AI algorithm process and service all-in-one machine based on edge computing

Abstract

The invention relates to the related field of artificial intelligence (AI) and edge computing, in particular to an AI algorithm process and service all-in-one machine based on edge computing, and efficient operation and dynamic adaptability of a system are achieved through cooperative work and intelligent resource management of edge equipment and a cloud end. By deploying the real-time resource monitoring module on the edge device, the computing resource use condition of the device can be monitored in real time, resource allocation is dynamically adjusted according to the current task requirement, and the working efficiency and the energy utilization rate of the device are improved. And the cloud end performs regular optimization and issues the updated model to the edge device, so that dynamic updating and optimization of the model are realized, and the system is ensured to maintain efficient and stable operation in a complex and changeable environment. Through preliminary reasoning of the edge device and auxiliary processing of the cloud end, the data transmission amount and the cloud computing burden are reduced, and meanwhile the robustness, maintainability and expandability of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to artificial intelligence (AI) and edge computing related fields, and specifically to an AI algorithm process based on edge computing and an all-in-one service machine. Background Art

[0002] With the rapid development of the Internet of Things (IoT), a large amount of data is generated from various edge devices and transmitted to the cloud for processing. However, due to issues such as network bandwidth, latency, and data privacy, the traditional method of transmitting all data to the cloud for AI processing faces many challenges. Edge computing reduces the amount of data transmitted, reduces latency, and enhances data privacy protection by sinking computing resources to edge nodes close to the source of data generation. However, most of the current edge computing AI algorithms cannot be effectively deployed and run on edge devices, and are difficult to flexibly adjust in different application scenarios. Summary of the invention

[0003] The purpose of the present invention is to provide an AI algorithm process and a service all-in-one machine based on edge computing to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solution: an AI algorithm process based on edge computing, comprising the following steps:

[0005] Step S1, data collection and preprocessing: Data collection is performed on the edge device, and the raw data is cleaned, normalized, and feature extracted through the data preprocessing module to reduce data redundancy and computational complexity;

[0006] Step S2, model segmentation and distribution: The AI ​​model is segmented into an edge-side model and a cloud-side model. The edge-side model is a lightweight model responsible for low-complexity reasoning tasks, while the cloud-side model is responsible for high-complexity reasoning and training tasks. The weight distribution of the edge and cloud models is dynamically adjusted according to the specific application scenarios.

[0007] Step S3, edge reasoning and decision-making: Run the edge model on the edge device for real-time reasoning to obtain preliminary reasoning results. By setting a threshold, when the confidence level of the edge reasoning result is lower than the threshold, upload the data to the cloud for further reasoning by the cloud model.

[0008] Step S4, edge device and cloud collaboration: through the edge device and cloud collaboration mechanism, the model on the edge device is updated and optimized, and the cloud model periodically sends the updated model parameters to the edge device to adapt to the dynamically changing environment and needs;

[0009] Step S5, local storage and privacy protection: For sensitive data, it is preferred to store and process it locally on the edge device to avoid uploading the data to the cloud to enhance data privacy protection. At the same time, encryption technology is used to encrypt the data in transit to further improve security;

[0010] Step S6, intelligent resource management and optimization: monitor the computing resource usage of edge devices in real time through the resource monitoring module, and dynamically adjust the allocation of computing resources according to the needs of the current task to improve the working efficiency and energy utilization of edge devices.

[0011] Preferably, the specific steps of data collection and preprocessing in step S1 are as follows:

[0012] Step S11, determine the data source according to the specific application scenario, and use edge devices to collect video and audio streams and sensor continuous data sources in real time, and use batch collection for environmental data sources to obtain collected data;

[0013] Step S12: For the above-mentioned collected data, filtering technology is used to remove noise in the data, wherein the image data is processed using a denoising algorithm; for the missing data points of the collected data, an interpolation method is used to supplement them, and a statistical method is used to identify and process outliers; then the data of different dimensions are converted to the same dimension to obtain preprocessed data;

[0014] Step S13: extract basic features from the preprocessed data, and use a convolutional neural network (CNN) or a spectrum analysis algorithm to extract advanced features.

[0015] Preferably, the specific steps of model segmentation and distribution in step S2 are as follows:

[0016] Step S21, model segmentation strategy determination: Determine the model segmentation strategy based on the complexity of the AI ​​model, task requirements, computing power of the edge device, and network bandwidth factors, where the initial segmentation point is selected at the middle layer of the AI ​​model, the early layers are deployed on the edge device, and the later layers are deployed in the cloud;

[0017] Step S22, model layer division: for deep learning models, the segmentation method is based on the network layer; for traditional machine learning models, segmentation is based on the complexity of feature processing;

[0018] Step S23, segmentation point selection algorithm: select the segmentation point through the optimization function. Assuming that the model has N layers in total and the computing power that the edge device can carry is C e , the cloud computing power is C c , then the optimization problem can be expressed as:

[0019]

[0020] Among them, k is the split point, C e (i) is the computational cost of the i-th layer on the edge device, C c (i) is the computational cost of the i-th layer in the cloud. The goal is to find a split point k that optimally allocates computing resources on the edge and in the cloud.

[0021] Step S24, model distribution strategy design: After the AI ​​model segmentation is completed, the edge-side model and the cloud-side model are distributed through network connection, where the model of the edge device is lightweight, and TensorFlow Lite or ONNX tools are used to compress and optimize the model to reduce storage space and running time;

[0022] Step S25, weight allocation and dynamic adjustment: In specific application scenarios, the edge and cloud weight allocation of the model is dynamically adjusted. Specifically, when the edge device load is light and the network conditions are good, more calculations are completed on the edge side; when the edge device load is heavy or the network delay is high, more cloud computing is used. The dynamic adjustment is controlled by the weight allocation coefficient. Assuming that the calculation weights of the edge device and the cloud device are a and β respectively, the dynamic adjustment problem can be expressed as:

[0023] a+β=1

[0024] Among them, when the computing load of the edge device increases: a↓, β↑; when the network conditions deteriorate: a↑, β↓.

[0025] Preferably, the specific processing steps of edge reasoning and decision making in step S3 are as follows:

[0026] Step S31: When the edge device is started, the allocated edge side model is loaded into the memory and initialized;

[0027] Step S32: input the collected data processed in step S1 into the edge side model, use the lightweight model to perform the reasoning task on the edge device, and obtain preliminary reasoning results;

[0028] Step S33: predict the confidence of the obtained inference result. The specific confidence evaluation content is as follows:

[0029] Confidence calculation: For classification tasks, the confidence is the probability value calculated by the Softmax function; for regression tasks, the confidence is measured by the difference between the model output value and the preset threshold, where the confidence of the model output is set to P i , where i is the category index, then:

[0030]

[0031] Among them, z i is the output score of the model on the i-th category, and N is the total number of categories;

[0032] Threshold setting: According to the application scenario and tolerance, a confidence threshold T is set. When the output confidence P of the edge model i When it is lower than T, the model is considered uncertain about the current inference result and data needs to be uploaded to the cloud for further processing. The threshold T is determined through offline testing and tuning to balance the efficiency and accuracy of edge inference.

[0033] Confidence evaluation: After each inference, the output confidence of the edge model is calculated and compared with the set threshold T; if the highest confidence The edge device directly makes decisions and takes corresponding actions; if the highest confidence The inference result is considered unreliable and the next step is entered;

[0034] Step S34: When the edge device detects that the inference confidence is lower than the set threshold T, the data upload condition is triggered, and the current input data or intermediate feature data is uploaded to the cloud. Before uploading the data, data compression can be performed to reduce the amount of transmitted data. At the same time, to ensure data security, encryption technology should be used when uploading data to protect the privacy of the data during transmission;

[0035] Step S35: After the cloud receives the data, it performs more complex reasoning through the preset AI model. After the cloud reasoning is completed, the final reasoning result is returned to the edge device. The edge device makes a final decision based on the reasoning result of the cloud and performs corresponding operations.

[0036] Preferably, the specific working steps of the collaborative work between the edge device and the cloud in step S4 are as follows:

[0037] Step S41: Model update and optimization requirements analysis

[0038] a. Detection of demand changes: The cloud collects and analyzes the operating data, user feedback, and environmental information of edge devices to detect whether the model needs to be updated;

[0039] b. Model performance monitoring: Real-time monitoring of the performance of models running on edge devices, including accuracy, response time, and resource usage, including CPU, memory, and power consumption. The cloud regularly evaluates these performance data to determine whether the current model needs to be updated or optimized.

[0040] c. Model update trigger conditions: When the model performance of the edge device drops below the set threshold, or the environmental changes exceed the model adaptation range, the model update requirement is triggered;

[0041] Step S42: Cloud model training and optimization

[0042] a. Data collection and training set generation: The cloud collects data from multiple edge devices, including raw data, inference results of edge devices, and misjudgment cases. Through data cleaning and annotation, a new training set is generated, especially focusing on data with misjudgments and low confidence of edge devices, to ensure the diversity and representativeness of the training set;

[0043] b. Model training and optimization: The cloud uses the latest collected data to retrain the model;

[0044] c. Model evaluation and verification: After training is completed, the cloud performs rigorous testing and verification on the new model to ensure its adaptability and robustness in different environments;

[0045] Step S43: Model parameter distribution and deployment

[0046] a. Model compression and optimization: Before sending the model to the edge device, the cloud compresses and optimizes the model, including model quantization, which converts floating-point numbers into fixed-point numbers, and pruning, which removes unimportant neurons or weights.

[0047] b. Model delivery mechanism: The cloud sends the optimized model parameters to the edge device through a secure network channel;

[0048] c. Edge device model reception and verification: After receiving the new model parameters, the edge device first verifies the integrity and security of the model, specifically checking the hash value and digital signature; after the verification is passed, the edge device loads the new model into the memory and performs initialization operations to prepare to replace the old model;

[0049] Step S44: Edge device feedback and collaborative optimization

[0050] a. Real-time feedback mechanism: During operation, edge devices feed back the model’s reasoning results, performance indicators, and abnormal situation data to the cloud in real time. Based on these feedbacks, the cloud adjusts the model’s parameters or structure in a timely manner and performs remote optimization.

[0051] b. Collaborative optimization strategy: The cloud integrates data from multiple edge devices to perform global optimization. The cloud pushes different optimization strategies or model versions to meet the special needs of different regions or devices. Through reinforcement learning or federated learning, the cloud learns from the operating data of multiple edge devices and extracts common optimization solutions to further improve the performance of the overall system.

[0052] Preferably, the resource monitoring module in step S6 is a key component on the edge device, which is used to monitor the device's CPU, GPU, memory, storage, network bandwidth, battery power resource usage and data in real time, wherein the resource monitoring module performs real-time analysis on the collected data, identifies changing trends in resource usage, including periodic fluctuations in CPU load and gradual increases in memory occupancy, and predicts changes in resource demand in the future through historical data analysis, and makes optimization decisions in advance, i.e., introduces a priority scheduling algorithm to allocate appropriate computing resources according to task priority and resource requirements.

[0053] Preferably, a service-in-one machine includes the AI ​​algorithm process based on edge computing.

[0054] Compared with the prior art, the present invention has the following beneficial effects: the present invention performs real-time reasoning tasks on edge devices, uses lightweight models to quickly obtain preliminary results, and can make decisions directly in most cases, thereby reducing data transmission delays and improving system response speed; by setting a confidence threshold, when the reasoning result is uncertain, it is automatically uploaded to the cloud for further processing, ensuring the accuracy and reliability of the reasoning result; the cloud of the present invention can periodically optimize the model parameters and send them to the edge device according to environmental changes and task requirements, ensuring that the device always runs the latest and most suitable model for the current scenario; through intelligent resource management and optimization strategies, the computing resource usage of the edge device is monitored in real time, and resource allocation is dynamically adjusted to ensure that high-priority tasks obtain sufficient resource support, thereby improving the overall efficiency of the system; by performing preliminary reasoning on the edge device and uploading data to the cloud only when needed, the present invention significantly reduces the amount of data uploaded and the computing pressure of the cloud, and reduces bandwidth usage and cloud computing costs; compression and encryption processing before data upload not only protects data privacy, but also further reduces the amount of transmitted data and improves transmission efficiency; the cloud integrates data from multiple edge devices and performs global optimization, which can support large-scale edge computing networks and improve the scalability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the workflow of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] See also Figure 1 , the present invention provides a technical solution: an AI algorithm process based on edge computing, comprising the following steps:

[0058] Step S1, data collection and preprocessing: Data collection is performed on the edge device, and the raw data is cleaned, normalized, and feature extracted through the data preprocessing module to reduce data redundancy and computational complexity;

[0059] Step S2, model segmentation and distribution: The AI ​​model is segmented into an edge-side model and a cloud-side model. The edge-side model is a lightweight model responsible for low-complexity reasoning tasks, while the cloud-side model is responsible for high-complexity reasoning and training tasks. The weight distribution of the edge and cloud models is dynamically adjusted according to the specific application scenarios.

[0060] Step S3, edge reasoning and decision-making: Run the edge model on the edge device for real-time reasoning to obtain preliminary reasoning results. By setting a threshold, when the confidence level of the edge reasoning result is lower than the threshold, upload the data to the cloud for further reasoning by the cloud model.

[0061] Step S4, edge device and cloud collaboration: through the edge device and cloud collaboration mechanism, the model on the edge device is updated and optimized, and the cloud model periodically sends the updated model parameters to the edge device to adapt to the dynamically changing environment and needs;

[0062] Step S5, local storage and privacy protection: For sensitive data, it is preferred to store and process it locally on the edge device to avoid uploading the data to the cloud to enhance data privacy protection. At the same time, encryption technology is used to encrypt the data in transit to further improve security;

[0063] Step S6, intelligent resource management and optimization: monitor the computing resource usage of edge devices in real time through the resource monitoring module, and dynamically adjust the allocation of computing resources according to the needs of the current task to improve the working efficiency and energy utilization of edge devices.

[0064] Furthermore, the specific steps of data collection and preprocessing in step S1 are as follows:

[0065] Step S11, determine the data source according to the specific application scenario, and use edge devices to collect video and audio streams and sensor continuous data sources in real time, and use batch collection for environmental data sources to obtain collected data;

[0066] Step S12: For the above-mentioned collected data, filtering technology is used to remove noise in the data, wherein the image data is processed using a denoising algorithm; for the missing data points of the collected data, an interpolation method is used to supplement them, and a statistical method is used to identify and process outliers; then the data of different dimensions are converted to the same dimension to obtain preprocessed data;

[0067] Step S13: extract basic features from the preprocessed data, and use a convolutional neural network (CNN) or a spectrum analysis algorithm to extract advanced features.

[0068] Furthermore, the specific steps of model segmentation and distribution in step S2 are as follows:

[0069] Step S21, model segmentation strategy determination: Determine the model segmentation strategy based on the complexity of the AI ​​model, task requirements, computing power of the edge device, and network bandwidth factors, where the initial segmentation point is selected at the middle layer of the AI ​​model, the early layers are deployed on the edge device, and the later layers are deployed in the cloud;

[0070] Step S22, model layer division: for deep learning models, the segmentation method is based on the network layer; for traditional machine learning models, segmentation is based on the complexity of feature processing;

[0071] Step S23, segmentation point selection algorithm: select the segmentation point through the optimization function. Assuming that the model has N layers in total and the computing power that the edge device can carry is C e , the cloud computing power is C c , then the optimization problem can be expressed as:

[0072]

[0073] Among them, k is the split point, C e (i) is the computational cost of the i-th layer on the edge device, C c (i) is the computational cost of the i-th layer in the cloud. The goal is to find a split point k that optimally allocates computing resources on the edge and in the cloud.

[0074] Step S24, model distribution strategy design: After the AI ​​model segmentation is completed, the edge-side model and the cloud-side model are distributed through network connection, where the model of the edge device is lightweight, and TensorFlow Lite or ONNX tools are used to compress and optimize the model to reduce storage space and running time;

[0075] Step S25, weight allocation and dynamic adjustment: In specific application scenarios, the edge and cloud weight allocation of the model is dynamically adjusted. Specifically, when the edge device load is light and the network conditions are good, more calculations are completed on the edge side; when the edge device load is heavy or the network delay is high, more cloud computing is used. The dynamic adjustment is controlled by the weight allocation coefficient. Assuming that the calculation weights of the edge device and the cloud device are a and β respectively, the dynamic adjustment problem can be expressed as:

[0076] a+β=1

[0077] Among them, when the computing load of the edge device increases: a↓, β↑; when the network conditions deteriorate: a↑, β↓.

[0078] Furthermore, the specific processing steps of edge reasoning and decision making in step S3 are as follows:

[0079] Step S31: When the edge device is started, the allocated edge side model is loaded into the memory and initialized;

[0080] Step S32: input the collected data processed in step S1 into the edge side model, use the lightweight model to perform the reasoning task on the edge device, and obtain preliminary reasoning results;

[0081] Step S33: predict the confidence of the obtained inference result. The specific confidence evaluation content is as follows:

[0082] Confidence calculation: For classification tasks, the confidence is the probability value calculated by the Softmax function; for regression tasks, the confidence is measured by the difference between the model output value and the preset threshold, where the confidence of the model output is set to P i , where i is the category index, then:

[0083]

[0084] Among them, z i is the output score of the model on the i-th category, and N is the total number of categories;

[0085] Threshold setting: According to the application scenario and tolerance, a confidence threshold T is set. When the output confidence P of the edge model i When it is lower than T, the model is considered uncertain about the current inference result and data needs to be uploaded to the cloud for further processing. The threshold T is determined through offline testing and tuning to balance the efficiency and accuracy of edge inference.

[0086] Confidence evaluation: After each inference, the output confidence of the edge model is calculated and compared with the set threshold T; if the highest confidence The edge device directly makes decisions and takes corresponding actions; if the highest confidence The inference result is considered unreliable and the next step is entered;

[0087] Step S34: When the edge device detects that the inference confidence is lower than the set threshold T, the data upload condition is triggered, and the current input data or intermediate feature data is uploaded to the cloud. Before uploading the data, data compression can be performed to reduce the amount of transmitted data. At the same time, to ensure data security, encryption technology should be used when uploading data to protect the privacy of the data during transmission;

[0088] Step S35: After the cloud receives the data, it performs more complex reasoning through the preset AI model. After the cloud reasoning is completed, the final reasoning result is returned to the edge device. The edge device makes a final decision based on the reasoning result of the cloud and performs corresponding operations.

[0089] Furthermore, the specific working steps of the edge device and the cloud in step S4 are as follows:

[0090] Step S41: Model update and optimization requirements analysis

[0091] a. Detection of demand changes: The cloud collects and analyzes the operating data, user feedback, and environmental information of edge devices to detect whether the model needs to be updated;

[0092] b. Model performance monitoring: Real-time monitoring of the performance of models running on edge devices, including accuracy, response time, and resource usage, including CPU, memory, and power consumption. The cloud regularly evaluates these performance data to determine whether the current model needs to be updated or optimized.

[0093] c. Model update trigger conditions: When the model performance of the edge device drops below the set threshold, or the environmental changes exceed the model adaptation range, the model update requirement is triggered;

[0094] Step S42: Cloud model training and optimization

[0095] a. Data collection and training set generation: The cloud collects data from multiple edge devices, including raw data, inference results of edge devices, and misjudgment cases. Through data cleaning and annotation, a new training set is generated, especially focusing on data with misjudgments and low confidence of edge devices, to ensure the diversity and representativeness of the training set;

[0096] b. Model training and optimization: The cloud uses the latest collected data to retrain the model;

[0097] c. Model evaluation and verification: After training is completed, the cloud performs rigorous testing and verification on the new model to ensure its adaptability and robustness in different environments;

[0098] Step S43: Model parameter distribution and deployment

[0099] a. Model compression and optimization: Before sending the model to the edge device, the cloud compresses and optimizes the model, including model quantization, which converts floating-point numbers into fixed-point numbers, and pruning, which removes unimportant neurons or weights.

[0100] b. Model delivery mechanism: The cloud sends the optimized model parameters to the edge device through a secure network channel;

[0101] c. Edge device model reception and verification: After receiving the new model parameters, the edge device first verifies the integrity and security of the model, specifically checking the hash value and digital signature; after the verification is passed, the edge device loads the new model into the memory and performs initialization operations to prepare to replace the old model;

[0102] Step S44: Edge device feedback and collaborative optimization

[0103] a. Real-time feedback mechanism: During operation, edge devices feed back the model’s reasoning results, performance indicators, and abnormal situation data to the cloud in real time. Based on these feedbacks, the cloud adjusts the model’s parameters or structure in a timely manner and performs remote optimization.

[0104] b. Collaborative optimization strategy: The cloud integrates data from multiple edge devices to perform global optimization. The cloud pushes different optimization strategies or model versions to meet the special needs of different regions or devices. Through reinforcement learning or federated learning, the cloud learns from the operating data of multiple edge devices and extracts common optimization solutions to further improve the performance of the overall system.

[0105] Furthermore, the resource monitoring module in step S6 is a key component on the edge device, which is used to monitor the device's CPU, GPU, memory, storage, network bandwidth, battery power resource usage and data in real time. The resource monitoring module analyzes the collected data in real time to identify the changing trend of resource usage, including the periodic fluctuation of CPU load and the gradual increase of memory usage. It also predicts the changes in resource demand in the future through historical data analysis, and makes optimization decisions in advance, that is, introduces a priority scheduling algorithm to allocate appropriate computing resources according to the priority and resource requirements of the task.

[0106] Furthermore, a service all-in-one machine includes an AI algorithm process based on edge computing.

[0107] The present invention performs real-time reasoning tasks on edge devices and uses lightweight models to quickly obtain preliminary results. In most cases, decisions can be made directly, thereby reducing data transmission delays and improving system response speed. By setting a confidence threshold, when the reasoning result is uncertain, it is automatically uploaded to the cloud for further processing, ensuring the accuracy and reliability of the reasoning result. The cloud of the present invention can periodically optimize model parameters and send them to edge devices according to environmental changes and task requirements to ensure that the device always runs the latest and most suitable model for the current scenario. Through intelligent resource management and optimization strategies, the computing resource usage of edge devices is monitored in real time, and resource allocation is dynamically adjusted to ensure that high-priority tasks obtain sufficient resource support, thereby improving the overall efficiency of the system. By performing preliminary reasoning on edge devices and uploading data to the cloud only when needed, the present invention significantly reduces the amount of data uploaded and the computing pressure of the cloud, and reduces bandwidth usage and cloud computing costs. The compression and encryption processing before data upload not only protects data privacy, but also further reduces the amount of transmitted data and improves transmission efficiency. The cloud integrates data from multiple edge devices and performs global optimization, which can support large-scale edge computing networks and improve the scalability of the system.

[0108] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI algorithm process based on edge computing, characterized in that: The following steps are involved: Step S1, data collection and preprocessing: Data collection is performed on the edge device, and the raw data is cleaned, normalized, and feature extracted through the data preprocessing module to reduce data redundancy and computational complexity; Step S2, model segmentation and distribution: The AI ​​model is segmented into an edge-side model and a cloud-side model. The edge-side model is a lightweight model responsible for low-complexity reasoning tasks, while the cloud-side model is responsible for high-complexity reasoning and training tasks. The weight distribution of the edge and cloud models is dynamically adjusted according to the specific application scenarios. Step S3, edge reasoning and decision-making: Run the edge model on the edge device for real-time reasoning to obtain preliminary reasoning results. By setting a threshold, when the confidence level of the edge reasoning result is lower than the threshold, upload the data to the cloud for further reasoning by the cloud model. Step S4, edge device and cloud collaboration: through the edge device and cloud collaboration mechanism, the model on the edge device is updated and optimized, and the cloud model periodically sends the updated model parameters to the edge device to adapt to the dynamically changing environment and needs; Step S5, local storage and privacy protection: For sensitive data, it is preferred to store and process it locally on the edge device to avoid uploading the data to the cloud to enhance data privacy protection. At the same time, encryption technology is used to encrypt the data in transit to further improve security; Step S6, intelligent resource management and optimization: monitor the computing resource usage of edge devices in real time through the resource monitoring module, and dynamically adjust the allocation of computing resources according to the needs of the current task to improve the working efficiency and energy utilization of edge devices.

2. According to the AI ​​algorithm process based on edge computing according to claim 1, it is characterized in that: The specific steps of data collection and preprocessing in step S1 are as follows: Step S11, determine the data source according to the specific application scenario, and use edge devices to collect video and audio streams and sensor continuous data sources in real time, and use batch collection for environmental data sources to obtain collected data; Step S12: For the above-mentioned collected data, filtering technology is used to remove noise in the data, wherein the image data is processed using a denoising algorithm; for the missing data points of the collected data, an interpolation method is used to supplement them, and a statistical method is used to identify and process outliers; then the data of different dimensions are converted to the same dimension to obtain preprocessed data; Step S13: extract basic features from the preprocessed data, and use a convolutional neural network (CNN) or a spectrum analysis algorithm to extract advanced features.

3. According to the AI ​​algorithm process based on edge computing according to claim 1, it is characterized in that: The specific steps of model segmentation and distribution in step S2 are as follows: Step S21, model segmentation strategy determination: Determine the model segmentation strategy based on the complexity of the AI ​​model, task requirements, computing power of the edge device, and network bandwidth factors, where the initial segmentation point is selected at the middle layer of the AI ​​model, the early layers are deployed on the edge device, and the later layers are deployed in the cloud; Step S22, model layer division: for deep learning models, the segmentation method is based on the network layer; for traditional machine learning models, segmentation is based on the complexity of feature processing; Step S23, segmentation point selection algorithm: select the segmentation point through the optimization function. Assuming that the model has N layers in total and the computing power that the edge device can carry is C e , the cloud computing power is C c , then the optimization problem can be expressed as: Among them, k is the split point, C e (i) is the computational cost of the i-th layer on the edge device, C c (i) is the computational cost of the i-th layer in the cloud. The goal is to find a split point k that optimally allocates computing resources on the edge and in the cloud. Step S24, model distribution strategy design: After the AI ​​model segmentation is completed, the edge-side model and the cloud-side model are distributed through network connection, where the model of the edge device is lightweight, and TensorFlow Lite or ONNX tools are used to compress and optimize the model to reduce storage space and running time; Step S25, weight allocation and dynamic adjustment: In specific application scenarios, the edge and cloud weight allocation of the model is dynamically adjusted. Specifically, when the edge device load is light and the network conditions are good, more calculations are completed on the edge side; when the edge device load is heavy or the network delay is high, more cloud computing is used. The dynamic adjustment is controlled by the weight allocation coefficient. Assuming that the calculation weights of the edge device and the cloud device are a and β respectively, the dynamic adjustment problem can be expressed as: a+β＝1 Among them, when the computing load of the edge device increases: a↓, β↑; when the network conditions deteriorate: a↑, β↓.

4. The AI ​​algorithm process based on edge computing according to claim 1 is characterized in that: The specific processing steps of edge reasoning and decision making in step S3 are as follows: Step S31: When the edge device is started, the allocated edge side model is loaded into the memory and initialized; Step S32: input the collected data processed in step S1 into the edge side model, use the lightweight model to perform the reasoning task on the edge device, and obtain preliminary reasoning results; Step S33: predict the confidence of the obtained inference result. The specific confidence evaluation content is as follows: Confidence calculation: For classification tasks, the confidence is the probability value calculated by the Softmax function; for regression tasks, the confidence is measured by the difference between the model output value and the preset threshold, where the confidence of the model output is set to P i , where i is the category index, then: Among them, z i is the output score of the model on the i-th category, and N is the total number of categories; Threshold setting: According to the application scenario and tolerance, a confidence threshold T is set. When the output confidence P of the edge model i When it is lower than T, the model is considered uncertain about the current inference result and data needs to be uploaded to the cloud for further processing. The threshold T is determined through offline testing and tuning to balance the efficiency and accuracy of edge inference. Confidence evaluation: After each inference, the output confidence of the edge model is calculated and compared with the set threshold T; if the highest confidence The edge device directly makes decisions and takes corresponding actions; if the highest confidence The inference result is considered unreliable and the next step is entered; Step S34: When the edge device detects that the inference confidence is lower than the set threshold T, the data upload condition is triggered, and the current input data or intermediate feature data is uploaded to the cloud. Before uploading the data, data compression can be performed to reduce the amount of transmitted data. At the same time, to ensure data security, encryption technology should be used when uploading data to protect the privacy of the data during transmission; Step S35: After the cloud receives the data, it performs more complex reasoning through the preset AI model. After the cloud reasoning is completed, the final reasoning result is returned to the edge device. The edge device makes a final decision based on the reasoning result of the cloud and performs corresponding operations.

5. The AI ​​algorithm process based on edge computing according to claim 1 is characterized in that: The specific working steps of the collaborative work between the edge device and the cloud in step S4 are as follows: Step S41: Model update and optimization requirements analysis a. Detection of demand changes: The cloud collects and analyzes the operating data, user feedback, and environmental information of edge devices to detect whether the model needs to be updated; b. Model performance monitoring: Real-time monitoring of the performance of models running on edge devices, including accuracy, response time, and resource usage, including CPU, memory, and power consumption. The cloud regularly evaluates these performance data to determine whether the current model needs to be updated or optimized. c. Model update trigger conditions: When the model performance of the edge device drops below the set threshold, or the environmental changes exceed the model adaptation range, the model update requirement is triggered; Step S42: Cloud model training and optimization a. Data collection and training set generation: The cloud collects data from multiple edge devices, including raw data, inference results of edge devices, and misjudgment cases. Through data cleaning and annotation, a new training set is generated, especially focusing on data with misjudgments and low confidence of edge devices, to ensure the diversity and representativeness of the training set; b. Model training and optimization: The cloud uses the latest collected data to retrain the model; c. Model evaluation and verification: After training is completed, the cloud performs rigorous testing and verification on the new model to ensure its adaptability and robustness in different environments; Step S43: Model parameter distribution and deployment a. Model compression and optimization: Before sending the model to the edge device, the cloud compresses and optimizes the model, including model quantization, which converts floating-point numbers into fixed-point numbers, and pruning, which removes unimportant neurons or weights. b. Model delivery mechanism: The cloud sends the optimized model parameters to the edge device through a secure network channel; c. Edge device model reception and verification: After receiving the new model parameters, the edge device first verifies the integrity and security of the model, specifically checking the hash value and digital signature; after the verification is passed, the edge device loads the new model into the memory and performs initialization operations to prepare to replace the old model; Step S44: Edge device feedback and collaborative optimization a. Real-time feedback mechanism: During operation, edge devices feed back the model’s reasoning results, performance indicators, and abnormal situation data to the cloud in real time. Based on these feedbacks, the cloud adjusts the model’s parameters or structure in a timely manner and performs remote optimization. b. Collaborative optimization strategy: The cloud integrates data from multiple edge devices to perform global optimization. The cloud pushes different optimization strategies or model versions to meet the special needs of different regions or devices. Through reinforcement learning or federated learning, the cloud learns from the operating data of multiple edge devices and extracts common optimization solutions to further improve the performance of the overall system.

6. The AI ​​algorithm process based on edge computing according to claim 1 is characterized in that: The resource monitoring module in step S6 is a key component on the edge device, which is used to monitor the device's CPU, GPU, memory, storage, network bandwidth, battery power resource usage and data in real time. The resource monitoring module analyzes the collected data in real time to identify the changing trend of resource usage, including the periodic fluctuation of CPU load and the gradual increase in memory occupancy. It also predicts the changes in resource demand in the future through historical data analysis and makes optimization decisions in advance, that is, introduces a priority scheduling algorithm to allocate appropriate computing resources according to the priority and resource requirements of the task.

7. A service all-in-one machine, characterized in that: This service all-in-one machine includes the edge computing-based AI algorithm process described in any one of claims 1-6.

Images