An intelligent fusion terminal real-time decision method based on edge computing

Abstract

The application discloses an intelligent fusion terminal real-time decision method based on edge computing and relates to the technical field of edge computing, which comprises the following steps: collecting multi-source state data, preprocessing and generating a terminal state set; calculating the continuous first-order difference of the terminal state set to obtain a state change rate and generating a dynamic state feature sequence; inputting the dynamic state feature sequence into an improved VanillaNet network to generate a multi-dimensional decision representation vector; calculating the business relationship according to the multi-dimensional decision representation vector, generating a target business set and an influence range identifier; performing influence domain contraction processing on the real-time decision influence domain to generate a real-time decision strategy set; performing edge task scheduling, data uploading control and collaborative processing, updating the terminal running state and entering the next round of decision process. Through edge side dynamic state perception, real-time decision influence domain construction and minimum stable decision domain generation, the application realizes real-time autonomous decision and efficient collaborative processing of the intelligent fusion terminal.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and in particular to a real-time decision-making method for intelligent fusion terminals based on edge computing. Background Technology

[0002] With the development of the power Internet of Things and the digital construction of distribution substations, intelligent converged terminals are widely deployed in power fields to collect data on electrical operation status, communication link status, terminal computing power status, and business load status. Existing intelligent converged terminals can typically complete data acquisition, protocol conversion, buffering and forwarding, and remote reporting, and transmit a large amount of field data to the main station platform or cloud for centralized analysis, providing data support for substation monitoring, anomaly alarms, and business processing.

[0003] However, existing intelligent converged terminals still primarily rely on main station platforms or cloud-based analysis for real-time decision-making. When on-site electrical conditions, communication quality, terminal load, and business urgency change rapidly, the terminal can typically only execute fixed rule judgments or simple threshold processing locally, making it difficult to promptly identify the correlation and impact between multiple states. Furthermore, existing methods often use fixed time windows to construct state features, failing to dynamically adjust the sliding window granularity according to the rate of state change. This easily leads to missing crucial change information in abrupt scenarios and generating redundant calculations in stable scenarios.

[0004] Existing edge inference methods rarely consider the constraints between terminal power consumption, computation latency, and service priority simultaneously, which can easily lead to problems such as excessively long inference times, excessive computational resource consumption, or unreasonable service processing order. When abnormal states affect multiple service tasks, existing methods usually directly generate a single processing strategy, lacking influence domain analysis of resource contention, communication dependencies, and execution succession relationships. They also lack a processing mechanism to shrink the scope of influence to form a minimum stable decision domain, resulting in response delays in abnormal service identification and local processing, making it difficult to generate real-time decision strategies adapted to the current terminal operating state in a timely manner.

[0005] Therefore, how to provide a real-time decision-making method for intelligent fusion terminals based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a real-time decision-making method for intelligent fusion terminals based on edge computing. This invention fully utilizes edge computing, dynamic sliding window reconstruction, improved VanillaNet network, and real-time decision influence domain analysis technology to perform local fusion processing on electrical status data, communication link data, terminal computing power data, and service load data collected by the intelligent fusion terminal. Through state change rate identification, adaptive sliding window construction, energy consumption delay constraint reasoning, service influence domain construction, and influence domain contraction processing, the intelligent fusion terminal can achieve real-time judgment, task scheduling, anomaly alarm, and collaborative processing in the field of power Internet of Things. It has the advantages of fast response speed, low cloud dependence, controllable resource consumption, high service processing stability, and strong field autonomy.

[0007] A real-time decision-making method for an intelligent fusion terminal based on edge computing according to an embodiment of the present invention includes: Collect multi-source status data from intelligent fusion terminals, perform preprocessing on the multi-source status data, and generate a terminal status set; Calculate the continuous first-order difference of various features for the terminal state set, obtain the state change rate, perform sliding window reconstruction processing according to the state change rate, generate an adaptive sliding window corresponding to the state change rate, and construct a dynamic state feature sequence. The dynamic state feature sequence is input into the improved VanillaNet network. A sliding window sensing state compression module is set at the input of VanillaNet to perform adaptive state compression on the dynamic state feature sequence. An energy consumption-delay dual-gated lattice layer is set after each VanillaNet block to output a multi-dimensional decision representation vector. Based on the multidimensional decision representation vector, the resource competition relationship, communication dependency relationship and execution succession relationship among various business tasks are calculated. The real-time decision influence domain is constructed, the target business set affected by the current abnormal state is identified, and the business processing priority sequence and influence scope identifier are generated. The impact domain of real-time decision is shrinked by performing impact domain shrinkage processing. Based on the business processing priority sequence and the impact scope identifier, the business nodes in the impact domain of real-time decision are iteratively filtered to construct the minimum stable decision domain and generate a set of real-time decision strategies. Based on the real-time decision-making strategy set, the system executes data upload control, edge task scheduling, abnormal alarm triggering, and collaborative processing instruction generation. It then writes back the execution results to update the terminal's operating status information and enters the next round of edge-side real-time decision-making process.

[0008] Optionally, the multi-source status data includes electrical status data, communication link data, terminal computing power data, and service load data.

[0009] Optionally, the preprocessing of multi-source state data to generate a terminal state set includes performing time alignment, anomaly removal, and unified format processing on the multi-source state data to generate a terminal state set.

[0010] Optionally, the generation of an adaptive sliding window corresponding to the state change rate to construct a dynamic state feature sequence includes: Read electrical status data, communication link data, terminal computing power data and service load data from the terminal status set, and establish corresponding time series according to terminal number, sampling time and feature type; The difference between the state values ​​at adjacent sampling times under the same feature type is calculated to obtain the change of each feature between consecutive sampling times, and a continuous first-order difference sequence is formed in chronological order. Within a set observation interval, the unit time variation amplitude, number of consecutive changes, and consistency of change direction of continuous first-order difference sequences are statistically analyzed, and the state change rate corresponding to various characteristics is determined based on the statistical results. Perform sliding window reconstruction processing based on the rate of state change to generate fine-grained sliding windows, coarse-grained sliding windows, and transitional sliding windows; Electrical operation characteristics, link quality characteristics, computing resource characteristics, and business urgency characteristics are extracted from each fine-grained sliding window, coarse-grained sliding window, and transitional sliding window. The sliding window length, sliding window step size, overlapping area, and state change rate identifier are written into the corresponding feature record and combined in the order of sampling time to form a dynamic state feature sequence.

[0011] Optionally, the step of performing sliding window reconstruction processing based on the state change rate includes: When the rate of state change exceeds the fast threshold, the corresponding feature is classified as a mutation state. The minimum window length and minimum step size are used, and the overlap ratio between sliding windows is increased to generate a fine-grained sliding window. When the rate of change of state is lower than the slow threshold, the corresponding feature is classified as a stable state. The maximum window length and maximum step size are used, and the overlap ratio between sliding windows is reduced to generate a coarse-grained sliding window. When the rate of state change is between a fast threshold and a slow threshold, the corresponding feature is divided into a gradual state. An intermediate value between the minimum window length and the maximum window length and a matching intermediate step size are used. The overlap ratio is set between the overlap ratios corresponding to the abrupt state and the stable state to generate a transition sliding window.

[0012] Optionally, the output multidimensional decision representation vector includes: Read the state feature records corresponding to each sliding window in the dynamic state feature sequence, extract the sliding window length, sliding window step size, overlap ratio and state change rate identifier in each state feature record, and generate a sliding window perception input record. The sliding window sensing input record is input into the sliding window sensing state compression module at the input end of VanillaNet. The retention ratio of the corresponding state features is determined according to the sliding window length, sliding window step size, overlap ratio and state change rate identifier. High-frequency change features and business urgency features corresponding to abrupt states are retained, and repeated sampling features corresponding to stable states are compressed to generate a compressed state feature sequence. The compressed state feature sequence is input into the improved VanillaNet network. Through VanillaNet blocks, layer-by-layer mapping, non-linear activation and state feature transfer are performed on the compressed state feature sequence to generate intermediate state representations corresponding to each network layer. An energy consumption gate is set after each VanillaNet block. The terminal processor current value, processor temperature value and current computing power occupancy status are read. The number of channels to be reserved is determined based on the processor current value, processor temperature value and current computing power occupancy status. Redundant channels in the intermediate state representation are pruned to generate an energy consumption constraint state representation. A delay gate is set after each VanillaNet block. The terminal response budget, service urgency, and communication link delay are read. Based on the terminal response budget, service urgency, and communication link delay, the number of layers the network continues to calculate or the number of layers is output in advance is determined. Deep control is performed on the energy consumption constraint state representation to generate a multi-dimensional decision representation vector.

[0013] Optionally, the generation of the service processing priority sequence and impact scope identifier includes: Read the multidimensional decision representation vector, and according to the business task type, split the multidimensional decision representation vector into resource occupancy representation, communication path representation, execution status representation and exception association representation, and associate each representation with the corresponding business task; Based on resource occupancy, the status of each business task's occupancy of processor, memory, cache queue, and local computing thread is read, and business tasks that simultaneously occupy the same terminal resources and have overlapping execution times are marked as having resource contention. Based on the communication path representation, read the communication interface, transmission link, reporting target and collaborative terminal corresponding to each business task, and mark the business tasks that share the same link, depend on the same reporting path or require the previous communication result as input as having a communication dependency relationship. Based on the execution status, the triggering conditions, execution preconditions, and processing result destinations of each business task are read. Business tasks whose execution results are called by another business task, require control processing after an exception alarm is triggered, or require supplementary reporting processing after the cached task is completed are marked as having an execution succession relationship. Using business tasks as nodes and resource competition relationships, communication dependencies, and execution acceptance relationships as connecting edges, a real-time decision-making influence domain is constructed. Based on the anomaly association representation, the target business set affected by the current anomaly state is located in the real-time decision-making influence domain. A business processing priority sequence is generated according to resource occupation conflict, communication dependency order, and execution acceptance order, and an influence scope identifier is written for the target business set.

[0014] Optionally, the construction of the real-time decision influence domain includes: Read the multidimensional decision representation vector and map the resource consumption data, communication path data and execution status data in it to the corresponding business tasks to form a set of business feature records; Compare the resource usage data in the business feature record set to identify business tasks that simultaneously occupy the same processor, memory, or cache resources and have overlapping execution times, generate resource contention connection edges and write them into the resource contention edge list; Compare the communication path data in the business feature record set to identify business tasks that share the same communication link, share the same upload channel, or depend on the same return channel, generate communication dependency connection edges and write them into the communication dependency edge list; The execution status data in the business feature record set is compared to identify business tasks with trigger-call, alarm-control, or cache-report relationships, and execution connection edges are generated and written into the execution connection edge list. Using business tasks as nodes and resource contention connection edges, communication dependency connection edges, and execution acceptance connection edges as edges, the resource contention edge list, communication dependency edge list, and execution acceptance edge list are merged to generate a real-time decision influence domain graph structure. Business nodes affected by the current abnormal state are marked in the real-time decision influence domain graph structure to form a real-time decision influence domain.

[0015] Optionally, the step of constructing a minimum stable decision domain and generating a set of real-time decision strategies includes: Read the real-time decision impact domain, target business set, business processing priority sequence and impact scope identifier, take the business tasks in the target business set as the business nodes to be filtered, and write the business processing priority sequence into the filtering order record of the corresponding business node; Read the business nodes to be filtered in order of priority sequence, and take the business nodes with the highest priority as the core candidate nodes. Read the resource competition connection edges, communication dependency connection edges and execution acceptance connection edges of the core candidate nodes in the real-time decision influence domain. Based on the influence scope identifier, the influence propagation boundary in the real-time decision influence domain is defined, and business nodes located within the influence propagation boundary that have resource competition connections, communication dependency connections, or execution takeover connections with core candidate nodes are marked as reserved business nodes; Iterative pruning is performed on the business nodes in the real-time decision influence domain, removing business nodes that are not referenced by the business processing priority sequence, do not fall within the influence scope identifier limited area, and have no resource competition connection, communication dependency connection, or execution acceptance connection with the retained business nodes from the real-time decision influence domain; The decision domain boundary is reconstructed based on the retained business nodes and their corresponding resource contention connections, communication dependency connections, and execution acceptance connections, generating a minimum stable decision domain. A set of real-time decision strategies is then generated according to the selection order, connection relationship, and influence scope of the business nodes in the minimum stable decision domain.

[0016] Optionally, the step of generating instructions for data upload control, edge task scheduling, anomaly alarm triggering, and collaborative processing based on a set of real-time decision-making strategies includes: Read the real-time decision strategy set and, according to the filtering order, connection relationship and impact range of business nodes in the minimum stable decision domain, split the real-time decision strategy set into data upload control strategy, edge task scheduling strategy, abnormal alarm triggering strategy and collaborative processing instruction strategy; Based on the data upload control strategy, the data to be uploaded is identified in terms of business type, upload priority marking, cache queue allocation and link selection. Abnormal alarm data, control feedback data and target business data within the scope of influence are written into the priority upload queue and non-target business data are written into the delayed upload queue. Based on the edge task scheduling strategy, the terminal processor usage status, memory usage status and local task queue status are read. Business tasks that need to be processed locally within the minimum stable decision domain are allocated to edge computing threads, and the task start time, task end time, task execution status and task output results are recorded. Based on the abnormal alarm triggering strategy, the abnormal type, impact scope identifier and business processing priority sequence in the target business set are read, an abnormal alarm record is generated, and the abnormal alarm record is sent to the main station platform, associated terminal or local control unit. Based on the collaborative processing instruction strategy, the communication dependencies and execution succession relationships in the real-time decision influence domain are read, collaborative processing instructions are generated, and collaborative processing instructions are sent to associated terminals or field execution devices, and the instruction number, sending time, recipient and execution feedback are recorded. The system receives data upload results, edge task execution results, anomaly alarm feedback results, and collaborative processing feedback results. It writes each execution result into the status write-back record, updates the communication link status, terminal computing power status, service load status, and anomaly handling status in the terminal status set, and then enters the next round of edge-side real-time decision-making process.

[0017] The beneficial effects of this invention are: This invention collects electrical status data, communication link data, terminal computing power data, and service load data from intelligent converged terminals. At the terminal side, it performs time alignment, anomaly removal, and unified format processing to form a terminal status set that can be used for edge analysis. This allows intelligent converged terminals to move beyond simply acting as data acquisition and forwarding devices and directly participate in status judgment and service processing in the power Internet of Things (IoT) field. By performing continuous first-order difference analysis on various features to obtain the state change rate, and reconstructing fine-grained sliding windows, transitional sliding windows, and coarse-grained sliding windows based on the state change rate, it can improve the ability to capture key changes under abrupt changes and reduce redundant calculations under stable conditions, thereby improving the accuracy and processing efficiency of real-time state modeling at the terminal side.

[0018] This invention improves the VanillaNet network by inputting dynamic state feature sequences. A sliding window-aware state compression module is added to the network input, and an energy-latency dual-gated lattice layer is set after each VanillaNet block. This allows the network inference process to be dynamically adjusted simultaneously based on the sliding window state, terminal processor current, processor temperature, terminal response budget, service urgency, and communication link latency. By constraining the network channel width through energy gates and controlling the network computation depth through latency gates, the computational load on the terminal can be reduced while ensuring the effectiveness of the decision representation. This avoids the problem of excessive inference latency at the edge due to limited computing power or insufficient response budget, and improves the real-time processing capability of intelligent fusion terminals in complex distribution network scenarios.

[0019] This invention constructs a real-time decision influence domain based on a multi-dimensional decision representation vector, incorporating resource competition, communication dependencies, and execution succession relationships among business tasks into a unified analysis. Furthermore, it constructs a minimum stable decision domain through influence domain contraction, enabling real-time decision-making strategies to move beyond isolated judgments of single anomalies or single business processes. Instead, it allows for scope limitation, node selection, and strategy generation around a set of affected businesses, reducing resource consumption caused by irrelevant business involvement in decision-making, minimizing unnecessary data uploads and scheduling conflicts, and improving the coherence of anomaly alarms, edge task scheduling, collaborative processing, and status write-back. This empowers intelligent converged terminals with stronger on-site autonomous decision-making capabilities and rapid response capabilities to anomalies. Attached Figure Description

[0020] 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 a real-time decision-making method for an intelligent fusion terminal based on edge computing proposed in this invention; Figure 2This is a diagram showing the state change rate and sliding window type classification of a real-time decision-making method for an intelligent fusion terminal based on edge computing proposed in this invention. Figure 3 This diagram illustrates the construction and contraction of the real-time decision influence domain in a real-time decision-making method for an intelligent fusion terminal based on edge computing proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A real-time decision-making method for intelligent fusion terminals based on edge computing, comprising: Collect multi-source status data from intelligent fusion terminals, perform preprocessing on the multi-source status data, and generate a terminal status set; Calculate the continuous first-order difference of various features for the terminal state set, obtain the state change rate, perform sliding window reconstruction processing according to the state change rate, generate an adaptive sliding window corresponding to the state change rate, and construct a dynamic state feature sequence. The dynamic state feature sequence is input into the improved VanillaNet network. A sliding window sensing state compression module is set at the input of VanillaNet to perform adaptive state compression on the dynamic state feature sequence. An energy consumption-delay dual-gated lattice layer is set after each VanillaNet block to output a multi-dimensional decision representation vector. Based on the multidimensional decision representation vector, the resource competition relationship, communication dependency relationship and execution succession relationship among various business tasks are calculated. The real-time decision influence domain is constructed, the target business set affected by the current abnormal state is identified, and the business processing priority sequence and influence scope identifier are generated. The impact domain of real-time decision is shrinked by performing impact domain shrinkage processing. Based on the business processing priority sequence and the impact scope identifier, the business nodes in the impact domain of real-time decision are iteratively filtered to construct the minimum stable decision domain and generate a set of real-time decision strategies. Based on the real-time decision-making strategy set, the system executes data upload control, edge task scheduling, abnormal alarm triggering, and collaborative processing instruction generation. It then writes back the execution results to update the terminal's operating status information and enters the next round of edge-side real-time decision-making process.

[0023] In this embodiment, the multi-source status data includes electrical status data, communication link data, terminal computing power data, and service load data.

[0024] In this embodiment, the preprocessing of multi-source state data to generate a terminal state set includes performing time alignment, anomaly removal, and unified format processing on the multi-source state data to generate a terminal state set.

[0025] In this embodiment, the step of generating an adaptive sliding window corresponding to the state change rate and constructing a dynamic state feature sequence includes: Read electrical status data, communication link data, terminal computing power data, and service load data from the terminal status set, and establish corresponding time series according to terminal number, sampling time, and feature type. Specifically: Electrical status data includes voltage, current, power, frequency, switch status, and abnormal event markers. Communication link data includes link delay, packet loss rate, retransmission count, signal strength, link occupancy rate, and communication interface number. Terminal computing power data includes processor occupancy rate, memory occupancy rate, cache queue length, thread occupancy status, processor temperature, and processor current. Service load data includes service type, service trigger time, service data volume, service processing status, service reporting target, and service urgency level. The system uses terminal number as the data attribution identifier, sampling time as the sorting basis, and feature type as the classification field to arrange data from the same terminal and the same feature type in ascending order of sampling time, forming a corresponding time series. The difference between the state values ​​at adjacent sampling times under the same feature type is calculated to obtain the change of each feature between consecutive sampling times, and a continuous first-order difference sequence is formed in chronological order. Specifically: After establishing the time series, the difference between the state values ​​at adjacent sampling times under the same feature type is calculated. For voltage, current, link delay, packet loss rate, processor utilization rate, buffer queue length, and service data volume, the difference between the state value at the next sampling time and the state value at the previous sampling time is directly calculated to obtain the change of the feature between adjacent sampling times. For switch state, communication interface state, and service processing state, different states are first mapped to state codes, and then it is calculated whether the state codes have changed at adjacent sampling times. When the state code changes, it is recorded as a state transition. When the state code does not change, it is recorded as a state hold. The change of various features or the state transition results are arranged in the order of sampling time to form a continuous first-order difference sequence. The continuous first-order difference sequence is used to characterize the change trend of the feature in the continuous sampling process. Within a defined observation interval, the unit-time variation amplitude, number of consecutive variations, and consistency of the direction of variation of a continuous first-order difference sequence are statistically analyzed. Based on the statistical results, the state change rate corresponding to each characteristic is determined. Specifically: Within a defined observation interval, the unit-time variation amplitude, number of consecutive changes, and consistency of change direction of a continuous first-order difference sequence are statistically analyzed. The observation interval is set according to the business processing cycle of the intelligent fusion terminal. In the real-time monitoring scenario of the distribution area, an observation interval is set to consist of 10 to 60 consecutive sampling points. When the sampling period is 1 second, the observation interval is set to 10 to 60 seconds. The unit-time variation amplitude is used to characterize the intensity of the feature change within the observation interval. The number of consecutive changes is used to characterize the frequency of the feature's continuous changes within the observation interval. The consistency of change direction is used to characterize whether the feature continuously increases, continuously decreases, or fluctuates repeatedly within the observation interval. The state change rate is determined based on the comprehensive results of the unit-time variation amplitude, the number of consecutive changes, and the consistency of change direction. The state change rate is divided into abrupt change state, gradual change state, and stable state. The abrupt change state indicates that the feature undergoes a large continuous change or state jump in a short period of time. The gradual change state indicates that the feature continues to change within the observation interval but the change amplitude does not reach the abrupt change state. The stable state indicates that the feature changes with a small amplitude and a small number of consecutive changes within the observation interval. Based on the rate of state change, a sliding window reconstruction process is performed to generate fine-grained sliding windows, coarse-grained sliding windows, and transitional sliding windows. Specifically: Sliding window reconstruction is performed based on the rate of state change. For features classified as abrupt changes, fine-grained sliding windows are generated. The fine-grained sliding windows use preset minimum window length, minimum sliding window step size, and maximum overlap area to retain more repeated sampling points between adjacent sliding windows, which are used to capture detailed changes before and after abrupt changes. For features classified as stable states, a coarse-grained sliding window is generated. The coarse-grained sliding window uses a preset maximum window length, maximum sliding window step size, and minimum overlap area to compress repeated sampling information in stable states and reduce the amount of computation at the edge. For features classified as gradual states, a transition window is generated. The transition window uses a window length, window step size, and overlapping area that are between fine-grained and coarse-grained windows, in order to balance the recognition of gradual trends and the control of computational load. Electrical operation characteristics, link quality characteristics, computing resource characteristics, and business urgency characteristics are extracted from each fine-grained sliding window, coarse-grained sliding window, and transitional sliding window. The sliding window length, sliding window step size, overlapping area, and state change rate identifier are written into the corresponding feature record and combined in the order of sampling time to form a dynamic state feature sequence.

[0026] In this embodiment, the step of performing sliding window reconstruction processing based on the state change rate includes: When the rate of state change exceeds the fast threshold, the corresponding feature is classified as a mutation state. The minimum window length and minimum step size are used, and the overlap ratio between sliding windows is increased to generate a fine-grained sliding window. When the rate of change of state is lower than the slow threshold, the corresponding feature is classified as a stable state. The maximum window length and maximum step size are used, and the overlap ratio between sliding windows is reduced to generate a coarse-grained sliding window. When the rate of state change is between the fast threshold and the slow threshold, the corresponding feature is divided into a gradual state. The intermediate value between the minimum window length and the maximum window length and the matching intermediate step size are adopted, and the overlap ratio is set between the overlap ratios corresponding to the abrupt state and the stable state to generate a transition sliding window. The rate of change of various features is normalized to the range of 0 to 1, with a fast threshold of 0.6 and a slow threshold of 0.2. When the rate of state change is greater than 0.6, the corresponding feature is classified as an abrupt state, and a fine-grained sliding window is generated using the minimum window length, minimum sliding window step size, and maximum overlap ratio. In one embodiment, the window length of the fine-grained sliding window is set to 5 to 10 sampling points, the sliding window step size is set to 1 to 2 sampling points, and the overlap area is set to 60% to 80% of the window length. When the rate of change of state is less than 0.2, the corresponding feature is classified as a steady state, and a coarse-grained sliding window is generated using the maximum window length, the maximum sliding window step size, and the minimum overlap ratio. In one embodiment, the window length of the coarse-grained sliding window is set to 30 to 60 sampling points, the sliding window step size is set to 10 to 20 sampling points, and the overlap area is set to 10% to 30% of the window length. When the rate of change of state is greater than or equal to 0.2 and less than or equal to 0.6, the corresponding feature is divided into a gradual state, and a transition window is generated using a window length, window step size and overlap ratio that are between fine-grained and coarse-grained sliding windows. In one embodiment, the window length of the transition window is set to 15 to 30 sampling points, the window step size is set to 5 to 10 sampling points, and the overlap area is set to 30% to 50% of the window length.

[0027] In this embodiment, the output multidimensional decision representation vector includes: Read the state feature records corresponding to each sliding window in the dynamic state feature sequence, extract the sliding window length, sliding window step size, overlap ratio, and state change rate identifier from each state feature record, and generate a sliding window sensing input record. Specifically: The dynamic state feature sequence consists of multiple state feature records arranged according to sampling time. Each state feature record corresponds to a fine-grained sliding window, a coarse-grained sliding window, or a transitional sliding window. When the intelligent fusion terminal reads each state feature record, it simultaneously reads the electrical operation features, link quality features, computing resource features, and business urgency features within the sliding window. It also reads the sliding window length, sliding window step size, overlap ratio, and state change rate identifier written during the construction of the sliding window. The sliding window length is used to indicate the number of sampling points covered by the state feature record. The sliding window step size is used to indicate the movement interval between adjacent sliding windows. The overlap ratio is used to indicate the proportion of repeated sampling points between adjacent sliding windows to the current sliding window length. The state change rate identifier is used to indicate whether the state corresponding to the sliding window is a sudden state, a gradual state, or a stable state. The intelligent fusion terminal binds the sliding window construction information with the state features within the sliding window to generate a sliding window perception input record. The sliding window sensing input record is input into the sliding window sensing state compression module at the input end of VanillaNet. Based on the sliding window length, sliding window step size, overlap ratio, and state change rate identifier, the retention ratio of corresponding state features is determined. High-frequency change features and urgent service features corresponding to abrupt changes are retained, while repetitive sampling features corresponding to stable states are compressed to generate a compressed state feature sequence. Specifically: The sliding window sensing state compression module is set at the input end of the VanillaNet network. Its input is the sliding window sensing input record, and its output is the compressed state feature sequence. The sliding window sensing state compression module first reads the state change rate identifier. When the state change rate identifier is a sudden state, it retains the voltage change, current change, link delay change, packet loss change, processor usage change, buffer queue change, and service urgency level change that fluctuate rapidly within the sliding window as key features. It also retains the service urgency features related to abnormal alarms, control feedback, and emergency reporting within the corresponding sliding window. When the rate of change of state is identified as a steady state, the repeated sampling features with small numerical changes, consistent change direction, or no change in state code in multiple consecutive sampling points are compressed. When the rate of change of state is identified as a gradual state, the initial state, intermediate state and final state that can represent the trend of change are retained, while the state features with high repetition between adjacent sampling points are compressed. When the sliding window length is shorter, the sliding window step size is smaller, and the overlap ratio is larger, it indicates that the sliding window is used to capture subtle changes. The sliding window perception state compression module improves the feature retention ratio. When the sliding window length is longer, the sliding window step size is larger, and the overlap ratio is smaller, it indicates that the sliding window is used to compress stable states. The sliding window perception state compression module reduces the retention ratio of repetitive features. Key change information in abrupt states is retained, and redundant sampling information in stable states is compressed to generate a compressed state feature sequence. The compressed state feature sequence is input into the improved VanillaNet network. Through VanillaNet blocks, layer-by-layer mapping, non-linear activation, and state feature transfer are performed on the compressed state feature sequence to generate intermediate state representations for each network layer. Specifically: The improved VanillaNet network includes a sliding window sensing state compression module at the input end, multiple VanillaNet blocks, energy consumption gates and delay gates set after each VanillaNet block, and an output mapping layer for outputting multi-dimensional decision representation vectors. After the compressed state feature sequence enters the VanillaNet block, it first undergoes layer-by-layer mapping processing to map electrical operation features, link quality features, computing power resource features, and service emergency features from different sources to the same state representation space. Then, nonlinear activation processing is performed to express the correlation changes between different features as distinguishable state patterns. Then, state feature transmission is performed so that the state change information extracted by the previous layer is passed to the next layer. Each VanillaNet block outputs an intermediate state representation, which contains a comprehensive expression of the current network layer's terminal operation state, communication link state, computing power load state, and service emergency state. For compressed state features generated by fine-grained sliding windows, the VanillaNet block retains more abrupt change details. For compressed state features generated by coarse-grained sliding windows, the VanillaNet block extracts stable trends. For compressed state features generated by transitional sliding windows, the VanillaNet block extracts continuous change trends. An energy consumption gate is set after each VanillaNet block. The terminal processor current value, processor temperature value, and current computing power occupancy status are read. Based on the processor current value, processor temperature value, and current computing power occupancy status, the number of channels to be reserved is determined. Redundant channels in the intermediate state representation are pruned to generate an energy consumption constraint state representation. Specifically: An energy consumption gate is set at the output of each VanillaNet block to dynamically constrain the network channel width based on the current hardware operating status of the intelligent fusion terminal. The energy consumption gate reads the terminal processor current value, processor temperature value, and current computing power occupancy status. The processor current value is used to characterize the load pressure of the current inference process on the terminal power supply side, the processor temperature value is used to characterize the thermal state of the terminal chip or processing unit, and the current computing power occupancy status is used to characterize the occupancy of processor threads, cache, and local computing queue. When the processor current value, processor temperature value, and current computing power occupancy status are all within the normal operating range, the energy consumption gate retains the main and auxiliary channels output by the current VanillaNet block. When the processor current value increases, the processor temperature rises, or the computing power occupancy increases, the energy consumption gate prioritizes retaining channels related to abnormal alarms, service urgency levels, link congestion, and electrical anomalies, and cuts off repetitive channels, stable channels, and low-change channels that contribute less to the current decision. In one implementation, the energy consumption gate can sort channels according to their correlation with the urgency of the business, the degree of correlation with anomalies, and the rate of state change. It retains the channels with higher rankings and sets the channels with lower rankings as invalid outputs, generating an energy consumption constraint state representation. The energy consumption constraint state representation reduces the network computing burden while retaining key decision information. A delay gate is set after each VanillaNet block. The terminal response budget, service urgency, and communication link delay are read. Based on the terminal response budget, service urgency, and communication link delay, the number of layers the network continues to compute or the number of layers is output in advance is determined. Deep control is performed on the energy consumption constraint state representation to generate a multi-dimensional decision representation vector. Specifically: The delay gate is set after the energy consumption gate to control the computation depth of the VanillaNet network according to the real-time service processing requirements. The delay gate reads the terminal response budget, the service urgency level, and the communication link latency. The terminal response budget represents the maximum processing time allowed for the edge terminal to complete the decision for the current service. The service urgency level indicates whether the current service belongs to an abnormal alarm, control feedback, emergency reporting, or ordinary data collection service. The communication link latency represents the link time required for the terminal to transmit the results to the main station platform, associated terminals, or field execution devices. When the service urgency level is high and the communication link latency is large, the delay gate reduces the number of layers the network continues to compute. When the current energy consumption constraint state indicates that the output conditions have been met, it triggers early output. When the service urgency level is low and the terminal response budget is sufficient, the delay gate allows the energy consumption constraint state to continue to enter the next VanillaNet block for computation. After each layer outputs, the delay gate reads the currently consumed computation time and compares it with the terminal response budget. If the remaining computation time can cover the computation of the next network layer, it proceeds to the next network layer to continue inference. If the remaining computation time is insufficient to cover the computation of the next network layer, it stops inference and triggers early output. For the layer that triggers early output, the delay gate extracts the current energy consumption constraint state representation as a candidate output feature. For the case where all allowed computation layers are completed, the delay gate extracts the energy consumption constraint state representation of the last level as a candidate output feature. After the candidate output features are processed by the output mapping layer, a multi-dimensional decision representation vector is generated. The multi-dimensional decision representation vector includes resource occupancy representation, communication path representation, execution state representation, and anomaly association representation.

[0028] This invention structurally optimizes the traditional VanillaNet model, proposing an improved VanillaNet network for real-time decision-making scenarios at the edge of intelligent fusion terminals. In terms of input modeling, a sliding window perception state compression module is introduced. Based on the sliding window length, sliding window step size, overlap ratio, and state change rate identifiers corresponding to each sliding window in the dynamic state feature sequence, differential compression processing is performed on electrical operation characteristics, link quality characteristics, computing resource characteristics, and business urgency characteristics under different state change types. This ensures that the state features of the input network retain key change information under abrupt changes while compressing repetitive sampling information under stable conditions, thereby reducing input redundancy at the edge and improving the model's efficiency in recognizing changes in the field state.

[0029] To address the limitations of computing power and energy consumption in intelligent converged terminals, this invention introduces an energy consumption gate after each VanillaNet block. By reading the terminal processor's current and temperature values, and the current computing power occupancy status, the number of channels retained representing intermediate network states is dynamically controlled. Since the energy consumption gate prioritizes channels related to anomaly alarms, link congestion, service urgency levels, and state change rates, and prunes redundant channels with high repetition in stable states, it can reduce the participation of invalid channels in computation without altering the primary real-time decision-making task. This makes the network inference process more suitable for deployment in the edge computing environment of intelligent converged terminals.

[0030] To address the varying response time requirements of field services in the power Internet of Things (IoT), this invention further incorporates a delay gate after each VanillaNet block. By reading the terminal response budget, service urgency, and communication link latency, it controls the number of layers the network continues to compute or outputs prematurely. When the service urgency is high and communication link latency consumes a significant portion of the response budget, the delay gate triggers premature output, enabling the model to generate a multi-dimensional decision representation vector within a limited time. When the response budget can cover the continued inference process, the delay gate allows the state representation to proceed to the next network layer for further computation. This allows the network depth to be dynamically adjusted according to the service time limit, avoiding response delays caused by fixed-depth inference.

[0031] The improved VanillaNet network of this invention not only retains the characteristics of VanillaNet's simple structure and suitability for edge deployment, but also achieves coordinated constraints on input features, channel width, and computational depth through a sliding window-aware state compression module, energy consumption gates, and delay gates. The sliding window-aware state compression module solves the problem of inconsistent input granularity of multi-source state data in abrupt, gradual, and stable scenarios; the energy consumption gate solves the problem of uncontrollable model computational burden when terminal processor temperature, current, and computing power occupancy change; and the delay gate solves the problem of incompatibility of fixed inference depth under different service response budgets. This improves the deployability, real-time performance, and effectiveness of state representation of the model in real-time decision-making scenarios of intelligent fusion terminals.

[0032] In this embodiment, the generation of the service processing priority sequence and the scope of influence identifier includes: Read the multidimensional decision representation vector, and according to the business task type, split the multidimensional decision representation vector into resource occupancy representation, communication path representation, execution status representation and exception association representation, and associate each representation with the corresponding business task; Based on resource occupancy, the system reads the occupancy status of each business task in terms of processor, memory, cache queue, and local computing threads. Business tasks that simultaneously occupy the same terminal resources and have overlapping execution times are marked as having resource contention. Specifically: The resource usage representation corresponding to each business task is read sequentially, and the processor core number, processor usage time period, memory usage space, cache queue number and local computing thread number are extracted. When two or more business tasks have the same processor core, the same cache queue, the same computing thread or the same memory space usage, and their execution time intervals overlap, it is determined that there is a resource competition relationship between the business tasks. For the identified resource competition relationship, the competition resource type, competition business number, competition start time and competition end time are recorded, and the corresponding resource competition connection edge is generated. Based on the communication path representation, the communication interface, transmission link, reporting target, and cooperating terminal corresponding to each business task are read. Business tasks that share the same link, depend on the same reporting path, or require the previous communication result as input are marked as having a communication dependency relationship. Specifically: Read the communication interface number, link number, reporting target address and cooperating terminal number corresponding to each business task, and establish the corresponding communication path. When multiple business tasks share the same communication interface, share the same wireless link, share the same master station reporting path, or need to read the uploaded result of the previous business task as communication input, determine that there is a communication dependency relationship between the business tasks, establish a communication dependency connection edge, and record the communication dependency source business, communication dependency target business, dependency link number and dependency direction. Based on the execution status, the triggering conditions, preconditions, and destination of processing results for each business task are read. Business tasks whose execution results are called by another business task, require control processing after an exception alarm is triggered, or require supplementary reporting processing after a cached task is completed are marked as having an execution succession relationship. Specifically: The execution status representation of each business task is read sequentially, and the business triggering event, execution pre-execution business, execution result output object and result receiving object are obtained. When the output result of a certain business task is used as the input condition of another business task after the execution is completed, an execution succession relationship is established. When the abnormal alarm business is completed and the control execution business is triggered, an alarm control succession relationship is established. When the cache replenishment business depends on the cache completion status before it can be started, a cache replenishment succession relationship is established. The source business of execution succession, the target business of execution succession, the succession type and the succession order are recorded, and the corresponding execution succession connection edge is generated. Using business tasks as nodes and resource competition relationships, communication dependencies, and execution succession relationships as connecting edges, a real-time decision-making influence domain is constructed. Based on the anomaly correlation representation, the set of target business processes affected by the current anomaly state is located within the real-time decision-making influence domain. A business processing priority sequence is generated according to resource occupation conflicts, communication dependency order, and execution succession order, and an influence scope identifier is written for the target business set. Specifically: Each business task is treated as a business node within the real-time decision-making influence domain. Resource contention connections, communication dependencies, and execution acceptance connections are all written into the same business relationship graph to form the real-time decision-making influence domain. Anomaly associations are read, and based on the business task number corresponding to the anomaly event, business nodes with direct or indirect connections are searched within the real-time decision-making influence domain. These business nodes are grouped into a target business set. The resource release order is determined based on resource contention connections, the communication execution order based on communication dependencies, and the business processing order based on execution acceptance connections. A business processing priority sequence is generated by comprehensively considering these three orders. Starting from the anomaly business node, the search proceeds layer by layer along resource contention connections, communication dependencies, and execution acceptance connections. Expansion stops when a business node without a connection edge is found or a preset search level is reached. All searched business nodes are written into the target business set. The entire range of business nodes covered by the target business set is considered the anomaly's influence scope, and a corresponding influence scope identifier is written for each target business node.

[0033] Among them, the real-time decision impact domain refers to a dynamic business association area constructed with business tasks as nodes and resource competition relationships, communication dependencies, and execution acceptance relationships as connecting edges. It is used to characterize the scope, path, and business association relationships of the current abnormal state on each business task within the intelligent fusion terminal.

[0034] In this embodiment, constructing the real-time decision influence domain includes: Read the multidimensional decision representation vector and map the resource consumption data, communication path data and execution status data in it to the corresponding business tasks to form a set of business feature records; Compare the resource usage data in the business feature record set to identify business tasks that simultaneously occupy the same processor, memory, or cache resources and have overlapping execution times, generate resource contention connection edges and write them into the resource contention edge list; Compare the communication path data in the business feature record set to identify business tasks that share the same communication link, share the same upload channel, or depend on the same return channel, generate communication dependency connection edges and write them into the communication dependency edge list; The execution status data in the business feature record set is compared to identify business tasks with trigger-call, alarm-control, or cache-report relationships, and execution connection edges are generated and written into the execution connection edge list. Using business tasks as nodes and resource contention connection edges, communication dependency connection edges, and execution acceptance connection edges as edges, the resource contention edge list, communication dependency edge list, and execution acceptance edge list are merged to generate a real-time decision influence domain graph structure. Business nodes affected by the current abnormal state are marked in the real-time decision influence domain graph structure to form a real-time decision influence domain.

[0035] In this embodiment, the step of constructing a minimum stable decision domain and generating a set of real-time decision strategies includes: Read the real-time decision impact domain, target business set, business processing priority sequence and impact scope identifier, take the business tasks in the target business set as the business nodes to be filtered, and write the business processing priority sequence into the filtering order record of the corresponding business node; The business nodes to be screened are read sequentially from front to back according to the business processing priority sequence. The business nodes with the highest ranking are selected as core candidate nodes. The resource competition connection edges, communication dependency connection edges, and execution acceptance connection edges of the core candidate nodes in the real-time decision influence domain are read. Specifically: The intelligent fusion terminal reads the business nodes to be screened from front to back according to the business processing priority sequence, and takes the currently read business node as the core candidate node. For each core candidate node, it searches for the connection edge directly connected to it in the real-time decision influence domain, including resource contention connection edge, communication dependency connection edge, and execution acceptance connection edge. The resource contention connection edge is used to indicate whether the core candidate node and other business nodes occupy the same processor, memory, cache queue or local computing thread. The communication dependency connection edge is used to indicate whether the core candidate node and other business nodes share the same communication link, upload channel or return channel. The execution acceptance connection edge is used to indicate whether there is an execution order relationship of triggering, calling, alarm control or cache supplementation between the core candidate node and other business nodes. After reading the connection edge, the directly associated business nodes of the core candidate node are obtained. Based on the influence scope identifier, the influence propagation boundary in the real-time decision influence domain is defined. Business nodes located within the influence propagation boundary and having resource contention connections, communication dependency connections, or execution takeover connections with core candidate nodes are marked as reserved business nodes. Specifically: The impact propagation boundary is determined by the impact scope identifier, which includes the anomaly source service number, impact level, impact link type, and impact termination condition. Starting from the anomaly source service node corresponding to the current anomaly state, the search is expanded along the resource contention connection edge, communication dependency connection edge, and execution acceptance connection edge in the real-time decision impact domain. When the searched service node is within the level defined by the impact scope identifier, and there is a resource contention connection, communication dependency connection, or execution acceptance connection between the service node and the core candidate node, the service node is marked as a reserved service node. For service nodes that exceed the impact level, are not within the current impact link type, or meet the impact termination condition, the expansion along that direction is stopped. The reserved service nodes only cover the service tasks actually affected by the current anomaly state. Iterative pruning is performed on the business nodes in the real-time decision influence domain. Business nodes that are not referenced by the business processing priority sequence, do not fall within the influence scope identifier area, and have no resource contention connection, communication dependency connection, or execution takeover connection with the retained business nodes are removed from the real-time decision influence domain. Specifically: After marking the retained business nodes, an iterative pruning process is performed on all business nodes in the real-time decision influence domain. In each iteration, a business node is read and it is determined whether the business node is referenced by the business processing priority sequence, whether it falls within the influence scope identifier limited area, whether it has a resource contention connection, a communication dependency connection, or an execution connection with any retained business node. For business nodes that do not meet the above conditions, the business node is determined to be an unrelated business node, and the business node and its connected edges are removed from the real-time decision influence domain. After completing one pruning, the connection relationship of the remaining business nodes is read again, and the same judgment is performed again until there are no unrelated business nodes that can be removed. Through iterative pruning, irrelevant business nodes in the real-time decision influence domain are deleted, and business nodes that are affected by abnormal states and are related to core business processing are retained. Based on reconstructing the decision domain boundary by retaining business nodes and their corresponding resource contention connections, communication dependency connections, and execution acceptance connections, a minimum stable decision domain is generated. A real-time decision strategy set is then generated according to the selection order, connection relationships, and influence scope of business nodes within the minimum stable decision domain. Specifically: The process of generating the minimum stable decision domain is as follows: Initialize the decision domain boundary: Using the business nodes retained after iterative pruning as the initial node set, read the resource contention connection, communication dependency connection, and execution acceptance connection corresponding to each business node, and construct three types of connection subgraphs; merge the three types of connection subgraphs to obtain the initial decision domain boundary; Determine necessary connection edges: Iterate through each connection edge within the initial decision domain boundary. If deleting a connection edge would cause the two business nodes to be inseparable in terms of resource occupation conflicts, break the communication dependency link, or interrupt the execution order, then mark the connection edge as a necessary connection edge; otherwise, mark the connection edge as an excluded connection edge. Remove excluded edges: Delete all excluded edges to obtain a compressed subgraph consisting of necessary edges and business nodes; if the compressed subgraph still meets the following three conditions: all business nodes are still connected and maintain the original business processing order; resource conflict relationships can be uniquely located; communication dependency links are unbroken and the execution order is complete, then replace the initial decision domain boundary with the compressed subgraph. Repeated boundary compression: Continue to determine necessary connecting edges and remove excluding connecting edges on the compressed subgraph until there are no excluding connecting edges. The resulting compressed subgraph is defined as the minimum stable decision region.

[0036] Based on the retained business nodes, the resource contention connection edges, communication dependency connection edges, and execution acceptance connection edges corresponding to the retained business nodes are reorganized to form a pruned business association structure. The minimum stable decision domain refers to the minimum set of business nodes and connection relationships that can maintain the continuous execution of the target business, the integrity of communication dependencies, and the effectiveness of execution acceptance relationships under the current abnormal state. The business node screening order is read in the minimum stable decision domain to determine the execution order of the real-time decision strategy. The resource contention connection edges are read to determine the local computing resource allocation strategy. The communication dependency connection edges are read to determine the data upload, caching, supplementary reporting, or collaborative transmission strategy. The execution acceptance connection edges are read to determine the abnormal alarm, control trigger, task scheduling, and collaborative processing strategy. The above strategies are combined according to the business node screening order to generate a real-time decision strategy set. The real-time decision strategy set includes data upload control strategy, edge task scheduling strategy, abnormal alarm trigger strategy, and collaborative processing instruction strategy.

[0037] In this embodiment, the step of generating instructions for data upload control, edge task scheduling, anomaly alarm triggering, and collaborative processing based on a set of real-time decision-making strategies includes: Read the real-time decision strategy set and, according to the filtering order, connection relationship and impact range of business nodes in the minimum stable decision domain, split the real-time decision strategy set into data upload control strategy, edge task scheduling strategy, abnormal alarm triggering strategy and collaborative processing instruction strategy; Based on the data upload control strategy, the data to be uploaded is identified in terms of business type, upload priority marking, cache queue allocation and link selection. Abnormal alarm data, control feedback data and target business data within the scope of influence are written into the priority upload queue and non-target business data are written into the delayed upload queue. Based on the edge task scheduling strategy, the terminal processor usage status, memory usage status and local task queue status are read. Business tasks that need to be processed locally within the minimum stable decision domain are allocated to edge computing threads, and the task start time, task end time, task execution status and task output results are recorded. Based on the abnormal alarm triggering strategy, the abnormal type, impact scope identifier and business processing priority sequence in the target business set are read, an abnormal alarm record is generated, and the abnormal alarm record is sent to the main station platform, associated terminal or local control unit. Based on the collaborative processing instruction strategy, the communication dependencies and execution succession relationships in the real-time decision influence domain are read, collaborative processing instructions are generated, and collaborative processing instructions are sent to associated terminals or field execution devices, and the instruction number, sending time, recipient and execution feedback are recorded. The system receives data upload results, edge task execution results, anomaly alarm feedback results, and collaborative processing feedback results. It writes each execution result into the status write-back record, updates the communication link status, terminal computing power status, service load status, and anomaly handling status in the terminal status set, and then enters the next round of edge-side real-time decision-making process.

[0038] Example 1: During a continuous power distribution area operation monitoring cycle, the intelligent fusion terminal receives multi-source status data from the low-voltage acquisition unit, communication module, local edge computing unit, and service queue, collecting a total of 900 sets of status records with a sampling period of 1 second. Each set of status records includes 15 categories of features: voltage, current, active power, frequency, link delay, packet loss rate, retransmission count, processor utilization, memory utilization, processor temperature, processor current, cache queue length, service trigger count, service waiting time, and service urgency level. Traditional processing methods use a fixed window of 30 sampling points and upload the anomaly judgment results to the main station for analysis. This invention performs status change rate identification, adaptive sliding window reconstruction, improved VanillaNet inference, real-time decision influence domain construction, and minimum stable decision domain generation locally on the terminal.

[0039] During the first 300 seconds of stable operation, the voltage fluctuated between 220.1V and 222.3V, the current fluctuated between 42.6A and 46.8A, the link latency ranged from 41ms to 58ms, the packet loss rate ranged from 0.3% to 0.8%, the processor utilization rate ranged from 29% to 37%, and the buffer queue length ranged from 16 to 34 entries. After calculating the difference between the state values ​​at adjacent sampling times under the same feature type, the system obtained a continuous first-order difference sequence, in which the average change in adjacent voltage was 0.28V, the average change in adjacent link latency was 2.1ms, and the average change in adjacent buffer queue entries was 0.4. The system statistically analyzes the unit-time variation amplitude, number of consecutive changes, and consistency of change direction within the observation interval of 40 sampling points. After normalizing the state change rate, it obtains a value of 0.07 to 0.18, which is lower than the slow threshold of 0.2. Therefore, the features of this stage are classified as a stable state, and a coarse-grained sliding window is adopted with a window length of 40 sampling points, a sliding window step size of 20 sampling points, and an overlapping area of ​​10 sampling points. This stage generates 14 coarse-grained sliding windows, compared to 18 windows generated by traditional fixed windows. In this invention, the number of input feature records in the stable state is reduced from 540 to 392, and the feature construction time is reduced from 46ms to 33ms.

[0040] Between 301 and 620 seconds, the business load gradually increased, with the number of business triggers rising from 11 to 42 per minute, the cache queue length increasing from 34 to 119, processor utilization rising from 37% to 63%, link latency rising from 58ms to 128ms, and packet loss rate rising from 0.8% to 2.9%. During this phase, the continuous first-order difference sequence exhibited a consistent unidirectional change, with the cache queue increasing 31 times consecutively and link latency increasing 24 times consecutively. The normalized state change rate was concentrated between 0.26 and 0.55, falling within the 0.2 to 0.6 range. The system divided this phase into a gradual transition state and adopted a transition sliding window with a window length of 20 sampling points, a sliding window step size of 10 sampling points, and an overlapping area of ​​10 sampling points. The transition sliding window retained the business waiting time, link latency growth trend, cache queue growth trend, and processor utilization change trend, generating 32 transition sliding windows. Traditional fixed windows only identify congestion trends 28 seconds after business congestion begins, while this invention identifies congestion trends at 11 seconds, triggering local analysis 17 seconds in advance.

[0041] Between 621 and 690 seconds, abrupt changes occurred in the communication link and service queue. Link latency increased from 128 ms to 316 ms, packet loss rate from 2.9% to 9.4%, retransmission frequency from 17 to 71 per minute, buffer queue length from 119 to 284, abnormal alarm frequency from 2 to 18 per minute, processor temperature from 58.6℃ to 69.5℃, and processor current from 0.83A to 1.15A. During this phase, multiple abrupt increases in amplitude appeared in the continuous first-order difference sequence, with the normalized state change rate reaching 0.74 to 0.92, exceeding the fast threshold of 0.6. The system classified this phase as a sudden change state and adopted a fine-grained sliding window with a window length of 10 sampling points, a sliding window step size of 2 sampling points, and an overlapping area of ​​8 sampling points. A total of 31 fine-grained sliding windows are generated, including points where link latency changes, packet loss rate changes, buffer queue changes, and service emergency level switching changes are all retained. Traditional fixed windows require waiting for 30 sampling points after an anomaly occurs before entering the judgment stage, while this invention generates usable mutation feature records after the 4th sampling point.

[0042] To verify the model's performance, this embodiment constructs 12,000 training samples and 3,000 test samples. The training samples include 4,200 stable states, 3,900 gradual states, and 3,900 abrupt states. In the stable samples, the link latency varies from 3ms to 18ms, the cache queue varies from 0 to 6 records, and the processor utilization varies from 1% to 7%. In the gradual states, the link latency varies from 19ms to 95ms, the cache queue varies from 7 to 88 records, and the processor utilization varies from 8% to 24%. In the abrupt states, the link latency varies from 96ms to 310ms, the cache queue varies from 89 to 260 records, and the processor utilization varies from 25% to 48%. Traditional VanillaNet directly inputs fixed window features; this invention improves VanillaNet by first processing the input through a sliding window-aware state compression module. For fine-grained sliding windows, the feature retention ratio is set to 80% to 100%; for transitional sliding windows, the feature retention ratio is set to 50% to 80%; and for coarse-grained sliding windows, the feature retention ratio is set to 20% to 50%. After 80 rounds of training, the traditional VanillaNet achieved an anomaly detection accuracy of 88.7%, a business priority detection accuracy of 84.9%, and a mutation state false negative rate of 8.9% on the test samples. The improved VanillaNet of this invention achieved an anomaly detection accuracy of 94.6%, a business priority detection accuracy of 92.8%, and a mutation state false negative rate of 2.7%.

[0043] In the anomaly handling after 621 seconds, the sliding window perception state compression module reads fine-grained sliding window records, retaining link latency, packet loss rate, retransmission count, buffer queue length, processor current, and changes in service urgency level. The energy consumption gate reads processor current (1.15A), processor temperature (69.5℃), and computing power utilization (74%), pruning stable-state repetitive channels and retaining channels related to anomaly alarms, link congestion, service urgency level, and state change rate. The latency gate reads terminal response budget (150ms), service urgency level (urgent alarm), and communication link latency (316ms), triggering early output after the third VanillaNet block. The average inference time for traditional fixed-depth networks is 82ms, while the average inference time for this invention is 48ms.

[0044] The system constructs a real-time decision influence domain based on a multi-dimensional decision representation vector. There are 12 business task nodes on-site, including status acquisition, normal upload, anomaly alarm, control feedback, cache replenishment, collaborative processing, link detection, and queue cleanup. Resource usage representation shows that the anomaly alarm and cache replenishment services share a cache queue; communication path representation shows that anomaly alarms, normal uploads, and cache replenishment share an upload link; execution status representation shows that control feedback must be triggered after an anomaly alarm is completed. The system generates 7 resource contention connection edges, 8 communication dependency connection edges, and 6 execution acceptance connection edges, forming a total of 21 connection edges. Based on the anomaly association representation, the system identifies anomaly alarms, control feedback, emergency uploads, cache replenishment, and collaborative processing as the target business set and generates a business processing priority sequence.

[0045] During the impact domain shrinking phase, the system reads target business nodes according to the priority sequence of business processing, using abnormal alarm nodes as core candidate nodes, and expands along resource contention connection edges, communication dependency connection edges, and execution acceptance connection edges. The impact scope identifier limits the search level to 3 layers. Ordinary statistical uploads, low-priority collections, and periodic report businesses that exceed this level and have no connection with the retained business nodes are pruned. After two rounds of iterative pruning, the number of business nodes was reduced from 12 to 5, and the number of connection edges was reduced from 21 to 8, forming a minimum stable decision domain. This decision domain retains abnormal alarms, control feedback, emergency uploads, cached reporting, and collaborative processing to ensure that the abnormal business chain is not broken, while reducing the terminal resources occupied by irrelevant businesses.

[0046] After executing the real-time decision-making strategy set, abnormal alarm data is written to the priority upload queue, ordinary collected data is written to the delayed upload queue, control feedback tasks are assigned to high-priority edge computing threads, and cached reporting tasks are scheduled to be executed after the link is restored. After execution, the terminal local decision-making time is 186ms, while the end-to-end time of the traditional cloud analysis method is 920ms; the peak cache queue count decreased from 342 to 284; within a 900-second period, the traditional method uploads 186MB of raw data, while this invention uploads 73MB of critical data; the traditional method has 17 business processing order conflicts, while this invention has 4 conflicts; the traditional method has 6 instances of missed exception handling, while this invention has 1 instance of missed exception handling.

[0047] 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. A real-time decision-making method for intelligent fusion terminals based on edge computing, characterized in that, include: Collect multi-source status data from intelligent fusion terminals, perform preprocessing on the multi-source status data, and generate a terminal status set; Calculate the continuous first-order difference of various features for the terminal state set, obtain the state change rate, perform sliding window reconstruction processing according to the state change rate, generate an adaptive sliding window corresponding to the state change rate, and construct a dynamic state feature sequence. The dynamic state feature sequence is input into the improved VanillaNet network. A sliding window sensing state compression module is set at the input of VanillaNet to perform adaptive state compression on the dynamic state feature sequence. An energy consumption-delay dual-gated lattice layer is set after each VanillaNet block to output a multi-dimensional decision representation vector. Based on the multidimensional decision representation vector, the resource competition relationship, communication dependency relationship and execution succession relationship among various business tasks are calculated. The real-time decision influence domain is constructed, the target business set affected by the current abnormal state is identified, and the business processing priority sequence and influence scope identifier are generated. The impact domain of real-time decision is shrinked by performing impact domain shrinkage processing. Based on the business processing priority sequence and the impact scope identifier, the business nodes in the impact domain of real-time decision are iteratively filtered to construct the minimum stable decision domain and generate a set of real-time decision strategies. Based on the real-time decision-making strategy set, the system executes data upload control, edge task scheduling, abnormal alarm triggering, and collaborative processing instruction generation. It then writes back the execution results to update the terminal's operating status information and enters the next round of edge-side real-time decision-making process. The output multidimensional decision representation vector includes: Read the state feature records corresponding to each sliding window in the dynamic state feature sequence, extract the sliding window length, sliding window step size, overlap ratio and state change rate identifier in each state feature record, and generate a sliding window perception input record. The sliding window sensing input record is input into the sliding window sensing state compression module at the input end of VanillaNet. The retention ratio of the corresponding state features is determined according to the sliding window length, sliding window step size, overlap ratio and state change rate identifier. High-frequency change features and business urgency features corresponding to abrupt states are retained, and repeated sampling features corresponding to stable states are compressed to generate a compressed state feature sequence. The compressed state feature sequence is input into the improved VanillaNet network. Through VanillaNet blocks, layer-by-layer mapping, non-linear activation and state feature transfer are performed on the compressed state feature sequence to generate intermediate state representations corresponding to each network layer. An energy consumption gate is set after each VanillaNet block. The terminal processor current value, processor temperature value and current computing power occupancy status are read. The number of channels to be reserved is determined based on the processor current value, processor temperature value and current computing power occupancy status. Redundant channels in the intermediate state representation are pruned to generate an energy consumption constraint state representation. A delay gate is set after each VanillaNet block. The terminal response budget, service urgency, and communication link delay are read. Based on the terminal response budget, service urgency, and communication link delay, the number of layers the network continues to calculate or the number of layers is output in advance is determined. Deep control is performed on the energy consumption constraint state representation to generate a multi-dimensional decision representation vector.

2. The real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 1, characterized in that, The multi-source status data includes electrical status data, communication link data, terminal computing power data, and service load data.

3. The real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 1, characterized in that, The preprocessing of multi-source state data to generate a terminal state set includes performing time alignment, anomaly removal, and unified format processing on the multi-source state data to generate a terminal state set.

4. The real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 1, characterized in that, The generation of an adaptive sliding window corresponding to the state change rate, constructing a dynamic state feature sequence, includes: Read electrical status data, communication link data, terminal computing power data and service load data from the terminal status set, and establish corresponding time series according to terminal number, sampling time and feature type; The difference between the state values ​​at adjacent sampling times under the same feature type is calculated to obtain the change of each feature between consecutive sampling times, and a continuous first-order difference sequence is formed in chronological order. Within a set observation interval, the unit time variation amplitude, number of consecutive changes, and consistency of change direction of continuous first-order difference sequences are statistically analyzed, and the state change rate corresponding to various characteristics is determined based on the statistical results. Perform sliding window reconstruction processing based on the rate of state change to generate fine-grained sliding windows, coarse-grained sliding windows, and transitional sliding windows; Electrical operation characteristics, link quality characteristics, computing resource characteristics, and business urgency characteristics are extracted from each fine-grained sliding window, coarse-grained sliding window, and transitional sliding window. The sliding window length, sliding window step size, overlapping area, and state change rate identifier are written into the corresponding feature record and combined in the order of sampling time to form a dynamic state feature sequence.

5. The real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 4, characterized in that, The step of performing sliding window reconstruction processing based on the state change rate includes: When the rate of state change exceeds the fast threshold, the corresponding feature is classified as a mutation state. The minimum window length and minimum step size are used, and the overlap ratio between sliding windows is increased to generate a fine-grained sliding window. When the rate of change of state is lower than the slow threshold, the corresponding feature is classified as a stable state. The maximum window length and maximum step size are used, and the overlap ratio between sliding windows is reduced to generate a coarse-grained sliding window. When the rate of state change is between a fast threshold and a slow threshold, the corresponding feature is divided into a gradual state. An intermediate value between the minimum window length and the maximum window length and a matching intermediate step size are used. The overlap ratio is set between the overlap ratios corresponding to the abrupt state and the stable state to generate a transition sliding window.

6. The real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 1, characterized in that, The generated service processing priority sequence and impact scope identifier include: Read the multidimensional decision representation vector, and according to the business task type, split the multidimensional decision representation vector into resource occupancy representation, communication path representation, execution status representation and exception association representation, and associate each representation with the corresponding business task; Based on resource occupancy, the status of each business task's occupancy of processor, memory, cache queue, and local computing thread is read, and business tasks that simultaneously occupy the same terminal resources and have overlapping execution times are marked as having resource contention. Based on the communication path representation, read the communication interface, transmission link, reporting target and collaborative terminal corresponding to each business task, and mark the business tasks that share the same link, depend on the same reporting path or require the previous communication result as input as having a communication dependency relationship. Based on the execution status, the triggering conditions, execution preconditions, and processing result destinations of each business task are read. Business tasks whose execution results are called by another business task, require control processing after an exception alarm is triggered, or require supplementary reporting processing after the cached task is completed are marked as having an execution succession relationship. Using business tasks as nodes and resource competition relationships, communication dependencies, and execution acceptance relationships as connecting edges, a real-time decision-making influence domain is constructed. Based on the anomaly association representation, the target business set affected by the current anomaly state is located in the real-time decision-making influence domain. A business processing priority sequence is generated according to resource occupation conflict, communication dependency order, and execution acceptance order, and an influence scope identifier is written for the target business set.

7. A real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 6, characterized in that, The construction of the real-time decision influence domain includes: Read the multidimensional decision representation vector and map the resource consumption data, communication path data and execution status data in it to the corresponding business tasks to form a set of business feature records; Compare the resource usage data in the business feature record set to identify business tasks that simultaneously occupy the same processor, memory, or cache resources and have overlapping execution times, generate resource contention connection edges and write them into the resource contention edge list; Compare the communication path data in the business feature record set to identify business tasks that share the same communication link, share the same upload channel, or depend on the same return channel, generate communication dependency connection edges and write them into the communication dependency edge list; The execution status data in the business feature record set is compared to identify business tasks with trigger-call, alarm-control, or cache-report relationships, and execution connection edges are generated and written into the execution connection edge list. Using business tasks as nodes and resource contention connection edges, communication dependency connection edges, and execution acceptance connection edges as edges, the resource contention edge list, communication dependency edge list, and execution acceptance edge list are merged to generate a real-time decision influence domain graph structure. Business nodes affected by the current abnormal state are marked in the real-time decision influence domain graph structure to form a real-time decision influence domain.

8. The real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 1, characterized in that, The construction of the minimum stable decision domain and the generation of a real-time decision strategy set include: Read the real-time decision impact domain, target business set, business processing priority sequence and impact scope identifier, take the business tasks in the target business set as the business nodes to be filtered, and write the business processing priority sequence into the filtering order record of the corresponding business node; Read the business nodes to be filtered in order of priority sequence, and take the business nodes with the highest priority as the core candidate nodes. Read the resource competition connection edges, communication dependency connection edges and execution acceptance connection edges of the core candidate nodes in the real-time decision influence domain. Based on the influence scope identifier, the influence propagation boundary in the real-time decision influence domain is defined, and business nodes located within the influence propagation boundary that have resource competition connections, communication dependency connections, or execution takeover connections with core candidate nodes are marked as reserved business nodes; Iterative pruning is performed on the business nodes in the real-time decision influence domain, removing business nodes that are not referenced by the business processing priority sequence, do not fall within the influence scope identifier limited area, and have no resource competition connection, communication dependency connection, or execution acceptance connection with the retained business nodes from the real-time decision influence domain; The decision domain boundary is reconstructed based on the retained business nodes and their corresponding resource contention connections, communication dependency connections, and execution acceptance connections, generating a minimum stable decision domain. A set of real-time decision strategies is then generated according to the selection order, connection relationship, and influence scope of the business nodes in the minimum stable decision domain.

9. A real-time decision-making method for an intelligent fusion terminal based on edge computing according to claim 1, characterized in that, The process of generating instructions for data upload control, edge task scheduling, anomaly alarm triggering, and collaborative processing based on a set of real-time decision-making strategies includes: Read the real-time decision strategy set and, according to the filtering order, connection relationship and impact range of business nodes in the minimum stable decision domain, split the real-time decision strategy set into data upload control strategy, edge task scheduling strategy, abnormal alarm triggering strategy and collaborative processing instruction strategy; Based on the data upload control strategy, the data to be uploaded is identified in terms of business type, upload priority marking, cache queue allocation and link selection. Abnormal alarm data, control feedback data and target business data within the scope of influence are written into the priority upload queue and non-target business data are written into the delayed upload queue. Based on the edge task scheduling strategy, the terminal processor usage status, memory usage status and local task queue status are read. Business tasks that need to be processed locally within the minimum stable decision domain are allocated to edge computing threads, and the task start time, task end time, task execution status and task output results are recorded. Based on the abnormal alarm triggering strategy, the abnormal type, impact scope identifier and business processing priority sequence in the target business set are read, an abnormal alarm record is generated, and the abnormal alarm record is sent to the main station platform, associated terminal or local control unit. Based on the collaborative processing instruction strategy, the communication dependencies and execution succession relationships in the real-time decision influence domain are read, collaborative processing instructions are generated, and collaborative processing instructions are sent to associated terminals or field execution devices, and the instruction number, sending time, recipient and execution feedback are recorded. The system receives data upload results, edge task execution results, anomaly alarm feedback results, and collaborative processing feedback results. It writes each execution result into the status write-back record, updates the communication link status, terminal computing power status, service load status, and anomaly handling status in the terminal status set, and then enters the next round of edge-side real-time decision-making process.

Images