the sampled historical samples are mixed in a 7:3 ratio. The historical samples are sampled according to hierarchical weights, as follows:

[0231]

[0232] Important levels that are at risk of being forgotten will receive more replay opportunities to strengthen memory through repeated training and prevent forgetting.

[0233] Based on the above, the total loss function is constructed by combining the training loss of newly added samples, the hierarchical EWC constraint loss, and the replay loss of historical samples:

[0234]

[0235] in: The cross-entropy loss for new samples drives the model to learn new knowledge; To consolidate the loss with tiered elastic weights, the variation of key parameters is constrained; To compensate for the loss of historical samples during playback, and to strengthen the memory of old knowledge; This is the weighting coefficient for the playback loss.

[0236] By optimizing the total loss function, the model is protected by both EWC and experience replay while learning new knowledge, effectively preventing catastrophic forgetting and achieving a good balance between new and old knowledge.

[0237] It should be noted that in application scenarios where system stability is extremely critical and frequent training is unacceptable, a hybrid triggering mode can be adopted: retaining the weekly routine training (similar to the periodic triggering in existing technologies), while adding an emergency triggering mechanism based on expected benefits. When model performance drops sharply (e.g., accuracy drops by more than 10% in a short period of time) or when a major new event occurs, emergency training can be initiated immediately after assessing the benefit-cost ratio.

[0238] The advantages of this alternative are more controllable and stable training frequency, smaller system load fluctuations, and suitability for critical business scenarios with extremely high stability requirements. The disadvantages are a slightly slower response speed to performance changes compared to the purely dynamic triggering scheme, and a slight decrease in the timeliness of knowledge solidification. It is expected that under this scheme, the average response cycle may extend from 3-5 days in the purely dynamic scheme to 5-7 days, but this is still better than the 30-day cycle of existing technologies.

[0239] After incremental training is completed, a rigorous verification and release process is required to ensure the quality and stability of the new model version. Therefore, the continuous learning method provided in this embodiment also includes a step of verifying the trained model agent, which includes:

[0240] The trained disposal model agent is evaluated using a test set based on multiple key metrics. If each key metric meets the set requirements, the offline verification is considered successful. The disposal model agent is then deployed to the test environment, and replay tests are performed using historical samples. The performance during the replay tests is monitored. If the performance meets the set requirements, the internal test is considered successful. The trained disposal model agent and the disposal model agents before and after training are used to perform tests. Multiple key metrics of the disposal model agents before and after training are compared. If the key metrics of the trained disposal model agent are better than those of the disposal model agent before training, the gray-scale test is considered successful.

[0241] This invention employs a five-stage incremental release strategy to reduce deployment risks. The model validation and canary release process is as follows: Figure 7 Show.

[0242] 1. Offline verification stage:

[0243] The new disposition model agent was evaluated on a standard test set (containing 200 manually labeled cases covering 8 event categories) based on five key metrics: Recognition accuracy: required Event recall rate: Required F1 score: required Current online version F1 score; inference latency: requirements seconds; Forgetting rate: required (Performance loss on the old task test set).

[0244] Only when all five indicators meet the requirements can the new disposal model agent pass offline verification and enter the next stage; if any indicator fails to meet the requirements, the training strategy is adjusted (such as adjusting the learning rate, increasing the replay ratio, strengthening EWC constraints, etc.) and retraining is performed.

[0245] 2. Internal testing phase:

[0246] Deploy the new version of the disposal model agent in an isolated test environment and perform replay testing using historical data, running it continuously for 24 hours. During the test, monitor the system's functional correctness (whether all interfaces are working properly), inference stability (whether there are crashes or anomalies), and resource consumption (whether GPU and memory usage are within reasonable ranges). If the test passes, proceed to the canary release phase.

[0247] 3. Small-scale gray-scale phase:

[0248] Switch 10% of online traffic to the new model version, while retaining the old version for 90% of traffic. Monitor the comparison of key metrics between the old and new versions of the handling model agent in real time, including accuracy, recall, false positive rate, and inference latency. Observe for 24 hours. If all metrics of the new version of the handling model agent are no worse than the old version (allowing for slight fluctuations, such as accuracy fluctuations not exceeding 2%), and the error rate is below 0.5%, then the small-scale rollout is considered successful, and the model proceeds to the next stage.

[0249] 4. Medium-volume grayscale stage:

[0250] Increase the grayscale ratio to 50%, meaning the old and new versions of the processing model agent will each handle 50% of the traffic. Continue monitoring for 24 hours, focusing on the stability of the new version under concurrent scenarios. If no anomalies are found, proceed to the full release phase.

[0251] 5. Full release phase:

[0252] Switch 100% of traffic to the new version of the handling model agent, and discontinue the old version. After the full release, the system continuously monitors model performance, establishes a long-term performance tracking mechanism, compiles performance metrics for various events daily, draws performance trend charts, and promptly detects signs of performance degradation.

[0253] An automatic rollback mechanism is configured so that at any stage of the canary release, if a key metric of the new version's handling model agent is detected to be significantly worse than the old version (e.g., an accuracy drop of more than 5%) or the error rate exceeds 0.5%, the automatic rollback mechanism is immediately triggered, switching all traffic back to the old version within 5 minutes to avoid impacting business operations. Simultaneously, an alert notification is generated, allowing technical personnel to analyze the cause of the failure, make adjustments, and attempt to release again.

[0254] In summary, the adaptive continuous learning method for the highway incident handling model agent provided in this embodiment mainly includes the following improvements:

[0255] 1. A value assessment method based on case diversity perception and knowledge coverage blind spot identification:

[0256] This method is the primary innovation of this invention, overcoming the limitations of existing technologies that assess case value in isolation. By using multimodal embedding representations of case features and cosine similarity calculation, it quantifies the degree of difference between cases and existing knowledge, calculating diversity gain. Through K-means clustering and cluster density analysis of the case space, it identifies blind spots in knowledge coverage, assigning additional weight to cases falling into these blind spots, guiding the system to prioritize filling knowledge gaps. This method evaluates the incremental contribution of each case to existing knowledge from the perspective of the global knowledge system, avoiding repeated learning of similar cases and achieving system optimization of knowledge diversity. Compared to existing technologies, this method can improve scene coverage by more than 30% and reduce knowledge blind spots by more than 25% with the same number of samples.

[0257] 2. An active sample selection method based on model uncertainty quantification:

[0258] This method represents the second core innovation of this invention, achieving a paradigm shift from passive to active learning. Employing Monte Carlo Dropout technology, it estimates the uncertainty (prediction variance) of model predictions through multiple random inferences, identifying scenarios where the model is "uncertain." A segmented weighting strategy is used based on the level of uncertainty (2x for high uncertainty, 1.5x for medium uncertainty, and 1x for low uncertainty), prioritizing cases where the model most needs to learn. Cases with incorrect predictions are given an additional 2x weighting, strengthening the focus on failure cases. This method actively selects training samples from the perspective of "learning what the model needs," significantly improving sample utilization efficiency. Theoretically, it is expected that, with the same number of training rounds, the recall rate for long-tail events can be increased by more than 40% compared to existing technologies.

[0259] 3. Training timing optimization method based on expected return prediction:

[0260] This method represents the third core innovation of this invention, addressing the problem of blindly selecting training timing in existing technologies. It constructs a meta-learning model (XGBoost regression model), inputting features such as current model performance, the quality and diversity of new samples, and historical training results to predict the expected performance improvement in the current training session. It quantifies the multi-dimensional costs of training (computation time, performance risk, system load); calculates the benefit-cost ratio (BCR), and only initiates training when the BCR exceeds a dynamically adjusted threshold, avoiding blind training with low returns; it dynamically adjusts the trigger threshold based on recent training results, increasing the threshold to reduce ineffective training when the failure rate is high and decreasing it to encourage active learning when the failure rate is low. This method achieves optimal allocation of training resources, and theoretical analysis shows that the number of training iterations can be reduced by 30%-40%, while training efficiency can be improved by more than 50%.

[0261] 4. Hierarchical forgetting control and selective experience replay methods:

[0262] This method represents the fourth core innovation of this invention, achieving more refined knowledge management. It defines a four-layer task hierarchy (basic features → general detection → routine events → long-tail events), constructs a parameter-task mapping through gradient analysis, and identifies the contribution of each parameter to tasks at different levels. Differentiated EWC constraint strengths are applied to different levels (L0 level 10000, L1 level 5000, L2 level 2000, L3 level 1000), with strong protection for key basic capabilities and allow for moderate forgetting of lower priority capabilities. The forgetting risk of each level is monitored in real time, and replay sampling weights are calculated based on task importance and forgetting risk for selective experience replay, focusing on strengthening levels with forgetting risk. This method achieves a refined balance between new and old knowledge, with the overall forgetting rate controlled within 3%, significantly better than the 10%-15% of existing technologies.

[0263] 5. Automatic training sample generation and quality assurance mechanism:

[0264] This mechanism automates the conversion from cases to training samples. It automatically extracts multimodal information (environment, event, and handling process) from incident logs, automatically constructs model input (video frame sequences + sensor data + environmental context), automatically generates annotation labels (event category + location coordinates + severity level), and performs automatic quality checks (completeness + consistency + reasonableness). Processing speed is reduced from 2-3 hours manually to less than 1 minute automatically, achieving an automation rate of 93%. This mechanism overcomes the bottleneck of manual annotation and forms the technological foundation for a significant improvement in knowledge accumulation efficiency.

[0265] 6. Gradual canary release and automatic rollback mechanism:

[0266] This mechanism ensures the safe deployment of the new version model. It employs a five-stage release process: offline verification (all five metrics meet the standards) → internal testing (24-hour historical data playback) → 10% small-scale gray-scale deployment (24-hour observation) → 50% medium-scale gray-scale deployment (24-hour observation) → 100% full release. Real-time monitoring compares the performance of the new and old versions; if the new version's metrics significantly deteriorate or the error rate exceeds the limit, it automatically rolls back to the old version within 5 minutes. This mechanism reduces deployment risks, ensures system stability, and provides reliable engineering support for continuous learning.

[0267] Based on the above improvements, the continuous learning method provided in this embodiment has at least the following beneficial effects:

[0268] 1. A dual improvement in the efficiency and quality of knowledge acquisition:

[0269] This invention significantly improves the efficiency and quality of knowledge acquisition by establishing a multi-dimensional value assessment and automatic sample generation mechanism. Regarding knowledge retention efficiency, the processing time for a single case is reduced from 2-3 hours using existing manual methods to less than 1 minute using automated methods, representing an efficiency improvement of over 100 times. Within the same timeframe, the knowledge retention rate can be increased from 4% using existing methods to over 80%, an improvement of 76 percentage points.

[0270] Regarding knowledge quality, by introducing case diversity gain calculation and knowledge coverage blind spot identification mechanisms, the system can prioritize cases that complement existing knowledge, avoiding repeated learning of similar cases. Theoretical analysis shows that, with the same number of training samples, the case library's coverage of the scene space can be improved by more than 30%, and knowledge blind spots can be reduced by more than 25%. This dual improvement of "high efficiency + high quality" enables the system to establish a more comprehensive and balanced knowledge system under limited resource conditions.

[0271] Existing technologies are severely limited by manual processing capabilities, allowing only a small number of cases to be handled. Furthermore, the selection criteria are often subjective and arbitrary, easily leading to a large number of similar cases and knowledge gaps in the case library. This invention overcomes the bottleneck of manual processing by automating information extraction, value scoring, and sample generation processes. Through case similarity calculation and cluster analysis, it optimizes the diversity distribution of samples from a global perspective, achieving a systematic improvement in the efficiency and quality of knowledge acquisition.

[0272] 2. Significantly enhanced ability to handle long-tail events:

[0273] This invention utilizes an active learning mechanism driven by model uncertainty to accurately identify the long-tail scenarios that the model most needs to learn from and prioritize their training. For low-frequency, high-risk events such as hazardous chemical leaks and chain collisions, the system automatically detects the high uncertainty of the model in these scenarios and assigns them a 2x learning weight; simultaneously, a scarcity weighting mechanism further enhances the priority of these long-tail events. Theoretically, with the same number of training rounds, the recall rate for long-tail events can be increased by more than 40% compared to existing technologies.

[0274] For example, assuming the initial recall rate of the model for long-tail events is 42%, using the existing passive learning method, the recall rate may increase to 50% (an increase of 8 percentage points) after 5 incremental training iterations; while using the active learning method of this invention, the recall rate can be increased to about 64% (an increase of 22 percentage points) after the same number of training iterations, which is 2.75 times that of the existing technology.

[0275] Existing technologies employ a passive learning approach, treating all cases equally and failing to identify "what the model most needs to learn." With a limited sample size, long-tail events are often marginalized due to their small number, failing to receive sufficient learning opportunities. This invention estimates the model's prediction uncertainty using Monte Carlo Dropout, allocating training resources to the model's weakest points, achieving "precise learning," and significantly improving sample utilization efficiency and the ability to handle long-tail events.

[0276] 3. Coordinated optimization of training efficiency and model generalization performance:

[0277] This invention achieves an optimal balance between training frequency and training benefits by establishing a training timing optimization mechanism based on expected returns. The meta-learning model can predict the expected performance improvement of the current training session and weigh it against training costs, initiating training only when the benefit-cost ratio meets the target. Theoretical analysis shows that, under the premise of achieving the same performance improvement goal, the number of training sessions can be reduced by 30%-40%, and the utilization efficiency of training resources can be improved by more than 50%.

[0278] Meanwhile, by avoiding low-return blind training, the model's generalization performance is also improved. Theoretically, after continuous learning, the model's generalization performance in unseen scenarios can be improved by 20%-30% compared to the initial version, and its robustness to environmental changes (such as changes in weather, lighting, and traffic patterns) is significantly enhanced. For example, the model's accuracy in rainy night scenarios may improve from the initial 58% to over 75%, and its accuracy in complex lighting scenarios at tunnel entrances and exits may improve from 55% to over 70%.

[0279] Existing technologies use fixed threshold triggering, which cannot determine the optimal training timing. This may lead to training starting even when sample quality is low and expected returns are low, resulting in wasted resources; conversely, it may delay triggering at the optimal training time, causing performance to remain low for extended periods. This invention uses a meta-learning model to predict training returns, achieving intelligent scheduling that "acts at the right time." Furthermore, it optimizes system learning through case diversity, covering a wider range of scenarios. These two approaches work synergistically to achieve dual optimization of training efficiency and generalization performance.

[0280] 4. Ensuring anti-forgetting properties and the stability of the knowledge system:

[0281] This invention effectively retains old knowledge while learning new tasks through hierarchical forgetting control and selective experience replay mechanisms. The strongest EWC constraints are applied to key fundamental capabilities such as basic visual features (L0 layer) and general object detection (L1 layer) to ensure these capabilities are never forgotten; moderate constraints are applied to routine event recognition (L2 layer); and a larger parameter adjustment space is allowed for long-tail event recognition (L3 layer). Theoretically, after learning a new task, the forgetting rate of basic visual features and routine event recognition capabilities can be controlled to within 1%, and the overall forgetting rate can be controlled to within 3%, significantly better than the 10%-15% forgetting rate of existing technologies.

[0282] For example, assuming the model initially has an accuracy of 87% in recognizing routine events (such as abnormal parking), after five incremental training iterations using existing technology, the accuracy may drop to around 75% (forgetting rate of about 14%) due to catastrophic forgetting. However, using the hierarchical forgetting control mechanism of this invention, the accuracy only drops to about 86% (forgetting rate of about 1.2%) after the same training, thus fully ensuring the stability of the knowledge system.

[0283] Existing technologies employ a uniform protection strategy for all existing knowledge, failing to consider the varying importance of different tasks. With limited parameter resources, uniform protection often leads to insufficient protection and the forgetting of key capabilities, or excessive protection and hindered learning of new knowledge. This invention identifies the contribution of each parameter to different levels of tasks through parameter-task mapping, implementing differentiated constraint strengths: strong protection for key capabilities and moderate adjustment allowed for secondary capabilities; and dynamically adjusting the replay ratio of each level through selective experience replay, focusing on strengthening levels at risk of forgetting. This refined knowledge management mechanism achieves an optimal balance between new and old knowledge, ensuring the long-term stability of the knowledge system.

[0284] 5. Improvement in overall system response efficiency and handling quality:

[0285] Due to the synergistic effect of the above four aspects, the model's recognition capability continues to improve, the accuracy and reliability of automatic event detection are constantly enhanced, the frequency of human intervention is gradually reduced, and the overall response efficiency and handling quality of the system are significantly improved. Theoretically, the end-to-end latency from event discovery to the generation of a handling plan can be shortened from 8-12 minutes in the current technology to less than 2 minutes, improving response efficiency by more than 75%.

[0286] In terms of handling quality, the model's ability to identify long-tail events has been significantly enhanced, and its generalization performance in complex scenarios has been greatly improved, resulting in a significant reduction in both false positive and false negative rates. Theoretically, the handling time for simple faults can be reduced from 25 minutes to approximately 12 minutes (an improvement of about 50%), the handling time for general incidents from 45 minutes to approximately 28 minutes (an improvement of about 40%), and the handling time for complex incidents from 90 minutes to approximately 62 minutes (an improvement of about 30%).

[0287] Existing technologies suffer from fixed model capabilities or low learning efficiency, leading to long-term reliance on manual judgment and intervention, with response time primarily consumed in manual processes. This invention, through a continuous learning mechanism, continuously improves model capabilities, resulting in a sustained increase in the proportion and accuracy of automated processing. This gradually reduces reliance on manual intervention, achieving a systematic improvement in response efficiency and processing quality.

[0288] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interface; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0289] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0290] Furthermore, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0291] It should be noted that if the functionality is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0292] The above are merely embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive continuous learning method for an intelligent agent in a highway incident handling model, characterized in that, The method includes: Extract highway incident handling cases from the highway incident process logs; Diversity gain is obtained based on the similarity between the treatment case and existing cases, and a blind spot filling weighting coefficient is obtained by cluster analysis of existing cases. The perceived value of diversity is obtained by combining the diversity gain, the blind spot filling weighting coefficient and the basic score of treatment effect. Evaluate the prediction results of the disposal case executed by the disposal model agent, and determine the prediction uncertainty of the disposal model agent for the disposal case based on the prediction results; The active learning value score is obtained by combining the perceived value of diversity and the uncertainty of prediction. Cases with active learning value scores greater than the preset score are added to the sample library. Evaluate the benefit-cost ratio of model training based on newly added samples in the sample library. If the benefit-cost ratio is greater than a dynamic threshold, then start model training. During model training, a total loss function is constructed that includes the training loss of new samples, the hierarchical EWC constraint loss, and the replay loss of historical samples. The model is trained under the guidance of the total loss function until the preset requirements are met. The disposal case has multiple levels of tasks, and the hierarchical EWC constraint loss is constructed in the following way: A mapping relationship is constructed between multiple parameters of the disposal model agent and multiple hierarchical tasks. Based on the mapping relationship, each parameter is assigned to the corresponding hierarchical task. Different constraint strength coefficients are set according to the hierarchical task to which each parameter belongs. The contribution of each parameter to each hierarchical task in historical cases is calculated. Based on the contribution of each parameter and the constraint strength coefficient, a hierarchical EWC constraint loss is constructed. The playback loss of the historical samples is constructed in the following way: Evaluate the performance changes of each level of tasks during model training; calculate the replay sampling weights of each level of tasks based on the importance weights of each level of tasks and the performance changes; sample from historical samples according to the replay sampling weights of each level of tasks, and construct the replay loss of historical samples based on the sampled historical samples.

2. The adaptive continuous learning method for the intelligent agent of the highway event handling model according to claim 1, characterized in that, The basic score for the treatment effect was obtained through the following methods: The actual response time is obtained based on the detection time and handling time of the aforementioned cases. A response timeliness score is obtained by combining the actual response time, the standard response time, and the maximum response time. The actual processing time of the case is obtained, and the standard processing time of historical cases is obtained. A processing efficiency score is obtained based on the actual processing time and the standard processing time. The total number of resources actually allocated in the aforementioned cases is statistically analyzed, and the optimal resource allocation number in the expert knowledge base is obtained. Based on the total number of resources and the optimal resource allocation number, a resource rationality score is obtained. The congestion length and duration caused by the aforementioned treatment case are obtained, and an impact control score is obtained based on the congestion length and duration and the set threshold parameters. The actual handling process of the aforementioned cases is matched with the standard handling process template, and a process compliance score is obtained based on the matching results. By combining the response timeliness score, handling efficiency score, resource rationality score, impact control score, and process standardization score, a basic score for handling effectiveness is obtained.

3. The adaptive continuous learning method for the intelligent agent of the highway event handling model according to claim 1, characterized in that, The step of obtaining diversity gain based on the similarity between the treatment case and existing cases includes: The case information of the disposed cases is encoded into a feature vector; Calculate the cosine similarity between the feature vector of the treatment case and the feature vector of existing cases; Based on the cosine similarity, determine the existing case that is most similar to the treatment case; The diversity gain is calculated based on the cosine similarity between the treatment case and the most similar existing case, wherein the diversity gain is negatively correlated with the cosine similarity.

4. The adaptive continuous learning method for the intelligent agent of the highway event handling model according to claim 1, characterized in that, The step of obtaining the blind spot filling weighting coefficient by performing cluster analysis on existing cases includes: Cluster analysis is performed on existing cases to divide them into multiple clusters, and the cluster density of each cluster is calculated. Determine the target cluster to which the treatment case is assigned, and obtain the cluster density of the target cluster; The blind spot filling weighting coefficient for the treatment case is set according to the cluster density of the target cluster, wherein the blind spot filling weighting coefficient is negatively correlated with the cluster density.

5. The adaptive continuous learning method for the intelligent agent of the highway event handling model according to claim 1, characterized in that, The step of obtaining the perceived value of diversity by combining the diversity gain, the blind spot filling weighting coefficient, and the basic score of treatment effect includes: The preliminary perceived value of diversity is obtained by combining the aforementioned diversity gain, blind spot filling weighting coefficient, and treatment effect basic score; If the case being handled is a long-tail event, the occurrence frequency of the long-tail event is obtained, and the perceptual value of diversity is calculated based on the occurrence frequency, the average frequency, and the preliminary perceived value of diversity.

6. The adaptive continuous learning method for the intelligent agent of the highway event handling model according to claim 1, characterized in that, The step of evaluating the predicted results of the disposal model agent's execution of the disposal case, and determining the prediction uncertainty of the disposal model agent for the disposal case based on the predicted results, includes: Input the case information of the treatment case into the existing model and output multiple prediction results; Calculate the variance of the multiple prediction results to obtain an uncertainty measure index; The uncertainty metric is normalized to a uniform range; The predicted uncertainty is obtained based on the normalized uncertainty metric and a segmented weighting strategy.

7. The adaptive continuous learning method for the intelligent agent of the highway event handling model according to claim 1, characterized in that, The step of evaluating the cost-benefit ratio of model training based on newly added samples in the sample library includes: The current state features are obtained based on the newly added samples in the sample library and the training effect of model training based on the newly added samples, and the prediction performance improvement value is obtained based on the current state features. Estimate the computation time cost, performance risk cost, and system load cost for model training, and combine the computation time cost, performance risk cost, and system load cost to obtain the total cost; The benefit-cost ratio is calculated based on the predicted performance improvement and the total cost.

8. The adaptive continuous learning method for the intelligent agent of the highway event handling model according to any one of claims 1-7, characterized in that, The method further includes a step of validating the trained disposal model agent, which includes: The trained processing model agent is evaluated using a test set based on multiple key metrics. If each of the key metrics meets the set requirements, the offline verification is deemed successful. The aforementioned processing model agent is deployed to the test environment, and historical samples are used for replay testing. The performance during the replay test is monitored. If the performance meets the set requirements, the internal test is deemed to have passed. The test is performed using the trained disposal model agent and the disposal model agent before and after training. Several key indicators of the disposal model agent before and after training are compared. If the key indicators of the trained disposal model agent are better than those of the disposal model agent before training, the grayscale test is considered to have passed.

Images