Iot data intelligent shunting and processing system and method based on edge computing

Abstract

The application discloses an Internet of Things data intelligent shunting and processing system and method based on edge computing, relates to the technical field of edge computing and artificial intelligence, and comprises an edge data sensing module, an intelligent shunting decision module, a lightweight edge inference module, a federal collaborative learning module and a cloud-edge collaborative optimization module. High-value data is identified through multi-dimensional importance evaluation, a shunting strategy is dynamically optimized based on deep reinforcement learning, cloud model capability is migrated to the edge end through knowledge distillation, edge device collaborative training is realized through federal learning, and data security is protected through differential privacy. The application solves the problems of single data processing strategy, insufficient utilization of edge computing capability, lack of dynamic adaptability and privacy protection in the prior art.

Description

Technical Field

[0001] This invention relates to the fields of edge computing and artificial intelligence, specifically to an IoT data intelligent offloading and processing system and method based on edge computing. Background Technology

[0002] With the rapid development of IoT technology and the large-scale deployment of 5G networks, the amount of data generated in fields such as smart cities, industrial IoT, and smart homes is growing exponentially. Traditional cloud computing architectures upload all data to the cloud for centralized processing, which faces problems such as high transmission latency, high bandwidth pressure, and difficulties in privacy protection. Edge computing, as a distributed computing paradigm, can effectively reduce data transmission latency and alleviate network burden by offloading computing tasks to the network edge.

[0003] Chinese patent application CN118113787A discloses a smart city IoT data processing method and system based on edge computing. This solution acquires data from a cloud database, calculates and analyzes the acquisition time based on device type to output an acquisition interval, then calculates the interaction time based on the numerical values ​​within the acquisition interval, outputting a preliminary adjustment time or first aggregation information. Next, it optimizes the preliminary adjustment time to output an optimized adjustment time, and finally aggregates data from edge devices based on the optimized adjustment time or the first aggregation information. This solution reduces data congestion and improves data transmission efficiency to some extent.

[0004] However, the aforementioned existing technologies still have the following shortcomings: First, the solution mainly focuses on optimizing data upload time, lacking in-depth analysis of the importance of data content, resulting in high-value data and low-value data being processed using the same strategy, and resource allocation is not refined enough; Second, the solution does not fully utilize the computing power of edge devices for intelligent data processing, with the edge only performing simple data aggregation operations, unable to achieve complex data analysis and reasoning tasks; Third, the processing strategy of the solution is a preset static rule, unable to be dynamically adjusted according to network conditions and data characteristics, and lacks adaptability to the complex and ever-changing IoT environment; Fourth, the solution does not consider collaborative learning between edge devices, with each edge device working independently, unable to share learning experience to continuously improve overall processing capabilities; Fifth, the solution lacks an effective privacy protection mechanism during data transmission, and sensitive data directly uploaded to the cloud poses a risk of privacy leakage.

[0005] Therefore, there is an urgent need for an IoT data intelligent diversion and processing technology solution that can intelligently analyze the importance of data, dynamically adjust processing strategies, fully utilize edge computing capabilities, support collaborative learning of edge devices, and protect data privacy. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as limited data processing strategies, insufficient utilization of edge computing capabilities, lack of dynamic adaptability and privacy protection mechanisms, this invention provides an IoT data intelligent diversion and processing system and method based on edge computing.

[0007] In a first aspect, the present invention provides an IoT data intelligent diversion and processing system based on edge computing, including an edge data perception module, an intelligent diversion decision module, a lightweight edge inference module, a federated collaborative learning module, and a cloud-edge collaborative optimization module.

[0008] The edge data perception module is used to collect raw data from IoT devices, preprocess and extract features from the raw data, generate data feature vectors, and calculate data importance scores based on the data’s temporal sensitivity, semantic relevance and anomaly degree through a multidimensional importance assessment model.

[0009] The intelligent traffic splitting decision module receives data feature vectors, data importance scores, and network state parameters obtained through network monitoring components. Based on a deep reinforcement learning network, it selects data processing actions according to the current system state and outputs a traffic splitting decision vector, which includes a processing location label and a transmission path identifier.

[0010] The lightweight edge inference module receives data to be processed from the intelligent triage decision module and performs local inference using a student model obtained from the cloud teacher model through knowledge distillation technology, and outputs the edge inference results.

[0011] The federated collaborative learning module is used to perform local model training based on local training data and edge inference results, calculate local gradient parameters, and upload the local gradient parameters, after differential privacy processing, to the cloud.

[0012] The cloud-edge collaborative optimization module receives local gradient parameters uploaded by each edge device, performs weighted aggregation to generate global model parameters, distributes the global model parameters to each edge device, performs knowledge distillation based on the output differences between the cloud teacher model and the edge student model, generates distilled knowledge parameters and updates the student model in the lightweight edge inference module, and transmits the system processing performance indicators as feedback reward signals to the intelligent diversion decision module to update the diversion strategy.

[0013] Preferably, the data importance score is obtained by weighted fusion calculation of time-series sensitivity index, semantic relevance index and anomaly index. The values ​​of the three indicators are all in the range of 0 to 1, and the value range of the data importance score is 0 to 1.

[0014] Preferably, the network state parameters include current bandwidth utilization, transmission latency, and packet loss rate, and the deep reinforcement learning network adopts a dual deep Q-network architecture.

[0015] Preferably, the number of parameters in the student model does not exceed 5% to 25% of the number of parameters in the teacher model, and the marginal inference results include inference output values ​​and confidence scores.

[0016] Preferably, the local model training employs a mini-batch stochastic gradient descent method, and differential privacy processing includes adding noise that follows a Laplace distribution to the local gradient parameters.

[0017] Preferably, the weighted aggregation adopts an adaptive weight allocation strategy based on data volume and model quality, and the knowledge distillation adopts a temperature-adaptive soft-label distillation method.

[0018] Secondly, the present invention also provides an intelligent data offloading and processing method for the Internet of Things based on edge computing, which is applied to the above-mentioned system. The method includes an edge data perception step, an intelligent offloading decision step, a lightweight edge inference step, a federated collaborative learning step, and a cloud-edge collaborative optimization step, with each step corresponding to the function of the corresponding module in the system.

[0019] The beneficial effects of this invention are as follows:

[0020] First, by using a multidimensional importance assessment model to analyze the importance of data, high-value data can be identified and prioritized, thus improving the targeting of data processing and the efficiency of resource utilization.

[0021] Secondly, by using deep reinforcement learning to achieve dynamic optimization of the traffic splitting strategy, the system can adaptively adjust the processing location and transmission path according to network conditions and data characteristics, significantly improving its adaptability to complex and ever-changing environments.

[0022] Third, by using knowledge distillation technology, the capabilities of complex models in the cloud are transferred to lightweight models at the edge, which significantly reduces the computing resource requirements of edge devices while maintaining high inference accuracy, enabling resource-constrained edge devices to perform complex intelligent analysis tasks.

[0023] Fourth, the federated learning framework enables collaborative training among edge devices, allowing each device to improve model performance without sharing raw data, thus protecting data privacy while achieving knowledge sharing.

[0024] Fifth, the differential privacy mechanism protects the uploaded parameters, effectively preventing the leakage of sensitive information and meeting the stringent privacy protection requirements of IoT applications. Attached Figure Description

[0025] Figure 1 This is an architecture diagram of an IoT data intelligent diversion and processing system based on edge computing, provided in an embodiment of the present invention.

[0026] Figure 2This is a flowchart of the IoT data intelligent diversion and processing method based on edge computing provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 As shown, this embodiment of the invention provides an intelligent IoT data offloading and processing system based on edge computing. This system is deployed in a layered IoT architecture comprising cloud servers and multiple edge computing nodes. In one embodiment, the edge computing nodes can be edge servers deployed on the roadside in smart cities, industrial gateways in industrial IoT scenarios, or home gateway devices in smart home environments. The system includes an edge data sensing module 1, an intelligent offloading decision module 2, a lightweight edge inference module 3, a federated collaborative learning module 4, and a cloud-edge collaborative optimization module 5. These five modules form a deeply coupled closed-loop collaborative architecture. The output of the edge data sensing module 1 serves as the key input to the intelligent offloading decision module 2. The output of the intelligent offloading decision module 2 determines the processing content of the lightweight edge inference module 3. The results of the lightweight edge inference module 3 are fed back to the federated collaborative learning module 4 for model optimization. The federated collaborative learning module 4 interacts with the cloud-edge collaborative optimization module 5, and the output of the cloud-edge collaborative optimization module 5 in turn affects the policy parameters of the intelligent offloading decision module 2 and the model parameters of the lightweight edge inference module 3, thus forming a complete closed-loop feedback mechanism.

[0029] Edge data sensing module 1 is deployed on each edge computing node and is responsible for the collection, preprocessing, and feature analysis of IoT data. This module is the data entry point for the entire system, and its processing quality directly affects the performance of subsequent modules.

[0030] In one embodiment of the present invention, the edge data sensing module 1 first collects raw data from IoT devices through a standardized interface. The types of raw data include, but are not limited to, temperature sensor data, humidity sensor data, video surveillance data, device operating status data, and energy consumption monitoring data. The data acquisition frequency is set according to the application scenario. Preferably, for industrial IoT scenarios, the acquisition frequency is set to 10 to 100 samples per second; for smart city scenarios, the acquisition frequency is set to 1 to 10 samples per second.

[0031] After acquiring the raw data, the data preprocessing unit in the edge data perception module 1 performs missing value imputation, outlier filtering, and standardization on the raw data. Missing value imputation uses either linear interpolation or nearest neighbor imputation, selecting an appropriate imputation strategy based on the temporal characteristics of the data. Outlier filtering employs a statistical method, marking data points exceeding three times the standard deviation of the mean as outliers and performing smoothing or removal. Standardization uses the Z-score standardization method, mapping the data to a standard normal distribution with a mean of 0 and a standard deviation of 1. Specifically, this is calculated by subtracting the mean of the data sequence from the raw data and then dividing by the standard deviation.

[0032] After preprocessing, the edge data perception module 1 extracts features from the data to generate data feature vectors. In one embodiment of the present invention, the feature extraction employs a lightweight convolutional neural network structure, which includes three convolutional layers and one fully connected layer. The first convolutional layer uses 32 3×3 convolutional kernels, the second convolutional layer uses 64 3×3 convolutional kernels, and the third convolutional layer uses 128 3×3 convolutional kernels. Each convolutional layer is followed by a ReLU activation function and a 2×2 max pooling operation. The fully connected layer maps the features to feature vectors of fixed dimensions; preferably, the feature vector dimension is set to 128 or 256 dimensions.

[0033] After feature extraction, the edge data perception module 1 calculates the data importance score based on the data's temporal sensitivity, semantic relevance, and anomaly degree using a multidimensional importance assessment model. The multidimensional data importance assessment algorithm proposed in this invention can be implemented using the following formula:

[0034] ,

[0035] in, Assign a score to the importance of the data, with a value ranging from 0 to 1; It is a time-series sensitivity index, which characterizes the sensitivity of data to changes over time, and its value ranges from 0 to 1; This is a semantic relevance index, representing the degree of relevance between data and key business semantics, with a value range of 0 to 1; This is an anomaly index, representing the degree to which data deviates from the normal pattern, with a value ranging from 0 to 1; , , For the weighting coefficients, satisfying The constraints. In a preferred embodiment of the invention, The value is 0.3. The value is 0.4. The value is 0.3. This parameter configuration has been experimentally verified to achieve good results in smart city application scenarios.

[0036] Time-series sensitivity index The calculation is based on the data change rate and timeliness requirements. The specific calculation method is as follows: First, calculate the difference between the current data and historical data; then, divide the difference by the timeliness window length and normalize it. The timeliness window length is set according to the data type. Preferably, for real-time monitoring data, the window length is set to 1 to 5 seconds; for status monitoring data, the window length is set to 10 to 60 seconds.

[0037] Semantic relevance index The calculation is based on the degree of matching between the data content and predefined key semantics. In one embodiment of the present invention, the system pre-establishes a key semantic lexicon, which includes important business terms and key indicator names in various application scenarios. The semantic relevance index is obtained by calculating the cosine similarity between the data description information and the key semantic lexicon.

[0038] Abnormality index The calculation is based on the degree of deviation of the data from historical normal patterns. The specific calculation method is as follows: First, a probability distribution model of the normal pattern is established based on historical data. Then, the log-likelihood value of the current data under this distribution is calculated. Finally, the log-likelihood value is converted into an anomaly score within the range of 0 to 1. The greater the deviation from the normal pattern, the higher the anomaly score.

[0039] When the importance score of data exceeds a preset importance threshold, the edge data awareness module 1 marks the corresponding data as high-priority data. Preferably, the importance threshold is set to 0.6 to 0.8, with the specific value determined based on the business requirements of the application scenario. High-priority data will receive priority processing and more computing resources in subsequent traffic allocation decisions and processing.

[0040] The intelligent traffic splitting decision module 2 is the core decision-making unit of the system of this invention. It is responsible for dynamically determining the processing location and transmission path of data based on data characteristics, importance, and network status. This module adopts deep reinforcement learning technology, which can adaptively optimize the traffic splitting strategy in complex and ever-changing IoT environments.

[0041] The intelligent traffic offloading decision module 2 receives three types of inputs: data feature vectors and data importance scores from the edge data perception module 1, network status parameters obtained through the network monitoring component, and feedback reward signals from the cloud-edge collaborative optimization module 5. Network status parameters include current bandwidth utilization, transmission latency, and packet loss rate. Current bandwidth utilization represents the degree of network transmission channel usage, ranging from 0% to 100%; transmission latency represents the time required for data to be transmitted from the edge to the cloud, measured in milliseconds; and packet loss rate represents the proportion of data packets lost during data transmission, ranging from 0% to 100%.

[0042] The intelligent traffic splitting decision module 2 combines the above input information into a system state vector, which is then used as input to the deep reinforcement learning network. In one embodiment of the present invention, the dimension of the state vector is the data feature vector dimension plus four dimensions (data importance score 1 dimension, bandwidth utilization 1 dimension, transmission delay 1 dimension, and packet loss rate 1 dimension). Preferably, when the data feature vector dimension is 128 dimensions, the total dimension of the state vector is 132 dimensions.

[0043] The deep reinforcement learning network in the intelligent traffic splitting decision module 2 adopts a Double Deep Q-Network (DDQN) architecture. This architecture consists of two neural networks with identical structures but different parameters: an evaluation network and a target network. The evaluation network is used to select actions, and the target network is used to calculate the target Q-value. This design can effectively alleviate the Q-value overestimation problem in traditional DQN.

[0044] The evaluation network and target network structure includes an input layer, three hidden layers, and an output layer. The input layer dimension is equal to the state vector dimension. The number of neurons in the three hidden layers are 256, 128, and 64 respectively, and each hidden layer uses the ReLU activation function. The number of neurons in the output layer is equal to the number of selectable actions. In one embodiment of the invention, the selectable actions include edge local processing, cloud processing, and multiple selectable transmission paths. The number of neurons in the output layer is set to 5 to 10, with the specific number determined according to the network topology.

[0045] The reinforcement learning-based decision-making algorithm proposed in this invention is implemented based on the following Q-value update formula:

[0046] ,

[0047] in, To evaluate the state-action value function of the network output, representing the state... Next action Expected cumulative rewards; The learning rate controls the step size of parameter updates, with a value ranging from 0.0001 to 0.01, preferably set to 0.001. The reward is calculated by the cloud-edge collaborative optimization module 5 based on the system's processing performance indicators. This is a discount factor, representing the degree of decay of future rewards, with a value ranging from 0.9 to 0.99, preferably set to 0.95; The state-action value function output by the target network; The next state to transition to after performing an action; Indicates the evaluation network in the next state The optimal action to choose from.

[0048] The intelligent triage decision module 2 also includes an experience replay cache unit for storing historical state transition samples. These historical state transition samples are in the form of four-tuples, including the current state. Actions performed Instant rewards received and the next state to which it is transitioned The capacity of the experience replay buffer unit is set to 10,000 to 100,000 samples, and a first-in, first-out (FIFO) elimination strategy is adopted. During training, a preset number of samples are randomly sampled from the experience replay buffer unit to update the network parameters. Preferably, the batch size of each sampling is set to 32 to 128 samples.

[0049] The training of deep reinforcement learning networks employs an ε-greedy strategy to balance exploration and exploitation. In the initial training phase, actions are randomly selected for exploration with a high probability ε; as training progresses, the value of ε is gradually decreased, increasing the proportion of learned strategies utilized. Preferably, the initial value of ε is set to 1.0, the final value to 0.01, and exponential decay is used with a decay rate of 0.995.

[0050] The parameters of the target network are copied from the evaluation network at preset update intervals. Preferably, the update interval is set to 100 to 1000 steps, that is, the target network parameters are updated once after every 100 to 1000 action selections. This soft update strategy can ensure the stability of the target Q-value and avoid drastic oscillations during training.

[0051] The intelligent traffic offloading decision module 2 outputs a traffic offloading decision vector, which includes a processing location label and a transmission path identifier. The processing location label includes an edge processing label and a cloud processing label, indicating whether the data should be processed locally at the edge or uploaded to the cloud for processing. The transmission path identifier specifies the path to be used when multiple transmission paths are available. When the processing location label is edge processing, the intelligent traffic offloading decision module 2 sends the data to be processed to the lightweight edge inference module 3; when the processing location label is cloud processing, the intelligent traffic offloading decision module 2 uploads the data to the cloud server through the specified transmission path.

[0052] The lightweight edge inference module 3 is deployed on edge computing nodes and is responsible for performing local intelligent inference on data diverted to the edge. This module uses a lightweight student model obtained from the cloud teacher model through knowledge distillation technology, which can achieve inference accuracy close to that of the cloud model on resource-constrained edge devices.

[0053] The lightweight edge inference module 3 receives the data to be processed, which is diverted to the edge by the intelligent diversion decision module 2, as input. After receiving the data, the lightweight edge inference module 3 first converts the data format to meet the input requirements of the student model, and then inputs the converted data into the student model to perform forward inference.

[0054] The student model's structural design follows a lightweight principle, minimizing the number of model parameters and computational complexity while maintaining high inference accuracy. In one embodiment of the invention, the student model employs a depthwise separable convolutional structure, which decomposes standard convolution into two steps: depthwise convolution and pointwise convolution, significantly reducing computational cost while maintaining similar receptive fields. The student model contains five depthwise separable convolutional blocks and two fully connected layers, with the number of output channels for each convolutional block being 16, 32, 64, 64, and 128, respectively.

[0055] The number of parameters in the student model does not exceed a preset compression ratio for the number of parameters in the teacher model. In a preferred embodiment of the invention, the preset compression ratio ranges from 5% to 25%. For example, when the number of parameters in the cloud-based teacher model is 50 million, the number of parameters in the edge-based student model is controlled between 2.5 million and 12.5 million. This compression ratio has been experimentally verified to compress the model size to one-fifth to one-twentieth of the original size without sacrificing more than 5% in inference accuracy.

[0056] The lightweight edge inference module 3 outputs edge inference results, including inference output values ​​and confidence scores. The inference output value is the student model's prediction of the input data; its format depends on the specific application scenario and can be a classification label, regression value, or feature representation. The confidence score represents the student model's degree of confidence in the current prediction result, ranging from 0 to 1, and is determined by the maximum value of the softmax activation value of the student model's output layer.

[0057] When the confidence score is lower than a preset confidence threshold, the lightweight edge inference module 3 marks the corresponding data as a low-confidence sample and forwards the sample to the cloud for verification. Preferably, the confidence threshold is set to 0.7 to 0.9, with the specific value determined according to the accuracy requirements of the application scenario. This adaptive verification mechanism effectively ensures the reliability of the inference results, allowing the cloud-based teacher model to handle samples that are difficult for the student model to process accurately.

[0058] The parameters of the student model are generated and periodically updated by the cloud-edge collaborative optimization module 5 using knowledge distillation technology. The update cycle is set according to the application scenario. Preferably, for scenarios where the model needs to quickly adapt to new data distributions, the update cycle is set to once per hour or once per day; for scenarios where the data distribution is relatively stable, the update cycle is set to once per week or once per month.

[0059] The Federated Collaborative Learning Module 4 is deployed across various edge computing nodes. It is responsible for training models based on local data and securely interacting with the cloud to exchange parameters, enabling collaborative learning among multiple edge devices. This module employs a federated learning framework, allowing edge devices to jointly train a global model without sharing raw data, thus protecting data privacy while achieving knowledge sharing.

[0060] The input to the Federated Co-learning Module 4 includes local training data and edge inference results from the Lightweight Edge Inference Module 3. The local training data, collected and stored locally by the Edge Data Awareness Module 1, contains the original data and its corresponding label information. The edge inference results serve as a reference for model optimization, used to evaluate the performance of the current model and guide adjustments to the training direction.

[0061] Federated Co-learning Module 4 performs local model training using the mini-batch stochastic gradient descent method. The specific process is as follows: First, a mini-batch of samples is randomly sampled from the local training data. Then, the samples are input into the local model to perform forward propagation to calculate the loss function value. Next, backpropagation is performed to calculate the gradients of the parameters at each layer. Finally, the model parameters are updated based on the gradients.

[0062] The number of local training rounds ranges from 1 to 10, with the specific number determined based on a trade-off between the amount of local data and the communication frequency. Preferably, when the amount of local data is large and the communication cost is high, the number of local training rounds is set to 5 to 10; when the amount of local data is small or more frequent synchronization of the global model is required, the number of local training rounds is set to 1 to 3. The mini-batch size is set to 16 to 64 samples, and the learning rate is set to 0.001 to 0.01.

[0063] After the local model training is complete, the Federated Co-learning Module 4 calculates the local gradient parameters. The local gradient parameters are the differences between the local model parameters and the global model parameters from the previous round, representing the amount of model updates made during local training. This method of transmitting gradients rather than complete model parameters effectively reduces communication overhead.

[0064] Before uploading the local gradient parameters, the Federated Co-learning Module 4 performs differential privacy processing to prevent sensitive information from being leaked through the gradient. Differential privacy processing involves adding noise following a Laplace distribution to the local gradient parameters. In the proposed differential privacy federated aggregation algorithm, the noise addition process can be implemented using the following formula:

[0065] ,

[0066] in, These are the gradient parameters after adding noise; These are the original local gradient parameters; This represents a random sample from a Laplace distribution, with the parameter being the noise scale; Sensitivity to gradients represents the maximum influence of a single sample on the gradient calculation result. It is calculated by clipping the gradient and taking the clipping threshold as the upper bound of the sensitivity. The privacy budget parameter controls the strength of privacy protection. The smaller the value, the stronger the privacy protection, but the greater the impact on model accuracy. The value range is from 0.1 to 10, and it is preferably set to 1 to 5.

[0067] Gradient clipping is a key step in differential privacy processing, used to limit the influence of individual samples on the gradient. The clipping operation is performed as follows: First, the L2 norm of the gradient vector is calculated. When the L2 norm exceeds a preset clipping threshold, the gradient vector is scaled proportionally until its L2 norm equals the clipping threshold. Preferably, the clipping threshold is set to 1 to 10, with the specific value determined based on the model size and data distribution.

[0068] After differential privacy processing, the local gradient parameters are uploaded to the cloud-edge collaborative optimization module 5. The upload uses an encrypted transmission channel to ensure that the parameters are not stolen or tampered with during transmission.

[0069] The Federated Collaborative Learning Module 4 also receives global model parameters from the Cloud-Edge Collaborative Optimization Module 5. Upon receiving the global model parameters, the Federated Collaborative Learning Module 4 replaces its local model parameters with them and then continues local training based on the updated model. This periodic parameter synchronization mechanism enables edge devices to share the learning experience of other devices, continuously improving the overall model performance.

[0070] The cloud-edge collaborative optimization module 5 is deployed on the cloud server and is responsible for the aggregation and updating of the global model, knowledge distillation, and monitoring and feedback of system performance. This module is the key hub for realizing cloud-edge collaboration, connecting various edge devices and coordinating the optimization direction of the overall system.

[0071] The cloud-edge collaborative optimization module 5 receives local gradient parameters uploaded by each edge device and performs weighted aggregation to generate global model parameters. The weighted aggregation employs an adaptive weight allocation strategy based on data volume and model quality. The aggregation weight for each edge device is determined according to the proportion of local data and the local model validation accuracy, calculated using the following formula:

[0072] ,

[0073] in, For the first The aggregate weight of each edge device ranges from 0 to 1, and the sum of the weights of all devices is 1. For the first Local data volume of each edge device; For the first The local model validation accuracy of each edge device is obtained by evaluating each edge device on the local validation set and uploaded along with the gradient parameters; The total number of edge devices participating in this round of aggregation.

[0074] The global model parameters are updated according to the following formula:

[0075] ,

[0076] in, These are the updated global model parameters; These are the global model parameters before the update. For the first Local gradient parameters uploaded by an edge device after differential privacy processing.

[0077] The cloud-edge collaborative optimization module 5 also includes a model version management unit for maintaining the global model version sequence. The version number is incremented after each global model update. When the difference between the model version corresponding to the local gradient parameters uploaded by the edge device and the current global model version exceeds a preset version threshold, the gradient parameter is discarded to ensure aggregation consistency. Preferably, the version threshold is set to 2 to 5, meaning that the model version of the edge device is allowed to lag behind the current global version by 2 to 5 versions.

[0078] The updated global model parameters are distributed to the federated collaborative learning module 4 of each edge device. The distribution adopts an incremental update method, transmitting only the parameter changes rather than the complete parameters to reduce communication overhead.

[0079] The cloud-edge collaborative optimization module 5 simultaneously performs knowledge distillation, transferring the capabilities of the cloud-based teacher model to the edge-based student model. Knowledge distillation employs a temperature-adaptive soft-label distillation method, and the adaptive knowledge distillation algorithm proposed in this invention can be implemented using the following formula:

[0080] The distillation loss function is defined as:

[0081] ,

[0082] in, For knowledge distillation loss; The distillation temperature parameter is used to soften the output distribution of the teacher and student models, and its value ranges from 1 to 20. This is the KL divergence function, used to measure the difference between two probability distributions; It is the softmax function; The output logits vector of the student model; This is the output logits vector of the teacher model; This is the temperature square factor, used to balance the gradient magnitude of distillation losses.

[0083] The distillation temperature parameter is dynamically adjusted based on the difference in output distribution between the teacher model and the student model. The specific adjustment strategy is as follows: First, the entropy value of the teacher model's output distribution is calculated. When the entropy value is high (the output distribution is relatively flat), a lower temperature parameter is used to retain more detailed information; when the entropy value is low (the output distribution is relatively sharp), a higher temperature parameter is used to soften the distribution and facilitate learning by the student model. Preferably, the temperature parameter is dynamically adjusted between 1 and 10.

[0084] The total training loss of the student model is a weighted combination of distillation loss and true label loss:

[0085] ,

[0086] in, The total training loss for the student model; The weighting coefficient for distillation loss ranges from 0 to 1, and is preferably set to 0.5 to 0.9. Output the cross-entropy loss between the student model and the true label.

[0087] The distilled knowledge parameters generated by knowledge distillation are distributed to the lightweight edge inference module 3 of each edge device to update the student model. The distilled knowledge parameters include the complete weight parameters of the student model, and the update method is to directly replace the existing parameters of the student model at the edge.

[0088] The cloud-edge collaborative optimization module 5 is also responsible for monitoring and providing feedback on system performance. System processing performance metrics include average processing latency, inference accuracy, resource utilization, and energy consumption. Average processing latency represents the average time from when data arrives at the edge to when the processing result is obtained; inference accuracy represents the correctness of the inference results at the edge and in the cloud; resource utilization represents the proportion of computing resources used by edge devices and cloud servers; and energy consumption represents the total energy consumed during system operation.

[0089] The cloud-edge collaborative optimization module 5 calculates and feeds back a reward signal based on the above performance indicators, which is then transmitted to the intelligent traffic allocation decision module 2 to update the traffic allocation strategy. The calculation of the reward signal comprehensively considers latency, accuracy, and resource efficiency, and the calculation formula is as follows:

[0090] ,

[0091] in, This is a reward signal, with a value ranging from 0 to 1; This is due to a current processing delay. For the maximum acceptable latency, it is preferably set to 100 to 500 milliseconds; The inference accuracy is represented by a value ranging from 0 to 1. Current energy consumption; The maximum acceptable energy consumption; , , For the weighting coefficients, satisfying Preferably set as , , .

[0092] Through the deep coupling and closed-loop collaboration of the above five modules, the system of this invention can achieve intelligent diversion and efficient processing of IoT data. The data importance assessment result of the edge data perception module 1 affects the decision input of the intelligent diversion decision module 2. The diversion result of the intelligent diversion decision module 2 determines the processing content of the lightweight edge inference module 3. The inference result of the lightweight edge inference module 3 is fed back to the federated collaborative learning module 4 for model optimization. The gradient parameters of the federated collaborative learning module 4 are converged to the cloud-edge collaborative optimization module 5 for global aggregation. The output of the cloud-edge collaborative optimization module 5 then updates the strategy of the intelligent diversion decision module 2 and the model of the lightweight edge inference module 3, forming a continuously optimized closed-loop system.

[0093] The system of this invention has demonstrated significant technical effects in practical deployment. In tests within a smart city scenario, the system was deployed in an environment comprising 100 edge nodes and one cloud server. The edge nodes were embedded devices configured with a quad-core ARM processor and 4GB of memory, while the cloud server was a high-performance server configured with a 32-core CPU, 128GB of memory, and four GPUs. The test dataset included mixed data from devices such as temperature sensors, humidity sensors, and video surveillance, totaling 1TB per day. Test results show that, compared to existing solutions, the average processing latency of the system of this invention is reduced by 45%, from 200 milliseconds to 110 milliseconds; data transmission bandwidth consumption is reduced by 60%, from 600GB per day to 240GB; inference accuracy is improved by 8%, from 87% to 95%; and the utilization rate of computing resources on edge devices is increased from 30% to 75%. These technical effects are attributed to the combined effect of the intelligent traffic allocation decision-making, knowledge distillation, and federated collaborative learning mechanisms of the system of this invention.

[0094] Please see Figure 2 As shown, this embodiment of the invention also provides an intelligent data offloading and processing method for the Internet of Things based on edge computing, applied to the system described in the above system embodiment, including the following steps.

[0095] Step S1, Edge Data Sensing Step: Collect raw data from IoT devices, preprocess and extract features from the raw data to generate data feature vectors, and calculate the data importance score based on the data’s temporal sensitivity, semantic relevance and anomaly degree through a multidimensional importance assessment model.

[0096] In one embodiment of the present invention, the specific execution process of step S1 corresponds to the working process of edge data perception module 1 in the system embodiment. After the raw data is acquired, preprocessing operations such as missing value imputation, outlier filtering, and Z-score normalization are first performed. Then, a 128-dimensional or 256-dimensional data feature vector is extracted through a lightweight convolutional neural network. Finally, a data importance score is calculated based on a multidimensional importance assessment algorithm. When the data importance score is greater than a preset importance threshold, the corresponding data is marked as high-priority data.

[0097] Step S2, Intelligent Triage Decision Step: Receive data feature vectors, data importance scores, and network state parameters obtained through network monitoring components. Based on the deep reinforcement learning network, select data processing actions according to the current system state and output a triage decision vector.

[0098] In one embodiment of the present invention, the specific execution process of step S2 corresponds to the working process of the intelligent traffic splitting decision module 2 in the system embodiment. Network state parameters include current bandwidth utilization, transmission delay, and packet loss rate. The deep reinforcement learning network adopts a dual deep Q-network architecture, selects the optimal action based on the current state vector, and outputs a traffic splitting decision vector containing processing location labels and transmission path identifiers. The training process employs an ε-greedy strategy to balance exploration and utilization, and improves sample utilization efficiency through an experience replay mechanism.

[0099] Step S3, Lightweight Edge Inference Step: For the data to be processed that has been diverted to the edge, local inference is performed using the student model obtained from the cloud teacher model through knowledge distillation technology, and the edge inference results are output.

[0100] In one embodiment of the present invention, the specific execution process of step S3 corresponds to the working process of the lightweight edge inference module 3 in the system embodiment. The student model adopts a depthwise separable convolutional structure, and the number of parameters does not exceed 5% to 25% of the number of parameters in the teacher model. The edge inference result includes inference output value and confidence score. When the confidence score is lower than the preset confidence threshold, the corresponding data is forwarded to the cloud for review processing.

[0101] Step S4, Federated Collaborative Learning Step: Perform local model training based on local training data and edge inference results, calculate local gradient parameters, and upload the local gradient parameters after differential privacy processing to the cloud.

[0102] In one embodiment of the present invention, the specific execution process of step S4 corresponds to the working process of the federated collaborative learning module 4 in the system embodiment. Local model training employs a mini-batch stochastic gradient descent method, with 1 to 10 local training rounds. Differential privacy processing includes two steps: gradient clipping and adding Laplacian noise. The noise scale parameter is determined according to a preset privacy budget.

[0103] Step S5, Cloud-Edge Collaborative Optimization Step: Perform weighted aggregation on the local gradient parameters uploaded by each edge device to generate global model parameters and distribute them. At the same time, perform knowledge distillation to generate distilled knowledge parameters to update the student model at the edge. And pass the system processing performance index as a feedback reward signal to the intelligent diversion decision step to update the diversion strategy.

[0104] In one embodiment of the present invention, the specific execution process of step S5 corresponds to the working process of cloud-edge collaborative optimization module 5 in the system embodiment. Weighted aggregation adopts an adaptive weight allocation strategy based on data volume and model quality, and knowledge distillation employs a temperature-adaptive soft-label distillation method. System processing performance indicators include average processing latency, inference accuracy, resource utilization, and energy consumption. The reward signal calculated based on these indicators is fed back to step S2 to continuously optimize the traffic distribution strategy.

[0105] The above steps S1 to S5 form a closed-loop execution flow. The output of step S1 serves as the input of step S2, the output of step S2 determines the processing content of step S3, the result of step S3 is fed back to step S4, the output of step S4 converges to step S5, and the output of step S5 in turn affects the execution parameters of steps S2 and S3. The system continuously runs the above closed-loop flow, and with the continuous accumulation of data and continuous optimization of the model, the processing performance gradually improves and tends to stabilize.

[0106] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions can also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means implemented in a flow...Figure 1 One or more processes and boxes Figure 1 The function specified in one or more boxes.

[0107] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. An IoT data intelligent offloading and processing system based on edge computing, characterized in that, include: The edge data sensing module is used to collect raw data from IoT devices, preprocess and extract features from the raw data to generate data feature vectors, and calculate data importance scores based on the data’s temporal sensitivity, semantic relevance and anomaly degree through a multidimensional importance assessment model. The intelligent traffic splitting decision module is used to receive the data feature vector, the data importance score, and network state parameters obtained through the network monitoring component. Based on the deep reinforcement learning network, it selects data processing actions according to the current system state and outputs a traffic splitting decision vector, which includes a processing location label and a transmission path identifier. The lightweight edge inference module is used to receive the data to be processed that is diverted to the edge by the intelligent diversion decision module, and to perform local inference using a student model obtained from the cloud teacher model through knowledge distillation technology, and output the edge inference results. The federated collaborative learning module is used to perform local model training based on local training data and the edge inference results, calculate local gradient parameters, and upload the local gradient parameters after differential privacy processing to the cloud. The cloud-edge collaborative optimization module is used to receive the local gradient parameters uploaded by each edge device, perform weighted aggregation to generate global model parameters, distribute the global model parameters to each edge device, and perform knowledge distillation based on the output difference between the cloud teacher model and the edge student model to generate distilled knowledge parameters and update the student model in the lightweight edge inference module. The system processing performance index is transmitted as a feedback reward signal to the intelligent diversion decision module to update the diversion strategy.

2. The system according to claim 1, characterized in that, In the edge data perception module, the data importance score is obtained by weighted fusion calculation of time-series sensitivity index, semantic relevance index and anomaly index; the value range of the time-series sensitivity index is 0 to 1, the value range of the semantic relevance index is 0 to 1, and the value range of the anomaly index is 0 to 1. The data importance score ranges from 0 to 1. When the data importance score is greater than a preset importance threshold, the corresponding data is marked as high priority data.

3. The system according to claim 1, characterized in that, In the intelligent traffic splitting decision module, the network status parameters include current bandwidth utilization, transmission latency, and packet loss rate; the deep reinforcement learning network adopts a dual deep Q-network architecture, which includes an evaluation network and a target network. The evaluation network is used to select actions, and the target network is used to calculate the target Q-value; the processing location labels include edge processing labels and cloud processing labels.

4. The system according to claim 1, characterized in that, In the lightweight edge inference module, the number of parameters of the student model does not exceed the preset compression ratio of the number of parameters of the teacher model, and the preset compression ratio ranges from 5% to 25%. The edge inference result includes inference output value and confidence score. When the confidence score is lower than the preset confidence threshold, the corresponding data is forwarded to the cloud for review.

5. The system according to claim 1, characterized in that, In the federated collaborative learning module, the local model training adopts the mini-batch stochastic gradient descent method, and the number of local training rounds ranges from 1 to 10 rounds; the differential privacy processing includes adding noise that follows a Laplace distribution to the local gradient parameters, and the noise scale parameter is determined according to a preset privacy budget.

6. The system according to claim 1, characterized in that, In the cloud-edge collaborative optimization module, the weighted aggregation adopts an adaptive weight allocation strategy based on data volume and model quality. The aggregation weight of each edge device is determined according to the proportion of local data volume and the local model validation accuracy. The knowledge distillation adopts a temperature-adaptive soft-label distillation method, and the distillation temperature parameter is dynamically adjusted according to the difference in output distribution between the teacher model and the student model.

7. The system according to claim 1, characterized in that, The edge data perception module also includes a data preprocessing unit, which is used to perform missing value imputation, outlier filtering and standardization on the raw data; the standardization process adopts the Z-score standardization method to map the data to a distribution with a mean of 0 and a standard deviation of 1.

8. The system according to claim 1, characterized in that, The intelligent traffic splitting decision module also includes an experience replay cache unit, which is used to store historical state transition samples. The historical state transition samples include a quadruple of state, action, reward and next state. During training, a preset number of samples are randomly sampled from the experience replay cache unit to update the network parameters.

9. The system according to claim 1, characterized in that, The cloud-edge collaborative optimization module also includes a model version management unit, which is used to maintain the version sequence of the global model. When it is detected that the model version corresponding to the local gradient parameter uploaded by the edge device differs from the current global model version by more than a preset version threshold, the gradient parameter is discarded to ensure aggregation consistency.

10. A method for intelligent data offloading and processing in the Internet of Things based on edge computing, applied to the system described in any one of claims 1 to 9, characterized in that, include: Edge data perception steps: Collect raw data from IoT devices, preprocess and extract features from the raw data to generate data feature vectors, and calculate the data importance score based on the data’s temporal sensitivity, semantic relevance and anomaly degree through a multidimensional importance assessment model; Intelligent traffic splitting decision steps: Receive the data feature vector, the data importance score, and the network state parameters obtained through the network monitoring component; select data processing actions based on the current system state using a deep reinforcement learning network; and output a traffic splitting decision vector. Lightweight edge inference steps: For the data to be processed that is diverted to the edge, the student model obtained from the cloud teacher model through knowledge distillation technology is used to perform local inference and output the edge inference results; Federated collaborative learning steps: Perform local model training based on local training data and the edge inference results, calculate local gradient parameters, and upload the local gradient parameters after differential privacy processing to the cloud; Cloud-edge collaborative optimization steps: The local gradient parameters uploaded by each edge device are weighted and aggregated to generate global model parameters and distributed. At the same time, knowledge distillation is performed to generate distilled knowledge parameters to update the student model at the edge. The system processing performance index is transmitted as a feedback reward signal to the intelligent diversion decision step to update the diversion strategy.

Images