A Deep Learning-Based Edge Computing Task Offloading and Scheduling Method for Port Safety Monitoring

Abstract

This invention discloses a deep learning-based edge computing task offloading and scheduling method for port safety monitoring, relating to the field of port safety monitoring technology. The method includes the following steps: performing deep learning inference on port safety monitoring videos to extract semantic features of abnormal behavior; quantifying the task processing process to extract task data scale features; and constructing semantic risk potential values ​​and load scale potential values ​​based on the semantic features of abnormal behavior and the task data scale features, respectively; constructing a potential energy adversarial domain based on the directional difference between the semantic risk potential value and the load scale potential value in a unified reference space; and determining whether there is a conflict between the semantic features of abnormal behavior and the task data scale features based on the potential energy adversarial domain. This invention solves the problem of not being able to adjust the scheduling focus direction based on the intensity of the conflict between semantic risk and task scale, achieving the effect of prioritizing high-risk tasks and dynamically optimizing scheduling strategies.

Description

Technical Field

[0001] This invention relates to the field of port safety monitoring technology, specifically to a deep learning-based method for offloading and scheduling edge computing tasks for port safety monitoring. Background Technology

[0002] Deep learning-based edge computing task offloading and scheduling for port safety monitoring refers to a method that, in port safety monitoring scenarios, utilizes deep learning models to model and analyze computing resources and network conditions in the edge computing environment for computationally intensive tasks such as video surveillance and anomaly detection. This allows for intelligent decision-making regarding the optimal offloading location and execution strategy for tasks between terminal devices, local edge nodes, or the remote cloud. Its core objective is to reduce system latency and resource consumption while improving overall response efficiency, while ensuring monitoring accuracy. Existing technologies typically construct deep neural networks or reinforcement learning models, using multi-dimensional state information from the port monitoring system as input, including video data complexity, task urgency, edge node computing load, and network conditions. By considering factors such as network bandwidth and latency, the model learns the mapping relationship between task scheduling and resource allocation through model training, and outputs the optimal offloading decision in real time during operation. The specific implementation process generally includes: first, multi-source data acquisition and state awareness are carried out to model the port monitoring task and computing environment; second, the monitoring video stream is decomposed into tasks and features are extracted to form a set of schedulable subtasks; then, relevant state parameters are input into a pre-trained deep learning model for inference to obtain the offloading strategy and resource allocation scheme for each task; next, collaborative scheduling and dynamic adjustment are performed among edge computing nodes; and finally, the model parameters or strategies are continuously updated through a feedback mechanism, thereby achieving efficient, intelligent and adaptive task offloading scheduling in a complex and ever-changing port environment.

[0003] The existing technology has the following shortcomings: In the process of implementing edge computing task offloading and scheduling for port safety monitoring based on deep learning, when low-data-volume but high-risk abnormal behaviors occur in the port safety monitoring scenario, the semantic features of abnormal behaviors extracted by deep learning and the task data scale features collected by edge computing form conflicting expressions in the same input. Because the deep learning model uniformly maps features at different levels during training and mainly models based on physical features such as data scale, it lacks an explicit characterization mechanism for semantic risk priority. As a result, the model still tends to make judgments based on the task data scale when making scheduling decisions. Existing technologies cannot adjust the decision-making focus of deep learning in edge computing task offloading and scheduling according to the conflict intensity when the semantic features of abnormal behaviors and the task data scale features conflict in port safety monitoring. This results in high-risk but low-computation tasks being incorrectly assigned to low-priority or low-computation-power nodes, thereby affecting the timeliness of abnormal event response and the overall reliability of the port safety monitoring system.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for offloading and scheduling edge computing tasks for port safety monitoring, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for offloading and scheduling edge computing tasks for port safety monitoring, specifically including the following steps: S1. Deep learning inference is performed on port safety monitoring videos to extract semantic features of abnormal behavior, the task processing process is quantified to extract task data scale features, and semantic risk potential value and load scale potential value are constructed based on the semantic features of abnormal behavior and the task data scale features, respectively. S2. Construct a potential energy adversarial domain based on the directional difference between semantic risk potential energy value and load scale potential energy value in a unified reference space, and determine whether there is a conflict between the semantic features of abnormal behavior and the scale features of task data based on the potential energy adversarial domain. S3. In the event of a conflict, calculate the potential energy offset gradient based on the offset relationship between the semantic risk potential energy value and the load scale potential energy value, and generate the conflict intensity through nonlinear mapping. S4. Map the conflict intensity to generate decision-driven guiding factors, and construct semantic-dominated migration trajectories and load-dominated migration trajectories based on the changing trends of decision-driven guiding factors. S5. Based on the positional relationship between decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories, adjust the decision-oriented direction of deep learning in edge computing task offloading scheduling, and update the decision-driven guiding factors based on the potential energy offset gradient.

[0007] Preferably, S1 specifically includes the following steps: S101. Perform continuous frame parsing on port safety monitoring video and construct a time-series image sequence. Input the time-series image sequence into a deep learning network for inference operations. Output a semantic vector containing abnormal behavior category information and abnormal behavior probability information. Perform category encoding processing on the abnormal behavior category information and numerical normalization processing on the abnormal behavior probability information. Perform weighted combination calculation on the category encoding result and the normalized abnormal behavior probability information to extract the semantic features of abnormal behavior. S102. Based on the input data stream during the task processing, the data transmission volume, processing latency and data change rate are numerically calculated to construct a multi-dimensional data scale vector. The data of each dimension in the multi-dimensional data scale vector are standardized and weighted fusion calculation is performed according to the distribution weight of each dimension in a unified numerical space to extract the task data scale features. S103. Based on the mapping relationship between the semantic features of abnormal behavior and the scale features of task data in a unified numerical space, perform weight mapping calculation on the semantic features of abnormal behavior to construct the semantic risk potential value, and perform scale transformation calculation on the scale features of task data to construct the load scale potential value.

[0008] Preferably, S103 is as follows: Based on the mapping relationship between the semantic features of abnormal behavior and the scale features of task data in a unified numerical space, the semantic features of abnormal behavior are decomposed into dimensions and a set of semantic components is constructed. The components in the set of semantic components are normalized numerically and weighted and superimposed according to the distribution weight of each component in the unified numerical space to form a weight mapping result. The weight mapping results are mapped and transformed along the directional gradient of the unified numerical space, and the cumulative calculation is performed by combining the weight contribution relationship of each component in the semantic component set to construct the semantic risk potential value. Based on the distribution range of the task data scale characteristics in a unified numerical space, the task data scale characteristics are scale segmented and a scale component set is constructed. The scale of each component in the scale component set is normalized, and the load scale potential energy value is constructed by proportional transformation based on the scale span of each component in the unified numerical space.

[0009] Preferably, S2 is as follows: The semantic risk potential value and the load scale potential value are mapped to a unified reference space. The directional components of the semantic risk potential value and the load scale potential value in each dimension of the unified reference space are extracted respectively, and a directional difference distribution sequence is constructed based on the directional components in each dimension. Based on the directional difference distribution sequence, the directional difference between semantic risk potential value and load scale potential value in the unified reference space is divided into intervals. The intervals with consistent directions and the intervals with opposite directions are separated. The directional components of each dimension in the intervals with opposite directions are aggregated to form a continuous interval set. The boundary is determined according to the coverage of the continuous interval set in the unified reference space. The potential energy adversarial domain is constructed based on the numerical interval corresponding to the coverage. Based on the distribution of semantic risk potential energy value and load scale potential energy value in the potential energy adversarial domain, the overlapping interval of semantic risk potential energy value and load scale potential energy value in the potential energy adversarial domain is identified, and the proportion of semantic risk potential energy value and load scale potential energy value in the overlapping interval is calculated. When the proportion of semantic risk potential energy value and load scale potential energy value in the overlapping interval is greater than zero, it is determined that the semantic features of abnormal behavior and the features of task data scale conflict. When the proportion of at least one of semantic risk potential energy value and load scale potential energy value in the overlapping interval is zero, it is determined that the semantic features of abnormal behavior and the features of task data scale do not conflict.

[0010] Preferably, S3 specifically includes the following steps: S301. In the event of a conflict, the semantic risk potential value and the load scale potential value are mapped to a unified reference space. The numerical differences between the semantic risk potential value and the load scale potential value in each dimension of the unified reference space are calculated dimension by dimension and arranged in dimensional order to form a sequence of offset relationships between the semantic risk potential value and the load scale potential value. S302. Perform differential calculation on the numerical differences between adjacent dimensions in the offset relationship sequence to form the offset change between each adjacent dimension. Arrange each offset change according to the dimensional order in the unified reference space. Combine the positional distribution of each offset change in the unified reference space, perform gradient classification processing on the offset change. Divide the offset change corresponding to different change amplitudes into different gradient levels. Reorganize and arrange the offset change according to the gradient level to form the potential energy offset gradient. S303. The potential energy offset gradient is segmented according to the numerical interval in the unified reference space. The distribution density of the potential energy offset gradient in each interval is statistically analyzed. Based on the distribution density, the potential energy offset gradient is nonlinearly mapped and calculated. The nonlinear mapping calculation results are accumulated and calculated to generate the conflict intensity.

[0011] Preferably, S303 is as follows: The potential energy shift gradient is divided into multiple continuous numerical intervals according to the numerical range in the unified reference space, and the potential energy shift gradient is segmented according to the distribution of the potential energy shift gradient in each numerical interval. For the potential energy shift gradient in each numerical interval, the frequency and distribution concentration of the potential energy shift gradient in each numerical interval are statistically analyzed, and the distribution density sequence of the corresponding numerical interval is constructed based on the frequency and distribution concentration. The potential energy shift gradient is nonlinearly mapped based on the distribution density sequence. The potential energy shift gradient corresponding to different distribution densities is segmented and weighted. The nonlinear mapping calculation results are accumulated segment by segment according to the numerical interval order to generate the conflict intensity.

[0012] Preferably, S4 specifically includes the following steps: S401. Divide the numerical distribution of conflict intensity in a unified reference space into intervals, map the conflict intensity to the corresponding numerical intervals, assign values ​​to the conflict intensity in each numerical interval according to the interval position, and perform numerical transformation of the conflict intensity according to the mapping relationship corresponding to each numerical interval. Combine the transformed values ​​in order according to the arrangement order of the numerical intervals to construct a decision-driven guiding factor. S402. Extract the numerical changes of decision-driven guiding factors in a unified reference space point by point, calculate the direction and magnitude of numerical changes between adjacent positions, classify and label them according to the direction of numerical changes, and classify them according to the distribution range of numerical change magnitude in the unified reference space. Combine and arrange the classification and labeling results with the classification and division results to construct a change trend sequence. S403. Based on the direction of change at each position in the trend sequence, divide the trend sequence into intervals. Construct the intervals where the direction of change is biased towards the semantic risk potential value as the semantic-dominated migration trajectory, and construct the intervals where the direction of change is biased towards the load scale potential value as the load-dominated migration trajectory.

[0013] Preferably, S403 is as follows: The direction of change at each position in the trend sequence is extracted point by point, and the direction of change is classified and marked according to the value sign of the direction of change in the unified reference space. The direction of change is divided into direction marks biased towards semantic risk potential value and direction marks biased towards load scale potential value. Based on the distribution of directional markers at continuous positions in the trend sequence, continuous positions with the same directional markers are aggregated to form multiple continuous intervals, and the intervals are divided according to the order of the continuous intervals in the trend sequence. Based on the interval division results, the continuous intervals containing the direction markers of the biased semantic risk potential value are connected in order of position to construct the semantically dominant migration trajectory, and the continuous intervals containing the direction markers of the biased load scale potential value are connected in order of position to construct the load-dominated migration trajectory.

[0014] Preferably, S5 is as follows: The positions of decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories are located point by point. The corresponding position intervals of decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories are extracted, and position relationship sequences are constructed according to the distribution order of position intervals in the two types of migration trajectories. Based on the distribution ratio of each position interval in the positional relationship sequence in the semantically dominant migration trajectory and the load-dominated migration trajectory, the decision-driven guidance factor is subjected to direction mapping processing. The decision-driven guidance factor corresponding to the position interval with the distribution ratio biased towards the semantically dominant migration trajectory is adjusted to the semantically biased direction, and the decision-driven guidance factor corresponding to the position interval with the distribution ratio biased towards the load-dominated migration trajectory is adjusted to the load-biased direction, so as to complete the adjustment of the decision-biased direction of deep learning in the offloading and scheduling of edge computing tasks based on positional relationship. The potential energy shift gradient is mapped to the corresponding position interval in the unified reference space. The potential energy shift gradient and the decision-driven guidance factor are matched point by point in the same position interval. The decision-driven guidance factor is numerically corrected according to the change of the potential energy shift gradient in each position interval, thus completing the update of the decision-driven guidance factor based on the potential energy shift gradient.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a multi-level associative processing link consisting of "abnormal behavior semantic features—task data scale features—potential energy mapping—conflict intensity—decision-driven guiding factors—migration trajectory—scheduling emphasis direction." This decouples and adversarially models semantic risk information and computational load information, which were originally difficult to prioritize within a unified modeling system. This achieves explicit characterization and quantitative description of the conflict relationship between the two types of features. Based on this, the conflict state is structurally analyzed through potential energy adversarial domain and potential energy offset gradient. Nonlinear mapping is used to transform the degree of conflict into a conflict intensity that can participate in decision-making. This allows deep learning to no longer rely solely on physical characteristics such as data scale for judgment during the offloading and scheduling of edge computing tasks. Instead, it can comprehensively consider the dynamic game relationship between semantic risk and computational load. This solves the problem in existing technologies where high-risk, low-load tasks are processed with low priority due to unified mapping, thus improving the semantic sensitivity and security response capability of scheduling decisions.

[0016] 2. This invention introduces decision-driven guiding factors and models their changing trends, and constructs semantic-dominated migration trajectories and load-dominated migration trajectories. This establishes a continuous evolutionary relationship between conflict intensity and scheduling decisions, enabling the scheduling focus to be dynamically adjusted according to changes in conflict state. Simultaneously, by combining potential energy offset gradients to continuously update the decision-driven guiding factors, the scheduling strategy possesses adaptive adjustment and continuous optimization capabilities. Through this series of processes, the offloading scheduling of edge computing tasks is transformed from a static rule-driven process to a dynamic control process based on multi-dimensional feature conflict relationships. This not only improves the response priority of high-risk abnormal events but also enhances the stability and robustness of the system in complex port environments, reduces the probability of misscheduling, and improves overall resource utilization efficiency and safety monitoring reliability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0019] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0020] This invention provides, for example Figure 1 The deep learning-based edge computing task offloading and scheduling method for port safety monitoring, as shown, specifically includes the following steps: S1. Deep learning inference is performed on port safety monitoring videos to extract semantic features of abnormal behavior, the task processing process is quantified to extract task data scale features, and semantic risk potential value and load scale potential value are constructed based on the semantic features of abnormal behavior and the task data scale features, respectively. In this embodiment, S1 specifically includes the following steps: S101. Perform continuous frame parsing on port safety monitoring video and construct a time-series image sequence. Input the time-series image sequence into a deep learning network for inference operations. Output a semantic vector containing abnormal behavior category information and abnormal behavior probability information. Perform category encoding processing on the abnormal behavior category information and numerical normalization processing on the abnormal behavior probability information. Perform weighted combination calculation on the category encoding result and the normalized abnormal behavior probability information to extract the semantic features of abnormal behavior. In practical implementation, the port security monitoring video can be read frame by frame first, splitting the video stream into continuous image frames according to time sequence, and arranging them according to timestamp order to form a time-series image sequence. For example, a monitoring video can be divided into a continuous image sequence at fixed time intervals to retain the time information of behavior changes. Then, the time-series image sequence is input into a trained deep learning network for inference operations. The deep learning network can use a convolutional structure combined with a temporal modeling structure to analyze each frame or the relationship between frames, and output the corresponding abnormal behavior category information and abnormal behavior probability information. For example, it can identify "personnel entering a restricted area" and give the corresponding probability value. Next, the abnormal behavior category information is processed by category encoding, for example, mapping different categories to unique numerical identifiers or vector identifiers. At the same time, the abnormal behavior probability information is processed by numerical normalization, mapping the probability values ​​to a unified numerical range to make different samples comparable. Finally, the category encoding results and the normalized abnormal behavior probability information are combined and calculated according to preset weights, for example, by weighting and superimposing according to the importance of the category and the probability value, thereby generating abnormal behavior semantic features, so that the semantic information includes both the behavior type and the credibility of the behavior.

[0021] Port safety monitoring video refers to continuous video data from camera equipment in the port area, used to record the operating status of personnel, vehicles, and equipment; continuous frame parsing refers to splitting video into discrete image frames according to time sequence; time-series image sequence represents a set of images arranged in time to reflect the process of behavioral changes; deep learning network is a computational model used to extract semantic information from images; inference operation refers to using a trained model to perform forward computation on input data to obtain the output result; abnormal behavior category information represents the identified behavior type, and abnormal behavior probability information represents the confidence level of the occurrence of the behavior; semantic vector is a data structure that unifies category information and probability information; category encoding is the process of converting behavior types into a computable form; numerical normalization is the process of mapping data of different ranges to a unified scale; weighted combination calculation is the process of fusing multiple information according to set weights; abnormal behavior semantic features are semantic description results formed by combining behavior type and occurrence probability for subsequent scheduling decisions.

[0022] S102. Based on the input data stream during the task processing, the data transmission volume, processing latency and data change rate are numerically calculated to construct a multi-dimensional data scale vector. The data of each dimension in the multi-dimensional data scale vector are standardized and weighted fusion calculation is performed according to the distribution weight of each dimension in a unified numerical space to extract the task data scale features. In the implementation process, input data streams can be collected in real time from the task processing chain. Statistical analysis is performed on the data streams during the transmission and processing stages. The total amount of data transmitted per unit time is accumulated to obtain the data transmission volume, and the time interval between data entering processing and processing completion is recorded to obtain the processing latency. Simultaneously, the data change rate is calculated by analyzing the changes in data volume within a continuous time window, for example, by statistically analyzing the difference in data volume between adjacent time slices to obtain the change trend. Subsequently, the data transmission volume, processing latency, and data change rate are uniformly converted into numerical forms and combined to form a multi-dimensional data scale vector. Then, the data in each dimension of the multi-dimensional data scale vector is standardized to map data of different dimensions to a unified range. Next, weights are assigned to each dimension based on historical statistical results or current data distribution. For example, a higher weight is given to processing latency to reflect real-time requirements. The standardized data in each dimension are then fused and calculated according to their weights to form a task data scale characteristic that can comprehensively reflect the task processing scale and processing pressure.

[0023] The input data stream in the task processing process refers to the continuous set of data entering the computing node for processing; data transmission volume represents the total scale of data transmitted per unit time; processing latency represents the time elapsed from data entering processing to completion of processing; data change rate represents the degree of change of data scale in the time dimension; numerical computation is the process of converting the above information into computable numerical values; multidimensional data scale vector is a vector structure composed of various quantitative indicators, used to describe the task scale; standardization processing is mapping data of different dimensions to a unified range to eliminate differences in units; unified numerical space means that the processed data has a consistent comparison scale; distribution weight reflects the importance of each dimension of data in the whole; weighted fusion computation is the process of combining data of each dimension according to weight; task data scale characteristics are a comprehensive quantity formed after fusing multidimensional information to characterize the task processing scale and resource requirements.

[0024] S103. Based on the mapping relationship between the semantic features of abnormal behavior and the scale features of task data in a unified numerical space, perform weight mapping calculation on the semantic features of abnormal behavior to construct the semantic risk potential value, and perform scale transformation calculation on the scale features of task data to construct the load scale potential value.

[0025] In this embodiment, S103 specifically refers to: Based on the mapping relationship between the semantic features of abnormal behavior and the scale features of task data in a unified numerical space, the semantic features of abnormal behavior are decomposed into dimensions and a set of semantic components is constructed. The components in the set of semantic components are normalized numerically and weighted and superimposed according to the distribution weight of each component in the unified numerical space to form a weight mapping result. After mapping the semantic features of abnormal behavior to the same numerical range as the scale features of task data, the semantic features of abnormal behavior can be finely decomposed, breaking down the original semantic information into multiple independent semantic dimensions, such as behavior type intensity, confidence level of behavior occurrence, and behavior change trend, thus forming a set of semantic components. Subsequently, the numerical values ​​of each component in the semantic component set are normalized, compressing the values ​​of different dimensions into a unified range to ensure comparability between components. Based on this, distribution weights are determined according to the distribution characteristics of each semantic component in the unified numerical space; for example, higher weights are assigned to components with larger fluctuations or significant changes, and lower weights are assigned to components with stable changes. Furthermore, each semantic component is further divided according to its corresponding... The distributed weights are weighted and superimposed, for example, different components are accumulated according to their weight ratios to form a weight mapping result with directional and intensity information. This allows semantic information to participate in subsequent processing in a quantified form. The unified numerical space represents the numerical range after scale unification, the mapping relationship represents the correspondence between different features within the numerical range, the dimensionality decomposition process represents the splitting of complex semantics into multiple independent components, the semantic component set represents the data structure composed of multiple semantic components, the numerical normalization process represents mapping each component to a unified scale, the distributed weights represent the contribution ratio of each component in the whole, the weighted superposition calculation represents the cumulative combination of each component according to the weights, and the weight mapping result represents the fused semantic quantification result.

[0026] The weight mapping results are mapped and transformed along the directional gradient of the unified numerical space, and the cumulative calculation is performed by combining the weight contribution relationship of each component in the semantic component set to construct the semantic risk potential value. The weight mapping result can be viewed as a multi-dimensional position vector in a unified numerical space. Based on this, the concept of directional gradient is introduced to describe the changing trend of this position in different semantic directions. The directional gradient distribution is determined by analyzing the increase or decrease of the weight mapping result in each dimension. Then, the weight mapping result is mapped and transformed along the directional gradient of the unified numerical space. For example, directions with significant changing trends are amplified, while directions with gentler changing trends are compressed, thus giving different semantic components directional differences in their influence on the overall expression. Furthermore, by combining the weight contribution relationships of each component in the semantic component set, the effect of each component on the directional gradient is accumulated dimension by dimension. The process involves multiplication, such as assigning higher cumulative values ​​to components with higher weights and larger directional gradients, and lower cumulative values ​​to components with lower weights and smaller directional gradients, ultimately forming a semantic risk potential value. This value is used to characterize the overall intensity of abnormal behavior in the risk dimension. Here, the directional gradient of the unified numerical space represents the directionality and trend of numerical changes in each dimension, the mapping transformation represents the process of redistributing the original values ​​according to the directional gradient, the weight contribution relationship of each component represents the proportion of influence of different semantic components on the overall result, the cumulative calculation represents the process of gradually superimposing each component according to its weight and direction, and the semantic risk potential value represents the risk quantification result after integrating directional and weight information.

[0027] Based on the distribution range of the task data scale characteristics in a unified numerical space, the task data scale characteristics are scale segmented and a scale component set is constructed. The scale of each component in the scale component set is normalized, and the load scale potential energy value is constructed by proportional transformation based on the scale span of each component in the unified numerical space.

[0028] The distribution of task data scale characteristics within a unified numerical space can be divided into intervals. Based on the data density and trend within this space, the continuous distribution can be divided into several scale intervals, thus completing scale segmentation. For example, smaller, more stable data segments can be divided into low-scale intervals, while larger, more dynamic data segments can be divided into high-scale intervals, thereby constructing a scale component set. Subsequently, scale normalization is performed on each component in the scale component set to map data within different scale intervals to a unified range, thus eliminating differences in the original data across different magnitudes. Furthermore, proportional transformation calculations are performed on each component based on its scale span within the unified numerical space. For example, components with larger spans can be amplified to enhance their representation within the overall scale. The impact is reduced by compressing components with smaller spans. Finally, the values ​​of each component after proportional adjustment are integrated to form the load scale potential value, which is used to characterize the load intensity of the task in terms of computing resources. The distribution range of the task data scale characteristics in the unified numerical space represents the value range of the data after standardization. The scale segmentation process represents the process of dividing the data into intervals according to the distribution characteristics. The scale component set represents the data subsets corresponding to different intervals. The scale normalization process represents mapping the data of each interval to a unified scale. The scale span represents the range of change of each component in the numerical space. The proportional transformation calculation represents the process of proportionally adjusting the data according to the scale span. The load scale potential value represents the load intensity quantification result formed after integrating the components.

[0029] S2. Construct a potential energy adversarial domain based on the directional difference between semantic risk potential energy value and load scale potential energy value in a unified reference space, and determine whether there is a conflict between the semantic features of abnormal behavior and the scale features of task data based on the potential energy adversarial domain. In this embodiment, S2 specifically refers to: The semantic risk potential value and the load scale potential value are mapped to a unified reference space. The directional components of the semantic risk potential value and the load scale potential value in each dimension of the unified reference space are extracted respectively, and a directional difference distribution sequence is constructed based on the directional components in each dimension. Semantic risk potential and load scale potential can be mapped to the same coordinate system, making the two types of data comparable under the same reference standard. Then, the semantic risk potential and load scale potential are processed in multiple dimensions, breaking them down into multiple dimensional components. The direction of numerical change in each dimension is analyzed, such as determining whether the value in a certain dimension is increasing or decreasing, thus obtaining the directional components in each dimension. Subsequently, the directions of the semantic risk potential and load scale potential in the corresponding dimensions are compared dimension by dimension, distinguishing between changes in the same direction and changes in opposite directions, and arranging them according to time or dimensional order to form a directional difference distribution sequence. For example, in a certain task, if the semantic risk potential shows an increasing trend in multiple dimensions while the load scale potential shows a decreasing trend in the same dimension, then the corresponding dimension is marked as having the opposite direction, thus forming a complete directional difference sequence. This process can transform the relationship between the two types of potential from a single numerical difference into a structured directional change relationship, providing a basis for subsequent conflict determination.

[0030] The unified reference space representation maps potential energy values ​​from different sources to a common representation space within the same numerical range through scaling transformation, enabling different features to have a unified comparison standard. The directional components of semantic risk potential energy values ​​and load scale potential energy values ​​in each dimension of the unified reference space refer to the directional information reflecting the trend of numerical change in each dimension, used to describe the change attribute of potential energy in that dimension. The directional difference distribution sequence is a set of directional relationships arranged in dimensional or temporal order, used to represent the consistency or opposition distribution of semantic risk potential energy values ​​and load scale potential energy values ​​in each dimension. Through this sequence, the directional difference structure between the two types of potential energy can be clearly characterized, thus providing a basis for subsequently constructing potential energy antagonistic relationships.

[0031] Based on the directional difference distribution sequence, the directional difference between semantic risk potential value and load scale potential value in the unified reference space is divided into intervals. The intervals with consistent directions and the intervals with opposite directions are separated. The directional components of each dimension in the intervals with opposite directions are aggregated to form a continuous interval set. The boundary is determined according to the coverage of the continuous interval set in the unified reference space. The potential energy adversarial domain is constructed based on the numerical interval corresponding to the coverage. After obtaining the directional difference distribution sequence, the directional relationship between semantic risk potential value and load scale potential value is segmented according to dimensional order. Positions with the same direction and positions with opposite directions are marked and distinguished. For example, if multiple dimensions in a continuous dimension show an increase in semantic risk potential value and a decrease in load scale potential value, then the continuous dimension segment is divided into opposite-direction intervals. Subsequently, the dimensional components in all opposite-direction intervals are aggregated, and adjacent or continuous opposite-direction intervals are merged to form a continuous interval set. For example, when multiple adjacent dimensions are all opposite in direction, they are merged into a whole interval. Next, the distribution range of the continuous interval set in a unified reference space is statistically analyzed. The boundaries are determined by determining the start and end positions of each continuous interval in the numerical space, and the numerical intervals corresponding to these boundaries are used as the range of the adversarial region. Finally, a potential adversarial domain is constructed based on these intervals, so that the opposition relationship between semantic risk potential value and load scale potential value is clearly limited to a specific numerical range, thus providing a clear spatial basis for subsequent conflict identification.

[0032] The directional difference between semantic risk potential and load scale potential in a unified reference space represents the difference in their changing trends across dimensions; interval partitioning is the process of segmenting discrete directional difference results according to dimensional continuity; directionally consistent intervals represent the dimensional range where the two types of potential change in the same direction, while directionally opposite intervals represent the dimensional range where the two types of potential change in opposite directions; aggregation processing is the process of merging adjacent or consecutive directionally opposite intervals; a continuous interval set is a set structure composed of multiple consecutive directionally opposite intervals; the coverage of the continuous interval set in the unified reference space represents the distribution position of these intervals in the overall numerical space; boundary determination refers to determining the start and end positions of each continuous interval; the numerical interval corresponding to the coverage represents the specific numerical range formed by the boundary determination; the potential adversarial domain is a region constructed based on these numerical intervals, used to represent the region where there is an adversarial relationship between semantic risk potential and load scale potential.

[0033] Based on the distribution of semantic risk potential energy value and load scale potential energy value in the potential energy adversarial domain, the overlapping interval of semantic risk potential energy value and load scale potential energy value in the potential energy adversarial domain is identified, and the proportion of semantic risk potential energy value and load scale potential energy value in the overlapping interval is calculated. When the proportion of semantic risk potential energy value and load scale potential energy value in the overlapping interval is greater than zero, it is determined that the semantic features of abnormal behavior and the features of task data scale conflict. When the proportion of at least one of semantic risk potential energy value and load scale potential energy value in the overlapping interval is zero, it is determined that the semantic features of abnormal behavior and the features of task data scale do not conflict.

[0034] After determining the adversarial domain, the semantic risk potential value and the load scale potential value are mapped to the corresponding numerical intervals within the domain. The overlapping intervals are identified by comparing their distribution within these intervals. For example, the ranges of the semantic risk potential value and the load scale potential value within the adversarial domain are represented as intervals, and the intersection of these two intervals is determined to obtain the overlapping intervals. Subsequently, the proportions of the semantic risk potential value and the load scale potential value within the overlapping intervals are statistically analyzed. For instance, the proportion of each value falling within the overlapping interval relative to the overall numerical range is calculated. If both are distributed within the overlapping interval, their corresponding proportions are non-zero, indicating a conflict between the abnormal behavior semantic feature and the task data scale feature. If one of them is not distributed within the overlapping interval, its corresponding proportion is zero, indicating no conflict. This method transforms conflict determination into a quantitative judgment based on interval overlap, transforming the relationship between different features from a subjective judgment into a computable process.

[0035] The distribution of semantic risk potential energy values ​​and load scale potential energy values ​​in the adversarial domain indicates the specific value location or range of the two types of potential energy within the adversarial region. The overlapping interval refers to the intersection area where semantic risk potential energy values ​​and load scale potential energy values ​​exist simultaneously within the same numerical range. The proportion of semantic risk potential energy values ​​and load scale potential energy values ​​within the overlapping interval indicates the proportion of each falling into the intersection area, used to measure the degree of overlap between the two in the adversarial domain. A proportion greater than zero indicates that the two types of potential energy have a common coverage area in the adversarial domain, reflecting that they overlap within the same interval. At least one of the proportions being zero indicates that one type of potential energy has not entered the intersection area, indicating that the two do not overlap in the adversarial domain. Through these quantitative indicators, it is possible to clearly determine whether there is a conflict between the semantic features of abnormal behavior and the scale features of task data.

[0036] S3. In the event of a conflict, calculate the potential energy offset gradient based on the offset relationship between the semantic risk potential energy value and the load scale potential energy value, and generate the conflict intensity through nonlinear mapping. In this embodiment, S3 specifically includes the following steps: S301. In the event of a conflict, the semantic risk potential value and the load scale potential value are mapped to a unified reference space. The numerical differences between the semantic risk potential value and the load scale potential value in each dimension of the unified reference space are calculated dimension by dimension and arranged in dimensional order to form a sequence of offset relationships between the semantic risk potential value and the load scale potential value. Once the conflict relationship has been determined, semantic risk potential and load scale potential can be mapped to the same numerical coordinate system, allowing for comparison of the two types of potential at the same scale. Then, the values ​​of the two types of potential in each dimension are compared one by one, and the numerical differences in the corresponding dimensions are calculated to form an offset. For example, if the semantic risk potential is higher than the load scale potential in a certain dimension, a positive offset is formed for that dimension; if the semantic risk potential is lower than the load scale potential in another dimension, a negative offset is formed. Subsequently, the offsets of all dimensions are arranged in dimensional order to form a sequence of offset relationships with a sequential structure. For example, in three-dimensional space, a sequence of three offsets can be obtained. This sequence can completely describe the distribution of differences between the two types of potential in each dimension. Through this dimensional calculation and sequential arrangement, the originally scattered difference information is transformed into a structured data sequence, providing an input basis for subsequent gradient calculation.

[0037] The unified reference space represents a common space formed by mapping potential energy values ​​from different sources to the same numerical range after scaling, making different potential energies comparable. The numerical differences between semantic risk potential energy values ​​and load scale potential energy values ​​in each dimension of the unified reference space represent the degree of numerical deviation between the two types of potential energy in the corresponding dimensions, used to characterize the magnitude of the difference. Dimensional calculation means that the numerical differences are calculated and processed sequentially according to each dimension. The offset relation sequence represents the difference set structure arranged in dimensional order, used to describe the overall offset distribution of semantic risk potential energy values ​​and load scale potential energy values ​​in multidimensional space. This sequence not only reflects the magnitude of the difference, but also retains the dimensional order information, providing continuous input for the subsequent calculation of potential energy offset gradient.

[0038] S302. Perform differential calculation on the numerical differences between adjacent dimensions in the offset relationship sequence to form the offset change between each adjacent dimension. Arrange each offset change according to the dimensional order in the unified reference space. Combine the positional distribution of each offset change in the unified reference space, perform gradient classification processing on the offset change. Divide the offset change corresponding to different change amplitudes into different gradient levels. Reorganize and arrange the offset change according to the gradient level to form the potential energy offset gradient. The offset sequence can be used as input data. The numerical differences between adjacent dimensions in the sequence are compared one by one. The offset change is obtained by calculating the change in offset between two adjacent dimensions. For example, in a three-dimensional offset sequence, if the offset difference between the first and second dimensions is small, while the offset difference between the second and third dimensions is large, then different magnitudes of offset change are formed. Then, all offset changes are arranged according to the dimensional order in the unified reference space. Based on the specific position of each offset change in the unified reference space, its change magnitude is divided. For example, offset changes with small magnitudes are divided into low-level, those with medium magnitudes into medium-level, and those with large magnitudes into high-level. Then, the offset changes are reordered according to the different gradient levels obtained. The offset changes of the same gradient level are grouped together to form a potential energy offset gradient with a hierarchical structure, which transforms the original discrete change information into a gradient expression with a continuous change trend.

[0039] The numerical difference between adjacent dimensions represents the variation in the difference between two adjacent dimensions in the offset relationship sequence; difference calculation refers to the process of obtaining the amount of change by comparing the difference values ​​of adjacent dimensions; the offset change represents the specific numerical value of the difference change between adjacent dimensions; the positional distribution of each offset change in the unified reference space represents the distribution position of these changes in the overall multidimensional space; gradient classification processing represents the process of classifying the offset change according to the magnitude of change; the offset change corresponding to different magnitudes of change represents a set of values ​​with different degrees of change; different gradient levels represent multiple hierarchical intervals formed according to the magnitude of change; recombination and arrangement represent the process of reorganizing the offset change according to the gradient level; the potential energy offset gradient represents the offset structure with continuous variation characteristics formed after gradient classification and recombination, used to characterize the changing trend between semantic risk potential energy value and load scale potential energy value.

[0040] S303. The potential energy offset gradient is segmented according to the numerical interval in the unified reference space. The distribution density of the potential energy offset gradient in each interval is statistically analyzed. Based on the distribution density, the potential energy offset gradient is nonlinearly mapped and calculated. The nonlinear mapping calculation results are accumulated and calculated to generate the conflict intensity.

[0041] In this embodiment, S303 specifically refers to: The potential energy shift gradient is divided into multiple continuous numerical intervals according to the numerical range in the unified reference space, and the potential energy shift gradient is segmented according to the distribution of the potential energy shift gradient in each numerical interval. In the implementation process, the potential energy offset gradient can be mapped to the overall numerical range of a unified reference space. First, the upper and lower boundaries of this range are determined, for example, by statistically analyzing the maximum and minimum values ​​of the potential energy offset gradient to define the complete numerical interval. Then, within this interval, it is divided into multiple adjacent and non-overlapping sub-intervals at fixed intervals, thus forming multiple continuous numerical intervals. Subsequently, the distribution of the potential energy offset gradient within each numerical interval is statistically analyzed, for example, by counting the number of potential energy offset gradients or their concentration within each interval to reflect the aggregation state of data in different intervals. Based on this, the distribution density of the potential energy offset gradient in each interval is analyzed. Segmentation processing, for example, separately identifying densely distributed intervals as key intervals and merging sparsely distributed intervals, divides continuous potential energy shift gradient data into several distinguishable segments, enabling subsequent processing to perform differentiated calculations based on the characteristics of different intervals. Here, the numerical range in the unified reference space represents the overall value range of the potential energy shift gradient, equal interval division means uniformly splitting the interval at fixed intervals, continuous numerical intervals represent multiple adjacent intervals formed after division, the distribution of the potential energy shift gradient in each numerical interval represents the distribution density or concentration of the data in different intervals, and segmentation processing means the process of dividing and classifying the data into intervals based on the distribution.

[0042] For the potential energy shift gradient in each numerical interval, the frequency and distribution concentration of the potential energy shift gradient in each numerical interval are statistically analyzed, and the distribution density sequence of the corresponding numerical interval is constructed based on the frequency and distribution concentration. After dividing the numerical intervals, a statistical analysis is performed on the potential energy shift gradients contained within each interval. First, the number of potential energy shift gradients falling into each interval is counted to reflect the frequency of data occurrence within that interval. For example, if a large number of potential energy shift gradients are found in a certain interval, it indicates that the corresponding range of change is relatively concentrated. Then, the distribution of these potential energy shift gradients within the intervals is further analyzed. For example, by observing whether the distribution of values ​​within the interval is concentrated in a certain local area, the degree of concentration is determined. If most values ​​are concentrated in the middle of the interval or near a certain endpoint, it indicates a high degree of concentration. After obtaining the... After considering frequency and distribution concentration, the two are combined for characterization. For example, intervals with high frequency and high concentration are assigned higher density labels, while intervals with low frequency or scattered distribution are assigned lower density labels. The density labels corresponding to each interval are arranged sequentially according to the numerical interval order to form a distribution density sequence, which is used to describe the distribution characteristics of the potential energy shift gradient throughout the entire numerical range. The frequency and distribution concentration of the potential energy shift gradient in each numerical interval represent the quantity distribution and aggregation state of the data within the interval. The distribution density sequence corresponding to the numerical interval represents the set of density labels arranged in interval order, which is used to characterize the overall distribution structure.

[0043] The potential energy shift gradient is nonlinearly mapped based on the distribution density sequence. The potential energy shift gradient corresponding to different distribution densities is segmented and weighted. The nonlinear mapping calculation results are accumulated segment by segment according to the numerical interval order to generate the conflict intensity.

[0044] After obtaining the distribution density sequence, the potential energy shift gradient is correlated with the distribution density corresponding to each numerical interval. Nonlinear mapping calculations are performed on the potential energy shift gradient based on different distribution densities. For example, an enhanced mapping is applied to the potential energy shift gradient in intervals with high distribution density, further amplifying its numerical variation, while a compressed mapping is applied to the potential energy shift gradient in intervals with low distribution density, making its numerical variation more gradual. This results in nonlinear differences in the overall calculation of the potential energy shift gradient in different intervals. Subsequently, the segmented weights of the potential energy shift gradient corresponding to each interval are adjusted according to the distribution density. For example, higher weights are assigned to high-density intervals, and lower weights to low-density intervals, making the influence of different intervals on the overall result different. Then, the potential energy shift gradient is processed according to the order of the numerical intervals. The potential energy shift gradient after nonlinear mapping and weight adjustment is accumulated segment by segment. The results of each interval are superimposed to form an overall value, which is used to represent the conflict intensity. For example, if high-density intervals appear consecutively, the accumulated value increases faster, reflecting a higher degree of conflict. In this way, the distribution characteristics are combined with the numerical changes, so that the conflict intensity can reflect the comprehensive situation of multidimensional shift. Among them, nonlinear mapping calculation represents the process of nonlinearly adjusting the value according to the distribution density, segmented weight adjustment represents the processing method of assigning different degrees of influence to different intervals, segment-by-segment accumulation calculation represents the process of gradually superimposing the value according to the interval order, and conflict intensity represents the quantitative index formed after integrating the results of all intervals, which is used to characterize the degree of conflict between semantic risk potential energy value and load scale potential energy value.

[0045] S4. Map the conflict intensity to generate decision-driven guiding factors, and construct semantic-dominated migration trajectories and load-dominated migration trajectories based on the changing trends of decision-driven guiding factors. In this embodiment, S4 specifically includes the following steps: S401. Divide the numerical distribution of conflict intensity in a unified reference space into intervals, map the conflict intensity to the corresponding numerical intervals, assign values ​​to the conflict intensity in each numerical interval according to the interval position, and perform numerical transformation of the conflict intensity according to the mapping relationship corresponding to each numerical interval. Combine the transformed values ​​in order according to the arrangement order of the numerical intervals to construct a decision-driven guiding factor. Conflict intensity can be viewed as a set of continuous numerical distributions in a unified reference space. First, the distribution range is divided into intervals, for example, the overall interval is divided into multiple continuous sub-intervals based on the numerical range of conflict intensity. Then, each conflict intensity value is mapped to its corresponding numerical interval. After the interval assignment is completed, the conflict intensity in each numerical interval is graded and assigned a value according to the position of the interval in the overall numerical range. For example, the conflict intensity in the high numerical interval is assigned a higher level value, and the conflict intensity in the low numerical interval is assigned a lower level value. Subsequently, the conflict intensity is numerically transformed according to the pre-set mapping relationship of each numerical interval. For example, the conflict intensity of different levels is converted into corresponding guidance weights, and the transformed values ​​of each interval are arranged and combined in sequence according to the interval order to form a decision-driven guidance factor with a sequential structure. For example, when multiple consecutive high interval values ​​appear, the corresponding guidance factor shows a continuous high value change, which is used for weight adjustment in subsequent scheduling decisions.

[0046] The numerical distribution of conflict intensity in a unified reference space represents the overall range of conflict intensity values ​​at a unified scale; interval division represents the segmentation of this range into multiple continuous intervals; corresponding numerical intervals represent the specific interval location to which each conflict intensity value belongs; hierarchical assignment represents the process of assigning different levels of values ​​to values ​​according to interval location; the mapping relationship between each numerical interval represents the transformation correspondence between different intervals and target values; numerical transformation represents the process of re-numerating the original conflict intensity according to the mapping relationship; the decision-driving guiding factor represents the data sequence formed by combining the transformed values ​​in interval order, used to characterize the degree to which conflict intensity drives subsequent decisions.

[0047] S402. Extract the numerical changes of decision-driven guiding factors in a unified reference space point by point, calculate the direction and magnitude of numerical changes between adjacent positions, classify and label them according to the direction of numerical changes, and classify them according to the distribution range of numerical change magnitude in the unified reference space. Combine and arrange the classification and labeling results with the classification and division results to construct a change trend sequence. The decision-driven guiding factor can be viewed as a set of continuous values ​​arranged sequentially in a unified reference space. Each position in this sequence is extracted point by point, and the values ​​between adjacent positions are compared and analyzed to determine the direction and magnitude of the numerical change. For example, when the value of a later position is higher than that of a previous position, it is marked as an upward direction; when the value of a later position is lower than that of a previous position, it is marked as a downward direction. Simultaneously, the magnitude of the change is determined by comparing the difference between adjacent positions. Then, each position is classified and labeled according to the direction of the numerical change, for example, an upward direction is labeled as one category, and a downward direction as another. Furthermore, the magnitude of the change is divided into intervals based on the distribution of the magnitude of the change in the unified reference space, for example, dividing the magnitude of the change into low-magnitude, medium-magnitude, and high-magnitude intervals. Next, the classification and labeling results are combined with the corresponding grading results of the magnitude of the change, so that each position simultaneously possesses both directional attributes and magnitude levels, for example, forming a combination label of "upward-high magnitude" or "downward-low magnitude". These combination labels are then arranged in their original order to form a trend sequence, used to describe the overall change process of the decision-driven guiding factor.

[0048] The numerical changes of the decision-driven guiding factor in a unified reference space represent the continuous changes of the factor at a unified scale; point-by-point extraction means obtaining each value sequentially according to its position; the direction and magnitude of numerical changes between adjacent positions represent the trend attribute and degree of change of the numerical changes, respectively; classification labels represent classifying the data according to the direction of change; the distribution interval of the magnitude of numerical changes in the unified reference space represents the interval structure that maps the magnitude of change to different numerical ranges; hierarchical division represents the hierarchical division of the magnitude of change according to the interval range; the combined arrangement of classification label results and hierarchical division results represents the process of fusing and encoding directional information and magnitude level; the trend sequence represents the data sequence formed by arranging these combined results in sequence, used to characterize the overall change pattern.

[0049] S403. Based on the direction of change at each position in the trend sequence, divide the trend sequence into intervals. Construct the intervals where the direction of change is biased towards the semantic risk potential value as the semantic-dominated migration trajectory, and construct the intervals where the direction of change is biased towards the load scale potential value as the load-dominated migration trajectory.

[0050] In this embodiment, S403 specifically refers to: The direction of change at each position in the trend sequence is extracted point by point, and the direction of change is classified and marked according to the value sign of the direction of change in the unified reference space. The direction of change is divided into direction marks biased towards semantic risk potential value and direction marks biased towards load scale potential value. Each position in the trend sequence is treated as an independent unit containing information about the direction of change. These positions are extracted point by point in sequence, and the corresponding direction of change identifiers are read. Then, the direction of change is mapped to a symbolic representation in a unified reference space. For example, an upward trend in value is mapped to a positive symbol, and a downward trend in value is mapped to a negative symbol. After completing the symbol mapping, the direction of change is classified and labeled according to the symbol type. For example, the direction of change corresponding to a positive symbol is labeled as a direction biased towards semantic risk potential value, and the direction of change corresponding to a negative symbol is labeled as a direction biased towards load scale potential value. For example, if multiple positive symbols appear consecutively in a sequence, the corresponding position is uniformly labeled as the direction dominated by semantic risk, while if a negative symbol appears, it is labeled as the direction dominated by load scale. In this way, continuous trend information is transformed into a sequence of direction identifiers with clear semantic orientation, providing a directional basis for subsequent migration trajectory construction.

[0051] The direction of change at each position in the trend sequence represents the trend attribute of the numerical change at each position; point-by-point extraction means obtaining the direction information of each position in sequence; the value symbol of the direction of change in the unified reference space means that the direction is uniformly expressed in symbol form; classification label means that the direction is classified according to the symbol type; the direction label biased towards semantic risk potential value indicates that the trend of change points to the direction of semantic risk enhancement; the direction label biased towards load scale potential value indicates that the trend of change points to the direction of load scale enhancement. These labels together constitute the data structure used to describe the direction attribute of the trend of change.

[0052] Based on the distribution of directional markers at continuous positions in the trend sequence, continuous positions with the same directional markers are aggregated to form multiple continuous intervals, and the intervals are divided according to the order of the continuous intervals in the trend sequence. The directional markers in the trend sequence are traversed in their original order. For each position, the continuity of the directional marker is assessed. When adjacent positions have consistent directional markers, these positions are grouped into the same set for aggregation. For example, if multiple consecutive positions have directional markers biased towards semantic risk potential, these positions are merged into a single interval. When a directional marker changes, the current aggregation ends and a new interval division begins. For example, when the bias shifts from semantic risk potential to load scale potential, a new continuous interval is formed. By scanning the entire sequence segment by segment, multiple continuous intervals composed of the same directional markers can be formed, and these intervals are then processed accordingly. The intervals are arranged in the order of the changing trend sequence to complete the interval division. For example, if a semantic risk direction marker interval appears first in a changing trend sequence, and then the load scale direction marker interval appears, multiple consecutive interval sets are formed in sequence. This processing can transform discrete directional change information into a structured interval expression, providing a clear segmentation basis for subsequent migration trajectory construction. The distribution of directional markers at consecutive positions in the changing trend sequence represents the arrangement of directional markers at adjacent positions. Aggregation processing represents the process of merging consecutive positions with the same directional marker. Multiple consecutive intervals represent multiple consecutive regions formed by segments with different directional markers.

[0053] Based on the interval division results, the continuous intervals containing the direction markers of the biased semantic risk potential value are connected in order of position to construct the semantically dominant migration trajectory, and the continuous intervals containing the direction markers of the biased load scale potential value are connected in order of position to construct the load-dominated migration trajectory.

[0054] After dividing the intervals, all continuous intervals are traversed according to their position in the trend sequence. For each continuous interval, it is classified and filtered according to the type of direction markers it contains. All continuous intervals containing direction markers biased towards semantic risk potential value are extracted and connected and combined according to their original position in the sequence. For example, multiple scattered intervals that are all semantic risk directions are spliced ​​together in sequence to form a continuous path, thereby constructing a semantically dominant migration trajectory. At the same time, all continuous intervals containing direction markers biased towards load scale potential value are extracted and connected in the same way to form a load-dominated migration trajectory. For example, if semantic direction intervals appear at the beginning and end of a sequence, while load direction intervals are concentrated in the middle, the corresponding intervals are connected in sequence to generate two independent trajectories. This processing method can integrate the originally scattered direction change intervals into a trajectory structure with continuous path characteristics, so that the evolution process of different dominant factors can be clearly expressed.

[0055] The interval partitioning result represents a set of multiple continuous intervals formed based on the distribution of direction markers; continuous intervals containing direction markers biased towards semantic risk potential value represent a set of intervals where all directions tend towards semantic risk change; semantically dominated migration trajectory represents a path structure formed by connecting these intervals in positional order, used to describe the change process under the dominance of semantic risk; continuous intervals containing direction markers biased towards load scale potential value represent a set of intervals where all directions tend towards load scale change; load-dominated migration trajectory represents a path structure formed by connecting these intervals in sequence, used to describe the change process under the dominance of load scale.

[0056] S5. Based on the positional relationship between decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories, adjust the decision-oriented direction of deep learning in edge computing task offloading scheduling, and update the decision-driven guiding factors based on the potential energy offset gradient.

[0057] In this embodiment, S5 specifically refers to: The positions of decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories are located point by point. The corresponding position intervals of decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories are extracted, and position relationship sequences are constructed according to the distribution order of position intervals in the two types of migration trajectories. In the specific implementation process, the distribution of decision-driven guiding factors in the semantically dominant migration trajectory and the load-dominated migration trajectory can be scanned point by point. First, according to the arrangement order of the guiding factors in the unified reference space, each position is located, and the landing interval of the corresponding position is found in the two types of migration trajectories. For example, when the decision-driven guiding factor of a certain position appears in the first interval of the semantically dominant migration trajectory and the middle interval of the load-dominated migration trajectory, its interval position in the two types of trajectories is recorded respectively. Then, the corresponding intervals of each position in the two types of migration trajectories are paired to form position interval pairs, and these interval pairs are arranged according to the order of the guiding factors in the original sequence. For example, the interval pairs of consecutive positions are connected in sequence to form a structured sequence, thereby constructing a position relationship sequence. This sequence can reflect the relative position changes of the decision-driven guiding factors in the two types of migration trajectories, providing a basis for subsequently determining the scheduling focus direction.

[0058] The position of the decision-driven guiding factor in the semantically dominant migration trajectory and the load-dominated migration trajectory represents the specific distribution location of the guiding factor in the two types of trajectories; point-by-point positioning means that each position is individually identified and labeled according to the sequence order; the corresponding position interval represents the range of intervals that match the same position in the two types of migration trajectories; the distribution order of position intervals represents the order of these intervals in the trajectory; the positional relationship sequence represents the data structure formed by the sequential combination of multiple position interval pairs, which is used to describe the overall positional correspondence between the decision-driven guiding factor in the semantically dominant migration trajectory and the load-dominated migration trajectory.

[0059] Based on the distribution ratio of each position interval in the positional relationship sequence in the semantically dominant migration trajectory and the load-dominated migration trajectory, the decision-driven guidance factor is subjected to direction mapping processing. The decision-driven guidance factor corresponding to the position interval with the distribution ratio biased towards the semantically dominant migration trajectory is adjusted to the semantically biased direction, and the decision-driven guidance factor corresponding to the position interval with the distribution ratio biased towards the load-dominated migration trajectory is adjusted to the load-biased direction, so as to complete the adjustment of the decision-biased direction of deep learning in the offloading and scheduling of edge computing tasks based on positional relationship. The distribution ratio of each position interval in the positional relationship sequence within the semantically dominant migration trajectory and the load-dominated migration trajectory can be calculated. Specifically, the distribution ratio is determined by statistically analyzing the coverage or frequency of the same position interval in the two types of migration trajectories. For example, if the coverage length of the decision-driven guiding factor in the semantically dominant migration trajectory is greater than its coverage length in the load-dominated migration trajectory, then the position interval is determined to be semantically dominant. Subsequently, the distribution ratio corresponding to each position interval is compared, and the decision-driven guiding factor is directionally mapped based on the comparison results. For example, the decision-driven guiding factor corresponding to the semantically dominant interval is marked as the semantically oriented direction, and the decision-driven guiding factor corresponding to the load-dominated interval is marked as the load-oriented direction. For example, in a set of continuous intervals, if the first half of the interval is semantically dominant and the second half is load-dominated, then the corresponding guiding factor is adjusted to the semantically oriented direction in the first half and to the load-oriented direction in the second half. Through this interval-by-interval mapping process, the decision-oriented direction can be dynamically adjusted as the position changes.

[0060] The distribution ratio of each position interval in the positional relationship sequence within the semantically dominated migration trajectory and the load-dominated migration trajectory represents the relative proportion of the same position interval in the two types of trajectories; the direction mapping process represents the process of adjusting the direction of decision-driven guiding factors according to the distribution ratio; position intervals with a distribution ratio biased towards the semantically dominated migration trajectory indicate that semantically related distributions dominate within that interval; semantically biased direction indicates that the decision tends to prioritize the direction of semantic risk factors; position intervals with a distribution ratio biased towards the load-dominated migration trajectory indicate that load-related distributions dominate within that interval; load biased direction indicates that the decision tends to prioritize the direction of computational load factors; the adjustment of the decision-oriented direction of deep learning in edge computing task offloading scheduling based on positional relationships represents directional control of scheduling decisions through the relationship between position intervals and distribution ratios.

[0061] The potential energy shift gradient is mapped to the corresponding position interval in the unified reference space. The potential energy shift gradient and the decision-driven guidance factor are matched point by point in the same position interval. The decision-driven guidance factor is numerically corrected according to the change of the potential energy shift gradient in each position interval, thus completing the update of the decision-driven guidance factor based on the potential energy shift gradient.

[0062] Potential energy shift gradients can be mapped according to their positional distribution in a unified reference space, establishing a one-to-one correspondence between each potential energy shift gradient value and its corresponding positional interval. Then, within the same positional interval, the potential energy shift gradient and the decision-driven guidance factor are matched point-by-point. For example, for a given positional interval, the potential energy shift gradient values ​​within that interval are aligned with the decision-driven guidance factor values ​​within the same interval. After matching, the decision-driven guidance factor is numerically corrected based on the magnitude of change in the potential energy shift gradient within that positional interval. For instance, when the potential energy shift gradient exhibits a large magnitude of change within a certain interval, the decision-driven guidance factor at the corresponding position is strengthened; conversely, when the magnitude of change is small, the decision-driven guidance factor at the corresponding position is weakened. For example, if the potential energy shift gradient continuously increases across multiple consecutive positional intervals, the decision-driven guidance factors within these intervals are progressively increased to ensure their trend aligns with the potential energy shift gradient, thereby achieving dynamic updates to the decision-driven guidance factor.

[0063] Mapping the potential energy offset gradient to the corresponding position interval in the unified reference space means allocating gradient values ​​according to spatial location, establishing a correspondence between them and the position intervals; point-by-point matching within the same position interval means performing one-to-one alignment processing on two data sequences within the same interval; the magnitude of change of the potential energy offset gradient within each position interval means the degree of change of the gradient value within the interval; numerical correction processing means adjusting the value of the decision-driven guiding factor according to the magnitude of change; updating the decision-driven guiding factor based on the potential energy offset gradient means continuously adjusting the guiding factor using gradient change information so that it can reflect the current state change.

[0064] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0065] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0066] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0067] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0068] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0069] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0070] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based port safety monitoring edge computing task offloading scheduling method, characterized in that, Specifically comprising the following steps: S1, deep learning inference is performed on the port safety monitoring video to extract abnormal behavior semantic features, the task processing process is quantified to extract task data scale features, and semantic risk potential values and load scale potential values are constructed based on the abnormal behavior semantic features and the task data scale features respectively; S2, based on the direction difference of the semantic risk potential values and the load scale potential values in the unified reference space, a potential energy confrontation domain is constructed, and whether the abnormal behavior semantic features and the task data scale features conflict is determined according to the potential energy confrontation domain; S3, in the case of conflict, the potential energy offset gradient is calculated based on the offset relationship between the semantic risk potential values and the load scale potential values, and the conflict intensity is generated through nonlinear mapping; S4, the conflict intensity is mapped and converted to generate a decision driving guide factor, and semantic dominant migration trajectories and load dominant migration trajectories are constructed according to the change trend of the decision driving guide factor; S5, based on the position relationship of the decision driving guide factor in the semantic dominant migration trajectory and the load dominant migration trajectory, the decision emphasis direction of deep learning in edge computing task offloading scheduling is adjusted, and the decision driving guide factor is updated based on the potential energy offset gradient.

2. The deep learning based port safety monitoring edge computing task offloading scheduling method according to claim 1, characterized in that, S1 specifically comprises the following steps: S101, the port safety monitoring video is analyzed in continuous frames and a time sequence image sequence is constructed, the time sequence image sequence is input into a deep learning network for inference operation, a semantic vector containing abnormal behavior category information and abnormal behavior probability information is output, the abnormal behavior category information is processed by category coding, the abnormal behavior probability information is processed by numerical normalization, and the category coding result and the normalized abnormal behavior probability information are combined and calculated to extract abnormal behavior semantic features; S102, based on the input data stream in the task processing process, the data transmission volume, the processing time delay and the data change rate are calculated, a multi-dimensional data scale vector is constructed, each dimension data in the multi-dimensional data scale vector is standardized, and the dimension data in the unified numerical space is weighted and fused to extract the task data scale features; S103, based on the mapping relationship of the abnormal behavior semantic features and the task data scale features in the unified numerical space, the abnormal behavior semantic features are weighted and mapped to construct the semantic risk potential values, and the task data scale features are scaled to construct the load scale potential values.

3. The deep learning based port safety monitoring edge computing task offloading scheduling method according to claim 2, characterized in that, S103 specifically comprises: Based on the mapping relationship of the abnormal behavior semantic features and the task data scale features in the unified numerical space, the abnormal behavior semantic features are dimensionally decomposed and a semantic component set is constructed, each component in the semantic component set is normalized, and the weighted mapping result is formed by weighted superposition calculation according to the distribution weight of each component in the unified numerical space; The weighted mapping result is mapped and converted along the direction gradient of the unified numerical space, and the semantic risk potential values are constructed by combining the weight contribution relationship of each component in the semantic component set and accumulating calculation; Based on the distribution range of the task data scale characteristics in a unified numerical space, the task data scale characteristics are scale segmented and a scale component set is constructed. The scale of each component in the scale component set is normalized, and the load scale potential energy value is constructed by proportional transformation based on the scale span of each component in the unified numerical space.

4. The deep learning based port safety monitoring edge computing task offloading scheduling method according to claim 1, characterized in that, S2 specifically refers to: The semantic risk potential value and the load scale potential value are mapped to a unified reference space. The directional components of the semantic risk potential value and the load scale potential value in each dimension of the unified reference space are extracted respectively, and a directional difference distribution sequence is constructed based on the directional components in each dimension. Based on the directional difference distribution sequence, the directional difference between semantic risk potential value and load scale potential value in the unified reference space is divided into intervals. The intervals with consistent directions and the intervals with opposite directions are separated. The directional components of each dimension in the intervals with opposite directions are aggregated to form a continuous interval set. The boundary is determined according to the coverage of the continuous interval set in the unified reference space. The potential energy adversarial domain is constructed based on the numerical interval corresponding to the coverage. Based on the distribution of semantic risk potential energy value and load scale potential energy value in the potential energy adversarial domain, the overlapping interval of semantic risk potential energy value and load scale potential energy value in the potential energy adversarial domain is identified, and the proportion of semantic risk potential energy value and load scale potential energy value in the overlapping interval is calculated. When the proportion of semantic risk potential energy value and load scale potential energy value in the overlapping interval is greater than zero, it is determined that the semantic features of abnormal behavior and the features of task data scale conflict. When the proportion of at least one of semantic risk potential energy value and load scale potential energy value in the overlapping interval is zero, it is determined that the semantic features of abnormal behavior and the features of task data scale do not conflict.

5. The deep learning based port safety monitoring edge computing task offloading scheduling method according to claim 1, characterized in that, S3 specifically includes the following steps: S301. In the event of a conflict, the semantic risk potential value and the load scale potential value are mapped to a unified reference space. The numerical differences between the semantic risk potential value and the load scale potential value in each dimension of the unified reference space are calculated dimension by dimension and arranged in dimensional order to form a sequence of offset relationships between the semantic risk potential value and the load scale potential value. S302. Perform differential calculation on the numerical differences between adjacent dimensions in the offset relationship sequence to form the offset change between each adjacent dimension. Arrange each offset change according to the dimensional order in the unified reference space. Combine the positional distribution of each offset change in the unified reference space, perform gradient classification processing on the offset change. Divide the offset change corresponding to different change amplitudes into different gradient levels. Reorganize and arrange the offset change according to the gradient level to form the potential energy offset gradient. S303. The potential energy offset gradient is segmented according to the numerical interval in the unified reference space. The distribution density of the potential energy offset gradient in each interval is statistically analyzed. Based on the distribution density, the potential energy offset gradient is nonlinearly mapped and calculated. The nonlinear mapping calculation results are accumulated and calculated to generate the conflict intensity.

6. The deep learning based port safety monitoring edge computing task offloading scheduling method according to claim 5, characterized in that, S303 specifically refers to: The potential energy shift gradient is divided into multiple continuous numerical intervals according to the numerical range in the unified reference space, and the potential energy shift gradient is segmented according to the distribution of the potential energy shift gradient in each numerical interval. For the potential energy shift gradient in each numerical interval, the frequency and distribution concentration of the potential energy shift gradient in each numerical interval are statistically analyzed, and the distribution density sequence of the corresponding numerical interval is constructed based on the frequency and distribution concentration. The potential energy shift gradient is nonlinearly mapped based on the distribution density sequence. The potential energy shift gradient corresponding to different distribution densities is segmented and weighted. The nonlinear mapping calculation results are accumulated segment by segment according to the numerical interval order to generate the conflict intensity.

7. The deep learning based port safety monitoring edge computing task offloading scheduling method according to claim 1, characterized in that, S4 specifically includes the following steps: S401. Divide the numerical distribution of conflict intensity in a unified reference space into intervals, map the conflict intensity to the corresponding numerical intervals, assign values ​​to the conflict intensity in each numerical interval according to the interval position, and perform numerical transformation of the conflict intensity according to the mapping relationship corresponding to each numerical interval. Combine the transformed values ​​in order according to the arrangement order of the numerical intervals to construct a decision-driven guiding factor. S402. Extract the numerical changes of decision-driven guiding factors in a unified reference space point by point, calculate the direction and magnitude of numerical changes between adjacent positions, classify and label them according to the direction of numerical changes, and classify them according to the distribution range of numerical change magnitude in the unified reference space. Combine and arrange the classification and labeling results with the classification and division results to construct a change trend sequence. S403. Based on the direction of change at each position in the trend sequence, divide the trend sequence into intervals. Construct the intervals where the direction of change is biased towards the semantic risk potential value as the semantic-dominated migration trajectory, and construct the intervals where the direction of change is biased towards the load scale potential value as the load-dominated migration trajectory.

8. The deep learning based port safety monitoring edge computing task offloading scheduling method according to claim 7, characterized in that, S403 specifically refers to: The direction of change at each position in the trend sequence is extracted point by point, and the direction of change is classified and marked according to the value sign of the direction of change in the unified reference space. The direction of change is divided into direction marks biased towards semantic risk potential value and direction marks biased towards load scale potential value. Based on the distribution of directional markers at continuous positions in the trend sequence, continuous positions with the same directional markers are aggregated to form multiple continuous intervals, and the intervals are divided according to the order of the continuous intervals in the trend sequence. Based on the interval division results, the continuous intervals containing the direction markers of the biased semantic risk potential value are connected in order of position to construct the semantically dominant migration trajectory, and the continuous intervals containing the direction markers of the biased load scale potential value are connected in order of position to construct the load-dominated migration trajectory. 9.The deep learning based port security monitoring edge computing task offloading scheduling method of claim 1, wherein, S5 specifically refers to: The positions of decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories are located point by point. The corresponding position intervals of decision-driven guiding factors in semantic-dominated migration trajectories and load-dominated migration trajectories are extracted, and position relationship sequences are constructed according to the distribution order of position intervals in the two types of migration trajectories. Based on the distribution ratio of each position interval in the positional relationship sequence in the semantically dominant migration trajectory and the load-dominated migration trajectory, the decision-driven guidance factor is subjected to direction mapping processing. The decision-driven guidance factor corresponding to the position interval with the distribution ratio biased towards the semantically dominant migration trajectory is adjusted to the semantically biased direction, and the decision-driven guidance factor corresponding to the position interval with the distribution ratio biased towards the load-dominated migration trajectory is adjusted to the load-biased direction, so as to complete the adjustment of the decision-biased direction of deep learning in the offloading and scheduling of edge computing tasks based on positional relationship. The potential energy shift gradient is mapped to the corresponding position interval in the unified reference space. The potential energy shift gradient and the decision-driven guidance factor are matched point by point in the same position interval. The decision-driven guidance factor is numerically corrected according to the change of the potential energy shift gradient in each position interval, thus completing the update of the decision-driven guidance factor based on the potential energy shift gradient.

Images