A cloud phone non-inductive adaptive synchronization method and device based on prediction

Abstract

The application discloses a cloud mobile phone non-inductive adaptive synchronization method and device based on prediction and belongs to the technical field of mobile cloud computing and edge cooperation. Through the collection of multi-source data fusion, user operation is predicted, semantic prediction instruction flow is generated, and a key degree level is marked. A mapping relationship between operation intention and edge resources is established, and calculation preheating, memory reservation and network preconfiguration are executed before real operation. High-probability candidate branches are screened in the cloud and are pre-rendered in parallel, picture difference data is generated and is cached to an edge node. Value evaluation and dynamic resource scheduling are performed on each branch. When real operation occurs, if matching is achieved, cached difference is directly applied, otherwise, fast correction and feedback of optimized prediction parameters are performed. The application solves the operation asynchronous problem of cloud mobile phones caused by network and calculation delay, and realizes low-delay, high-robustness non-inductive adaptive synchronization.

Description

Technical Field

[0001] This invention relates to the field of mobile cloud computing and edge collaboration technology, and in particular to a prediction-based cloud phone seamless adaptive synchronization method and device. Background Technology

[0002] With the rapid development of mobile internet and cloud computing technologies, cloud phones run their operating systems and applications on cloud servers, transmitting rendered images to user terminals via real-time audio and video streams, enabling even low-spec devices to achieve a high-performance experience. The core challenge lies in achieving low-latency, high-synchronization interaction between user operations, cloud computing, and network transmission. Traditional cloud phones use a request-response model, where user actions are transmitted over the network to the cloud for processing and results are returned. While this model works under ideal network conditions, in real-world mobile networks, limitations in latency, jitter, and bandwidth fluctuations make it difficult to guarantee real-time performance and smoothness. This is especially true in latency-sensitive scenarios such as gaming and real-time collaboration, where issues like stuttering and screen tearing become prominent.

[0003] To mitigate network latency, existing technologies have proposed predictive pre-computation schemes, which pre-cache rendering results based on user behavior predictions. Additionally, some schemes involve state synchronization across multiple devices to improve system responsiveness. However, most existing schemes are reactive mechanisms, meaning system actions are triggered by events that have already occurred, failing to establish a complete closed loop from predicting user terminal intent, to multi-branch parallel pre-computation in the cloud, and then to edge caching and dynamic transmission. Their state synchronization is typically a complete replication of the already occurred state, rather than proactive resource allocation and content pre-generation based on predictions and potential future states. Existing schemes do not utilize prediction time windows to mask network latency, nor do they intelligently adjust computation and transmission strategies based on prediction probabilities and network conditions. Furthermore, local optimization measures on the client side are often not linked to the global resource state in the cloud, resulting in limited effectiveness or even conflicts under heavy loads or network degradation. Therefore, a predictive-driven, seamless adaptive synchronization method is urgently needed. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] To address this, this invention proposes a prediction-based, seamless adaptive synchronization method for cloud phones. This method involves collecting multi-source data at the user device end and predicting user actions to generate a semantic prediction command stream with criticality levels. Before the actual operation occurs, resource pre-allocation is performed at the target edge node based on the command stream. High-probability candidate branches are selected and pre-rendered in parallel on the cloud, generating screen difference data and caching it at the edge node. The pre-computation tasks for each branch are evaluated for value and dynamically resource-scheduled. When the actual operation occurs, if a match is found, the cached differences are directly applied; otherwise, the error is quickly corrected and the prediction parameters are optimized. This invention achieves prediction-driven closed-loop synchronization, significantly reducing interaction latency and improving the real-time performance and robustness of cloud phones in dynamic network environments.

[0006] Another objective of this invention is to propose a prediction-based cloud phone seamless adaptive synchronization device.

[0007] To achieve the above objectives, this invention proposes a prediction-based, seamless adaptive synchronization method for cloud phones, comprising: On the user device side, user interaction behavior data, device sensor data and current application interface data are collected in real time. By fusing multi-source data, multiple operations that the user may perform in the future short time window and their corresponding probabilities are predicted. The prediction results are compressed into a semantic prediction instruction stream and the operation criticality level is marked. The prediction instruction stream is synchronously sent to the cloud rendering service and global resource coordination function. Establish a mapping relationship between user operation intentions and computing, storage and network resources in the cloud and edge nodes. Before the actual operation is confirmed, based on the operation type and criticality level in the predicted instruction stream, initiate resource pre-allocation actions on the target edge node, such as computing instance preheating, memory reservation and network service quality policy pre-configuration. Continuously monitor global resource load and network status, and feed back system status information to the user device in real time to dynamically adjust prediction parameters. In the cloud rendering service, the prediction instruction stream is received, multiple high-probability prediction instructions are selected as candidate branches to be processed, independent rendering task instances are created in parallel in the edge computing environment corresponding to each candidate branch, future screen frames of each candidate branch are generated, each pre-rendered result is compared with the current baseline screen, screen difference data containing only the changed parts is generated and cached in the edge storage node closest to the user. The value and cost of the pre-computation tasks for each candidate branch are evaluated, the value-efficiency ratio of each pre-computation task is calculated, and all parallel pre-computation tasks are dynamically prioritized and differentiated resource allocation is performed accordingly. The coding and transmission strategies are dynamically adjusted based on the operational criticality level and the real-time network link quality. When a user actually performs an action, the actual action is compared with the pre-calculated prediction instruction. If the match is successful, the corresponding screen difference data is directly extracted from the edge cache and applied. If the match fails, a fast error correction process is triggered, which only calculates the difference between the actual action and the closest prediction to restore state consistency. The prediction deviation data is then fed back to the prediction function on the user's device to optimize the prediction parameters.

[0008] The cloud phone seamless adaptive synchronization method based on prediction according to an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, multi-source data fusion prediction includes: Capture and calculate the kinematic characteristics of instantaneous velocity, acceleration, and trajectory curvature from raw touch events on the touchscreen; read sensor data of device acceleration and rotation speed to sense changes in device attitude; Obtain complete interface information of the currently running application in the foreground, parse the screen coordinates, shape, text and semantic tags of all interactive interface elements, and construct the spatial layout diagram and semantic relationship diagram of the current interface; By integrating temporal correlation analysis, spatial correlation analysis, and attention guidance, a joint prediction distribution of multiple user operations and their corresponding probabilities within a future short-term window is output.

[0009] In one embodiment of the present invention, the prediction result is compressed into a semantic prediction instruction stream, including: The underlying data of coordinate sequences and action types are abstracted into high-level instructions that include action type, start coordinates, end coordinates, and estimated duration; The criticality level of the operation is dynamically marked based on the application context that generated the prediction instruction and the criticality of the prediction operation in the application. The prediction probability, semantic instructions, and operation criticality level are encapsulated to form a prediction instruction stream.

[0010] In one embodiment of the present invention, the resource pre-allocation action includes: On the edge node closest to the corresponding user location and where the load allows, create or preheat computing resource instances that meet the corresponding performance requirements according to the operation criticality level, reserve memory capacity, and pre-configure network service quality policies for the transmission path from the corresponding computing instance to the user device.

[0011] In one embodiment of the present invention, the comparison of the pre-rendered result with the current reference image specifically involves: performing pixel-by-pixel or block-by-block difference calculations on each pre-rendered result and the current reference image, and extracting image difference data containing only the changed parts. The image difference data adopts an incremental encoding method to reduce storage and transmission overhead.

[0012] In one embodiment of the present invention, the differentiated resource allocation and encoded transmission strategy is as follows: High-value-efficiency pre-computation tasks are given the highest scheduling priority, and a layered, progressive preloading strategy is adopted to prioritize loading low-resolution resources and core logic states. For the pre-calculated results corresponding to P0 level critical operations, when network bandwidth is limited, priority is given to ensuring smoothness of the screen, reducing the resolution and increasing the key frame frequency, and enabling forward error correction on the transmission side to enhance the protection of core data. For P2 level normal operations, the frame rate can be appropriately reduced or a higher compression ratio encoding configuration can be used.

[0013] In one embodiment of the present invention, the value efficiency ratio is calculated as follows: A value scoring model is constructed to calculate a dynamic value score for each predicted intent branch. The value score integrates the prediction probability, the operational criticality level, real-time feedback from the client side, and global resource load and network status information. Simultaneously build a computational overhead model to estimate the rendering processing time and memory usage of each intent branch; calculate the ratio of the value score of each branch to the estimated computational overhead in real time as the value efficiency ratio.

[0014] In one embodiment of the present invention, the fast error correction process includes: The client generates correction information containing only the core differences between the actual operation and the closest prediction and sends it to the edge nodes; The cloud rendering service at the edge node locates the common computing starting point on which the current application state and the prediction branch are based. It only calculates the difference between the corresponding common starting point and the actual operation target state, and quickly encodes the corrected image and sends it to the client.

[0015] In one embodiment of the present invention, it further includes: During the window period while waiting for the cloud confirmation image to arrive, the client runs a local lightweight image prediction based on the recently received multiple frames and real-time touch data, generates the next frame of the local predicted image and displays it, and updates it by smooth replacement or fusion after receiving the cloud confirmation image. When a sudden increase in network round-trip latency is detected, the client simulates an instantaneous animation feedback that conforms to the law of inertia based on the physical properties of the operation when it leaves the screen. The physical properties are obtained by fitting the last few sampling points of the touch event stream.

[0016] To achieve the above objectives, another aspect of the present invention proposes a prediction-based cloud phone seamless adaptive synchronization device, comprising: The perception and prediction module is used to collect user interaction behavior data, device sensor data and current application interface data in real time on the user device. It predicts multiple operations that the user may perform in a short time window and their corresponding probabilities through multi-source data fusion. The prediction results are compressed into a semantic prediction instruction stream and the operation criticality level is marked. The prediction instruction stream is synchronously sent to the cloud rendering service and global resource coordination function. The resource preheating module is used to establish a mapping relationship between user operation intentions and computing, storage and network resources in the cloud and edge nodes. Before the actual operation is confirmed, it initiates resource pre-allocation actions such as computing instance preheating, memory reservation and network service quality policy pre-configuration on the target edge node according to the operation type and criticality level in the predicted instruction stream. It also continuously monitors the global resource load and network status and feeds the system status information back to the user device in real time to dynamically adjust the prediction parameters. The branch pre-rendering module is used to receive the prediction instruction stream in the cloud rendering service, filter out multiple high-probability prediction instructions as candidate branches to be processed, create independent rendering task instances in parallel in the edge computing environment corresponding to each candidate branch, generate future screen frames for each candidate branch, compare the differences between each pre-rendering result and the current baseline screen, generate screen difference data containing only the changed parts and cache it to the edge storage node closest to the user. The value scheduling module is used to evaluate the value and cost of the pre-computation tasks of each candidate branch, calculate the value-efficiency ratio of each pre-computation task, and dynamically prioritize all parallel pre-computation tasks accordingly and perform differentiated resource allocation. It also dynamically adjusts the encoding and transmission strategies based on the criticality level of the operation and the real-time network link quality. The error correction feedback module is used to match and compare the actual user operation with the pre-calculated prediction instruction when the actual operation occurs. If the match is successful, the corresponding screen difference data is directly extracted from the edge cache and applied. If the match fails, a fast error correction process is triggered, which only calculates the difference between the actual operation and the closest prediction to restore state consistency, and feeds the prediction deviation data back to the prediction function on the user device to optimize the prediction parameters.

[0017] This invention discloses a prediction-based, seamless adaptive synchronization method and apparatus for cloud phones. It predicts user operations and generates semantic command streams through multi-source data fusion. Before the actual operation occurs, it performs edge resource preheating and multi-branch parallel pre-rendering, generating screen difference data and caching it to edge nodes. Combined with dynamic scheduling based on value-efficiency ratio and a rapid error correction feedback mechanism, it effectively solves the problem of interaction asynchrony caused by network latency and computational overhead in existing reactive architectures. It achieves a complete closed loop from user intent perception, predicted command generation, edge resource pre-allocation, multi-branch pre-computation to real-time synchronization and adaptive error correction, significantly reducing end-to-end interaction latency, improving screen smoothness and operational response robustness in dynamic network environments, and enhancing the adaptability and engineering practical value of cloud phones under complex conditions such as weak networks and high loads.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a cloud phone seamless adaptive synchronization method based on prediction according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the overall process according to an embodiment of the present invention; Figure 3 This is a flowchart of step S101 of the present invention according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a prediction-based cloud phone seamless adaptive synchronization device according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes a prediction-based, seamless adaptive synchronization method and apparatus for cloud phones according to embodiments of the present invention.

[0023] The core idea of ​​this invention is to address the high interaction latency issue of traditional cloud phone request-response architectures under dynamic networks. It solves the problem by constructing a seamless adaptive synchronization method covering the entire process of "perception prediction—resource preheating—branch pre-rendering—value scheduling—error correction feedback," thus overcoming the limitation of existing reactive models in not being able to mask transmission and computation latency using prediction time windows. First, multi-source fusion predicts user operations and generates semantic command streams. Before the actual operation, edge computing preheating, memory reservation, and network pre-configuration are performed based on the command stream. Then, high-probability candidate branches are pre-rendered in parallel, generating and caching screen difference data. Resources are dynamically scheduled based on the value-efficiency ratio, and encoding and transmission strategies are adaptively adjusted. When the actual operation occurs, if a match is successful, the cached differences are directly applied; otherwise, rapid correction and feedback optimization of prediction parameters are implemented. This invention transforms the traditional passive response synchronization model into a prediction-driven, multi-branch preparation, and adaptive error correction intelligent synchronization system, significantly reducing end-to-end interaction latency and improving smoothness and robustness in weak network and high-load environments.

[0024] The following description, with reference to the accompanying drawings, describes a prediction-based, seamless adaptive synchronization method and apparatus for cloud phones according to embodiments of the present invention.

[0025] like Figure 1 As shown, a prediction-based cloud phone seamless adaptive synchronization method of the present invention includes the following steps: S1. On the user device side, real-time collection of user interaction behavior data, device sensor data and current application interface data. Through multi-source data fusion, prediction of multiple operations that the user may perform in the future short window and their corresponding probabilities of occurrence. The prediction results are compressed into a semantic prediction instruction stream and the operation criticality level is marked. The prediction instruction stream is synchronously sent to the cloud rendering service and global resource coordination function. S2, establish a mapping relationship between user operation intentions and computing, storage and network resources in the cloud and edge nodes. Before the actual operation is confirmed, according to the operation type and criticality level in the predicted instruction stream, initiate resource pre-allocation actions such as computing instance preheating, memory reservation and network service quality policy pre-configuration on the target edge node, and continuously monitor the global resource load and network status, and feed back the system status information to the user device in real time to dynamically adjust the prediction parameters. S3, in the cloud rendering service, receive the prediction instruction stream, filter out multiple high-probability prediction instructions as candidate branches to be processed, create independent rendering task instances in parallel in the edge computing environment corresponding to each candidate branch, generate future screen frames for each candidate branch, compare the differences between each pre-rendered result and the current baseline screen, generate screen difference data containing only the changed parts and cache it to the edge storage node closest to the user. S4 evaluates the value and cost of the pre-computation tasks for each candidate branch, calculates the value-efficiency ratio of each pre-computation task, and dynamically prioritizes all parallel pre-computation tasks accordingly, and performs differentiated resource allocation. It also dynamically adjusts the coding and transmission strategies based on the operational criticality level and real-time network link quality. S5: When a user performs a real operation, the real operation is matched and compared with the pre-calculated prediction instruction. If the match is successful, the corresponding screen difference data is directly extracted from the edge cache and applied. If the match fails, a fast error correction process is triggered, which only calculates the difference between the real operation and the closest prediction to restore state consistency. The prediction deviation data is then fed back to the prediction function on the user device to optimize the prediction parameters.

[0026] Furthermore, such as Figure 2 As shown, the method of this embodiment of the invention specifically includes: S101. Establish a mapping relationship between intent and cloud-edge resources. Based on this mapping relationship, pre-allocate computing, storage and network resources before the user's actual operation occurs, and dynamically adjust the prediction parameters according to the global resource status. When the system is running, this step S101 functions as a global resource coordination function.

[0027] S102. On the user device, analyze real-time touch, sensor and interface data to predict the user's operation intention and its probability of occurrence in the near future. Compress the prediction results and mark them as semantic prediction command streams. Then, send the command streams synchronously to steps S101 and S103.

[0028] S103. In the cloud, the prediction instruction stream from step S2 is received, and multiple high-probability future states are pre-calculated in parallel based on this, generating screen difference data and caching it to edge nodes; at the same time, the encoding and transmission strategies are dynamically adjusted according to the real-time network conditions and the criticality of the instructions.

[0029] S104. Under the resource constraints set in step S101, the value and cost of the pre-computation tasks generated in step S103 are evaluated, the value-efficiency ratio of each task is calculated, and resources are scheduled and rendering and transmission tasks are executed in stages accordingly.

[0030] S105. On the user equipment side, local prediction is used to fill the network transmission gap, network latency is compensated by simulating physical inertia, and operational feedback is provided to the user when there is network latency or fluctuation, in conjunction with the control instructions from step S101.

[0031] S106. When a prediction deviation occurs or a network anomaly occurs, the application state is corrected and consistency is restored through difference correction calculation, incremental state snapshot and synchronization mechanism. At the same time, the prediction deviation data is fed back to step S102 to optimize its prediction parameters.

[0032] Among them, such as Figure 3 As shown, step S101 specifically includes the following steps: S1011. Perform unified abstraction and quantitative modeling of computing, storage, and network resources distributed across the central cloud and edge nodes to form measurable and schedulable basic resource units.

[0033] S1012. Construct a dynamic intent-resource mapping relationship. This mapping relationship combines the operation criticality level (P0, P1, P2) defined in step S102 with the application context to clarify the precise resource requirements of different user intents. For example, the requirement for firing operation in a P0-level shooting game can be defined as: needing a graphics processing unit with specific performance and a high-priority, low-latency, and bandwidth-guaranteed data path.

[0034] S1013. After the prediction function in step S102 is activated, the aforementioned global resource coordination function is activated and receives a prediction instruction stream with prediction probability and criticality labels from step S102. Based on the mapping relationship established in step S1012, a resource pre-allocation action is initiated within hundreds of milliseconds before the user's actual operation (such as lifting a finger) is confirmed. The pre-allocation action includes: creating or preheating computing resource instances, reserving memory, and pre-configuring network service quality policies for the transmission path from the computing instance to the user on the edge node closest to the predicted user location and with allowable load. For example, when step S102 reports that the probability of clicking the settlement button of an e-commerce application is 70% and is marked as P1 level, the global resource coordination function prepares a lightweight computing environment in advance on a suitable edge node and pre-configures a low-latency policy for the network exit of the environment.

[0035] S1014. The global resource coordination function continuously monitors the real-time load and status of all edge nodes and network links to form system status information. The system status information includes, for example, that the utilization rate of a node's graphics processing unit has reached 85%, or that the network path to a user is expected to be congested in 300 milliseconds. The system status information is fed back to the prediction function in step S102 in real time to adjust its decision parameters, such as adaptively changing the prediction confidence threshold or prediction time window under different resource conditions.

[0036] Specifically, step S102 includes the following steps: S1021. First, synchronously collect multi-source data streams, specifically including: Interactive Flow: Captures raw touch events on the device's touchscreen and calculates kinematic characteristics such as instantaneous velocity, acceleration, and trajectory curvature.

[0037] Sensor stream: Reads sensor data on device acceleration and rotation speed to sense changes in device posture and gripping status.

[0038] Context flow: Through system interfaces, obtain complete interface information of the currently running application in the foreground, parse all interactive interface elements, obtain their screen coordinates, shape, text and semantic tags, and thus construct the spatial layout diagram and semantic relationship diagram of the current interface.

[0039] S1022 Next, the interaction stream, sensor stream and context stream collected in step S21 are fused and correlated through a lightweight prediction process that runs locally on the user device.

[0040] Among them, temporal correlation analysis refers to performing time series analysis on interaction flow and sensor flow to capture operation rhythm and gesture trends.

[0041] Spatial correlation analysis refers to treating the current interface structure as a graph model composed of interface elements and their relationships, analyzing the spatial distance and layout hierarchy between elements, and inferring the user's potential operation goals.

[0042] After an attention-guided fusion phase, the process outputs a joint predicted distribution of multiple actions that the user may perform within a short future time window and their probabilities; for example, clicking the submit button: 65%, swiping right about 300 pixels: 25%, no action: 10%.

[0043] To filter out valid user action prediction results, the prediction validity judgment formula is defined as follows:

[0044] in, To predict the effective margin; Predict the probability for a single operation; This is a preset probability threshold.

[0045] S1023. Semantically compress the coordinate sequence and action type underlying data output by the prediction process in step S1022 and abstract them into higher-order instructions; for example, the abstract representation is: action type: sliding, starting coordinates: (x1, y1), ending coordinates: (x2, y2), estimated duration: 120ms.

[0046] S1024. Based on the application context that generated the prediction instruction and the criticality of the prediction operation itself in the application, dynamically assign an operation criticality level (P0 / P1 / P2) to each semantic instruction; for example, the firing operation in a game is marked as P0, and the page-turning operation in a reading application may be marked as P1.

[0047] S1025. The probability predicted in step S1022, the semantic instructions generated in step S1023, and the operation criticality level marked in step S1024 are encapsulated to form a prediction instruction stream. The prediction instruction stream is distributed in real time in two paths: one path is reported to the global resource coordination function in step S101; the other path is sent directly to the cloud rendering service in step S103.

[0048] Specifically, step S103 includes the following steps: S1031. Receive the prediction instruction stream from step S102, which carries the prediction probability and the operational criticality level, through the cloud rendering service. Select multiple high-probability prediction instructions as candidate branches to be processed. In the edge computing environment pre-configured or started in step S101, ensure that the latest state snapshot of the current application has been loaded.

[0049] S1032. Based on the latest state snapshot of the edge computing environment and application prepared in step S1031, create independent rendering task instances in parallel for each high-probability prediction instruction; each task instance simulates the user operation described by the semantic instruction to generate the corresponding future screen frame, i.e., the pre-rendering result.

[0050] S1033. Compare each pre-rendered result with the current baseline image to generate image difference data containing only the changed parts; store these image difference data in the edge storage node closest to the target user, which is coordinated by step S101; at the same time, according to the next application state pointed to by the prediction instruction, actively preload the resource files that may be used into the memory of the computing environment.

[0051] S1034. Through transmission control logic in collaboration with the rendering service, continuously monitor the quality of the end-to-end transmission link from the edge node to the target user device, and make short-term predictions on the recent quality; the monitored indicators include real-time bandwidth, round-trip latency, jitter, and packet loss rate.

[0052] S1035. Based on the network monitoring information from step S1034 and the operational criticality level of the instructions in step S102, dynamically adjust the encoding and transmission strategies, specifically: For the pre-calculated results corresponding to P0 level critical operations, when it is predicted that the network bandwidth will tighten, priority is given to ensuring the smoothness of the picture, quickly reducing the resolution, adjusting the quantization parameters to compress the single frame size, and increasing the frequency of key frames. On the transmission side, forward error correction or imbalance protection strategies are enabled to provide stronger protection for core data such as key frames and motion vectors.

[0053] For P2 level normal operations, the frame rate is appropriately reduced or a more efficient encoding configuration is adopted.

[0054] Specifically, step S104 includes the following steps: S1041. Construct a value scoring model to calculate a dynamic value score for each intent branch predicted in step S102. The dynamic value score integrates multiple input factors, including: the probability of the intent occurring provided in step S102, the operation criticality level (P0 / P1 / P2) marked in step S102, real-time feedback from the client side in step S105, and global resource load and network status information from step S101.

[0055] Simultaneously build a computational overhead model: estimate the computational overhead required to render each intent branch, including processing time, memory usage, data bandwidth consumption, and result generation latency.

[0056] S1042. Based on the value score and estimated computational cost of each intent branch, calculate the value efficiency ratio of each branch in real time, and prioritize all parallel pre-computation tasks in real time and dynamically.

[0057] S1043. Based on the priority queue generated in step S1042, execute differentiated resource allocation and task execution strategies, specifically including: For high-value, low-overhead branches, the highest scheduling priority is given; for high-value, high-overhead branches, a layered, progressive preloading strategy is adopted, prioritizing the loading of low-resolution resources, core logic states, and basic model information, or rendering only a rough preview; for low-value branches, the scheduling priority is reduced, and a strategy of not performing pre-computation or only preparing the necessary logic states is adopted.

[0058] S1044. Implement prediction-based local rendering: When the prediction operation is continuous content browsing, prioritize rendering the static background and easily compressible parts of the text on the next screen to achieve on-demand local rendering.

[0059] Implement computation and transmission interleaving: After generating a partial frame block during the rendering process, encoding and transmission begin immediately without waiting for the entire frame to be rendered.

[0060] Specifically, step S105 includes the following steps: S1051. During the brief window period before the cloud confirmation screen arrives, the client runs a local lightweight screen prediction function. Based on the recently received frames and real-time touch data, it quickly generates and displays the next frame of the local prediction screen. Once the cloud confirmation screen is received, it is updated by smooth replacement or fusion.

[0061] S1052. When a sudden increase in network round-trip latency is detected, the client simulates an instantaneous animation feedback that conforms to the law of inertia based on the physical properties of the operation when it leaves the screen. For example, when scrolling through a list, the list continues to move and decelerates according to the final speed of the scroll.

[0062] S1053. The strength parameters of the above local strategy are dynamically adjusted by receiving the adjustment instructions from step S101. Step S101 dynamically generates and issues control instructions based on a comprehensive judgment of the load of edge nodes of the entire network, the prediction of transmission link quality and the queuing situation of cloud computing.

[0063] Specifically, step S106 includes the following steps: S1061. When a user performs a real operation, the client immediately performs a quick match and comparison with the high-probability prediction instruction previously reported in step S102 and pre-calculated in step S103. If the overall score of the matching degree between a prediction instruction and the real operation exceeds the preset hit threshold, it is determined to be a prediction hit. The client directly extracts the corresponding screen difference data from the edge cache and applies it.

[0064] To assess the degree of agreement between actual operations and predicted instructions, the comprehensive matching score formula is defined as follows:

[0065] in, A comprehensive score for operational matching. , , The scores are for operation type, coordinate distance, and direction / angle, respectively. , , These are the corresponding sub-item weight coefficients.

[0066] S1062. If the matching degree between the actual operation and all predicted instructions is lower than the hit threshold, then the fast error correction process is triggered: The client generates a very simple correction message, containing only the core difference between the actual operation and the closest prediction, and immediately sends it to the edge node.

[0067] After receiving the correction information, the cloud rendering service at the edge node quickly locates the common computing starting point on which the current application state and the prediction branch are based. It only calculates the difference between the starting point and the actual operation target state, and the calculated correction screen is quickly encoded and sent out.

[0068] The client applies this correction screen to synchronize the state with the user's intent.

[0069] S1063. The system collects comparison data between prediction and actual operation, and after anonymization, forms a prediction deviation sample, which is sent back to the prediction function in step S102 as feedback information for online parameter adjustment and long-term optimization.

[0070] S1064. In a cloud computing environment, an incremental application state persistence mechanism is established. The application state persistence mechanism periodically or when a key state change is detected, the state snapshot is saved. Each time a snapshot is saved, incremental storage is performed based on the previous snapshot, and a unique state version number is generated. All state snapshots are mapped to the timestamp or frame number of the video stream output at the corresponding time and are persistently stored.

[0071] In step S1013, the determination of the predicted user location is based on the GPS, base station, or Wi-Fi positioning information reported by the user equipment in real time, combined with historical movement trajectories for short-term prediction; the load allowance judgment criterion is that the current CPU / GPU utilization, memory margin, and network egress bandwidth of the edge node are all lower than a preset dynamic threshold; in step S1014, the feedback of the system status information is based on a preset triggering strategy, which actively pushes the information when the resource load or network quality changes exceed a set range, or retrieves it on demand by the prediction function in step S102.

[0072] In step S1022, the attention-guided fusion stage calculates attention weights based on the temporal urgency of interaction flow features, the degree of abrupt change in sensor flow features, and the spatial salience and historical interaction frequency of interface elements in the context flow. The output of the joint prediction distribution retains only prediction results with a probability exceeding a preset threshold (e.g., 5%) to control the scale of the instruction flow. In step S1024, the dynamic marking logic incorporates a predefined rule library for different application types, such as marking operations that cause large-scale screen redraws or trigger time-consuming background logic as having higher criticality. In step S1031, the criterion for selecting high-probability prediction instructions is instructions with a probability greater than a preset threshold that are not semantically completely covered by other higher-probability instructions. The accuracy of the state snapshot is ensured by strictly aligning the version number with the frame last confirmed by the client.

[0073] In step S1041, the value scoring model is a weighted summation function, wherein the weight of the real-time feedback factor from the client side in step S105 is dynamically adjusted according to the reported latency and confidence level; the computational overhead estimation is based on historical performance analysis data of different types of rendering tasks (such as 2DUI and 3D scenes).

[0074] In step S1043, the layered progressive preloading includes a logical state layer, a basic geometry layer, a standard texture layer, and a high-definition texture layer, with the number of layers to be loaded determined based on task priority and resource availability.

[0075] In step S1051, the smooth replacement or fusion adopts alpha mixing based on time or content similarity, and the mixing ratio is dynamically calculated according to the confidence of the local predicted image and the network latency estimate.

[0076] In step S1052, the physical properties are obtained by fitting the last few sampling points of the touch event stream. The duration of the simulated animation is positively correlated with the estimated network latency, but there is a maximum value limit.

[0077] In step S1053, the control command is generated based on the global policy table maintained in step S101, which defines the recommended local policy parameters corresponding to different load levels and network quality ranges.

[0078] In step S1061, the comprehensive matching score is calculated by multiplying the scores of sub-items such as operation type, coordinate distance, and direction angle by preset weights; the hit threshold can be dynamically adjusted according to the application type or network conditions.

[0079] In step S1062, the location of the common computing starting point is achieved by tracking and comparing the derivation relationship graph and state version number of each parallel rendering task instance.

[0080] In step S1064, the detection of the critical state change is based on the state machine defined in the application logic or the monitoring of write operations to a specific memory region.

[0081] Therefore, the method of this invention has the following significant advantages compared to the prior art: First, by predicting intent, cloud computing and image transmission are executed before the user confirms the operation. When the actual operation occurs, the target image is already ready at the edge node, achieving subjective negative latency perception and eliminating the stuttering caused by the traditional serial path. Second, based on predictive execution resource preheating and value efficiency ratio scheduling, computing instances, memory, and network pre-configuration are completed before the user operation, suppressing low-value branch resource occupation and significantly improving resource utilization in high-concurrency scenarios. Third, by combining the criticality of the operation and network status to implement differentiated encoding transmission and local predictive animation compensation, the user experience is smoothly degraded rather than interrupted in weak network environments. Finally, through prediction deviation feedback and incremental state correction, the prediction parameters are continuously converged, the system has self-evolution capabilities, and the overall energy efficiency and robustness are significantly enhanced.

[0082] The method of this invention also provides application scenarios, including: Scenario 1. E-commerce live streaming flash sale scenario (high concurrency, short bursts); Application scenario description: Millions of users watch the live stream simultaneously, and after the host announces the link, they all click "Buy Now"; Technical solution implementation: Local recognition of accelerated finger swiping + lingering on the "Buy Now" button area, predicting a 90% probability of clicking "Buy Now" (P1); Global coordination function completed within 100ms: Edge nodes expanded by 40% lightweight containers; Entry bandwidth reserved for predicted users; Pre-generated static skeleton of the order page + key button area difference screens; Low-value branches (back, refresh) are downgraded to logic pre-calculation, not pre-rendered; The actual operation is adding to the shopping cart, only the difference fields are corrected.

[0083] The technical effects are shown in Table 1: Table 1

[0084] Scenario 2. Office Collaboration Scenario (Weak Network, Cross-Regional); Application Scenario Description: Users use cloud phones to edit documents and turn pages in weak network environments such as subways and elevators; Technical Solution Implementation: Recognize page-turning gestures + document structure, predict the probability of the next page with 72% (P1); When network RTT>200ms, simulate page slide-in animation locally; Only pre-render text and structure layers, images are delayed in loading; When the actual page turn deviates from the predicted page number, incrementally correct the content block.

[0085] The technical effects are shown in Table 2: Table 2

[0086] The method of this invention demonstrates significant technical effects in typical application scenarios. In e-commerce live-streaming flash sale scenarios, by recognizing user gestures and predicting purchase intentions based on interface areas, and pre-executing edge scaling, bandwidth reservation, and order page skeleton pre-rendering, the response time of the purchase button is reduced from 420ms to 135ms, the first-screen rendering failure rate is reduced from 12.4% to 0.9%, the peak CPU usage of edge nodes is reduced from 92% to 67%, the bandwidth consumption per request is reduced from 1.8MB to 0.6MB, and the concurrency support capability is increased by 2.8 times. In weak network office collaboration scenarios, by predicting page-turning intentions, local animation compensation, and incremental content loading, the response latency of weak network operations is reduced from 310ms to 95ms, the single-page traffic consumption is reduced from 2.1MB to 0.7MB, and the user interruption rate is reduced from 14% to 2%. This invention effectively solves the problems of response latency and resource waste in traditional solutions under high concurrency and weak network environments, achieving a systematic improvement in interactive response speed, system concurrency capability, and weak network robustness, and enhancing the engineering applicability of cloud phones in complex working conditions.

[0087] To achieve the above invention, such as Figure 4 As shown, this embodiment also provides a prediction-based cloud phone seamless adaptive synchronization device 10, which includes: The perception and prediction module 100 is used to collect user interaction behavior data, device sensor data and current application interface data in real time on the user device. It predicts multiple operations that the user may perform in a short time window and their corresponding probabilities through multi-source data fusion. The prediction results are compressed into a semantic prediction instruction stream and the operation criticality level is marked. The prediction instruction stream is synchronously sent to the cloud rendering service and global resource coordination function.

[0088] The resource preheating module 200 is used to establish a mapping relationship between user operation intentions and computing, storage and network resources in the cloud and edge nodes. Before the actual user operation is confirmed, it initiates resource pre-allocation actions such as computing instance preheating, memory reservation and network service quality policy pre-configuration on the target edge node according to the operation type and criticality level in the predicted instruction stream. It also continuously monitors the global resource load and network status and feeds the system status information back to the user device in real time to dynamically adjust the prediction parameters.

[0089] The branch pre-rendering module 300 is used to receive the prediction instruction stream in the cloud rendering service, filter out multiple high-probability prediction instructions as candidate branches to be processed, create independent rendering task instances in parallel in the edge computing environment corresponding to each candidate branch, generate future screen frames for each candidate branch, compare the differences between each pre-rendering result and the current baseline screen, generate screen difference data containing only the changed parts and cache it to the edge storage node closest to the user.

[0090] The value scheduling module 400 is used to evaluate the value and cost of the pre-computation tasks of each candidate branch, calculate the value-efficiency ratio of each pre-computation task, and dynamically prioritize all parallel pre-computation tasks accordingly and perform differentiated resource allocation. It also dynamically adjusts the encoding and transmission strategies based on the operational criticality level and the real-time network link quality.

[0091] The error correction feedback module 500 is used to match and compare the actual operation with the pre-calculated prediction instruction when the user's actual operation occurs. If the match is successful, the corresponding screen difference data is directly extracted from the edge cache and applied. If the match fails, a fast error correction process is triggered, which only calculates the difference between the actual operation and the closest prediction to restore state consistency, and feeds the prediction deviation data back to the prediction function on the user device to optimize the prediction parameters.

[0092] The device in this invention effectively solves the problem of asynchronous interaction caused by network latency and computational overhead in traditional reactive architectures. It achieves integrated synchronization of the entire process from intent perception, resource preheating, branch precomputation to dynamic scheduling and self-correction, significantly reducing end-to-end interaction latency, improving screen smoothness and system resource utilization in weak network environments, and enhancing the engineering practical value of cloud phones in high-concurrency and dynamic network conditions.

[0093] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0094] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A prediction-based, seamless adaptive synchronization method for cloud phones, characterized in that, include: On the user device side, user interaction behavior data, device sensor data and current application interface data are collected in real time. By fusing multi-source data, multiple operations that the user may perform in the future short time window and their corresponding probabilities are predicted. The prediction results are compressed into a semantic prediction instruction stream and the operation criticality level is marked. The prediction instruction stream is synchronously sent to the cloud rendering service and global resource coordination function. Establish a mapping relationship between user operation intentions and computing, storage and network resources in the cloud and edge nodes. Before the actual operation is confirmed, based on the operation type and criticality level in the predicted instruction stream, initiate resource pre-allocation actions on the target edge node, such as computing instance preheating, memory reservation and network service quality policy pre-configuration. Continuously monitor global resource load and network status, and feed back system status information to the user device in real time to dynamically adjust prediction parameters. In the cloud rendering service, the prediction instruction stream is received, multiple high-probability prediction instructions are selected as candidate branches to be processed, independent rendering task instances are created in parallel in the edge computing environment corresponding to each candidate branch, future screen frames of each candidate branch are generated, each pre-rendered result is compared with the current baseline screen, screen difference data containing only the changed parts is generated and cached in the edge storage node closest to the user. The value and cost of the pre-computation tasks for each candidate branch are evaluated, the value-efficiency ratio of each pre-computation task is calculated, and all parallel pre-computation tasks are dynamically prioritized and differentiated resource allocation is performed accordingly. The coding and transmission strategies are dynamically adjusted based on the operational criticality level and the real-time network link quality. When a user actually performs an action, the actual action is compared with the pre-calculated prediction instruction. If the match is successful, the corresponding screen difference data is directly extracted from the edge cache and applied. If the match fails, a fast error correction process is triggered, which only calculates the difference between the actual action and the closest prediction to restore state consistency. The prediction deviation data is then fed back to the prediction function on the user's device to optimize the prediction parameters.

2. The method as described in claim 1, characterized in that, Multi-source data fusion prediction includes: Capture and calculate the kinematic characteristics of instantaneous velocity, acceleration, and trajectory curvature from raw touch events on the touchscreen; read sensor data of device acceleration and rotation speed to sense changes in device attitude; Obtain complete interface information of the currently running application in the foreground, parse the screen coordinates, shape, text and semantic tags of all interactive interface elements, and construct the spatial layout diagram and semantic relationship diagram of the current interface; By integrating temporal correlation analysis, spatial correlation analysis, and attention guidance, a joint prediction distribution of multiple user operations and their corresponding probabilities within a future short-term window is output.

3. The method as described in claim 2, characterized in that, The prediction results are compressed into a semantic prediction instruction stream, including: The underlying data of coordinate sequences and action types are abstracted into high-level instructions that include action type, start coordinates, end coordinates, and estimated duration; The criticality level of the operation is dynamically marked based on the application context that generated the prediction instruction and the criticality of the prediction operation in the application. The prediction probability, semantic instructions, and operation criticality level are encapsulated to form a prediction instruction stream.

4. The method as described in claim 1, characterized in that, Resource pre-allocation actions include: On the edge node closest to the corresponding user location and where the load allows, create or preheat computing resource instances that meet the corresponding performance requirements according to the operation criticality level, reserve memory capacity, and pre-configure network service quality policies for the transmission path from the corresponding computing instance to the user device.

5. The method as described in claim 1, characterized in that, The comparison of the pre-rendered result with the current reference image is specifically as follows: each pre-rendered result is compared with the current reference image pixel by pixel or block by block, and the image difference data containing only the changed parts is extracted. The image difference data adopts an incremental encoding method to reduce storage and transmission overhead.

6. The method as described in claim 1, characterized in that, The differentiated resource allocation and encoding transmission strategy is as follows: High-value-efficiency pre-computation tasks are given the highest scheduling priority, and a layered, progressive preloading strategy is adopted to prioritize loading low-resolution resources and core logic states. For the pre-calculated results corresponding to P0 level critical operations, when network bandwidth is limited, priority is given to ensuring smoothness of the screen, reducing the resolution and increasing the key frame frequency, and enabling forward error correction on the transmission side to enhance the protection of core data. For P2 level normal operations, the frame rate can be appropriately reduced or a higher compression ratio encoding configuration can be used.

7. The method as described in claim 6, characterized in that, The value efficiency ratio is calculated as follows: A value scoring model is constructed to calculate a dynamic value score for each predicted intent branch. The value score integrates the prediction probability, the operational criticality level, real-time feedback from the client side, and global resource load and network status information. Simultaneously build a computational overhead model to estimate the rendering processing time and memory usage of each intent branch; calculate the ratio of the value score of each branch to the estimated computational overhead in real time as the value efficiency ratio.

8. The method as described in claim 1, characterized in that, The rapid error correction process includes: The client generates correction information containing only the core differences between the actual operation and the closest prediction and sends it to the edge nodes; The cloud rendering service at the edge node locates the common computing starting point on which the current application state and the predicted branch are based. It only calculates the difference between the corresponding common starting point and the actual operation target state, and quickly encodes the corrected image and sends it to the client.

9. The method as described in claim 1, characterized in that, Also includes: During the window period while waiting for the cloud confirmation image to arrive, the client runs a local lightweight image prediction based on the recently received multiple frames and real-time touch data, generates the next frame of the local predicted image and displays it, and updates it by smooth replacement or fusion after receiving the cloud confirmation image. When a sudden increase in network round-trip latency is detected, the client simulates an instantaneous animation feedback that conforms to the law of inertia based on the physical properties of the operation when it leaves the screen. The physical properties are obtained by fitting the last few sampling points of the touch event stream.

10. A prediction-based, seamless adaptive synchronization device for cloud phones, characterized in that, include: The perception and prediction module is used to collect user interaction behavior data, device sensor data and current application interface data in real time on the user device. It predicts multiple operations that the user may perform in a short time window and their corresponding probabilities through multi-source data fusion. The prediction results are compressed into a semantic prediction instruction stream and the operation criticality level is marked. The prediction instruction stream is synchronously sent to the cloud rendering service and global resource coordination function. The resource preheating module is used to establish a mapping relationship between user operation intentions and computing, storage and network resources in the cloud and edge nodes. Before the actual operation is confirmed, it initiates resource pre-allocation actions such as computing instance preheating, memory reservation and network service quality policy pre-configuration on the target edge node according to the operation type and criticality level in the predicted instruction stream. It also continuously monitors the global resource load and network status and feeds the system status information back to the user device in real time to dynamically adjust the prediction parameters. The branch pre-rendering module is used to receive the prediction instruction stream in the cloud rendering service, filter out multiple high-probability prediction instructions as candidate branches to be processed, create independent rendering task instances in parallel in the edge computing environment corresponding to each candidate branch, generate future screen frames for each candidate branch, compare the differences between each pre-rendering result and the current baseline screen, generate screen difference data containing only the changed parts and cache it to the edge storage node closest to the user. The value scheduling module is used to evaluate the value and cost of the pre-computation tasks of each candidate branch, calculate the value efficiency ratio of each pre-computation task, and dynamically prioritize all parallel pre-computation tasks accordingly and perform differentiated resource allocation. It also dynamically adjusts the encoding and transmission strategies based on the criticality level of the operation and the real-time network link quality. The error correction feedback module is used to match and compare the actual user operation with the pre-calculated prediction instruction when the actual operation occurs. If the match is successful, the corresponding screen difference data is directly extracted from the edge cache and applied. If the match fails, a fast error correction process is triggered, which only calculates the difference between the actual operation and the closest prediction to restore state consistency, and feeds the prediction deviation data back to the prediction function on the user device to optimize the prediction parameters.

Images