Efficient data processing system for Internet of Things equipment

Abstract

The invention relates to the technical field of Internet of Things data processing, and discloses an efficient data processing system for Internet of Things equipment. According to the method, multi-node cooperative training and reasoning are supported through the federated learning driven edge computing cluster, and localized efficient processing is realized while data privacy is protected; the cloud edge collaborative intelligent center solves the compatibility problem of heterogeneous equipment and various scenes through a lightweight model dynamic adaptation technology, the communication overhead is reduced, and the model generalization ability is improved; the cross-layer dynamic scheduling subsystem is combined with space-time flow prediction and a global optimization algorithm to realize elastic allocation of computing resources, so that burst loads are effectively dealt with and node overload is avoided; the full-link trusted execution system integrates data desensitization, security aggregation and block chain evidence storage, constructs hierarchical protection from equipment to the cloud, resists hostile attacks and ensures operation traceability; the AI-driven data processing assembly line greatly reduces the redundant data transmission and storage cost through adaptive compression and exception filtering.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things (IoT) data processing, and particularly to an efficient IoT device data processing system. Background Art

[0002] IoT devices refer to intelligent terminals connected through the Internet and equipped with data collection, transmission, and interaction capabilities, such as sensors, smart home devices, etc. They collect physical environment or user behavior data in real time. The IoT data processing system is responsible for receiving, storing, and analyzing this vast amount of data. Usually, it adopts cloud computing or edge computing technologies, combined with artificial intelligence algorithms to achieve data cleaning, pattern recognition, and real-time feedback. Finally, the processing results are used for device control, business decision-making, or service optimization, forming a closed-loop system of "perception - transmission - analysis - application".

[0003] However, the existing systems face significant challenges: centralized cloud computing has high latency due to remote data transmission, making it difficult to meet the millisecond-level response requirements in scenarios such as industrial control; although edge computing reduces latency, the computing power of a single node is limited and the cooperation efficiency among multiple devices is low, resulting in insufficient complex task processing capabilities. At the same time, the diversity of data modalities and dynamic environmental changes make the adaptability of traditional models poor, and it is difficult to balance privacy protection and computing efficiency in the static encryption mechanism. The rigid resource allocation strategy exacerbates load imbalance and energy waste, restricting the reliability and security of large-scale IoT applications. Therefore, an efficient IoT device data processing system is proposed. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides an efficient IoT device data processing system to solve the problems in the background art.

[0005] To achieve the above object, the present invention provides the following technical solution: An efficient IoT device data processing system, including the following modules:

[0006] Federated learning-enabled edge computing cluster: It includes multiple geographically distributed edge nodes. Each node is configured with heterogeneous computing units (CPU + FPGA + AI accelerator) to achieve local data processing and privacy protection, and reports the resource status vector [computing load, memory occupancy, network bandwidth] to the central coordinator in real time. Among them, the AI accelerator is built-in with an adaptive scheduling engine to optimize the energy efficiency ratio and dynamically switch the DNN model inference mode (fixed-point calculation / mixed precision / sparse calculation) according to the real-time workload.

[0007] Cloud-edge collaborative intelligent center: The model optimization engine is deployed in the cloud to receive the local model parameters uploaded by the edge nodes, generate lightweight models adapted to the characteristics of each region through neural architecture search, and send them to the corresponding edge nodes through a secure channel.

[0008] Cross-layer dynamic scheduling subsystem: It includes a traffic awareness module in the device access layer, a resource status monitoring module in the edge layer, and a global optimization module in the cloud. The three cooperate to achieve:

[0009] In the device access layer, predict the data traffic in the next 5 minutes based on the LSTM network;

[0010] In the edge layer, construct a dynamic topology graph containing the resource status of each node in real time;

[0011] Full-link trusted execution system: It runs through the data path from the device to the cloud and integrates three-layer protection mechanisms, including: a data desensitization module at the device end, a secure aggregation protocol in the edge layer, and a blockchain evidence storage service in the cloud;

[0012] AI-driven data processing pipeline: It is deployed between the edge node and the cloud and includes an adaptive encoder, a configurable DNN feature extraction layer, and a data quality assessment module.

[0013] Preferably, the adaptive scheduling engine adopts a two-layer Q-learning strategy:

[0014] The first-layer decision maker selects the computing precision mode according to the input data features;

[0015] The second-layer decision maker adjusts the parallel computing granularity based on the real-time power consumption budget;

[0016] The two-layer decision makers share a state encoder and use a graph neural network to extract the resource association features of heterogeneous computing units.

[0017] Preferably, the reinforcement learning-driven neural architecture search includes:

[0018] The environment simulator generates an edge node feature matrix [device type distribution, data modality, network latency];

[0019] The reward function synthesizes model accuracy, inference latency, and communication overhead;

[0020] The policy network outputs the architecture search action space, including the combined optimization of operator types, connection topologies, and quantization strategies.

[0021] Preferably, the dynamic topology graph construction adopts:

[0022] The graph convolutional network extracts the spatio-temporal dependence relationship between nodes;

[0023] The dynamic edge weight calculation module fuses the historical load pattern and real-time bandwidth data;

[0024] The topology compression algorithm generates a hierarchical graph representation to support fast optimization and solution in the cloud.

[0025] Preferably, the device-end data desensitization module includes:

[0026] A variational autoencoder generative adversarial network (VAE-GAN) is used to construct a data anonymization model;

[0027] A dynamic privacy budget allocation algorithm is used to balance data availability and privacy protection.

[0028] Preferably, the global optimization module includes:

[0029] Modeling of multi-objective optimization problem: min (total delay, maximum node load, communication energy consumption);

[0030] The Pareto front solver uses the NSGA-III algorithm;

[0031] The solution recommendation engine performs transfer learning based on historical decisions in similar scenarios.

[0032] Preferably, the federated learning process includes:

[0033] Gradient obfuscation mechanism: Inject a noise matrix during local model update;

[0034] The model difference evaluation module detects malicious nodes;

[0035] The adaptive aggregation algorithm dynamically adjusts the aggregation weight according to the node credibility.

[0036] Preferably, the edge node includes:

[0037] Online knowledge distillation module: Use the lightweight model sent from the cloud as the teacher model;

[0038] Local adaptive module: Quickly adjust the model parameters through meta-learning.

[0039] Preferably, the resource status monitoring module includes:

[0040] A multi-dimensional time series prediction model (LSTM+Transformer) predicts resource bottlenecks;

[0041] The elastic resource allocator realizes dynamic partitioning of computing units;

[0042] The fault prediction module gives early warnings of hardware anomalies based on the survival analysis model.

[0043] Preferably, the AI-driven data processing pipeline specifically includes:

[0044] The adaptive encoder dynamically selects a compression algorithm according to the data type;

[0045] The data quality evaluation module uses an anomaly detection model to filter out low-value data.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention supports multi-node collaborative training and inference through a federated learning-driven edge computing cluster, achieving efficient local processing while protecting data privacy; the cloud-edge collaborative intelligent center solves the compatibility problems of heterogeneous devices and diverse scenarios through lightweight model dynamic adaptation technology, reducing communication overhead and improving the model generalization ability; the cross-layer dynamic scheduling subsystem combines spatio-temporal traffic prediction and global optimization algorithms to achieve elastic allocation of computing resources, effectively coping with sudden loads and avoiding node overload; the full-link trusted execution system integrates data desensitization, secure aggregation, and blockchain-based evidence storage to build hierarchical protection from devices to the cloud, resisting malicious attacks and ensuring traceability of operations; the AI-driven data processing pipeline significantly reduces redundant data transmission and storage costs through adaptive compression and anomaly filtering.

[0048] Other features and advantages of the present invention will be described in the following specification, and some of them will be obvious from the specification, or can be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the structures pointed out in the specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is the architecture diagram of the efficient data processing system for Internet of Things devices of the present invention;

[0050] Figure 2 is the data processing flow chart of the present invention;

[0051] Figure 3 is the flow chart of the security protection mechanism of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art in the technical field of the present invention without creative efforts based on the embodiments of the present invention belong to the protection scope of the present invention.

[0053] Please refer to Figures 1-3 , an efficient data processing system for Internet of Things devices in the present invention includes the following:

[0054] 1. Edge computing cluster enabled by federated learning

[0055] Real-time monitor the computing, storage, and network resource status of edge nodes to provide data support for dynamic scheduling.

[0056] Hardware configuration:

[0057] Edge node: Deploy heterogeneous computing units, including a multi-core CPU (such as ARM Cortex-A72) to execute general computing tasks (such as data preprocessing), a programmable FPGA (such as Xilinx Zynq UltraScale+) to accelerate customized algorithms (such as real-time video encoding), and a dedicated AI accelerator (such as Google Edge TPU) dedicated to DNN model inference (such as TensorRT engine), forming a distributed computing network.

[0058] Each node reports the resource status vector to the central coordinator every 5 seconds, expressed as:

[0059]

[0060] where C is the number of CPU cores; u c (t) is the utilization rate (%) of the c-th CPU core at time t, e.g., u1(10:00) = 85%; m uscd (t) is the amount of memory used (GB). For example, when the total node memory is 64GB and 32GB is used, the value is 0.5; b actual (t) is the currently actual occupied bandwidth (Mbps). If the total bandwidth is 1Gbps (i.e., 1000Mbps) and the current occupancy is 200Mbps, the value is 0.2;

[0061] Example:

[0062] The vector reported by a certain edge node at 10:00:00 is Indicating:

[0063] The average CPU load is 80% (3.2 cores out of 4 are fully loaded);

[0064] The memory occupancy is 60% (38.4GB out of 64GB is used);

[0065] The bandwidth occupancy is 20% (200Mbps / 1000Mbps).

[0066] 1.2 Adaptive scheduling engine

[0067] Dynamically adjust the computing mode and parallel strategy according to data characteristics and resource status.

[0068] 1.21 Input feature extraction

[0069] Parse the meta-information of the input data to generate a feature vector:

[0070]

[0071] Example: The features of the 4K video stream generated by an intelligent camera are f = [1200, 50, 1.2×109 , indicating:

[0072] Data volume: 1200MB; Requirement: to be processed within 50ms; Require 120 million floating-point operations.

[0073] 1.22 Double-layer Q-learning dynamic decision-making:

[0074] In edge computing, it is necessary to dynamically adjust the computing mode and parallel strategy according to data characteristics (such as precision requirements, latency sensitivity) and resource status (such as power, temperature).

[0075] The first layer: Computing mode selection (precision / energy efficiency)

[0076] Input state: s1 = [data volume, latency constraint, remaining power];

[0077] Optional actions:

[0078] FP32: 32-bit floating point (high precision, high energy consumption);

[0079] FP16: 16-bit floating point (balanced precision and energy consumption);

[0080] INT8: 8-bit integer (low precision, low energy consumption).

[0081] Reward function:

[0082]

[0083] Example:

[0084] A certain video analysis task:

[0085] FP32 mode: Precision 98%, energy consumption if

[0086] FP16 mode: Precision 95%, energy consumption

[0087] Decision: Select the FP16 mode (higher reward value).

[0088] The second layer: Parallel granularity adjustment (throughput vs temperature)

[0089] Input state: s2 = [length of the queue to be processed, chip temperature];

[0090] Optional actions: Select the number of parallel threads (1 / 2 / 4 / 8);

[0091] Reward function:

[0092]

[0093] Example:

[0094] 4 threads: Throughput 100 tasks / second, temperature rise 0.5 °C / second →

[0095] ≈166.7;

[0096] 8 threads: Throughput 150 tasks / second, temperature rise 2.0 °C / second →

[0097] Decision: Select 4 threads (higher reward value).

[0098] Graph neural network resource modeling:

[0099] Dynamic topology graph g t : The nodes are edge nodes, and the edge weights are communication efficiency.

[0100] Graph convolution formula:

[0101]

[0102] Among them, W sclf , W ncigh are trainable weight matrices; d i is the degree (number of connections) of node i; is the communication efficiency weight between nodes i and j; is the hidden state vector of node j at the l-th layer; d j is the degree of node j, that is, the number of edges connected to node j;

[0103] Resource state vector Drives adaptive scheduling decisions, and the scheduling results are fed back to the cloud-edge center to optimize the model distribution strategy.

[0104] 2. Cloud-edge collaborative intelligent center

[0105] Generates a lightweight model adapted to edge resources through neural architecture search (NAS) to balance accuracy and efficiency.

[0106] 2.1 Environment modeling and feature matrix:

[0107] Construct an edge node feature matrix containing:

[0108] Device density (units / ㎡), data modality ratio, average latency (ms), computing capacity (TFLOPS), storage margin (GB);

[0109] Example: The features of a certain node are [0.6, 0.8, 40, 10, 50], indicating:

[0110] Device density 0.6 units / ㎡, 80% of the data is images;

[0111] The average processing delay is 40 ms, the computing capacity is 10 TFLOPS, and the remaining storage is 50 GB.

[0112] 2.2 Reinforcement learning policy network:

[0113] Policy network structure: 6-layer Transformer encoder, with a hidden layer dimension of 512;

[0114] Action space: includes 3 types of actions:

[0115] Operator type: standard convolution, depthwise separable convolution, dilated convolution;

[0116] Connection method: residual connection, dense connection;

[0117] Quantization policy: 8-bit dynamic quantization, 16-bit half-precision.

[0118] Reward function:

[0119]

[0120] Among them, the denominator 100: is used to normalize the dimensions of delay (ms) and volume (MB);

[0121] Example:

[0122] If a certain model has an accuracy of 95%, a delay of 20 ms, and a volume of 15 MB, then the reward value:

[0123] R = 0.6×0.95 - 0.3×0.2 - 0.1×0.15 = 0.57 - 0.06 - 0.015 = 0.495.

[0124] 2.3 Model distribution and deployment:

[0125] Secure transmission protocol: Adopt key agreement based on elliptic curve cryptography:

[0126] K session = ECDH(SK edge , PK cloud )

[0127] Among them, SK edge is the private key of the edge node; PK cloud is the public key of the cloud;

[0128] The model volume generated by NAS directly affects the memory occupancy rate (m uscd ) of the edge node, and the delay index is closely related to the bandwidth utilization rate (b actual ) in the resource status vector. ) is closely related.

[0129] 3. Cross - layer Dynamic Scheduling Subsystem

[0130] 3.1 Device - layer Traffic Prediction

[0131] Use spatio - temporal attention LSTM to predict the future data traffic peak and pre - allocate computing resources.

[0132] 3.11 Model Structure:

[0133] Input: Traffic sequence in the past 60 minutes {x t-60 ,...,x t};

[0134] Spatio - temporal Attention Mechanism:

[0135]

[0136] Among them, is the query vector (Query) of the current node; is the key vector (Key) of neighboring nodes; d = 64 is the vector dimension, used to scale the dot - product result;

[0137] Output Prediction: Traffic in the next 5 minutes

[0138] 3.12 Training Loss Function:

[0139]

[0140] Among them, N is the number of training samples; is the predicted value of the model for the i - th sample (such as traffic in the next 5 minutes); y i is the true value of the i - th sample; λ is the weight coefficient (such as 0.1), controlling the importance of the KL - divergence term; KL(p data ||p model ) is the Kullback - Leibler divergence, measuring the difference between the true data distribution p data and the model - predicted distribution p model .

[0141] 3.13 Cloud - side Global Optimization

[0142] Use mixed - integer programming to solve the multi - objective optimization problem, minimizing the total delay, maximum energy consumption, and communication cost;

[0143] Model formula:

[0144] min 0.5T total +0.3E max +0.2C comm

[0145]

[0146] Among them, T total = ∑ i,j t ij x ij is the total time delay (t ij is the time delay of task j at node i); is the maximum node energy consumption (e ij is the energy consumption of task j at node i); C comm = ∑ i,j c ij x ij The total communication cost (c ij is the communication overhead of node i for processing task j); x ij is the number of resource units allocated from node i to task j (such as the number of CPU cores); D j is the resource requirement of task j; R i is the upper limit of the resource capacity of node i; M is the total number of nodes; J is the total number of tasks.

[0147] Solution process:

[0148] NSGA-III algorithm parameters:

[0149] Population size is 200, crossover probability is 0.9, mutation probability is 0.1;

[0150] Reference point generation method: 120 uniformly distributed points are generated by the Das-Dennis method.

[0151] Example:

[0152] Input: 3 nodes (resource capacities 20, 15, 25), 5 tasks (requirements 8, 12, 10, 6, 9);

[0153] Output: The optimal allocation scheme reduces the total time delay by 25% and the energy consumption by 18%.

[0154] 4. Full-link trusted execution system

[0155] 4.1 Data desensitization

[0156] VAE-GAN is adopted to eliminate sensitive information and retain data availability.

[0157] 4.11 VAE-GAN joint training:

[0158] Encoder E(x): Maps the input data to the latent space z ~ N(μ, σ 2 );

[0159] Generator G(z): Reconstructs the desensitized data x from the latent space;

[0160] Discriminator D(x): Distinguish real data from generated data;

[0161] Loss function:

[0162]

[0163] Among them, β is the regularization coefficient, controlling the regularization strength of the latent space distribution; λ is the adversarial loss weight, balancing the reconstruction error and the generation quality; q(z|x) is the latent variable distribution output by the encoder; p(z) is the prior distribution (usually the standard normal distribution N(0, I)).

[0164] 4.12 Dynamic privacy budget allocation:

[0165]

[0166] Among them, i is the privacy budget allocated to data x i ; is the total privacy budget (such as 1.0), set by the administrator; sensitivity level (x i ) is the sensitivity degree of data x i (such as face data = 10, temperature data = 1); x j represents the j-th data item in the dataset.

[0167] Example:

[0168] The sensitivity level of face data is set to 10, and the environmental data is 1, and the total budget = 1;

[0169] If there are 10 face images and 90 environmental images in the dataset, then for the face data

[0170] 4.2 Federated learning secure aggregation

[0171] Prevent malicious nodes from contaminating the global model.

[0172] 4.21 Gradient obfuscation: Add Gaussian noise to protect parameter privacy:

[0173]

[0174] Among them, is the local model parameter update (gradient) of node i; is Gaussian noise with a mean of 0 and a variance of σ 2 ; Δf is the query sensitivity (the maximum L2 norm of parameter changes); δ = 10 -5 is the failure probability; is the privacy budget (the smaller the value, the stronger the privacy protection).

[0175] 4.22 Anomaly Detection: Calculate the Mahalanobis distance of parameter updates:

[0176]

[0177] where d i is the Mahalanobis distance of the parameter update of node i, measuring the degree of its deviation from the average value; is the parameter update vector of node i; μ is the mean vector of all node parameter updates; Σ is the covariance matrix of parameter updates.

[0178] Decision rule: If d i > 3σ, mark it as an abnormal node and reduce the aggregation weight to 0.1.

[0179] 5. AI-Driven Data Processing Pipeline

[0180] 5.1 Adaptive Encoder

[0181] Dynamically select the compression algorithm according to the data type to balance quality and efficiency;

[0182] Decision strategy:

[0183]

[0184] where A is the set of candidate algorithms (such as {JPEG2000, HEVC, Zstandard}); w1 is the quality weight (positively correlated with the task urgency); w2 is the efficiency weight (w2 = 1 - w2); PSNR(a) is the peak signal-to-noise ratio of algorithm a; Compression Ratio(a) is the compression efficiency of algorithm a.

[0185] Example:

[0186] Emergency video stream (urgency = 8): Select HEVC (PSNR = 38dB, compression ratio 15:1);

[0187] Ordinary sensor data (urgency = 2): Select Zstandard (PSNR = ∞, compression ratio 8:1).

[0188] 5.2 Anomaly Detection

[0189] Adopt the Isolation Forest algorithm to quickly identify low-value or abnormal data and reduce redundant processing.

[0190] Specifically:

[0191] Build isolation trees: Randomly select features and split values to isolate abnormal points;

[0192] Path length calculation: The path of abnormal data in the tree is shorter. Among them, A is the set of candidate algorithms (such as {JPEG2000, HEVC, Zstandard}); w1 is the quality weight (positively correlated with the urgency of the task); w2 is the efficiency weight (w2 = 1 -).

[0194] Abnormal score:

[0195]

[0196] Decision threshold: If Score(x) > 0.7, it is determined as abnormal; E(h(x)) is the average path length of the data point x in the isolation tree; c(n) is the normalization coefficient, related to the sample number n; 0.5772 is the Euler - Mascheroni constant, used to correct the expected value of the path length.

[0197] Example:

[0198] Temperature sensor data: The normal value is 25°C ± 2°C. If 50°C is detected and the score is 0.85, it is discarded.

[0199] 6. Example of the system working process:

[0200] Scenario 1: Processing of large - scale sensor data

[0201] Goal: Real - time process sensor data in the factory environment, optimize resource allocation and model update.

[0202] 1. Data acquisition

[0203] Input:

[0204] Temperature sensor: 25°C (normal range), humidity sensor: 60% RH.

[0205] The edge node receives the raw data [25, 60].

[0206] 2. Desensitization and compression

[0207] Operations:

[0208] VAE - GAN desensitization: Blur the sensor location information (e.g., "Area A - 05" → "Area").

[0209] Compression: Compress the data with Zstandard, from the original 10KB to 0.83KB after compression (compression ratio 12:1).

[0210] 3. Feature extraction and anomaly detection

[0211] Operations:

[0212] Load the TinyLSTM model and extract temporal features (such as the temperature change slope of 0.2 °C / min).

[0213] The humidity is detected to have suddenly increased to 85% (abnormal threshold of 70%), and it is marked as abnormal.

[0214] 4. Resource Scheduling

[0215] Operations:

[0216] The LSTM predicts that the data volume will increase by 150% in the next 5 minutes (from 100 records / second to 250 records / second).

[0217] The node load is detected to be unbalanced (node A load is 90%, node B load is 30%).

[0218] Migrate 30% of the tasks on node A to node B (such as moving 50 computing tasks to node B).

[0219] 5. Model Update and Security

[0220] Operations:

[0221] The lightweight model is sent from the cloud (volume from 5MB → 2MB, accuracy from 92% → 94.3%).

[0222] The parameters of node C are detected to be abnormal (update distance from the mean is 3.5σ), and the node is isolated.

[0223] The blockchain records the event hash: 0x3a7d...f1c2.

[0224] Scenario 2: Real-time Video Analysis

[0225] Objective: Monitor the production line video stream and detect product defects in real time.

[0226] 1. Data Input

[0227] Input:

[0228] The camera captures a 1080p video stream (30 frames / second).

[0229] The edge node enables INT8 quantization, and the single-frame inference latency changes from 50ms → 12ms.

[0230] 2. Dynamic Scheduling

[0231] Operations:

[0232] The FPGA temperature is detected to reach 82 °C (threshold of 80 °C), and it automatically switches to the mixed-precision mode (FP16 + INT8).

[0233] The power consumption changes from 20W → 12W (a 40% reduction).

[0234] 3. Model Collaboration

[0235] Operation:

[0236] The cloud generates a dedicated detection model (supporting multi-scale defect recognition), which is encrypted and then distributed.

[0237] The encryption key is rotated every 24 hours (e.g., from Key-A → Key-B).

[0238] 4. Exception Handling

[0239] Operation:

[0240] Isolation Forest detects that a certain frame is blurred (score 0.82 > threshold 0.7), and it is marked as a low-value frame.

[0241] Discard this frame (45 frames are discarded out of 1000 total frames, accounting for 4.5%).

[0242] This embodiment provides an efficient Internet of Things data processing system. Through federated learning to coordinate edge nodes to achieve local model training, combined with dynamic resource scheduling and cloud-edge collaboration optimization, it realizes multi-level adaptive processing from the device side to the cloud while ensuring data privacy. The system integrates anomaly detection, lightweight model compression, and secure aggregation mechanisms, supports real-time traffic prediction and elastic resource allocation, significantly improves data processing efficiency and system robustness, and is applicable to complex scenarios such as industrial monitoring and smart cities.

Claims

1. An efficient data processing system for Internet of Things devices, characterized in that, It includes the following modules: Federated learning-enabled edge computing cluster: It contains multiple edge nodes distributed geographically. Each node is configured with heterogeneous computing units (CPU+FPGA+AI accelerator) to achieve data local processing and privacy protection, and reports the resource status vector [computing load, memory occupancy, network bandwidth] to the central coordinator in real time. The AI accelerator is built-in with an adaptive scheduling engine to optimize the energy efficiency ratio, and dynamically switches the DNN model inference mode (fixed-point calculation / mixed precision / sparse calculation) according to the real-time workload; Cloud-edge collaborative intelligent center: The model optimization engine is deployed in the cloud to receive the local model parameters uploaded by the edge nodes, generates lightweight models adapted to the characteristics of each region through neural architecture search, and distributes them to the corresponding edge nodes through a secure channel; Cross-layer dynamic scheduling subsystem: It includes a traffic awareness module at the device access layer, a resource status monitoring module at the edge layer, and a global optimization module at the cloud. The three cooperate to achieve: At the device access layer, predict the data traffic in the next 5 minutes based on the LSTM network; At the edge layer, build a dynamic topology graph containing the resource status of each node in real time; Full-link trusted execution system: Integrate three-layer protection mechanisms throughout the data path from the device to the cloud, including: device-side data desensitization module, edge-layer secure aggregation protocol, and cloud blockchain evidence storage service; AI-driven data processing pipeline: Deployed between the edge nodes and the cloud, it includes an adaptive encoder, a configurable DNN feature extraction layer, and a data quality assessment module.

2. The efficient data processing system for an Internet of Things device according to claim 1, characterized in that The adaptive scheduling engine adopts a two-layer Q-learning strategy: The first-layer decision maker selects the computing precision mode according to the input data characteristics; The second-layer decision maker adjusts the parallel computing granularity based on the real-time power consumption budget; The two-layer decision makers share a state encoder, and use a graph neural network to extract the resource association characteristics of heterogeneous computing units.

3. An efficient data processing system for Internet of Things devices according to claim 1, characterized in that The reinforcement learning-driven neural architecture search includes: The environment simulator generates an edge node feature matrix [device type distribution, data modality, network latency]; The reward function synthesizes model accuracy, inference latency, and communication overhead; The policy network outputs the architecture search action space, including the combined optimization of operator types, connection topologies, and quantization strategies.

4. An efficient data processing system for an Internet of Things device according to claim 1, characterized in that, The construction of the dynamic topology graph adopts: The graph convolutional network extracts the spatio-temporal dependence relationship between nodes; The dynamic edge weight calculation module fuses the historical load pattern and real-time bandwidth data; The topology compression algorithm generates a hierarchical graph representation to support fast optimization and solution in the cloud.

5. An efficient data processing system for Internet of Things devices according to claim 1, characterized in that The device-side data desensitization module includes: The variational autoencoder generative adversarial network (VAE-GAN) constructs a data anonymization model; The dynamic privacy budget allocation algorithm adjusts the balance between data availability and privacy protection.

6. An efficient data processing system for Internet of Things devices according to claim 1, characterized in that, The global optimization module includes: Modeling of multi-objective optimization problems: min (total latency, maximum node load, communication energy consumption); The Pareto front solver adopts the NSGA-III algorithm; The solution recommendation engine performs transfer learning based on historical decisions in similar scenarios.

7. An efficient data processing system for Internet of Things devices according to claim 1, characterized in that, The federated learning process includes: Gradient confusion mechanism: Inject a noise matrix during local model update; The model difference evaluation module detects malicious nodes; The adaptive aggregation algorithm dynamically adjusts the aggregation weight according to the node credibility.

8. An efficient data processing system for an Internet of Things device according to claim 1, wherein, The edge node includes: Online knowledge distillation module: using the lightweight model sent from the cloud as the teacher model; Local adaptive module: quickly adjusts the model parameters through meta-learning.

9. An efficient data processing system for Internet of Things devices according to claim 1, characterized in that The resource status monitoring module includes: Multidimensional time series prediction model (LSTM + Transformer) to predict resource bottlenecks; Elastic resource allocator to achieve dynamic partitioning of computing units; Fault prediction module to early warn of hardware anomalies based on the survival analysis model.

10. An efficient data processing system for an Internet of Things device according to claim 1, characterized in that, The AI-driven data processing pipeline specifically includes: Adaptive encoder to dynamically select compression algorithms according to data types; Data quality assessment module to filter low-value data using the anomaly detection model.

Images