Behavior recognition model adaptive learning method under dynamic delivery data monitoring

Abstract

The invention 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, aligning the delivery features, and mapping the delivery features of different dimensions to a unified semantic space; based on the feature space, designing a feature weight dynamic allocation strategy driven by an attention mechanism; on the basis of the fusion features, using an incremental mutual information screening mechanism to realize feature space online expansion; dynamically monitoring the distribution offset of the features in the feature space by using JS divergence, and triggering updating of the behavior recognition model when the distribution offset of the features reaches a set condition; and judging whether the delivery behavior is abnormal or not based on the updated behavior recognition model and the feature set. According to the invention, through a cooperation mechanism of dynamic feature alignment, distribution drift monitoring and meta-learning parameter updating, accurate identification and real-time response of abnormal delivery behaviors are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an adaptive learning method for a behavior recognition model under dynamic delivery data monitoring. Background Art

[0002] In recent years, with the intelligent transformation of the express delivery industry, express delivery safety risk monitoring faces the dual challenges of adapting to dynamic environments and maximizing data utilization efficiency. On the one hand, the differences in the spatiotemporal characteristics of multi-source heterogeneous data make it difficult for traditional models to accurately identify risks. 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 positives or false negatives. Current model adaptive learning technologies have significant shortcomings in terms of adaptability to dynamic environments and data stability. Specifically, the following issues exist:

[0003] First, dynamic adaptability is limited. Current adaptive learning technologies, such as those based on collaborative filtering and rules, rely too heavily on static associations with historical data. When faced with data sparseness, they are unable to effectively establish a dynamic mapping between new features and knowledge networks, resulting in decision paths that deviate from actual needs.

[0004] Second, model updates lack stability. When data quality fluctuates, adaptive learning methods based on fixed-interval updates lack a mechanism to monitor changes in feature distribution, directly triggering high-frequency model iterations. This disordered update process exacerbates performance fluctuations.

[0005] Third, there is a lag in capturing spatiotemporal evolution patterns. Existing evaluation methods rely solely on a single behavioral dimension and fail to establish multidimensional correlations across spatiotemporal neighborhoods, making it impossible to analyze the evolution of delivery behavior patterns in real time.

[0006] Fourth, the framework has weak adaptability to unevenly distributed interaction data. Due to its static sampling strategy, it is unable to achieve feature alignment through semantic space reconstruction when faced with unevenly distributed interaction data, resulting in a decrease in anomaly recognition accuracy. Summary of the Invention

[0007] To address the aforementioned technical issues, the present invention provides an adaptive learning method for behavior recognition models under dynamic delivery data monitoring. This method constructs a continuously iterative behavior recognition model through dynamic feature alignment, distribution drift monitoring, and meta-learning parameter updates. This framework integrates spatiotemporal features based on a multimodal association network. Combining dynamic weight allocation with an online feature screening mechanism, it captures data evolution patterns in real time and triggers model parameter self-adjustment through spatiotemporal neighborhood analysis. This method maintains discrimination stability despite fluctuating data quality, achieving a dynamic balance between precise interception of abnormal behavior and risk prevention and control.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A method for adaptive learning of a behavior recognition model under dynamic delivery data monitoring, comprising:

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

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

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

[0013] Based on the fused features, an incremental mutual information screening mechanism is used to achieve online expansion of the feature space. This involves calculating the mutual information between the newly added features and the target variable within the sliding window, assessing the importance of the newly added features in real time, and updating the feature set and the feature space. The target variable is the label of the abnormal behavior.

[0014] Use JS divergence to dynamically monitor the distribution offset of features in the feature space. When the distribution offset of features reaches the set conditions, trigger the update of the behavior recognition model: Use meta-learning to dynamically update the behavior recognition model.

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

[0016] In one embodiment, the multi-dimensional delivery features of the delivery behavior are collected to form a feature set, specifically including:

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

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

[0019] Construct the cross-modal mapping matrix W by optimizing the objective To align the delivery features, represents the i-th source modal feature, represents the jth target modal feature; λ||W|| F is a regularization term used to control the complexity of the behavior recognition model, and λ is a regularization coefficient used to control the penalty intensity of the regularization term; ||·|| F is the F-norm.

[0020] In one embodiment, the multi-projection of the source modality features and the target modality features is performed respectively, the association weights between different features in the feature space are dynamically adjusted, and the fusion features are output, specifically including:

[0021] Source modal characteristics and target modal features Perform multi-head projection to generate query vector Q, key vector K, and value vector V, and use the multi-head attention mechanism to calculate the cross-modal association weight:

[0022]

[0023] Among them, d is the dimension of the feature in the feature space, α ij is the source modal feature and target modal features The association weight between them. Output fusion features

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

[0025] In 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 newly added 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 of the target variable Y when the newly added 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 is the feature set at time t.

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

[0032] Initialize the feature base distribution to Q0.

[0033] Update the feature base distribution at time t to:

[0034] Q t =(1-ρ)Q t +ρP t ;

[0035] Q t represents the feature space distribution state referenced by the behavior recognition model at the tth moment, P t represents the latest feature sample distribution collected at time t, and ρ represents the degree of trust in the latest feature sample distribution when the reference distribution is updated;

[0036] Based on the updated feature reference distribution, the Gaussian mixture model is used to fit the current feature distribution P t , through JS divergence D JS (·) Quantify the degree of feature distribution deviation D JS (P t |Q t ):

[0037]

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

[0039] If D JS >θ t , then the behavior recognition model is updated; θ t is the offset threshold.

[0040] In one embodiment, the offset threshold θ t According to the degree of feature distribution deviation, the degree of feature distribution deviation D quantified by the historical JS divergence is JS (P t |Q t ) mean Standard deviation And 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) of the delivery behavior in the spatiotemporal neighborhood:

[0044]

[0045] Combined with the PID controller to dynamically adjust the threshold coefficient m, generate an adaptive alarm threshold τ t :τ t =μ t-1 +mσ t-1;μ t-1 represents the mean of the abnormal score A(x) at time t-1, σ t-1 represents the variance of the abnormal score A(x) at time t-1;

[0046] When the anomaly score A(x) of the delivery behavior exceeds the adaptive alarm threshold τ t , it will be considered that the delivery behavior is abnormal, thereby starting the update of 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 behaviors through the collaborative mechanism of dynamic feature alignment, distribution drift monitoring, and meta-learning parameter updating, significantly improving delivery security protection capabilities and risk management efficiency. Specifically, this invention has the following technical effects:

[0049] First, through the cross-modal dynamic alignment mechanism and incremental feature expansion technology, the modal differences of multi-source data are eliminated and burst features are incorporated in real time, which is conducive to providing high-precision and high-consistency feature representation for anomaly detection.

[0050] Second, based on adaptive distribution offset quantification and dynamic threshold control, it can perceive feature distribution changes in real time and balance detection sensitivity and false alarm rate, which is conducive to improving 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 the rapid iteration of model parameters through a two-layer optimization strategy, it is beneficial to improve the response efficiency and decision-making accuracy of sudden risk scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0053] Figure 2 This is a technical roadmap for the adaptive learning framework for the behavior recognition model in the embodiment of the present invention;

[0054] Figure 3 This is a technical roadmap for the dynamic feature alignment mechanism in an embodiment of the present invention. DETAILED DESCRIPTION

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

[0056] Based on the analysis of data evolution characteristics and model degradation mechanism in dynamic delivery monitoring scenarios, the technical goal of this invention is to build an adaptive learning framework with cross-modal dynamic association and real-time self-optimization capabilities, such as Figure 2 This framework solves the problem of feature drift in unbalanced data environments by establishing a multimodal semantic mapping mechanism and a distribution change perception system, and achieves dynamic decoupling of model performance and data quality fluctuations.

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

[0058] 1. Design a dynamic feature alignment mechanism to construct a cross-modal mapping matrix based on the spatiotemporal feature differences of multi-source heterogeneous data, solving the semantic alignment problem of 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 behavior recognition model adaptive learning framework improves the risk identification accuracy and stability of the prediction model during iterative evolution through dynamic feature alignment, distribution drift monitoring, and meta-learning parameter updates, thereby improving the overall security protection level and operational efficiency of delivery. The framework integrates spatiotemporal feature flows, constructs an adaptive feature space, and realizes continuous model iteration, such as Figure 1 As shown, the specific steps include:

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

[0062] S2, constructs a cross-modal mapping matrix, aligns the delivery features, maps the delivery features of different dimensions to a unified semantic space, and obtains the feature space.

[0063] S3, based on the feature space, designs a dynamic feature weight allocation strategy driven by the attention mechanism: perform multi-head projection on the source modality features and the target modality features respectively, dynamically adjust the association weights between different features in the feature space, and output fusion features.

[0064] S4, based on fused features, uses an incremental mutual information screening mechanism to achieve online expansion of the feature space: calculate the mutual information between the newly added features and the target variable in 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 is the label of abnormal behavior.

[0065] S5 uses JS divergence to dynamically monitor the distribution offset of features in the feature space. When the distribution offset of the features reaches the set conditions, 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 sequences of the delivery features specifically includes:

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

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

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

[0071] Construct the cross-modal mapping matrix W by optimizing the objective To align the delivery features, represents the i-th source modal feature, represents the jth target modal feature; λ||W|| F is a regularization term used to control the complexity of the behavior recognition model, ||·|| F is the F-norm.

[0072] The source modal features and the target modal features are respectively delivery features of one dimension, or features extracted or projected from the delivery features.

[0073] In one embodiment, the multi-projection of the source modality features and the target modality features is performed respectively, the association weights between different features are dynamically adjusted, and the fusion features are output, specifically including:

[0074] Source modal characteristics and target modal features Perform multi-head projection to generate query vector Q, key vector K, and value vector V, and use the multi-head attention mechanism to calculate the cross-modal association weight:

[0075]

[0076] Among them, d is the feature dimension, αij is the source modal feature and target modal features The association weight between them. Output fusion features

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

[0078] In 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 is the feature set at time t.

[0083] New features are those derived from new data during model execution, or generated through fusion, processing, or transformation. Features here refer to vectors within the feature space. New features are derived 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 offset of features in the feature space and triggering the update of the behavior recognition model when the distribution offset of the features reaches a set condition specifically includes:

[0085] Initialize the feature base distribution to Q0.

[0086] Update the feature base distribution at time t to:

[0087] Q t =(1-ρ)Q t +ρP t ;

[0088] Based on the updated feature reference distribution, the Gaussian mixture model is used to fit the current feature distribution P t , through JS divergence D JS (·) Quantify the degree of feature distribution deviation D JS (P t |Q t ):

[0089]

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

[0091] If D JS >θ t , then the behavior recognition model is updated; θ t is the offset threshold.

[0092] In one embodiment, the offset threshold θ t According to the degree of feature distribution deviation, combined with the historical JS divergence mean μD JS , standard deviation σD JS And 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; set 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) of the delivery behavior in the spatiotemporal neighborhood:

[0096]

[0097] Combined with the PID controller to dynamically adjust the threshold coefficient m, generate an adaptive alarm threshold τ 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 , it will be considered that the delivery behavior is abnormal, thereby starting the update of 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 that it can automatically adjust its recognition strategy to new environmental characteristics, promptly detecting and responding to anomalous delivery behavior. Based on alarm results, manually verified anomalous data is injected into the training set, triggering an incremental learning process and forming a closed-loop detection-feedback-optimization chain. Anomalous data refers to delivery behavior data that is identified as anomalous by the model and confirmed to be anomalous by manual review. For example, if the package shape, packaging method, or labeling method significantly deviates from standard procedures, the model will automatically adjust its recognition strategy to meet the requirements of the delivery model.

[0101] In this invention, the behavior recognition model updates with two parallel trigger conditions, each monitoring system status from different perspectives to ensure the behavior recognition model can adapt promptly to the dynamically changing delivery environment. Trigger condition one is feature distribution shift, based on statistical changes in the feature spatial distribution; trigger condition two is the frequent occurrence of abnormal behavior, which directly monitors the anomaly score at the behavioral level.

[0102] The update to the behavior recognition model primarily involves updating the feature extraction network within the model. As a core component of the behavior recognition model, the feature extraction network performs key functions such as multimodal feature fusion, semantic representation extraction, and generation of representations for identifying abnormal behaviors. It runs throughout the entire model training and inference process. During the model update phase, the parameters of the feature extraction network are optimized using a meta-learning strategy on the neighborhood construction task to enhance adaptive learning capabilities. Specifically, the feature extraction network is involved in the following aspects: after cross-modal mapping and feature alignment, it connects the transitive features within the unified semantic space to extract deep semantic representations. In the multi-head attention mechanism, the feature extraction network implements the multi-head projection structure, generating query, key, and value vectors to calculate association weights. During the abnormal behavior identification phase, the model uses the representations output by the feature extraction network for behavior recognition and scoring. During the model update process, the feature extraction network is updated through gradient optimization based on task samples constructed in spatiotemporal neighborhoods to improve the model's adaptability to new abnormal behaviors. The feature extraction network serves as a bridge between input features and behavior recognition decisions throughout the entire method, forming the foundation for the dynamic adaptive learning of the behavior recognition model.

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

[0104] By constructing a multimodal semantic mapping network and fusing multi-source heterogeneous features, semantic alignment and dynamic mapping can be achieved in data-sparse scenarios.

[0105] Based on the distribution deviation quantification of adaptive benchmarks, a delivery behavior anomaly quantification algorithm is designed to capture dynamic evolution patterns and trigger model updates.

[0106] Combined with meta-learning methods, real-time incremental optimization of parameters is achieved, promoting rapid model iteration.

[0107] The present invention integrates the aforementioned mechanisms into an end-to-end adaptive learning framework that can dynamically adapt to data sparsity, uneven distribution, and spatiotemporal evolution patterns in dynamic environments. This framework addresses the limitations of traditional methods in dynamic scenarios and improves model stability and adaptability. Through this innovative adaptive framework, the present invention is expected to achieve breakthroughs in dynamic environment adaptability and data stability, providing a new solution for model adaptive learning in postal delivery scenarios.

[0108] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps 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 multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times. The order of execution of these steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0109] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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 embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. It is intended that all variations within the meaning and range of equivalents of the claims be embraced herein, and any reference signs in the claims should not be construed as limiting the claims to which they relate.

[0111] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A behavior recognition model adaptive learning method under dynamic delivery data monitoring, characterized by: include: Collect multi-dimensional delivery features in delivery behavior to form a feature set; the delivery features include regional risk features, item status features, and delivery behavior features; Construct a cross-modal mapping matrix, align the delivery features, and map the delivery features of different dimensions into a unified semantic space to obtain the feature space; Based on the feature space, a dynamic feature weight allocation strategy driven by an attention mechanism is designed: multi-head projection is performed on the source modality features and the target modality features respectively, the association weights between different features in the feature space are dynamically adjusted, and the fused features are output; Based on the fusion features, an incremental mutual information screening mechanism is used to achieve online expansion of the feature space: the mutual information between the newly added features and the target variable in 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; Use JS divergence to dynamically monitor the distribution offset of features in the feature space. When the distribution offset of features reaches the set conditions, trigger the update of the behavior recognition model: Use meta-learning to dynamically update the behavior recognition model. Based on the updated behavior recognition model and feature set, determine whether the delivery behavior is abnormal.

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

3. The method for adaptive learning of a behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that: The cross-modal mapping matrix is ​​constructed to align delivery features and map delivery features of different dimensions into a unified semantic space, specifically including: Construct the cross-modal mapping matrix W by optimizing the objective To align the delivery features, represents the i-th source modal feature, represents the jth target modal feature; λ||W|| F is a regularization term used to control the complexity of the behavior recognition model, and λ is a regularization coefficient used to control the penalty intensity of the regularization term; ||·|| F is the F-norm.

4. The method for adaptive learning of a behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that: The method of performing multi-head projection on the source modality features and the target modality features, dynamically adjusting the association weights between different features in the feature space, and outputting fusion features specifically includes: Source modal characteristics and target modal features Perform multi-head projection to generate query vector Q, key vector K, and value vector V, and use the multi-head attention mechanism to calculate the cross-modal association weight: Among them, d is the dimension of the feature in the feature space, α ij is the source modal feature and target modal features The association weight between them; output fusion features 5. The method for adaptive learning of a behavior recognition model under dynamic delivery data monitoring according to claim 1, characterized in that: The calculation of the mutual information between the newly added features and the target variable in 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: In the sliding window, calculate the mutual information between the newly added feature Δf and the target variable Y: I(Δf;Y)=H(Y)-H(Y∣Δf); I(Δf; Y) represents the degree of information sharing between the newly added 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 of the target variable Y when the newly added feature Δf is known. When I(Δf; Y)>β, add Δf to the feature set: F t =F t-1 ∪{Δf|I(Δf)>β}, β is the set threshold; F t is the feature set at time t.

6. The method for adaptive learning of a 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 offset of features in the feature space and triggering the update of the behavior recognition model when the distribution offset of the features reaches a set condition specifically includes: Initialize the feature base distribution to Q0; Update the feature base distribution at time t to: Q t =(1-ρ)Q t +ρP t ; Q t represents the feature space distribution state referenced by the behavior recognition model at the tth moment, P t represents the latest feature sample distribution collected at time t, and ρ represents the degree of trust in the latest feature sample distribution when the reference distribution is updated; Based on the updated feature reference distribution, the Gaussian mixture model is used to fit the current feature distribution P t , through JS divergence D JS (·) Quantify the degree of feature distribution deviation D JS (P t |Q t ): D KL (·) represents KL divergence; If D JS >θ t , then the behavior recognition model is updated; θ t is the offset threshold.

7. The method for adaptive learning of a behavior recognition model under dynamic delivery data monitoring according to claim 6, characterized in that: The offset threshold θ t According to the degree of feature distribution deviation, the degree of feature distribution deviation D quantified by the historical JS divergence is JS (P t |Q t ) mean Standard deviation And characteristic entropy is determined.

8. The method for adaptive learning of a 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 the spatiotemporal neighborhood N(x) of the delivery behavior x; Calculate the Euclidean distance anomaly score A(x) of the delivery behavior in the spatiotemporal neighborhood: Combined with the PID controller to dynamically adjust the threshold coefficient m, generate an adaptive alarm threshold τ t :τ t =μ t-1 +mσ t-1 ;μ t-1 represents the mean of the abnormal score A(x) at time t-1, σ t-1 represents the variance of the abnormal score A(x) at time t-1; When the anomaly score x(x) of the delivery behavior exceeds the adaptive alarm threshold τ t , it will be considered that the delivery behavior is abnormal, thereby triggering the update of 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