An end-edge-cloud three-level distributed layered autonomous intelligent control architecture and collaborative scheduling method

Abstract

The application relates to the technical field of Internet of Things intelligent control, and discloses an end-edge-cloud three-level distributed layered autonomous intelligent control architecture and a collaborative scheduling method, wherein the method comprises the following steps: a terminal extracts a feature vector through a lightweight sensing model; an edge gateway generates a control instruction based on an edge decision model; the edge gateway aggregates operation data and uploads the operation data to a cloud end; the cloud end trains a global optimization model and issues updated parameters; a calculation task is dynamically scheduled according to a multi-constraint optimization model; and a network connection state is detected, and a local autonomous mode is switched to when the network connection state is interrupted. The application solves the technical problems of poor real-time performance, high bandwidth occupation, unbalanced computing power and fragile business continuity in the cloud centralized calculation mode.

Description

Technical Field

[0001] This invention relates to the field of IoT intelligent control technology, specifically to a three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling method for the edge-cloud. Background Technology

[0002] The Internet of Things (IoT) system is widely used in smart buildings, industrial automation, smart cities, and other scenarios, requiring real-time sensing and control of a large number of terminal devices. Existing technologies generally adopt a cloud-based centralized computing model, where terminal devices upload the collected raw data to a cloud server, where the cloud performs data processing, model inference, and decision generation before issuing control commands to the terminal for execution.

[0003] The aforementioned centralized cloud computing model suffers from the following drawbacks: Data transmission between the terminal and the cloud, along with cloud-based queuing computation, results in end-to-end latency reaching the second level, failing to meet the millisecond-level response requirements of real-time control scenarios; continuous uploading of massive amounts of raw data consumes significant network bandwidth, and packet loss caused by network congestion leads to a decrease in the accuracy of perception algorithms; there is an overabundance of computing power in the cloud and a shortage of computing power at the terminal, resulting in unreasonable resource allocation; and the terminal lacks local decision-making capabilities during network outages, causing the system to malfunction. These drawbacks collectively make it difficult for the centralized cloud computing model to simultaneously meet the comprehensive requirements of real-time performance, low bandwidth consumption, balanced computing power, and business continuity. Summary of the Invention

[0004] This invention provides a three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling, which solves the technical problems of strong network dependence, high response latency and low reliability of traditional centralized control systems.

[0005] This invention discloses a three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling, comprising the following steps: A terminal device collects environmental data through sensors, inputs the environmental data into a lightweight perception model, outputs a feature vector, and discards the environmental data; an edge gateway receives the feature vector, inputs the feature vector into an edge decision model, and generates control commands; the edge gateway periodically performs desensitization processing and data aggregation on the running data, generates aggregated data, and uploads it to a cloud server; the cloud server trains a global optimization model based on the aggregated data, generates optimized decision strategy parameters, calculates model parameter difference packets, and sends them to the edge gateway and the terminal device; a multi-constraint optimization model is constructed according to the current system state to solve for the optimal execution position of each computational task, and the computational tasks are allocated to the terminal, edge, or cloud for execution; the edge gateway periodically sends heartbeat messages to the cloud server, and when the heartbeat timeout duration exceeds the timeout threshold, it determines that the network connection is interrupted, switches to local autonomous mode to independently generate control commands, and when the network connection is restored, it synchronizes the model and switches back to collaborative mode.

[0006] Furthermore, the lightweight perception model is obtained by sequentially performing model distillation, quantization compression, and structured pruning on the original perception model. The model distillation uses the original perception model as the teacher model and the native lightweight network as the student model, and trains it by jointly optimizing the hard label loss and soft label loss. The hard label loss is calculated based on the student model's predicted output and the true label, while the soft label loss is calculated based on the student model's predicted output and the soft label output by the teacher model under the temperature parameter. The two losses are weighted and summed to obtain the total loss.

[0007] Furthermore, the quantization compression maps model parameters and activation values ​​from floating-point numbers to integers. It statistically analyzes the numerical range of tensors in each layer using a calibration dataset, calculates a scaling factor based on the ratio of the difference between the maximum and minimum values ​​within the numerical range to the integer representation range, and calculates the zero-point offset based on the difference between the minimum integer representation and the minimum value within the numerical range divided by the scaling factor. The structured pruning calculates the weights for each convolutional channel. The norm is used as an importance measure, and channels whose importance measure is less than a threshold are removed.

[0008] Furthermore, the edge decision model includes a lightweight neural network and a rule engine. The lightweight neural network takes the feature vector as input, and its output layer uses a softmax activation function to output the probability distribution of each pattern category, selecting the category with the highest probability as the pattern recognition result. The rule engine generates the control command based on the pattern recognition result and preset rules, which include conditional judgment logic and action mapping relationships.

[0009] Furthermore, the edge gateway aggregates feature vectors uploaded by multiple terminals within the same area, extracts spatiotemporal correlation features, inputs the spatiotemporal correlation features into the edge decision model, and generates global control commands for that area.

[0010] Furthermore, the objective function of the multi-constraint optimization model is to minimize the weighted sum of the end-to-end latency of each computational task, the reciprocal of the weighted sum of resource utilization, and the task failure probability. The end-to-end latency is standardized to eliminate dimensional differences before being substituted into the objective function. The task failure probability is calculated based on the priority weight of each task, the network packet loss rate, and the number of data packets to be transmitted by each task. The constraints are that the available computing power at the execution location is not less than the computing power required by the task, and the currently available bandwidth is not less than the minimum bandwidth requirement.

[0011] Furthermore, the multi-constraint optimization model is periodically resolved, with a solution period of 1 to 10 seconds. When the network packet loss rate exceeds a threshold or the network latency exceeds a threshold, immediate scheduling is triggered, and the multi-constraint optimization model is resolved.

[0012] Furthermore, in the local autonomous mode, the edge gateway loads pre-cached decision strategies and lightweight decision models from local storage to independently generate control commands, and caches the runtime data in a local time-series database using a circular buffer structure. When the network connection is restored, the edge gateway sends a model version number to the cloud server. If the model version number is inconsistent with the latest model version number, it receives a model parameter differential packet to complete the model update, and uploads the runtime data cached in the local time-series database to the cloud server via a breakpoint resume method.

[0013] Furthermore, the model parameter differential packet includes a version number, differential data, and a checksum, wherein the checksum is calculated using a hash algorithm. After receiving the model parameter differential packet, the edge gateway and the terminal device merge the differential data with the local current model parameters to generate a candidate model, calculate the hash value of the candidate model, and compare it with the checksum. If they match, the candidate model replaces the local current model; if they do not match, the local current model remains unchanged, and an update failure notification is sent to the cloud server.

[0014] This invention discloses a three-level distributed hierarchical autonomous intelligent control system (edge-cloud) for executing the aforementioned method, comprising: a terminal feature extraction module for collecting environmental data through sensors, inputting the environmental data into a lightweight perception model to output feature vectors, and discarding the environmental data; an edge decision generation module for receiving the feature vectors, inputting the feature vectors into an edge decision model to generate control commands; a data aggregation and uploading module for periodically desensitizing and aggregating the running data, generating aggregated data, and uploading it to a cloud server; a cloud strategy optimization module for training a global optimization model based on the aggregated data, generating optimized decision strategy parameters, and calculating model parameter difference packets to be sent to edge gateways and terminal devices; a task dynamic scheduling module for constructing a multi-constraint optimization model based on the current system state, solving for the optimal execution position of each computational task, and allocating the computational tasks to the terminal, edge, or cloud for execution; and a disconnection self-healing management module for periodically sending heartbeat messages to the cloud server, determining network connection interruption when the heartbeat timeout duration exceeds a timeout threshold, switching to local autonomous mode to independently generate control commands, and synchronizing the model and switching back to collaborative mode when the network connection is restored.

[0015] This invention deploys the perception algorithm to the terminal after model distillation, quantization compression, and structured pruning, deploys the decision-making algorithm to the edge gateway, and retains the global optimization algorithm in the cloud. This achieves hierarchical processing of computing tasks, solves the technical problem of computing power imbalance in the centralized cloud computing model, and achieves the technical effect of reasonable allocation of computing resources. After the terminal extracts feature vectors, it only transmits low-dimensional feature vectors, solving the technical problem of bandwidth consumption from uploading large amounts of raw data and achieving the technical effect of reducing the amount of data transmitted over the network. The edge gateway completes inference and generates control commands locally, solving the technical problem of excessive latency caused by data traveling to and from the cloud and achieving the technical effect of shortening end-to-end control latency. When the network connection is interrupted, the edge gateway switches to local autonomous mode and independently generates control commands based on pre-cached decision strategies, solving the technical problem of system malfunction due to network interruption and achieving the technical effect of ensuring business continuity. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the end-edge-cloud three-level collaborative architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the dual-algorithm relationship in an embodiment of the present invention, illustrating the decoupling and collaborative mechanism between the perception algorithm and the decision-making algorithm; Figure 3 This is a simplified schematic diagram of the core system process in an embodiment of the present invention. Detailed Implementation

[0017] I. Implementation Examples Traditional IoT intelligent control systems employ a centralized cloud computing model. Terminal devices upload collected raw data to a cloud server, where data processing, model inference, and decision generation are performed before control commands are sent back to the terminals for execution. This centralized cloud computing model suffers from the following technical problems: data transmission and cloud computing between terminals result in end-to-end latency reaching the second level, failing to meet millisecond-level real-time control requirements; the massive upload of raw data consumes significant bandwidth resources, and packet loss caused by network congestion leads to decreased accuracy of sensing algorithms; cloud computing power is excessive but maintenance costs are high, while terminal computing power is insufficient but model computation is heavy, resulting in unreasonable resource allocation; and terminals lack local decision-making capabilities during network outages, leading to system paralysis.

[0018] To address the aforementioned technical challenges, this invention proposes a three-tiered distributed hierarchical autonomous intelligent control architecture and collaborative scheduling system. This architecture decouples the perception and decision-making algorithms. Through model lightweighting technology, the perception algorithm is deployed to the terminal, the real-time decision-making algorithm is deployed to the edge gateway, and the global optimization algorithm is retained in the cloud, enabling hierarchical processing and collaborative scheduling of computational tasks.

[0019] According to an embodiment of the present invention, a three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling based on edge-cloud includes the following steps: Step S1: Acquire environmental data and extract feature vectors The terminal device collects environmental data through sensors, including at least one of video streams, audio signals, temperature values, humidity values, and light intensity. This environmental data is input into a lightweight perception model, which outputs a feature vector. The feature vector includes at least one of target detection results, skeleton keypoint coordinates, frequency domain features, and environmental parameter statistics.

[0020] It should be noted that the lightweight perception model is obtained by performing model distillation, quantization compression, and structured pruning on the original perception model. The original perception model is a high-precision perception model pre-trained on a cloud server, with a parameter count of [missing information]. The number of parameters in the lightweight perception model is ,satisfy .

[0021] It should be noted that the model distillation process uses the original perception model as the teacher model and the native lightweight network as the student model. The student model takes environmental data as input and feature vectors as output, and is trained using the following loss function:

[0022] in, The predicted output of the student model. For the true labels of the training data, For hard-label loss function, For the teacher model in temperature parameters The output soft tags, This is the output logarithm of the teacher model. For soft-label loss function, and For the weighting coefficients, satisfying The training process employs the Adam optimization algorithm. The output layer of the student model is a fully connected layer corresponding to the dimension of the feature vector. Using the cross-entropy loss function, The KL divergence loss function is used.

[0023] It should be noted that the quantization compression maps model parameters and activation values ​​from 32-bit floating-point numbers to 8-bit integers, with the mapping relationship as follows:

[0024] in, For real values, The quantized integer value. Scaling factor This represents the zero-point offset. The numerical range of tensors in each layer is statistically analyzed using the calibration dataset. Calculate the scaling factor and zero offset ,in and These are the maximum and minimum values ​​represented by integers, respectively.

[0025] It should be noted that the structured pruning removes channels with low contribution from the convolutional neural network. The weights of each convolutional channel are calculated. norm As a measure of importance, among For the first channel One weight parameter, This represents the total number of weight parameters for this channel. Channels are sorted according to their importance metric, and channels with an importance metric below a threshold are removed. The channel, the threshold Determined based on target sparsity.

[0026] In this embodiment, to reduce the computational load and network transmission overhead of the terminal, the terminal discards the original environmental data immediately after extracting the feature vector, retaining only the feature vector for subsequent transmission. The data volume of the feature vector is less than one-tenth of the data volume of the original environmental data.

[0027] Step S2: Receive feature vectors and generate control commands. The edge gateway receives feature vectors uploaded by terminals through a communication interface. These feature vectors are then input into an edge decision model to generate control commands. The edge decision model includes a lightweight neural network and a rule engine. The lightweight neural network is used to identify complex patterns, and the rule engine is used to perform deterministic logical judgments. The control commands include at least one of the following: device on / off status, device parameter adjustment values, and scene linkage trigger signals.

[0028] It should be noted that the lightweight neural network takes feature vectors as input and pattern recognition results as output. The output layer is a fully connected layer corresponding to the number of pattern categories. It is trained using supervised learning, with cross-entropy loss as the optimization objective, and the Adam optimization algorithm is used for parameter updates. The rule engine generates control commands based on the pattern recognition results and preset rules. The preset rules include conditional judgment logic and action mapping relationships. When the conditional judgment logic is met, the corresponding action mapping relationship is triggered. The inference latency of the edge decision model is less than 100 milliseconds.

[0029] Furthermore, the output layer of the lightweight neural network uses the softmax activation function to output the probability distribution of each pattern category. The category with the highest probability is taken as the pattern recognition result, and then the rule engine maps the pattern recognition result into specific control instructions.

[0030] In this embodiment, to improve decision-making accuracy, the edge decision model integrates feature vectors uploaded by multiple terminals when generating control commands. For multiple terminals within the same area, the edge gateway aggregates the feature vectors of each terminal, extracts spatiotemporal correlation features, and inputs these features into the edge decision model to generate global control commands for that area.

[0031] Step S3: Aggregate runtime data and upload it to the cloud. The edge gateway periodically collects local operational data, which includes at least one of the following: feature vector statistics uploaded by terminals, control command execution records, device operating status parameters, and network communication quality indicators. The operational data is anonymized and aggregated to generate aggregated data. This aggregated data is then uploaded to a cloud server.

[0032] It should be noted that the upload cycle for the aggregated data is 5 to 30 minutes. The data anonymization process removes user and device identification information from the operational data. The data aggregation calculates statistical measures of the operational data over time, including at least one of the following: mean, variance, maximum value, minimum value, and quantiles.

[0033] In this embodiment of the application, in order to reduce network bandwidth usage, the edge gateway uses a lossless compression algorithm to encode the aggregated data before uploading it, generates a compressed data packet, and uploads the compressed data packet to the cloud server through the transmission control protocol.

[0034] Step S4: Optimize the decision-making strategy and update the model. A cloud server receives aggregated data uploaded from multiple edge gateways and stores the aggregated data in a time-series database. Based on the aggregated data, a global optimization model is trained to generate optimized decision strategy parameters. The global optimization model includes at least one of a reinforcement learning model, a deep neural network model, and an optimization algorithm model.

[0035] It should be noted that the reinforcement learning model uses the system state represented by aggregated data as the state space, the adjustment scheme of decision policy parameters as the action space, and system performance indicators as the basis for calculating the reward function. These system performance indicators include at least one of average response latency, resource utilization, and task success rate. The reinforcement learning model learns the optimal policy through interaction with the environment and is trained using the policy gradient method.

[0036] Furthermore, the decision strategy parameters output by the reinforcement learning model are task scheduling weight vectors for each execution location. The task scheduling weight vectors need to be decoded into specific task allocation schemes through the argmax operation before being sent to the edge gateway and terminal devices for execution.

[0037] Furthermore, the parameters of the edge decision-making model and the lightweight perception model are updated. The difference between the optimized model parameters and the current model parameters is calculated to generate a model parameter difference package. The model parameter difference package is then sent to the edge gateway and terminal devices. The model parameter difference package includes a version number, difference data, and a checksum. The checksum is calculated using the SHA-256 hash algorithm to verify the integrity of the model parameter update.

[0038] In this embodiment, to ensure the consistency of model updates, the edge gateway and terminal device perform verification after receiving the model parameter differential packet. The received differential data is merged with the local current model parameters to generate a candidate model. The hash value of the candidate model is calculated and compared with the verification value in the model parameter differential packet. When the hash value of the candidate model matches the verification value, the candidate model replaces the local current model, completing the model update; when the hash value of the candidate model does not match the verification value, the local current model remains unchanged, and an update failure notification is sent to the cloud server.

[0039] Step S5: Dynamically schedule computing tasks to the terminal, edge, or cloud. The execution location of the computing task is determined based on the current system state. The system state includes at least one of the following: terminal computing resources, edge computing resources, cloud computing resources, network bandwidth, network latency, and packet loss rate. A multi-constraint optimization model is constructed, and the objective function of the multi-constraint optimization model is:

[0040] in, To calculate the total number of tasks, For the first The execution location of each computational task. For the index of the computation task, Indicates the terminal. Indicates the edge, Indicates cloud computing. For execution location End-to-end delay, For execution location Resource utilization rate This represents the probability of task failure. , , These are the weighting coefficients.

[0041] Furthermore, in the objective function The unit is milliseconds. The dimensionless normalized utilization rate has a range of values. , This is a dimensionless probability value, with a range of values ​​of [value missing]. To eliminate the influence of dimensional differences on the weighted summation operation, the values ​​are adjusted before substituting them into the objective function. Z-score standardization is used to make all terms dimensionless before weighted summation.

[0042] The end-to-end delay is defined as: Calculate the time for the terminal; It is the sum of the terminal-to-edge transmission time and the edge computing time; It is the sum of the terminal-to-cloud transmission time and the cloud computing time.

[0043] The probability of the task failing is:

[0044] in, To calculate the total number of tasks, For the first The priority weight of each task. For network packet loss rate, For the first The number of data packets that need to be transmitted for each task.

[0045] Under constraints and The multi-constraint optimization model is then solved to obtain the optimal execution position for each computational task. For execution location Available computing power For the first The computing power required for each task To meet the minimum bandwidth requirements, This represents the currently available bandwidth.

[0046] Computational tasks are allocated based on the optimal execution location. When a computational task is allocated to a terminal, the terminal executes the task and outputs the result. When a computational task is allocated to the edge, the terminal uploads the task data to the edge gateway, the edge gateway executes the task and sends the result back to the terminal. When a computational task is allocated to the cloud, either the terminal or the edge gateway uploads the task data to the cloud server, the cloud server executes the task and sends the result back.

[0047] In this embodiment, to adapt to dynamic changes in network conditions, the multi-constraint optimization model is periodically re-solved, with a period of 1 to 10 seconds. When the network packet loss rate exceeds a threshold or the network latency exceeds a threshold, immediate scheduling is triggered to re-solve the multi-constraint optimization model and adjust the execution position of the computation task.

[0048] In this embodiment, to ensure the execution of high-priority tasks, different weight coefficients are assigned to tasks of different priorities when solving the multi-constraint optimization model. The weight coefficients of high-priority tasks are greater than those of low-priority tasks. When resources are insufficient to simultaneously satisfy all tasks, the resource requirements of high-priority tasks are prioritized.

[0049] Step S6: Detect network connection status and trigger disconnection self-healing. The edge gateway periodically sends heartbeat messages to the cloud server every 3 seconds. Upon receiving a heartbeat message, the cloud server returns an acknowledgment message to the edge gateway. The edge gateway records the timestamp of the received acknowledgment message and calculates the difference between the current time and the timestamp of the most recent received acknowledgment message to obtain the heartbeat timeout duration. When the heartbeat timeout duration exceeds a timeout threshold of 9 seconds, the network connection is considered interrupted.

[0050] When a network connection is determined to be interrupted, the edge gateway loads a pre-cached decision strategy and a lightweight decision model from its local storage. The pre-cached decision strategy is determined based on the current time and historical operational data. The control command generation process is then switched to a local autonomous mode, in which the edge gateway independently generates control commands based on the pre-cached decision strategy and the lightweight decision model, without relying on a cloud server.

[0051] It should be noted that in the local autonomous mode, the edge gateway manages the priority of resource allocation for computing tasks. High-priority tasks include at least one of security alarms, fire warnings, and equipment fault detection. The highest computing resources and communication channels are allocated to these high-priority tasks to ensure their successful execution.

[0052] It should be noted that the runtime data generated by the terminal and edge gateway is cached in a local time-series database. The local time-series database adopts a circular buffer structure, and when the buffer is full, it overwrites the oldest data record.

[0053] The edge gateway continues to periodically send heartbeat messages to the cloud server to detect when network connectivity has been restored. It determines that network connectivity has been restored upon receiving a confirmation message from the cloud server.

[0054] When the network connection is determined to be restored, the edge gateway sends a model version number to the cloud server. This model version number identifies the versions of the currently used edge decision model and lightweight perception model. The cloud server compares the received model version number with the latest model version number. If the two version numbers are inconsistent, the cloud server calculates the parameter difference between the latest model and the current local model, generates a model parameter difference packet, and sends it to the edge gateway. The edge gateway receives the model parameter difference packet and completes the model update according to the model update process in step S4.

[0055] The control command generation process is switched back to collaborative mode. In collaborative mode, the edge gateway receives optimization strategy parameters from the cloud server and generates control commands based on these parameters. It then reads cached runtime data from the local time-series database and uploads this data to the cloud server using a breakpoint resume upload method. The breakpoint resume upload records the timestamps of already uploaded data and resumes uploads from data after those timestamps to avoid duplicate uploads.

[0056] In this embodiment of the application, in order to speed up the model synchronization speed after network recovery, the edge gateway divides the model parameter differential packet into multiple data blocks when receiving the model parameter differential packet, and uses multiple transmission channels to download each data block in parallel. After all data blocks have been downloaded, they are merged and verified.

[0057] Technical effects of this embodiment This implementation deploys a lightweight perception algorithm to the terminal, a decision-making algorithm to the edge gateway, and an optimization algorithm to the cloud, achieving hierarchical processing of computational tasks. After extracting feature vectors, the terminal immediately discards the original data, transmitting only low-dimensional feature vectors, reducing data transmission volume to less than one-tenth of the original data volume and resolving bandwidth resource bottlenecks. The edge gateway generates control commands locally, with inference latency less than 100 milliseconds, avoiding second-level delays caused by data traveling to and from the cloud and resolving real-time issues. Through model distillation, quantization compression, and structured pruning, the number of model parameters is compressed to less than three percent of the original model, enabling the lightweight perception model to run on low-computing-power terminal chips and resolving the problem of computing power imbalance. When network connectivity is interrupted, the edge gateway switches to local autonomous mode, independently generating control commands based on pre-cached decision-making strategies, ensuring the execution of high-priority tasks and resolving the issue of fragile business continuity.

[0058] On the other hand, this invention also proposes a three-level distributed hierarchical autonomous intelligent control system consisting of: The terminal feature extraction module is used to acquire environmental data and extract feature vectors; The edge decision generation module is used to receive feature vectors and generate control commands; The data aggregation and upload module is used to aggregate runtime data and upload it to the cloud. The cloud-based strategy optimization module is used to optimize decision-making strategies and update models. The task dynamic scheduling module is used to dynamically schedule computing tasks to the terminal, edge, or cloud. The disconnection self-healing management module is used to detect network connection status and trigger disconnection self-healing.

[0059] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.

[0060] II. Detailed Explanation of Specific Applications 1. System Architecture and Dual-Algorithm Co-design This invention follows the principles of "proximity perception, hierarchical decision-making, and global cloud computing," deconstructing the traditional coupled end-to-end algorithm into a three-level collaborative architecture: End-Device Layer – Lightweight Perception and Feature Extraction: The end-device's task is no longer complete inference, but rather to run an extremely lightweight feature extraction model. This model is derived from the original perception model (such as YOLOv8) through a three-level lightweight pipeline, which can efficiently identify and output high-dimensional, low-redundancy key feature vectors such as "human body exists in the scene," "target detected and its skeletal key point coordinates extracted," and "specific audio-visual domain features captured," rather than the original video stream.

[0061] Edge Layer – Local Real-Time Inference and Control: The edge gateway acts as the intelligent core within a region (such as a home or an intersection), undertaking millisecond-level real-time control tasks. It deploys a hybrid decision-making model consisting of a lightweight neural network (for complex pattern recognition) and a rule engine (for deterministic logical judgment). The gateway receives feature vectors uploaded by terminals and locally executes complex scene linkages such as "if a family member is detected and the indoor illuminance is <50 lux, automatically turn on the living room light strip and adjust it to a warm color mode," or local dynamic scheduling such as "if the queue length of vehicles in phase A is > a threshold, extend the green light passage time."

[0062] Cloud Layer – Global Optimization and Strategy Evolution: The cloud no longer processes raw data streams, but instead receives anonymized statistical data (such as regional traffic flow trends and device health logs) periodically aggregated and reported by edge gateways. On top of this, the cloud runs global optimization models such as reinforcement learning to conduct in-depth analysis of long-term decisions such as macro-level scheduling of cross-regional traffic flow and energy demand response across households. As the "strategy brain" and "model factory," the cloud continuously trains and optimizes global strategies, and incrementally distributes the lightweight model parameters to the edge and terminals, forming an intelligent closed loop.

[0063] 2. Module and Interface Design (1) Module Design Dual-algorithm decomposition module: The perception algorithm is decoupled from the decision / optimization algorithm. The perception algorithm is deployed on the terminal after being lightweighted, while the decision / optimization algorithm is split into edge and cloud according to complexity, and the computational boundaries and input / output of each module are clearly defined.

[0064] Lightweight Model Module: Integrates three techniques: model distillation, quantization compression, and pruning and sparsification. Model distillation: Using a large cloud-based model as the "teacher," a lightweight "student model" is trained, transferring complex dual-algorithm knowledge to the small model, maintaining accuracy while significantly reducing size.

[0065] Quantization compression: Compresses model parameters from 32-bit floating-point to 8-bit integer, reducing storage and computational overhead and adapting to low-performance chips.

[0066] Pruning and sparsification: Remove redundant neurons and connections in the model, retain the core feature extraction capability, and further improve inference speed.

[0067] Edge-Cloud Collaborative Inference Module: Defines the task scheduling and data interaction process between three layers of devices. 1. Terminal trigger: The terminal runs a lightweight perception algorithm to complete preliminary detection (such as recognizing personnel movement).

[0068] 2. Edge decision-making: The terminal uploads feature data to the edge gateway, and the edge side runs a decision-making algorithm to generate a local response (such as automatically turning on the lights).

[0069] 3. Cloud optimization: The edge gateway aggregates data and uploads it to the cloud. The cloud runs optimization algorithms, updates model parameters, and distributes them to the edge and terminals.

[0070] Communication and Resource Management Module: Prioritizes data transmission, allocates bandwidth, and schedules computing power between the management terminal, edge, and cloud to ensure the continuity of core services in weak network environments.

[0071] (2) Interface design Terminal-edge interface: Supports efficient transmission of feature data, adopts lightweight communication protocols (such as MQTT), and has a transmission rate of no less than 10Mbps to ensure real-time uploading of sensing results.

[0072] Edge-to-cloud interface: Supports high-speed transmission of aggregated data and model parameters, using TCP / IP protocol with a transmission rate of no less than 100Mbps, ensuring the timeliness of cloud-based optimization and iteration.

[0073] Model delivery interface: The cloud-based lightweight model is delivered to the edge and terminals via incremental updates, with an update time of no more than 5 minutes to avoid business interruption.

[0074] Reliability Interface: The modules employ heartbeat and retransmission mechanisms to ensure the stability and integrity of data transmission, adapting to network fluctuation scenarios.

[0075] 3. Implementation of Intelligent Control Algorithms (1) Detailed explanation of the dual-algorithm lightweight pipeline: After decoupling the "perception" and "decision" algorithms, this invention adopts a three-stage pipeline of "model distillation → quantization compression → structured pruning" for the perception model that needs to be deployed on the terminal, thereby achieving an order-of-magnitude reduction in model size and computational load.

[0076] Level 1: Knowledge Distillation A high-precision, large-scale model pre-trained on a cloud GPU cluster is used as the teacher model (e.g., YOLOv8x, 87.7M parameters), and a native, lightweight but less accurate network is used as the student model (e.g., MobileNetV3-Small, 2.5M parameters). The distillation process not only teaches the student model the hard labels (Ground Truth) of the training data but also the soft labels (Probability Distributions of Each Class) output by the teacher model, thereby transferring the generalization ability of the teacher model. The loss function is defined as:

[0077] in The teacher model is in temperature The output soft tags, This is the output logits of the teacher model. By adjusting... With hyperparameters, we were able to reduce the number of model parameters from 87.7M to 2.5M while maintaining an accuracy decrease of less than 2%, resulting in a volume reduction of over 97%. However, the computational load at this point is still too high for chips with less than 1 TOPS.

[0078] Level 2: Quantization Furthermore, we linearly map the weights and activation values ​​of the distilled student model from FP32 (32-bit floating-point) space to INT8 (8-bit integer) space, which is an affine transformation:

[0079] Where r is a real value, q is the quantized integer value, S is the scaling factor, and Z is the zero point. To minimize accuracy loss, we use a calibration set for quantization error calibration, and statistically analyze the numerical distribution range of each layer's tensors by traversing the data. It automatically calculates the optimal S and Z. After the model was upgraded from FP32 to INT8, the model size was reduced by about 75%, and the INT8 calculation can be accelerated by hardware using the NPU (Neural Processing Unit) instruction set on the chip, improving the energy efficiency by several times.

[0080] Level 3: Structured Pruning Finally, we perform channel-level structured pruning on the quantized INT8 model to remove redundant channels with low contribution, achieving real hardware acceleration. We do this by analyzing the L1 norm of the weights in each convolutional channel. As an importance criterion, a dynamic sparse pruning strategy is adopted, which allows for differences in sparsity between different layers while maintaining the overall sparsity, and restores accuracy through retraining (fine-tuning). After pruning, the computational cost (FLOPs) and size of the model are further compressed by 40%-60%, ultimately reducing the overall size of the perception model by 75%~82% (from 87.7M to less than 3MB), enabling it to run efficiently at >30FPS on terminal chips with computing power <1 TOPS.

[0081] (2) Implementation of intelligent control and dynamic scheduling algorithms To achieve a dynamic optimal balance between system resources and performance, a multi-constraint adaptive scheduling model and a hierarchical autonomous strategy are constructed.

[0082] Multi-constraint collaborative reasoning scheduling model Task scheduling is defined as a constrained optimization problem, with the goal of finding the execution location (terminal d, edge e, or cloud c) that minimizes latency and maximizes resource utilization. The mathematical model is as follows: Objective function:

[0083] in, This represents the execution position of the i-th task; It is the end-to-end latency at that execution location, for the terminal. For the edges, Regarding the cloud, .

[0084] It is the resource pooling utilization rate of the execution node.

[0085] It is a penalty item based on the current packet loss rate. and task priority calculate: , where k is the number of data packets that the task needs to transmit.

[0086] These are weighting factors, which are dynamically configured by the administrator based on the business scenario.

[0087] Constraints:

[0088]

[0089] Hierarchical autonomous control and disconnection self-healing strategy When the MQTT long-lived connection heartbeat between the edge gateway and the cloud times out (heartbeat cycle 3 seconds, timeout threshold 9 seconds), it is determined to be in a network outage state and immediately triggers the "disconnection self-healing" mode. The edge gateway takeover process is as follows: 1. Loading Pre-Cached Policies: The gateway immediately reads a pre-cached rule base and lightweight decision model from local flash storage that matches the current scenario (identified by time and historical data). For example, the smart intersection edge gateway loads a "flashy yellow" or "all red" safety fallback control policy.

[0090] 2. Data Bus Switching: The data stream originally directed to the cloud is switched to a local time-series database (such as InfluxDB) for high-speed caching. All generated decision commands are directly sent to the terminal for execution via the local MQTT Broker, ensuring that loopback latency remains <100ms.

[0091] 3. Priority Degradation Execution: For high-priority safety linkage tasks (such as fire alarm valve shut-off), maintain the highest computing resources and communication channels to ensure 100% execution.

[0092] 4. Connection Restoration and Data Retransmission: Once the gateway detects that the MQTT connection with the cloud has been restored, it first performs model version alignment. A "parameter differential update and version hash verification" mechanism is used: the cloud compares the model version hash values ​​(SHA-256) of the local and gateway versions, and only sends incremental differential packets (Patches). The gateway applies the patch and re-verifies the hash; after confirming consistency, it switches back to collaborative mode. During this process, historical operation logs and sensor data cached locally will be asynchronously transmitted back to the cloud via a breakpoint resume method for post-event analysis and updating the historical case library.

[0093] (3) Data and model iteration process Feature transmission optimization: After the terminal completes the perception process, the original data is immediately discarded, and only the extracted low-dimensional key feature vectors (such as JSON format) are transmitted.

[0094] Encapsulate and upload data. Depending on the urgency of the business, the transport layer adaptively switches protocols: status and log data are transmitted via TCP (such as MQTT QoS 1 / 2) to ensure reliable transmission; alarm feature vectors with extremely high real-time requirements are sent quickly via UDP (or UDP-based MQTT-SN), tolerating a very small amount of packet loss in exchange for extremely low latency.

[0095] Incremental Model Updates and Version Management: After training a new model in the cloud, the entire model file is not distributed. Instead, a Delta Patch (XOR difference packet) is calculated between the parameter matrices of the old and new models. The distribution process strictly follows: The cloud distributes update_manifest.json (containing the new version number v1.2, patch size, and SHA-256 checksum) → The edge gateway confirms receipt and downloads the Delta Patch → Locally, the candidate model v1.2_candidate is generated by merging the two packages → The SHA-256 value is calculated and compared → If the verification passes, the old model file is atomically replaced and the inference process is restarted; if the verification fails, the system automatically rolls back to version v1.1 and reports an alert. This mechanism ensures strong consistency of the distributed system during continuous evolution.

[0096] Beneficial (1) Improved real-time performance: Key sensing and decision-making tasks are completed on the edge side, reducing response latency from seconds to milliseconds, meeting the real-time requirements of IoT services.

[0097] (2) Reduced bandwidth usage: Only feature data is transmitted instead of raw data, reducing network traffic by more than 90%, making it suitable for low-bandwidth IoT environments.

[0098] (3) Enhanced reliability: When the cloud connection is lost, the edge and terminal can still independently complete core tasks, avoiding system paralysis and improving business continuity.

[0099] (4) Cost optimization: Reduce the computing power requirements of terminal devices and adopt more economical hardware solutions; the cloud only needs to undertake optimization tasks, and server resource consumption is reduced by more than 50%.

[0100] (5) Enhanced scalability: Modular architecture and algorithm decomposition facilitate the addition of new perception / decision algorithms and adapt to multiple scenarios such as smart home and smart city.

[0101] The advantages of the algorithms are compared in Table 1: Table 1: Comparison of Algorithm Advantages

[0102] III. Experimental Data 1. Experimental Environment and Testing Methods This invention constructs a highly simulated, full-element verification environment in two typical scenarios: smart home and smart city.

[0103] Physical Topology: An experimental bed was built consisting of 20 terminal nodes (based on ARM Cortex-M7 and Cortex-A53 core boards, simulating IP cameras and environmental sensors), 5 edge gateways (based on NVIDIA Jetson Orin NX, with a computing power of 70 TOPS, simulating home gateways / roadside units) and 1 cloud server cluster (simulator).

[0104] Network Simulation: Using the WANem network impairment tester, a realistic and complex network environment is accurately simulated between the edge gateway and the cloud server. Preset parameters include bandwidth fluctuations of 1Mbps-100Mbps, random packet loss rates of 0%-30%, network jitter of ±50ms, and transient network outages lasting from several minutes to several hours.

[0105] Stress testing scheme: A continuous 72-hour stress and stability test was conducted. A mixed load containing 80% regular events (timed uploads) and 20% emergency events (security alarms, traffic accidents) was continuously injected through a script to comprehensively test the system's performance under various extreme communication conditions.

[0106] 2. Quantitative comparative data analysis A strict control group (traditional cloud-based centralized architecture) and an experimental group (the distributed architecture of this invention) were established, and the performance index matrix comparison is as follows: Real-time leap: such as Figure 1 As shown, the average end-to-end control latency of the experimental group decreased from 2.3 seconds (standard deviation ± 1.5 s) in the control group to 85 milliseconds (standard deviation ± 22 ms). The advantage was even more pronounced at the 99th percentile latency (i.e., tail latency), decreasing from 5.8 s in the control group to 150 ms. This demonstrates the system's performance stability under peak load conditions and fully meets the requirements for millisecond-level real-time control.

[0107] Network and Cost Optimization: Uplink bandwidth consumption was reduced by 92.8%, from a peak of 85Mbps to below 6Mbps. This reduced monthly traffic operation costs from several thousand yuan to the hundreds of yuan in remote 4G / 5G weak coverage scenarios. Since the terminal only needs to run lightweight models, the high-performance GPU core can be replaced with a low-cost MCU / SoC with a built-in NPU, reducing the hardware BOM cost per node by 60%. Because cloud servers only handle aggregated data and model training, the number of GPU computing instances rented / purchased can be reduced to one-third of the original, reducing resource consumption by 62.4%.

[0108] High business reliability: In a simulated 30-minute network outage test, all linkage functions of the control group system instantly failed, and business availability dropped to 0. The experimental group, on the other hand, smoothly switched to edge autonomous mode, and all core linkages (such as security alarms and traffic light security status switching) resumed execution within 120ms, with core business availability reaching over 98.5%.

[0109] IV. Key Points of the Invention 1. Extraction of core innovation points Distributed Three-Tier Collaborative Architecture: The core innovation of this invention lies in proposing and implementing a three-tiered distributed intelligent control architecture: "terminal characterization - edge real-time processing - cloud globalization." This breaks through the centralized paradigm of traditional cloud computing's "edge data acquisition - cloud computing," creatively decentralizing perception and decision-making intelligence on demand. It fundamentally solves a series of interconnected bottlenecks such as latency, bandwidth, computing power, and reliability, achieving a revolutionary improvement in system-level computing resource utilization and service quality.

[0110] Dual Decoupling of Perception and Decision-Making and Multi-Level Lightweight Technology: This invention is the first to apply the concept of "decoupling" from industrial control logic to AI model deployment, explicitly separating the "perception" and "decision-making" algorithms. It also innovatively proposes a three-stage pipeline composite technology solution of "distillation-quantization-pruning," rather than using a single technology. This achieves extremely high compression efficiency (75%-82% volume reduction) and negligible accuracy loss, paving the way for large models to be deployed on low-power IoT chips with computing power below 1 TOPS. This is the core technological support of this invention.

[0111] Self-healing dynamic task scheduling mechanism: This invention not only defines the optimal collaborative reasoning scheduling model under normal conditions, but more importantly, protects its "disconnection self-healing" autonomous logic under extreme weak network / disconnection conditions. This scheduling strategy, which uses "computing power-bandwidth-priority" as multiple constraints, can achieve a smooth switch of control from the cloud to the edge, and is a unique invention for achieving industrial-grade high availability (98.5%) business continuity.

[0112] Model Consistency Collaborative Iteration Scheme: The synchronization mechanism based on parameter differential update and version hash verification proposed in this invention is key to ensuring the safe, efficient, and consistent iteration of models in large-scale distributed systems. This mechanism solves the problem of continuous operation and maintenance of AI models on heterogeneous, distributed nodes with minimal bandwidth overhead and strong verification logic, representing a closed-loop innovation that ensures the system's continued vitality.

[0113] 2. Summary of Economic and Social Benefits The large-scale application of this invention will generate enormous economic and social value. Economically, by lightweighting terminals and offloading cloud loads, it can help industries such as smart cities and the Industrial Internet of Things save 30%-50% on infrastructure investment (CapEx), while significantly reducing operating costs (OpEx) from bandwidth leasing and power consumption. Socially, by deploying distributed nodes, it significantly reduces the computing energy consumption of cloud data centers. Simultaneously, its inherent high reliability ensures the safe operation of critical systems such as transportation and security, possessing profound macro-value for protecting people's lives and property and improving social governance.

[0114] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling, characterized in that, Includes the following steps: The terminal device collects environmental data through sensors, inputs the environmental data into a lightweight perception model, outputs a feature vector, and discards the environmental data. The edge gateway receives the feature vector, inputs the feature vector into the edge decision model, and generates control commands; The edge gateway periodically performs de-identification processing and data aggregation on the operational data, generates aggregated data, and uploads it to the cloud server; The cloud server trains a global optimization model based on the aggregated data, generates optimized decision strategy parameters, and calculates model parameter difference packets to be sent to the edge gateway and the terminal device. Based on the current system state, a multi-constraint optimization model is constructed to solve the optimal execution position of each computing task, and the computing tasks are allocated to the terminal, edge or cloud for execution. The edge gateway periodically sends heartbeat messages to the cloud server. When the heartbeat timeout duration exceeds the timeout threshold, it determines that the network connection is interrupted, switches to local autonomous mode to independently generate control commands, and synchronizes the model and switches back to collaborative mode when the network connection is restored.

2. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 1, characterized in that, The lightweight perception model is obtained by sequentially performing model distillation, quantization compression, and structured pruning on the original perception model. The model distillation uses the original perception model as the teacher model and the native lightweight network as the student model. Training is performed by jointly optimizing the hard label loss and the soft label loss. The hard label loss is calculated based on the predicted output of the student model and the true label, and the soft label loss is calculated based on the predicted output of the student model and the soft label output of the teacher model under the temperature parameter. The two losses are weighted and summed to obtain the total loss.

3. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 2, characterized in that, The quantization compression maps model parameters and activation values ​​from floating-point numbers to integers. It uses a calibration dataset to statistically analyze the numerical range of tensors in each layer. It calculates a scaling factor based on the ratio of the difference between the maximum and minimum values ​​of the numerical range to the integer representation range. It calculates the zero-point offset based on the difference between the minimum value represented by the integer and the minimum value of the numerical range divided by the scaling factor. The structured pruning calculates the weights of each convolutional channel. The norm is used as an importance measure, and channels whose importance measure is less than a threshold are removed.

4. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 1, characterized in that, The edge decision model includes a lightweight neural network and a rule engine. The lightweight neural network takes the feature vector as input, and the output layer uses the softmax activation function to output the probability distribution of each pattern category. The category with the highest probability is taken as the pattern recognition result. The rule engine generates the control command based on the pattern recognition result and preset rules. The preset rules include condition judgment logic and action mapping relationship.

5. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 1, characterized in that, The edge gateway aggregates feature vectors uploaded by multiple terminals in the same area, extracts spatiotemporal correlation features, inputs the spatiotemporal correlation features into the edge decision model, and generates global control commands for the area.

6. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 1, characterized in that, The objective function of the multi-constraint optimization model is to minimize the weighted sum of the end-to-end latency of each computational task, the reciprocal of the weighted sum of resource utilization, and the task failure probability. The end-to-end latency is standardized to eliminate dimensional differences before being substituted into the objective function. The task failure probability is calculated based on the priority weight of each task, the network packet loss rate, and the number of data packets to be transmitted by each task. The constraints are that the available computing power at the execution location is not less than the computing power required by the task, and the currently available bandwidth is not less than the minimum bandwidth requirement.

7. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 1, characterized in that, The multi-constraint optimization model is periodically resolved, with a solution period of 1 to 10 seconds. When the network packet loss rate exceeds a threshold or the network latency exceeds a threshold, immediate scheduling is triggered and the multi-constraint optimization model is resolved.

8. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 1, characterized in that, In the local autonomous mode, the edge gateway loads pre-cached decision strategies and lightweight decision models from local storage to independently generate control commands and caches the running data in a local time-series database with a ring buffer structure. When the network connection is restored, the edge gateway sends the model version number to the cloud server. When the model version number is inconsistent with the latest model version number, it receives the model parameter differential packet to complete the model update and uploads the running data cached in the local time series database to the cloud server through the breakpoint resume method.

9. The three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of end-edge-cloud as described in claim 1, characterized in that, The model parameter differential packet includes a version number, differential data, and a checksum. The checksum is calculated using a hash algorithm. After receiving the model parameter differential packet, the edge gateway and the terminal device merge the differential data with the local current model parameters to generate a candidate model, calculate the hash value of the candidate model, and compare it with the checksum. If they match, the candidate model replaces the local current model. If they do not match, the local current model remains unchanged, and an update failure notification is sent to the cloud server.

10. A three-level distributed hierarchical autonomous intelligent control system (edge-cloud), used to execute the three-level distributed hierarchical autonomous intelligent control architecture and collaborative scheduling of claims 1 to 9, characterized in that, include: The terminal feature extraction module is used to collect environmental data through sensors, input the environmental data into a lightweight perception model to output a feature vector, and discard the environmental data. An edge decision generation module is used to receive the feature vector and input the feature vector into the edge decision model to generate control commands. The data aggregation and upload module is used to periodically de-identify and aggregate the running data, generate aggregated data, and upload it to the cloud server. The cloud-based strategy optimization module is used to train a global optimization model based on the aggregated data, generate optimized decision strategy parameters, and calculate model parameter difference packets to be sent to edge gateways and terminal devices. The task dynamic scheduling module is used to construct a multi-constraint optimization model based on the current system state, solve the optimal execution position of each computing task, and allocate the computing tasks to the terminal, edge, or cloud for execution. The disconnection self-healing management module is used to periodically send heartbeat messages to the cloud server. When the heartbeat timeout duration exceeds the timeout threshold, it determines that the network connection is interrupted and switches to local autonomous mode to independently generate control commands. When the network connection is restored, it synchronizes the model and switches back to collaborative mode.

Images