Lightweight abnormal behavior recognition system and method based on edge calculation

Abstract

The invention relates to the technical field of behavior recognition, and discloses a lightweight abnormal behavior recognition system and method based on edge computing, and the method comprises the steps: obtaining multi-dimensional heterogeneous behavior observation data covering a target region through an edge side behavior collection node system, and constructing a behavior dynamic feature vector set which can be iteratively updated; performing structure self-adaptive decoupling processing on the behavior dynamic feature vector set, and recording the fluctuation convergence rate of an abnormal category in real time in a model iteration process; judging the behavior recognition stability in the model reasoning stage, and constructing a behavior transfer trajectory map in combination with the environmental transaction interference factors; extracting density disturbance parameters of the abnormal behaviors, and generating a behavior intervention evaluation result set; and performing hierarchical risk judgment on the current identification behavior result, and automatically generating an edge execution regulation and control instruction set. The method has the advantage of improving the efficiency.

Description

Technical Field

[0001] This invention relates to the field of behavior recognition technology, specifically to a lightweight abnormal behavior recognition system and method based on edge computing. Background Technology

[0002] With the increasing adoption of smart construction sites and smart industrial parks, video surveillance and behavior recognition technologies have been widely applied in areas such as monitoring construction worker behavior and providing early warnings for work safety. Traditional video behavior recognition systems often rely on transmitting collected video data to the cloud or a central server for processing and analysis. However, in scenarios like smart construction sites, due to limited edge device resources, the deployment of existing behavior recognition algorithms at the edge still faces challenges such as large model size, slow inference speed, and high energy consumption. Especially at construction sites, abnormal worker behaviors are often characterized by short-term suddenness and wide spatial distribution, such as not wearing safety ropes at heights, illegally crossing dangerous areas, or sudden falls during work. These behaviors are not only time-sensitive but also require the system to be able to identify them simultaneously at multiple points. However, existing technologies often focus on high-precision models or complex multimodal analysis, neglecting the importance of "lightweight models" and "multi-edge node collaboration" in real-world scenarios, making it difficult to achieve fast, distributed, and efficient abnormal behavior recognition. Therefore, it is essential to design a lightweight abnormal behavior recognition system and method based on edge computing to improve efficiency. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a lightweight abnormal behavior recognition system and method based on edge computing, which has the advantage of improved efficiency and solves the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of improving efficiency, this invention provides the following technical solution: a lightweight abnormal behavior recognition system and method based on edge computing, comprising the following steps:

[0005] By using the edge-side behavior acquisition node system, multi-dimensional heterogeneous behavior observation data covering the target area is acquired, and combined with the perturbation-aware coding mechanism and scene-dependent weight analysis, a set of dynamic behavior feature vectors that can be iteratively updated is constructed.

[0006] The structure is adaptively decoupled for the dynamic feature vector set of behavior. A behavior nested pattern model is constructed based on a lightweight multidimensional attention fusion network. An inference stability observation factor is introduced to record the fluctuation convergence rate of the anomaly category in real time during the model iteration process.

[0007] Based on the nonlinear perturbation trend of the convergence rate, the stability of behavior recognition in the model inference stage is judged. If it is determined to be stable, the key feature transfer segments of the current abnormal behavior are marked, and a behavior transfer trajectory map is constructed in combination with environmental anomaly interference factors.

[0008] Based on the behavior transfer trajectory map, a multi-scale differential recognition operator is applied to dynamically reorganize the edge behavior data sequence, extract the density perturbation parameters of abnormal behavior, and combine the sudden interference response mechanism to correct the recognition probability output, thereby generating a behavior intervention evaluation result set.

[0009] Based on the behavioral intervention assessment result set and combined with the preset behavioral abnormality threshold standard cluster, the current identified behavioral results are used to determine the hierarchical risk and automatically generate a set of edge execution control instructions.

[0010] Preferably, the process of constructing an iteratively updatable set of dynamic behavioral feature vectors is as follows:

[0011] By deploying multiple types of behavior sensing nodes at the edge, real-time multi-source behavior observation of people or equipment within the target area can be performed;

[0012] The disturbance-aware coding mechanism is used to dynamically label the timestamp alignment, scale normalization and disturbance sensitivity of various types of data sources to form an initial behavioral fragment representation.

[0013] By combining scene-dependent weight analysis, weight matrices are assigned to different behavioral segments according to scene categories, and similarity matching and label correction are performed based on a historical scene pattern library;

[0014] The encoded behavior fragments are jointly embedded into the vector space along with scene labels, and an incremental dynamic feature fusion mechanism is used to generate a set of dynamic behavior feature vectors that supports iterative updates.

[0015] Preferably, the adaptive decoupling process for the behavioral dynamic feature vector set is as follows:

[0016] Based on the time window and spatial domain partitioning rules, the behavioral dynamic feature vector set is initially classified to form multiple sets of structurally heterogeneous feature subsets.

[0017] An adaptive decoupling algorithm is introduced within each feature subset to identify the dominant dependency path between features based on a Bayesian network model and to perform suppression operations on weakly correlated dimensions.

[0018] The high-coupling dimension is orthogonalized using a sparse transformation method, and the output is a decoupled feature subset that meets the requirements of subsequent multi-dimensional attention fusion.

[0019] Preferably, the process of constructing a behavior nesting pattern model based on a lightweight multidimensional attention fusion network is as follows:

[0020] For the subset of behavioral features after structural adaptive decoupling, a multi-dimensional fusion mechanism combining channel attention and spatial attention is introduced to perform information weight reconstruction and context state expression.

[0021] By loading the pruned and optimized Transformer structure, the parameter scale is compressed while the core feature capture is preserved.

[0022] A behavior nesting pattern model output through multi-dimensional attention mechanism fusion.

[0023] Preferably, the process of introducing the inference stability observation factor is as follows:

[0024] During the training and inference process of the behavior nesting pattern model, observation channels are dynamically inserted;

[0025] Extract the maximum fluctuation difference of the predicted probabilities of different categories within each inference period, and combine it with the convergence trend in the time dimension to form an inference stability curve;

[0026] The curve is mapped to a convergence rate index factor and compared and verified with known stability models in the historical anomaly pattern library.

[0027] If the stability factor fluctuates within the preset tolerance range, it can be confirmed that the model is currently in the stable identification phase.

[0028] Preferably, the process of constructing a behavior transfer trajectory map by combining environmental anomaly disturbance factors is as follows:

[0029] Based on key abnormal segments marked in the stable phase, extract the behavior state change patterns before and after them, and construct the smallest behavior transition unit.

[0030] Synchronously access the abnormal interference data of environmental monitoring nodes, use the correlation coefficient between nodes as the weight of the connection edge, and establish a heterogeneous graph structure that includes behavioral state nodes and environmental factor nodes.

[0031] The time-series window translation method is used to identify the crossing paths of abnormal trajectories in the map, forming a behavior transfer trajectory map.

[0032] Preferably, the process of dynamically reshaping the edge behavior data sequence using a multi-scale difference recognition operator is as follows:

[0033] For the original observation data sequence corresponding to each behavioral state node in the trajectory map, a combination of fixed step size difference and variable step size difference is used for perturbation analysis;

[0034] A multi-scale window mechanism is introduced to evaluate the smoothness of density abrupt changes in both short and long periods, and to calculate the rate of change of the perturbation gradient.

[0035] Based on the temporal variation trend of the perturbation rate, resampling and outlier correction are performed on the original data sequence to reconstruct the edge behavior data sequence under a unified scale.

[0036] Preferably, the process for generating the behavioral intervention assessment result set is as follows:

[0037] Based on the matching error between the reorganized behavioral data sequence and the corresponding trajectory segment, a preliminary output of the anomaly identification probability is calculated;

[0038] By integrating a sudden interference response mechanism, interference correction processing is applied to the identification probability, and the influence factor of the current environmental state is added for normalization adjustment.

[0039] The evaluation scoring sub-models are established based on three categories of indicators: intervention cost, behavior persistence, and potential risk level. Finally, the results are integrated and output as a multi-dimensional behavioral intervention evaluation set.

[0040] Preferably, the process of generating the edge execution control instruction set is as follows:

[0041] Based on the risk level labels extracted from the behavioral intervention assessment results, the impact factors of various perception and execution nodes on the edge side are ranked to identify the set of key nodes that need to trigger control responses.

[0042] By combining a cluster of behavioral anomaly threshold standards, the identified categories are classified according to the urgency and controllability of the abnormal behavior, and a multi-level response control strategy parameter library is generated.

[0043] Based on the functional role of nodes in the edge system, corresponding control configuration parameters are matched for key nodes, and data caching strategies and computation scheduling priorities are adjusted for fusion-type nodes;

[0044] At the control strategy scheduling layer, the control template is adaptively optimized by integrating the current node running status, resource consumption level and predicted behavior trend.

[0045] Finally, the edge execution control instruction set is generated through the control instruction integration module.

[0046] A lightweight anomaly behavior recognition system based on edge computing includes:

[0047] Feature construction module: Acquires multidimensional heterogeneous behavioral data at the edge and constructs an updatable set of dynamic behavioral feature vectors through perturbation coding and scene weight analysis;

[0048] Inference monitoring module: Decouples and fuses behavioral feature vectors to build a lightweight behavioral pattern model, and monitors the inference stability of anomaly categories in real time during model iteration;

[0049] Trajectory Analysis Module: Determines the stability of anomaly recognition based on inference convergence rate, marks the anomaly feature transfer process, and generates corresponding behavior transfer trajectory maps;

[0050] Disturbance compensation module: Performs differential reshaping and disturbance modeling on behavioral trajectories, and generates corrected abnormal behavior intervention evaluation results by combining the disturbance response mechanism;

[0051] Risk assessment module: Based on intervention assessment results and abnormal threshold standards, it completes risk level judgment and generates a set of control instructions for edge nodes.

[0052] Compared with existing technologies, the present invention provides a lightweight abnormal behavior recognition system and method based on edge computing, which has the following beneficial effects:

[0053] This invention achieves efficient perception and dynamic updating of multi-source heterogeneous data in complex behavioral environments by deploying multi-dimensional sensing nodes at the edge and introducing a disturbance-aware coding mechanism. It effectively improves the accuracy of abnormal behavior pattern extraction and model computation efficiency by employing a structurally adaptive decoupling and multi-dimensional attention fusion network. By introducing an inference stability observation factor, it enhances the model's robustness and reliability judgment capabilities during non-stationary behavior evolution. It constructs a behavior transfer trajectory map by combining environmental anomaly interference factors, enabling visualized modeling of behavior evolution paths and the impact of external interventions. The integration of multi-scale differential operators and a sudden disturbance response mechanism improves the sensitivity and recognition accuracy to minor anomalies. Finally, it generates personalized edge execution control instruction sets based on risk level and node status, significantly reducing the load on the central server and achieving closed-loop control of behavior recognition and response. This invention possesses good deployment flexibility, resource controllability, and broad application and promotion value. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the method of the present invention;

[0055] Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a lightweight abnormal behavior identification method based on edge computing includes the following steps:

[0058] S1: By using the edge-side behavior acquisition node system, multi-dimensional heterogeneous behavior observation data covering the target area is acquired, and combined with the perturbation-aware coding mechanism and scene-dependent weight analysis, a set of dynamic behavior feature vectors that can be iteratively updated is constructed.

[0059] The process of constructing the iteratively updatable dynamic feature vector set in S1 is as follows:

[0060] By deploying various types of behavior sensing nodes at the edge, including visual capture devices, temperature probes, and inertial measurement units, real-time multi-source behavior observation of people or equipment within the target area can be performed.

[0061] Deploy various types of behavior sensing terminals at the edge nodes of the target area, including but not limited to visible light / infrared vision devices, thermal imaging temperature sensing probes, inertial measurement units such as accelerometers and gyroscopes, and configure local synchronization clocks to collect real-time behavioral data streams of personnel or equipment in the space, forming a raw multi-dimensional heterogeneous dataset.

[0062] The disturbance-aware coding mechanism is used to dynamically label the timestamp alignment, scale normalization and disturbance sensitivity of various types of data sources to form an initial behavioral fragment representation.

[0063] For the collected heterogeneous behavioral data, timestamp alignment is performed to ensure the consistency of data between different sensing channels. Then, the data of each channel is scaled and a perturbation sensing factor is introduced to mark high-frequency abrupt or discontinuous signal segments with perturbation-sensitive labels, generating a structured initial behavioral segment representation.

[0064] By combining scene-dependent weight analysis, weight matrices are assigned to different behavioral segments according to scene categories, and similarity matching and label correction are performed based on a historical scene pattern library;

[0065] By combining the pre-built scene recognition module in the edge node, the scene type to which the current behavior segment belongs is identified; for different behavior segments, the scene dependency weight matrix is ​​calculated according to their typicality and importance in the corresponding scene; at the same time, the historical behavior pattern library is used to perform similarity matching and automatic correction on the segment labels to improve the consistency and robustness of the data labels.

[0066] The encoded behavior fragments and scene labels are jointly embedded into the vector space, and an incremental dynamic feature fusion mechanism is used to generate a set of dynamic behavior feature vectors that supports iterative updates.

[0067] The perturbation-encoded and weight-corrected behavior fragment data, along with scene labels, are input into the embedding module. A unified behavior representation vector is generated using a low-dimensional mapping method based on sparse embedding. By introducing an incremental update strategy, the newly collected behavior vectors are dynamically fused with the existing feature set. A sliding window is used to perform weight coordination and trend compensation on conflict vectors, continuously improving the vector set structure. Based on the above dynamic fusion, the elements in the feature vector set are updated periodically or in an event-driven manner, forming a dynamic feature vector set of behavior with timeliness, adaptability, and iterability, providing input support for subsequent structural decoupling and nested modeling.

[0068] S2: Perform structural adaptive decoupling processing on the dynamic feature vector set of behavior, construct a behavior nested pattern model based on a lightweight multidimensional attention fusion network, and introduce an inference stability observation factor to record the fluctuation convergence rate of the abnormal category in real time during the model iteration process.

[0069] The adaptive decoupling process for the behavioral dynamic feature vector set in S2 is as follows:

[0070] Based on the time window and spatial domain partitioning rules, the behavioral dynamic feature vector set is initially classified to form multiple sets of structurally heterogeneous feature subsets.

[0071] Based on the preset time sliding window length and spatial domain partitioning criteria, the behavioral dynamic feature vector set is grouped according to time period and monitoring area; the feature data in each spatiotemporal group is preliminarily statistically aggregated to construct a feature subset with local consistency, thereby reducing cross-regional interference and mixing between data.

[0072] An adaptive decoupling algorithm is introduced within each feature subset to identify the dominant dependency path between features based on a Bayesian network model and to perform suppression operations on weakly correlated dimensions.

[0073] For each structurally heterogeneous feature subset, a dependency modeling algorithm based on Bayesian structural learning is applied to construct a directed acyclic graph that reflects the causal relationships between each feature dimension. The dominant feature dimension with a strong correlation path is identified in the graph structure as the core feature pathway. For dimensions with low dependency or weak correlation with the main path, an inhibition factor is introduced to reduce their fusion influence.

[0074] The highly coupled dimension is orthogonalized using a sparse transformation method, and the output is a decoupled feature subset that meets the requirements of subsequent multidimensional attention fusion.

[0075] For feature dimension groups with high coupling in Bayesian graphs, sparse transformation methods such as sparse principal component analysis or independent component analysis are used to transform the feature space. The original highly coupled dimensions are projected onto a set of orthogonal bases to improve the separability and independence between feature vectors. The processed feature subsets with sparsity, orthogonality and clear dependencies are uniformly encoded and output. This provides a structurally optimized and highly discriminative data input channel for the fusion network input, improving the overall accuracy and stability of abnormal behavior modeling.

[0076] The process of constructing the behavior nesting pattern model based on the lightweight multidimensional attention fusion network in S2 is as follows:

[0077] For the subset of behavioral features after structural adaptive decoupling, a multi-dimensional fusion mechanism combining channel attention and spatial attention is introduced to perform information weight reconstruction and context state expression.

[0078] Channel-dimensional attention and spatial-dimensional attention modules are constructed for the feature subsets after structural adaptive decoupling. In the channel attention mechanism, a weight distribution is generated by combining global average pooling and max pooling among features, and higher semantic focus weights are assigned to high-response channels. In the spatial attention mechanism, the importance distribution of local feature regions is extracted based on two-dimensional convolution, and the focus is dynamically on spatial regions with significant behavioral features. The channel and spatial attention are fused in parallel or serially and then embedded into the context encoding path to form a multi-dimensional information cross-enhanced expression, thereby improving the ability to perceive the spatiotemporal evolution of behavior.

[0079] In scenarios where edge computing resources are limited, the core feature capture can be retained while compressing the parameter scale by loading a pruned and optimized Transformer structure.

[0080] The standard Transformer module is pruned and distilled to remove redundant attention heads and shallow feedforward substructures, reducing the number of parameters and computational complexity. A lightweight positional encoding strategy and low-rank matrix factorization technique are introduced to preserve the model's ability to express temporal features. The optimized Transformer is loaded into the edge-side inference engine, and its computation path is dynamically adjusted during model invocation to improve response speed. This ensures that effective modeling of key state transition points and evolutionary trends can still be achieved in nested behavior recognition scenarios.

[0081] The model results fused through the multi-dimensional attention mechanism exhibit a hierarchical and nested behavioral structure.

[0082] The intermediate feature results output by the fusion network are decoded according to semantic hierarchy to form a bottom-up hierarchical nested behavior representation. In each nested structure, the activation path and context dependency structure of key feature nodes are recorded to construct a nested behavior graph. Finally, a nested pattern model that reflects the behavior evolution logic, has clear hierarchy, and has a stable structure is formed, providing structural support for subsequent abnormal behavior identification and path intervention analysis.

[0083] The process of introducing the inference stability observation factor in S2 is as follows:

[0084] During the training and inference process of the behavior nesting pattern model, observation channels are dynamically inserted;

[0085] An auxiliary observation module is embedded in the output layer and key intermediate layer of the behavior nested pattern model to record the predicted probability distribution of each inference output. This module continuously collects the confidence output of the model for each category in different inference cycles using a sliding window mechanism. It supports local storage of sampled data with a lightweight cache structure, adapting to the low latency and limited storage conditions of edge computing environments.

[0086] Extract the maximum fluctuation difference of the predicted probabilities of different categories within each inference period, and combine it with the convergence trend in the time dimension to form an inference stability curve;

[0087] Within each inference window period, the probability changes of each predicted category are statistically analyzed, and the difference between its maximum and minimum values ​​is calculated. For all output categories, a fluctuation matrix is ​​constructed along the category dimension to reflect the instability of the current model's recognition results. Timestamp alignment is introduced to ensure that the fluctuation calculation is strictly synchronized with the inference cycle, eliminating the influence of instantaneous interference or impulsive anomalies. Time series modeling is performed on the fluctuation difference to plot the confidence fluctuation trend curve that changes over time. Smoothing algorithms such as exponential moving average filters are applied to denoise and trend fit the original fluctuation curve. A stability curve describing the change in the stability of the model output during the inference process is obtained, serving as the basis for subsequent stability analysis.

[0088] The curve is mapped to a convergence rate index factor and compared and verified with known stability models in the historical anomaly pattern library.

[0089] Statistical features such as slope, amplitude, and frequency are extracted from the stability curve to construct a convergence rate index factor. This factor is standardized and matched with a historical anomaly behavior identification stability model library to compare the consistency of the current model's behavior stability with historical mature models. The model stability level is evaluated using methods such as minimum distance matching, divergence, or Mahalanobis distance.

[0090] If the stability factor fluctuates within the preset tolerance range, it can be confirmed that the model is currently in the stability identification phase.

[0091] Set the tolerance range threshold for model recognition stability; when the convergence rate index factor stably falls within this tolerance range for several consecutive periods, it is determined that the model has entered a stable recognition state.

[0092] S3: Based on the nonlinear perturbation trend of the convergence rate, determine the stability of behavior recognition during the model inference stage. If it is determined to be stable, mark the key feature transfer segments of the current abnormal behavior, and construct the behavior transfer trajectory map in combination with environmental anomaly interference factors.

[0093] The process of constructing the behavior transfer trajectory map by combining environmental anomaly interference factors in S3 is as follows:

[0094] Based on key abnormal segments marked in the stable phase, extract the behavior state change patterns before and after them, and construct the smallest behavior transition unit.

[0095] Using the stable reasoning phase time period identified in the previous steps, the key behavioral segments that the model determines to be abnormal are located; the abnormal segments are extended forward / backward by a fixed duration window, and the behavioral state feature vectors before and after their occurrence are extracted to form behavioral change pairs; each pair of state changes constitutes a minimum behavioral transition unit, which serves as the basic building block of subsequent graph nodes.

[0096] Synchronously access environmental monitoring node data on anomalies and disturbances, including highly sensitive factors such as temperature anomalies, crowd gatherings, and changes in light intensity;

[0097] It connects to various environmental monitoring devices deployed at the edge, such as temperature and humidity sensors, noise detectors, infrared / photosensitive probes, and crowd density detectors; it performs time-series alignment on environmental sensor data adjacent to the time period of the behavior segment and extracts abnormal fluctuation signals; it marks external interference factors that may exist during the occurrence of the behavior as the background of the behavior state transition.

[0098] Using the correlation coefficient between node anomalies as the weight of the connection edges, a heterogeneous graph structure containing behavioral state nodes and environmental factor nodes is established;

[0099] Each minimum behavior transition unit is represented as a pair of behavior state nodes; the extracted highly sensitive environmental factors are represented as environmental state nodes; the correlation coefficient between each behavior state change and the environmental factors is calculated and used as the weight of the connecting edge; a heterogeneous behavior transition graph containing two types of nodes and their associated edges is constructed.

[0100] The cross-path of abnormal trajectories in the map is identified by the temporal window translation method, forming a behavior transfer trajectory map;

[0101] In a heterogeneous graph, a sliding time window is set up, and graph traversal is performed starting from the abnormal behavior node. Based on the edge weight priority, a complete path is searched that starts from the initial abnormal state and crosses multiple state transition nodes and environmental factor nodes. Potential abnormal propagation paths or behavior transition chains are constructed using graph traversal strategies. These traversal paths constitute a complete behavior transition trajectory graph.

[0102] S4: Based on the behavior transfer trajectory map, apply the multi-scale differential recognition operator to dynamically reorganize the edge behavior data sequence, extract the density perturbation parameters of abnormal behavior, and combine the sudden interference response mechanism to correct the recognition probability output, and generate a behavior intervention evaluation result set.

[0103] In step S4, the process of dynamically reorganizing the edge behavior data sequence using a multi-scale differential recognition operator is as follows:

[0104] For the original observation data sequence corresponding to each behavioral state node in the trajectory map, a combination of fixed step size difference and variable step size difference is used for perturbation analysis;

[0105] The system iterates through the constructed behavior transition trajectory map, extracting the raw sensing data associated with each behavior state node in the map; maps each behavior segment to a time sequence, and organizes them in order according to timestamps to form a set of raw behavior observation data streams; applies a fixed step-size difference operator to each behavior data sequence to identify basic change trends and abrupt change points; and a variable step-size difference operator to adaptively adjust the difference interval according to the volatility of the behavior segment to enhance the detection capability of sensitive areas of local disturbances; and generates a differential perturbation sequence for subsequent multi-scale analysis of perturbation intensity and distribution.

[0106] A multi-scale window mechanism is introduced to evaluate the smoothness of density abrupt changes in both short and long periods, and to calculate the rate of change of the perturbation gradient.

[0107] Set up time sliding windows of multiple scales. Within each window: calculate the local smoothness index of the differential perturbation sequence; evaluate the rate of change of the perturbation gradient to quantify the severity of the sequence fluctuations at each scale; and identify the locations where significant perturbation abrupt changes occur at different scales.

[0108] Based on the temporal variation trend of the perturbation rate, resampling and outlier correction are performed on the original data sequence to reconstruct the edge behavior data sequence under a unified scale;

[0109] For periods with abnormally high disturbance rates, dense sampling is performed to preserve key mutation details of the original data; for periods with low disturbance rates or stable changes, sparse sampling is performed to reduce redundancy. For points with excessively large differential gradient changes, outlier correction is performed in conjunction with historical behavior distribution models; a unified reorganization behavior sequence after disturbance regulation is constructed; the results of multi-scale analysis are fused and uniformly mapped to a standardized time axis and data scale; and a time-consistent, fluctuation-balanced, and structurally optimized edge behavior data sequence is output, providing a standardized data foundation for subsequent identification model input.

[0110] The process of generating the behavioral intervention assessment result set in S4 is as follows:

[0111] Based on the matching error between the reorganized behavioral data sequence and the corresponding trajectory segment, a preliminary output of the anomaly identification probability is calculated;

[0112] The dynamically reorganized edge behavior data sequence is input into the constructed behavior nesting pattern model; the structural features of the current data and the corresponding behavior segment in the trajectory map are compared frame by frame to calculate the sequence matching error; based on the size of the matching error, the soft discriminant function is used to output the preliminary identification probability value of the current behavior as abnormal; and a preliminary list of time period-behavior category-abnormal probability triplet is obtained.

[0113] By integrating a sudden interference response mechanism, interference correction processing is applied to the identification probability, and the influence factor of the current environmental state is added for normalization adjustment.

[0114] The system acquires interference factor data for the current time period from the environmental perception channel; constructs an interference impact function to quantify the correction coefficients of various sudden interferences on the recognition accuracy; combines the initial recognition probability with the interference correction coefficients, performs disturbance sensitivity normalization processing, and obtains a more robust anomaly recognition probability output behavior-time-interference correction probability matrix.

[0115] An assessment scoring sub-model is established based on three categories of indicators: intervention cost, behavior persistence, and potential risk level. Finally, a multi-dimensional behavioral intervention assessment result set is integrated and output.

[0116] Intervention cost scoring model: assesses the types and quantities of resources required to trigger an intervention, and provides an intervention cost score based on a resource consumption table; Behavior persistence scoring model: calculates the duration score of the current abnormal state based on the length of the behavior segment and the frequency of recurrence; weights are increased for abnormalities that have not recovered for a long time; Potential risk level scoring model: integrates historical behavior labels, environmental status level, and current abnormal probability, and uses a risk coefficient library to output a potential harm level score caused by the behavior; the outputs of the above three scoring sub-models are standardized into a unified scoring range; combined to form a ternary scoring vector for each abnormal behavior: each abnormal behavior instance is labeled, sorted, and prioritized; a structured behavioral intervention assessment result set is output.

[0117] S5: Based on the behavioral intervention assessment result set and combined with the preset behavioral abnormality threshold standard cluster, the current identified behavioral results are used to determine the hierarchical risk and automatically generate the edge execution control instruction set.

[0118] The process of generating the edge execution control instruction set in S5 is as follows:

[0119] Based on the risk level labels extracted from the behavioral intervention assessment results, the impact factors of various perception and execution nodes on the edge side are ranked to identify the set of key nodes that need to trigger control responses.

[0120] Read the risk level labels of each abnormal behavior item in the behavioral intervention assessment result set; perform impact factor assessment on edge nodes to quantify the correlation score between each node and high-risk behavior; sort by impact score, set threshold for screening, and form a set of key edge nodes that need to be responded to; mark the functional roles and perception scope of key nodes to prepare for strategy binding.

[0121] By combining a cluster of behavioral anomaly threshold standards, the identified categories are classified according to the urgency and controllability of the abnormal behavior, and a multi-level response control strategy parameter library is generated.

[0122] The cluster classifies abnormal behaviors into multiple levels based on the system's built-in abnormal threshold standard. For each abnormal level, a set of response control strategy templates is designed, covering: the sampling frequency of sensing nodes; the response actions of execution nodes; and the adjustment parameters of model inference accuracy. The template parameters are stored in the hierarchical control strategy parameter library for subsequent dynamic matching.

[0123] Based on the functional role of nodes in the edge system, corresponding control configuration parameters are matched for key nodes, and data caching strategies and computation scheduling priorities are adjusted for fusion-type nodes;

[0124] For each key node, specific parameters are matched according to its functional type: Sensing nodes: configure sampling interval, sensing sensitivity, and data upload frequency; Execution nodes: load corresponding control commands, such as initiating device linkage and executing behavioral intervention commands; Fusion nodes: adjust data caching strategies and set edge computing task priorities; establish a "node type - parameter configuration" mapping table to ensure the feasibility and compatibility of strategy application.

[0125] At the control strategy scheduling layer, the control template is adaptively optimized by integrating the current node running status, resource consumption level and predicted behavior trend.

[0126] Obtain the current operating status information of each key node in the edge system; use the behavior trend prediction module to analyze the behavior risk fluctuation trend in the next period of time; adaptively adjust the matched control template according to the above information: dynamically reduce the task intensity of nodes with high load; adjust the scheduling window period; delete or rewrite inapplicable parameters in the strategy; output an optimized control configuration set that conforms to the current system status and behavior prediction.

[0127] Finally, the control instruction integration module generates an edge execution control instruction set, which includes strategy optimization suggestions, feature data compression schemes, model dynamic reconstruction instructions, and perception task reordering plans.

[0128] All optimized control parameter templates, node configuration information, and behavioral risk contexts are integrated into the control instruction integration module. Following a modular structure, the following types of control instructions are generated: strategy optimization suggestions (simplification of behavioral intervention processes, reordering of strategy execution paths, etc.); feature data compression schemes (low-priority feature dimension compression encoding methods); model dynamic reconstruction instructions (lightweight inference schemes based on model pruning and parameter freezing); perception task reordering plans (adjusting the division of perception tasks among nodes to avoid resource bottlenecks); and an edge execution control instruction set with schedulable, deployable, and traceable characteristics is output.

[0129] Example 2: Figure 2As shown in the embodiment of the present invention, a lightweight abnormal behavior recognition system based on edge computing includes:

[0130] Feature construction module: Acquires multidimensional heterogeneous behavioral data at the edge and constructs an updatable set of dynamic behavioral feature vectors through perturbation coding and scene weight analysis;

[0131] Inference monitoring module: Decouples and fuses behavioral feature vectors to build a lightweight behavioral pattern model, and monitors the inference stability of anomaly categories in real time during model iteration;

[0132] Trajectory Analysis Module: Determines the stability of anomaly recognition based on inference convergence rate, marks the anomaly feature transfer process, and generates corresponding behavior transfer trajectory maps;

[0133] Disturbance compensation module: Performs differential reshaping and disturbance modeling on behavioral trajectories, and generates corrected abnormal behavior intervention evaluation results by combining the disturbance response mechanism;

[0134] Risk assessment module: Based on intervention assessment results and abnormal threshold standards, it completes risk level judgment and generates a set of control instructions for edge nodes.

[0135] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0136] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A lightweight abnormal behavior recognition method based on edge computing, characterized in that, Includes the following steps: By using the edge-side behavior acquisition node system, multi-dimensional heterogeneous behavior observation data covering the target area is acquired, and combined with the perturbation-aware coding mechanism and scene-dependent weight analysis, a set of dynamic behavior feature vectors that can be iteratively updated is constructed. The structure is adaptively decoupled for the dynamic feature vector set of behavior. A behavior nested pattern model is constructed based on a lightweight multidimensional attention fusion network. An inference stability observation factor is introduced to record the fluctuation convergence rate of the anomaly category in real time during the model iteration process. Based on the nonlinear perturbation trend of the convergence rate, the stability of behavior recognition in the model inference stage is judged. If it is determined to be stable, the key feature transfer segments of the current abnormal behavior are marked, and a behavior transfer trajectory map is constructed in combination with environmental anomaly interference factors. Based on the behavior transfer trajectory map, a multi-scale differential recognition operator is applied to dynamically reorganize the edge behavior data sequence, extract the density perturbation parameters of abnormal behavior, and combine the sudden interference response mechanism to correct the recognition probability output, thereby generating a behavior intervention evaluation result set. Based on the behavioral intervention assessment result set and combined with the preset behavioral abnormality threshold standard cluster, the current identified behavioral results are used to determine the hierarchical risk and automatically generate a set of edge execution control instructions.

2. The lightweight abnormal behavior recognition method based on edge computing according to claim 1, characterized in that, The process of constructing an iteratively updatable set of dynamic behavioral feature vectors is as follows: By deploying multiple types of behavior sensing nodes at the edge, real-time multi-source behavior observation of people or equipment within the target area can be performed; The disturbance-aware coding mechanism is used to dynamically label the timestamp alignment, scale normalization and disturbance sensitivity of various types of data sources to form an initial behavioral fragment representation. By combining scene-dependent weight analysis, weight matrices are assigned to different behavioral segments according to scene categories, and similarity matching and label correction are performed based on a historical scene pattern library; The encoded behavior fragments are jointly embedded into the vector space along with scene labels, and an incremental dynamic feature fusion mechanism is used to generate a set of dynamic behavior feature vectors that supports iterative updates.

3. The lightweight abnormal behavior recognition method based on edge computing according to claim 2, characterized in that, The process of performing structural adaptive decoupling on the behavioral dynamic feature vector set is as follows: Based on the time window and spatial domain partitioning rules, the behavioral dynamic feature vector set is initially classified to form multiple sets of structurally heterogeneous feature subsets. An adaptive decoupling algorithm is introduced within each feature subset to identify the dominant dependency path between features based on a Bayesian network model and to perform suppression operations on weakly correlated dimensions. The high-coupling dimension is orthogonalized using a sparse transformation method, and the output is a decoupled feature subset that meets the requirements of subsequent multi-dimensional attention fusion.

4. The lightweight abnormal behavior recognition method based on edge computing according to claim 3, characterized in that, The process of constructing a behavior nested pattern model based on a lightweight multidimensional attention fusion network is as follows: For the subset of behavioral features after structural adaptive decoupling, a multi-dimensional fusion mechanism combining channel attention and spatial attention is introduced to perform information weight reconstruction and context state expression. By loading the pruned and optimized Transformer structure, the parameter scale is compressed while the core feature capture is preserved. The output behavior nesting pattern model is fused through a multi-dimensional attention mechanism.

5. A lightweight abnormal behavior recognition method based on edge computing according to claim 4, characterized in that, The process of introducing the inference stability observation factor is as follows: During the training and inference process of the behavior nesting pattern model, observation channels are dynamically inserted; Extract the maximum fluctuation difference of the predicted probabilities of different categories within each inference period, and combine it with the convergence trend in the time dimension to form an inference stability curve; The inference stability curve is mapped to a convergence rate index factor and compared and verified with known stability models in the historical anomaly pattern library. If the stability factor fluctuates within the preset tolerance range, it can be confirmed that the model is currently in the stable identification phase.

6. A lightweight abnormal behavior recognition method based on edge computing according to claim 5, characterized in that, The process of constructing a behavior transfer trajectory map by combining environmental anomaly disturbance factors is as follows: Based on key abnormal segments marked in the stable phase, extract the behavior state change patterns before and after them, and construct the smallest behavior transition unit. Synchronously access the abnormal interference data of environmental monitoring nodes, use the correlation coefficient between nodes as the weight of the connection edge, and establish a heterogeneous graph structure that includes behavioral state nodes and environmental factor nodes. The time-series window translation method is used to identify the crossing paths of abnormal trajectories in the map, forming a behavior transfer trajectory map.

7. The lightweight abnormal behavior identification and method based on edge computing according to claim 6, characterized in that, The process of dynamically reshaping edge behavior data sequences using multi-scale difference recognition operators is as follows: For the original observation data sequence corresponding to each behavioral state node in the trajectory map, a combination of fixed step size difference and variable step size difference is used for perturbation analysis; A multi-scale window mechanism is introduced to evaluate the smoothness of density abrupt changes in both short and long periods, and to calculate the rate of change of the perturbation gradient. Based on the temporal variation trend of the perturbation rate, resampling and outlier correction are performed on the original data sequence to reconstruct the edge behavior data sequence under a unified scale.

8. The lightweight abnormal behavior identification and method based on edge computing according to claim 7, characterized in that, The process of generating the behavioral intervention assessment result set is as follows: Based on the matching error between the reorganized behavioral data sequence and the corresponding trajectory segment, a preliminary output of the anomaly identification probability is calculated; By integrating a sudden interference response mechanism, interference correction processing is applied to the identification probability, and the influence factor of the current environmental state is added for normalization adjustment. The evaluation scoring sub-models are established based on three categories of indicators: intervention cost, behavior persistence, and potential risk level. Finally, the results are integrated and output as a multi-dimensional behavioral intervention evaluation set.

9. A lightweight abnormal behavior identification method based on edge computing according to claim 8, characterized in that, The process of generating the edge execution control instruction set is as follows: Based on the risk level labels extracted from the behavioral intervention assessment results, the impact factors of various perception and execution nodes on the edge side are ranked to identify the set of key nodes that need to trigger control responses. By combining a cluster of behavioral anomaly threshold standards, the identified categories are classified according to the urgency and controllability of the abnormal behavior, and a multi-level response control strategy parameter library is generated. Based on the functional role of nodes in the edge system, corresponding control configuration parameters are matched for key nodes, and data caching strategies and computation scheduling priorities are adjusted for fusion-type nodes; At the control strategy scheduling layer, the control template is adaptively optimized by integrating the current node running status, resource consumption level and predicted behavior trend. Finally, the edge execution control instruction set is generated through the control instruction integration module.

10. A lightweight abnormal behavior recognition system based on edge computing, applied to the method described in any one of claims 1-9, characterized in that, include: Feature construction module: Acquires multidimensional heterogeneous behavioral data at the edge and constructs an updatable set of dynamic behavioral feature vectors through perturbation coding and scene weight analysis; Inference monitoring module: Decouples and fuses behavioral feature vectors to build a lightweight behavioral pattern model, and monitors the inference stability of anomaly categories in real time during model iteration; Trajectory Analysis Module: Determines the stability of anomaly recognition based on inference convergence rate, marks the anomaly feature transfer process, and generates corresponding behavior transfer trajectory maps; Disturbance compensation module: Performs differential reshaping and disturbance modeling on behavioral trajectories, and generates corrected abnormal behavior intervention evaluation results by combining the disturbance response mechanism; Risk assessment module: Based on intervention assessment results and abnormal threshold standards, it completes risk level judgment and generates a set of control instructions for edge nodes.

Images