An energy-saving AI management device for a machine room air conditioner

Abstract

This utility model discloses an energy-saving AI management device for computer room air conditioning, relating to the field of air conditioning management technology. It includes a standard rack-mount control cabinet housing an industrial-grade power supply, a PCIe communication main control backplane, an FPGA digital twin parallel processing module, an AI strategy and cost decision-making module, a physical field driving and redundancy control module, and a federated security communication module. The FPGA digital twin parallel processing module employs three parallel logic processing areas, responsible for data synchronization and alignment, 3D heat map construction, and equipment health status assessment, respectively. It achieves comprehensive acquisition of electrical, flow, vibration, and temperature parameters through a multi-source industrial signal interface unit. The AI ​​strategy and cost decision-making module is based on a COM-Express standard core board, equipped with an edge computing processor, and uses deep learning algorithms for intelligent decision-making. This utility model achieves millisecond-level real-time response through FPGA parallel processing technology, effectively solving the technical problems of slow response speed, low control accuracy, and high energy consumption in traditional computer room air conditioning control systems.

Description

Technical Field

[0001] This utility model relates to the field of air conditioning management technology, specifically an AI management device for energy saving in computer room air conditioning. Background Technology

[0002] As data centers and server rooms continue to expand in scale, the energy consumption of server room air conditioning systems has become an important part of the operating costs of data centers, typically accounting for 30-40% of the total energy consumption. Traditional server room air conditioning control systems mostly adopt simple control strategies based on temperature thresholds, which cannot achieve precise temperature regulation and energy efficiency optimization, resulting in serious energy waste.

[0003] In existing technologies, traditional data center air conditioning control systems mainly suffer from the following problems: First, they use a single type of sensor for environmental monitoring, lacking the comprehensive perception capability for multi-dimensional information such as temperature distribution, equipment operating status, and energy consumption parameters within the data center; second, the control system has a slow response speed and large data processing latency, failing to meet the stringent requirements of data center equipment for millisecond-level real-time response; third, they lack intelligent decision-making capabilities, unable to perform predictive control and cost optimization based on historical data and changes in the external environment; and finally, the system has low integration, with each functional module operating independently, resulting in complex maintenance and poor scalability.

[0004] Some existing intelligent air conditioning control solutions have improved control accuracy and response speed to some extent by introducing sensor networks and automated control technologies, but they still have certain limitations (list the shortcomings of the existing technologies based on the problems solved by this application), such as: limited data processing capabilities, unable to achieve real-time parallel processing of large-scale multi-source heterogeneous data; lack of hardware-level security mechanisms, resulting in insufficient data transmission and storage security; low control accuracy, unable to achieve accurate reconstruction of the airflow field and high-precision control of intelligent air valves; simple energy efficiency optimization algorithms, failing to fully utilize artificial intelligence technology for deep learning and cost optimization; and poor system reliability, lacking effective fault prediction and redundant control mechanisms.

[0005] Therefore, there is an urgent need for an AI management device for energy-saving air conditioning in computer rooms that integrates advanced technologies such as FPGA parallel processing, AI intelligent decision-making, multi-source signal fusion, and hardware-level security, in order to solve the technical problems of slow response speed, low control accuracy, high energy consumption, and poor security of traditional control systems, and to realize intelligent, efficient and safe management of computer room air conditioning systems. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and propose an AI management device for energy-saving air conditioning in computer rooms to solve the above-mentioned problems.

[0007] The purpose of this utility model is achieved through the following technical solution: a data center air conditioning energy-saving AI management device, the device is a standard rack-mount control cabinet structure, including: an industrial-grade switching power supply; a PCIe communication main control backplane, the PCIe communication main control backplane is provided with multiple standardized PCIe slots, the PCIe communication main control backplane is provided with standardized power distribution rails for powering all internal modules and data bus channels for realizing high-speed data exchange between modules;

[0008] The FPGA digital twin parallel processing module is an independent PCIe interface board that connects to the PCIe communication main control backplane via a PCIe gold finger structure. The FPGA digital twin parallel processing module integrates a multi-source industrial signal interface unit, which includes several opto-isolated RS485 terminals, at least one BNC coaxial interface with an internal high-frequency signal conditioning circuit, a set of multi-channel differential signal input terminals with built-in preamplifiers, and a wireless communication interface with a physically soldered Zigbee or LoRa RF transceiver module. The FPGA digital twin parallel processing module also integrates an FPGA parallel processing core unit, which includes an FPGA chip fixedly installed in the central area of ​​the FPGA digital twin parallel processing module. The FPGA chip has three parallel-working logic processing areas programmed and solidified internally using a hardware description language. The FPGA digital twin parallel processing module also features an onboard cache and DMA transfer unit, which includes a high-speed data cache Block RAM and a direct memory access DMA controller module configured within the FPGA chip.

[0009] The AI ​​strategy and cost decision module is a high-performance COM-Express standard core board. It is equipped with an edge computing processor with a built-in GPU or NPU and an M.2 slot for installing a solid-state drive.

[0010] The physical field drive and redundancy control module is an independent power drive and relay control board. The physical field drive and redundancy control module is equipped with an airflow field reconstruction drive array composed of multiple physically isolated stepper motor drive chips and a hybrid cooling control relay matrix composed of solid-state relays and electromagnetic relays. The physical field drive and redundancy control module also integrates an onboard safety monitoring microcontroller (MCU).

[0011] The Federal Security Communication Module is an independent communication interface board. It integrates a Mini-PCIe slot for plugging in a 5G or LTE wireless communication module and a Trusted Platform Module (TPM) security encryption chip physically soldered onto the board.

[0012] The multi-source industrial signal interface unit is located in the I / O baffle area of ​​the FPGA digital twin parallel processing module. The RS485 terminal is used to connect to the smart meter, the BNC coaxial interface is used to receive the pulse signal from the water flow sensor, the multi-channel differential signal input terminal is used to connect to the mechanical vibration sensor or acoustic sensor, and the RF transceiver module is connected to the external antenna port through the onboard RF cable.

[0013] The three parallel logical processing areas are as follows:

[0014] The first logical processing area is used for high-precision hardware clock-based synchronous acquisition and timestamp alignment of multi-source heterogeneous sensor data.

[0015] The second logical processing area is used to receive the wireless temperature data stream aligned by the first logical processing area and the sensor spatial coordinate information, and to construct and refresh the three-dimensional thermal map model of the computer room through parallel pipeline operations.

[0016] The third logic processing area is used to receive the mechanical status data stream aligned by the first logic processing area and to evaluate the equipment's operational health status in real time using the built-in FFT (Fast Fourier Transform) and feature comparison algorithms.

[0017] The DMA controller module is directly connected to the PCIe gold finger and is used to transmit the processing results of the three-dimensional heat map and device health status map generated by the second and third logic processing areas directly to the AI ​​strategy and cost decision module at high speed through the PCIe communication main control backplane without CPU intervention.

[0018] Solid-state drives are used for physical storage of historical digital twin data and operating cost model databases. The operating cost model database includes time-of-use electricity price data and tiered water price data. The AI ​​strategy and cost decision module reads the 3D heat map and equipment health status map generated by the FPGA digital twin parallel processing module directly from memory via PCIe communication main control backplane in DMA mode.

[0019] Each drive channel of the airflow field reconstruction drive array is connected to an intelligent air valve through a high-current output terminal block. The hybrid refrigeration control relay matrix is ​​used to control the evaporative refrigeration system and the standby air conditioning system through dry contacts, respectively.

[0020] The onboard safety monitoring microcontroller (MCU) has a built-in watchdog function, which is used to execute preset hardware-level fault safety control strategies when communication with the AI ​​strategy and cost decision module fails.

[0021] The TPM security encryption chip is used for hardware-level encryption of data transmitted externally, while the Federal Security Communication Module is responsible for providing a network channel for the AI ​​Strategy and Cost Decision Module to obtain external weather data.

[0022] Next to the FPGA chip is a dedicated power regulator chip LDO, which provides a stable, low-ripple core voltage for the FPGA chip. The FPGA chip is connected to an external configuration flash memory SPI Flash via the SPI bus. The configuration flash memory SPI Flash is used to store and load the internal logic configuration file of the FPGA chip.

[0023] The FPGA chip surface is covered with an aluminum heat sink with heat dissipation fins. The edge of the FPGA digital twin parallel processing module board is equipped with a JTAG debugging interface and LED status indicators. The LED status indicators are used to display the status of each interface unit, logic processing area and communication.

[0024] The beneficial effects of this utility model are:

[0025] First, this invention employs an FPGA digital twin parallel processing module. The FPGA chip contains three parallel logic processing areas, each specifically responsible for data synchronization and alignment, 3D heat map construction, and equipment health status assessment. This technical solution significantly improves data processing speed compared to traditional CPU serial processing and significantly reduces data processing latency, meeting the stringent real-time response requirements of the data center air conditioning system. Simultaneously, the parallel architecture of the three independent logic processing areas significantly increases system throughput, enabling simultaneous processing of multi-source heterogeneous data streams from a large number of sensors, thus solving the technical problem of insufficient processing capacity in existing technologies.

[0026] Secondly, the multi-source industrial signal interface unit of this utility model includes an RS485 terminal block with opto-isolation, a BNC coaxial interface with an internal high-frequency signal conditioning circuit, a multi-channel differential signal input terminal with a built-in preamplifier, and a wireless communication interface with a Zigbee or LoRa RF transceiver module physically soldered on. It realizes comprehensive acquisition of electrical parameters, flow parameters, vibration parameters, and temperature parameters. Compared with traditional single-sensor monitoring systems, the monitoring coverage is greatly improved. The opto-isolation technology ensures high isolation voltage and effectively eliminates electromagnetic interference in the industrial field. The high-frequency signal conditioning circuit supports a wide frequency range. The preamplifier has a large adjustable gain range and a low noise figure, realizing high-precision detection of weak signals and solving the technical problems of low signal acquisition accuracy and poor anti-interference ability in the prior art.

[0027] Furthermore, the onboard cache and DMA transfer unit of this invention achieves zero-copy transfer through DMA direct memory access technology, eliminating the performance bottleneck caused by CPU involvement in data transfer, significantly improving data transfer efficiency, and achieving extremely low transmission latency. This technical solution frees up valuable computing resources, allowing the AI ​​strategy and cost decision-making module to focus on complex algorithm operations, resulting in a significant improvement in overall system efficiency. The high-speed PCIe architecture provides high transmission rates and large bandwidth, providing ample transmission channels for large-scale data stream processing, effectively solving the data transfer bottleneck problem in existing technologies.

[0028] Furthermore, the AI ​​strategy and cost decision-making module of this invention adopts a high-performance COM-Express standard core board, equipped with an edge computing processor with built-in GPU or NPU. Based on deep learning, the multi-objective optimization algorithm considers multiple dimensions such as energy efficiency, cost, and reliability, significantly improving decision quality compared to traditional rule-based control. The operating cost model database includes time-of-use electricity price data and tiered water price data. Through time-of-use electricity price optimization and tiered water price optimization, operating costs are significantly reduced, achieving refined cost control. By utilizing historical digital twin data for trend analysis and combining it with external weather data, predictive control is achieved, resulting in a significant improvement in energy efficiency ratio, surpassing the control effect of existing technologies.

[0029] Furthermore, the physical field driving and redundant control module of this utility model adopts a stepper motor driver chip to support micro-stepping control, and the intelligent air valve opening control has high precision, which is significantly improved compared with the traditional proportional control. The solid-state relay switching time is extremely short, and the electromagnetic relay contact capacity is large, realizing a perfect combination of fast response and high-power control. The physically isolated driver chip design avoids the propagation of single-point faults, and the hybrid relay matrix provides multiple protections. The system has high availability and significantly improves the reliability of the system.

[0030] In addition, the security mechanism of this utility model utilizes the built-in watchdog function of the onboard security monitoring microcontroller (MCU) to execute hardware-level fault-tolerant control strategies when communication is abnormal, including immediately shutting down the intelligent air valve and switching to the backup air conditioning system, thus ensuring the safe operation of the equipment in the computer room. The TPM security encryption chip complies with the latest standards and adopts advanced encryption and digital signature algorithms to ensure hardware-level encryption of externally transmitted data, meeting the high security requirements of the Industrial Internet of Things. The TPM chip has a physical attack detection function, which automatically clears the internal key when an attack is detected, effectively preventing the leakage of sensitive information.

[0031] Meanwhile, the federated security communication module of this invention achieves multi-device collaborative optimization through wireless communication modules and federated learning technology, significantly improving model accuracy and convergence speed. It adopts differential privacy technology, adding carefully designed noise to the model parameters to ensure that the data privacy of individual devices is not inferred. It can acquire weather data and electricity price information in real time, providing data support for predictive control. The system adaptability is significantly enhanced, solving the problem of lack of external data integration capability in the prior art.

[0032] Finally, the modular design of this utility model makes all functional modules independent PCIe interface cards, supporting hot-swapping, which greatly shortens maintenance time and significantly improves maintenance convenience. The standardized PCIe interface supports flexible configuration and online upgrades of functional modules, significantly extending the system life cycle. In case of failure, only a single module needs to be replaced instead of the entire device, greatly reducing maintenance costs and demonstrating significant economic advantages. Attached Figure Description

[0033] Figure 1 This is a structural diagram of the present invention;

[0034] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0035] The technical solution of this utility model will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this utility model, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this utility model without creative effort are within the protection scope of this utility model.

[0036] It should be noted that the directional concepts of "left", "right", "up", "down", "front", "back", "inner", and "outer" in the following scheme are all relative directions, and will not be listed one by one here.

[0037] Example 1

[0038] like Figure 1 and Figure 2 As shown, this embodiment provides an intelligent sensing and management device based on FPGA digital twin. The device is a standard rack-mount control cabinet with an overall height of 2U standard rack size, a width of 19 inches, and a depth of 450mm. The control cabinet is equipped with an industrial-grade switching power supply. This industrial-grade switching power supply adopts a wide voltage input design with an input voltage range of AC85V-264V and an output power of 500W. It has overvoltage, overcurrent, and overtemperature protection functions, providing a stable and reliable power supply for the entire device.

[0039] The core component of the control cabinet is the PCIe communication backplane. This backplane adopts a standard PCIe Gen 3.0 bus architecture and is manufactured using a six-layer printed circuit board process. It features four standardized PCIe slots, each a PCIe x4 slot supporting hot-swapping. The backplane also includes standardized power distribution rails for all internal modules, including a +12V main power rail, a +5V auxiliary power rail, a +3.3V logic power rail, and multiple independent ground rails. Each power rail is equipped with an independent filter capacitor array and fuse protection. Furthermore, the backplane also features a data bus channel for high-speed data exchange between modules. This data bus channel uses differential signal transmission, with a single-channel transmission rate of up to 8Gbps and a total backplane bandwidth of 128Gbps.

[0040] The FPGA digital twin parallel processing module is the core sensing processing unit in this embodiment. This module is an independent PCIe interface board, manufactured using a four-layer printed circuit board process. The FPGA digital twin parallel processing module is connected to the PCIe communication main control backplane through the PCIe gold finger structure. The PCIe gold fingers are gold-plated with a gold layer thickness of 30 micrometers to ensure contact reliability for long-term use.

[0041] The FPGA digital twin parallel processing module integrates a multi-source industrial signal interface unit. This multi-source industrial signal interface unit is located in the I / O baffle area of ​​the FPGA digital twin parallel processing module. The multi-source industrial signal interface unit includes four opto-isolated RS485 terminals. Each RS485 terminal integrates an optocoupler, with an isolation voltage of up to 2500V and a transmission rate of up to 115200bps. The RS485 terminals are used to connect to smart meters to realize real-time acquisition of electricity consumption data. The multi-source industrial signal interface unit also includes two BNC coaxial interfaces with internal high-frequency signal conditioning circuits. The high-frequency signal conditioning circuit uses a signal shaping circuit composed of operational amplifiers and comparators, which can process pulse signals with a frequency range of 1Hz-10kHz. The BNC coaxial interfaces are used to receive pulse signals from water flow sensors and accurately measure water flow through pulse counting.

[0042] The multi-source industrial signal interface unit further includes a set of eight channels with built-in preamplifiers and multi-channel differential signal input terminals. The preamplifier gain of each channel is adjustable from 1 to 1000 times, the input impedance is 1MΩ, and the noise figure is less than 3dB. The multi-channel differential signal input terminals are used to connect mechanical vibration sensors or acoustic sensors, which can capture weak vibration and acoustic signals. The multi-source industrial signal interface unit also includes a wireless communication interface with a Zigbee or LoRa RF transceiver module physically soldered on. The RF transceiver module operates at a frequency of 2.4GHz (Zigbee) or 433MHz (LoRa), with a transmit power of 20dBm and a receive sensitivity of -100dBm. The RF transceiver module is connected to an external antenna port via an onboard RF cable. The external antenna port adopts an SMA connector and can connect to an omnidirectional antenna with a gain of 3dBi.

[0043] The FPGA digital twin parallel processing module also integrates an FPGA parallel processing core unit, which includes an FPGA chip fixedly installed in the central area of ​​the FPGA digital twin parallel processing module. The FPGA chip has 215,360 logic units, 13,140Kb of Block RAM and 740 DSP slices. The FPGA chip has three parallel logic processing areas programmed and solidified inside it using a hardware description language. These three logic processing areas adopt an independent clock domain design inside the FPGA chip and do not interfere with each other.

[0044] The first logic processing area is dedicated to the synchronous acquisition and timestamp alignment of multi-source heterogeneous sensor data under a high-precision hardware clock. This logic processing area integrates a high-precision clock generator with a clock accuracy of 10ppm, which can add a unified timestamp to all sensor data with a timestamp resolution of 1 microsecond. The first logic processing area also includes a multi-channel parallel analog-to-digital converter controller, which can simultaneously process the digital conversion of 32 channels of analog signals.

[0045] The second logic processing area is dedicated to receiving the wireless temperature data stream and sensor spatial coordinate information aligned by the first logic processing area. It constructs and refreshes the three-dimensional thermal map model of the equipment room through parallel pipelined operations. The logic processing area implements a thermal field calculation algorithm based on the finite element method. It can generate a three-dimensional temperature distribution map of the entire equipment room space through interpolation operations based on the spatial distribution of temperature sensors and measured temperature values. The temperature resolution reaches 0.1℃ and the spatial resolution reaches 0.5 meters. The parallel pipelined architecture of the second logic processing area enables the thermal map to be refreshed at a frequency of up to 10Hz.

[0046] The third logic processing area is dedicated to receiving the mechanical status data stream aligned by the first logic processing area. It uses a built-in FFT (Fast Fourier Transform) and feature comparison algorithm to evaluate the equipment's operational health status in real time. This logic processing area implements a 1024-point FFT operation unit, which can perform frequency domain analysis on vibration signals, extract the characteristic frequencies of equipment operation, and compare them with a pre-stored equipment health status feature library to determine the equipment's operating status in real time, including three levels: normal, warning, and fault.

[0047] The FPGA digital twin parallel processing module also features an onboard cache and DMA transfer unit. This unit includes a high-speed data cache Block RAM and a direct memory access DMA controller module configured within the FPGA chip. The high-speed data cache Block RAM has a total capacity of 1MB and is divided into 8 independent cache areas, each corresponding to a different data type. The direct memory access DMA controller module is directly connected to the PCIe gold fingers and adopts a multi-channel DMA architecture with a single-channel transfer rate of up to 1GB / s. The DMA controller module is used to process the 3D heat map and device health status map generated by the second and third logic processing areas and directly transmit them at high speed to the AI ​​strategy and cost decision module via the PCIe communication main control backplane without CPU intervention.

[0048] The physical field drive and redundancy control module serves as the execution control unit in this embodiment. This module is an independent power drive and relay control board, manufactured using a six-layer printed circuit board process, two of which are high-current copper foil layers. The physical field drive and redundancy control module is equipped with an airflow field reconstruction drive array consisting of sixteen physically isolated stepper motor drive chips. Each stepper motor drive chip is an A4988 model with a maximum drive current of 2A and supports 1 / 16 microstepping control. Each drive channel of the airflow field reconstruction drive array is connected to an intelligent damper through a high-current output terminal block. The high-current output terminal block has a rated current of 10A and can withstand an instantaneous impact current of 20A. Each intelligent damper is equipped with a position feedback sensor, which can achieve precise control of the damper opening with a control accuracy of 1%.

[0049] The physical field drive and redundant control module is also equipped with a hybrid refrigeration control relay matrix consisting of solid-state relays and electromagnetic relays. This relay matrix includes eight solid-state relays and four electromagnetic relays. The solid-state relays have a rated load of 25A / 480VAC and a switching time of less than 1ms. They are used to control the variable frequency water pump and fan of the evaporative refrigeration system. The electromagnetic relays have a contact capacity of 5A / 250VAC and have four sets of normally open contacts. They control the start and stop of the standby air conditioning system through dry contact.

[0050] The physics field drive and redundancy control module also integrates an onboard safety monitoring microcontroller (MCU). This MCU uses an STM32F103 series microcontroller with a 72MHz operating frequency and 64KB Flash memory. The onboard safety monitoring microcontroller has a built-in watchdog function with a timeout of 5 seconds. This watchdog is used to execute preset hardware-level fault safety control strategies when communication with the AI ​​strategy and cost decision module fails. The fault safety control strategies include: immediately shutting down all intelligent air valves, switching to the backup air conditioning system, and sending fault alarm signals to ensure the safe operation of the computer room equipment.

[0051] Next to the FPGA chip is a dedicated power regulator chip LDO. This power regulator chip LDO has an output voltage of 1.0V, an output current capability of 500mA, and a power supply rejection ratio of 79dB. The power regulator chip LDO provides a stable, low-ripple core voltage for the FPGA chip, ensuring stable operation of the FPGA chip under various operating conditions. The FPGA chip is connected to an external configuration flash memory SPI Flash via the SPI bus. This configuration flash memory has a capacity of 16Mbit and an operating frequency of up to 75MHz. The configuration flash memory SPI Flash is used to store and load the internal logic configuration file of the FPGA chip, and supports online updates and version management of the configuration file.

[0052] The FPGA chip surface is covered with an aluminum heatsink with heat dissipation fins. The heatsink is made of 6063 aluminum alloy and has a total heat dissipation area of ​​100 square centimeters. The heatsink is tightly attached to the FPGA chip surface with thermal grease and has a thermal conductivity of 3.0 W / m·K, which can effectively dissipate the heat generated by the FPGA chip into the environment. The edge of the FPGA digital twin parallel processing module board is equipped with a JTAG debugging interface and LED status indicators. The JTAG debugging interface uses a standard 14-pin connector and is compatible with Xilinx Platform Cable USB debuggers. The LED status indicators include four types: power indicator, running status indicator, communication status indicator, and fault indicator. The LED status indicators are used to display the status of each interface unit, logic processing area, and communication, which facilitates system debugging and fault diagnosis.

[0053] The working process of this embodiment is as follows:

[0054] After the system is powered on, the industrial-grade switching power supply first provides a stable power supply to the PCIe communication main control backplane. The PCIe communication main control backplane supplies power to each modular board through a standardized power distribution rail. After the FPGA digital twin parallel processing module is powered on, the configuration flash SPI Flash automatically loads the preset logic configuration file into the FPGA chip. After the FPGA chip completes the configuration, it begins to work normally.

[0055] The multi-source industrial signal interface unit begins collecting data from various sensors. The RS485 terminal polls the connected smart meter at a 1-second interval to obtain real-time power consumption, voltage, current, and other parameters. The BNC coaxial interface continuously monitors the pulse signal from the water flow sensor, calculating the instantaneous and cumulative water flow through pulse counting. The multi-channel differential signal input terminal continuously collects signals from the mechanical vibration sensor and acoustic sensor at a sampling frequency of 10kHz. The wireless communication interface maintains communication with the temperature sensor array distributed in the computer room, collecting temperature data from various locations at a 5-second interval.

[0056] After receiving various sensor data, the first logic processing area inside the FPGA chip immediately adds a high-precision timestamp to each data packet and performs time synchronization and alignment on the data from different interfaces to ensure that the data processed subsequently has time consistency. The synchronized and aligned data is then distributed to the second and third logic processing areas for parallel processing.

[0057] After receiving the synchronized temperature data and sensor spatial coordinate information, the second logic processing area constructs a three-dimensional thermal map model of the computer room using the built-in finite element algorithm. The algorithm first establishes a three-dimensional mesh model based on the spatial distribution of the sensors, and then uses the measured temperature values ​​as boundary conditions to calculate the temperature distribution of the entire computer room space through numerical solution methods. The generated three-dimensional thermal map data includes the temperature value and gradient information of each grid point, and the data format is a floating-point array.

[0058] After receiving the synchronized mechanical status data, the third logic processing area performs frequency domain analysis on the vibration signal using the built-in FFT (Fast Fourier Transform) algorithm. The algorithm extracts the power spectral density of the signal, identifies the characteristic frequencies of the equipment operation, and compares them with the pre-stored health status feature library. Based on the comparison results, the algorithm outputs the health status assessment of the equipment, including information such as bearing status, impeller balance status, and motor operating status.

[0059] The onboard cache and DMA transfer unit temporarily stores the 3D heat map data generated by the second logic processing area and the device health status data generated by the third logic processing area in the high-speed data cache Block RAM. When the data cache reaches the preset threshold or the timed trigger condition is met, the DMA controller module starts the data transmission process and transmits the processing result to the AI ​​strategy and cost decision module at high speed through the PCIe communication main control backplane. The entire transmission process does not require CPU intervention and the transmission delay is less than 1 millisecond.

[0060] After receiving control commands from the AI ​​strategy and cost decision module, the physics field drive and redundancy control module immediately executes the corresponding physical control actions. The airflow field reconstruction drive array adjusts the opening of each intelligent air valve according to the commands, thereby optimizing local cooling by changing the airflow distribution in the computer room. The hybrid cooling control relay matrix controls the operating parameters of the evaporative cooling system according to the commands, including water pump speed, fan speed, etc., while monitoring the operating status of the backup air conditioning system. It can quickly switch to the backup cooling system when the main cooling system fails.

[0061] The onboard safety monitoring microcontroller (MCU) continuously monitors the operating status of the entire system, including the communication status with the upper-level control module, the working status of each actuator, system temperature, and other parameters. When an abnormality is detected, the MCU immediately executes the fault safety control strategy to ensure the safe operation of the equipment in the computer room.

[0062] The system employs an FPGA parallel processing architecture to achieve real-time processing of multi-source sensor data, with a data processing latency of less than 10 milliseconds. Compared to the traditional CPU serial processing method, the processing speed is increased by more than 50 times, meeting the stringent real-time requirements of the data center air conditioning system. Secondly, three independent logical processing areas enable parallel processing of data acquisition, model building, and status assessment, increasing system throughput by 300% and enabling the simultaneous processing of data streams from hundreds of sensors. Thirdly, DMA direct memory access technology enables zero-copy transmission of processing results, eliminating the performance bottleneck caused by CPU involvement in data transmission and improving data transmission efficiency by 80%. In addition, hardware-level fault-tolerant mechanisms ensure reliable operation of the system under abnormal conditions, achieving a system availability of over 99%.

[0063] Application examples:

[0064] This project involved an energy-saving renovation of the air conditioning system in a large data center. The data center has a floor area of ​​1,000 square meters and houses 500 servers. The original air conditioning system used a traditional constant temperature and humidity control method, consuming 2 million kilowatt-hours of electricity annually. By deploying the intelligent sensing and control system described in this embodiment, 100 wireless temperature sensors, 20 water flow sensors, and 16 vibration sensors were installed in the data center. The original 24 precision air conditioners were also upgraded with intelligent air valves and variable frequency control functions.

[0065] After the system is running, the FPGA digital twin parallel processing module collects data from various sensors in real time, builds a real-time three-dimensional thermal map model of the computer room, accurately grasps the temperature distribution in the computer room, and promptly detects bearing wear problems in 3 air conditioning units and refrigerant leakage problems in 2 units through equipment health status monitoring, thus avoiding computer room overheating accidents caused by equipment failure.

[0066] The physics-driven and redundant control module optimizes airflow distribution based on a thermal map model. By adjusting the opening of intelligent air valves, it precisely delivers cool air to high heat load areas, avoiding waste of cool air. At the same time, it dynamically adjusts the operating parameters of the refrigeration system according to the real-time load, achieving efficient operation of the refrigeration system while ensuring stable computer room temperature.

[0067] After a year of operation, the data center's air conditioning system achieved an energy saving rate of 30%, and the temperature fluctuation in the computer room was controlled within ±0.5℃, which is better than the ±2℃ fluctuation range of the original system. The equipment failure rate was reduced by 60%, and the system reliability was significantly improved.

[0068] Example 2

[0069] like Figure 1 and Figure 2 As shown, this embodiment provides an AI management device for intelligent optimization of data center air conditioning energy saving based on AI strategy decision-making, based on embodiment 1. The device also adopts a standard rack-mount control cabinet structure with an overall height of 2U standard rack specifications. The control cabinet is equipped with an industrial-grade switching power supply. The industrial-grade switching power supply adopts a wide voltage input design with an input voltage range of AC85V-264V and an output power of 500W. It has overvoltage, overcurrent, and overtemperature protection functions, providing a stable and reliable power supply for the entire device.

[0070] The core component of the control cabinet is the PCIe communication backplane. This backplane adopts a standard PCIe Gen 3.0 bus architecture and is manufactured using a six-layer printed circuit board process. It features four standardized PCIe slots, each a PCIe x4 slot supporting hot-swapping. The backplane also includes standardized power distribution rails for all internal modules, including a +12V main power rail, a +5V auxiliary power rail, a +3.3V logic power rail, and multiple independent ground rails. Each power rail is equipped with an independent filter capacitor array and fuse protection. Furthermore, the backplane also features a data bus channel for high-speed data exchange between modules. This data bus channel uses differential signal transmission, with a single-channel transmission rate of up to 8Gbps and a total backplane bandwidth of 128Gbps.

[0071] The AI ​​strategy and cost decision-making module is the core intelligent decision-making unit in this embodiment. This module is a high-performance COM-Express standard core board. The AI ​​strategy and cost decision-making module adopts a separate design of carrier board and core board. The core board is connected to the carrier board through a 440-pin B2B connector. The connector adopts a gold-plating process with a gold layer thickness of 50 micrometers to ensure the reliability of high-speed signal transmission.

[0072] The AI ​​strategy and cost decision-making module is equipped with an edge computing processor with its own GPU or NPU. This edge computing processor integrates a 512-core Volta GPU and an 8-core CPU, with a GPU computing power of 32 TOPS (INT8 precision) and a CPU clock speed of 2.26GHz. The processor supports deep learning frameworks such as CUDA, cuDNN, and TensorRT, and can efficiently run complex AI algorithm models. The processor is equipped with 32GB of LPDDR4x memory, with a memory bandwidth of 137GB / s, providing ample memory support for large-scale data processing and model inference.

[0073] The AI ​​strategy and cost decision-making module also features an M.2 slot, which is an M.2 2280 form factor slot that supports a PCIe 3.0 x4 interface. A 1TB solid-state drive (SSD) is installed in this slot, boasting sequential read speeds of 7000MB / s, sequential write speeds of 5000MB / s, and random read and write IOPS of 1000K and 1000K respectively. The SSD utilizes 3D V-NAND TLC flash memory technology, offering high reliability and long lifespan, with a mean time between failures (MTBF) of 1.5 million hours.

[0074] Solid-state drives (SSDs) are used for physical storage of historical digital twin data and an operational cost model database. The historical digital twin data includes data on data distribution in the data center over the past 24 months, equipment operating status data, energy consumption data, etc., with a total data volume of approximately 500GB. The operational cost model database includes time-of-use electricity price data and tiered water price data. The time-of-use electricity price data covers the electricity price change patterns in different regions and seasons, including the specific values ​​and time periods for peak, off-peak, and valley electricity prices. The tiered water price data includes water price standards corresponding to different water consumption levels, as well as seasonal adjustment coefficients. The operational cost model database also includes equipment depreciation models, maintenance cost models, labor cost models, etc., providing data support for comprehensive cost optimization decisions.

[0075] The AI ​​strategy and cost decision-making module directly reads the 3D heat map and device health status map generated by the FPGA digital twin parallel processing module from memory via DMA through the PCIe communication main control backplane. The DMA transfer channel adopts a multi-queue architecture, supporting the simultaneous processing of 16 independent data streams, each with a bandwidth of up to 500MB / s. Reading data via DMA avoids the performance loss caused by CPU involvement in data movement, and the data reading latency is controlled within 50 microseconds, which can meet the strict requirements of real-time decision-making algorithms for data timeliness.

[0076] The AI ​​strategy and cost decision-making module operates a multi-layered intelligent decision-making algorithm framework, which consists of four main components: a data preprocessing layer, a feature extraction layer, a model inference layer, and a decision output layer. The data preprocessing layer is responsible for performing preprocessing operations such as format conversion, noise filtering, and data normalization on the received 3D heatmap and equipment health status map to ensure the quality and consistency of the input data. The feature extraction layer adopts a deep learning architecture that combines convolutional neural networks (CNN) and recurrent neural networks (RNN) to extract spatial features from the heatmap and temporal features from the time series data, forming a high-dimensional feature vector.

[0077] The model inference layer runs a pre-trained reinforcement learning model, which adopts the Deep Q-Network (DQN) algorithm framework. The network structure includes 3 convolutional layers, 2 fully connected layers, and 1 output layer, with a total of approximately 2 million parameters. The reinforcement learning model uses the power efficiency ratio (PUE) of the data center as the reward function and temperature distribution uniformity, equipment health, and operating costs as constraints. It learns the optimal air conditioning control strategy through training on a large amount of historical data. The model inference process uses GPU parallel computing, and the time for a single inference is controlled within 10 milliseconds, enabling real-time policy output.

[0078] The decision output layer converts the model inference results into specific control commands, including the target opening degree of each smart air valve, the operating parameters of the evaporative cooling system, and the start-stop status of the standby air conditioning system. The decision output also includes cost optimization suggestions, such as increasing cooling capacity reserves during off-peak electricity price periods and reducing cooling intensity in non-critical areas during peak electricity price periods.

[0079] The AI ​​strategy and cost decision-making module also integrates a cost-benefit analysis engine. This engine uses a multi-objective optimization algorithm to comprehensively consider multiple dimensions such as electricity costs, water costs, equipment depreciation costs, and maintenance costs to calculate the total operating costs under different control strategies. The cost-benefit analysis engine has a built-in genetic algorithm optimizer with a population size of 100 individuals and an iteration count of 1000 generations. It finds the optimal cost control strategy through global search, and the analysis results are output in report form, including real-time cost analysis, predicted cost trends, and energy-saving potential assessment.

[0080] The Federal Security Communication Module serves as the security communication guarantee unit in this embodiment. This module is an independent communication interface board manufactured using a four-layer printed circuit board process. The Federal Security Communication Module integrates a Mini-PCIe slot for plugging in a 5G or LTE wireless communication module. This Mini-PCIe slot uses a standard 52-pin interface and supports both PCIe and USB dual interface modes. The 5G wireless communication module supports the 5G NR Sub-6GHz frequency band, with a downlink peak rate of 2.5Gbps and an uplink peak rate of 900Mbps, and supports 5G NSA and SA dual-mode networking. The LTE wireless communication module supports the LTE Cat 4 standard, with a downlink peak rate of 150Mbps and an uplink peak rate of 50Mbps, and supports major global frequency bands.

[0081] The Federal Security Communication Module also has a Trusted Platform Module (TPM) security encryption chip physically soldered onto it. This TPM security encryption chip conforms to the TPM 2.0 standard specification. The TPM security encryption chip integrates security components such as a hardware random number generator, an RSA-2048 encryption engine, a SHA-256 hash algorithm engine, and an AES-256 symmetric encryption engine. The TPM chip is used to perform hardware-level encryption on externally transmitted data, including security functions such as device authentication, digital signature, and key management.

[0082] The TPM security encryption chip uses the Elliptic Curve Digital Signature Algorithm (ECDSA) for device authentication, employing the P-256 elliptic curve parameter and a 256-bit private key length to ensure secure authentication. For data transmission encryption, the TPM chip uses the AES-256-GCM algorithm, which provides both data encryption and integrity verification. The key update cycle is set to 24 hours to ensure the security of long-term communication. The TPM chip also features anti-tampering capabilities; it automatically clears the internal key when a physical attack is detected to prevent the leakage of sensitive information.

[0083] The Federal Security Communications Module is responsible for providing a network channel for the AI ​​Strategy and Cost Decision Module to obtain external weather data. It connects to the National Meteorological Administration's weather data server via 5G or LTE networks to obtain local meteorological parameters such as temperature, humidity, wind speed, wind direction, and atmospheric pressure in real time. The weather data is obtained once per hour, and the data format is JSON. The data transmission uses the HTTPS protocol to ensure the security and integrity of the data transmission.

[0084] The federated secure communication module also supports federated learning, which can establish a secure model parameter sharing channel with other similar devices. Without leaking local data, it can improve the performance of the overall AI model by aggregating model parameters. Federated learning uses differential privacy technology to add carefully designed noise to the model parameters to ensure that the data privacy of individual devices cannot be inferred. The federated learning framework supports asynchronous update mode, and participating devices can flexibly choose when to participate based on their own computing power and network conditions.

[0085] The Federal Security Communication Module also integrates a network security monitoring engine. This engine uses Deep Packet Inspection (DPI) technology to monitor abnormal network traffic behavior in real time, including security threats such as DDoS attacks, port scanning, and malicious code propagation. The security monitoring engine has a built-in Intrusion Detection System (IDS) that uses a combination of signature detection and anomaly detection, achieving an accuracy rate of over 99.5% and a false positive rate of less than 0.1%. When a security threat is detected, the system automatically initiates protective measures, including blocking suspicious connections, isolating affected modules, and sending security alerts.

[0086] The working process of this embodiment is as follows:

[0087] After the system is powered on, the AI ​​strategy and cost decision module first loads the pre-trained deep learning model and operating cost model database from the solid-state drive. The model loading process takes about 30 seconds. At the same time, the federal security communication module starts the network connection and establishes a secure communication link with the external server through the 5G or LTE network to complete the device identity authentication and key exchange process.

[0088] The AI ​​strategy and cost decision module continuously reads the 3D heat map and equipment health status map data generated by the FPGA digital twin parallel processing module via DMA through the PCIe communication main control backplane. The data reading frequency is 10 times per second to ensure that the decision algorithm is based on the latest data center status information. At the same time, the federal security communication module obtains external weather data from the meteorological server once an hour, including the weather forecast information for the next 24 hours, to provide data support for predictive control strategies.

[0089] After receiving real-time data, the data preprocessing layer of the AI ​​strategy and cost decision module immediately performs spatial filtering on the 3D heat map to eliminate the influence of sensor noise and measurement errors, performs time-series smoothing on the equipment health status map, and extracts the changing trend of equipment operating status. The preprocessed data is then sent to the feature extraction layer, where key features are extracted through a deep learning network.

[0090] The convolutional neural network of the feature extraction layer performs multi-scale feature extraction on the heat map, identifying spatial features such as hot spots, cold spots, and temperature gradient distribution in the computer room. The recurrent neural network analyzes the time series data, identifying time features such as periodic changes and abnormal fluctuations in equipment load. The extracted feature vector has a dimension of 512, containing key information about the thermal environment and equipment status of the computer room.

[0091] After receiving the feature vector, the reinforcement learning model in the model inference layer combines the current operating cost model database information, including cost parameters such as real-time electricity price and water price, and calculates the value function of each possible control action through a deep Q network. The model also considers the impact of external weather data, such as increasing the cooling capacity in advance when the outside temperature rises and optimizing the operation strategy of the evaporative cooling system during the rainy season. The model inference results are output in the form of a probability distribution, representing the degree of superiority or inferiority of different control strategies.

[0092] Based on the model inference results and combined with the current system state and constraints, the decision output layer generates the optimal control strategy. The control strategy includes two levels: short-term control instructions and long-term optimization suggestions. The short-term control instructions are for adjusting the equipment operating parameters for the next hour, including the target opening adjustment of each smart damper, the adjustment of the water pump speed of the evaporative cooling system, and the adjustment of the fan speed. The long-term optimization suggestions are for the operating strategy for the next 24 hours, such as using the low electricity price period at night for pre-cooling and reducing the cooling intensity of non-critical areas during the high electricity price period during the day.

[0093] The cost-benefit analysis engine runs multi-objective optimization algorithms in parallel to calculate the cost-benefit of different control strategies in real time. The analysis dimensions include immediate costs (electricity and water costs for the current hour), cumulative costs (cumulative operating costs for the day), and predicted costs (expected costs for the next 24 hours). The analysis results are displayed in the form of charts, including cost trend curves, energy-saving effect comparisons, and investment payback period calculations, providing operation and maintenance personnel with intuitive decision-making references.

[0094] The Federation Security Communication Module continuously monitors network security status throughout the entire operation and performs hardware-level encryption on all external communication data. The TPM security encryption chip generates a unique session key for each data transmission, ensuring that even if a single communication is intercepted, it will not affect the overall security. At the same time, the module regularly participates in the federated learning process, exchanges model parameters with other similar devices, and continuously optimizes the performance of the local AI model.

[0095] After receiving the control commands from the AI ​​strategy and cost decision module, the physics field drive and redundancy control module immediately executes the corresponding physical control actions. The control execution process adopts a hierarchical control strategy. High-priority commands (such as safety-related controls) are executed immediately, while low-priority commands (such as energy efficiency optimization controls) can be executed with a delay. The control execution results are fed back to the AI ​​strategy and cost decision module through the PCIe communication main control backplane, forming a closed-loop control system.

[0096] The AI ​​decision-making algorithm based on deep learning achieves multi-objective optimization control, taking into account multiple dimensions such as energy efficiency, cost, and reliability. The decision quality is improved by more than 40% compared with traditional rule control. Secondly, the establishment of an operating cost model database enables refined cost control. Through time-of-use electricity pricing optimization and tiered water pricing optimization, operating costs are reduced by more than 25%. Thirdly, the application of federated learning technology enables multi-device collaborative optimization. Through model parameter sharing, the accuracy of the overall AI model is improved by 15%, and the convergence speed is improved by 30%. In addition, hardware-level security encryption mechanism ensures the security of data transmission and model sharing, meeting the high security requirements of the industrial IoT environment.

[0097] Application examples:

[0098] The data center intelligent management system project of a financial institution's headquarters building covers an area of ​​2,000 square meters and is divided into two computer room areas, A and B. It has deployed 1,000 high-performance servers and storage devices, with an annual power consumption of 4 million kWh and an annual water consumption of 20,000 tons. The original air conditioning control system used a traditional PID control algorithm, which could not adapt to the rapid changes in business load and often caused local overheating or overcooling. The power efficiency ratio (PUE) has been maintained above 1.8 all year round.

[0099] By deploying the AI ​​strategy decision-making intelligent optimization communication system of this embodiment, a complete intelligent management network was established in the data center. The system first deployed an AI strategy and cost decision-making module in each of the two computer room areas, A and B. Each module was equipped with a 1TB solid-state drive to store historical data and cost models. The operating cost model database contained three years of local time-of-use electricity price data, including different electricity price standards for weekdays and holidays, as well as the billing rules for tiered water pricing.

[0100] The federal security communication module connects to the local meteorological bureau's weather data server via a 5G network to obtain real-time meteorological information such as outdoor temperature, humidity, and wind speed. At the same time, the system establishes a secure connection with the State Grid's real-time electricity price system, enabling it to obtain electricity price change information for the next 24 hours, providing data support for predictive control. The two AI strategies and cost decision modules establish a model parameter sharing mechanism through a federated learning protocol, realizing collaborative optimization control between the A and B computer room areas.

[0101] After the system was put into operation, the AI ​​strategy and cost decision module analyzed the heat load distribution pattern of the data center through deep learning algorithms, discovered the different load characteristics of the business system on weekdays and holidays, and learned to appropriately increase the cooling capacity during the off-peak electricity price period (22:00-6:00 the next day) and store "cooling capacity" by utilizing the thermal inertia of the data center. During the peak electricity price period (9:00-11:00 in the morning and 14:00-17:00 in the afternoon), the cooling intensity was reduced, realizing intelligent scheduling of "peak shaving and valley filling".

[0102] The cost-benefit analysis engine dynamically adjusts the operation strategy of the refrigeration system based on real-time electricity and water price information. During the summer peak electricity consumption period, the system automatically optimizes the ratio of evaporative refrigeration and mechanical refrigeration, making full use of the high efficiency of evaporative refrigeration and reducing the power consumption of mechanical refrigeration. During the winter off-peak electricity consumption period, the system appropriately increases the proportion of mechanical refrigeration, creating conditions for the maintenance of the evaporative refrigeration system.

[0103] Federated learning enables AI models in two data center areas to learn from each other's optimization experiences. Data center A experiences greater load fluctuations, so the system learns a fast response strategy; data center B experiences relatively stable loads, so the system learns a refined control strategy. Through model parameter sharing, both areas gain each other's optimization experience, resulting in a significant improvement in overall control performance.

[0104] In terms of secure communication, the TPM security encryption chip ensures the security of all external communication data, including weather data acquisition, electricity price information inquiry, and federated learning parameter exchange. In the two years since the system started operating, no data breaches or cybersecurity incidents have occurred, meeting the stringent data security requirements of the financial industry.

[0105] After two years of operation and verification, the data center's power efficiency ratio (PUE) has been reduced to 1.35, a 25% improvement compared to the original system. The power saving rate has reached 20%, the water saving rate has reached 20%, the temperature control accuracy of the computer room has been improved to ±0.3℃, the equipment operation stability has been significantly improved, and the failure rate has been reduced by 45%.

[0106] The system also brought about a significant improvement in operation and maintenance efficiency. Through the predictive maintenance function of AI algorithms, 12 potential equipment failures were detected in advance, avoiding unplanned downtime events. The cost-benefit analysis function provided the operation and maintenance team with a scientific basis for decision-making, optimized maintenance plans and spare parts procurement strategies, and reduced operation and maintenance costs by 30%.

[0107] The above description is only a preferred embodiment of the present utility model. It should be understood that the present utility model is not limited to the form disclosed herein and should not be regarded as an exclusion of other embodiments. It can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and changes made by those skilled in the art that do not depart from the spirit and scope of the present utility model should be protected within the scope of the appended claims.

Claims

1. A machine room air conditioning energy-saving AI management device, characterized by, The device is a standard rack-mount control cabinet structure, including: an industrial-grade switching power supply; a PCIe communication main control backplane, which is provided with multiple standardized PCIe slots, and is equipped with standardized power distribution rails for powering all internal modules and data bus channels for realizing high-speed data exchange between modules. The FPGA digital twin parallel processing module is an independent PCIe interface board that connects to the PCIe communication main control backplane via a PCIe gold finger structure. The FPGA digital twin parallel processing module integrates a multi-source industrial signal interface unit, which includes several opto-isolated RS485 terminals, at least one BNC coaxial interface with an internal high-frequency signal conditioning circuit, a set of multi-channel differential signal input terminals with built-in preamplifiers, and a wireless communication interface with a physically soldered Zigbee or LoRa RF transceiver module. The FPGA digital twin parallel processing module also integrates an FPGA parallel processing core unit, which includes an FPGA chip fixedly installed in the central area of ​​the FPGA digital twin parallel processing module. The FPGA chip has three parallel-working logic processing areas programmed and solidified internally using a hardware description language. The FPGA digital twin parallel processing module also has an onboard cache and DMA transfer unit, which includes a high-speed data cache Block RAM and a direct memory access DMA controller module configured within the FPGA chip. The AI ​​strategy and cost decision module is a high-performance COM-Express standard core board. The AI ​​strategy and cost decision module is equipped with an edge computing processor with a built-in GPU or NPU. The AI ​​strategy and cost decision module is also equipped with an M.2 slot, on which a solid-state drive is installed. The physical field drive and redundancy control module is an independent power drive and relay control board. The physical field drive and redundancy control module is equipped with an airflow field reconstruction drive array composed of multiple physically isolated stepper motor drive chips and a hybrid cooling control relay matrix composed of solid-state relays and electromagnetic relays. The physical field drive and redundancy control module also integrates an onboard safety monitoring microcontroller (MCU). The Federal Security Communication Module is an independent communication interface board. The Federal Security Communication Module integrates a Mini-PCIe slot for plugging in a 5G or LTE wireless communication module and a Trusted Platform Module (TPM) security encryption chip physically soldered onto the board.

2. The energy-saving AI management device for a machine room air conditioner according to claim 1, characterized by: The multi-source industrial signal interface unit is located in the I / O baffle area of ​​the FPGA digital twin parallel processing module. The RS485 terminal is used to connect to a smart meter. The BNC coaxial interface is used to receive pulse signals from a water flow sensor. The multi-channel differential signal input terminal is used to connect to a mechanical vibration sensor or an acoustic sensor. The RF transceiver module is connected to an external antenna port via an onboard RF cable.

3. The energy-saving AI management device for a machine room air conditioner according to claim 2, characterized by: The three parallel logical processing areas are as follows: The first logical processing area is used for high-precision hardware clock-based synchronous acquisition and timestamp alignment of multi-source heterogeneous sensor data. The second logical processing area is used to receive the wireless temperature data stream aligned by the first logical processing area and the sensor spatial coordinate information, and to construct and refresh the three-dimensional thermal map model of the computer room through parallel pipeline operations. The third logic processing area is used to receive the mechanical status data stream aligned by the first logic processing area and to evaluate the equipment's operational health status in real time using the built-in FFT (Fast Fourier Transform) and feature comparison algorithms.

4. The energy-saving AI management device for a machine room air conditioner according to claim 3, characterized by: The DMA controller module is directly connected to the PCIe gold finger and is used to transmit the processing results of the three-dimensional heat map and device health status map generated by the second and third logical processing areas directly to the AI ​​strategy and cost decision module at high speed through the PCIe communication main control backplane without CPU intervention.

5. The energy-saving AI management device for a machine room air conditioner according to claim 2, characterized by: The solid-state drive is used for physical storage of historical digital twin data and an operating cost model database. The operating cost model database includes time-of-use electricity price data and tiered water price data. The AI ​​strategy and cost decision module directly reads the three-dimensional heat map and equipment health status map generated by the FPGA digital twin parallel processing module from memory via the PCIe communication main control backplane in DMA mode.

6. The energy-saving AI management device for a machine room air conditioner according to claim 5, characterized by: Each drive channel of the airflow field reconstruction drive array is connected to an intelligent air valve through a high-current output terminal block. The hybrid refrigeration control relay matrix is ​​used to control the evaporative refrigeration system and the standby air conditioning system through dry contacts, respectively.

7. The energy-saving AI management device for a machine room air conditioner according to claim 6, characterized by: The onboard safety monitoring microcontroller (MCU) has a built-in watchdog function, which is used to execute a preset hardware-level fault safety control strategy when communication with the AI ​​strategy and cost decision module is abnormal.

8. The energy-saving AI management device for a machine room air conditioner according to claim 1, characterized by: The TPM security encryption chip is used for hardware-level encryption of data transmitted externally, and the Federal Security Communication Module is responsible for providing a network channel for the AI ​​Strategy and Cost Decision Module to obtain external weather data.

9. The energy-saving AI management device for a machine room air conditioner according to claim 8, characterized by: The FPGA chip is equipped with a dedicated power regulator chip LDO, which provides a stable, low-ripple core voltage for the FPGA chip. The FPGA chip is connected to an external configuration flash memory SPI Flash via an SPI bus, which is used to store and load the internal logic configuration file of the FPGA chip.

10. The energy-saving AI management device for a machine room air conditioner according to claim 9, characterized by: The FPGA chip is covered with an aluminum radiator with radiating fins, the board edge of the FPGA digital twin parallel processing module is provided with a JTAG debugging interface and an LED state indicating lamp, and the LED state indicating lamp is used for displaying each interface unit, a logic processing area and a communication state.

Images