A Multimodal Fusion Intelligent Operation and Maintenance Method and System for Container Data Centers

Abstract

This invention discloses a multimodal fusion intelligent operation and maintenance method and system for container data centers, belonging to the field of container intelligent operation and maintenance technology. The method includes: S1, collecting multi-source operational data from each container module to form an original multi-source time-series dataset; S2, cleaning the original data and outputting a standardized dataset; S3, extracting operational features from the cleaned data, quantifying the overall health, and generating an equipment health matrix and feature vector; S4, performing fault probability prediction and micro-module-level root cause localization; S5, constructing a multi-objective optimization function, combining optimization algorithms and adaptive PID control to solve for and output the optimal control command; and S6, issuing control commands to the terminal for execution and simultaneously transmitting real-time operational data from the field. This invention effectively improves the quality of operation and maintenance data, completes multi-objective collaborative optimization and control, forms a closed-loop intelligent operation and maintenance process, is suitable for container data center scenarios, and has strong engineering practicality.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for containers, and in particular to a multimodal integrated intelligent operation and maintenance method and system for container data centers. Background Technology

[0002] Containerized data centers, with their advantages of modular integration, flexible deployment, and compact space, are now widely used in edge computing, field sites, and emergency computing scenarios. Due to their high degree of internal equipment integration and enclosed space, environmental and electrical parameters are highly coupled with equipment operating status, requiring significantly higher precision in operation and maintenance than traditional data centers. Existing operation and maintenance solutions mostly directly adopt conventional data center operation and maintenance techniques, often using fixed sampling frequencies for data collection, which easily leads to data redundancy or missing critical data. Furthermore, data cleaning and fusion methods are relatively simple, resulting in a high level of noise and outliers in the raw data, leading to poor data source quality. Simultaneously, existing fault analysis often relies on a single predictive model without considering the physical operating mechanisms of the equipment for constraint correction, resulting in large fault prediction errors and insufficient root cause location accuracy, making it difficult to adapt to the high-density micro-module operation and maintenance needs.

[0003] Furthermore, traditional operation and maintenance optimization often focuses solely on reducing energy efficiency, failing to consider both equipment operational reliability and the safety boundaries of the cabin environment. It lacks a multi-objective collaborative optimization and control mechanism, the overall operation and maintenance process cannot form a complete closed loop, the model lacks adaptive iteration capabilities, and the system architecture has poor layered linkage, lacks offline emergency response capabilities at the edge, and has a low overall level of intelligence, poor on-site adaptability, and poor engineering practicality. Most existing solutions are designed for single-container integrated container data centers and are not adapted to modular container data center architectures with multiple independent functions that can be flexibly spliced ​​on demand. They cannot achieve a balance between independent operation and maintenance of a single module and cross-module collaborative scheduling, making it difficult to meet the long-term stable, refined, unattended operation and maintenance requirements of container data centers. Summary of the Invention

[0004] The purpose of this invention is to provide a multimodal integrated intelligent operation and maintenance method and system for container data centers, thereby solving the aforementioned technical problems.

[0005] To achieve the above objectives, this invention provides a multimodal fusion-based intelligent operation and maintenance method for containerized data centers, comprising the following steps: S1. Collect multi-source operating data of each container module, dynamically adjust the sampling frequency according to the equipment load conditions, and form the original multi-source time series dataset. S2. Clean the original data based on the adaptive downsampling and spatiotemporal weight fusion algorithm, and output a standardized dataset; S3. Extract the operational characteristics of the cleaned data, quantify the overall health of each device, and generate a device health matrix and feature vector. S4. Combine the physical operation mechanism of the equipment with the LSTM-RF-SVR hybrid model to perform fault probability prediction and micro-module-level root cause localization; S5. Construct a multi-objective optimization function under multiple constraints of environment, energy efficiency, and equipment safety, and solve and output the optimal control command by combining optimization algorithm and adaptive PID control. S6 sends control commands to the terminal to execute the control, and at the same time transmits back the real operation data on site to realize the closed loop of operation and maintenance data throughout the entire process.

[0006] Preferably, S1 specifically includes: S11. The sensing layer collects parameters covering four categories of operational data within the container: environment, electrical, equipment, and liquid cooling. Environmental parameters include cabinet inlet and outlet air temperature, relative humidity, and internal pressure difference. Electrical parameters include input voltage, operating current, active power, and UPS load rate. Equipment parameters include server CPU load, hard drive temperature, air conditioning cooling power, and equipment vibration amplitude. Liquid cooling parameters include coolant supply and return temperatures, supply and return pressure difference, flow rate, heat exchange efficiency, pump power, pipeline leakage monitoring values, and cold plate temperature. S12. The sampling frequency is dynamically switched according to the real-time load and operating status of the equipment. High-frequency sampling is used during high load and high-temperature abnormal operating conditions. Low-frequency sampling is performed when the equipment is operating stably, and the sampling frequency is: The data are then aggregated to form the original multi-source time-series dataset.

[0007] Preferably, S2 specifically includes: S21. Using the original multi-source time-series dataset as the processing object, adaptive downsampling is used to perform data stratification and filtering, eliminating redundant and messy data. Simultaneously, a spatiotemporal weighted fusion algorithm is employed to perform weighted fusion operations on multi-sensor data of the same dimension. The fusion formula is as follows: ; In the formula, for Real-time fusion and processing of data; The number of sensors under the same monitoring dimension; To correspond to the dynamic spatiotemporal weighting coefficients, and to satisfy the constraints. ; For the first Raw data collected by each sensor; S22. An improved adaptive 3σ criterion is used to identify and remove outliers, and the fused data is used to complete and repair missing data, finally generating a standardized dataset. The formula for determining outliers is: ; In the formula, This is the deviation amount for identifying data anomalies. for Single-dimensional raw data that is constantly awaiting verification; This represents the time-series moving average of the data. The adaptive decision coefficient has a value range of 1.5 to 3. The moving standard deviation of the data time series.

[0008] Preferably, S3 specifically includes: S31. Conduct data mining based on standardized datasets to extract multi-dimensional time-domain operational feature data, including mean, variance, peak value, parameter change rate, and load volatility. S32. Combining operational characteristic data, calculate the operational health status of each device using a health quantification model. The calculation formula is as follows: ; In the formula, The overall health status of the equipment is measured in the range [0,1]. This represents the total dimension of the extracted operational features; The actual monitored values ​​are characteristic. The reference value for the corresponding feature; S33. Integrate the generated equipment health matrix and feature vector as the basic input data for fault diagnosis analysis.

[0009] Preferably, S4 specifically includes: S41. Based on the basic input data, four types of equipment physical operation mechanisms are introduced: heat flow conduction, electrical loss, airflow coupling in the duct, and heat exchange-flow resistance-pipeline leakage coupling in the liquid cooling system, to correct the prediction bias that is prone to occur in pure data models. S42. The overall failure probability of the equipment is obtained by using a hybrid LSTM-RF-SVR model. The model weight fusion formula is as follows: ; In the formula, This represents the overall probability of equipment failure. , , The fusion weight proportions for the three models LSTM, RF, and SVR are respectively, and the constraints are satisfied. ; S43. Set fault trigger threshold When the overall failure probability exceeds the threshold At that time, the root cause localization process is initiated to achieve four levels of localization: container module level, cabinet module level, sensor location, and actuator level, and output fault risk information including fault type, level, and location information.

[0010] Preferably, S5 specifically includes: S51. Combining container cabin environmental data and fault risk information, define the hard operational constraints for the entire process, including ambient temperature. Ambient humidity Energy efficiency indicators Equipment failure rate Liquid cooling pump unit operating power Coolant flow rate Liquid cooling supply and return hydraulic differential The constraint ranges are as follows: ; ; ; ; ; ; ; In the formula, The lower limit of the cabin temperature, This is the upper limit of the cabin temperature. This is the lower limit of relative humidity. This is the maximum permissible relative humidity limit. This represents the upper limit of the allowable energy efficiency value for data centers. The upper limit of the allowable failure rate of the equipment. This is the lower limit of the pump set power. This is the upper limit of the pump set power. This is the lower limit of coolant flow rate. This is the upper limit of coolant flow rate. This is the lower limit of the differential pressure between the liquid cooling supply and return pressures. This is the upper limit of the differential pressure between the liquid cooling supply and return pressures. S52. Construct a multi-objective optimization function within the constraints, with the function expression as follows: ; In the formula, To comprehensively optimize the target value, , , , The optimized weights are for four indicators: energy efficiency, failure probability, temperature deviation, and liquid cooling supply and return liquid temperature difference deviation. Real-time temperature inside the cabin. The reference temperature for the equipment. To provide real-time supply and return liquid temperature difference for liquid cooling, This is the rated heat exchange temperature difference for liquid cooling; S53. Call the adaptive PID control algorithm to complete the tuning of the closed-loop control parameters of the equipment. The control parameters include the proportional coefficient. Integral coefficient Differential coefficients The control parameters are adaptively tuned according to the amplitude of thermal fluctuations within the cabin, ultimately outputting the optimal control command. The PID adjustment formula is as follows: ; In the formula, The controller outputs the adjustment amount; To control the real-time deviation of parameters.

[0011] Preferably, S6 specifically includes: S61. Receive the optimal control command and send it to the field terminal of the corresponding container module; S62. The entire control process synchronously collects real-time operating parameters, actual equipment operating status, and real fault data on-site, and transmits them back to the cloud. S63: The cloud-based system updates the internal parameters of the model on a monthly cycle based on the returned data, optimizing only the weights, judgment thresholds, and control parameters, while retaining the original main structure of the model. At the same time, the edge device has offline emergency control capabilities, and can autonomously maintain the basic environmental safety within the container module when the cloud network is interrupted.

[0012] A multimodal integrated intelligent operation and maintenance system for container data centers includes a perception layer, an edge computing layer, a cloud-based intelligent analysis layer, and an operation and maintenance execution layer that are connected in a bidirectional communication manner. The perception layer is used for multi-source operational data acquisition and adaptive sampling frequency adjustment; The edge computing layer is used for preliminary data preprocessing, command forwarding, and local offline emergency control; The cloud-based intelligent analysis layer integrates a data fusion and cleaning module, a status assessment module, a dual-modal fault diagnosis module, a multi-objective optimization decision-making module, and a model iteration module to complete data processing, fault analysis, optimization instruction generation, and model parameter iteration. The operation and maintenance execution layer is used to receive optimal control commands to complete equipment regulation and fault alarms, and to transmit real-time on-site operating data back to achieve closed-loop intelligent operation and maintenance throughout the entire process.

[0013] Preferably, the sensing layer includes a temperature and humidity sensor, a differential pressure sensor, a power monitoring module, a vibration sensor, a UPS status acquisition unit, a server BMC data acquisition module, and a liquid cooling monitoring unit. Each hardware unit establishes a stable data interaction link with the edge computing layer of the corresponding container module through RS485 and Ethernet communication protocols. The edge computing layer includes an industrial edge gateway, a local embedded computing power module, and a local data storage unit; it can cache raw data and standardized data throughout the entire process, undertake instruction forwarding functions, and realize independent emergency operation and maintenance control of a single container module in the event of a network outage. The edge computing nodes of each module are independent of each other, and the computing power and data are isolated. The cloud-based intelligent analysis layer also integrates a digital twin mapping module to build a three-dimensional model of the container cabin, internal equipment, air ducts and liquid cooling pipelines, enabling a visualized display of the entire process of fault location and operation and maintenance control. The operation and maintenance execution layer includes a precision air conditioning drive unit, a UPS power regulation module, a server load scheduling unit, an audible and visual fault alarm, a liquid cooling drive control unit, and a human-machine interface terminal.

[0014] Therefore, the present invention employs the above-mentioned multimodal fusion intelligent operation and maintenance method and system for container data centers, which has the following beneficial effects: 1. In terms of data processing, this invention uses adaptive sampling combined with a spatiotemporal weighted fusion cleaning algorithm to process data in domains according to independent container modules. This effectively suppresses noise in the original data, fills in missing data, and significantly improves the quality of multi-source operational data, providing a reliable data foundation for subsequent operation and maintenance decisions.

[0015] 2. In terms of fault diagnosis, the method integrates the physical operation mechanism of equipment with multi-model hybrid prediction to effectively reduce the false alarm rate, achieve refined root cause localization, improve fault early warning and analysis capabilities, and reduce the risk of equipment failure and downtime.

[0016] 3. In terms of operation and maintenance control, it achieves multi-objective collaborative optimization of energy consumption, fault risk, environmental status, and liquid cooling operation status under multiple operational constraints. Combined with adaptive closed-loop control, it achieves balanced operation and maintenance adjustment. At the same time, it has the ability of model self-iteration and edge emergency control, forming a closed-loop intelligent operation and maintenance system throughout the entire process.

[0017] 4. The system adopts a layered architecture design and uses all mature industrial hardware, which has good compatibility and low transformation cost. It can be directly adapted to existing single-container and modular assembled container data centers, and is compatible with both air-cooled and liquid-cooled scenarios, making it highly practical for engineering implementation.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart of a multimodal fusion intelligent operation and maintenance method for container data centers according to the present invention; Figure 2 This is an architecture diagram of a multimodal fusion intelligent operation and maintenance system for container data centers according to the present invention; Figure 3 This is a schematic diagram of the modular, assembled container data center hardware deployment of the present invention; Figure 4 This is a graph showing the performance indicators of the intelligent operation and maintenance of the container data center according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0021] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] like Figures 1-4 As shown, a multimodal fusion-based intelligent operation and maintenance method and system for container data centers is presented. This method is suitable for container data centers composed of multiple functionally independent container modules that can be flexibly assembled as needed. Each container module has independent functions and decoupled operation and maintenance management. The method includes the following steps: S1, collecting multi-source operating data from each container module, dynamically adjusting the sampling frequency according to equipment load conditions to form an original multi-source time-series dataset; S2, cleaning the original data based on an adaptive downsampling and spatiotemporal weight fusion algorithm to output a standardized dataset. Adaptive downsampling refers to performing equal-interval sampling on low-rate-of-change data during stable operating periods based on the data's temporal fluctuation amplitude. S3. Downsampling: A hierarchical screening method that retains all sampled points for high-rate-of-change data during periods of abnormal fluctuations; S4. Extracting operational characteristics from the cleaned data, quantifying the overall health of each device, and generating a device health matrix and feature vector; S5. Combining the physical operating mechanism of the equipment with the LSTM-RF-SVR hybrid model to perform fault probability prediction and micro-module-level root cause localization; S6. Constructing a multi-objective optimization function under multiple constraints of environment, energy efficiency, and equipment safety, combining optimization algorithms and adaptive PID control to solve and output the optimal control command; S7. Issuing control commands to the terminal for execution and control, while simultaneously transmitting real-time operational data back to achieve a closed-loop operation and maintenance data throughout the entire process.

[0024] S1 specifically includes: S11, the sensing layer collects parameters covering four categories of operational data within the container: environment, electrical, equipment, and liquid cooling. Environmental parameters include cabinet inlet and outlet air temperatures, relative humidity, and internal pressure differential. Electrical parameters include input voltage, operating current, active power, and UPS load rate. Equipment parameters include server CPU load, hard drive temperature, air conditioning cooling power, and equipment vibration amplitude. Liquid cooling parameters include coolant supply and return temperatures, supply and return pressure differential, flow rate, heat exchange efficiency, pump power, pipeline leak monitoring values, and cold plate temperature. S12, based on the real-time load and operating status of equipment within a single container module, the sampling frequency is dynamically switched independently for each module. High-frequency sampling is used during high load, high temperature anomalies, and liquid cooling anomalies. The sampling frequency is... The value is 10Hz. Low-frequency sampling is performed during stable equipment operation, and its sampling frequency is... The value is set to 1Hz to reduce the amount of invalid data. The sampling frequency is dynamically switched to avoid redundancy of steady-state data and omission of key data under high load, ensuring that the original data collection is comprehensive and efficient. Finally, the original multi-source time series dataset is formed.

[0025] S2 specifically includes: S21, using the original multi-source time-series dataset as the processing object, relying on adaptive downsampling to complete data stratification and filtering, removing redundant and messy data, and simultaneously employing a spatiotemporal weighted fusion algorithm to perform weighted fusion operations on multi-sensor data of the same dimension. The spatiotemporal weighted fusion algorithm refers to a method of weighted fusion of multi-sensor data of the same monitoring dimension based on the historical acquisition accuracy of the sensors, the spatial deployment location and the distance to the measured object, and dynamically allocating weight coefficients. The fusion formula is: ; In the formula, for Real-time fusion and processing of data; The number of sensors under the same monitoring dimension; To correspond to the dynamic spatiotemporal weighting coefficients, and to satisfy the constraints. The weighting coefficients are adaptively updated based on the historical acquisition accuracy and spatial deployment distance of the sensors; For the first The system collects raw data from several sensors; S22, it uses an improved adaptive 3σ criterion to identify and remove outliers, and uses fused data to complete and repair missing data, ultimately generating a standardized dataset directly for subsequent feature extraction processing. The formula for determining outliers is: ; In the formula, This is the deviation amount for identifying data anomalies. for Single-dimensional raw data that is constantly awaiting verification; for The time-series moving average of the data sequence at each given time point; The adaptive decision coefficient has a value range of 1.5 to 3. for The moving standard deviation of the data sequence at any given time is used to determine outlier data, based on the inequality. This processing method effectively eliminates sensor noise interference from the sealed chamber, corrects missing data, unifies heterogeneous data formats, significantly improves the overall quality of the data source, and eliminates the root cause of data deviations in subsequent algorithm calculations.

[0026] S3 specifically includes: S31, conducting data mining based on standardized datasets to extract multi-dimensional time-domain operational characteristic data, including mean, variance, peak value, parameter change rate, and load volatility; S32, combining the operational characteristic data, calculating the operational health status of each device through a health measurement model, the calculation formula being: ; In the formula, The overall health score of the equipment is measured in the range [0,1]. The closer the value is to 1, the better the equipment's operating status. This represents the total dimension of the extracted operational features; The actual monitored values ​​are characteristic. The corresponding baseline values ​​are used for the features; S33, the final integrated equipment health matrix and feature vector are used as the basic input data for subsequent fault diagnosis and analysis. They can quantitatively characterize the comprehensive operating status of the equipment, transform the original multidimensional data into effective feature data that can be used for model calculation, and realize the graded evaluation of equipment operating status.

[0027] S4 specifically includes: S41. Based on the basic input data composed of the equipment health matrix and feature vectors, four types of physical operation mechanisms of the equipment are introduced: heat conduction, electrical loss, airflow coupling in the duct, and heat exchange-flow resistance-pipeline leakage coupling in the liquid cooling system. The constraint boundaries formed by these four types of physical operation mechanisms are the inherent parameter variation limits and internal energy and airflow coupling law limits in the actual operation of the equipment. These are used to limit the reasonable range of the model prediction results and eliminate abnormal prediction values ​​that do not conform to the actual physical operation law of the equipment, thereby correcting the prediction bias that is prone to occur in pure data-driven models; S42. The comprehensive failure probability of the equipment is obtained by using an LSTM-RF-SVR hybrid fusion model. The model weight fusion formula is: ; In the formula, This represents the overall probability of equipment failure. , , The fusion weight proportions for the three models LSTM, RF, and SVR are respectively, and the constraints are satisfied. The weight parameters are iteratively optimized and adapted using a grid search algorithm combined with a historical container failure dataset; S43, Set the failure trigger threshold. The value is set to 0.85. When the overall failure probability exceeds 0.85, the root cause localization process is initiated, achieving four levels of localization: container module level, cabinet module level, sensor location, and actuator level. It outputs fault risk information including fault type, level, and location, providing constraints for subsequent optimization decisions. This diagnostic method overcomes the shortcomings of single models, such as poor generalization ability and large prediction bias, effectively reducing the probability of false alarms and missed alarms. Simultaneously, it achieves refined multi-level root cause localization, narrowing the scope of fault investigation and shortening maintenance and handling time.

[0028] S5 specifically includes: S51, combining container cabin environmental data and fault risk information to define hard operational constraints for the entire process. The minimum permissible temperature inside the cabin is set at 18°C. The maximum permissible temperature inside the cabin is set at 27°C. The minimum permissible relative humidity is set at 40%. The maximum permissible relative humidity is set at 60%. This represents the maximum permissible energy efficiency value for the data center, with a value of 1.3. The maximum permissible failure rate for the equipment is 0.1%. The lower limit of the pump set power is set at 0.5kW. The upper limit of the pump set power is set at 5kW. The lower limit of coolant flow rate is set at 2m. 3 / h, The upper limit for coolant flow rate is set at 20m. 3 / h, The lower limit of the differential pressure between the liquid cooling supply and return pressures is set at 50 kPa. The upper limit of the liquid cooling supply and return pressure differential is set at 200 kPa. Therefore, the ambient temperature... Ambient humidity Energy efficiency indicators Equipment failure rate Liquid cooling pump unit operating power Coolant flow rate Liquid cooling supply and return hydraulic differential The constraint ranges are as follows: ; ; ; ; ; ; ; S52. Within the constraints, construct a multi-objective optimization function that considers energy consumption, failure risk, and environmental stability. The function expression is: ; In the formula, To comprehensively optimize the target value, , , , The optimized weights are for four indicators: energy efficiency, failure probability, temperature deviation, and liquid cooling supply and return liquid temperature difference deviation. Real-time temperature inside the cabin. This is the equipment's reference temperature, set at 22℃, and is also the optimal reference temperature for the equipment. To provide real-time supply and return liquid temperature difference for liquid cooling, The rated heat exchange temperature difference for liquid cooling is set to 5℃. The optimal parameter solution for this function can be obtained using the particle swarm optimization algorithm. S53: The adaptive PID control algorithm is called to complete the tuning of the equipment's closed-loop control parameters, including the proportional coefficient. Integral coefficient Differential coefficients The PID control formula is as follows: ; In the formula, The controller outputs the adjustment amount; To control real-time deviations in control parameters, the control parameters are adaptively tuned according to the amplitude of thermal fluctuations within the cabin, ultimately outputting the optimal control command. This optimized control method can balance energy consumption, failure risk, and environmental conditions under strict equipment safety boundary constraints, avoiding operational imbalances caused by single optimization. Combined with adaptive closed-loop control, it achieves dynamic adaptation of control parameters, balancing data center energy efficiency reduction with stable equipment operation.

[0029] S6 specifically includes: S61, receiving optimal control commands and issuing them to the field terminals of the corresponding container modules to complete the coordinated control of the refrigeration unit, power supply module, and server computing load; S62, synchronously collecting real-time operating parameters, actual equipment operating status, and real fault data throughout the entire control process, and transmitting them back to the cloud layer by layer; S63, the cloud uses the transmitted data to iterate and update the internal parameters of the model on a monthly cycle, optimizing only the weights, judgment thresholds, and control parameters, while retaining the original main structure of the model to ensure stable operation; at the same time, the edge terminal has offline emergency control capabilities, and can autonomously maintain the basic environmental safety within the container module when the cloud network is interrupted. The entire process realizes the implementation of operation and maintenance commands and the iterative feedback of data, the model accuracy continuously improves with operation, and has the capability of offline emergency support at the edge, solving the problem of operation and maintenance interruption due to cloud network outages, forming a long-term stable closed-loop operation and maintenance system.

[0030] A multimodal fusion intelligent operation and maintenance system for container data centers is used to realize intelligent operation and maintenance methods for container data centers, including a perception layer, an edge computing layer, a cloud intelligent analysis layer, and an operation and maintenance execution layer that are connected in a bidirectional communication manner. The sensing layer is used to complete the acquisition of multi-source operating data and adaptive sampling frequency adjustment. It integrates temperature and humidity sensors, differential pressure sensors, power monitoring modules, vibration sensors, UPS status acquisition units, server BMC data acquisition modules, and liquid cooling monitoring units. Each hardware unit establishes a stable data interaction link with the edge computing layer of the corresponding container module through RS485 and Ethernet communication protocols. It can realize the full coverage acquisition of all-dimensional operating parameters inside the container, and is compatible with various sensing devices through multiple protocols, ensuring the integrity of the original data acquisition dimensions and stable transmission.

[0031] The edge computing layer is used for preliminary data preprocessing, command relay and distribution, and offline emergency control in the event of cloud network outages. Internally, it includes an industrial edge gateway, a local embedded computing power module, and a local data storage unit. It can cache raw and standardized data throughout the process, undertake command forwarding functions, and realize independent emergency operation and maintenance control of a single container module in the event of a network outage. The edge computing nodes of each module are independent of each other, and the computing power and data are isolated, effectively sharing the computing pressure of the cloud, reducing data transmission latency, and improving the overall fault tolerance and emergency security of the system.

[0032] The cloud-based intelligent analysis layer can perform data cleaning and fusion, operational status assessment, dual-modal fault diagnosis, multi-objective optimization decision-making, model iteration and updates, and complete all algorithm calculations and operation and maintenance command generation. It is also equipped with a digital twin mapping module to build three-dimensional models of container cabins, internal equipment, air ducts and liquid cooling pipelines, realize the visualization of the entire process of fault location and operation and maintenance control, and intuitively present abstract operation and maintenance data and fault information, so as to facilitate operation and maintenance personnel to quickly verify the control effect and assist on-site manual review.

[0033] The operation and maintenance execution layer receives control commands to complete equipment regulation and fault alarms, and transmits on-site operating data back to the upper layer to form a complete system data closed loop. Internally, it includes a precision air conditioning drive unit, a UPS power regulation module, a server load scheduling unit, an audible and visual fault alarm, a liquid-cooled drive control unit, and a human-machine interface terminal. All system hardware uses mature industrial-grade components, which can be directly connected to existing deployed container data centers to complete access adaptation, realize integrated linkage regulation of multiple devices and multi-level fault alarms, and have low on-site deployment costs and good engineering practicality.

[0034] In this embodiment, the system hardware utilizes mature industrial-grade components and is adapted for integration with existing modular containerized data centers. The data center is a cohesive system comprised of multiple functionally independent container modules that can be flexibly assembled as needed. Specifically, it includes independent IT cabinet container modules, liquid cooling system container modules, power supply and distribution container modules, and refrigeration and air conditioning container modules. These modules are connected via standardized interfaces, allowing for flexible increases or decreases in module quantity based on computing power requirements. The modules are functionally independent and their operation and maintenance are decoupled. Intelligent operation and maintenance are implemented for the aforementioned modular containerized data center deployed outdoors. The IT cabinet container modules use standard 20-foot specifications and house multiple cabinet servers. The liquid cooling system container modules incorporate variable frequency pumps, heat exchange units, liquid replenishment devices, and leakage monitoring components. The power supply and distribution container modules incorporate a UPS power supply system, and the refrigeration and air conditioning container modules incorporate precision air conditioning. The entire operation and maintenance process is performed according to the aforementioned methods and system procedures.

[0035] First, the sensing devices within each independent container module collect real-time operating data. The IT cabinet container module collects data on cabinet inlet and outlet air temperature, humidity, pressure difference, server CPU load, and hard drive temperature. The power supply and distribution container module collects input voltage, operating current, active power, and UPS load rate. The refrigeration and air conditioning container module collects air conditioning cooling power, equipment vibration amplitude, and cabin environmental parameters. The liquid cooling system container module collects coolant supply and return temperatures, supply and return pressure difference, flow rate, heat exchange efficiency, pump power, pipeline leakage monitoring values, and cold plate temperature. The system automatically switches the sampling frequency independently for each container module based on equipment load conditions. When the server is under high load and high temperature or the liquid cooling system is in abnormal condition, it uses 10Hz high-frequency sampling. When the system is running smoothly under low load, it switches to 1Hz low-frequency sampling, forming a raw multi-source time-series dataset archived by module. After receiving the raw data, the edge gateway performs layered filtering of the data by independent container modules using adaptive downsampling. It then combines a spatiotemporal weighted fusion algorithm to fuse multi-sensor data of the same dimension within the same module, adaptively allocating weights based on sensor acquisition accuracy and spatial distance. Subsequently, an improved adaptive 3σ criterion is used to identify and remove abnormal noise data, verify the liquid cooling operation data, and use the fused data to fill in missing data, outputting a standardized, high-quality, modular dataset. Based on the standardized dataset, the mean, variance, peak value, parameter change rate, and load fluctuation rate features are extracted. Simultaneously, the flow fluctuation rate, differential pressure change rate, and heat exchange efficiency decay rate features of the liquid cooling system are extracted. The comprehensive health of each device within a single module is calculated using a health metric formula, and the overall health of the liquid cooling system and its core components is calculated simultaneously, generating a health matrix and feature vector for each module.

[0036] Combining the constraints of heat conduction, electrical losses, airflow physics, and the coupling mechanism of heat exchange, flow resistance, and pipe leakage in the liquid cooling system, a hybrid LSTM-RF-SVR model is used to calculate the failure probability of corresponding equipment. The model fusion weights are optimized through grid search combined with historical failure data. When the failure probability exceeds a threshold of 0.85, the system automatically performs root cause localization at four levels: container module, cabinet, sensor location, and actuator. Under the hard constraints of temperature 18~27℃, humidity 40%~60%, PUE≤1.3, and liquid cooling system operating parameters, the optimal solution of the multi-objective optimization function is solved using the particle swarm optimization algorithm. Combined with adaptive PID control to tune the adjustment parameters of air conditioning and power supply equipment, the optimal operation and maintenance control commands are generated.

[0037] After receiving instructions from the cloud, the edge layer distributes them to the execution terminal of the corresponding container module to complete the adjustment of air conditioning cooling capacity, UPS power control, server load balancing migration, and liquid cooling system flow and pressure regulation. During operation, all real-time field data is transmitted back to the cloud, and the system automatically updates model weights, judgment thresholds, and PID control parameters on a monthly basis. When the cloud network is interrupted, the edge computing layer of the corresponding module autonomously completes the basic environmental control within the cabin based on locally cached data. Emergency operation and maintenance of each module are independent and do not interfere with each other, ensuring uninterrupted and safe operation of the equipment.

[0038] Verified through actual field operation, this solution significantly improves data efficiency and greatly reduces false alarm rate. It optimizes operation and maintenance energy consumption while ensuring safe equipment operation. The entire system can be directly adapted to the transformation and deployment of single-container and modular assembled container data centers, is compatible with both air-cooled and liquid-cooled scenarios, and features fully automated operation and maintenance, meeting the long-term stable operation requirements of unattended outdoor sites.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multimodal fusion-based intelligent operation and maintenance method for containerized data centers, characterized in that, Includes the following steps: S1. Collect multi-source operating data of each container module, dynamically adjust the sampling frequency according to the equipment load conditions, and form the original multi-source time series dataset. S2. Clean the original data based on the adaptive downsampling and spatiotemporal weight fusion algorithm, and output a standardized dataset; S3. Extract the operational characteristics of the cleaned data, quantify the overall health of each device, and generate a device health matrix and feature vector. S4. Combine the physical operation mechanism of the equipment with the LSTM-RF-SVR hybrid model to perform fault probability prediction and micro-module-level root cause localization; S5. Construct a multi-objective optimization function, combine optimization algorithm with adaptive PID control, solve for and output the optimal control command; S6 sends control commands to the terminal to execute the control, and at the same time transmits back the real operation data on site to realize the closed loop of operation and maintenance data throughout the entire process.

2. The intelligent operation and maintenance method for a multimodal fusion container data center according to claim 1, characterized in that, S1 specifically includes: S11. The sensing layer collects parameters covering four categories of operational data within the container: environment, electrical, equipment, and liquid cooling. Environmental parameters include cabinet inlet and outlet air temperature, relative humidity, and internal pressure difference. Electrical parameters include input voltage, operating current, active power, and UPS load rate. Equipment parameters include server CPU load, hard drive temperature, air conditioning cooling power, and equipment vibration amplitude. Liquid cooling parameters include coolant supply and return temperatures, supply and return pressure difference, flow rate, heat exchange efficiency, pump power, pipeline leakage monitoring values, and cold plate temperature. S12. The sampling frequency is dynamically switched according to the real-time load and operating status of the equipment. High-frequency sampling is used during high load and high-temperature abnormal operating conditions. Low-frequency sampling is performed when the equipment is operating stably, and the sampling frequency is: The data are then aggregated to form the original multi-source time-series dataset.

3. The intelligent operation and maintenance method for a multimodal fusion container data center according to claim 2, characterized in that, S2 specifically includes: S21. Using the original multi-source time-series dataset as the processing object, adaptive downsampling is used to perform data stratification and filtering, eliminating redundant and messy data. Simultaneously, a spatiotemporal weighted fusion algorithm is employed to perform weighted fusion operations on multi-sensor data of the same dimension. The fusion formula is as follows: ； In the formula, for Real-time fusion and processing of data; The number of sensors under the same monitoring dimension; To correspond to the dynamic spatiotemporal weighting coefficients, and to satisfy the constraints. ; For the first Raw data collected by each sensor; S22. An improved adaptive 3σ criterion is used to identify and remove outliers, and the fused data is used to complete and repair missing data, finally generating a standardized dataset. The formula for determining outliers is: ； In the formula, This is the deviation amount for identifying data anomalies. for Single-dimensional raw data that is constantly awaiting verification; This represents the time-series moving average of the data. The adaptive decision coefficient has a value range of 1.5 to 3. The moving standard deviation of the data time series.

4. The intelligent operation and maintenance method for a multimodal fusion container data center according to claim 3, characterized in that, S3 specifically includes: S31. Conduct data mining based on standardized datasets to extract multi-dimensional time-domain operational feature data, including mean, variance, peak value, parameter change rate, and load volatility. S32. Combining operational characteristic data, calculate the operational health status of each device using a health quantification model. The calculation formula is as follows: ； In the formula, The overall health status of the equipment is measured in the range [0,1]. This represents the total dimension of the extracted operational features; The actual monitored values ​​are characteristic. The reference value for the corresponding feature; S33. Integrate the generated equipment health matrix and feature vector as the basic input data for fault diagnosis analysis.

5. The intelligent operation and maintenance method for a multimodal fusion container data center according to claim 4, characterized in that, S4 specifically includes: S41. Based on the basic input data, four types of equipment physical operation mechanisms are introduced: heat flow conduction, electrical loss, airflow coupling in the duct, and heat exchange-flow resistance-pipeline leakage coupling in the liquid cooling system, to correct the prediction bias that is prone to occur in pure data models. S42. The overall failure probability of the equipment is obtained by using a hybrid LSTM-RF-SVR model. The model weight fusion formula is as follows: ； In the formula, This represents the overall probability of equipment failure. , , The fusion weight proportions for the three models LSTM, RF, and SVR are respectively, and the constraints are satisfied. ; S43. Set fault trigger threshold When the overall failure probability exceeds the threshold At that time, the root cause localization process is initiated to achieve four levels of localization: container module level, cabinet module level, sensor location, and actuator level, and output fault risk information including fault type, level, and location information.

6. The intelligent operation and maintenance method for a multimodal fusion container data center according to claim 5, characterized in that, S5 specifically includes: S51. Combining container cabin environmental data and fault risk information, define the hard operational constraints for the entire process, including ambient temperature. Ambient humidity Energy efficiency indicators Equipment failure rate Liquid cooling pump unit operating power Coolant flow rate Liquid cooling supply and return hydraulic differential The constraint ranges are as follows: ； ； ； ； ； ； ； In the formula, The lower limit of the cabin temperature, This is the upper limit of the cabin temperature. This is the lower limit of relative humidity. This is the maximum permissible relative humidity limit. This represents the upper limit of the allowable energy efficiency value for data centers. The upper limit of the allowable failure rate of the equipment. This is the lower limit of the pump set power. This is the upper limit of the pump set power. This is the lower limit of coolant flow rate. This is the upper limit of coolant flow rate. This is the lower limit of the differential pressure between the liquid cooling supply and return pressures. This is the upper limit of the differential pressure between the liquid cooling supply and return pressures. S52. Construct a multi-objective optimization function within the constraints, with the function expression as follows: ； In the formula, To comprehensively optimize the target value, , , , The optimized weights are for four indicators: energy efficiency, failure probability, temperature deviation, and liquid cooling supply and return liquid temperature difference deviation. Real-time temperature inside the cabin. The reference temperature for the equipment. To provide real-time supply and return liquid temperature difference for liquid cooling, This is the rated heat exchange temperature difference for liquid cooling; S53. Call the adaptive PID control algorithm to complete the tuning of the closed-loop control parameters of the equipment. The control parameters include the proportional coefficient. Integral coefficient Differential coefficients The control parameters are adaptively tuned according to the amplitude of thermal fluctuations within the cabin, ultimately outputting the optimal control command. The PID adjustment formula is as follows: ； In the formula, The controller outputs the adjustment amount; To control the real-time deviation of parameters.

7. The intelligent operation and maintenance method for a multimodal fusion container data center according to claim 6, characterized in that, S6 specifically includes: S61. Receive the optimal control command and send it to the field terminal of the corresponding container module; S62. The entire control process synchronously collects real-time operating parameters, actual equipment operating status, and real fault data on-site, and transmits them back to the cloud. S63: The cloud-based system updates the internal parameters of the model on a monthly cycle based on the returned data, optimizing only the weights, judgment thresholds, and control parameters, while retaining the original main structure of the model. At the same time, the edge device has offline emergency control capabilities, and can autonomously maintain the basic environmental safety within the container module when the cloud network is interrupted.

8. A multimodal fusion intelligent operation and maintenance system for container data centers, used to implement the intelligent operation and maintenance method for container data centers according to any one of claims 1-7, characterized in that: It includes a perception layer, an edge computing layer, a cloud-based intelligent analysis layer, and an operation and maintenance execution layer, which are connected in a bidirectional manner. The perception layer is used for multi-source operational data acquisition and adaptive sampling frequency adjustment; The edge computing layer is used for preliminary data preprocessing, command forwarding, and local offline emergency control; The cloud-based intelligent analysis layer integrates a data fusion and cleaning module, a status assessment module, a dual-modal fault diagnosis module, a multi-objective optimization decision-making module, and a model iteration module to complete data processing, fault analysis, optimization instruction generation, and model parameter iteration. The operation and maintenance execution layer is used to receive optimal control commands to complete equipment regulation and fault alarms, and to transmit real-time on-site operating data back to achieve closed-loop intelligent operation and maintenance throughout the entire process.

9. The intelligent operation and maintenance system for a multimodal fusion container data center according to claim 8, characterized in that: The sensing layer includes temperature and humidity sensors, differential pressure sensors, power monitoring modules, vibration sensors, UPS status acquisition units, server BMC data acquisition modules, and liquid cooling monitoring units. Each hardware unit establishes a stable data interaction link with the edge computing layer of the corresponding container module through RS485 and Ethernet communication protocols. The edge computing layer includes an industrial edge gateway, a local embedded computing power module, and a local data storage unit; it can cache raw data and standardized data throughout the entire process, undertake instruction forwarding functions, and realize independent emergency operation and maintenance control of a single container module in the event of a network outage. The edge computing nodes of each module are independent of each other, and the computing power and data are isolated. The cloud-based intelligent analysis layer also integrates a digital twin mapping module to build a three-dimensional model of the container cabin, internal equipment, air ducts and liquid cooling pipelines, enabling a visualized display of the entire process of fault location and operation and maintenance control. The operation and maintenance execution layer includes a precision air conditioning drive unit, a UPS power regulation module, a server load scheduling unit, an audible and visual fault alarm, a liquid cooling drive control unit, and a human-machine interface terminal.

Images