Dynamic adaptive edge-cloud collaborative computing split point and compression strategy adjustment method

Abstract

The application discloses a dynamic self-adaptive edge-cloud collaborative computing segmentation point and compression strategy adjustment method. The core of the method is to deploy a dynamic decision engine on the edge device. The engine real-time perceives network state (such as bandwidth, delay) and application specified task priority, and according to the preset decision logic (such as query table or lightweight prediction model), dynamically selects the optimal model segmentation point and feature compression level combination from a deep learning model supporting multiple segmentation points and equipped with variable bit rate layered quantization feature compression model for each segmentation point. The combination is transmitted to the cloud, and the corresponding backend model is loaded to complete the calculation. Through the dynamic adjustment mechanism, the application realizes the real-time optimal balance among inference delay, task accuracy and network overhead, and significantly improves the performance stability, resource utilization and adaptability of the system in a variable environment.

Description

Technical Field

[0001] This invention belongs to the fields of edge computing, deep learning model deployment and network communication technology, and in particular, a method for dynamically adaptive edge-cloud collaborative computing segmentation point and compression strategy adjustment. Background Technology

[0002] With the widespread application of edge-cloud collaborative computing architecture in the field of real-time video analytics, how to efficiently allocate computing tasks and transmit data between resource-constrained edge devices (such as smart cameras and mobile terminals) and powerful cloud servers has become a key technical challenge.

[0003] Internationally accepted mainstream technical solutions, such as the collaborative inference frameworks proposed in pioneering studies like Neurosurgeon (ASPLOS '17) and Edger (MobiCom '18), generally employ a model partitioning strategy. This strategy splits a large deep learning model at an intermediate layer, forming a front-end and a back-end. The relatively less computationally intensive front-end (usually the feature extraction backbone network) is deployed on edge devices, while the computationally intensive back-end (containing deeper networks and task processing heads) is deployed on cloud servers. After the edge devices process the raw data (such as video frames), the generated intermediate feature maps are transmitted to the cloud via the network, where the final analysis and inference are performed.

[0004] To address the issue of massive amounts of intermediate feature map data and reduce latency and bandwidth pressure caused by network transmission, academia and industry have further researched various deep feature compression techniques. These techniques, including advanced algorithms based on transform coding, quantization, and autoencoder-based methods, are used to compress feature maps before transmission.

[0005] However, most of these existing collaborative computing frameworks adopt a static deployment strategy. Specifically, the model's split points and the compression ratio of the feature compression algorithm are usually determined once during the system design and deployment phase based on preset or average network conditions, and remain fixed during system operation.

[0006] This static strategy performs well in ideal scenarios where the network environment is stable and the task requirements remain unchanged. However, in real-world industrial and daily applications, the dynamic nature of networks and tasks is the norm. On the one hand, the bandwidth and latency of wireless networks (such as 5G / Wi-Fi) can fluctuate dramatically due to factors such as signal strength, channel congestion, and network switching. On the other hand, video analytics systems often need to dynamically adjust the priority and accuracy requirements of different analysis tasks (such as face recognition in routine inspections and abnormal behavior detection in the event of an emergency) based on the urgency of the event or the real-time needs of the user.

[0007] Fixed segmentation points and compression rates cannot adapt to this dual dynamic change. When the network is congested, continuing to transmit a fixed amount of data will lead to unacceptable latency, potentially missing the opportunity to respond to critical events. Conversely, when network conditions are good, fixed low bit rate compression wastes valuable bandwidth resources and cannot improve analysis accuracy by transmitting higher quality features. Therefore, there is an urgent need for an intelligent collaborative computing and transmission scheme that can perceive changes in the environment and task in real time and intelligently and dynamically adjust model segmentation and data transmission strategies to continuously provide stable, efficient, and flexible intelligent video analytics services in a changing environment. Summary of the Invention

[0008] The purpose of this invention is to address the problems in existing edge-cloud collaborative computing schemes, where the statically fixed model segmentation points and feature compression rates fail to adapt to dynamic changes in network environments and task requirements, leading to unstable system performance and low resource utilization. This invention proposes an adaptive model segmentation and variable bit-rate feature compression method for dynamic network environments. This method constructs a dynamically adaptive edge-cloud collaborative computing framework, the core of which is a dynamic decision engine deployed at the edge. This engine intelligently selects the optimal combination of model segmentation points and feature compression rates based on real-time perceived network conditions and task priorities, thereby achieving a dynamic and real-time optimal balance between inference latency, task accuracy, and network overhead.

[0009] The technical solution to achieve the purpose of this invention is: a dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy adjustment method. This method requires system deployment before execution, specifically including:

[0010] At the edge device, a common front-end model, VQ-VAE encoders corresponding to all segmentation points, and a dynamic decision module are deployed. VQ-VAE is a variable bit rate vector quantization variational autoencoder that has been pre-trained. The dynamic decision module is responsible for dynamically selecting the combination of model segmentation points and feature compression rates based on the real-time changing system state.

[0011] Deploy all backend models, all VQ-VAE codebooks, and VQ-VAE decoders in the cloud;

[0012] The adjustment method includes the following steps:

[0013] Step 1, Data Acquisition and Status Assessment: The edge device captures the input data to be processed. At the same time, the dynamic decision module deployed on the edge side assesses the current network status and the task priority P specified by the upper layer application in real time.

[0014] Step 2, Dynamic Decision Generation: Based on the collected network state information and its internal decision-making logic, the dynamic decision-making module determines the optimal configuration combination, i.e., selects a model split point index. and a feature compression level index ;

[0015] Step 3, Edge-side front-end calculation: The edge device calculates based on the decision results. Execute the common front-end model until the split point is reached. This generates an intermediate feature map;

[0016] Step 4, Feature Compression: Calling and Segmentation Points The corresponding VQ-VAE encoder, and the feature compression level index based on the decision. The intermediate feature map is compressed to obtain a codebook index with a relatively smaller data volume;

[0017] Step 5, Data Packaging and Transmission: The edge device will compress the codebook index, along with the configuration information for this decision. Package them together and transmit them to the cloud server over the network;

[0018] Step 6, Cloud Reception and Task Preparation: After receiving the data packet, the cloud server first parses out the configuration information. Then based on the split point index Load and prepare the corresponding backend model and VQ-VAE decoder;

[0019] Step 7, Feature Decompression and Reconstruction: The cloud server uses the corresponding VQ-VAE decoder and the received codebook index to decompress and reconstruct the feature based on the feature compression level index. Reconstruct intermediate feature maps;

[0020] Step 8, Cloud-based backend computation and result output: Input the reconstructed intermediate feature map into the prepared backend model, complete the remaining inference computation, and finally output the task result.

[0021] Furthermore, the common front-end model and back-end model are obtained by constructing a deep learning model that supports multiple segmentation points. Specifically, the deep learning model that supports multiple segmentation points is divided into a common front-end feature extraction network, i.e., the common front-end model. and A different backend task processing network, i.e., a backend model Meanwhile, both the public front-end model and the back-end model have been trained.

[0022] Furthermore, the construction and partitioning process of the deep learning model supporting multiple segmentation points specifically includes:

[0023] Select a deep learning model suitable for multi-task analysis;

[0024] In the backbone network of a deep learning model, selection is made along its depth direction. Each logical layer serves as a potential dividing point. , , , , For any split point The model is divided into:

[0025] Front-end: From input layer to split point All layers, denoted as ;

[0026] Backend part: From the split point All layers from the final output layer onwards are denoted as... All front-end components Share from input layer to shallowest split point The network weights are allowed to be independent in their subsequent stages.

[0027] Furthermore, the training process for the public front-end model and back-end model includes:

[0028] Ensure that the front-end features at any segmentation point can support the back-end model to complete near-optimal inference;

[0029] Using a joint training or phased training strategy, the overall objective loss function of training is... It is a weighted sum of the task losses corresponding to all split points, with each split point only responsible for the task output at its own depth:

[0030]

[0031] In the formula, It is the input data. These are the corresponding truth labels. Indicates input data After front-end model processing to the segmentation point Intermediate features of the output It is the loss function of the downstream task. These are hyperparameter weights used to balance the training contributions of different split points. This indicates a summation.

[0032] Furthermore, the variable bit rate vector quantization variational autoencoder adopts a hierarchical quantization structure and includes an encoder. and a decoder ;

[0033] encoder : Divide the point intermediate feature maps output Encoded as Residual latent representation at each level ;

[0034] Hierarchical quantization: Residual representation at each level Each uses its own independent codebook Perform vector quantization to obtain the quantized representation. and the corresponding codebook index The quantization process is represented as ;

[0035] decoder : Able to receive any Quantization representation at each level The sum of the values ​​and the original feature map are then reconstructed from them. ;in, .

[0036] Furthermore, the training process of the variable bit rate vector quantization variational autoencoder is as follows:

[0037] By selecting before transmission Layer codebook index This can be achieved. Different levels of compression ratio;

[0038] Total loss function during training It is necessary to ensure good reconstruction quality at all compression levels, and the formula is as follows:

[0039]

[0040] In the formula, Before use Reconstruction loss when reconstructing layer information; It is the first The codebook commitment loss of the layer is used to induce the residual representation of the encoder output. Approximating its quantized value ; This indicates that the gradient operation has stopped; and These are weighting coefficients used to balance the various losses. This indicates a summation.

[0041] Furthermore, the dynamic decision-making module has a built-in status monitoring unit for collecting key input information required for decision-making;

[0042] Network status : By performing routine network probes or querying network statistics provided by the underlying operating system, metrics for evaluating network quality, including bandwidth, can be obtained. and delay ;

[0043] Task status : Receive the priority or type of the current task from the upper-layer application or system configuration. .

[0044] Furthermore, the dynamic decision-making module selects a suitable configuration from a preset action space based on the input state. ,in, For the split point index, This is a compression level index; the dynamic decision-making module makes decisions based on the following:

[0045] Total delay for:

[0046]

[0047] In the formula, It is the time for the edge side to perform front-end calculations. It is the network time for transmitting compressed features. This refers to the time it takes for the cloud to complete subsequent calculations.

[0048] Furthermore, the decision-making logic of the dynamic decision-making module is implemented in the following way:

[0049] Rule-based decision table: Through offline testing and analysis, a lookup table is pre-calculated and filled in for different network state ranges and task priority combinations; at runtime, the dynamic decision module looks up the table based on the current state to obtain the recommended configuration. .

[0050] Furthermore, the decision-making logic of the dynamic decision-making module is implemented in the following way:

[0051] Learning-based decision-making models: Train a lightweight prediction model that learns the optimal action from the input state. The model is trained offline using data collected in various simulated or real-world environments, with the goal of learning a decision-making strategy that effectively balances accuracy and latency.

[0052] Compared with the prior art, the significant advantages of this invention are:

[0053] (1) High adaptability and performance stability: This invention can sense the dynamic changes in network bandwidth, latency, and upper-layer task priorities in real time, and intelligently adjust the model segmentation point and feature compression rate accordingly. This enables the system to continuously provide stable service quality in variable and uncertain real-world environments (such as video backhaul analysis under mobile networks), effectively avoiding high latency caused by network congestion or resource waste when network conditions are good, and ensuring the timeliness of response to critical tasks.

[0054] (2) Dynamic balance between resource utilization and task accuracy: This invention achieves a real-time, optimal trade-off between system resources (edge ​​computing power, network bandwidth) and task performance (inference accuracy, response latency) through dynamic decision-making. When network conditions are favorable, the system can automatically select to transmit higher quality features (lower compression ratio) to improve analysis accuracy; when network conditions deteriorate or task priority is low, it prioritizes low latency and selects a higher compression ratio or a earlier segmentation point. This flexibility maximizes the overall system performance at any given moment.

[0055] (3) The framework’s versatility and scalability: The collaborative framework with multiple segmentation points and multiple compression levels proposed in this invention has good versatility. This framework is not limited to specific deep learning models or tasks and can be easily applied to scenarios that include various visual analysis tasks such as object detection, pose estimation, and image segmentation. By adding or removing segmentation points and compression levels, the granularity of the strategy can be flexibly adjusted, making it easy to extend and upgrade.

[0056] (4) Low overhead and high efficiency in decision-making: The dynamic decision-making module deployed at the edge is lightweight. Whether based on a lookup table or a small predictive model, the computational overhead and memory usage of its decision-making process are extremely low, and it will not impose a significant burden on the edge device. The decision-making process is completed locally in an instant, ensuring the high efficiency of the entire adaptive adjustment mechanism and enabling it to respond quickly to environmental changes.

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the overall system flow of a method for dynamically adaptive edge-cloud collaborative computing segmentation points and compression strategies in one embodiment.

[0059] Figure 2 This is a structural diagram of a variable bit rate hierarchical quantization VQ-VAE feature compression model in one embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0062] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0063] In one embodiment, combined Figure 1 This paper presents a dynamic and adaptive method for adjusting edge-cloud collaborative computing split points and compression strategies. This method requires hardware deployment before execution, specifically including:

[0064] At the edge device, deploy a common front-end model and VQ-VAE encoders corresponding to all segmentation points. The system includes a dynamic decision module; wherein, VQ-VAE is a variable bit rate vector quantization variational autoencoder, and is pre-trained; the dynamic decision module is responsible for dynamically selecting the combination of model segmentation points and feature compression rates according to the real-time changing system state.

[0065] Deploy all backend models in the cloud. All VQ-VAE codebooks and VQ-VAE decoders ;

[0066] The method includes the following steps:

[0067] Step 1, Data Acquisition and Status Assessment: The edge device captures the input data to be processed (e.g., a frame of video image). At the same time, the dynamic decision module deployed on the edge side assesses the current network status (e.g., bandwidth B, latency D) and the task priority P specified by the upper layer application in real time.

[0068] Step 2, Dynamic Decision Generation: Based on the collected network state information and its internal decision-making logic (e.g., by querying a preset decision table or running a lightweight decision model), the dynamic decision-making module determines the optimal configuration combination, i.e., selects a model split point index. and a feature compression level index ;

[0069] Step 3, Edge-side front-end calculation: The edge device calculates based on the decision results. Execute the common front-end model until the split point is reached. This generates an intermediate feature map;

[0070] Step 4, Feature Compression: Calling and Segmentation Points The corresponding VQ-VAE encoder, and the feature compression level index based on the decision. The intermediate feature map is compressed to obtain a codebook index with a relatively smaller data volume;

[0071] Step 5, Data Packaging and Transmission: The edge device will compress the codebook index, along with the configuration information for this decision. Package them together and transmit them to the cloud server over the network;

[0072] Step 6, Cloud Reception and Task Preparation: After receiving the data packet, the cloud server first parses out the configuration information. Then based on the split point index Load and prepare the corresponding backend model and VQ-VAE decoder;

[0073] Step 7, Feature Decompression and Reconstruction: The cloud server uses the corresponding VQ-VAE decoder and the received codebook index to decompress and reconstruct the feature based on the feature compression level index. Reconstruct intermediate feature maps;

[0074] Step 8, Cloud-side backend computation and result output: Input the reconstructed intermediate feature map into the prepared backend model, complete the remaining inference computation, and finally output the task result (e.g., the target's position information or pose estimation result).

[0075] Furthermore, in one embodiment, the common front-end model and the back-end model are obtained by constructing a deep learning model that supports multiple segmentation points. Specifically, the deep learning model that supports multiple segmentation points is divided into a common front-end feature extraction network, i.e., a common front-end model. and A different backend task processing network, i.e., a backend model Meanwhile, both the public front-end model and the back-end model have been trained.

[0076] Furthermore, in one embodiment, the construction and partitioning process of the deep learning model supporting multiple segmentation points specifically includes:

[0077] Select a deep learning model suitable for multi-task analysis (e.g., a combined model with ResNet / MobileNet as the backbone network, followed by a YOLOv8 task head for object detection and an HRNet task head for pose estimation).

[0078] In the backbone network of a deep learning model, selection is made along its depth direction. Each logical layer serves as a potential dividing point. , , , , For any split point The model is divided into:

[0079] Front-end: From input layer to split point All layers, denoted as ;

[0080] Backend part: From the split point All layers from then until the final output layer are denoted as... All front-end components Share from input layer to shallowest split point The network weights are allowed to be independent in their subsequent stages.

[0081] Furthermore, in one embodiment, the training process of the common front-end model and back-end model includes:

[0082] Ensure that the front-end features at any segmentation point can support the back-end model to complete near-optimal inference;

[0083] Using a joint training or phased training strategy, the overall objective loss function of training is... It is a weighted sum of the task losses corresponding to all split points, with each split point only responsible for the task output at its own depth:

[0084]

[0085] In the formula, It is the input data (such as an image). These are the corresponding truth labels. Indicates input data After front-end model processing to the segmentation point Intermediate features of the output It is the loss function for downstream tasks (e.g., for detection and pose estimation, it is a combination of classification loss, regression loss, and keypoint heatmap loss). These are hyperparameter weights used to balance the training contributions of different split points, and can typically be set to... Or adjust based on experience. This indicates a summation.

[0086] Furthermore, in one embodiment, combined with Figure 2 The variable bit rate vector quantization variational autoencoder adopts a hierarchical quantization structure and includes an encoder. and a decoder ;

[0087] encoder : Divide the point intermediate feature maps output Encoded as Residual latent representation at each level ;

[0088] Hierarchical quantization: Residual representation at each level They all use their own independent codebooks. Perform vector quantization to obtain the quantized representation. and the corresponding codebook index The quantization process is represented as ;

[0089] decoder : Able to receive any Quantization representation at each level The sum, and reconstruct the original feature map from it. ;in, .

[0090] Furthermore, in one embodiment, the training process of the variable bit rate vector quantization variational autoencoder is as follows:

[0091] By selecting before transmission Layer codebook index This can be achieved. Different levels of compression ratio;

[0092] Total loss function during training It is necessary to ensure good reconstruction quality at all compression levels, and the formula is as follows:

[0093]

[0094] In the formula, Before use Reconstruction loss when reconstructing layer information; It is the first The codebook commitment loss of a layer is used to induce the residual representation of the encoder output. Approximating its quantized value ; This represents the Stop Gradient operation; and These are weighting coefficients used to balance the various losses. This indicates a summation.

[0095] Furthermore, in one embodiment, the dynamic decision-making module incorporates a status monitoring unit for collecting key input information required for decision-making;

[0096] Network status By performing routine network probing (such as periodic ping tests and file transfer rate estimations) or querying network statistics provided by the underlying operating system, metrics for evaluating network quality can be obtained, primarily including bandwidth. and delay ;

[0097] Task status : Receive the priority or type of the current task from the upper-layer application or system configuration. For example, define the task as "normal mode" (low priority) or "focus mode" (high priority).

[0098] Furthermore, in one embodiment, the dynamic decision-making module determines the state based on the input state ( , Select a suitable configuration from the preset motion space. ,in, For the split point index, This is a compression level index; the dynamic decision-making module makes decisions based on the following:

[0099] Total delay for:

[0100]

[0101] In the formula, It is the time for the edge side to perform front-end calculations. The network time for transmitting compressed features (as opposed to feature size and bandwidth) (related) This refers to the time it takes for subsequent calculations to be completed in the cloud. These values ​​can be pre-measured through offline profiling or estimated through online measurement.

[0102] Here is an example of a strategy: the decision logic will be based on task priority. Adjust its focus. When in a "high priority" state, the strategy will tend to choose configurations that provide higher accuracy. As long as the estimated total delay Within acceptable thresholds. When in a "low priority" state or under network conditions. In poor cases, the strategy will prioritize selecting the option that minimizes the total latency. The configuration is optimized to ensure basic system responsiveness and conserve resources.

[0103] Preferably, in some embodiments, the decision logic of the dynamic decision module is implemented in, but not limited to, the following ways:

[0104] Rule-based decision table: This is a simple and efficient implementation. Through offline testing and analysis, a lookup table is pre-calculated and filled for different network state ranges and task priority combinations. At runtime, the dynamic decision module looks up the table based on the current state to obtain the recommended configuration. .

[0105] Preferably, in some embodiments, the decision logic of the dynamic decision module is implemented in, but not limited to, the following ways:

[0106] Learning-based decision models: Train a lightweight prediction model (such as a small neural network or gradient boosting tree model) that learns from the input states. , To the optimal action The model is trained offline using data collected in various simulated or real-world environments, with the goal of learning a decision-making strategy that effectively balances accuracy and latency.

[0107] In one embodiment, a dynamically adaptive edge-cloud collaborative computing split point and compression strategy adjustment system is provided, the system comprising sequentially executing:

[0108] The first module is used to deploy a common front-end model, VQ-VAE encoders for all segmentation points, and dynamic decision modules on edge devices; it is also used to deploy all back-end models, all VQ-VAE codebooks, and VQ-VAE decoders on the cloud.

[0109] The second module is used to realize data acquisition and status assessment: the edge device captures the input data to be processed, and at the same time, the dynamic decision module deployed on the edge side assesses the current network status and the task priority P specified by the upper layer application in real time.

[0110] The third module is used to generate dynamic decisions: Based on the collected network state information and its internal decision-making logic, the dynamic decision-making module determines the optimal configuration combination, that is, selects a model split point index. and a feature compression level index ;

[0111] The fourth module is used to implement edge-side front-end computing: the edge device performs calculations based on the decision results. Execute the common front-end model until the split point is reached. This generates an intermediate feature map;

[0112] The fifth module is used to implement feature compression: calling and segmentation points. The corresponding VQ-VAE encoder, and the feature compression level index based on the decision. The intermediate feature map is compressed to obtain a codebook index with a relatively smaller data volume;

[0113] The sixth module is used for data packaging and transmission: the edge device will compress the codebook index, along with the configuration information for this decision. Package them together and transmit them to the cloud server over the network;

[0114] The seventh module is used for cloud reception and task preparation: After receiving the data packet, the cloud server first parses out the configuration information. Then based on the split point index Load and prepare the corresponding backend model and VQ-VAE decoder;

[0115] The eighth module is used for feature decompression and reconstruction: the cloud server uses the corresponding VQ-VAE decoder and the received codebook index to perform feature decompression and reconstruction based on the feature compression level index. Reconstruct intermediate feature maps;

[0116] The ninth module is used to realize the cloud-side backend computing and result output: the reconstructed intermediate feature map is input into the prepared backend model to complete the remaining inference computing and finally output the task results.

[0117] Specific limitations regarding the dynamic adaptive edge-cloud collaborative computing partitioning and compression strategy adjustment system can be found in the limitations of the dynamic adaptive edge-cloud collaborative computing partitioning and compression strategy adjustment method described above, and will not be repeated here. Each module in the aforementioned dynamic adaptive edge-cloud collaborative computing partitioning and compression strategy adjustment system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device, or stored in the memory of the computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0118] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0119] At the edge device, a common front-end model, VQ-VAE encoders corresponding to all segmentation points, and a dynamic decision module are deployed. VQ-VAE is a variable bit rate vector quantization variational autoencoder that has been pre-trained. The dynamic decision module is responsible for dynamically selecting the combination of model segmentation points and feature compression rates based on the real-time changing system state.

[0120] Deploy all backend models, all VQ-VAE codebooks, and VQ-VAE decoders in the cloud;

[0121] Step 1, Data Acquisition and Status Assessment: The edge device captures the input data to be processed. At the same time, the dynamic decision module deployed on the edge side assesses the current network status and the task priority P specified by the upper layer application in real time.

[0122] Step 2, Dynamic Decision Generation: Based on the collected network state information and its internal decision-making logic, the dynamic decision-making module determines the optimal configuration combination, i.e., selects a model split point index. and a feature compression level index ;

[0123] Step 3, Edge-side front-end calculation: The edge device calculates based on the decision results. Execute the common front-end model until the split point is reached. This generates an intermediate feature map;

[0124] Step 4, Feature Compression: Calling and Segmentation Points The corresponding VQ-VAE encoder, and the feature compression level index based on the decision. The intermediate feature map is compressed to obtain a codebook index with a relatively smaller data volume;

[0125] Step 5, Data Packaging and Transmission: The edge device will compress the codebook index, along with the configuration information for this decision. Package them together and transmit them to the cloud server over the network;

[0126] Step 6, Cloud Reception and Task Preparation: After receiving the data packet, the cloud server first parses out the configuration information. Then based on the split point index Load and prepare the corresponding backend model and VQ-VAE decoder;

[0127] Step 7, Feature Decompression and Reconstruction: The cloud server uses the corresponding VQ-VAE decoder and the received codebook index to decompress and reconstruct the feature based on the feature compression level index. Reconstruct intermediate feature maps;

[0128] Step 8, Cloud-based backend computation and result output: Input the reconstructed intermediate feature map into the prepared backend model, complete the remaining inference computation, and finally output the task result.

[0129] For specific limitations on each step, please refer to the limitations on the dynamic adaptive edge-cloud collaborative computing split point and compression strategy adjustment method mentioned above, which will not be repeated here.

[0130] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0131] At the edge device, a common front-end model, VQ-VAE encoders corresponding to all segmentation points, and a dynamic decision module are deployed. VQ-VAE is a variable bit rate vector quantization variational autoencoder that has been pre-trained. The dynamic decision module is responsible for dynamically selecting the combination of model segmentation points and feature compression rates based on the real-time changing system state.

[0132] Deploy all backend models, all VQ-VAE codebooks, and VQ-VAE decoders in the cloud;

[0133] Step 1, Data Acquisition and Status Assessment: The edge device captures the input data to be processed. At the same time, the dynamic decision module deployed on the edge side assesses the current network status and the task priority P specified by the upper layer application in real time.

[0134] Step 2, Dynamic Decision Generation: Based on the collected network state information and its internal decision-making logic, the dynamic decision-making module determines the optimal configuration combination, i.e., selects a model split point index. and a feature compression level index ;

[0135] Step 3, Edge-side front-end calculation: The edge device calculates based on the decision results. Execute the common front-end model until the split point is reached. This generates an intermediate feature map;

[0136] Step 4, Feature Compression: Calling and Segmentation Points The corresponding VQ-VAE encoder, and the feature compression level index based on the decision. The intermediate feature map is compressed to obtain a codebook index with a relatively smaller data volume;

[0137] Step 5, Data Packaging and Transmission: The edge device will compress the codebook index, along with the configuration information for this decision. Package them together and transmit them to the cloud server over the network;

[0138] Step 6, Cloud Reception and Task Preparation: After receiving the data packet, the cloud server first parses out the configuration information. Then based on the split point index Load and prepare the corresponding backend model and VQ-VAE decoder;

[0139] Step 7, Feature Decompression and Reconstruction: The cloud server uses the corresponding VQ-VAE decoder and the received codebook index to decompress and reconstruct the feature based on the feature compression level index. Reconstruct intermediate feature maps;

[0140] Step 8, Cloud-based backend computation and result output: Input the reconstructed intermediate feature map into the prepared backend model, complete the remaining inference computation, and finally output the task result.

[0141] For specific limitations on each step, please refer to the limitations on the dynamic adaptive edge-cloud collaborative computing split point and compression strategy adjustment method mentioned above, which will not be repeated here.

[0142] In summary, this invention achieves a real-time optimal balance between inference latency, task accuracy, and network overhead through a dynamic adjustment mechanism, significantly improving the system's performance stability, resource utilization, and adaptability in variable environments.

[0143] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A method for dynamically adaptive edge-cloud collaborative computing segmentation point and compression strategy adjustment, characterized in that, This method requires system deployment before execution, specifically including: At the edge device, a common front-end model, VQ-VAE encoders corresponding to all segmentation points, and a dynamic decision module are deployed. VQ-VAE is a variable bit rate vector quantization variational autoencoder that has been pre-trained. The dynamic decision module is responsible for dynamically selecting the combination of model segmentation points and feature compression rates based on the real-time changing system state. Deploy all backend models, all VQ-VAE codebooks, and VQ-VAE decoders in the cloud; The adjustment method includes the following steps: Step 1, Data Acquisition and Status Assessment: The edge device captures the input data to be processed. At the same time, the dynamic decision module deployed on the edge side assesses the current network status and the task priority P specified by the upper layer application in real time. Step 2, Dynamic Decision Generation: Based on the collected network state information and its internal decision-making logic, the dynamic decision-making module determines the optimal configuration combination, i.e., selects a model split point index. and a feature compression level index ; Step 3, Edge-side front-end calculation: The edge device calculates based on the decision results. Execute the common front-end model until the split point is reached. This generates an intermediate feature map; Step 4, Feature Compression: Calling and Segmentation Points The corresponding VQ-VAE encoder, and the feature compression level index based on the decision. The intermediate feature map is compressed to obtain a codebook index with a relatively smaller data volume; Step 5, Data Packaging and Transmission: The edge device will compress the codebook index, along with the configuration information for this decision. Package them together and transmit them to the cloud server over the network; Step 6, Cloud Reception and Task Preparation: After receiving the data packet, the cloud server first parses out the configuration information. Then based on the split point index Load and prepare the corresponding backend model and VQ-VAE decoder; Step 7, Feature Decompression and Reconstruction: The cloud server uses the corresponding VQ-VAE decoder and the received codebook index to decompress and reconstruct the feature based on the feature compression level index. Reconstruct intermediate feature maps; Step 8, Cloud-based backend computation and result output: Input the reconstructed intermediate feature map into the prepared backend model, complete the remaining inference computation, and finally output the task result; The dynamic decision-making module selects a suitable configuration from the preset action space based on the input state. ,in, For the split point index, This is a compression level index; the dynamic decision-making module makes decisions based on the following: Total delay for: ； In the formula, It is the time for the edge side to perform front-end calculations. It is the network time for transmitting compressed features. This refers to the time it takes for the cloud to complete subsequent calculations.

2. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 1, characterized in that, The common front-end model and back-end model are obtained by constructing a deep learning model that supports multiple segmentation points. Specifically, the deep learning model that supports multiple segmentation points is divided into a common front-end feature extraction network, i.e., the common front-end model. and A different backend task processing network, i.e., a backend model Meanwhile, both the public front-end model and the back-end model have been trained.

3. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 2, characterized in that, The construction and partitioning process of the deep learning model supporting multiple segmentation points specifically includes: Select a deep learning model suitable for multi-task analysis; In the backbone network of a deep learning model, selection is made along its depth direction. Each logical layer serves as a potential dividing point. , , , For any split point The model is divided into: Front-end: From input layer to split point All layers, denoted as ; Backend part: From the split point All layers from the final output layer onwards are denoted as... All front-end components Share from input layer to shallowest split point The network weights are allowed to be independent in their subsequent stages.

4. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 2, characterized in that, The training process for the common front-end model and back-end model includes: Ensure that the front-end features at any segmentation point can support the back-end model to complete near-optimal inference; Using a joint training or phased training strategy, the overall objective loss function of training is... It is a weighted sum of the task losses corresponding to all split points, with each split point only responsible for the task output at its own depth: ； In the formula, It is the input data. These are the corresponding truth labels. Indicates input data After front-end model processing to the segmentation point Intermediate features of the output It is the loss function of the downstream task. These are hyperparameter weights used to balance the training contributions of different split points. This indicates a summation.

5. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 1, characterized in that, The variable bit rate vector quantization variational autoencoder adopts a hierarchical quantization structure and includes an encoder. and a decoder ; encoder : Divide the point intermediate feature maps output Encoded as Residual latent representation at each level ; Hierarchical quantization: Residual representation at each level Each uses its own independent codebook Perform vector quantization to obtain the quantized representation. and the corresponding codebook index The quantization process is represented as ; decoder : Able to receive any Quantization representation at each level The sum of the values ​​and the original feature map are then reconstructed from them. ;in, .

6. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 5, characterized in that, The training process of the variable bit rate vector quantization variational autoencoder is as follows: By selecting before transmission Layer codebook index This can be achieved. Different levels of compression ratio; Total loss function during training It is necessary to ensure good reconstruction quality at all compression levels, and the formula is as follows: ； In the formula, Before use Reconstruction loss when reconstructing layer information; It is the first The codebook commitment loss of the layer is used to induce the residual representation of the encoder output. Approximating its quantized value ; This indicates that the gradient operation has stopped; and These are weighting coefficients used to balance the various losses. This indicates a summation.

7. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 1, characterized in that, The dynamic decision-making module has a built-in status monitoring unit for collecting key input information required for decision-making; Network status : By performing routine network probes or querying network statistics provided by the underlying operating system, metrics for evaluating network quality, including bandwidth, can be obtained. and delay ; Task status : Receive the priority or type of the current task from the upper-layer application or system configuration. .

8. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 1, characterized in that, The decision-making logic of the dynamic decision-making module is implemented in the following way: Rule-based decision table: Through offline testing and analysis, a lookup table is pre-calculated and filled in for different network state ranges and task priority combinations; at runtime, the dynamic decision module looks up the table based on the current state to obtain the recommended configuration. .

9. The method for adjusting the dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy according to claim 1, characterized in that, The decision-making logic of the dynamic decision-making module is implemented in the following way: Learning-based decision-making models: Train a lightweight prediction model that learns the optimal action from the input state. The model is trained offline using data collected in various simulated or real-world environments, with the goal of learning a decision-making strategy that effectively balances accuracy and latency.

10. A dynamic adaptive edge-cloud collaborative computing segmentation point and compression strategy adjustment system, characterized in that, The system implements the method according to any one of claims 1 to 9.

Images