Real-time hedge-oriented airborne heterogeneous edge computing terminal

Abstract

This invention discloses an airborne heterogeneous edge computing terminal for real-time risk avoidance, comprising: a data acquisition and preprocessing module to obtain synchronous situation monitoring data; a masking module to obtain masked situation monitoring data; a situational hidden state reconstruction module to output a situational hidden state reconstruction model; a state reconstruction output module to obtain a normalized situational state reconstruction feature vector set; a risk-sensitive feature module to obtain a risk-sensitive feature vector set; a mapping scheduling control module to generate risk avoidance edge scheduling control vectors based on the risk-sensitive feature vector set through the mapping scheduling control model; and a drive execution module to merge the situation monitoring data acquired after the action with the synchronous situation monitoring data as the synchronous situation monitoring data for the next cycle, achieving closed-loop iteration. This invention achieves structural consistency reconstruction for airborne multi-source heterogeneous situational information, improving the sensitivity and robustness of risk identification.

Description

Technical Field

[0001] This invention relates to the field of heterogeneous edge computing technology, and in particular to an airborne heterogeneous edge computing terminal for real-time risk avoidance. Background Technology

[0002] With the continuous improvement of airborne intelligence, key functions such as mission planning, perception fusion, and avoidance control increasingly rely on the real-time computing capabilities of airborne computing platforms. Traditional airborne processing architectures are mostly based on relatively fixed single-node or weakly cooperative processing units, which are difficult to adaptively adjust to the computing power and communication pressure brought by sudden threats, dynamic environmental disturbances, and multi-task concurrency. Currently, many systems still rely on preset path planning or static resource allocation models. Under conditions of high-maneuverability flight, complex weather, and significant link fluctuations, the computing latency, mission failure rate, and reconstruction capability often fail to meet the stringent response requirements for real-time risk avoidance.

[0003] Existing airborne edge computing frameworks typically suffer from three main weaknesses. First, faced with heterogeneous situational data from multiple sources of sensors, communication links, and computing nodes, current systems generally employ simple filtering, statistical compression, or single-channel anomaly detection methods. They lack a unified modeling capability for the correlations between multiple variables and the topology, making it difficult for the system to accurately extract key factors influencing risk avoidance decisions from high-dimensional situational information. Second, current common scheduling strategies often fail to simultaneously address task deadlines, communication uncertainties caused by link fluctuations, and the time-varying performance characteristics of heterogeneous computing nodes. When link jitter is severe, computing resources are limited, or the threat situation suddenly deteriorates, traditional scheduling methods can easily lead to task blocking, amplified latency, or even failure of critical risk avoidance tasks. Third, existing airborne systems largely rely on manually set safety rules or fixed models trained offline. In dynamic risk-changing scenarios, they lack effective closed-loop feedback and adaptive adjustment mechanisms, resulting in slow system responses to sudden events such as link failures, computing power reduction, sensor anomalies, or distortion of local computing branches. This makes it difficult to meet the stability and continuity requirements of real-time risk avoidance. Summary of the Invention

[0004] One objective of this invention is to propose an airborne heterogeneous edge computing terminal for real-time risk avoidance. This invention achieves structural consistency reconstruction for airborne multi-source heterogeneous situational information, thereby improving the sensitivity and robustness of risk identification.

[0005] An airborne heterogeneous edge computing terminal for real-time risk avoidance according to an embodiment of the present invention includes: The data acquisition and preprocessing module acquires and preprocesses real-time situational data from the airborne heterogeneous edge computing platform to obtain synchronous situational monitoring data. The mask module performs noise suppression processing on the synchronous situation monitoring data and randomly selects a complete time window or a complete computation branch according to the task-resource-link topology relationship to generate a risk-related mask, thereby obtaining the masked situation monitoring data; The situational hidden state reconstruction module trains the mask contrast learning network using masked situational monitoring data and corresponding unmasked situational monitoring data, and outputs the situational hidden state reconstruction model. The state reconstruction output module inputs the synchronous situation monitoring data into the situation hidden state reconstruction model to obtain a set of normalized situation state reconstruction feature vectors. The risk-sensitive feature module inputs the set of normalized situational state reconstruction feature vectors into the hierarchical risk avoidance and focus enhancement module, which sequentially generates global task risk avoidance and focus features, node link risk avoidance and focus features, and micro-area risk mutation and focus features, and then performs weighted fusion to obtain the set of risk-sensitive feature vectors. The mapping scheduling control module generates risk-averse edge scheduling control vectors based on the risk-sensitive feature vector set through the mapping scheduling control model. The drive execution module converts the edge avoidance scheduling control vector into heterogeneous computing power control signals, link service quality control signals, and avoidance guidance signals. This drives the airborne heterogeneous edge computing platform to complete task unloading, resource allocation, link scheduling, and avoidance guidance actions. The module then merges the situation monitoring data collected after the actions with the synchronous situation monitoring data as the synchronous situation monitoring data for the next cycle, thus achieving closed-loop iteration.

[0006] Optionally, the real-time situational data preprocessing includes: The system collects flight status data, navigation and positioning data, environmental and threat perception data, link communication data, computing resource monitoring data, and task load data. It performs time synchronization, sampling rate alignment, missing data completion, and amplitude normalization on the collected data. It establishes a unified timestamp index for multi-source asynchronous data, and uses the sampling time index as the primary key to concatenate data from different sources into a multivariate channel vector with the same sampling time. It also performs threshold pruning and monotonicity consistency verification on outliers and outputs synchronized situational monitoring data.

[0007] Optionally, the mask module includes: Noise suppression processing is performed on the synchronous situation monitoring dataset to obtain a denoised synchronous situation monitoring dataset. Based on the task-resource-link topology relationship, construct a set of topology mappings; Generate risk-related masks under the constraints of the topological mapping set; Based on the denoised synchronous situational monitoring dataset and the risk-related mask, masked situational monitoring data is generated.

[0008] Optionally, the situational hidden state reconstruction module includes: A mask contrast learning network is constructed, with the masked situation monitoring data as input and the output mask representation vector set. At the same time, a target representation network is constructed, with the unmasked situation monitoring data as input and the output target representation vector set. The mask representation vector set and the target representation vector set are normalized so that each mask representation vector and target representation vector is transformed into a unit norm representation vector. Construct a comparison similarity matrix based on the unit norm representation vector; Based on the contrast similarity matrix, a mask contrast learning loss function is constructed. The gradient optimization method is used to minimize the mask contrast learning loss function. When the training process meets the preset training termination condition, the hidden state reconstruction model is obtained.

[0009] Optionally, the state reconstruction output module includes: The synchronous situation monitoring dataset is used as the input to the situation hidden state reconstruction model. The situation hidden state reconstruction model is used to perform forward feature mapping operation on the synchronous situation monitoring dataset to obtain the situation state reconstruction feature vector set. The components of each dimension in the situation state reconstruction feature vector set are uniformly normalized to obtain the normalized situation state reconstruction feature vector set.

[0010] Optionally, the risk-sensitive feature module includes: The set of normalized situational state reconstructed feature vectors is used as the input to the hierarchical risk avoidance focusing enhancement module; The topology mapping set is used to constrain the feature organization method of the hierarchical risk avoidance and focus enhancement module. The hierarchical risk avoidance and focus enhancement module establishes the correspondence between the task subgraph index, the computing node index and the link index based on the topology mapping set. The task load features, resource features and link features in the normalized situation state reconstruction feature vector are grouped and aggregated on different computing branches. Within the layered risk avoidance and focus enhancement module, for each sampling time index, the corresponding normalized situation state reconstruction feature vector is input into the global task risk avoidance and focus feature generation operator. The global task risk avoidance and focus feature generation operator forms the global task risk avoidance and focus feature vector by applying the global risk avoidance and focus gating vector to the normalized situation state reconstruction feature vector. Within the layered risk avoidance and focus enhancement module, a node link risk avoidance and focus feature generation operator is constructed based on the topology mapping set. The node link risk avoidance and focus feature generation operator forms the branch aggregation scalar of each computation branch. The branch aggregation scalar is normalized and used as the branch focus weight. The normalized situation state reconstruction feature vector is weighted and aggregated in the branch dimension in combination with the branch feature mapping of each computation branch to obtain the node link risk avoidance and focus feature vector. Within the layered risk avoidance and focus enhancement module, the difference between the normalized situation state reconstruction feature vector at the current sampling time and the normalized situation state reconstruction feature vector at the previous sampling time is calculated to form the micro-area risk mutation residual vector. The micro-area risk mutation residual vector is then subjected to a gating operation to generate the micro-area risk mutation focus feature vector. The risk-sensitive feature vector is generated by weighted fusion of the global task risk avoidance focus feature vector, the node link risk avoidance focus feature vector and the micro-area risk mutation focus feature vector at the same sampling time. The risk-sensitive feature vectors obtained at each sampling time are summarized in the order of the sampling time index to form a risk-sensitive feature vector set.

[0011] Optionally, the mapping scheduling control module includes: A mapping scheduling control model is constructed, and the risk-sensitive feature vector set is input into the mapping scheduling control model in the order of sampling time index. The mapping scheduling control model performs feature mapping and scheduling decision calculation on the risk-sensitive feature vector corresponding to each sampling time index, and generates a risk avoidance edge scheduling control vector that corresponds one-to-one with the sampling time index. The task unloading and placement strategy is determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model; The heterogeneous resource configuration parameters are determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model; The link service quality and routing strategy are determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model.

[0012] Optionally, the rules for determining the task unloading and placement strategy, heterogeneous resource configuration parameters, and link service quality and routing strategy include: When the proportion of feature components in the risk-sensitive feature vector that reflect increased threat proximity, reduced track safety margin, or tightened risk avoidance calculation deadline exceeds the first proportion threshold, the local priority of critical risk avoidance tasks is increased and they are placed on computing power nodes that meet deterministic latency, while at least one redundant execution mechanism is enabled. When the proportion of feature components in the risk-sensitive feature vector that reflect the increase in link packet loss rate, jitter, or bandwidth decrease is greater than the second proportion threshold, the proportion of external offloading is reduced and transmission-sensitive tasks are migrated to local or near-end nodes where the link converges faster. At the same time, the queue priority, bandwidth reservation ratio, and forward error correction redundancy of critical data streams are increased. When the proportion of feature components in the risk-sensitive feature vector that reflect computing power temperature rise approaching the upper limit, power consumption margin decrease, or cache / queue congestion exceeds the third proportion threshold, load migration across heterogeneous devices is triggered, migrating high-parallelism tasks to GPUs / accelerators and time-sensitive tasks to CPU real-time cores or FPGA pipelines. Resource allocation hysteresis and rate limiting mechanisms are introduced to limit the number of task migrations and the frequency / quota change amplitude per unit time, thereby suppressing scheduling oscillations. When the risk-sensitive feature vector reflects a stable situation and sufficient security margin, an energy consumption optimization strategy is implemented to reduce the resource quota for non-critical tasks and release the reserved bandwidth of the link for background updates or cache prefetching.

[0013] The beneficial effects of this invention are: This invention achieves structural consistency reconstruction for airborne multi-source heterogeneous situational information, improving the sensitivity and robustness of risk identification. It introduces a topological constraint mask contrast learning mechanism in airborne edge computing scenarios, explicitly encoding the task graph, link topology, and computing node structure into the sample mask strategy. This enables the model to learn structural risk patterns that cannot be characterized by existing technologies, such as the co-occurrence of risks in the same branch and the correlation of distortion in the same time period. By aligning the masked view and the unmasked view in the same feature space, it achieves high-confidence reconstruction of link packet loss, local node down-frequency, and random disturbances of intermittent sensor distortion, and maintains the correct dependencies between variables in the reconstructed features.

[0014] The hierarchical risk avoidance focusing model proposed in this invention utilizes three types of operators: global focusing gating, branch-level focusing weights, and micro-regional mutation gating. Under a unified framework, it decouples risk factors by scale and selectively amplifies them, enabling the model to simultaneously focus on long-term accumulated scheduling pressure and transient local risks.

[0015] The mapping scheduling control model constructed in this invention can map risk-sensitive features into a set of continuously adjustable and decomposable scheduling vectors, realizing cross-level linkage decision-making that was difficult for previous airborne systems to achieve. It can explicitly reflect the coupling relationship between risk distribution in task dependency structure, computing node operating status and link quality, and automatically concentrate resources to threat-related nodes, move loads out of restricted nodes, and detour task paths from unstable links. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an airborne heterogeneous edge computing terminal for real-time risk avoidance proposed in this invention. Detailed Implementation

[0017] Example 1: Reference Figure 1 An airborne heterogeneous edge computing terminal for real-time risk avoidance includes: The data acquisition and preprocessing module acquires and preprocesses real-time situational data from the airborne heterogeneous edge computing platform to obtain synchronous situational monitoring data. The mask module performs noise suppression processing on the synchronous situation monitoring data and randomly selects a complete time window or a complete computation branch according to the task-resource-link topology relationship to generate a risk-related mask, thereby obtaining the masked situation monitoring data; The situational hidden state reconstruction module trains the mask contrast learning network using masked situational monitoring data and corresponding unmasked situational monitoring data, and outputs the situational hidden state reconstruction model. The state reconstruction output module inputs the synchronous situation monitoring data into the situation hidden state reconstruction model to obtain a set of normalized situation state reconstruction feature vectors. The risk-sensitive feature module inputs the set of normalized situational state reconstruction feature vectors into the hierarchical risk avoidance and focus enhancement module, which sequentially generates global task risk avoidance and focus features, node link risk avoidance and focus features, and micro-area risk mutation and focus features, and then performs weighted fusion to obtain the set of risk-sensitive feature vectors. The mapping scheduling control module generates risk-averse edge scheduling control vectors based on the risk-sensitive feature vector set through the mapping scheduling control model. The drive execution module converts the edge avoidance scheduling control vector into heterogeneous computing power control signals, link service quality control signals, and avoidance guidance signals. This drives the airborne heterogeneous edge computing platform to complete task unloading, resource allocation, link scheduling, and avoidance guidance actions. The module then merges the situation monitoring data collected after the actions with the synchronous situation monitoring data as the synchronous situation monitoring data for the next cycle, thus achieving closed-loop iteration.

[0018] In this embodiment, real-time situational data preprocessing includes: The system collects flight status data, navigation and positioning data, environmental and threat perception data, link communication data, computing resource monitoring data, and task load data. It performs time synchronization, sampling rate alignment, missing data completion, and amplitude normalization on the collected data. It establishes a unified timestamp index for multi-source asynchronous data, and uses the sampling time index as the primary key to concatenate data from different sources into a multivariate channel vector with the same sampling time. It also performs threshold pruning and monotonicity consistency verification on outliers and outputs synchronized situational monitoring data.

[0019] In Example 1, flight status data includes inertial measurement unit attitude angles, airspeed, altitude, overload, and control surface commands; navigation and positioning data includes GNSS position, waypoint sequence, and heading angle; environmental and threat perception data includes target observations from airborne radar / electro-optical sensors, terrain database matching residuals, weather echo intensity, and turbulence indications; link communication data includes round-trip delay, jitter, packet loss rate, available bandwidth, and channel occupancy for each data link; computing resource monitoring data includes CPU / GPU / FPGA / AI accelerator utilization, temperature, power consumption, cache / video memory usage, queue length, and task queuing delay; and task load data includes task graph node arrival rate, message size, deadline, priority, and security level.

[0020] In this embodiment, the mask module includes: Noise suppression processing is performed on the synchronous situation monitoring dataset to obtain a denoised synchronous situation monitoring dataset. Based on the task-resource-link topology relationship, construct a set of topology mappings; Generate risk-related masks under the constraints of the topological mapping set; Based on the denoised synchronous situational monitoring dataset and the risk-related mask, masked situational monitoring data is generated.

[0021] In Example 1, for noise suppression processing, the local median and local absolute deviation are calculated within a sliding window for each variable channel as a robust metric. If the deviation of a data cell from the local median exceeds the product of a threshold coefficient and the local absolute deviation, the local median is used instead; otherwise, the original value is maintained to suppress spike noise caused by electromagnetic interference, machine vibration, and sudden jitter in the link.

[0022] The set of topology maps consists of several subsets of topology maps, each subset corresponding to a computational branch or a task subgraph, containing elements related to that branch or task subgraph. Figure 1 A corresponding variable channel index, compute node index, and link index are used to characterize task dependencies, heterogeneous computing power constraints, and link transmission constraints.

[0023] Risk-related masks include two types: full-time window masks and full-computation branch masks. Full-time window masks mask all data units of the selected variable channel set within the window at a randomly determined start time and window length. Full-computation branch masks mask all data units for the entire time period on the variable channels corresponding to randomly selected computing nodes and their input / output links. This ensures that the masks conform to the risk-related characteristics of common distortion in the same branch / common occlusion in the same time period, and generates a set of masked situation monitoring data.

[0024] In this embodiment, the situational hidden state reconstruction module includes: A mask contrast learning network is constructed, with the masked situation monitoring data as input and the output mask representation vector set. At the same time, a target representation network is constructed, with the unmasked situation monitoring data as input and the output target representation vector set. The mask representation vector set and the target representation vector set are normalized so that each mask representation vector and target representation vector is transformed into a unit norm representation vector. Construct a comparison similarity matrix based on the unit norm representation vector; Based on the contrast similarity matrix, a mask contrast learning loss function is constructed. The gradient optimization method is used to minimize the mask contrast learning loss function. When the training process meets the preset training termination condition, the hidden state reconstruction model is obtained.

[0025] In Example 1, the mask contrast learning network and the target representation network perform joint feature mapping on the multivariable channel data under the same sampling time index, ensuring that the mask view and the target view are comparable in the same semantic space.

[0026] The elements of the contrast similarity matrix are composed of the dot product similarity between the mask representation vector and the target representation vector, and a temperature coefficient is introduced to adjust the smoothness of the similarity distribution. The mask contrast learning loss function includes a positive sample pair similarity maximization term and a hard negative sample separation term across time / branch. Hard negative samples are composed of sample pairs with similar task load or similar link quality at different sampling times, which improves the model's ability to distinguish between similar appearances but different risks.

[0027] Training termination conditions include the loss decreasing below a threshold, the consistency of the validation set reconstruction meeting a threshold, or the number of training rounds reaching the upper limit, and the output being a situational hidden state reconstruction model for online inference.

[0028] In this embodiment, the state reconstruction output module includes: The synchronous situation monitoring dataset is used as the input to the situation hidden state reconstruction model. The situation hidden state reconstruction model is used to perform forward feature mapping operation on the synchronous situation monitoring dataset to obtain the situation state reconstruction feature vector set. The components of each dimension in the situation state reconstruction feature vector set are uniformly normalized to obtain the normalized situation state reconstruction feature vector set.

[0029] In Example 1, the set of situational state reconstruction feature vectors consists of multiple fixed-dimensional vectors that correspond one-to-one with the sampling time index. Each vector is determined by the flight state, threat observation, link quality, computing load, and task arrival characteristics at the same sampling time.

[0030] Consistency normalization processing includes mean-variance online update based on sliding time window, group normalization based on dimensional grouping, and weighted normalization based on task priority, so that the differences in dimensions of different sensors and the differences in heterogeneous computing power measurement can be compared in the same representation space; and norm constraints and drift suppression are applied to the normalized vectors to maintain representation stability under strong maneuvering, strong interference or sudden temperature rise of nodes.

[0031] In this embodiment, the risk-sensitive feature module includes: The set of normalized situational state reconstructed feature vectors is used as the input to the hierarchical risk avoidance focusing enhancement module; The topology mapping set is used to constrain the feature organization method of the hierarchical risk avoidance and focus enhancement module. The hierarchical risk avoidance and focus enhancement module establishes the correspondence between the task subgraph index, the computing node index and the link index based on the topology mapping set. The task load features, resource features and link features in the normalized situation state reconstruction feature vector are grouped and aggregated on different computing branches. Within the layered risk avoidance and focus enhancement module, for each sampling time index, the corresponding normalized situation state reconstruction feature vector is input into the global task risk avoidance and focus feature generation operator. The global task risk avoidance and focus feature generation operator forms the global task risk avoidance and focus feature vector by applying the global risk avoidance and focus gating vector to the normalized situation state reconstruction feature vector. Within the layered risk avoidance and focus enhancement module, a node link risk avoidance and focus feature generation operator is constructed based on the topology mapping set. The node link risk avoidance and focus feature generation operator forms the branch aggregation scalar of each computation branch. The branch aggregation scalar is normalized and used as the branch focus weight. The normalized situation state reconstruction feature vector is weighted and aggregated in the branch dimension in combination with the branch feature mapping of each computation branch to obtain the node link risk avoidance and focus feature vector. Within the layered risk avoidance and focus enhancement module, the difference between the normalized situation state reconstruction feature vector at the current sampling time and the normalized situation state reconstruction feature vector at the previous sampling time is calculated to form the micro-area risk mutation residual vector. The micro-area risk mutation residual vector is then subjected to a gating operation to generate the micro-area risk mutation focus feature vector. The risk-sensitive feature vector is generated by weighted fusion of the global task risk avoidance focus feature vector, the node link risk avoidance focus feature vector and the micro-area risk mutation focus feature vector at the same sampling time. The risk-sensitive feature vectors obtained at each sampling time are summarized in the order of the sampling time index to form a risk-sensitive feature vector set.

[0032] In Example 1, the global risk avoidance focusing gating vector is used to adaptively enhance or suppress the feature components representing the urgency of the mission deadline, the proximity of the threat, the deviation of the trajectory, and the safety margin. The branch focusing weight is used to assign higher weights to the feature components corresponding to the nodes where the critical mission is located, the bottleneck links, the energy consumption / temperature rise limited nodes, and the high packet loss links. The micro-area risk mutation gating is used to highlight the micro-area risk changes caused by cross-time-lapse link jitter, computing power reduction, target sudden appearance, or a surge in terrain matching residuals, so that the risk-sensitive feature vector can simultaneously reflect the macro-risk avoidance situation, branch-level scheduling risk, and transient mutation risk.

[0033] In this embodiment, the mapping scheduling control module includes: A mapping scheduling control model is constructed, and the risk-sensitive feature vector set is input into the mapping scheduling control model in the order of sampling time index. The mapping scheduling control model performs feature mapping and scheduling decision calculation on the risk-sensitive feature vector corresponding to each sampling time index, and generates a risk avoidance edge scheduling control vector that corresponds one-to-one with the sampling time index. The task unloading and placement strategy is determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model; The heterogeneous resource configuration parameters are determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model; The link service quality and routing strategy are determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model.

[0034] In Example 1, the task offloading and placement strategy is used to select the target execution location and determine the task fragmentation ratio among CPU / GPU / FPGA / AI accelerators and collaborative nodes. The heterogeneous resource configuration parameters include CPU frequency and core configuration, GPU streaming multiprocessor quota, FPGA reconfiguration level, accelerator batch size and real-time thread priority. The link service quality and routing strategy includes queue priority, bandwidth reservation ratio, retransmission / encoding redundancy and multi-link traffic splitting ratio.

[0035] The risk avoidance edge scheduling control vector is a set of control parameters consisting of a task placement vector, a resource allocation vector, a link scheduling vector, and a risk avoidance guidance vector. Each dimension of the task placement vector corresponds one-to-one with a task node and a target computing node, representing the selection and proportion of local execution, neighboring collaborative execution, or upper-layer link forwarding execution. The resource allocation vector is used to meet constraints on task deadlines, end-to-end latency, power consumption limits, and temperature rise limits. The link scheduling vector is used to meet the latency and reliability constraints of key perception data and risk avoidance decision data. The model parameters of the mapping scheduling control model are obtained through offline calibration using historical flight mission logs, known threat event replays, terrain risk annotations, and system telemetry data. During online operation, small-step adaptive corrections are performed based on closed-loop feedback, balancing real-time performance and stability.

[0036] In this embodiment, the rules for determining the task unloading and placement strategy, heterogeneous resource configuration parameters, and link service quality and routing strategy include: When the proportion of feature components in the risk-sensitive feature vector that reflect increased threat proximity, reduced track safety margin, or tightened risk avoidance calculation deadline exceeds the first proportion threshold, the local priority of critical risk avoidance tasks is increased and they are placed on computing power nodes that meet deterministic latency, while at least one redundant execution mechanism is enabled. Redundant execution mechanisms include parallel inference across multiple computing nodes, mirroring of primary and backup tasks, or dual writing of critical intermediate results.

[0037] When the proportion of feature components in the risk-sensitive feature vector that reflect the increase in link packet loss rate, jitter, or bandwidth decrease is greater than the second proportion threshold, the proportion of external offloading is reduced and transmission-sensitive tasks are migrated to local or near-end nodes where the link converges faster. At the same time, the queue priority, bandwidth reservation ratio, and forward error correction redundancy of critical data streams are increased. When the proportion of feature components in the risk-sensitive feature vector that reflect computing power temperature rise approaching the upper limit, power consumption margin decrease, or cache / queue congestion exceeds the third proportion threshold, load migration across heterogeneous devices is triggered, migrating high-parallelism tasks to GPUs / accelerators and time-sensitive tasks to CPU real-time cores or FPGA pipelines. Resource allocation hysteresis and rate limiting mechanisms are introduced to limit the number of task migrations and the frequency / quota change amplitude per unit time, thereby suppressing scheduling oscillations. When the risk-sensitive feature vector reflects a stable situation and sufficient security margin, an energy consumption optimization strategy is implemented to reduce the resource quota for non-critical tasks and release the reserved bandwidth of the link for background updates or cache prefetching.

[0038] Example 2: This example selects a certain type of twin-engine unmanned fixed-wing flight platform as the application object. The test flight location is in a plateau hilly airspace at an altitude of approximately 3100 meters. The terrain is significantly undulating, with local areas containing canyons, wind shear, and sudden crosswinds. The platform is equipped with an onboard heterogeneous edge computing system, including an eight-core CPU module, an embedded GPU module, a reconfigurable FPGA module, and a dedicated AI acceleration unit. It is also equipped with millimeter-wave radar, an electro-optical pod, an inertial measurement unit, a GNSS navigation module, and a dual-link communication module. The test mission is low-altitude flight path crossing and temporary threat avoidance verification. The flight speed is approximately 260 kilometers per hour, and the cruising altitude is approximately 220 meters above the ground.

[0039] As the flight entered the seventh segment, a sudden weather disturbance occurred in the airspace. Weather radar data showed that the local turbulence index increased from 0.42 to 0.87, the packet loss rate of Link 1 surged from 0.8% to 7.4%, communication latency rose from an average of 28 milliseconds to 96 milliseconds, and the onboard GPU temperature increased from 68 degrees Celsius to 84 degrees Celsius, causing automatic frequency reduction. At the same time, an electro-optical pod detected a suspected moving ground target approximately 1.8 kilometers ahead, with a speed of about 12 meters per second and a trajectory that intersected the flight path of the aircraft by about 20 degrees.

[0040] Under traditional airborne edge scheduling strategies, the system employs a fixed priority queue with static offload ratio control. Flight records show that after the aforementioned incident, the end-to-end processing latency of critical hazard avoidance tasks increased from the original 63 milliseconds to 152 milliseconds, exceeding the preset real-time threshold of 100 milliseconds by 52 milliseconds. The task overdue rate rose from 0.6% to 18.3% within 3 seconds, and the number of link retransmissions increased 3.7 times. Because GPU downclocking was not promptly identified as a scheduling risk source, some visual fusion tasks were still assigned to the GPU for execution, leading to task backlog and delayed issuance of critical trajectory correction commands. The peak flight trajectory deviation reached 11.6 meters, approaching the lower limit of the safety margin.

[0041] Under the deployment conditions of the method of this invention, a mask contrast learning model was established based on three months of historical flight logs before flight. The total training sample consisted of 186,000 time window samples, each with a window length of 2 seconds, containing 64-dimensional situational variables. The samples included approximately 112,000 normal cruise data, 37,000 link fluctuation scenario data, 19,000 computing power temperature rise scenario data, and 18,000 composite threat scenario data. The mask ratio was set to randomly mask 30% of the time period or the entire computation branch per window. Model training was completed in a ground validation environment, the loss function converged in the 42nd round, and the validation set reconstruction consistency reached 92.4%.

[0042] When the aforementioned unforeseen events occur during flight testing, the synchronous situation monitoring data first undergoes noise suppression and topology mapping processing. The system detects joint abrupt changes in three characteristic components: link quality, computing power temperature rise, and threat proximity. The reconstruction model outputs the situational hidden state reconstruction feature vector within 37 milliseconds. The hierarchical risk avoidance focusing module further calculates the global mission pressure index (0.76), branch-level risk weight (0.83), and micro-area mutation residual amplitude (0.41), representing improvements of 1.9 times, 2.4 times, and 3.1 times respectively compared to normal cruise conditions.

[0043] The mapping scheduling control model generates a risk avoidance edge scheduling control vector accordingly. In the control vector, the proportion of critical risk avoidance tasks executed locally is increased from the original 60% to 92%, the visual fusion task is partially migrated to the FPGA pipeline for execution, the GPU load is reduced from the original 82% to 61%, the bandwidth reservation ratio for critical data flow in Link 1 is increased from 20% to 45%, while the priority of non-critical tasks is reduced and background data synchronization is temporarily suspended. The above scheduling changes are completed within 64 milliseconds.

[0044] The results show that, under the same threat and link disturbance conditions, the end-to-end latency of critical risk avoidance tasks using the method of this invention is consistently below 78 milliseconds, with a peak latency not exceeding 91 milliseconds, a reduction of approximately 41% compared to traditional methods. The mission overdue rate remains below 2.3% throughout the entire emergency window, a decrease of approximately 87% compared to traditional methods. The peak flight trajectory deviation is controlled within 4.2 meters, and the safety margin remains above 1.8 times the design margin.

[0045] Further comparison of the three sets of simulation experimental data. In the second set of experiments, at the same location but with wind speeds increased to 17 meters per second, the average latency for critical tasks using the traditional method was 167 milliseconds, while the latency using the method of this invention was 85 milliseconds. In the third set of experiments, under the scenario of simultaneous fluctuations in both links, the task loss rate using the traditional method reached 6.8%, while the present invention controlled it at 1.2%. In a continuous 20-minute simulation of a strong interference environment, the traditional scheduling scheme experienced 5 scheduling oscillations, while the present invention's scheme experienced only 1 oscillation, with the oscillation amplitude reduced by approximately 58%.

[0046] Further comparative verification on the training samples shows that, in a dataset containing 3000 composite threat verification samples, the traditional static threshold-based anomaly detection has an accuracy of 73.6% in identifying the co-occurrence of link interruption and computing power reduction, while the accuracy reaches 91.8% after using the mask contrast learning model of this invention. In 1000 samples of sudden target proximity, the average response time for capturing micro-area risk mutations is reduced from 122 milliseconds in the traditional method to 48 milliseconds.

[0047] Log analysis after the flight ended showed that the system ran continuously for 33 minutes without any critical hazard avoidance link failures. The peak CPU temperature rise was 9 degrees Celsius lower than that of the traditional scheduling strategy, the number of GPU frequency reduction triggers decreased by 46%, and the average power consumption decreased by approximately 8.7%. Throughout the entire flight cycle, the total number of scheduling parameter adjustments was 38, and the scheduling change rate per unit time was less than half that of the traditional strategy, indicating that the method of this invention maintains high scheduling stability while ensuring response speed.

[0048] Example 2 verifies the effectiveness of the method of the present invention in multi-source heterogeneous situation reconstruction, hierarchical risk focusing, and cross-computing power-link-task collaborative scheduling in a real plateau and hilly complex environment. Specific data comparisons demonstrate that the method of the present invention significantly reduces the latency and overdue rate of critical tasks, reduces the number of scheduling oscillations, improves the accuracy of risk identification, and maintains continuous and stable system operation under strong interference environments.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An airborne heterogeneous edge computing terminal for real-time risk avoidance, characterized in that, include: The data acquisition and preprocessing module acquires and preprocesses real-time situational data from the airborne heterogeneous edge computing platform to obtain synchronous situational monitoring data. The mask module performs noise suppression processing on the synchronous situation monitoring data and randomly selects a complete time window or a complete computation branch according to the task-resource-link topology relationship to generate a risk-related mask, thereby obtaining the masked situation monitoring data; The situational hidden state reconstruction module trains the mask contrast learning network using masked situational monitoring data and corresponding unmasked situational monitoring data, and outputs the situational hidden state reconstruction model. The state reconstruction output module inputs the synchronous situation monitoring data into the situation hidden state reconstruction model to obtain a set of normalized situation state reconstruction feature vectors. The risk-sensitive feature module inputs the set of normalized situational state reconstruction feature vectors into the hierarchical risk avoidance and focus enhancement module, which sequentially generates global task risk avoidance and focus features, node link risk avoidance and focus features, and micro-area risk mutation and focus features, and then performs weighted fusion to obtain the set of risk-sensitive feature vectors. The mapping scheduling control module generates risk-averse edge scheduling control vectors based on the risk-sensitive feature vector set through the mapping scheduling control model. The drive execution module converts the edge avoidance scheduling control vector into heterogeneous computing power control signals, link service quality control signals, and avoidance guidance signals. This drives the airborne heterogeneous edge computing platform to complete task unloading, resource allocation, link scheduling, and avoidance guidance actions. The module then merges the situation monitoring data collected after the actions with the synchronous situation monitoring data as the synchronous situation monitoring data for the next cycle, thus achieving closed-loop iteration.

2. The airborne heterogeneous edge computing terminal for real-time risk avoidance according to claim 1, characterized in that, The real-time situational data and preprocessing include: The system collects flight status data, navigation and positioning data, environmental and threat perception data, link communication data, computing resource monitoring data, and task load data. It performs time synchronization, sampling rate alignment, missing data completion, and amplitude normalization on the collected data. It establishes a unified timestamp index for multi-source asynchronous data, and uses the sampling time index as the primary key to concatenate data from different sources into a multivariate channel vector with the same sampling time. It also performs threshold pruning and monotonicity consistency verification on outliers and outputs synchronized situational monitoring data.

3. The airborne heterogeneous edge computing terminal for real-time risk avoidance according to claim 1, characterized in that, The mask module includes: Noise suppression processing is performed on the synchronous situation monitoring dataset to obtain a denoised synchronous situation monitoring dataset. Based on the task-resource-link topology relationship, construct a set of topology mappings; Generate risk-related masks under the constraints of the topological mapping set; Based on the denoised synchronous situational monitoring dataset and the risk-related mask, masked situational monitoring data is generated.

4. The airborne heterogeneous edge computing terminal for real-time risk avoidance according to claim 1, characterized in that, The situational hidden state reconstruction module includes: A mask contrast learning network is constructed, with the masked situation monitoring data as input and the output mask representation vector set. At the same time, a target representation network is constructed, with the unmasked situation monitoring data as input and the output target representation vector set. The mask representation vector set and the target representation vector set are normalized so that each mask representation vector and target representation vector is transformed into a unit norm representation vector. Construct a comparison similarity matrix based on the unit norm representation vector; Based on the contrast similarity matrix, a mask contrast learning loss function is constructed. The gradient optimization method is used to minimize the mask contrast learning loss function. When the training process meets the preset training termination condition, the hidden state reconstruction model is obtained.

5. An airborne heterogeneous edge computing terminal for real-time risk avoidance according to claim 1, characterized in that, The state reconstruction output module includes: The synchronous situation monitoring dataset is used as the input to the situation hidden state reconstruction model. The situation hidden state reconstruction model is used to perform forward feature mapping operation on the synchronous situation monitoring dataset to obtain the situation state reconstruction feature vector set. The components of each dimension in the situation state reconstruction feature vector set are uniformly normalized to obtain the normalized situation state reconstruction feature vector set.

6. The airborne heterogeneous edge computing terminal for real-time risk avoidance according to claim 1, characterized in that, The risk-sensitive feature module includes: The set of normalized situational state reconstructed feature vectors is used as the input to the hierarchical risk avoidance focusing enhancement module; The topology mapping set is used to constrain the feature organization method of the hierarchical risk avoidance and focus enhancement module. The hierarchical risk avoidance and focus enhancement module establishes the correspondence between the task subgraph index, the computing node index and the link index based on the topology mapping set. The task load features, resource features and link features in the normalized situation state reconstruction feature vector are grouped and aggregated on different computing branches. Within the layered risk avoidance and focus enhancement module, for each sampling time index, the corresponding normalized situation state reconstruction feature vector is input into the global task risk avoidance and focus feature generation operator. The global task risk avoidance and focus feature generation operator forms the global task risk avoidance and focus feature vector by applying the global risk avoidance and focus gating vector to the normalized situation state reconstruction feature vector. Within the layered risk avoidance and focus enhancement module, a node link risk avoidance and focus feature generation operator is constructed based on the topology mapping set. The node link risk avoidance and focus feature generation operator forms the branch aggregation scalar of each computation branch. The branch aggregation scalar is normalized and used as the branch focus weight. The normalized situation state reconstruction feature vector is weighted and aggregated in the branch dimension in combination with the branch feature mapping of each computation branch to obtain the node link risk avoidance and focus feature vector. Within the layered risk avoidance and focus enhancement module, the difference between the normalized situation state reconstruction feature vector at the current sampling time and the normalized situation state reconstruction feature vector at the previous sampling time is calculated to form the micro-area risk mutation residual vector. The micro-area risk mutation residual vector is then subjected to a gating operation to generate the micro-area risk mutation focus feature vector. The risk-sensitive feature vector is generated by weighted fusion of the global task risk avoidance focus feature vector, the node link risk avoidance focus feature vector and the micro-area risk mutation focus feature vector at the same sampling time. The risk-sensitive feature vectors obtained at each sampling time are summarized in the order of the sampling time index to form a risk-sensitive feature vector set.

7. An airborne heterogeneous edge computing terminal for real-time risk avoidance according to claim 1, characterized in that, The mapping scheduling control module includes: A mapping scheduling control model is constructed, and the risk-sensitive feature vector set is input into the mapping scheduling control model in the order of sampling time index. The mapping scheduling control model performs feature mapping and scheduling decision calculation on the risk-sensitive feature vector corresponding to each sampling time index, and generates a risk avoidance edge scheduling control vector that corresponds one-to-one with the sampling time index. The task unloading and placement strategy is determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model; The heterogeneous resource configuration parameters are determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model; The link service quality and routing strategy are determined based on the risk avoidance edge scheduling control vector output by the mapping scheduling control model.

8. An airborne heterogeneous edge computing terminal for real-time risk avoidance according to claim 7, characterized in that, The rules for determining the task unloading and placement strategy, heterogeneous resource configuration parameters, and link service quality and routing strategy include: When the proportion of feature components in the risk-sensitive feature vector that reflect increased threat proximity, reduced track safety margin, or tightened risk avoidance calculation deadline exceeds the first proportion threshold, the local priority of critical risk avoidance tasks is increased and they are placed on computing power nodes that meet deterministic latency, while at least one redundant execution mechanism is enabled. When the proportion of feature components in the risk-sensitive feature vector that reflect the increase in link packet loss rate, jitter, or bandwidth decrease is greater than the second proportion threshold, the proportion of external offloading is reduced and transmission-sensitive tasks are migrated to local or near-end nodes where the link converges faster. At the same time, the queue priority, bandwidth reservation ratio, and forward error correction redundancy of critical data streams are increased. When the proportion of feature components in the risk-sensitive feature vector that reflect computing power temperature rise approaching the upper limit, power consumption margin decrease, or cache / queue congestion exceeds the third proportion threshold, load migration across heterogeneous devices is triggered, migrating high-parallelism tasks to GPUs / accelerators and time-sensitive tasks to CPU real-time cores or FPGA pipelines. Resource allocation hysteresis and rate limiting mechanisms are introduced to limit the number of task migrations and the frequency / quota change amplitude per unit time, thereby suppressing scheduling oscillations. When the risk-sensitive feature vector reflects a stable situation and sufficient security margin, an energy consumption optimization strategy is implemented to reduce the resource quota for non-critical tasks and release the reserved bandwidth of the link for background updates or cache prefetching.

Images