Intelligent fsu edge computing terminal dual-channel ai security control system

Abstract

The application provides a smart FSU edge computing terminal double-channel AI safety control system, relates to the technical field of intelligent control, and comprises a GD32F303 collection communication chip configured to collect and upload telemetry based on a priority-adjustable double communication interface and issue corresponding control instructions; an RK3568 master control unit configured to perform AI trend prediction based on the telemetry and generate intelligent control instructions; the control instructions are used to adjust the operating state of the computer room equipment; a RISC-V coprocessor configured to perform basic safety calculation based on the telemetry and generate safety control instructions when the AI trend prediction function of the RK3568 master control unit is in an abnormal state; the safety control instructions are used to adjust the operating state of the computer room equipment. The application solves the problem of lacking a computer room environment guarantee mechanism in the case of AI prediction failure through the above system.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a smart FSU edge computing terminal dual-path AI security control system. Background Technology

[0002] In the power and environmental monitoring systems of communication base stations, equipment rooms, and edge data centers, the field monitoring unit (FSU) is the core monitoring equipment. Traditional FSUs mainly perform basic data acquisition, alarm information uploading, and simple logic control. Their functions are relatively fixed and lack artificial intelligence analysis and flexible deployment capabilities.

[0003] With the advancement of intelligent communication infrastructure, smart FSUs (Fuel Supply Units) are gradually becoming a focus of industry technology exploration. Some solutions attempt to introduce edge computing chips into FSUs and leverage artificial intelligence algorithms and containerized deployment technologies to achieve advanced functions such as intelligent group control of air conditioning systems, environmental trend judgment, and energy efficiency optimization. However, the application of artificial intelligence and containerization technologies in FSU devices is still in its early exploratory stage, and a stable and reliable industry-wide solution has not yet been formed. In actual deployment and trial operation, such systems often face many non-steady-state operational risks, causing their intelligent control functions to easily fail, specifically in the following aspects: The containerized architecture has low maturity on the embedded platform used by FSU. The AI ​​containers that carry energy-saving control logic are prone to malfunctions, unresponsiveness, or competition for computing resources, which makes it impossible to issue control commands to equipment such as air conditioners. The adaptation of artificial intelligence models to the FSU edge environment is still in the experimental stage, and problems such as model loading failure, inference timeout, and abnormal output confidence often occur, which makes it impossible to generate effective control strategies. Intelligent FSUs often adopt a multi-container collaborative architecture that separates data acquisition, artificial intelligence, and management. However, the design of cross-container communication mechanisms and public service interfaces is not yet mature, which can easily lead to communication interruptions and command forwarding failures. In edge deployment environments, issues such as fluctuating network connectivity and abnormal data collection at the front end occur frequently. The AI ​​control module, which is still in the exploratory stage, lacks sufficient fault tolerance mechanisms and is prone to interrupting control output due to abnormal input data.

[0004] More importantly, the industry currently lacks a reliable safety fallback technology solution for the failure of intelligent FSU AI control systems. Traditional FSUs do not have artificial intelligence control functions, so there is no safety risk caused by the failure of the artificial intelligence module, nor is there a corresponding backup design; Currently, intelligent FSUs, which are in the exploratory stage, mainly focus on the implementation of artificial intelligence functions and have not yet designed a hardware-level security backup control mechanism that is independent of the artificial intelligence module, container service and software system. Existing reliability measures mostly rely on passive recovery methods such as container restart and process monitoring, which have a slow response speed, are difficult to effectively prevent critical parameters from exceeding limits during system anomalies, and cannot bypass faulty containers or service interfaces to directly execute hardware-level protection actions. The failure to establish an effective isolation and mutual exclusion mechanism between the AI ​​energy-saving control mode and the backup control mode can easily lead to control command conflicts, which may cause equipment to go out of control in severe cases.

[0005] More importantly, the AI ​​containers deployed in intelligent FSUs differ fundamentally from traditional control logic: AI containers (artificial intelligence models) rely on NPU to predict temperature trends. Their inference cycle is as long as 800-1500ms, and they have unique failure modes such as model loading failure, inference timeout, and abnormal confidence. Traditional watchdog mechanisms can only detect system-level crashes and cannot detect the "fake death" state of AI containers (i.e., the container process exists but there is no effective output), which makes it impossible to detect the risk of exceeding the temperature limit in time. Summary of the Invention

[0006] This application provides a smart FSU edge computing terminal dual-path AI security control system to solve the problem of lack of data center environment protection mechanism in the event of AI prediction failure.

[0007] The system includes: The GD32F303 data acquisition and communication chip is configured to acquire and upload telemetry data and issue corresponding control commands based on priority-adjustable dual communication interfaces. The telemetry data includes data on the data center environment and equipment. The control commands include intelligent control commands and security control commands. The RK3568 main control unit is configured to perform AI trend prediction based on the telemetry data and generate the intelligent control commands; the control commands are used to adjust the operating status of the data center equipment. A RISC-V coprocessor is configured to perform basic security calculations based on the telemetry data and generate the security control instructions when the AI ​​trend prediction function of the RK3568 main control unit is in an abnormal state; the security control instructions are used to adjust the operating status of the data center equipment.

[0008] In some implementations, the RK3568 main control unit includes: An information acquisition container is configured to receive the telemetry data uploaded by the GD32F303 acquisition and communication chip, and to standardize the telemetry data and AI trend prediction demand data to obtain telemetry AI calculation data. The AI ​​container is built on an artificial intelligence model and is configured to perform AI trend prediction based on the telemetry AI calculation data, generate the intelligent control command, and send it to the GD32F303 acquisition and communication chip.

[0009] In some implementations, the RK3568 main control unit further includes: A system management container is configured to periodically monitor the AI ​​container and issue the running status of the AI ​​container.

[0010] In some implementations, the RISC-V coprocessor is further configured as follows: The operating status of the AI ​​container is periodically monitored, and when the AI ​​container is in an abnormal state, the communication channel between it and the GD32F303 acquisition and communication chip is switched.

[0011] In some implementations, the RISC-V coprocessor has a built-in RPMsg shared memory, which is configured as follows: Periodically acquire the operating status data of the AI ​​container; The system independently determines whether the AI ​​has failed based on the operating status data; if the AI ​​fails, it switches the communication channel with the GD32F303 acquisition and communication chip.

[0012] In some implementations, the GD32F303 acquisition and communication chip is connected to: A temperature and humidity sensor, configured to acquire the temperature and humidity inside the computer room environment, as well as the current data of the computer room equipment; A smart air conditioner, configured to deliver air to the environment of the computer room according to control commands.

[0013] In some implementations, the dual communication interface of the GD32F303 acquisition and communication chip includes: A USB interface is provided, which connects the GD32F303 data acquisition and communication chip to the RK3568 main control unit. The USART interface connects the GD32F303 acquisition and communication chip to the RISC-V coprocessor.

[0014] In some implementations, the GD32F303 acquisition and communication chip further includes: An RS485 interface is provided, which connects the GD32F303 data acquisition and communication chip to the smart air conditioner. An ADC channel is provided, which connects the GD32F303 data acquisition and communication chip to the temperature and humidity sensor.

[0015] In some implementations, the GD32F303 acquisition and communication chip is further configured as follows: When the AI ​​container is in normal operation, the communication priority is switched to USB priority mode, and the acquired intelligent control commands are sent out. When the AI ​​container is in an abnormal operating state, the communication priority is switched to USART priority mode, and the acquired security control command is sent out.

[0016] In some implementations, the operational status data includes the status flags of the AI ​​container, inference confidence, and instruction output heartbeats.

[0017] As described above, this application provides a smart FSU edge computing terminal dual-path AI security control system. The system includes a GD32F303 acquisition and communication chip, configured to acquire and upload telemetry data and issue corresponding control commands via a priority-adjustable dual communication interface. The telemetry data includes data from the data center environment and equipment. The control commands include intelligent control commands and security control commands. An RK3568 main control unit is configured to perform AI trend prediction based on the telemetry data and generate the intelligent control commands. The control commands are used to adjust the operating status of the data center equipment. A RISC-V coprocessor is configured to perform basic security calculations based on the telemetry data and generate the security control commands when the AI ​​trend prediction function of the RK3568 main control unit is in an abnormal state. The security control commands are used to adjust the operating status of the data center equipment. This application solves the problem of lacking a data center environment protection mechanism in the event of AI prediction failure through the above system. Attached Figure Description

[0018] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a schematic diagram of a dual-path AI security control system for an intelligent FSU edge computing terminal according to this application; Figure 2 This is a schematic diagram of the RK3568 main control unit in a dual-path AI security control system for an intelligent FSU edge computing terminal according to this application; Figure 3 This application presents a failure determination and state machine diagram for a dual-path AI security control system for an intelligent FSU edge computing terminal. Figure 4 This is a flowchart of the firmware instruction filtering process of the GD32F303 of the intelligent FSU edge computing terminal dual-channel AI security control system of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0022] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] See Figures 1 to 4 As can be seen, this embodiment provides a dual-path AI security control system for intelligent FSU edge computing terminals, the system comprising: The GD32F303 data acquisition and communication chip is configured to acquire and upload telemetry data and issue corresponding control commands based on priority-adjustable dual communication interfaces. The telemetry data includes data on the data center environment and equipment. The control commands include intelligent control commands and security control commands.

[0024] Specifically, in this embodiment, the GD32F303 acquisition and communication chip can be understood as the information interaction core of the system, that is, the uploading and downloading of information is completed through the GD32F303 acquisition and communication chip.

[0025] The system also includes: The RK3568 main control unit is configured to perform AI trend prediction based on the telemetry data and generate the intelligent control commands; the control commands are used to adjust the operating status of the equipment in the computer room.

[0026] Specifically, in this embodiment, the RK3568 main control unit is built on a quad-core Cortex-A55 main core and runs on Linux 5.10. The RK3568 main control unit collects data uploaded by the GD32F303 acquisition and communication chip and performs AI trend prediction to obtain the intelligent control instructions for trend regulation, thereby adjusting the operating status of the data center equipment.

[0027] The system also includes: A RISC-V coprocessor is configured to perform basic security calculations based on the telemetry data and generate the security control instructions when the AI ​​trend prediction function of the RK3568 main control unit is in an abnormal state; the security control instructions are used to adjust the operating status of the data center equipment.

[0028] Specifically, in this embodiment, since AI trend prediction often fails, and the basic requirements of the data center environment and equipment still need to be maintained in such cases, the RISC-V coprocessor is set up. The RISC-V coprocessor is a component independent of the RK3568 main control unit. When the AI ​​trend prediction function is in an abnormal state, the RISC-V coprocessor collects the telemetry data, performs basic safety calculations, and generates the safety control commands, thereby providing the most basic environmental safety guarantee for the data center and its equipment.

[0029] It should be noted that this embodiment uses temperature and humidity as the basic protection data for the data center environment. Therefore, the security control instructions generated by the RISC-V coprocessor are mainly used to adjust the ambient temperature of the data center. AI trend prediction can be understood as long-term planning and control of the data center environment, while basic security calculation is to prevent the data center environment data from exceeding the set threshold. For example, if the data center temperature exceeds the upper limit threshold, cooling measures need to be activated or increased. The specific method of performing basic security calculation is not limited here.

[0030] The dual communication interfaces of the GD32F303 acquisition and communication chip include a USB interface connecting the GD32F303 acquisition and communication chip to the RK3568 main control unit; a USART interface connecting the GD32F303 acquisition and communication chip to the RISC-V coprocessor; an RS485 interface connecting the GD32F303 acquisition and communication chip to the smart air conditioner; and an ADC channel connecting the GD32F303 acquisition and communication chip to the temperature and humidity sensor.

[0031] Furthermore, in some embodiments, the RK3568 main control unit includes: An information acquisition container is configured to receive the telemetry data uploaded by the GD32F303 acquisition and communication chip, and to standardize the telemetry data and AI trend prediction demand data to obtain telemetry AI calculation data.

[0032] Specifically, in this embodiment, the telemetry data uploaded by the GD32F303 acquisition and communication chip is received through the information acquisition container, and before performing AI trend prediction, the telemetry data is preprocessed by the information acquisition container to obtain the AI ​​calculation data of the telemetry data.

[0033] The RK3568 main control unit also includes: The AI ​​container is built on an artificial intelligence model and is configured to perform AI trend prediction based on the telemetry AI calculation data, generate the intelligent control command, and send it to the GD32F303 acquisition and communication chip.

[0034] Specifically, in this embodiment, the AI ​​container is built based on an AI artificial intelligence model. The AI ​​container performs AI trend prediction based on the collected telemetry AI calculation data, thereby predicting the trend of the environmental conditions of the area where the computer room is located, and generating corresponding intelligent control commands, so as to perform "forward-looking" control of the environment where the computer room is located.

[0035] Furthermore, in some embodiments, the RK3568 main control unit further includes: A system management container is configured to periodically monitor the AI ​​container and issue the running status of the AI ​​container.

[0036] Specifically, in this embodiment, since the AI ​​container is based on the NPU to predict temperature trends, its inference cycle is as long as 800-1500ms, and it has unique failure modes such as model loading failure, inference timeout, and confidence abnormality, the system management container is set up to perform the most basic supervision of the running status of the AI ​​container.

[0037] Based on the above, in some embodiments, the RISC-V coprocessor is further configured as follows: The operating status of the AI ​​container is periodically monitored, and when the AI ​​container is in an abnormal state, the communication channel between it and the GD32F303 acquisition and communication chip is switched.

[0038] Specifically, in this embodiment, the system management container can be understood as a "traditional watchdog mechanism." However, the traditional watchdog mechanism can only detect system-level crashes and cannot detect the "fake death" state of the AI ​​container (i.e., the container process exists but has no effective output), which leads to the inability to detect the risk of exceeding the temperature limit in time. Therefore, the RISC-V coprocessor is set to monitor the AI ​​container. The RISC-V coprocessor independently monitors the running status of the main processor's AI container and implements a technical solution of switching the control path in milliseconds when the AI ​​fails.

[0039] Furthermore, in some embodiments, the RISC-V coprocessor has a built-in RPMsg shared memory, which is configured as follows: Periodically acquire the operating status data of the AI ​​container; The system independently determines whether the AI ​​has failed based on the operating status data; if the AI ​​fails, it switches the communication channel with the GD32F303 acquisition and communication chip.

[0040] Specifically, in this embodiment, the RISC-V coprocessor runs independently of the RK3568 main control unit (Cortex-A55 main core), and is dedicated to periodically obtaining the AI ​​container status flag, inference confidence, and instruction output heartbeat through the RPMsg shared memory to achieve independent determination of AI failure, rather than indirectly obtaining it through the main core software.

[0041] Furthermore, in some embodiments, the GD32F303 acquisition and communication chip is connected to: A temperature and humidity sensor, configured to acquire the temperature and humidity inside the computer room environment, as well as the current data of the computer room equipment; A smart air conditioner, configured to deliver low-temperature air to the environment of the computer room according to control commands.

[0042] Specifically, in this embodiment, temperature and humidity are used as parameters for controlling the environment of the computer room. Therefore, the corresponding temperature and humidity sensors and the smart air conditioner are required as data providers and specific devices for the actual implementation of the scheme.

[0043] It should be noted that this embodiment only uses temperature and humidity as examples. Depending on different needs, parameters such as the current and voltage of the equipment in the computer room can be used as one of the parameters for monitoring the computer room, and the corresponding data acquisition equipment can also be determined according to different needs.

[0044] Furthermore, in some embodiments, the GD32F303 acquisition and communication chip is also configured as follows: When the AI ​​container is in normal operation, the communication priority is switched to USB priority mode, and the acquired intelligent control commands are sent out. When the AI ​​container is in an abnormal operating state, the communication priority is switched to USART priority mode, and the acquired security control command is sent out.

[0045] Specifically, in this embodiment, the GD32F303 acquisition and communication chip selects to send control commands output by the AI ​​container or the RISC-V coprocessor by adjusting the priority mode of the channel interface.

[0046] It should be noted that when the system is in the USART priority mode or the USB priority mode, the data reading function is still retained in the other interface link.

[0047] As can be seen from the above, this embodiment includes the following links: AI energy-saving control path (main path): Logical link: AI container → public service interface → virtual device layer → USB communication → GD32F303 acquisition and communication chip (USB priority mode) → RS485 → air conditioner execution hardware; Dependencies: The AI ​​container is running normally, and USB communication is uninterrupted; Core data: Intelligent control commands output by the AI ​​container.

[0048] Hardware-level security path (backup control path): Logical link: RISC-V coprocessor → USART communication → GD32F303 (USART priority mode) → RS485 → air conditioner execution hardware; Independence: It does not rely on any software from the Cortex-A55 main core, and this is ensured through two core mechanisms: Data independence: The RISC-V coprocessor periodically acquires raw telemetry data from the GD32F303; State independence: The RISC-V coprocessor periodically synchronizes the running status of AI containers from the Cortex-A55 main core; Isolated and independent: The GD32F303 internal firmware switches priorities, and USB control commands are hardware-level blocked, responding only to RISC-V commands.

[0049] The communication connection methods of the system provided in this embodiment are shown in Table 1.

[0050] Table 1 System Communication Connection Methods

[0051] Interface parameter definition: RPMsg shared memory interface (inter-core state synchronization): Address range: RK3568 reserved-memory region (0x07800000~0x07BFFFFF), with 2KB allocated for state interaction; Data format: [AI container status flag (1 byte: 0 = normal, 1 = invalid) + inference confidence (1 byte) + timestamp (4 bytes) + checksum (1 byte)]; Synchronization period: 500ms / time, consistent with the RISC-V acquisition period; Synchronization mechanism: The Cortex-A55 main core sets the flag bit after writing data, and the RISC-V coprocessor sets it to zero after reading, which, together with HardwareSpinlock, enables atomic operations.

[0052] USART interface (RISC-V and GD32F303): Pin configuration: RISC-V coprocessor PB10 (TX), PB11 (RX), and corresponding pins of GD32F303USART2; Response latency: ≤10ms, meeting the 500ms cycle requirement.

[0053] Data format: The communication protocol format adopts the departmental format, and the basic format is shown in Table 2.

[0054] Table 2 USART Interface Data Format

[0055] The meanings of the symbols in Table 2 are shown in Table 3.

[0056] Table 3 Symbol Meanings

[0057] The command information definitions in Table 3 are shown in Table 4.

[0058] Table 4 Command Information Definition

[0059] The definition of the return code RTN in Table 3 is shown in Table 5.

[0060] Table 5. Return Code RTN Definition

[0061] The LENGTH data format in Table 3 is shown in Table 6.

[0062] Table 6 LENGTH Data Format

[0063] LENGTH consists of 2 bytes, composed of LENID and LCHKSUM. LENID indicates the number of ASCII bytes transmitted for the INFO item. When LENID=0, INFO is empty, meaning there is no such item. LENGTH is transmitted by sending the high byte first, followed by the low byte, in four ASCII bytes.

[0064] The check digit is calculated as follows: D11D10D9D8+D7D6D5D4+D3D2D1D0. After summing, the remainder modulo 16 is inverted and then 1 is added.

[0065] CHKSUM data format: CHKSUM is calculated by summing the ASCII values ​​of all characters except SOI, EOI, and CHKSUM, then inverting the remainder when the result is modulo 65536 and adding 1. Example: If the received or sent byte sequence is "~1203400456ABCDFEFC72\R", then FC72 in the last five characters "FC72\R" is CHKSUM. The calculation method is as follows: '1'+'2'+'0'+···+'A'+'B'+···+'F'+'E' =31H+32H+30H+···+41H+42H+···+46H+45H =038EH; Here, '1' represents the ASCII code value of 1, and 'E' represents the ASCII code value of E. The remainder of 038EH modulo 65536 is 038EH. Inverting 038EH and adding 1 gives 'FC72'.

[0066] Example message: Acquire telemetry data frames: 0x7E+0x10+0x01+0x90+0x42 (acquire telemetry data)+length information+checksum+0x0D; Switching data command frames: 0x7E+0x10+0x01+0x90+0x61 (Switch data channel)+Length information+Instruction information (0x01=Backup control path, 0x02=Normal mode)+Checksum+0x0D.

[0067] GD32F303 internal priority switching parameters: USB Priority Mode: The default mode, which prioritizes responding to the Cortex-A55 master core's USB control commands and disables USART control commands (only data reading is retained). USART Priority Mode: Forced mode, only responds to RISC-V coprocessor USART control commands, and filters USB control commands at the hardware level (reserves data reporting). Switching delay: ≤5ms, no jitter during switching.

[0068] The GD32F303 uses firmware logic to filter the priority of control commands internally, eliminating the need for external hardware switches. The GD32F303 internally sets a global control priority flag, g_control_priority.

[0069] When this flag is 0x00, the GD32F303 is in USB priority mode and only responds to USB control commands; When this flag is 0x01, the GD32F303 is in USART priority mode and only responds to USART control commands; Discard the control command for the USB channel (returns an error code).

[0070] The mode switching command is received via USART and the flag bit is modified directly, with a switching delay of ≤5ms.

[0071] RT-Thread task scheduling strategy: The RISC-V coprocessor runs the RT-Thread 3.1 real-time system, employing a multi-threaded architecture to achieve dual-task scheduling with a 500ms cycle. See Table 7 for thread configuration.

[0072] Table 7 Thread Configuration

[0073] Scheduling sequence: 0ms hardware timer interrupt triggered → 0-2ms data acquisition thread runs → 2-4ms state synchronization thread runs → 4-10ms control logic thread runs (triggered by semaphore) → 10-500ms idle wait.

[0074] AI container status monitoring mechanism: Status flag monitoring: The RISC-V coprocessor reads the AI ​​container status flag (0=normal, 1=failure) in the RPMsg shared memory every 500ms. This flag is set by the system management container when it detects that the AI ​​process has crashed or that the NPU resources are overloaded. Inference confidence monitoring: Read the confidence value (0-100) output by the AI ​​container. When the confidence is <50% or the confidence shows a downward trend for two consecutive periods, it is determined that the model inference has failed. Command output heartbeat monitoring: Detects whether the AI ​​container outputs control commands according to the expected cycle (800-1500ms). If there are no valid commands for two consecutive inference cycles, it is determined that the AI ​​is in a state of suspension. Inter-core communication effectiveness monitoring: The RPMsg shared memory update status is detected by Hardware Spinlock. If it is not updated for 3 consecutive cycles (1500ms), it is determined that the inter-core communication is interrupted.

[0075] The specific implementation steps of this embodiment are as follows: Step 1: Heterogeneous architecture initialization configuration: Cortex-A55 main core startup: Load Linux system and containerized environment, initialize AI container, system management container, and virtual device layer; send telemetry acquisition parameters to GD32F303 via USB (sampling frequency 500ms / time); send backup configuration parameters to RISC-V coprocessor via RPMsg shared memory (high temperature threshold 30℃, low temperature threshold 18℃, acquisition / status synchronization cycle 500ms).

[0076] RISC-V coprocessor startup: Load the RT-Thread3.1 real-time system, initialize the USART interface and RPMsg communication protocol stack, create data_acquisition_thread (priority 10), status_sync_thread (priority 11), and control_logic_thread (priority 12), start a 500ms hardware timer, and enter "AI power saving control mode".

[0077] GD32F303 initialization: Start telemetry acquisition (500ms cycle), enter "USB priority mode" by default (g_control_priority=0x00), enable parallel communication between USB and USART (data can be exchanged, control commands only respond to USB).

[0078] Step 2: Timed data acquisition and status synchronization: The RISC-V coprocessor executes dual tasks in 500ms cycles: Data acquisition task (data_acquisition_thread): Sends telemetry read commands to GD32F303 via USART, receives and parses raw temperature and humidity data, and stores it in local cache; Status synchronization task (status_sync_thread): Reads the running status of AI containers in RPMsg shared memory, including container running status, inference confidence (≥80% is valid), and instruction output status.

[0079] Step 3: AI Container Failure Determination: The RISC-V coprocessor determines that the AI ​​container has failed if any of the following conditions are met based on the synchronization data: Condition 1: The AI ​​container status flag is 1 (the system manages the container and writes it). Condition 2: Inference confidence < 50% or no valid instruction output for two consecutive cycles; Condition 3: No shared memory update is read for 3 consecutive cycles (inter-core communication interruption); Condition 4: Telemetry data shows that the temperature exceeds the threshold (≥30℃ or ≤18℃) and no AI container control command is received within 5 consecutive acquisition cycles (2500ms).

[0080] Note: The single inference cycle of an AI container is typically 800-1500ms. Using a 2500ms (5 cycles) decision window can cover the entire inference cycle, ensuring timely detection while avoiding misjudging normal inference delays as failures.

[0081] Step 4: Force switch to hardware-level security mode (software-level isolation): The RISC-V coprocessor sends a "standby control mode activation instruction" (0x7E 0x100x01 0x90 0x61 0x01 0x00 0x01 [CHKSUM] 0x0D) to the GD32F303 via USART. After receiving the command, the GD32F303 modifies the internal firmware to g_control_priority=0x01, switches to "USART priority mode", and hardware-level blocks all control commands of the USB channel, retaining only the telemetry data reporting function. The RISC-V coprocessor sends a "standby control mode start notification" to the Cortex-A55 main core via RPMsg, and the system management container disables the USB write permission of the AI ​​container, forming a dual guarantee.

[0082] Step 5: The RISC-V coprocessor performs standby control: The RISC-V coprocessor reads the latest telemetry (temperature and humidity) data from the GD32F303 and compares it with preset thresholds: High temperature ≥30℃: Send a "Turn on cooling to 26℃" command to GD32F303 via USART; Low temperature ≤18℃: Issue "shutdown" command; Temperature normal: Maintain the current state and continue to collect data and synchronize status.

[0083] After receiving the USART instruction, the GD32F303 forwards it to the air conditioner via RS485, executes the corresponding operation, and feeds back the status to the RISC-V coprocessor.

[0084] Step 6: RISC-V Coprocessor Failure Detection and Secondary Security Protection: The Cortex-A55 main core sends a heartbeat query (reading specific registers) to the RISC-V coprocessor via RPMsg every 1000ms. If there is no response for 3 consecutive times (3000ms), the RISC-V coprocessor is determined to have failed, triggering a secondary safety protection mechanism: The Cortex-A55 master core sends an "emergency mode command" (0x7E 0x10 0x01 0x900x71 0x00 0x00 [CHKSUM] 0x0D) to the GD32F303 via USB, forcibly switching the GD32F303 to "Cortex-A55 direct control mode"; Simplified control logic is enabled: forced power-on when the temperature is >30℃, forced power-off when the temperature is <18℃, no longer relying on AI inference; Report a "RISC-V failure alarm" and request manual intervention; It supports manually triggering recovery through the intelligent FSU Web interface, bypassing the automatic judgment logic.

[0085] Step 7: Automatic recovery switching (anti-vibration design): The AI ​​power-saving control recovery process is triggered when the RISC-V coprocessor detects all of the following conditions for five consecutive cycles (2500ms): Condition 1: The AI ​​container status flag is 0, and the confidence level is ≥70% or the confidence level shows an upward trend for 3 consecutive periods; Condition 2: Receive valid control commands from the AI ​​container for 3 consecutive cycles (1500ms); Condition 3: Remote measurement data shows that the temperature is within the safe range (18℃~28℃); Condition 4: Inter-core communication is normal (shared memory is updated every 500ms).

[0086] Recovery process: The RISC-V coprocessor sends a "normal mode recovery instruction" (0x7E 0x100x01 0x90 0x61 0x01 0x00 0x02 0x03 0x8E 0x0D) to the GD32F303 via USART. Modifying g_control_priority=0x00 on the GD32F303 switches back to "USB priority mode" and re-responds to Cortex-A55 master core control instructions; The RISC-V coprocessor notifies the Cortex-A55 main core to restore USB write permissions for the AI ​​container via RPMsg; The RISC-V coprocessor exits standby control and returns to "AI power saving control mode" to continue performing timed data acquisition and status synchronization. 2-second anti-oscillation delay: Before exiting the backup control mode, the RISC-V coprocessor continuously monitors for 2 seconds. Only after confirming that there are no abnormalities will it exit the backup control state and switch to the normal state.

[0087] The failure criteria for AI containers are shown in Table 8.

[0088] Table 8 Failure Judgment Criteria for AI Containers

[0089] A comparison of the dual control pathways is shown in Table 9 and see also [reference]. Figure 4 .

[0090] Table 9 Comparison of Dual Control Pathways

[0091] This embodiment has the following advantages: Dual-path hardware-level protection ensures extremely high reliability: The RISC-V coprocessor operates independently, acquiring dual information through timed data collection and status synchronization, without relying on the Cortex-A55 main core software, thus increasing system availability from 99% to 99.99%. Software-level isolation with zero hardware changes: The internal priority switching of the GD32F303 replaces the external analog switch, eliminating the need for additional hardware such as relays and bus switches. The switching delay is ≤5ms, and the reliability is better than that of physical switches, resulting in extremely low productization costs. Extremely fast response speed: data acquisition and status synchronization with a 500ms cycle, failure detection → mode switching → command issuance process ≤100ms, which is much faster than traditional watchdog recovery (≥3s), and can effectively prevent temperature from exceeding the limit; The anti-vibration design is comprehensive: multiple recovery conditions + confidence trend judgment + 5-cycle confirmation + 2-second delay, which completely avoids system oscillations caused by AI container instability and supports manual forced recovery to AI energy-saving control mode via the web interface; Secondary safety protection mechanism: When the RISC-V coprocessor fails, the Cortex-A55 main core can take over control, forming a dual hardware-level protection. Highly adaptable and cost-controllable: Fully compatible with RK3568 heterogeneous architecture, and compatible with Linux 5.10 system and containerized deployment; No new hardware is required; only the firmware of the GD32F303 and RISC-V coprocessor needs to be upgraded, resulting in low productization costs. Supports mainstream air conditioning protocols such as Modbus-RTU, and is compatible with air conditioning equipment from different brands; Compliant with heterogeneous architecture development specifications: All resource allocation, inter-core communication, and startup process follow the Rockchip RK3568AMP development guidelines, and the technical feasibility is supported by clear chip documentation; Quantitative testing and verification: After 30 days of continuous operation and testing, simulating 20 AI container failures, the dual-path AI security control strategy was triggered with a 100% success rate and 0 instances of temperature exceeding limits. Compared to traditional watchdog solutions (system reset time 3-5s, temperature fluctuation ±3℃), this solution has a control switching delay ≤100ms and temperature fluctuation controlled within ±0.5℃.

[0092] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the discussion in some embodiments is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the contents of this disclosure, thereby enabling those skilled in the art to better utilize the embodiments.

Claims

1. A dual-path AI security control system for an intelligent FSU edge computing terminal, characterized in that, The system includes: The GD32F303 data acquisition and communication chip is configured to acquire and upload telemetry data and issue corresponding control commands based on priority-adjustable dual communication interfaces. The telemetry data includes data on the data center environment and equipment. The control commands include intelligent control commands and security control commands. The RK3568 main control unit is configured to perform AI trend prediction based on the telemetry data and generate the intelligent control commands; the control commands are used to adjust the operating status of the data center equipment. A RISC-V coprocessor is configured to perform basic security calculations based on the telemetry data and generate the security control instructions when the AI ​​trend prediction function of the RK3568 main control unit is in an abnormal state; the security control instructions are used to adjust the operating status of the data center equipment. The RK3568 main control unit includes: An information acquisition container is configured to receive the telemetry data uploaded by the GD32F303 acquisition and communication chip, and to standardize the telemetry data and AI trend prediction demand data to obtain telemetry AI calculation data. The AI ​​container is built on an artificial intelligence model and is configured to perform AI trend prediction based on the telemetry AI calculation data, generate the intelligent control command, and send it to the GD32F303 acquisition and communication chip. The RISC-V coprocessor is also configured to: The operating status of the AI ​​container is periodically monitored, and when the AI ​​container is in an abnormal state, the communication channel between it and the GD32F303 acquisition and communication chip is switched. The RISC-V coprocessor has a built-in RPMsg shared memory, which is configured as follows: The real-time operating status data of the AI ​​container is acquired periodically; The system independently determines whether the AI ​​has failed based on the operating status data; if the AI ​​fails, it switches the communication channel with the GD32F303 acquisition and communication chip.

2. The intelligent FSU edge computing terminal dual-path AI security control system according to claim 1, characterized in that, The RK3568 main control unit also includes: A system management container is configured to periodically monitor the AI ​​container and issue the running status of the AI ​​container.

3. The intelligent FSU edge computing terminal dual-path AI security control system according to claim 1, characterized in that, The GD32F303 data acquisition and communication chip is connected to: A temperature and humidity sensor, configured to acquire the temperature and humidity inside the computer room environment, as well as the current data of the computer room equipment; A smart air conditioner, configured to deliver air to the environment of the computer room according to control commands.

4. The intelligent FSU edge computing terminal dual-path AI security control system according to claim 1, characterized in that, The dual communication interfaces of the GD32F303 acquisition and communication chip include: A USB interface is provided, which connects the GD32F303 data acquisition and communication chip to the RK3568 main control unit. The USART interface connects the GD32F303 acquisition and communication chip to the RISC-V coprocessor.

5. The intelligent FSU edge computing terminal dual-path AI security control system according to claim 3, characterized in that, The GD32F303 data acquisition and communication chip also includes: An RS485 interface is provided, which connects the GD32F303 data acquisition and communication chip to the smart air conditioner. An ADC channel is provided, which connects the GD32F303 data acquisition and communication chip to the temperature and humidity sensor.

6. The intelligent FSU edge computing terminal dual-path AI security control system according to claim 1, characterized in that, The GD32F303 data acquisition and communication chip is also configured as follows: When the AI ​​container is in normal operation, the communication priority is switched to USB priority mode, and the acquired intelligent control commands are sent out. When the AI ​​container is in an abnormal operating state, the communication priority is switched to USART priority mode, and the acquired security control command is sent out.

7. The intelligent FSU edge computing terminal dual-path AI security control system according to claim 1, characterized in that, The operational status data includes the status flags of the AI ​​container, inference confidence, and command output heartbeats.

Images