Low-power operation control method and system for intelligent converged terminals

Abstract

This invention discloses a low-power operation control method and system for intelligent converged terminals, relating to the field of terminal operation data processing technology. The method includes: collecting asynchronous state data from the terminal, preprocessing to generate standardized data, and constructing an asynchronous event sequence; inputting an improved GraphRNN model, performing node expansion, edge-gated aggregation, and weight injection to generate a debt ripple graph; filtering an object set based on the debt ripple graph, constructing a subset of eligible nodes for sealing, and forming a verification structure from the redemption reachable subset; generating a sealing capsule based on the verification structure, constructing a wake-up bridge topology, and generating a redemption trigger node; performing low-frequency keep-alive based on the wake-up bridge, with nodes within the bridge operating at low frequency and objects outside the bridge switching to sealing; restoring the object upon reaching the redemption trigger node, and correcting the low-power control strategy based on the operation results. This invention achieves refined low-power operation control of intelligent converged terminals by constructing an asynchronous wake-up backflush debt ripple graph and combining sealing-redemption cross-validation with a minimum wake-up bridge.

Description

Technical Field

[0001] This invention relates to the field of terminal operation data processing technology, and in particular to a low-power operation control method and system for intelligent converged terminals. Background Technology

[0002] Intelligent converged terminals are typically deployed in scenarios involving power distribution, energy management, edge data acquisition, and field device access. They are used to perform processing tasks such as data acquisition, protocol parsing, edge computing, communication reporting, cache management, and local model inference. As the types of services carried by the terminals increase, the communication modules, acquisition interfaces, edge containers, model inference threads, and cache queues within the terminals will be in a state of alternating operation for extended periods. Low-power operation control has become an important technical means to improve the continuous working capability and operational stability of the terminals.

[0003] Existing low-power control methods mostly rely on CPU utilization, module idle time, battery status, or fixed sleep cycles to determine whether to enter a low-power mode, and reduce energy consumption by reducing processing frequency, shutting down communication threads, pausing data collection polling, or freezing edge applications. While these methods can reduce power consumption in single-task or fixed-cycle task scenarios, their judgment is mainly based on the current running state, lacking comprehensive processing of the asynchronous relationships between task deadlines, cache expiration times, communication keep-alive time limits, container recovery time, and model inference deadlines.

[0004] Therefore, in the scenario of multi-task asynchronous operation of intelligent fusion terminals, existing methods are prone to misjudging communication modules, edge containers, acquisition threads or model inference threads that appear to be idle but are about to be called as objects that can be hibernated. This leads to communication reconnection, cache invalidation, container reloading, model reloading and chained wake-ups after sealing, resulting in power consumption backlash, task timeout and peak power consumption problems. Existing methods lack redemption reachability verification before sealing and minimum wake-up path construction after sealing, making it difficult to achieve stable low-power operation while ensuring data timeliness, communication continuity and model context integrity.

[0005] Therefore, how to provide a low-power operation control method and system for intelligent converged terminals 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 low-power operation control method and system for intelligent converged terminals. This invention fully utilizes improved GraphRNN models, asynchronous event modeling, backflush debt pattern generation, cap-and-redemption cross-validation, and minimum wake-up bridge topology skeleton construction technologies to process asynchronous state data of the intelligent converged terminal during operation. It identifies power backflush risks such as communication reconnection, cache failure, container overload, model overload, and chained wake-ups. It achieves timely recovery and low-frequency keep-alive of capped objects through redeemable capped capsules and low-frequency clock islands. It has the advantages of fine power control, fewer repeated wake-ups, strong communication continuity, low risk of task timeout, and high terminal operation stability.

[0007] A low-power operation control method for an intelligent fusion terminal according to an embodiment of the present invention includes:

[0008] Collect asynchronous state data of the intelligent fusion terminal during operation, preprocess the asynchronous state data to generate a standardized operation dataset, and construct an asynchronous event sequence;

[0009] The asynchronous event sequence is input into the improved GraphRNN model, and the time-series node adaptive expansion, dependency edge gating aggregation and backflush weight dynamic injection are performed to generate an asynchronous wake-up backflush debt pattern.

[0010] Based on the asynchronous wake-up backflip debt pattern, a set of backflip lockout objects and a set of controllable backflip candidate sealing objects are generated, and a sealing qualification subset and a redemption reachable subset are constructed to form a sealing-redemption cross-verification structure.

[0011] Based on the cap-redemption cross-validation structure, redeemable captive capsules are generated for objects that exist simultaneously in the captive eligibility subset and the redemption reachable subset. A minimum wake-up bridge topology skeleton is constructed based on the redeemable captive capsules, and redemption trigger nodes are generated according to the redemption time limit and triggering conditions of each redeemable captive capsule.

[0012] Based on the minimum wake-up bridge topology skeleton, perform low-frequency keep-alive and archive switching, keep the nodes connected to the minimum wake-up bridge topology skeleton running in the low-frequency clock island, and switch the objects not connected to the minimum wake-up bridge topology skeleton to the archive state;

[0013] When the redemption trigger node is reached, the corresponding sealed object is restored according to the recovery action sequence in the redeemable sealed capsule. The low-power operation strategy for the next control cycle is adjusted based on the task delay, number of communication reconnections, number of cache invalidations, number of module wake-ups, and actual power consumption reduction after restoration.

[0014] Optionally, the terminal asynchronous state data includes task running data, module state data, cache state data, communication session data, edge container state data, and model inference state data.

[0015] Optionally, the preprocessing of the terminal asynchronous state data to generate a standardized runtime dataset and construct an asynchronous event sequence includes:

[0016] The asynchronous state data of the terminal is processed for time alignment, anomaly removal, field completion and state normalization. The task deadline, cache validity period, communication keep-alive deadline, data reporting window time, collection trigger time, edge container recovery time and model inference deadline are extracted and constructed into an asynchronous event sequence in chronological order.

[0017] Optionally, generating the asynchronous wake-up backflush ripple pattern includes:

[0018] An improved GraphRNN model is constructed, which includes an asynchronous phase boundary node expansion layer, a sealed dependency gated aggregation layer, and a backflush ripple weight injection layer.

[0019] The asynchronous event sequence is input into the asynchronous phase boundary node augmentation layer in the order of trigger time. The event type, trigger time, latest response time, associated task number, associated module number and recovery time corresponding to each asynchronous event are read to generate task expiration node, cache invalidation node, communication keep-alive node, reporting window node, collection trigger node, container recovery node, model loading node and abnormal interruption node.

[0020] After generating the current recoil twig node in the asynchronous phase boundary node augmentation layer, read the time sequence relationship, task call relationship, module occupation relationship, cache read and write relationship, communication session relationship, container context relationship and model inference relationship between the recoil twig node and the generated recoil twig nodes, and generate candidate recoil twig edges.

[0021] The candidate backflip debt edges are input into the encapsulation dependency gating aggregation layer. The candidate backflip debt edges are gating and filtering according to task call relationship, module occupation relationship, cache read and write relationship, communication session relationship, container context relationship and model inference relationship. Candidate backflip debt edges with any relationship are retained, and candidate backflip debt edges without any relationship are deleted.

[0022] The retained candidate backflip tween edges are input into the backflip tween weight injection layer, and recovery time, recovery power consumption, latest response time and historical backflip records are written to generate communication reconnection backflip edges, cache invalidation backflip edges, container reload backflip edges, model reload backflip edges, task timeout backflip edges and chained wake-up backflip edges.

[0023] Based on the generation order of the anti-reverse bond pattern nodes, the connection relationship of the anti-reverse bond pattern edges, and the attribute fields written to the anti-reverse bond pattern edges, all anti-reverse bond pattern nodes and anti-reverse bond pattern edges are combined into an asynchronous wake-up anti-reverse bond pattern map.

[0024] Optionally, the construction of the sealed eligibility subset and the redemption reachable subset to form a sealed-redemption cross-validation structure includes:

[0025] Read the counter-current debt nodes, counter-current debt edges and edge attribute fields in the asynchronous wake-up counter-current debt pattern graph, and identify the objects that have connection relationships with the communication module, cache page, edge container, model inference thread, acquisition interface and task thread as objects to be judged and sealed.

[0026] For each object to be sealed, read the associated communication reconnection backflip edge, cache invalidation backflip edge, container reload backflip edge, model reload backflip edge, task timeout backflip edge, and chained wake-up backflip edge, extract the corresponding recovery time, recovery power consumption, latest response time, historical backflip count, and associated task deadline, and generate an object backflip judgment record.

[0027] Based on the object backflip determination record, objects to be determined and sealed that have communication session reconstruction, cache validity period expiration, edge container reloading, model parameter reloading, task deadline failure or continuous module wake-up are written into the backflip lockout object set. Objects to be determined and sealed that have not been written into the backflip lockout object set and are in an idle and sealable state are written into the controllable backflip candidate sealed object set.

[0028] For objects in the controllable backflush candidate sealing object set, read the current running status, task emergency flag, module idle status, cache occupancy status, communication reporting status and container context saving status, generate a sealing qualification subset, and read the latest redemption time, recovery action time, cache validity period, communication keep-alive deadline, container recovery entry and model context recovery status to generate a redemption reachable subset;

[0029] Based on the object number, object type, associated module number, latest redemption time, recovery action time, cache validity period, and communication keep-alive deadline, an object-by-object matching is performed on the sealed eligibility subset and the redemption reachable subset. Objects that exist in both the sealed eligibility subset and the redemption reachable subset, objects that exist only in the sealed eligibility subset, and their corresponding matching fields are combined to form a sealed-redemption cross-validation structure.

[0030] Optionally, the step of constructing a minimum wake-up bridge topology skeleton based on redeemable sealed capsules, and generating redemption trigger nodes according to the redemption time limit and triggering conditions of each redeemable sealed capsule, includes:

[0031] Read objects that exist in both the sealed eligibility subset and the redemption reachable subset in the sealed-redemption cross-validation structure, and write the object number, object type, associated module number, sealed start time, redemption time limit, recovery action time, cache validity status, communication keep-alive status, and container context status into the capsule basic record;

[0032] Based on the object type in the capsule's basic record, write the communication session token, protocol sequence number, and keep-alive thread identifier to the communication session; write the cache page index, data generation time, and dirty page bitmap to the cache page; write the container context pointer, container recovery entry, and input cache pointer to the edge container; write the model parameter mapping status, inference intermediate state summary, and model recovery entry to the model inference thread; and write the acquisition trigger condition, last acquisition time, and next acquisition time limit to the acquisition interface to generate a redeemable sealed capsule.

[0033] Read the redemption time limit, triggering condition, keep-alive thread identifier, cache page index, container recovery entry, model recovery entry and recovery action sequence from each redeemable sealed capsule, and determine the low-frequency timing node, communication keep-alive node, cache validity monitoring node, abnormal interruption listening node, capsule redemption scheduling node and recovery execution node used to maintain the recoverable state of the sealed object as wake-up bridge nodes;

[0034] The wake-up bridge edges are established according to the timing triggering order, communication keep-alive relationship, cache expiration monitoring relationship, abnormal interruption triggering relationship, capsule redemption scheduling relationship and recovery action execution order among the wake-up bridge nodes. The low-frequency timing nodes, communication keep-alive nodes, cache expiration monitoring nodes, abnormal interruption listening nodes, capsule redemption scheduling nodes, recovery execution nodes and corresponding wake-up bridge edges are combined into the minimum wake-up bridge topology skeleton.

[0035] Based on the redemption time limit and triggering conditions of each redeemable sealed capsule, a redemption trigger node is generated in the minimum wake-up bridge topology skeleton, and the redemption trigger node is connected to the corresponding capsule redemption scheduling node, recovery execution node and sealed object respectively.

[0036] Optionally, the low-frequency keep-alive and archive switching based on the minimum wake-up bridge topology skeleton includes:

[0037] Read the wake-up bridge nodes and wake-up bridge edges in the minimum wake-up bridge topology skeleton, mark the running objects that have a connection relationship with the redemption trigger node, capsule redemption scheduling node, communication keep-alive node, cache validity monitoring node and abnormal interruption listening node as bridge keep-alive objects, and mark the running objects that have not established a connection relationship with the minimum wake-up bridge topology skeleton as bridge sealed objects.

[0038] For objects kept alive within the bridge, retain the periodic timing task of the low-frequency timing node, the session persistence task of the communication keep-alive node, the cache expiration check task of the cache expiration monitoring node, the trigger listening task of the abnormal interruption listening node, and the redemption queuing task of the capsule redemption scheduling node, and bind the tasks to the low-frequency clock island.

[0039] For objects sealed outside the bridge, the scheduling entry of non-urgent task threads is suspended, the active sending task of communication reporting thread is stopped, the periodic collection task of collection polling thread is turned off, the unnecessary execution context of edge containers is frozen, and the non-urgent inference task of model inference thread is suspended.

[0040] Before switching the external sealed object to the sealed state, write the cache page index, communication session token, protocol sequence number, container context pointer, model parameter mapping state and recovery action sequence into the corresponding redeemable sealed capsule, and write the sealed state, sealed start time and corresponding redemption trigger node into the low power operation strategy table.

[0041] During the operation of the low-frequency clock island, low-frequency keep-alive inspections are performed according to the timing trigger edge, communication keep-alive edge, cache expiration trigger edge, abnormal interruption trigger edge and capsule redemption edge in the minimum wake-up bridge topology skeleton. For bridge-external sealed objects that have reached the redemption trigger node, recovery instructions are output, and bridge-external sealed objects that have not reached the redemption trigger node are kept in a sealed state.

[0042] Optionally, the low-power operation strategy for adjusting the next control cycle based on the recovered task latency, number of communication reconnections, number of cache invalidations, number of module wake-ups, and actual power consumption decrease includes:

[0043] When the low-frequency timing node, communication keep-alive node, cache validity monitoring node, or abnormal interruption listening node detects the arrival of the corresponding redemption trigger node, it reads the recovery action sequence, sealed object number, cache page index, communication session token, protocol sequence number, container context pointer, and model parameter mapping status from the corresponding redeemable sealed capsule, and generates a sealed object recovery queue.

[0044] According to the order of recovery actions in the sealed object recovery queue, cache page recovery, communication session recovery, protocol sequence recovery, edge container context recovery, model parameter mapping recovery, acquisition interface recovery, and task thread recovery are executed sequentially. The recovered objects are then reconnected to the running node corresponding to the minimum wake-up bridge topology skeleton.

[0045] After the archived object is restored, read the restored task execution time, number of communication reconnections, number of cache invalidations, number of module wake-ups, number of edge container reloads, number of model reloads, and actual power consumption reduction results, and generate a restoration result record;

[0046] The recovery results are compared with the current low-power operation strategy in terms of sealing duration, redemption time limit, wake-up trigger order, communication keep-alive cycle, cache retention cycle and low-frequency operation cycle. The sealing limit level is increased for sealed objects that have experienced task timeout, communication reconnection, cache invalidation, container reload, model reload or frequent wake-up. The sealing duration of the next control cycle is extended for sealed objects that have no abnormal recovery results and whose power consumption has decreased to the preset requirement. The revised low-power operation strategy is generated.

[0047] A low-power operation control system for an intelligent fusion terminal according to an embodiment of the present invention includes:

[0048] The asynchronous event building module is used to collect asynchronous state data from the terminal, generate a standardized runtime dataset, and construct an asynchronous event sequence.

[0049] The anti-rebound ripple generation module is used to input asynchronous event sequences into the improved GraphRNN model and generate asynchronous wake-up anti-rebound ripple maps;

[0050] The redemption verification module is used to generate a set of redemption lock-up objects and a set of controllable redemption candidate sealed objects based on the asynchronous wake-up redemption debt pattern, construct a subset of sealed eligibility and a subset of redemption reachability, and form a redemption cross-verification structure.

[0051] The wake-up bridge construction module is used to generate redeemable sealed capsules based on the seal-redemption cross-validation structure, construct the minimum wake-up bridge topology skeleton, and generate redemption trigger nodes;

[0052] The low-frequency sealing and control module is used to perform low-frequency keep-alive and sealing switching based on the minimum wake-up bridge topology skeleton;

[0053] The redemption correction module is used to restore the sealed objects based on the redemption trigger node and correct the low-power operation strategy for the next control cycle.

[0054] The beneficial effects of this invention are:

[0055] This invention collects asynchronous state data of intelligent fusion terminals during operation and uses an improved GraphRNN model to generate asynchronous wake-up backflow ripple maps. It can identify power backflow risks such as communication reconnection, cache failure, container overload, model overload, task timeout, and chained wake-ups before low power consumption regulation. This avoids the coarse sleep judgment based solely on CPU utilization or module idle time in existing technologies, and reduces false storage, frequent wake-ups, and peak power consumption.

[0056] This invention constructs a subset of sealing eligibility and a subset of redemption reachability to form a sealing-redemption cross-validation structure. It generates redeemable sealing capsules for objects that simultaneously meet the sealing eligibility and redemption reachability conditions. This allows communication sessions, cache pages, edge containers, model inference threads, and acquisition interfaces to retain necessary recovery information before entering the sealing state. This reduces the probability of communication session reconstruction, cache data invalidation, model parameter reloading, and edge container restart after sealing, and improves data timeliness, communication continuity, and model context integrity during low-power operation.

[0057] This invention constructs a minimal wake-up bridge topology skeleton through redeemable sealed capsules, keeping nodes connected to the minimal wake-up bridge topology skeleton running within a low-frequency clock island, and switching unconnected objects to a sealed state. This reduces the continuous power consumption of unnecessary threads and modules while retaining necessary timing, keep-alive, listening, and redemption capabilities. Simultaneously, by triggering nodes to restore sealed objects through redemption and adjusting the strategy for the next control cycle based on the restored task latency, communication reconnection count, cache invalidation count, module wake-up count, and actual power consumption reduction, low-power control has a closed-loop correction capability, improving the long-term stability and energy-saving effect of intelligent converged terminals. Attached Figure Description

[0058] 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:

[0059] Figure 1 This is a flowchart of the low-power operation control method for the intelligent fusion terminal proposed in this invention;

[0060] Figure 2 This is a schematic diagram of the low-power operation control system for the intelligent fusion terminal proposed in this invention. Detailed Implementation

[0061] 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.

[0062] refer to Figure 1 Low-power operation control methods for intelligent converged terminals include:

[0063] Collect asynchronous state data of the intelligent fusion terminal during operation, preprocess the asynchronous state data to generate a standardized operation dataset, and construct an asynchronous event sequence;

[0064] The asynchronous event sequence is input into the improved GraphRNN model, and the time-series node adaptive expansion, dependency edge gating aggregation and backflush weight dynamic injection are performed to generate an asynchronous wake-up backflush debt pattern.

[0065] Based on the asynchronous wake-up backflip debt pattern, a set of backflip lockout objects and a set of controllable backflip candidate sealing objects are generated, and a sealing qualification subset and a redemption reachable subset are constructed to form a sealing-redemption cross-verification structure.

[0066] Based on the cap-redemption cross-validation structure, redeemable captive capsules are generated for objects that exist simultaneously in the captive eligibility subset and the redemption reachable subset. A minimum wake-up bridge topology skeleton is constructed based on the redeemable captive capsules, and redemption trigger nodes are generated according to the redemption time limit and triggering conditions of each redeemable captive capsule.

[0067] Based on the minimum wake-up bridge topology skeleton, perform low-frequency keep-alive and archive switching, keep the nodes connected to the minimum wake-up bridge topology skeleton running in the low-frequency clock island, and switch the objects not connected to the minimum wake-up bridge topology skeleton to the archive state;

[0068] When the redemption trigger node is reached, the corresponding sealed object is restored according to the recovery action sequence in the redeemable sealed capsule. The low-power operation strategy for the next control cycle is adjusted based on the task delay, number of communication reconnections, number of cache invalidations, number of module wake-ups, and actual power consumption reduction after restoration.

[0069] In this embodiment, the asynchronous state data of the terminal includes task running data, module state data, cache state data, communication session data, edge container state data, and model inference state data.

[0070] In this embodiment, the preprocessing of the terminal asynchronous state data to generate a standardized runtime dataset and construct an asynchronous event sequence includes:

[0071] The asynchronous state data of the terminal is processed for time alignment, anomaly removal, field completion and state normalization. The task deadline, cache validity period, communication keep-alive deadline, data reporting window time, collection trigger time, edge container recovery time and model inference deadline are extracted and constructed into an asynchronous event sequence in chronological order.

[0072] In this embodiment, generating the asynchronous wake-up backflush ripple pattern includes:

[0073] An improved GraphRNN model is constructed, which includes an asynchronous phase boundary node expansion layer, a sealed dependency gated aggregation layer, and a backflush ripple weight injection layer.

[0074] The improved GraphRNN model is constructed as follows:

[0075] This paper retains the node-level and edge-level recursive frameworks of the GraphRNN model, which recursively generate graph structures according to the node generation order. The node generation part of the original GraphRNN, which only generates ordinary graph nodes based on historical node sequences, is improved into an asynchronous boundary node augmentation layer. This layer enables the generation of corresponding backflip fringe nodes based on task deadlines, cache expiration times, communication keep-alive deadlines, data reporting window times, acquisition trigger times, edge container recovery times, and model inference deadlines in asynchronous event sequences. The edge generation part of the original GraphRNN, which is oriented towards general node connections, is improved into a dependency-gated aggregation layer. After generating candidate backflip fringe edges, this layer removes candidate edges without real dependencies based on task call relationships, module occupancy relationships, cache read / write relationships, communication session relationships, container context relationships, and model inference relationships. A backflip fringe weight injection layer is added after the edge-level recursive output of the original GraphRNN and connected to the output of the dependency-gated aggregation layer. This layer is used to write recovery time, recovery power consumption, latest response time, and historical backflip records to the retained candidate backflip fringe edges, resulting in the improved GraphRNN model.

[0076] The asynchronous event sequence is input into the asynchronous phase boundary node augmentation layer in the order of trigger time. The event type, trigger time, latest response time, associated task number, associated module number and recovery time corresponding to each asynchronous event are read to generate task expiration node, cache invalidation node, communication keep-alive node, reporting window node, collection trigger node, container recovery node, model loading node and abnormal interruption node.

[0077] After generating the current recoil twig node in the asynchronous phase boundary node augmentation layer, read the time sequence relationship, task call relationship, module occupation relationship, cache read and write relationship, communication session relationship, container context relationship and model inference relationship between the recoil twig node and the generated recoil twig nodes, and generate candidate recoil twig edges.

[0078] The candidate backflip debt edges are input into the encapsulation dependency gating aggregation layer. The candidate backflip debt edges are gating and filtering according to task call relationship, module occupation relationship, cache read and write relationship, communication session relationship, container context relationship and model inference relationship. Candidate backflip debt edges with any relationship are retained, and candidate backflip debt edges without any relationship are deleted.

[0079] The retained candidate backflip tween edges are input into the backflip tween weight injection layer, and recovery time, recovery power consumption, latest response time and historical backflip records are written to generate communication reconnection backflip edges, cache invalidation backflip edges, container reload backflip edges, model reload backflip edges, task timeout backflip edges and chained wake-up backflip edges.

[0080] Based on the generation order of the anti-reverse bond tread nodes, the connection relationship of the anti-reverse bond tread edges, and the attribute fields written to the anti-reverse bond tread edges, all anti-reverse bond tread nodes and anti-reverse bond tread edges are combined into an asynchronous wake-up anti-reverse bond tread graph. Specifically, the combination of all anti-reverse bond tread nodes and anti-reverse bond tread edges into an asynchronous wake-up anti-reverse bond tread graph is as follows:

[0081] Read the backflush ripple node sequence output by the phase boundary sequence cache unit, and concatenate the trigger time, node type, associated object number, latest response time, and recovery time of each node in sequence to generate a 5-dimensional node vector; write all node vectors into the node matrix N in the order of node generation; calculate the absolute value of the trigger time difference between adjacent nodes. If the time difference does not exceed 3 seconds and the two nodes have a task call or module occupancy relationship, write 1 at the corresponding position in the node matrix N to generate a temporal dependency adjacency matrix; for every two nodes in the node matrix N, if they share a communication session, cache page, or container context, add them to the communication dependency adjacency matrix... Write 1 to the corresponding position in the cache dependency adjacency matrix or container dependency adjacency matrix; perform element-wise union of the three types of dependency adjacency matrices to obtain the comprehensive dependency adjacency matrix A; traverse the positions with a value of 1 in the comprehensive dependency adjacency matrix A, and generate backflip debt edges with the corresponding node index as the start and end points. Write recovery time, recovery power consumption, latest response time difference, and historical backflip count to each edge, and mark them as communication reconnection, cache invalidation, container reload, model reload, task timeout, or chained wake-up backflip edges according to the node types at both ends; finally, combine the node matrix N, the comprehensive dependency adjacency matrix A, and the edge attribute table to save it as an asynchronous wake-up backflip debt graph.

[0082] The asynchronous phase boundary node expansion layer includes:

[0083] The event vector receiving unit is used to receive asynchronous event sequences and convert task deadlines, cache validity periods, communication keep-alive deadlines, data reporting window times, acquisition trigger times, edge container recovery times, and model inference deadlines into event node input vectors.

[0084] The phase boundary type determination unit is used to determine the event node input vector as a task expiration phase boundary, cache invalidation phase boundary, communication keep-alive phase boundary, reporting window phase boundary, acquisition trigger phase boundary, container recovery phase boundary, model loading phase boundary, or abnormal interruption phase boundary based on the event source and triggering object.

[0085] The node adaptive expansion unit is used to expand and generate corresponding anti-friction ripple nodes in the order of task, cache, communication, container and model when multiple terminal running objects are associated with the same asynchronous event;

[0086] The redemption time limit writing unit is used to write the latest response time, recovery time, associated task number, associated module number, cache validity period, communication keep-alive time limit and model inference deadline to the expanded backflush debt tread node.

[0087] The phase boundary order caching unit is used to cache the generated reversal ripple nodes in the order of trigger time, and to provide node order, node type and node attributes for the generation of reversal ripple edges;

[0088] The asynchronous phase boundary node expansion layer receives a sequence of asynchronous events ordered by trigger time, and reads the event type, trigger time, latest response time, associated task number, associated module number, and recovery time of each asynchronous event item by item. When the event source is the task deadline, a task expiration node is generated; when the event source is the cache validity period, a cache invalidation node is generated; when the event source is the communication keep-alive deadline, a communication keep-alive node is generated; when the event source is the data reporting window time, a reporting window node is generated; when the event source is the acquisition trigger time, an acquisition trigger node is generated; when the event source is the edge container recovery time, a container recovery node is generated; when the event source is the model inference deadline, a model loading node is generated; and when the event source is an external alarm or master station request, an abnormal interruption node is generated. If the same asynchronous event is associated with multiple terminal running objects, it is expanded into multiple backflip ripple nodes in the order of task, cache, communication, container, and model. A common trigger time and different associated object numbers are written into each expanded node. After the node expansion is completed, all backflip ripple nodes are written into the phase boundary sequence cache unit in the order of generation.

[0089] The dependency-gated aggregation layer includes:

[0090] The historical node reading unit is used to read the historical backflip debt pattern nodes generated in the phase boundary sequence cache unit, and obtain the node type, associated task, associated module, cache page, communication session and container context information of the historical backflip debt pattern nodes;

[0091] The candidate edge generation unit is used to establish candidate backflip debt tread edges between the current backflip debt tread node and the historical backflip debt tread nodes. The candidate backflip debt tread edges include communication reconnection candidate edges, cache invalidation candidate edges, container reload candidate edges, model reload candidate edges, task timeout candidate edges, and chained wake-up candidate edges.

[0092] The real dependency gating unit is used to gating and filtering candidate backflush debt edges based on task call relationships, module usage relationships, cache read and write relationships, communication session relationships, container context relationships, and model inference relationships;

[0093] The archived impact aggregation unit is used to aggregate candidate backflush debt edges that have passed the gating screening, and to record the recovery usage, session usage, cache usage, and inference usage of the archived object on the associated object;

[0094] The invalid edge removal unit is used to delete candidate backflip edge patterns that do not have a real task module dependency, do not share cache pages, do not share communication sessions, do not share container contexts, and do not belong to the same model inference link.

[0095] After the asynchronous phase boundary node expansion layer generates the current backflip debt tread node, the dependency gating aggregation layer reads the temporal continuity relationship between the current backflip debt tread node and the previously generated historical backflip debt tread nodes, and reads whether they call the same communication module, access the same cache page, occupy the same edge container, depend on the same model inference engine, or belong to the same task execution chain. The candidate edge generation unit establishes candidate backflip debt tread edges between the current backflip debt tread node and historical backflip debt tread nodes. The real dependency gating unit checks the task call relationship, module occupation relationship, cache read / write relationship, communication session relationship, container context relationship, and model inference relationship of each candidate backflip debt tread edge item by item. When a candidate backflip debt tread edge satisfies any real dependency relationship, the candidate backflip debt tread edge is retained and sent to the impact aggregation unit. When a candidate backflip debt tread edge does not satisfy any real dependency relationship, it is deleted by the invalid edge removal unit. The impact aggregation unit records the connection relationship between the retained edge and the affected object.

[0096] The anti-hedging bond weight injection layer includes:

[0097] The archived object attribute reading unit is used to read the recovery time, recovery power consumption, latest response time, archive start time, cache validity period, communication keep-alive time limit and model inference deadline of the archived object corresponding to the retained candidate backflush debt edge;

[0098] The historical backflip record reading unit is used to read the number of communication reconnections, cache invalidations, container reloads, model reloads, task timeouts, and module wake-ups of the corresponding sealed object in the historical control cycle.

[0099] The backflip type marking unit is used to mark the candidate backflip edge as a communication reconnection backflip edge, cache invalidation backflip edge, container reload backflip edge, model reload backflip edge, task timeout backflip edge or chain wake-up backflip edge according to the node type and sealed object type of the candidate backflip edge connection.

[0100] The recoil attribute writing unit is used to write recovery time, recovery power consumption, latest response time, historical recoil records, associated sealed objects, and recoil trigger source to the recoil debt edge;

[0101] The graph assembly unit is used to assemble all the backflip twill nodes and backflip twill edges into an asynchronous wake-up backflip twill graph according to the generation order of the backflip twill nodes, the connection relationship of the backflip twill edges, and the attribute fields of the backflip twill edges.

[0102] The backflip twig weight injection layer receives candidate backflip twig edges preserved by the archived dependency-gated aggregation layer. First, the archived object attribute reading unit reads the recovery time, recovery power consumption, latest response time, cache validity period, communication keep-alive time limit, and model inference deadline of the archived object corresponding to the candidate backflip twig edge. Then, the historical backflip record reading unit reads records of communication reconnection, cache invalidation, container reload, model reload, task timeout, and module wake-up that occurred for the same object within the historical control cycle. The backflip type marking unit determines the backflip edge category based on the type of the nodes at both ends of the candidate backflip twig edge. When the edge connects the communication keep-alive node and the reporting window node, it is marked as a communication reconnection backflip edge; when the edge... When an edge connects a cache invalidation node and a task expiration node, it is marked as a cache invalidation backflip edge. When an edge connects a container recovery node and a model loading node, it is marked as a container reload backflip edge. When an edge connects a model loading node and an inference deadline node, it is marked as a model reload backflip edge. When an edge connects a task expiration node and a recovery execution node, it is marked as a task timeout backflip edge. When multiple nodes are triggered consecutively through the same cache object, it is marked as a chained wake-up backflip edge. Subsequently, the backflip attribute writing unit writes the recovery time, recovery power consumption, latest response time, and historical backflip records into the corresponding backflip debt pattern edge. The graph assembly unit combines all backflip debt pattern nodes and backflip debt pattern edges to generate an asynchronous wake-up backflip debt pattern graph.

[0103] In this embodiment, the construction of the sealed eligibility subset and the redemption reachable subset to form a sealed-redemption cross-validation structure includes:

[0104] Read the counter-current debt nodes, counter-current debt edges and edge attribute fields in the asynchronous wake-up counter-current debt pattern graph, and identify the objects that have connection relationships with the communication module, cache page, edge container, model inference thread, acquisition interface and task thread as objects to be judged and sealed.

[0105] For each object to be sealed, read the associated communication reconnection backflip, cache invalidation backflip, container reload backflip, model reload backflip, task timeout backflip, and chained wake-up backflip, and extract the corresponding recovery time, recovery power consumption, latest response time, historical backflip count, and associated task deadline to generate an object backflip determination record. Specifically, the generation of the object backflip determination record involves:

[0106] For the object to be sealed, retrieve the communication reconnection backflush edge, cache invalidation backflush edge, container reload backflush edge, model reload backflush edge, task timeout backflush edge and chained wake-up backflush edge with the object as the sealing endpoint from the edge index table of the asynchronous wake-up backflush debt graph, and write the six types of backflush edges into the local edge buffer in sequence.

[0107] Traverse the side buffer, read the difference between the recovery time and the sealing start time for each backflip edge to obtain the one-sided recovery delay, write the one-sided recovery delay into the recovery time field, read the difference between the latest response time and the current system time to obtain the one-sided response margin, write the response margin into the latest response time field, accumulate the number of times the same type of backflip edge appears in the last three control cycles to obtain the historical backflip count, and write it into the historical backflip count field;

[0108] Read the peak power consumption of the sealed object in the most recent backflush type, subtract the current static power consumption to get the one-sided recovery power consumption, write the value to the recovery power consumption field, read the task deadline of the backflush edge starting task node, and write it to the associated task deadline field.

[0109] After completing the traversal of all backflip edges, the recovery time, recovery power consumption, latest response time, historical backflip count and associated task deadline corresponding to the six types of backflip edges are merged in the form of a sequential vector and stored in the object backflip judgment record table, which corresponds to the unique index of the object to be judged and sealed.

[0110] Based on the object backflip determination record, objects to be determined and sealed that have communication session reconstruction, cache validity period expiration, edge container reloading, model parameter reloading, task deadline failure or continuous module wake-up are written into the backflip lockout object set. Objects to be determined and sealed that have not been written into the backflip lockout object set and are in an idle and sealable state are written into the controllable backflip candidate sealed object set.

[0111] For objects in the controllable backflush candidate sealing object set, read the current running status, task urgent flag, module idle status, cache occupancy status, communication reporting status, and container context saving status to generate a sealing eligibility subset. Then, read the latest redemption time, recovery action time, cache validity period, communication keep-alive deadline, container recovery entry point, and model context recovery status to generate a redemption reachable subset, where:

[0112] Generate a subset of the eligibility for sealing, specifically as follows:

[0113] For each object in the controllable backflush candidate sealing object set, read the current running status; if the object's task urgency is marked as non-urgent, the module idle status is idle, the queue length corresponding to the cache occupancy status does not exceed 30% of the total cache capacity, the communication reporting status is no data to be reported, and the container context saving status is marked as persistent, then add a new record to the local sealing qualification index table. The record content includes the object identifier, sealing start timestamp, and current static standby power consumption value; after completing the traversal, gather all records in the sealing qualification index table to obtain the sealing qualification subset;

[0114] Generate a redemption reachable subset, specifically:

[0115] The set of controllable backflush candidate archived objects is traversed again. For each object, the latest redemption timestamp, total recovery time, cache expiration, communication keep-alive deadline, remaining storage time of container context, and remaining storage time of model context are read. The difference between the current system time and the latest redemption timestamp is calculated. If the difference is greater than the total recovery time, and the total recovery time does not exceed the cache expiration, remaining communication keep-alive time, remaining storage time of container context, and remaining storage time of model context, respectively, a new record is added to the redemption reachable index table. The record contains the object identifier, the estimated remaining redemption time, and the estimated recovery energy consumption. After the traversal is completed, all records in the redemption reachable index table are aggregated to obtain the redemption reachable subset.

[0116] Based on object ID, object type, associated module ID, latest redemption time, recovery action time, cache validity period, and communication keep-alive deadline, a per-object matching process is performed on the archive eligibility subset and the redemption reachable subset. Objects that exist in both the archive eligibility subset and the redemption reachable subset, objects that exist only in the archive eligibility subset, and their corresponding matching fields are combined to form an archive-redemption cross-validation structure. Specifically, the archive-redemption cross-validation structure is as follows:

[0117] The eligible subsets for sealing are written into the first index table in ascending order of object number. The object type, associated module number, sealing start time and static standby power consumption are appended to each record. The redemption reachable subsets are written into the second index table in ascending order of object number. The latest redemption time, recovery action time, cache validity period and communication keep-alive deadline are appended to each record.

[0118] Using the object number as the primary key, perform row-by-row matching between the first and second index tables. If the same object number exists in both tables, add a new record of redeemable sealed objects in the cross-match result table and merge the fields of the two tables. If the object number only appears in the first index table, add a new record of sealed locked objects in the cross-match result table and fill in the redemption field with a missing flag.

[0119] After the matching is completed, all records in the cross-result table are reordered in ascending order by object type and associated module number. A verification flag is written for each record: a flag of 1 indicates redeemable sealing, and a flag of 0 indicates sealing and locking. The cross-result table and the verification flag are saved as a sealed-redemption cross-verification structure.

[0120] In this embodiment, the step of constructing a minimum wake-up bridge topology skeleton based on redeemable sealed capsules, and generating redemption trigger nodes according to the redemption time limit and triggering conditions of each redeemable sealed capsule, includes:

[0121] Read objects that exist in both the sealed eligibility subset and the redemption reachable subset in the sealed-redemption cross-validation structure, and write the object number, object type, associated module number, sealed start time, redemption time limit, recovery action time, cache validity status, communication keep-alive status, and container context status into the capsule basic record;

[0122] Based on the object types in the capsule's basic record, the communication session token, protocol sequence number, and keep-alive thread identifier are written to the communication session; the cache page index, data generation time, and dirty page bitmap are written to the cache page; the container context pointer, container recovery entry, and input cache pointer are written to the edge container; the model parameter mapping state, inference intermediate state summary, and model recovery entry are written to the model inference thread; and the acquisition trigger condition, last acquisition time, and next acquisition time limit are written to the acquisition interface. A redeemable sealed capsule is then generated. Specifically, the generation of a redeemable sealed capsule involves:

[0123] Read the object type, object identifier, sealing start time, and redemption time limit from the capsule's basic record, and establish the capsule's primary key;

[0124] When the object type is a communication session, write the communication session token field to the communication control area, write the protocol sequence number field and verify the increment integrity, write the keep-alive thread identifier field and record the thread priority;

[0125] When the object type is a cache page, write the cache page index field in the cache control area, write the data generation time field and align the timestamp to the second level, write the dirty page bitmap field and mark the start and end offsets of the dirty page;

[0126] When the object type is an edge container, write the container context pointer field to the container control area, write the container recovery entry field and mark the entry offset address, write the input buffer pointer field and record the current write pointer position;

[0127] When the object type is a model inference thread, write the model parameter mapping status field in the model control area, write the inference intermediate state summary field and perform a 32-byte CRC check, write the model recovery entry field and record the entry offset address.

[0128] When the object type is a data collection interface, write the data collection trigger condition field in the data collection control area, write the previous data collection time field and align the timestamp to the second level, and write the next data collection time limit field.

[0129] After writing all fields, the starting offsets and lengths of the communication control area, cache control area, container control area, model control area, and acquisition control area are written into the segment index table and encapsulated together with the capsule primary key into a binary structure to generate a redeemable sealed capsule.

[0130] Read the redemption time limit, triggering condition, keep-alive thread identifier, cache page index, container recovery entry, model recovery entry and recovery action sequence from each redeemable sealed capsule, and determine the low-frequency timing node, communication keep-alive node, cache validity monitoring node, abnormal interruption listening node, capsule redemption scheduling node and recovery execution node used to maintain the recoverable state of the sealed object as wake-up bridge nodes;

[0131] Wake-up bridge edges are established according to the timing triggering order, communication keep-alive relationship, cache expiration monitoring relationship, abnormal interruption triggering relationship, capsule redemption scheduling relationship, and recovery action execution order among wake-up bridge nodes. Low-frequency timing nodes, communication keep-alive nodes, cache expiration monitoring nodes, abnormal interruption listening nodes, capsule redemption scheduling nodes, recovery execution nodes, and their corresponding wake-up bridge edges are combined to form a minimal wake-up bridge topology skeleton. Specifically, this minimal wake-up bridge topology skeleton is as follows:

[0132] Read the low-frequency timing node, communication keep-alive node, cache expiration monitoring node, abnormal interruption listening node, capsule redemption scheduling node, and recovery execution node in sequence. Write the six types of nodes into the bridge node sequence table in a fixed order of timing, keep-alive, monitoring, listening, scheduling, and recovery. Using the bridge node sequence table as an index, establish timing trigger edges, keep-alive acceptance edges, cache expiration monitoring edges, abnormal interruption trigger edges, and scheduling-recovery acceptance edges between nodes in sequence. Write the start number, end number, and trigger condition of each edge into the bridge edge list. Traverse the bridge edge list and keep only one edge with the same start and end points and the same trigger condition. Delete the duplicate edges from the list.

[0133] Perform a topological traversal on the node sequence list within the bridge. If a closed loop path is detected, delete the lowest priority edge in the closed loop and regenerate the traversal order until there are no closed loops. After completing the acyclic process, write an edge number for each edge within the bridge. Record the edge number at the starting node and the ending node to generate an index matrix within the bridge. Integrate the node sequence list, edge list, and index matrix within the bridge and save them as the minimum wake-up bridge topological skeleton.

[0134] Based on the redemption time limit and triggering conditions of each redeemable sealed capsule, a redemption trigger node is generated in the minimum wake-up bridge topology skeleton. This redemption trigger node is then connected to the corresponding capsule redemption scheduling node, recovery execution node, and sealed object. Specifically, the generation of the redemption trigger node in the minimum wake-up bridge topology skeleton is as follows:

[0135] Sequentially read each capsule redemption scheduling node in the minimum wake-up bridge topology skeleton, locate the associated redeemable sealed capsules, extract the redemption time limit and triggering conditions for the sealed capsules, convert the redemption time limit into an absolute timestamp in integer milliseconds, and write the triggering conditions into trigger flags according to three categories: communication keep-alive expiration, cache expiration, and external abnormal triggering. Create a new redemption trigger node record for the current capsule redemption scheduling node, and write the node number, associated sealed object number, redemption timestamp, and trigger flag into the record;

[0136] Point the incoming edge of the redemption trigger node to the capsule redemption scheduling node. Establish a recovery trigger edge between the redemption trigger node and the recovery execution node. Write the recovery action sequence index into the recovery trigger edge. If the trigger conditions corresponding to the same capsule object include both communication keep-alive expiration and cache expiration, they are merged into a single redemption trigger node. Write both types of tags into the trigger tag field at the same time.

[0137] After processing all capsule redemption scheduling nodes, all redemption trigger nodes are sorted in ascending order of redemption timestamp. Time-connected edges are established for redemption trigger nodes with adjacent redemption intervals of less than 500 milliseconds. Finally, the redemption trigger nodes, recovery trigger edges, and time-connected edges are written into the bridge index matrix.

[0138] In this embodiment, the step of performing low-frequency keep-alive and archive switching based on the minimum wake-up bridge topology skeleton includes:

[0139] Read the wake-up bridge nodes and wake-up bridge edges in the minimum wake-up bridge topology skeleton, mark the running objects that have a connection relationship with the redemption trigger node, capsule redemption scheduling node, communication keep-alive node, cache validity monitoring node and abnormal interruption listening node as bridge keep-alive objects, and mark the running objects that have not established a connection relationship with the minimum wake-up bridge topology skeleton as bridge sealed objects.

[0140] For bridge-based keep-alive objects, the following tasks are retained: periodic timing tasks of low-frequency timing nodes, session persistence tasks of communication keep-alive nodes, cache expiration check tasks of cache expiration monitoring nodes, trigger listening tasks of abnormal interruption listening nodes, and redemption queuing tasks of capsule redemption scheduling nodes. These tasks are bound to the low-frequency clock island, which refers to:

[0141] Within the intelligent converged terminal, a separate set of processor clock domains is allocated for low-power keep-alive tasks:

[0142] The master clock frequency of this clock domain is locked at a fixed value far below the normal operating frequency and remains constant without fluctuating with the dynamic frequency modulation strategy.

[0143] Within the clock domain, only the lightweight threads or interrupt service routines corresponding to the low-frequency timing node, communication keep-alive node, cache validity monitoring node, abnormal interrupt listening node, and capsule redemption scheduling node are connected, while all other high-load threads are disconnected from the peripheral clock logic.

[0144] The clock island and the main frequency clock domain are separated by an isolation gated clock buffer. Tasks entering the clock island will not trigger high-frequency phase-locked loop up-frequency during execution, nor will they wake up the sealed processing core. By maintaining a low and stable clock frequency and a minimal number of active units, the low-frequency clock island can maintain the necessary timing, keep-alive, listening and queuing functions at microwatt-level power consumption.

[0145] For objects sealed outside the bridge, the scheduling entry of non-urgent task threads is suspended, the active sending task of communication reporting thread is stopped, the periodic collection task of collection polling thread is turned off, the unnecessary execution context of edge containers is frozen, and the non-urgent inference task of model inference thread is suspended.

[0146] Before switching the external sealed object to the sealed state, the cache page index, communication session token, protocol sequence number, container context pointer, model parameter mapping state, and recovery action sequence in the corresponding redeemable sealed capsule are written. The sealed state, sealed start time, and corresponding redemption trigger node are written to the low-power operation policy table. The construction of the low-power operation policy table is as follows:

[0147] Create a new low-power operation strategy master table in the system storage area. The fields of the master table are, in order: object identifier, sealed status bit, sealed start timestamp, redemption trigger node number, cache page index, communication session token, protocol sequence number, container context pointer, model parameter mapping status, and recovery action sequence index. In addition to the master table, three sub-tables are created: a cache mapping sub-table for quickly retrieving cache page location by object identifier-cache page index key-value pair, a communication mapping sub-table for quickly retrieving session status by object identifier-communication session token key-value pair, and a model mapping sub-table for quickly retrieving model mapping information by object identifier-model parameter mapping status key-value pair.

[0148] Before an object outside the bridge is switched to the sealed state, the object identifier is first written to the main table and the sealed state position is set to 1. The sealed start timestamp and redemption trigger node number are then written. The corresponding fields are then written according to the object type: if it is a communication session object, the communication session token and protocol sequence number are written and registered in the communication mapping sub-table; if it is a cache page object, the cache page index is written and registered in the cache mapping sub-table; if it is an edge container object, the container context pointer is written; if it is a model inference thread, the model parameter mapping state is written and registered in the model mapping sub-table; finally, the offset index of the corresponding recovery action sequence in the action sequence table is written.

[0149] After all fields have been written, a foreign key relationship is established between the main table and the three sub-tables through the object identifier. An ascending index is created for the archive start timestamp field, and a hash index is created for the redemption trigger node number field to support batch scanning by time and fast location by trigger node.

[0150] During the operation of the low-frequency clock island, low-frequency keep-alive inspections are performed according to the timing trigger edge, communication keep-alive edge, cache expiration trigger edge, abnormal interruption trigger edge and capsule redemption edge in the minimum wake-up bridge topology skeleton. For bridge-external sealed objects that have reached the redemption trigger node, recovery instructions are output, and bridge-external sealed objects that have not reached the redemption trigger node are kept in a sealed state.

[0151] In this embodiment, the low-power operation strategy for adjusting the next control cycle based on the recovered task latency, number of communication reconnections, number of cache invalidations, number of module wake-ups, and actual power consumption decrease includes:

[0152] When a low-frequency timing node, communication keep-alive node, cache expiration monitoring node, or abnormal interruption listening node detects the arrival of the corresponding redemption trigger node, it reads the recovery action sequence, sealed object number, cache page index, communication session token, protocol sequence number, container context pointer, and model parameter mapping status from the corresponding redeemable sealed capsule, and generates a sealed object recovery queue. Specifically, generating the sealed object recovery queue involves:

[0153] The system sequentially reads the redemption trigger node numbers triggered by the low-frequency timing node, communication keep-alive node, cache expiration monitoring node, and abnormal interruption listening node. It then writes the object indices triggered within the same time period into the recovery queue buffer according to their trigger times. For each sealed object in the buffer, it first retrieves a redeemable sealed capsule from the capsule pool by object number, parses and caches the recovery action sequence entry address, and sequentially reads the cache page index, communication session token, protocol sequence number, container context pointer, and model parameter mapping status. These are then written into the recovery field record of the same object in a fixed order: cache, communication, container, model, and task. After writing the fields, a one-to-one mapping is established between the recovery action sequence entry address and the recovery field record using the object number, and this mapping is stored in the recovery object dictionary. Finally, all objects in the recovery queue buffer are reordered in ascending order of trigger time and descending order of object priority to generate a sealed object recovery queue.

[0154] According to the order of recovery actions in the sealed object recovery queue, cache page recovery, communication session recovery, protocol sequence recovery, edge container context recovery, model parameter mapping recovery, acquisition interface recovery, and task thread recovery are executed sequentially. The recovered objects are then reconnected to the running node corresponding to the minimum wake-up bridge topology skeleton.

[0155] After the archived object is restored, read the restored task execution time, number of communication reconnections, number of cache invalidations, number of module wake-ups, number of edge container reloads, number of model reloads, and actual power consumption reduction results, and generate a restoration result record;

[0156] The recovery results are compared with the current low-power operation strategy in terms of sealing duration, redemption time limit, wake-up trigger order, communication keep-alive cycle, cache retention cycle and low-frequency operation cycle. The sealing limit level is increased for sealed objects that have experienced task timeout, communication reconnection, cache invalidation, container reload, model reload or frequent wake-up. The sealing duration of the next control cycle is extended for sealed objects that have no abnormal recovery results and whose power consumption has decreased to the preset requirement. The revised low-power operation strategy is generated.

[0157] refer to Figure 2 The low-power operation control system for intelligent converged terminals includes:

[0158] The asynchronous event building module is used to collect asynchronous state data from the terminal, generate a standardized runtime dataset, and construct an asynchronous event sequence.

[0159] The anti-rebound ripple generation module is used to input asynchronous event sequences into the improved GraphRNN model and generate asynchronous wake-up anti-rebound ripple maps;

[0160] The redemption verification module is used to generate a set of redemption lock-up objects and a set of controllable redemption candidate sealed objects based on the asynchronous wake-up redemption debt pattern, construct a subset of sealed eligibility and a subset of redemption reachability, and form a redemption cross-verification structure.

[0161] The wake-up bridge construction module is used to generate redeemable sealed capsules based on the seal-redemption cross-validation structure, construct the minimum wake-up bridge topology skeleton, and generate redemption trigger nodes;

[0162] The low-frequency sealing and control module is used to perform low-frequency keep-alive and sealing switching based on the minimum wake-up bridge topology skeleton;

[0163] The redemption correction module is used to restore the sealed objects based on the redemption trigger node and correct the low-power operation strategy for the next control cycle.

[0164] Example 1: During a continuous 24-hour low-power operation cycle of an intelligent fusion terminal, the terminal simultaneously runs 4 acquisition interfaces, 3 communication sessions, 1 protocol parsing process, 2 edge containers, 6 model inference threads, and 512 pages of local cache. The system collects asynchronous status data of the terminal every 5 seconds, obtaining a total of 17,280 raw records. Among the raw records, 10,342 records (59.8%) have a CPU utilization rate below 20%, but among these, 936 records have a communication keep-alive time of less than 30 seconds, 1,284 records have a cache validity period of less than 60 seconds, and 427 records have an edge container recovery time of more than 4 seconds. If sleep mode is judged solely based on CPU utilization and module idle time, a large number of critical communication, critical cache, and critical container states will be misjudged as sealable states.

[0165] The system first preprocesses 17,280 raw records, unifying task timestamps, cache timestamps, communication timestamps, and container status timestamps to a 5-second granularity. It removes 214 duplicate records, completes 73 missing fields, and eliminates 31 abnormal records with empty recovery times, ultimately obtaining 17,035 standardized operation records. Subsequently, the system extracts task deadlines, cache expiration dates, communication keep-alive deadlines, data reporting window times, data acquisition trigger times, edge container recovery times, and model inference deadlines, constructing an asynchronous event sequence in ascending order of time. This sequence includes 2,436 task expiration events, 1,817 cache expiration events, 692 communication keep-alive events, 288 reporting window events, 8,640 data acquisition trigger events, 156 container recovery events, 214 model loading events, and 37 abnormal interruption events.

[0166] In the improved GraphRNN model processing stage, the asynchronous phase boundary node augmentation layer reads the asynchronous event sequence item by item. Taking a communication keep-alive event at 3815 seconds as an example, its event type is communication keep-alive, the associated communication session number is C-02, the associated communication module number is M-01, the remaining communication keep-alive time is 24 seconds, the communication module recovery time is 7 seconds, and the protocol recovery time is 6 seconds. Based on this, the system generates one communication keep-alive node. Taking a cache expiration event at 3820 seconds as an example, its cache page number is P-147, the remaining validity period is 42 seconds, the cache read time is 3 seconds, and the associated task's remaining processing time is 18 seconds. The system generates one cache invalidation node. After completing the full sequence traversal, the system generates a total of 1348 backflip ripple nodes.

[0167] The dependency-gated aggregation layer establishes 3526 candidate backflip edge diagrams among 1348 backflip edge nodes. The system reads the task call relationship, module usage relationship, cache read / write relationship, communication session relationship, container context relationship, and model inference relationship of each candidate edge's two endpoints. Of these, 2094 candidate edges with no shared relationships between nodes are deleted, leaving 1432 candidate edges. The backflip edge weight injection layer writes recovery time, recovery power consumption, latest response time, and historical backflip records to the retained edges, ultimately generating 221 communication reconnection backflip edges, 386 cache invalidation backflip edges, 74 container reload backflip edges, 96 model reload backflip edges, 185 task timeout backflip edges, and 470 chained wake-up backflip edges, forming an asynchronous wake-up backflip edge graph.

[0168] The system filters 312 objects to be sealed based on the asynchronous wake-up backflip debt pattern. These objects include 126 task threads, 108 cache pages, 18 communication sessions, 36 acquisition interfaces, 8 edge containers, and 16 model inference threads. For communication session C-02, the system reads 3 associated communication reconnection backflip edges, a remaining keep-alive time of 19 seconds, a total recovery time of 15 seconds, and 2 reconnections within the past 3 control cycles; therefore, it is added to the backflip lockout object set. For cache page P-147, the system reads its remaining cache validity period of 126 seconds, a recovery read time of 4 seconds, a remaining processing time of 31 seconds for the associated task, and no associated task timeout backflip edge; therefore, it is added to the controllable backflip candidate sealing object set. Ultimately, the system obtains 97 backflip lockout objects and 215 controllable backflip candidate sealing objects.

[0169] The system generates a subset of objects eligible for sealing and a subset of objects reachable for redemption from 215 controllable backflush candidate sealing objects. The system reads the task urgency flag, module idle status, cache occupancy status, communication reporting status, and container context storage status of each object. Objects that are non-urgent, have an idle module, cache queue occupancy not exceeding 30%, no data to be reported, and whose context is persistent are written into the subset of objects eligible for sealing; a total of 168 objects are written into this subset. The system then reads the latest redemption time, recovery action time, cache validity period, communication keep-alive deadline, remaining storage time of the container context, and remaining storage time of the model context to determine if the total recovery action time is less than the remaining time limits for each category, resulting in 149 objects reachable for redemption. After matching the object numbers, 137 objects exist in both subsets, and 31 objects exist only in the subset of objects eligible for sealing, forming a sealing-redemption cross-validation structure.

[0170] For 137 redeemable sealed objects, the system generates redeemable sealed capsules. These include 3 communication session capsules, which contain a communication session token, protocol sequence number, and keep-alive thread identifier; 79 cache page capsules, which contain cache page index, data generation time, and dirty page bitmap; 5 edge container capsules, which contain a container context pointer, container recovery entry point, and input cache pointer; 9 model inference capsules, which contain model parameter mapping status, inference intermediate state summary, and model recovery entry point; and 18 acquisition interface capsules, which contain acquisition trigger conditions, last acquisition time, and next acquisition time limit. Based on the redemption time limits and trigger conditions of these capsules, the system generates 42 redemption trigger nodes and constructs a minimal wake-up bridge topology skeleton containing 58 in-bridge nodes and 71 wake-up bridge edges.

[0171] During low-power execution, the system keeps the low-frequency timing node, communication keep-alive node, cache expiration monitoring node, abnormal interruption listening node, and capsule redemption scheduling node running within the low-frequency clock island, with the low-frequency clock island frequency set to 8% of the normal main frequency. 82 task threads, 9 communication reporting threads, 18 data acquisition polling threads, 5 edge containers, and 9 model inference threads not connected to the minimum wake-up bridge topology are switched to a sealed state. During operation, the system triggered 166 recoveries, including 43 joint redemptions. The average recovery time was 3.2 seconds, and the maximum recovery time was 6.9 seconds, all within the corresponding redemption time limit.

[0172] Under the same 17,035 standardized operation records, the traditional method employs a strategy of hibernating when CPU utilization is below 20% and the module is idle for more than 60 seconds. The traditional method has an average power consumption of 7.6W, a peak power consumption of 14.2W, 168 communication reconnections, 291 cache misses, 76 container reloads, 103 model reloads, 1945 module wake-ups, an average task latency of 1.86 seconds, and a task timeout rate of 3.4%. The method of this invention has an average power consumption of 5.0W, a peak power consumption of 9.1W, 39 communication reconnections, 62 cache misses, 14 container reloads, 21 model reloads, 826 module wake-ups, an average task latency of 1.13 seconds, and a task timeout rate of 0.7%. As can be seen from the above data, the method of the present invention reduces average power consumption by 34.2%, peak power consumption by 35.9%, communication reconnection times by 76.8%, cache invalidation times by 78.7%, container reload times by 81.6%, model reload times by 79.6%, and module wake-up times by 57.5%. It can achieve stable and low-power operation of intelligent fusion terminals while ensuring communication continuity, cache effectiveness, and model context integrity.

[0173] 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 low-power operation control method for intelligent converged terminals, characterized in that, include: Collect asynchronous state data of the intelligent fusion terminal during operation, preprocess the asynchronous state data to generate a standardized operation dataset, and construct an asynchronous event sequence; The asynchronous event sequence is input into the improved GraphRNN model, and the time-series node adaptive expansion, dependency edge gating aggregation and backflush weight dynamic injection are performed to generate an asynchronous wake-up backflush debt pattern. Based on the asynchronous wake-up backflip debt pattern, a set of backflip lockout objects and a set of controllable backflip candidate sealing objects are generated, and a sealing qualification subset and a redemption reachable subset are constructed to form a sealing-redemption cross-verification structure. Based on the cap-redemption cross-validation structure, redeemable captive capsules are generated for objects that exist simultaneously in the captive eligibility subset and the redemption reachable subset. A minimum wake-up bridge topology skeleton is constructed based on the redeemable captive capsules, and redemption trigger nodes are generated according to the redemption time limit and triggering conditions of each redeemable captive capsule. Based on the minimum wake-up bridge topology skeleton, perform low-frequency keep-alive and archive switching, keep the nodes connected to the minimum wake-up bridge topology skeleton running in the low-frequency clock island, and switch the objects not connected to the minimum wake-up bridge topology skeleton to the archive state; When the redemption trigger node is reached, the corresponding sealed object is restored according to the recovery action sequence in the redeemable sealed capsule. The low-power operation strategy for the next control cycle is adjusted based on the task delay, number of communication reconnections, number of cache invalidations, number of module wake-ups, and actual power consumption reduction after restoration.

2. The low-power operation control method for an intelligent fusion terminal according to claim 1, characterized in that, The asynchronous state data of the terminal includes task running data, module state data, cache state data, communication session data, edge container state data, and model inference state data.

3. The low-power operation control method for an intelligent fusion terminal according to claim 1, characterized in that, The preprocessing of asynchronous terminal state data to generate a standardized runtime dataset and construct an asynchronous event sequence includes: The asynchronous state data of the terminal is processed for time alignment, anomaly removal, field completion and state normalization. The task deadline, cache validity period, communication keep-alive deadline, data reporting window time, collection trigger time, edge container recovery time and model inference deadline are extracted and constructed into an asynchronous event sequence in chronological order.

4. The low-power operation control method for an intelligent fusion terminal according to claim 1, characterized in that, The generation of asynchronous wake-up backflush ripple patterns includes: An improved GraphRNN model is constructed, which includes an asynchronous phase boundary node expansion layer, a sealed dependency gated aggregation layer, and a backflush ripple weight injection layer. The asynchronous event sequence is input into the asynchronous phase boundary node augmentation layer in the order of trigger time. The event type, trigger time, latest response time, associated task number, associated module number and recovery time corresponding to each asynchronous event are read to generate task expiration node, cache invalidation node, communication keep-alive node, reporting window node, collection trigger node, container recovery node, model loading node and abnormal interruption node. After generating the current recoil twig node in the asynchronous phase boundary node augmentation layer, read the time sequence relationship, task call relationship, module occupation relationship, cache read and write relationship, communication session relationship, container context relationship and model inference relationship between the recoil twig node and the generated recoil twig nodes, and generate candidate recoil twig edges. The candidate backflip debt edges are input into the encapsulation dependency gating aggregation layer. The candidate backflip debt edges are gating and filtering according to task call relationship, module occupation relationship, cache read and write relationship, communication session relationship, container context relationship and model inference relationship. Candidate backflip debt edges with any relationship are retained, and candidate backflip debt edges without any relationship are deleted. The retained candidate backflip tween edges are input into the backflip tween weight injection layer, and recovery time, recovery power consumption, latest response time and historical backflip records are written to generate communication reconnection backflip edges, cache invalidation backflip edges, container reload backflip edges, model reload backflip edges, task timeout backflip edges and chained wake-up backflip edges. Based on the generation order of the anti-reverse bond pattern nodes, the connection relationship of the anti-reverse bond pattern edges, and the attribute fields written to the anti-reverse bond pattern edges, all anti-reverse bond pattern nodes and anti-reverse bond pattern edges are combined into an asynchronous wake-up anti-reverse bond pattern map.

5. The low-power operation control method for an intelligent fusion terminal according to claim 1, characterized in that, The construction of the sealed eligibility subset and the redemption reachable subset, forming a sealed-redemption cross-validation structure, includes: Read the counter-current debt nodes, counter-current debt edges and edge attribute fields in the asynchronous wake-up counter-current debt pattern graph, and identify the objects that have connection relationships with the communication module, cache page, edge container, model inference thread, acquisition interface and task thread as objects to be judged and sealed. For each object to be sealed, read the associated communication reconnection backflip edge, cache invalidation backflip edge, container reload backflip edge, model reload backflip edge, task timeout backflip edge, and chained wake-up backflip edge, extract the corresponding recovery time, recovery power consumption, latest response time, historical backflip count, and associated task deadline, and generate an object backflip judgment record. Based on the object backflip determination record, objects to be determined and sealed that have communication session reconstruction, cache validity period expiration, edge container reloading, model parameter reloading, task deadline failure or continuous module wake-up are written into the backflip lockout object set. Objects to be determined and sealed that have not been written into the backflip lockout object set and are in an idle and sealable state are written into the controllable backflip candidate sealed object set. For objects in the controllable backflush candidate sealing object set, read the current running status, task emergency flag, module idle status, cache occupancy status, communication reporting status and container context saving status, generate a sealing qualification subset, and read the latest redemption time, recovery action time, cache validity period, communication keep-alive deadline, container recovery entry and model context recovery status to generate a redemption reachable subset; Based on the object number, object type, associated module number, latest redemption time, recovery action time, cache validity period, and communication keep-alive deadline, an object-by-object matching is performed on the sealed eligibility subset and the redemption reachable subset. Objects that exist in both the sealed eligibility subset and the redemption reachable subset, objects that exist only in the sealed eligibility subset, and their corresponding matching fields are combined to form a sealed-redemption cross-validation structure.

6. The low-power operation control method for an intelligent fusion terminal according to claim 1, characterized in that, The minimum wake-up bridge topology framework constructed based on redeemable sealed capsules generates redemption trigger nodes according to the redemption time limit and triggering conditions of each redeemable sealed capsule, including: Read objects that exist in both the sealed eligibility subset and the redemption reachable subset in the sealed-redemption cross-validation structure, and write the object number, object type, associated module number, sealed start time, redemption time limit, recovery action time, cache validity status, communication keep-alive status, and container context status into the capsule basic record; Based on the object type in the capsule's basic record, write the communication session token, protocol sequence number, and keep-alive thread identifier to the communication session; write the cache page index, data generation time, and dirty page bitmap to the cache page; write the container context pointer, container recovery entry, and input cache pointer to the edge container; write the model parameter mapping status, inference intermediate state summary, and model recovery entry to the model inference thread; and write the acquisition trigger condition, last acquisition time, and next acquisition time limit to the acquisition interface to generate a redeemable sealed capsule. Read the redemption time limit, triggering condition, keep-alive thread identifier, cache page index, container recovery entry, model recovery entry and recovery action sequence from each redeemable sealed capsule, and determine the low-frequency timing node, communication keep-alive node, cache validity monitoring node, abnormal interruption listening node, capsule redemption scheduling node and recovery execution node used to maintain the recoverable state of the sealed object as wake-up bridge nodes; The wake-up bridge edges are established according to the timing triggering order, communication keep-alive relationship, cache expiration monitoring relationship, abnormal interruption triggering relationship, capsule redemption scheduling relationship and recovery action execution order among the wake-up bridge nodes. The low-frequency timing nodes, communication keep-alive nodes, cache expiration monitoring nodes, abnormal interruption listening nodes, capsule redemption scheduling nodes, recovery execution nodes and corresponding wake-up bridge edges are combined into the minimum wake-up bridge topology skeleton. Based on the redemption time limit and triggering conditions of each redeemable sealed capsule, a redemption trigger node is generated in the minimum wake-up bridge topology skeleton, and the redemption trigger node is connected to the corresponding capsule redemption scheduling node, recovery execution node and sealed object respectively.

7. The low-power operation control method for an intelligent fusion terminal according to claim 1, characterized in that, The low-frequency keep-alive and save-by switching based on the minimum wake-up bridge topology includes: Read the wake-up bridge nodes and wake-up bridge edges in the minimum wake-up bridge topology skeleton, mark the running objects that have a connection relationship with the redemption trigger node, capsule redemption scheduling node, communication keep-alive node, cache validity monitoring node and abnormal interruption listening node as bridge keep-alive objects, and mark the running objects that have not established a connection relationship with the minimum wake-up bridge topology skeleton as bridge sealed objects. For objects kept alive within the bridge, retain the periodic timing task of the low-frequency timing node, the session persistence task of the communication keep-alive node, the cache expiration check task of the cache expiration monitoring node, the trigger listening task of the abnormal interruption listening node, and the redemption queuing task of the capsule redemption scheduling node, and bind the tasks to the low-frequency clock island. For objects sealed outside the bridge, the scheduling entry of non-urgent task threads is suspended, the active sending task of communication reporting thread is stopped, the periodic collection task of collection polling thread is turned off, the unnecessary execution context of edge containers is frozen, and the non-urgent inference task of model inference thread is suspended. Before switching the external sealed object to the sealed state, write the cache page index, communication session token, protocol sequence number, container context pointer, model parameter mapping state and recovery action sequence into the corresponding redeemable sealed capsule, and write the sealed state, sealed start time and corresponding redemption trigger node into the low power operation strategy table. During the operation of the low-frequency clock island, low-frequency keep-alive inspections are performed according to the timing trigger edge, communication keep-alive edge, cache expiration trigger edge, abnormal interruption trigger edge and capsule redemption edge in the minimum wake-up bridge topology skeleton. For bridge-external sealed objects that have reached the redemption trigger node, recovery instructions are output, and bridge-external sealed objects that have not reached the redemption trigger node are kept in a sealed state.

8. The low-power operation control method for an intelligent fusion terminal according to claim 1, characterized in that, The low-power operation strategy for adjusting the next control cycle based on the recovered task latency, number of communication reconnections, number of cache invalidations, number of module wake-ups, and actual power consumption decrease includes: When the low-frequency timing node, communication keep-alive node, cache validity monitoring node, or abnormal interruption listening node detects the arrival of the corresponding redemption trigger node, it reads the recovery action sequence, sealed object number, cache page index, communication session token, protocol sequence number, container context pointer, and model parameter mapping status from the corresponding redeemable sealed capsule, and generates a sealed object recovery queue. According to the order of recovery actions in the sealed object recovery queue, cache page recovery, communication session recovery, protocol sequence recovery, edge container context recovery, model parameter mapping recovery, acquisition interface recovery, and task thread recovery are executed sequentially. The recovered objects are then reconnected to the running node corresponding to the minimum wake-up bridge topology skeleton. After the archived object is restored, read the restored task execution time, number of communication reconnections, number of cache invalidations, number of module wake-ups, number of edge container reloads, number of model reloads, and actual power consumption reduction results, and generate a restoration result record; The recovery results are compared with the current low-power operation strategy in terms of sealing duration, redemption time limit, wake-up trigger order, communication keep-alive cycle, cache retention cycle and low-frequency operation cycle. The sealing limit level is increased for sealed objects that have experienced task timeout, communication reconnection, cache invalidation, container reload, model reload or frequent wake-up. The sealing duration of the next control cycle is extended for sealed objects that have no abnormal recovery results and whose power consumption has decreased to the preset requirement. The revised low-power operation strategy is generated.

9. A low-power operation control system for an intelligent converged terminal, comprising executing the low-power operation control method for an intelligent converged terminal as described in any one of claims 1 to 8, characterized in that, include: The asynchronous event building module is used to collect asynchronous state data from the terminal, generate a standardized runtime dataset, and construct an asynchronous event sequence. The anti-rebound ripple generation module is used to input asynchronous event sequences into the improved GraphRNN model and generate asynchronous wake-up anti-rebound ripple maps; The redemption verification module is used to generate a set of redemption lock-up objects and a set of controllable redemption candidate sealed objects based on the asynchronous wake-up redemption debt pattern, construct a subset of sealed eligibility and a subset of redemption reachability, and form a redemption cross-verification structure. The wake-up bridge construction module is used to generate redeemable sealed capsules based on the seal-redemption cross-validation structure, construct the minimum wake-up bridge topology skeleton, and generate redemption trigger nodes; The low-frequency sealing and control module is used to perform low-frequency keep-alive and sealing switching based on the minimum wake-up bridge topology skeleton; The redemption correction module is used to restore the sealed objects based on the redemption trigger node and correct the low-power operation strategy for the next control cycle.

Images