A behavior recognition model adaptive learning method under dynamic delivery data monitoring

Abstract

The application relates to the technical field of artificial intelligence, and discloses a behavior recognition model adaptive learning method under dynamic delivery data monitoring, which comprises the following steps: collecting multi-dimensional delivery features in a delivery scene to form a feature set; constructing a cross-modal mapping matrix to align the delivery features and map the delivery features of different dimensions to a unified semantic space; designing a feature weight dynamic distribution strategy driven by an attention mechanism based on the feature space; using an incremental mutual information screening mechanism to realize online expansion of the feature space based on the fused features; using JS divergence to dynamically monitor the distribution deviation of the features in the feature space; when the distribution deviation of the features reaches a set condition, triggering updating of the behavior recognition model; and judging whether the delivery behavior is abnormal based on the updated behavior recognition model and the feature set. Through the cooperative mechanism of dynamic feature alignment, distribution deviation monitoring and meta-learning parameter updating, the application realizes accurate identification and real-time response of abnormal delivery behavior.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an adaptive learning method for behavior recognition models under dynamic delivery data monitoring. Background Technology

[0002] In recent years, with the intelligent transformation of the express delivery industry, express delivery security risk monitoring faces the dual challenges of adapting to dynamic environments and improving data utilization efficiency. On the one hand, the spatiotemporal differences in multi-source heterogeneous data make it difficult for traditional models to achieve accurate risk identification; on the other hand, the data distribution in express delivery scenarios fluctuates frequently with environmental changes, and static models are prone to performance degradation due to feature evolution, leading to false alarms or missed alarms. Current adaptive learning technology for models has significant deficiencies in terms of dynamic environment adaptability and data stability. Specifically, the main problems are as follows:

[0003] First, the dynamic adaptability is limited. Current adaptive learning techniques based on collaborative filtering and rule-based adaptive learning rely excessively on the static association of historical data. When encountering data sparse scenarios, they cannot effectively establish a dynamic mapping between new features and knowledge networks, causing the generated decision paths to deviate from actual needs.

[0004] Second, the model update stability is insufficient. When data quality fluctuates, adaptive learning methods based on fixed time interval updates lack a monitoring mechanism for changes in feature distribution, directly triggering high-frequency model iterations. This disordered update exacerbates performance oscillations.

[0005] Third, the capture of spatiotemporal evolution patterns is lagging. Existing assessment methods rely only on a single behavioral dimension and fail to establish multidimensional relationships within the spatiotemporal neighborhood, resulting in the inability to analyze the evolution of delivery behavior patterns in real time.

[0006] Fourth, it has poor adaptability to imbalanced data. Due to the use of a static sampling strategy, the framework cannot achieve feature alignment through semantic space reconstruction when faced with imbalanced interactive data, resulting in a decrease in anomaly recognition accuracy. Summary of the Invention

[0007] To address the aforementioned technical challenges, this invention provides an adaptive learning method for behavior recognition models under dynamic postal data monitoring. This method constructs a continuously iterative behavior recognition model through dynamic feature alignment, distribution drift monitoring, and meta-learning parameter updates. The framework integrates spatiotemporal features based on a multimodal association network, combining dynamic weight allocation and online feature filtering mechanisms to capture data evolution patterns in real time. Furthermore, it triggers model parameter self-adjustment through spatiotemporal neighborhood analysis, thereby maintaining discrimination stability even under fluctuating data quality conditions and achieving a dynamic balance between accurate interception of abnormal behavior and risk prevention.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] An adaptive learning method for behavior recognition models under dynamic delivery data monitoring includes:

[0010] Collect multi-dimensional delivery characteristics from delivery behavior to form a feature set; the delivery characteristics include regional risk characteristics, item status characteristics, and delivery behavior characteristics;

[0011] Construct a cross-modal mapping matrix, align delivery features, and map delivery features of different dimensions to a unified semantic space to obtain the feature space;

[0012] Based on the feature space, an attention mechanism-driven dynamic feature weight allocation strategy is designed: multi-head projection is performed on the source modality features and the target modality features respectively, the correlation weights between different features in the feature space are dynamically adjusted, and the fused features are output.

[0013] Based on fusion features, an incremental mutual information filtering mechanism is used to achieve online expansion of the feature space: the mutual information between the newly added features and the target variable within the sliding window is calculated, the importance of the newly added features is evaluated in real time, and the feature set and the feature space are updated; the target variable refers to the label of abnormal behavior.

[0014] The behavior recognition model is updated dynamically using JS divergence to monitor the distribution shift of features in the feature space. When the distribution shift of features reaches a set condition, the behavior recognition model is updated dynamically using meta-learning.

[0015] Based on the updated behavior recognition model and feature set, it is determined whether the delivery behavior is abnormal.

[0016] In one embodiment, the collection of multi-dimensional delivery features in the delivery behavior forms a feature set, specifically including:

[0017] A dynamic feature expansion algorithm is used to align the sequences of postal features with different sampling rates to eliminate time axis bias.

[0018] In one embodiment, the construction of the cross-modal mapping matrix, aligning delivery features, and mapping delivery features of different dimensions to a unified semantic space specifically includes:

[0019] Construct a cross-modal mapping matrix W, and optimize the objective. To align delivery characteristics, Represents the i-th source mode feature. λ||W|| represents the j-th target modal feature; F λ is the regularization term used to control the complexity of the behavior recognition model, and λ is the regularization coefficient used to control the penalty strength of the regularization term; F It is the F-norm.

[0020] In one embodiment, the step of performing multi-head projection on the source modality features and the target modality features respectively, dynamically adjusting the correlation weights between different features in the feature space, and outputting fused features specifically includes:

[0021] source modal features and target modal features Perform multi-head projection to generate query vector Q, key vector K, and value vector V, and calculate cross-modal association weights using a multi-head attention mechanism.

[0022]

[0023] Where d is the dimension of the feature in the feature space, and α ij Source modal features and target modal features The correlation weights between them. Output fused features.

[0024] In one embodiment, the calculation of the mutual information between newly added features and the target variable within the sliding window, the real-time evaluation of the importance of the newly added features, and the updating of the feature set and the feature space specifically include:

[0025] Within the sliding window, calculate the mutual information between the newly added feature Δf and the target variable Y:

[0026] I(Δf;Y)=H(Y)-H(Y∣Δf);

[0027] I(Δf; Y) represents the degree of information sharing between the new feature Δf and the target variable Y, H(Y) represents the uncertainty of the target variable Y itself, and H(Y|Δf) represents the uncertainty that still exists in the target variable Y even when the new feature Δf is known.

[0028] When I(Δf; Y) > β, add Δf to the feature set:

[0029] F t =F t-1 ∪{Δf|I(Δf)>β},

[0030] β is the set threshold; F t Let be the feature set at time t.

[0031] In one embodiment, the method of using JS divergence to dynamically monitor the distribution shift of features in the feature space, and triggering an update of the behavior recognition model when the distribution shift of features reaches a set condition, specifically includes:

[0032] The characteristic baseline distribution is initialized to Q0.

[0033] Update the feature baseline distribution at time t as follows:

[0034] Q t =(1-ρ)Q t +ρP t ;

[0035] Q t P represents the feature space distribution state referenced by the behavior recognition model at time t. t Let represent the latest feature sample distribution collected at time t, and ρ represent the degree of confidence in the latest feature sample distribution when updating the reference distribution;

[0036] Based on the updated feature baseline distribution, a Gaussian mixture model is used to fit the current feature distribution P. t Through JS divergence D JS (·) Quantitative feature distribution offset degree D JS (P t |Q t ):

[0037]

[0038] D KL (·) represents the KL divergence;

[0039] If D JS >θ t This triggers an update to the behavior recognition model; θ t This is the offset threshold.

[0040] In one embodiment, the offset threshold θ t Based on the degree of feature distribution shift, combined with the feature distribution shift quantified by historical JS divergence, D... JS (P t |Q t mean Standard deviation And the characteristic entropy is determined.

[0041] In one embodiment, dynamically updating the behavior recognition model using meta-learning specifically includes:

[0042] Define the spatiotemporal neighborhood N(x) of the delivery behavior x;

[0043] Calculate the Euclidean distance anomaly score A(x) for delivery behavior within the spatiotemporal neighborhood:

[0044]

[0045] By combining the PID controller to dynamically adjust the threshold coefficient m, an adaptive alarm threshold τ is generated. t τ t =μ t-1 +mσ t-1μ t-1 Let σ represent the mean of the anomaly scores A(x) at time t-1. t-1 Let A(x) represent the variance of the outlier scores at time t-1.

[0046] When the anomaly score A(x) of the delivery behavior exceeds the adaptive alarm threshold τ t The system will consider the delivery behavior to be abnormal and thus initiate an update to the behavior recognition model.

[0047] After triggering the update of the behavior recognition model, a task is constructed based on the neighborhood delivery behavior data of the delivery behavior, and the parameters of the behavior recognition model are updated through gradient optimization.

[0048] This invention achieves accurate identification and real-time response to abnormal delivery behavior through a collaborative mechanism of dynamic feature alignment, distribution drift monitoring, and meta-learning parameter updates, significantly improving delivery security and risk management efficiency. Specifically, this invention offers the following key technical advantages:

[0049] First, by using a cross-modal dynamic alignment mechanism and incremental feature expansion technology, modal differences in multi-source data can be eliminated and burst features can be incorporated in real time, which is beneficial for providing high-precision and highly consistent feature representations for anomaly detection.

[0050] Second, based on adaptive distribution offset quantization and dynamic threshold adjustment, the model can perceive changes in feature distribution in real time, balance detection sensitivity and false alarm rate, and improve the long-term stability and reliability of the model in complex delivery scenarios.

[0051] Third, by combining historical drift patterns with online feedback data to construct meta-tasks, and driving rapid iteration of model parameters through a two-layer optimization strategy, it is beneficial to improve response efficiency and decision-making accuracy in sudden risk scenarios. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0053] Figure 2 This is a roadmap of the adaptive learning framework technology for behavior recognition models in this embodiment of the invention.

[0054] Figure 3 This is a technical roadmap for the dynamic feature alignment mechanism in this embodiment of the invention. Detailed Implementation

[0055] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0056] Based on the analysis of data evolution characteristics and model degradation mechanisms in dynamic delivery monitoring scenarios, the technical objective of this invention is to construct an adaptive learning framework with cross-modal dynamic correlation and real-time self-optimization capabilities, such as... Figure 2 As shown, this framework addresses the feature drift problem in imbalanced data environments by establishing a multimodal semantic mapping mechanism and a distribution change perception system, thereby achieving dynamic decoupling between model performance and data quality fluctuations.

[0057] Specifically, the core technical problems to be solved by this invention include:

[0058] 1. Design a dynamic feature alignment mechanism to construct a cross-modal mapping matrix under the spatiotemporal feature differences of multi-source heterogeneous data, and solve the semantic alignment problem between new features and knowledge networks in sparse scenarios;

[0059] 2. Design a collaborative mechanism for distribution drift monitoring to perceive changes in feature distribution in real time, balance detection sensitivity and false alarm rate, and improve the long-term stability and reliability of the model in complex delivery scenarios.

[0060] The adaptive learning framework for behavior recognition models improves risk assessment accuracy and stability through dynamic feature alignment, distribution drift monitoring, and meta-learning parameter updates, thereby enhancing the overall security and operational efficiency of the delivery service. The framework integrates spatiotemporal feature streams to construct an adaptive feature space, enabling continuous model iteration. Figure 1 As shown, the specific steps include:

[0061] S1, collect multi-dimensional delivery characteristics from delivery behavior to form a feature set; the delivery characteristics include regional risk characteristics, item status characteristics and delivery behavior characteristics.

[0062] S2. Construct a cross-modal mapping matrix, align the delivery features, and map the delivery features of different dimensions to a unified semantic space to obtain the feature space.

[0063] S3. Based on the feature space, design an attention mechanism-driven dynamic feature weight allocation strategy: perform multi-head projection on the source modality features and the target modality features respectively, dynamically adjust the correlation weights between different features in the feature space, and output fused features.

[0064] S4. Based on the fusion features, an incremental mutual information filtering mechanism is used to realize the online expansion of the feature space: calculate the mutual information between the newly added features and the target variable within the sliding window, evaluate the importance of the newly added features in real time, and update the feature set and the feature space; the target variable refers to the label of abnormal behavior.

[0065] S5 uses JS divergence to dynamically monitor the distribution shift of features in the feature space. When the distribution shift of features reaches a set condition, it triggers the update of the behavior recognition model: the behavior recognition model is dynamically updated using meta-learning.

[0066] S6, based on the updated behavior recognition model and feature set, determines whether the delivery behavior is abnormal.

[0067] In one embodiment, aligning the feature sequence composed of each delivery feature specifically includes:

[0068] A dynamic feature expansion algorithm is used to align feature sequences with different sampling rates to eliminate time axis bias.

[0069] The design of dynamic feature alignment mechanism, such as Figure 3 As shown, in order to achieve semantic consistency and dynamic evolution capabilities of multi-source heterogeneous data, a dynamic feature alignment mechanism is designed, which includes four main parts: cross-modal mapping, attention fusion, incremental expansion and distribution constraints. The stability and scalability of the feature space are ensured by joint optimization objectives and online verification.

[0070] In one embodiment, the construction of the cross-modal mapping matrix, aligning delivery features, and mapping delivery features of different dimensions to a unified semantic space specifically includes:

[0071] Construct a cross-modal mapping matrix W, and optimize the objective. To align delivery characteristics, Represents the i-th source mode feature. λ||W|| represents the j-th target modal feature; F The regularization term is used to control the complexity of the behavior recognition model. F It is the F-norm.

[0072] The source modality features and the target modality features are respectively one-dimensional postal features, or features extracted or projected from postal features.

[0073] In one embodiment, the step of performing multi-head projection on the source modality features and the target modality features respectively, dynamically adjusting the correlation weights between different features, and outputting fused features specifically includes:

[0074] source modal features and target modal features Perform multi-head projection to generate query vector Q, key vector K, and value vector V, and calculate cross-modal association weights using a multi-head attention mechanism.

[0075]

[0076] Where d is the feature dimension, αij Source modal features and target modal features The correlation weights between them. Output fused features.

[0077] In one embodiment, the calculation of the mutual information between newly added features and the target variable within the sliding window, the real-time evaluation of the importance of the newly added features, and the updating of the feature set and the feature space specifically include:

[0078] Within the sliding window, calculate the mutual information between the newly added feature Δf and the target variable Y:

[0079] I(Δf;Y)=H(Y)-H(Y∣Δf);

[0080] When I(Δf; Y) > β, add Δf to the feature set:

[0081] F t =F t-1 ∪{Δf|I(Δf)>β};

[0082] β is the set threshold; F t Let be the feature set at time t.

[0083] New features refer to features derived from new data during model operation, or new features generated through fusion, processing, or transformation. Here, features refer to vectors within the feature space. New features come partly from new data and partly from weighted combinations of multiple original features.

[0084] In one embodiment, the method of using JS divergence to dynamically monitor the distribution shift of features in the feature space, and triggering an update of the behavior recognition model when the distribution shift of features reaches a set condition, specifically includes:

[0085] The characteristic baseline distribution is initialized to Q0.

[0086] Update the feature baseline distribution at time t as follows:

[0087] Q t =(1-ρ)Q t +ρP t ;

[0088] Based on the updated feature baseline distribution, a Gaussian mixture model is used to fit the current feature distribution P. t Through JS divergence D JS (·) Quantitative feature distribution offset degree D JS (P t |Q t ):

[0089]

[0090] D KL (·) represents the KL divergence;

[0091] If D JS >θ t This triggers an update to the behavior recognition model; θ t This is the offset threshold.

[0092] In one embodiment, the offset threshold θ t Based on the degree of feature distribution shift, combined with the historical JS divergence mean μD JS Standard deviation σD JS And the characteristic entropy is determined.

[0093] In one embodiment, dynamically updating the behavior recognition model using meta-learning specifically includes:

[0094] Define the spatiotemporal neighborhood N(x) of the delivery behavior x; delineate the spatial radius r = 500m and the time window ΔT = 10min to provide a contextual benchmark for anomaly scoring.

[0095] Calculate the Euclidean distance anomaly score A(x) for delivery behavior within the spatiotemporal neighborhood:

[0096]

[0097] By combining the PID controller to dynamically adjust the threshold coefficient m, an adaptive alarm threshold τ is generated. t τ t =μ t-1 +mσ t-1 .

[0098] When the anomaly score A(x) of the delivery behavior exceeds the adaptive alarm threshold τ t The system will consider the delivery behavior to be abnormal and thus initiate an update to the behavior recognition model.

[0099] After triggering the update of the behavior recognition model, a task is constructed based on the neighborhood delivery behavior data of the delivery behavior, and the parameters of the behavior recognition model are updated through gradient optimization to improve the model's environmental adaptability.

[0100] By calculating anomaly scores and adjusting dynamic thresholds, the model ensures it can automatically adjust its identification strategy under new environmental features, promptly detecting and responding to abnormal delivery behaviors. Based on alarm results, manually verified anomaly data is injected into the training set, triggering an incremental learning process and forming a closed-loop chain of detection-feedback-optimization. Anomaly data refers to delivery behavior data that is identified as abnormal by the model and confirmed by manual verification, such as package shape, packaging method, and labeling method that significantly deviate from standard procedures.

[0101] In this invention, the behavior recognition model update has two parallel triggering conditions, which monitor the system status from different perspectives to ensure that the behavior recognition model can adapt to the dynamically changing delivery environment in a timely manner. The first triggering condition is feature distribution shift, which is based on the statistical changes in the feature space distribution; the second triggering condition is the frequency of abnormal behavior, which directly monitors abnormal scores at the behavior level.

[0102] The primary update to the behavior recognition model lies in updating the feature extraction network (FIN). As a core component of the behavior recognition model, the FIN undertakes crucial functions such as multimodal feature fusion, semantic representation extraction, and abnormal behavior discrimination representation generation, spanning the entire model training and inference process. During the model update phase, the FIN parameters are optimized on the neighborhood construction task through a meta-learning strategy to enhance adaptive learning capabilities. Specifically, the FIN participates in: after cross-modal mapping and feature alignment, it connects with post-feedback features within a unified semantic space to extract deep semantic representations; in the multi-head attention mechanism, it implements the multi-head projection structure, generating query, key, and value vectors to calculate association weights; in the abnormal behavior discrimination phase, the model performs behavior recognition and scoring based on the representation output by the FIN; and during the model update process, gradient optimization updates are performed on the FIN based on task samples constructed from the spatiotemporal neighborhood to improve the model's adaptability to novel abnormal behaviors. The FIN acts as a bridge connecting input features and behavior recognition decisions throughout the method, forming the foundation for the dynamic adaptive learning of the behavior recognition model.

[0103] Compared to existing methods that rely on static feature engineering and batch retraining, this invention proposes an adaptive learning framework for cross-modal dynamic association and incremental optimization. This framework systematically addresses issues such as semantic alignment, model update stability, and spatiotemporal evolution pattern capture in dynamic environments through the following key mechanisms:

[0104] By constructing a multimodal semantic mapping network and integrating heterogeneous features from multiple sources, semantic alignment and dynamic mapping can be achieved in data-sparse scenarios.

[0105] Based on adaptive baseline distribution offset quantization, an algorithm for quantifying delivery behavior anomalies is designed to capture dynamic evolution patterns and trigger model updates.

[0106] By combining meta-learning methods, real-time incremental optimization of parameters is achieved, driving rapid model iteration.

[0107] This invention integrates the aforementioned mechanisms into an end-to-end adaptive learning framework capable of dynamically adapting to data sparsity, imbalanced distribution, and spatiotemporal evolution patterns in dynamic environments. This framework addresses the limitations of traditional methods in dynamic scenarios, enhancing model stability and adaptability. Through this innovative adaptive framework, this invention aims to achieve breakthroughs in dynamic environment adaptability and data stability, providing a novel solution for adaptive learning in delivery scenarios.

[0108] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0109] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0111] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An adaptive learning method for behavior recognition models under dynamic delivery data monitoring, characterized in that, include: Collect multi-dimensional delivery characteristics from delivery behavior to form a feature set; the delivery characteristics include regional risk characteristics, item status characteristics, and delivery behavior characteristics; Construct a cross-modal mapping matrix, align delivery features, and map delivery features of different dimensions to a unified semantic space to obtain the feature space; Based on the feature space, an attention mechanism-driven dynamic feature weight allocation strategy is designed: multi-head projection is performed on the source modality features and the target modality features respectively, the correlation weights between different features in the feature space are dynamically adjusted, and the fused features are output. Based on fusion features, an incremental mutual information filtering mechanism is used to achieve online expansion of the feature space: calculate the mutual information between the newly added features and the target variable within the sliding window, evaluate the importance of the newly added features in real time, and update the feature set and the feature space. The target variable refers to the label of the abnormal behavior; The behavior recognition model is updated dynamically using JS divergence to monitor the distribution shift of features in the feature space. When the distribution shift of features reaches a set condition, the behavior recognition model is updated dynamically using meta-learning. Based on the updated behavior recognition model and feature set, it is determined whether the delivery behavior is abnormal.

2. The adaptive learning method for behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that, The collection of multi-dimensional delivery features in the delivery process constitutes a feature set, specifically including: A dynamic feature expansion algorithm is used to align the sequences of postal features with different sampling rates to eliminate time axis bias.

3. The adaptive learning method for behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that, The construction of the cross-modal mapping matrix, aligning delivery features and mapping delivery features of different dimensions to a unified semantic space, specifically includes: Constructing a cross-modal mapping matrix By optimizing the target To align delivery characteristics, Indicates the first Individual modal features, Indicates the first One target modal feature; This is a regularization term used to control the complexity of the behavior recognition model. is the regularization coefficient, used to control the intensity of the penalty for the regularization term; It is the F-norm.

4. The adaptive learning method for behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that, The process of performing multi-head projection on the source modality features and the target modality features, dynamically adjusting the correlation weights between different features in the feature space, and outputting fused features specifically includes: source modal features and target modal features Perform multi-head projection to generate query vector Q, key vector K, and value vector V, and calculate cross-modal association weights using a multi-head attention mechanism. Where d is the dimension of the feature in the feature space. Source modal features and target modal features The correlation weights between them; output fused features .

5. The adaptive learning method for behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that, The calculation of mutual information between newly added features and the target variable within the sliding window, the real-time evaluation of the importance of newly added features, and the updating of the feature set and the feature space specifically include: Calculate the new features within the sliding window. With target variable Mutual information: ； Indicates new features With target variable The degree of information sharing between them Represent the target variable The inherent uncertainty Indicates the addition of known features In this case, the target variable Uncertainties still exist; when At that time, Add feature set: To set a threshold; For the first The feature set of a given moment.

6. The adaptive learning method for behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that, The method of using JS divergence to dynamically monitor the distribution shift of features in the feature space, and triggering an update of the behavior recognition model when the distribution shift of features reaches a set condition, specifically includes: Initialize the characteristic baseline distribution to ; The first The feature baseline distribution at time step is updated as follows: ； The behavior recognition model is represented by the first The feature space distribution state referenced at any given time. Indicates the first The latest feature sample distribution collected at any time It indicates the degree of confidence in the latest feature sample distribution when the reference distribution is updated; Based on the updated feature baseline distribution, a Gaussian mixture model is used to fit the current feature sample distribution. Through JS divergence Quantitative feature sample distribution offset : ； Indicates KL divergence; like This triggers an update to the behavior recognition model; This is the offset threshold.

7. The adaptive learning method for behavior recognition model under dynamic delivery data monitoring according to claim 6, characterized in that, The offset threshold Based on the degree of feature distribution shift, combined with the feature distribution shift quantified by historical JS divergence. mean Standard deviation And the characteristic entropy is determined.

8. The adaptive learning method for behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that, The method of dynamically updating the behavior recognition model using meta-learning specifically includes: Define delivery behavior spatiotemporal neighborhood ; Calculate the Euclidean distance anomaly score for delivery behavior within the spatiotemporal neighborhood. : ； Dynamically adjust the threshold coefficient using a PID controller Generate adaptive alarm thresholds : ; Indicates the first Abnormal scores within a given time period mean Indicates the first Abnormal scores within a given time period variance; When the delivery behavior is abnormally scored Exceeding the adaptive alarm threshold The system will consider the delivery behavior to be abnormal, thereby triggering an update to the behavior recognition model. After triggering the update of the behavior recognition model, a task is constructed based on the neighborhood delivery behavior data of the delivery behavior, and the parameters of the behavior recognition model are updated through gradient optimization.

Images