Content distillation extraction method based on edge computing system

Abstract

The invention belongs to the technical field of data content extraction methods, and particularly relates to a content distillation extraction method based on an edge computing system, comprising an edge device which is a hardware device close to data sources such as a sensor, a camera and intelligent hardware and is responsible for collecting original data from a physical world; the edge computing node is a computing unit deployed at the edge position of the network and is responsible for receiving data from a plurality of edge devices, executing more complex data processing, analysis, computing or reasoning tasks and feeding back results to the edge devices or uploading the results to the cloud according to needs; the cloud server is used for processing a content distillation task which cannot be completed on the edge computing node; by transferring data processing to the edge device, the data can be quickly preprocessed and analyzed, the requirement for transmitting the data to a cloud server in a traditional mode is reduced, delay is remarkably reduced, and dependence on network broadband is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of data content extraction methods, and more particularly to a content distillation extraction method based on an edge computing system. Background Art

[0002] With the rapid development of technologies such as the Internet of Things, smart homes, and intelligent transportation, more and more devices and systems need to perform data analysis and real-time decision-making processing. In many application scenarios, devices not only need to collect a large amount of data, but this data often needs to be quickly processed and analyzed on edge devices to ensure low latency, high efficiency, and real-time performance. Especially for applications such as video surveillance, speech recognition, and environmental monitoring, the real-time requirement for data processing is very high. Transmitting all data to the cloud for processing not only causes waste of network bandwidth, but may also result in untimely responses due to network latency, affecting system performance.

[0003] Therefore, edge computing, as an emerging computing architecture, is widely used to solve these problems. Edge computing reduces the dependence on the cloud by moving data processing to edge devices close to the data source, and can process, analyze data locally, and make immediate decisions. The core idea of edge computing is to extend computing resources from the cloud to the edge of the network, getting closer to the location where data is generated, reducing the demand for network bandwidth, and being able to respond to user needs faster.

[0004] However, edge devices themselves usually have limitations in computing resources and storage. Traditional large-scale deep learning models cannot run efficiently on edge devices due to their huge computing requirements and storage overhead. Therefore, how to compress deep learning models without losing too much performance and make them adaptable to the resource limitations of edge computing devices has become a key issue.

[0005] In this context, model compression technology has emerged. Model compression reduces the scale and complexity of the model so that it can run on edge devices while maintaining relatively high accuracy as much as possible. The techniques of model compression include weight pruning, quantization, knowledge distillation, etc. Among them, knowledge distillation, as an effective model compression method, has received extensive attention in recent years. Knowledge distillation extracts the knowledge from a large and complex deep learning model and transfers it to a lightweight model, which can greatly reduce the number of model parameters and computational amount and retain high performance.

[0006] In the edge computing environment, combining knowledge distillation with optimization techniques can significantly improve the speed and efficiency of model inference on edge devices. By extracting important features and information from deep learning models into lightweight models, edge computing devices can complete complex tasks, such as real-time video analysis, speech recognition, and sensor data processing, with less computational resources and storage overhead.

[0007] However, although content distillation has achieved good applications in large-scale data centers and cloud computing, applying it to resource-constrained edge devices still faces many challenges, including the efficiency of model distillation, hardware acceleration support for edge devices, real-time requirements, etc. Therefore, how to design an efficient content distillation extraction method based on edge computing systems to adapt to the resource limitations of edge devices and improve processing speed has become a hot issue in current technology research.

[0008] Currently, there have been many studies on the combination of edge computing and knowledge distillation. Applying knowledge distillation in edge computing can help solve the problem of how to run complex deep learning models on edge devices.

[0009] However, the existing studies still face the following challenges:

[0010] Computing and storage limitations: Despite various model compression techniques, edge devices still face limitations in computational resources and storage capacity. How to design a more efficient knowledge distillation method to ensure its efficient operation on edge devices remains a challenge.

[0011] Real-time requirements: Edge computing systems usually need to process a large amount of data in real time. Under such real-time requirements, how to design a low-latency content distillation method and ensure its fast inference is an urgent problem to be solved.

[0012] Hardware acceleration: Edge devices often come with dedicated hardware acceleration modules, such as FPGA or ASIC. How to combine hardware acceleration technology for model optimization to further improve inference speed and efficiency is also an important research direction currently.

[0013] Therefore, a content distillation extraction method based on edge computing systems is proposed to address the above issues. Summary of the Invention

[0014] (1) Technical problems to be solved

[0015] In view of the deficiencies of the prior art, the present invention provides a content distillation extraction method based on edge computing systems to solve the problems raised in the background art.

[0016] (2) Technical solutions

[0017] To achieve the above object, the present invention provides the following technical solutions: A content distillation extraction method based on an edge computing system includes:

[0018] A terminal device, including built-in sensors, cameras, and microphones, for collecting different types of data;

[0019] An edge device, responsible for collecting raw data from the physical world;

[0020] An edge computing node, which is a computing unit deployed at the edge of the network. It is responsible for receiving data from multiple edge devices, performing more complex data processing, analysis, calculation, or inference tasks, and feeding back the results to the edge devices or uploading them to the cloud as needed;

[0021] A content distillation algorithm, which extracts key features and knowledge from complex deep learning models and transfers this knowledge to lightweight models to ensure that the models can operate efficiently on resource-constrained edge devices;

[0022] A cloud server, used to process content distillation tasks that cannot be completed on edge computing nodes;

[0023] A data communication network, used to transmit requests, data, and distillation results between edge computing nodes, terminal devices, and the cloud server;

[0024] A management module, used to decide whether to execute the content distillation task on the edge computing node or forward it to the cloud server according to the real-time load situation and network status of the system;

[0025] Optimization and acceleration, when running the lightweight model on the edge device, combined with the hardware acceleration module, and adopting optimization means such as model compression, pruning, and quantization to reduce latency and computational resource consumption, and improve processing speed and real-time response ability.

[0026] Preferably, the content distillation algorithm includes,

[0027] Text distillation, TF-IDF is a common feature extraction method in text data processing, and content distillation is performed by calculating the importance of each word in the document;

[0028] Image feature distillation, using convolutional neural networks and deep generative adversarial networks. The convolutional neural network extracts the features of the image layer by layer and generates a high-dimensional representation of the image in the high-level convolutional layer to help extract the core information of the image. The deep generative adversarial network is used for image generation and content distillation. The generative adversarial network can recover higher-quality details from low-resolution images or remove irrelevant information from the images and retain important content;

[0029] Video and audio distillation, video summarization, content distillation by extracting key frames or detecting specific events in the video, audio feature extraction: in audio processing, extract from audio data.

[0030] Preferably, the content distillation algorithm is based on deep learning models, natural language processing techniques, image recognition algorithms, or video processing algorithms, and combines the characteristics of data content for data compression, extraction, and optimization.

[0031] Preferably, the deep learning model adopts a multi-layer neural network structure, simulating the processing mechanism of human brain neurons, which includes

[0032] Convolutional layer, extracting spatial features;

[0033] Pooling layer, for dimensionality reduction;

[0034] Fully connected layer, outputting decision results, and extracting features and information layer by layer for subsequent distillation extraction.

[0035] Preferably, the edge computing node includes multiple computing units, which can dynamically adjust computing resources for task allocation according to the complexity and resource requirements of the content distillation task.

[0036] Preferably, the cloud server includes

[0037] Cloud collection module, receiving preprocessed data from edge nodes or actively collecting data through API interfaces;

[0038] Cloud analysis engine, using deep learning models to perform operations such as classification, behavior recognition, and speech emotion analysis on audio-visual data;

[0039] Cloud event impact module, the cloud can analyze data in real time and trigger relevant operations through an event-driven architecture;

[0040] Cloud classification and storage module, compressing and storing content and performing intelligent classification according to data types and access frequencies, optimizing storage space and reading efficiency.

[0041] Preferably, the distillation process is based on a multi-stage strategy. First, complete offline distillation training on the cloud server, and then deploy the distillation intermediate results to edge devices to accelerate model fine-tuning, improve distillation efficiency, and reduce computational load.

[0042] Preferably, the edge device is integrated with a hardware acceleration unit, including but not limited to GPU, NPU, FPGA, or ASIC, for accelerating the inference and distillation calculations of deep learning models and improving real-time processing capabilities.

[0043] A content distillation extraction method based on an edge computing system, the steps include:

[0044] S1: Data acquisition. The terminal device collects various types of data from different sources through built-in sensors, cameras, microphones and other hardware devices;

[0045] S2: Data preprocessing. After receiving the data from the terminal device, the edge device first preprocesses the data. The edge device may use a hardware acceleration unit to accelerate the data preprocessing to ensure the data processing speed and efficiency;

[0046] S3: Execution of lightweight content distillation algorithm. After receiving the preprocessed data from multiple edge devices, the edge computing node performs more complex content distillation tasks;

[0047] S4: Content optimization and acceleration. The lightweight model running on the edge device further optimizes the computing speed and response ability by combining with a hardware acceleration module. At the same time, the model uses technologies such as compression, pruning, and quantization to reduce the consumption of computing resources and improve the real-time response ability;

[0048] S5: Feedback and storage of content distillation results. The edge computing node sends the processed distillation results back to the edge device for further decision-making or control operations. If more complex processing is required, the edge computing node transmits the results to the cloud server;

[0049] S6: Cloud processing. If the edge computing node cannot complete some complex tasks, it forwards the tasks to the cloud server for further analysis and processing;

[0050] S7: Multi-stage strategy and distillation optimization. Multi-stage distillation: After the cloud server completes the offline distillation training, it deploys the intermediate distillation results to the edge computing node, thereby accelerating the model fine-tuning on the edge device. The offline distillation training in the cloud will be optimized through a deep learning framework, enabling the final model to adapt to the resource limitations of the edge device and maintain high efficiency during actual operation.

[0051] Beneficial effects

[0052] Compared with the prior art, the present invention provides a content distillation extraction method based on an edge computing system, having the following beneficial effects:

[0053] 1. In this invention, by moving the data processing to the edge device, it can quickly preprocess and analyze the data, reducing the need to transmit data to the cloud server in the traditional way, significantly reducing the latency, and reducing the dependence on network bandwidth.

[0054] 2. In this invention, by extracting and migrating the key information in a complex deep learning model to a lightweight model, the edge device can operate efficiently, reducing the number of model parameters and computational requirements while ensuring high performance.

[0055] 3. In this invention, the inference process of the deep learning model is optimized by combining with a hardware acceleration module. The hardware acceleration unit can effectively reduce the inference time and improve the real-time response ability. By combining with the hardware acceleration technology, the edge device can efficiently run the lightweight deep learning model, thus enhancing the overall performance and processing efficiency of the edge computing system.

[0056] 4. In this invention, a multi-stage distillation method is adopted. First, offline distillation training is completed in the cloud, and then the intermediate distillation results are deployed to the edge device for fine-tuning, effectively improving the efficiency of the distillation process and reducing the computational burden on the edge device. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The drawings described herein are used to provide a further understanding of the present invention and form a part of this application. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0058] Figure 1 is the system flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] Specific embodiments are given below.

[0061] Embodiment 1

[0062] As Figure 1 shown, a content distillation and extraction method based on an edge computing system. The edge node collects video data in real time through a video surveillance camera. The video data is first preprocessed by a sensor acquisition module. The preprocessing includes operations such as image scaling, noise filtering, and motion detection to ensure that the noise is minimized and unnecessary data volume is reduced before the data is transmitted to the edge computing device;

[0063] The edge computing device further performs a preliminary analysis on the preprocessed video data, including moving target detection, region annotation, etc., reducing the burden of data transmission to the cloud server;

[0064] To optimize and adapt a deep learning model for edge devices, the present invention uses content distillation. In the cloud, a large-scale deep learning model is used to perform offline training on video data to extract important features and knowledge;

[0065] The trained large model is transferred to a lightweight model through distillation technology, reducing the number of model parameters and the computational amount to ensure that the lightweight model can operate efficiently on edge devices;

[0066] After optimization, the lightweight model can process video data in real time on edge devices and complete real-time object recognition and event detection;

[0067] Hardware acceleration modules such as neural network processing units and GPUs are integrated on edge devices to accelerate the inference calculation in the video analysis process;

[0068] Combined with hardware acceleration technology, the lightweight model can operate efficiently, significantly reducing the inference time and improving the real-time response ability of the system;

[0069] In addition, the optimization of the model further reduces the computational overhead of the model through techniques such as pruning and quantization to ensure that edge devices can operate efficiently under resource constraints;

[0070] On edge devices, the preprocessed and preliminarily analyzed data will be stored in the local cache module and uploaded to the cloud server for in-depth analysis and archiving regularly or on demand as needed;

[0071] The cloud server receives data from edge nodes and further analyzes the video content through the cloud analysis engine to perform more complex tasks such as behavior recognition and event impact analysis;

[0072] The cloud server also triggers alarms or other corresponding operations through an event-driven architecture based on the analysis results, such as automatically locking the monitoring area or sending notifications;

[0073] Due to the adoption of a multi-stage distillation strategy, first, offline distillation training is carried out on the cloud server, and then the intermediate results of distillation are deployed to edge devices. The edge devices accelerate the model update and adaptive optimization through fine-tuning, thereby reducing the computational burden on edge devices while improving the processing efficiency;

[0074] This system significantly improves the efficiency and response ability of real-time video analysis and meets the requirements of high efficiency and real-time performance in video surveillance.

[0075] Embodiment 2

[0076] The edge device collects audio data in the environment through a microphone. The collected audio signal will go through a preprocessing module for noise cancellation and format standardization to reduce the impact of environmental noise on the accuracy of speech recognition and ensure that the data format meets the requirements of subsequent analysis;

[0077] Through the distillation technique, the cloud server first uses a complex deep learning model to train the audio data and extracts the key information and features of the speech;

[0078] Then, the cloud server converts the trained complex model into a lightweight model, which contains most of the key information and knowledge required for speech recognition, reducing the computational overhead;

[0079] After being fine-tuned, this lightweight model can operate efficiently on the edge device, process audio data in real time, recognize the speech content and perform corresponding operations;

[0080] The edge device integrates a hardware acceleration module to improve the inference speed of the speech recognition task;

[0081] By combining hardware acceleration and model optimization techniques, the speech recognition system can achieve efficient speech recognition with extremely low latency, ensuring the real-time response ability of the system.

[0082] The audio data processed by the edge device can perform simple operations locally, such as speech command recognition and sentiment analysis. If the system cannot handle complex tasks, the edge device will upload the audio data to the cloud server for deeper analysis;

[0083] The cloud server performs complex speech sentiment analysis and keyword extraction operations based on the audio data and triggers relevant operations according to the analysis results;

[0084] The multi-stage distillation method ensures the offline training of the speech recognition model on the cloud server and the rapid fine-tuning on the edge device, enabling speech recognition to be carried out efficiently with low latency;

[0085] The edge computing device ensures that the speech recognition system can operate efficiently and in real time on resource-constrained devices through hardware acceleration, model compression and optimization strategies.

[0086] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed.

Claims

1. An edge computing-based system, characterized in that, including: A terminal device, including built-in sensors, cameras, and microphones, for collecting different types of data; An edge device, responsible for collecting raw data from the physical world; An edge computing node, which is a computing unit deployed at the edge of the network. It is responsible for receiving data from multiple edge devices, performing more complex data processing, analysis, calculation, or inference tasks, and feeding back the results to the edge devices or uploading them to the cloud as needed; A content distillation algorithm, which extracts key features and knowledge from complex deep learning models and transfers this knowledge to lightweight models to ensure that the models can operate efficiently on resource-constrained edge devices; A cloud server, used to process content distillation tasks that cannot be completed on edge computing nodes; A data communication network, used to transmit requests, data, and distillation results between edge computing nodes, terminal devices, and the cloud server; A management module, used to decide whether to execute the content distillation task on the edge computing node or forward it to the cloud server according to the real-time load situation and network status of the system; Optimization and acceleration. When running a lightweight model on an edge device, combine with a hardware acceleration module and adopt optimization means such as model compression, pruning, and quantization to reduce latency and computational resource consumption, and improve processing speed and real-time response ability.

2. The edge computing system according to claim 1, wherein The content distillation algorithm includes, Text distillation. TF-IDF is a common feature extraction method in text data processing, and content distillation is performed by calculating the importance of each word in the document; Image feature distillation. Use a convolutional neural network and a deep generative adversarial network. The convolutional neural network extracts the features of the image layer by layer and generates a high-dimensional representation of the image in the high-level convolutional layer to help extract the core information of the image. The deep generative adversarial network is used for image generation and content distillation. The generative adversarial network can recover higher-quality details from low-resolution images or remove irrelevant information from the image and retain important content; Video and audio distillation. Video summarization, content distillation is performed by extracting key frames or detecting specific events in the video. Audio feature extraction: In audio processing, extract from audio data.

3. The edge computing system according to claim 1, characterized in that, The content distillation algorithm is based on deep learning models, natural language processing technologies, image recognition algorithms, or video processing algorithms, and combines the characteristics of data content to perform data compression, extraction, and optimization.

4. The edge computing system according to claim 3, wherein The deep learning model adopts a multi-layer neural network structure, simulating the processing mechanism of human brain neurons, and it includes, A convolutional layer, extracting spatial features; A pooling layer, for dimensionality reduction; A fully connected layer, outputting decision results, and extracting features and information layer by layer for subsequent distillation extraction.

5. A kind of edge computing system according to claim 1, characterized in that, The edge computing node includes multiple computing units, which can dynamically adjust computing resources for task allocation according to the complexity and resource requirements of the content distillation task.

6. The edge computing system according to claim 1, wherein, The cloud server includes, A cloud acquisition module, receiving preprocessed data from edge nodes or actively acquiring data through an API interface; A cloud analysis engine, using a deep learning model to perform operations such as classification, behavior recognition, and speech emotion analysis on audio and video data; Cloud Event Impact Module. The cloud can analyze data in real time and trigger relevant operations through an event-driven architecture; Cloud Classification and Storage Module. Compress and store content and perform intelligent classification based on data type and access frequency to optimize storage space and reading efficiency.

7. A kind of edge computing system according to claim 1, characterized in that, The distillation process is based on a multi-stage strategy. First, perform offline distillation training on the cloud server, and then deploy the intermediate distillation results to edge devices to accelerate model fine-tuning, improve distillation efficiency, and reduce the computational load.

8. A kind of edge computing system according to claim 7, characterized in that, The edge device is integrated with a hardware acceleration unit, including but not limited to GPU, NPU, FPGA, or ASIC, for accelerating the inference and distillation calculations of deep learning models and enhancing real-time processing capabilities.

9. According to any one of claims 1-8, a content distillation extraction method based on an edge computing system, characterized in that the steps Including: S1: Data collection. The terminal device collects various types of data from different sources through built-in hardware devices such as sensors, cameras, and microphones; S2: Data preprocessing. After receiving the data from the terminal device, the edge device first preprocesses the data. The edge device may use the hardware acceleration unit to accelerate data preprocessing to ensure data processing speed and efficiency; S3: Execution of lightweight content distillation algorithm. After receiving the preprocessed data from multiple edge devices, the edge computing node performs more complex content distillation tasks; S4: Content optimization and acceleration. The lightweight model running on the edge device further optimizes the computing speed and response ability by combining the hardware acceleration module. At the same time, the model uses technologies such as compression, pruning, and quantization to reduce computational resource consumption and improve real-time response ability; S5: Feedback and storage of content distillation results. The edge computing node sends the processed distillation results back to the edge device for further decision-making or control operations. If more complex processing is required, the edge computing node transmits the results to the cloud server; S6: Cloud processing. If the edge computing node cannot complete some complex tasks, it forwards the tasks to the cloud server for further analysis and processing; S7: Multi-stage strategy and distillation optimization. Multi-stage distillation: After completing offline distillation training on the cloud server, deploy the intermediate distillation results to the edge computing node to accelerate the model fine-tuning on the edge device. The offline distillation training on the cloud will be optimized through a deep learning framework, enabling the final model to adapt to the resource limitations of the edge device and maintain high efficiency during actual operation.

Images