Intelligent gateway device based on AI edge computing

Abstract

The invention belongs to the technical field of Internet of Things, and discloses intelligent gateway equipment based on AI edge computing, which comprises a main processor, an AI acceleration processor, a self-adaptive task scheduling engine, a heterogeneous resource dynamic arrangement module, a self-adaptive power consumption management module and a two-way model verification and execution module running in a hardware trusted execution environment. Wherein the task scheduling engine queues according to task priorities; the resource arrangement module accurately schedules tasks to the CPU or the AI processor according to task characteristics; the power consumption management module dynamically adjusts power consumption according to the load state of the task queue; and the security module performs source and function dual verification on the AI model before execution. Through deep collaboration and closed-loop control of the four modules, the high real-time processing performance, the extreme energy efficiency ratio and the firm safety guarantee of the intelligent gateway equipment are realized.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology, specifically relating to a smart gateway device based on AI edge computing. Background Technology

[0002] With the rapid development of Internet of Things (IoT) and Artificial Intelligence (AI) technologies, edge computing gateways, as the core hub connecting physical world terminal devices with the cloud digital world, are facing increasingly higher functional and performance requirements. Modern edge computing gateways are no longer merely data relay stations, but are endowed with powerful local computing capabilities, particularly through the integration of high-performance main processors (CPUs) and dedicated AI accelerators (such as TPUs and NPUs), enabling them to perform complex AI inference tasks near the data source. This architecture effectively reduces network latency, saves transmission bandwidth, and enhances data privacy.

[0003] In the process of realizing this invention, the inventors discovered that existing edge computing gateways still exhibit several technical bottlenecks in practical applications:

[0004] 1. The CPU and AI accelerator processor inside the gateway are typical heterogeneous computing resources, but most systems lack a fine-grained collaborative scheduling mechanism. When multiple tasks such as AI inference, data transformation, and network communication are performed concurrently, indiscriminate resource contention often leads to delays in response to time-sensitive critical AI tasks (such as real-time quality inspection on the production line), failing to meet the stringent requirements of industrial scenarios.

[0005] 2. Traditional power management strategies often rely on passive adjustments based on temperature or general CPU utilization, failing to accurately match the actual peak and trough of business load. This results in gateways operating at higher power consumption during idle periods, leading to energy waste; while during peak periods, simple power limits may prevent the hardware from reaching its full peak performance.

[0006] 3. AI models on edge gateways are typically distributed and updated remotely from the cloud. During this process, the models may be at risk of being tampered with or replaced. Traditional digital signature verification alone is insufficient to prevent more sophisticated attacks (such as model poisoning attacks). Once a malicious AI model is deployed, it could lead to flawed production decisions, system crashes, or even security incidents, with serious consequences. Summary of the Invention

[0007] This invention aims to address, at least to some extent, the problems of low resource scheduling efficiency, mismatch between power consumption and performance, and insufficient security and reliability of AI model execution in existing technologies. Therefore, the purpose of this invention is to provide an intelligent gateway device based on AI edge computing, which can intelligently coordinate internal heterogeneous resources, achieve a dynamic optimal balance between performance and power consumption, and establish a trusted end-to-end environment from model deployment to execution.

[0008] To achieve the above objectives, the present invention provides an intelligent gateway device based on AI edge computing, comprising: a main processor and an AI accelerator processor; an adaptive task scheduling engine configured to identify the type and priority of incoming tasks according to preset rules and allocate the tasks to one of multiple priority queues; a heterogeneous resource dynamic orchestration module connected to the adaptive task scheduling engine, configured to obtain tasks from the multiple priority queues and schedule the tasks to be executed on the main processor or the AI ​​accelerator processor according to the computational characteristics of the tasks; an adaptive power management module connected to the adaptive task scheduling engine, configured to dynamically adjust the power consumption mode of at least one of the main processor and the AI ​​accelerator processor according to the real-time load status of the multiple priority queues; and a dual-path model verification and execution module running in a hardware trusted execution environment, configured to perform dual-path verification of the AI ​​model before the AI ​​accelerator processor executes the AI ​​model.

[0009] Furthermore, the multiple priority queues include: a real-time queue for processing real-time AI inference tasks, a near-real-time queue for processing data preprocessing or result aggregation, and a non-real-time queue for processing data reporting or system logs.

[0010] Furthermore, the heterogeneous resource dynamic orchestration module is specifically configured to schedule AI model inference tasks that are determined to require large-scale parallel computing to the AI ​​acceleration processor for execution, and to schedule tasks that are determined to involve complex logic control or serial processing to the main processor for execution.

[0011] Furthermore, the adaptive power management module is specifically configured to: switch the power consumption mode of the AI ​​acceleration processor to high-performance mode when the load of the real-time queue is higher than a first preset threshold; and switch the main processor and the AI ​​acceleration processor to low-power mode when the load of all priority queues is lower than a second preset threshold.

[0012] Furthermore, the hardware trusted execution environment is a secure region built based on processor hardware isolation technology.

[0013] Furthermore, the dual-path verification includes: a first-path verification that verifies the digital signature of the AI ​​model, and a second-path verification that performs functional reasoning tests on the AI ​​model using preset security test data within the trusted execution environment of the hardware; the AI ​​model is only allowed to execute after both the first-path verification and the second-path verification have passed.

[0014] Furthermore, the adaptive task scheduling engine and the heterogeneous resource dynamic orchestration module work together to ensure that tasks in the real-time queue are scheduled with priority over tasks in other queues.

[0015] Furthermore, the adaptive task scheduling engine is configured to periodically or event-triggeredly notify the adaptive power management module of the load status information of the multiple priority queues.

[0016] Furthermore, the smart gateway device also includes at least one network interface, at least one RS-485 interface, and at least one digital input interface for collecting data from external devices or networks as the source of the incoming task.

[0017] Furthermore, the output information of the adaptive task scheduling engine serves as the input to the heterogeneous resource dynamic orchestration module and the adaptive power management module, and the execution of all AI models is controlled by the dual-path model verification and execution module, thereby forming an integrated closed-loop processing system.

[0018] The beneficial effects of this invention are as follows:

[0019] This invention, through the collaboration of task priority queues and precise orchestration of heterogeneous resources, ensures that high-priority AI tasks always receive the most suitable computing resources (AI accelerator processors) first, greatly reducing the processing latency of critical business operations and improving the effective throughput of the entire smart gateway device.

[0020] This invention's power management strategy is directly linked to the front-end service load, enabling on-demand power supply. The system operates at full capacity when services are busy and enters deep sleep mode when services are idle, avoiding energy waste and minimizing long-term operating costs while ensuring performance.

[0021] This invention employs a dual-path model verification mechanism based on hardware TEE, which not only prevents models from unknown sources but also detects maliciously tampered models that have been signed. It constructs a full-link trusted chain from cloud distribution to edge execution, ensuring the reliability and security of edge decision-making.

[0022] This invention does not simply stack four functions together, but rather forms an intelligent, adaptive internal management operating system by defining clear data and control flows between them. The overall performance, energy efficiency, and security improvements resulting from this system-level innovative design are unattainable through independent optimization of individual modules in existing technologies. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the system architecture of an AI edge computing-based smart gateway device according to the present invention.

[0024] Figure 2 This is a schematic diagram of the task scheduling and resource orchestration process in this invention.

[0025] Figure 3 This is a schematic diagram of the dual-path verification process for the AI ​​model in this invention. Detailed Implementation

[0026] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] like Figure 1 As shown in the figure, this invention provides an AI edge computing-based smart gateway device. Its hardware layer may include a main processor, an AI accelerator processor, memory, and various physical interfaces for data acquisition and communication. The main processor uses a multi-core ARM Cortex-A series CPU; the AI ​​accelerator processor uses a dedicated TPU or NPU, providing computing power of several to tens of TOPS; the memory includes RAM and embedded storage, with LPDDR4 RAM and eMMC flash memory; the physical interfaces include a Gigabit Ethernet port, an RS-485 serial port, a USB port, and digital / analog input / output interfaces.

[0028] This invention also includes four collaborative software modules running on the main processor, which together constitute the intelligent management core of the smart gateway device. These four modules include an adaptive task scheduling engine, a heterogeneous resource dynamic orchestration module, an adaptive power management module, and a dual-path model verification and execution module.

[0029] like Figure 2As shown, when a new task arrives at the smart gateway device through a physical interface (e.g., receiving a frame of video data from a camera connected to the network port), the adaptive task scheduling engine intervenes first. This engine identifies task attributes based on a preset set of rules. For example, a rule can be defined as:

[0030] All RTSP video stream parsing tasks originating from port 8554 are classified as real-time AI inference tasks, have the highest priority, and are placed in the real-time queue.

[0031] Sensor reading tasks originating from the RS-485 interface and using Modbus RTU data format are classified as near real-time data processing tasks, with medium priority, and are placed in the near real-time queue.

[0032] The tasks of the smart gateway device itself, which are used to send heartbeat packets to the cloud platform, are determined to be non-real-time tasks, have the lowest priority, and are placed in the non-real-time queue.

[0033] The adaptive task scheduling engine continuously monitors the status of the three queues and strictly ensures that the heterogeneous resource dynamic orchestration module always prioritizes tasks from the real-time queues for processing, achieving absolute priority protection for critical business operations. Simultaneously, the adaptive task scheduling engine periodically (every 100 milliseconds) or when queue lengths change significantly, broadcasts load status information such as the current length of each queue and the average task waiting time to the adaptive power management module as a basis for its decision-making.

[0034] The heterogeneous resource dynamic orchestration module is the computing power allocation center. After retrieving the highest priority task from the queue of the adaptive task scheduling engine, it further analyzes its computational characteristics.

[0035] If the task is "to perform face detection on a single frame of 1080p video", the heterogeneous resource dynamic orchestration module recognizes this as a typical AI computationally intensive task suitable for large-scale parallel processing. Therefore, it packages the task, data pointers, and the required face detection model information together and schedules them to the AI ​​accelerator processor for execution.

[0036] If the task is to "parse Modbus messages and execute an IF-THEN-ELSE control logic based on their values," the heterogeneous resource dynamic orchestration module recognizes that this is a task primarily involving logical judgments and serial processing, unsuitable for execution on a parallel computing unit. Therefore, it schedules the task to a general-purpose computing core on the main processor for execution.

[0037] This precise division avoids using expensive AI computing power to process simple logic and also avoids using general-purpose CPUs to inefficiently simulate parallel computing, thereby maximizing the utilization of hardware resources of the entire smart gateway device.

[0038] The adaptive power management module receives real-time queue load information from the adaptive task scheduling engine and executes a power consumption strategy tightly coupled with the service requirements. This power consumption strategy can be specifically described as follows:

[0039] 1. High-performance mode trigger: The adaptive power management module sets a first preset threshold, such as "the real-time queue length exceeds 5 tasks for 1 second". Once this condition is met, it means that there is a large backlog of real-time AI tasks. The adaptive power management module will immediately send an instruction to the hardware driver to increase the frequency and voltage of the AI ​​acceleration processor to the highest performance mode to ensure that the backlog of tasks is cleared quickly.

[0040] 2. Low-power mode trigger: The adaptive power management module sets a second preset threshold, such as "the length of all three queues (real-time, near real-time, and non-real-time) remains 0 for 10 seconds." This indicates that the smart gateway device is in a completely idle state. At this time, the adaptive power management module triggers a deep power-saving strategy, putting multiple cores of the main processor into sleep mode and significantly reducing the clock frequency of the AI ​​accelerator processor or even gating its power supply, thereby achieving the lowest power standby.

[0041] 3. Between the two extreme states mentioned above, the adaptive power management module can also make smoother power level adjustments (DVFS, Dynamic Voltage Frequency Adjustment) based on the continuous change of queue length.

[0042] The dual-path model verification and execution module runs in a trusted execution environment (TEE) built on processor hardware isolation technology (such as ARM TrustZone). Figure 3 As shown, when a new AI model is delivered from the cloud, the dual-path model verification and execution module is configured to execute a dual-path verification process:

[0043] First, the module performs the first-way verification, namely signature verification. Within the TEE, it uses a pre-configured trusted public key to decrypt and verify the digital signature of the model file stored in memory. This function ensures that the model indeed originates from a trusted cloud repository and has not been tampered with during transmission. If the signature is invalid, the module rejects subsequent operations and reports a security alert.

[0044] Next, after successful signature verification, the module performs a second verification: functional testing. It loads the AI ​​model within the TEE's secure sandbox environment and uses a pre-set set of harmless security test data for inference. For example, for a quality inspection model, the test data could be a standard image of a qualified product and a standard image of a defective product. The module checks whether the model's output falls within the expected range (e.g., the confidence level for the qualified product image should be higher than 0.9, and the confidence level for the defective product image should be lower than 0.1). This function aims to capture fake models that, while having a correct signature, have had their internal logic maliciously modified (model poisoning) to produce harmful outputs.

[0045] Ultimately, the module was configured to authorize and securely load the AI ​​model into the AI ​​accelerator's dedicated memory only after all the above dual-path verifications have passed, for subsequent real-time inference tasks to call upon.

[0046] This invention solves many pain points in the prior art through the deep coupling and closed-loop collaboration of the above four modules, and provides a high-performance, high-energy-efficiency, and high-security smart gateway device.

[0047] This invention is not limited to the above-described optional embodiments. Anyone can derive other various forms of products under the guidance of this invention. However, regardless of any changes made in their shape or structure, any technical solution that falls within the scope of the claims of this invention shall be protected by this invention.

Claims

1. A smart gateway device based on AI edge computing, characterized in that, include: One main processor and one AI accelerator processor; An adaptive task scheduling engine is configured to identify the type and priority of incoming tasks according to preset rules, and assign the tasks to one of multiple priority queues; A heterogeneous resource dynamic orchestration module is connected to the adaptive task scheduling engine and configured to obtain tasks from the multiple priority queues and schedule the tasks to be executed on the main processor or the AI ​​acceleration processor according to the computational characteristics of the tasks. An adaptive power management module, connected to the adaptive task scheduling engine, is configured to dynamically adjust the power consumption mode of at least one of the main processor and the AI ​​accelerator processor according to the real-time load status of the multiple priority queues. A dual-path model verification and execution module, running in a hardware trusted execution environment, is configured to perform dual-path verification of the AI ​​model before the AI ​​accelerator processor executes the AI ​​model.

2. The intelligent gateway device according to claim 1, characterized in that, The multiple priority queues include: a real-time queue for processing real-time AI inference tasks, a near-real-time queue for processing data preprocessing or result aggregation, and a non-real-time queue for processing data reporting or system logs.

3. The intelligent gateway device according to claim 1 or 2, characterized in that, The heterogeneous resource dynamic orchestration module is specifically configured to schedule AI model inference tasks that are determined to require large-scale parallel computing to the AI ​​acceleration processor for execution, and to schedule tasks that are determined to involve complex logic control or serial processing to the main processor for execution.

4. The intelligent gateway device according to claim 2, characterized in that, The adaptive power management module is specifically configured to: switch the power mode of the AI ​​accelerator to high-performance mode when the load of the real-time queue is higher than a first preset threshold; and switch the main processor and the AI ​​accelerator to low-power mode when the load of all priority queues is lower than a second preset threshold.

5. The intelligent gateway device according to claim 1, characterized in that, The trusted execution environment is a secure region built based on processor hardware isolation technology.

6. The intelligent gateway device according to claim 1 or 5, characterized in that, The dual-path verification includes: a first-path verification that verifies the digital signature of the AI ​​model, and a second-path verification that performs functional reasoning tests on the AI ​​model using preset security test data within the trusted execution environment of the hardware; the AI ​​model is only allowed to execute after both the first-path verification and the second-path verification have passed.

7. The intelligent gateway device according to claim 2, characterized in that, The adaptive task scheduling engine and the heterogeneous resource dynamic orchestration module work together to ensure that tasks in the real-time queue are scheduled with priority over tasks in other queues.

8. The intelligent gateway device according to claim 1, characterized in that, The adaptive task scheduling engine is configured to periodically or event-triggeredly notify the adaptive power management module of the load status information of the multiple priority queues.

9. The intelligent gateway device according to claim 1, characterized in that, It also includes at least one network interface, at least one RS-485 interface, and at least one digital input interface for collecting data from external devices or networks as the source of the incoming task.

10. The intelligent gateway device according to claim 1, characterized in that, The output information of the adaptive task scheduling engine serves as the input to the heterogeneous resource dynamic orchestration module and the adaptive power management module, and the execution of all AI models is controlled by the dual-path model verification and execution module, thereby forming an integrated closed-loop processing system.

Images