Sensor network-based data center air quality intelligent early warning method and system

Abstract

The present application relates to the technical field of sensor network and internet of things, in particular to a data center air quality intelligent early warning method and system based on sensor network; the method comprises the following steps: constructing a self-organizing cooperative network by deploying multiple types of sensor nodes in key areas of the data center, collecting and correcting multi-dimensional air data in real time; improving data reliability through dynamic weight fusion and local anomaly recognition between nodes; establishing an air parameter and machine room structure correlation model by using spatial mapping and time series analysis to identify micro-scale diffusion trends; constructing a self-adaptive dynamic threshold mechanism to realize rolling optimization of the threshold combined with historical statistics and real-time feedback; generating a hierarchical alarm strategy based on multi-level early warning judgment and trigger conditions, and continuously optimizing the early warning accuracy and response efficiency through closed-loop feedback. The present application realizes all-around, intelligent and highly reliable early warning and regulation of the air quality in the data center.

Description

Technical Field

[0001] This invention relates to the fields of sensor networks and the Internet of Things (IoT), specifically to a method and system for intelligent early warning of air quality in data centers based on sensor networks. Background Technology

[0002] As a key physical facility supporting digital operations, the stability of the internal environment of data centers directly affects the lifespan of computing equipment and business continuity. In recent years, with the continuous increase in server power density, the unevenness of airflow and pollutant distribution in the rack-level microenvironment has become increasingly significant, placing higher demands on the precision and real-time nature of air quality monitoring.

[0003] Chinese invention patent application CN109596492A discloses a real-time intelligent indoor air quality monitoring system. This system includes a sensor sensing subsystem, a computer analysis terminal, and intelligent terminals. The sensor sensing subsystem collects indoor air quality information and comprises a wireless sensor network consisting of a convergence node and multiple sensor nodes deployed indoors. The sensor nodes collect air quality information at their respective monitoring locations. The convergence node aggregates the air quality information collected by each sensor node and sends it to the computer analysis terminal. The computer analysis terminal is connected to multiple intelligent terminals to send the received air quality information to these terminals.

[0004] Meanwhile, with the advancement of sensor technology, edge computing, and intelligent analysis methods, building an air quality early warning system capable of real-time perception, intelligent analysis, and autonomous decision-making has become an important research direction for improving the intelligence level of data center operation and maintenance. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a smart early warning method and system for air quality in data centers based on sensor networks.

[0006] The technical solution of this invention: A smart early warning method for air quality in data centers based on sensor networks, comprising the following specific implementation steps:

[0007] S1. Deploy multiple types of sensor nodes in key areas of the data center to build a self-organizing collaborative network, collect multi-dimensional air data in real time, and perform node-end correction and local anomaly identification on the collected data.

[0008] S2. Improve data reliability by dynamic weight fusion between nodes and local anomaly identification, and establish a correlation model between air parameters and computer room structure by using spatial mapping and time series analysis to identify microscale diffusion trends.

[0009] S3. Construct an adaptive dynamic threshold mechanism, combining historical statistics and real-time feedback to achieve threshold rolling optimization;

[0010] S4. Generates a graded alarm strategy based on multi-level early warning judgment and triggering conditions, and continuously optimizes the early warning accuracy and response efficiency through closed-loop feedback.

[0011] Preferably, step S1 specifically includes:

[0012] Temperature and humidity sensors, particulate matter sensors, gas concentration sensors, and wind speed and direction sensors are installed in the cold aisle, hot aisle, server rack aisle, air conditioning outlet, and underfloor return air area of ​​the computer room.

[0013] Construct an air state vector for a single node, which includes ambient temperature, relative humidity, particulate matter concentration, gas concentration, air velocity, and wind direction;

[0014] Construct a node self-organizing and cooperative network. Each node forms a self-organizing network through low-power wireless communication. The cooperative weight is dynamically calculated based on the distance between nodes and their health status. Local data fusion and preliminary anomaly identification are then performed.

[0015] The nodes collect air state vectors at an adaptive sampling frequency and attach a unique identifier and timestamp. Key areas are collected at high frequency and redundantly supplemented by neighboring nodes.

[0016] The collected data undergoes drift correction, filtering, delay compensation, and airflow smoothing at the node end, and also supports preliminary anomaly warning at the node end.

[0017] Preferably, a self-organizing and collaborative network of nodes is constructed, specifically including:

[0018] Before each sensor node collaborates with its neighboring nodes in real time, a collaboration weight is calculated. This weight takes into account the distance between nodes, the health status of nodes, and the reliability of data.

[0019] Each node selects the optimal set of neighbors based on the cooperation weight, and performs weighted fusion of the data of the neighboring nodes to generate a local cooperative air state vector;

[0020] The network performs dynamic topology maintenance based on node health status and environmental changes, including removing failed nodes, adding new nodes, and adaptively reconstructing the neighbor set and weights.

[0021] Preferably, step S2 specifically includes:

[0022] The calibration parameters of each node are mapped to the rack coordinates, air supply and exhaust paths and cold aisle structure of the data center, thus constructing a dynamic correspondence framework between multi-dimensional air conditions and spatial layout.

[0023] By using a continuous time window sliding method to analyze the rate of change, direction of change, and persistence of local disturbances of air parameters, microscale diffusion or aggregation trends can be identified within a spatial distribution framework.

[0024] By constructing a cross-regional factor chain, external factors such as cabinet heat load changes, air duct resistance, and fluctuations in the operation of neighboring equipment are coupled with air quality trends for evaluation, thereby quantifying the potential driving forces for the formation and spread of anomalies.

[0025] Based on the perturbation trend and factor risk quantification results, a fusion probability model is used to output the comprehensive credibility and possible causes of air quality anomalies.

[0026] Preferably, the calibration parameters of each node are mapped to the rack coordinates, air supply and exhaust paths, and cold aisle structure of the data center, constructing a dynamic correspondence framework between multi-dimensional air conditions and spatial layout, specifically including:

[0027] Extract the corrected multidimensional vector and local anomaly marker from the node data packet, and calculate the node-neighborhood consistency residual;

[0028] Calculate the initial confidence score and perform local smoothing using the adjacency weights between nodes to obtain the final confidence score;

[0029] An observation weight matrix is ​​formed based on the adjacency weights and credibility between nodes. Spatial and temporal analysis is then performed by combining graph Laplace smoothing and temporal smoothing.

[0030] Preferably, step S3 specifically includes:

[0031] Based on historical fusion index sequences and outliers, calculate the regional mean and standard deviation, and generate the initial threshold for each region through weighted combination.

[0032] The contribution weights of each component to the threshold are adjusted in real time using regional uncertainty, and stability is ensured by combining historical weights.

[0033] The thresholds are updated on a rolling basis based on the adjusted weights and fusion index, while amplification corrections are made for sudden abnormal events.

[0034] By statistically analyzing historical false alarms and missed alarms, and adjusting the rolling update rate and safety factor based on the false alarm and missed alarm indicators, closed-loop optimization can be achieved.

[0035] Preferably, based on historical fusion index sequences and outliers, the regional mean and standard deviation are calculated, and an initial threshold for each region is generated through weighted combination, specifically including:

[0036] Select the fusion index sequence and corresponding abnormal event labels within the historical time period, and calculate the statistical distribution for each region, including the mean and standard deviation;

[0037] The initial threshold is defined based on a weighted combination of the mean and standard deviation, with a safety factor taken into account.

[0038] An initial threshold is generated by weighting the multidimensional components, with the component weights based on historical statistics and risk correlation.

[0039] Preferably, step S4 specifically includes:

[0040] By comparing regional indices with dynamic thresholds and combining uncertainty correction rules, environmental conditions are classified into normal, mild, moderate, and severe warning levels.

[0041] By setting a time accumulation coefficient and a neighborhood coordination factor, the continuity and spatial consistency of the early warning trigger are judged to avoid false alarms caused by short-term fluctuations or local sporadic events.

[0042] Different levels of execution actions are matched according to the warning level and the triggering result, and the strategy intensity is dynamically adjusted according to the importance of the area and the real-time status.

[0043] Based on the actual alarm effect, false alarms and missed alarms, and air quality recovery, the threshold range, trigger parameters, and alarm priority are automatically optimized.

[0044] Preferably, by setting a time accumulation coefficient and a neighborhood coordination factor, the persistence and spatial consistency of the early warning trigger are judged, specifically including:

[0045] Introduce a time accumulation factor to calculate the proportion of times the threshold condition is met within the past sliding window;

[0046] Introduce a neighborhood collaboration factor to calculate the proportion of regions exceeding the threshold within the neighborhood at the current time;

[0047] By combining time accumulation and neighborhood collaboration, trigger conditions are set, and an early warning is triggered when either continuous judgment or neighborhood resonance is satisfied.

[0048] The technical solution of this invention: A data center air quality intelligent early warning system based on sensor networks, which is used to execute the above-mentioned data center air quality intelligent early warning method based on sensor networks, including:

[0049] The sensor node and self-organizing network module is responsible for actual physical sensing and reliable data acquisition. It includes multiple types of sensor nodes and low-power wireless self-organizing gateways. The nodes have local self-testing, health scoring, positioning identification, adjustable sampling rate and node-end calibration capabilities, and support inter-node collaborative weight calculation and redundant acquisition strategies.

[0050] The edge preprocessing and multidimensional fusion module is deployed on the edge server. It receives the corrected multidimensional state and local anomaly markers, completes credibility assessment, neighborhood credibility propagation, graph time-series weighted interpolation and robust fusion, and outputs a spatially continuous fusion estimate with uncertainty quantification and a region-level index.

[0051] The adaptive threshold and prediction decision module uses regional indicators, uncertainty and historical event knowledge base to build and evolve dynamic thresholds. It includes component weight adaptor, threshold rolling updater, burst amplification and buffering strategy, and short-term trend prediction engine based on time series model to realize threshold verification feedback and online learning mechanism.

[0052] The multi-level early warning and intelligent alarm execution module is responsible for transforming threshold decisions into executable operation and maintenance actions. It includes an early warning level determiner, time accumulation and neighborhood collaboration triggers, alarm policy library, alarm issuance and execution interface, and closed-loop feedback acquisition unit.

[0053] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0054] This invention designs an intelligent early warning method and system for data center air quality based on sensor networks. By constructing a self-organizing collaborative sensor network, it achieves high-precision and high-reliability real-time monitoring of air quality in key areas of the data center, effectively overcoming the limitations of single-point sensing being susceptible to interference and data silos. Through dynamic weight fusion between nodes and a local anomaly identification mechanism, the quality and reliability of the raw data are significantly improved, laying a solid foundation for subsequent analysis. A dynamic correlation model is established between air quality parameters and the spatial structure and airflow organization of the computer room, which can accurately analyze the diffusion path and trend of abnormal pollution, achieving a leap from perception to cognition. Utilizing an adaptive dynamic threshold mechanism, it can continuously self-optimize based on environmental changes and historical feedback, effectively balancing the sensitivity and stability of the early warning and significantly reducing false alarms and missed alarms. Through intelligent hierarchical early warning and closed-loop feedback execution, the system transforms monitoring results into operable operation and maintenance instructions, realizing full automation and intelligence from risk perception to control and disposal, significantly improving the initiative and operation and maintenance efficiency of data center environmental management. Attached Figure Description

[0055] Figure 1 This is a flowchart of a data center air quality intelligent early warning method based on sensor networks proposed in this invention.

[0056] Figure 2 This is a system architecture diagram of a data center air quality intelligent early warning system based on sensor networks proposed in this invention. Detailed Implementation

[0057] Example 1, as Figure 1 As shown, the present invention proposes a data center air quality intelligent early warning method based on sensor networks, which includes the following specific implementation steps:

[0058] S1. By deploying multiple types of sensor nodes in key areas of the data center, multi-dimensional air data such as temperature, humidity, particulate matter, gas concentration, and wind speed and direction are collected. Combined with a self-organizing collaborative network of nodes, real-time acquisition, node-end calibration, and preliminary anomaly identification, high precision, real-time performance, and reliability of air quality monitoring are achieved, providing a foundation for subsequent data fusion and intelligent early warning. The specific implementation process is as follows:

[0059] S11. Deploy temperature, humidity, particulate matter, gas, and airflow sensors in the cold aisle, hot aisle, and rack aisles of the computer room. Integrate multi-dimensional information into node state vectors, and achieve refined spatial monitoring through high-density deployment in key areas to improve the integrity of air quality characterization. Specifically:

[0060] Multiple types of sensor nodes are deployed in key areas of the computer room (cold aisle, hot aisle, server rack aisle, air conditioning vents, and underfloor return air area). Each node collects the following information:

[0061] Ambient temperature relative humidity Concentration of particulate matter (PM2.5, PM10) in the air ,monitor VOC and other gas concentrations Measuring air velocity and wind direction ;

[0062] Construct the air state vector of a single node :

[0063] ;

[0064] Where n represents the number of chemical gases collected; This indicates the plane coordinates of the sensor node in the computer room; t represents the data acquisition timestamp.

[0065] S12. Construct a node self-organizing and cooperative network. Each node forms a self-organizing network through low-power wireless communication. The cooperative weight is dynamically calculated based on the distance between nodes and their health status. Local data fusion and preliminary anomaly identification are performed to improve the reliability and response speed of monitoring in key areas. Specifically:

[0066] Before each sensor node engages in real-time data collaboration with its neighboring nodes, the collaboration weight is calculated. Taking into account factors such as distance between nodes, node health status, and data reliability:

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] in, This represents the status score of sensor j, evaluated based on the normal operation of each sensor at the node and error drift. 1 indicates that all sensors are operating normally. ; This represents the power supply status score, reflecting the node's power supply stability and remaining power. 1 indicates stable power supply. ; This represents a communication status score, reflecting the stability of communication between a node and its central or neighboring nodes. A score of 1 indicates stable communication with no packet loss. ; , and Indicates the weighting coefficient. N represents the total number of historical data collections; Indicates the sensor's history factor; This represents the cooperation weight of node i with respect to its neighbor node j; This represents the Euclidean distance between node i and node j; This represents the health status score of node j. ; This represents the historical data reliability score for node j, calculated based on the stability and anomaly rate of previously collected data. ;

[0072] Each node selects the optimal set of neighbors based on its collaboration weight. : ;

[0073] Node i performs weighted fusion of data from its neighboring nodes to generate a local cooperative air state vector:

[0074] ;

[0075] in, Represents the set of cooperative neighbors of node i; This represents the set weight threshold used to filter highly reliable neighbor nodes; Represents the local cooperative air state vector of node i; Represents the air state vector of neighbor node j;

[0076] It should be noted that after the inter-node collaboration is completed, the network performs dynamic topology maintenance based on the node health status and environmental changes: nodes can automatically adjust their collaboration relationships based on their own health status and the status of their neighbors, and failed nodes are temporarily removed; when a new node joins or a critical node fails, the network adaptively reconstructs the neighbor set and weights to ensure data coverage continuity; and periodically evaluates the contribution of each node, optimizes the communication path between nodes, and reduces network latency and energy consumption.

[0077] S13. Nodes collect air state vectors at an adaptive sampling frequency, attaching a unique ID and timestamp. High-frequency sampling is performed in key areas, with redundant sampling possible from neighboring nodes, ensuring data continuity and spatiotemporal accuracy. This provides high-quality raw data for subsequent analysis. Specifically:

[0078] Each node is based on a local cooperative state vector. Automatically adjust the sampling frequency based on historical air quality fluctuation characteristics : ;

[0079] in, This indicates the basic sampling frequency of the node, ensuring the minimum sampling interval; This represents the adjustment coefficient, used to amplify the impact of air quality fluctuations on the sampling frequency; The standard deviation of the local cooperative state vector is used to measure the magnitude of air quality changes.

[0080] Based on a self-organizing network, when a node experiences S consecutive data collection anomalies or communication failures, a redundant data collection mechanism involving neighboring nodes is automatically activated. ;

[0081] in, This indicates the redundant air status data collected when node i is abnormal;

[0082] The node performs self-monitoring during the data collection process and generates a complete data collection data packet. :

[0083] ;

[0084] in, Indicates node i at time The collected data packets; Indicates a unique identifier for the node;

[0085] S14. The collected data undergoes drift correction, filtering, delay compensation, and airflow smoothing at the node end, and is then uploaded to the central processing unit. Simultaneously, it supports preliminary anomaly warnings at the node end, improving data reliability and response efficiency while reducing the central processing load. Specifically:

[0086] Data collection Preliminary processing and correction are performed at the node: signal filtering (such as temperature and humidity filtering, particulate matter smoothing), data format standardization and timestamp correction, and local anomaly marking (such as local peaks exceeding the threshold).

[0087] Nodes use the local fusion state output by the cooperative network and the corrected data to determine local anomalies: ;

[0088] in, This indicates a local abnormality; 1 indicates abnormality, and 0 indicates normality. This represents the multidimensional air state vector after node correction; and This indicates an adaptive threshold, dynamically determined based on historical statistical characteristics and the risk level of key areas.

[0089] The node will correct the multidimensional air state Local anomaly markers The data packet consists of node status information: ;

[0090] in, This indicates the final data packet uploaded by the node.

[0091] S2. By constructing a hierarchical air condition analysis mechanism oriented towards the operational characteristics of data centers, the system transforms raw local monitoring results into interpretable and inferable indoor aerodynamic structures. A dynamic correlation framework is established between air quality parameters and rack distribution and air duct structure using spatial mapping. Then, the spread trend of local anomalies within a micro-scale space is extracted using a continuous temporal variation pattern. An interference level assessment model is constructed based on cross-regional factor influence chains to structurally quantify the risk of anomaly spread. Finally, probabilistic fusion is used to comprehensively determine air disturbance trends and possible sources. The specific implementation process is as follows:

[0092] S21. By mapping the calibration parameters of each node to the rack coordinates, air supply and exhaust paths, and cold aisle structure of the data center, a dynamic correspondence framework between multi-dimensional air conditions and spatial layout is constructed to achieve structured positioning of air quality parameters, specifically:

[0093] Extract the corrected multidimensional vector from the node packet. and local anomaly markers And calculate the consistency residual between the node and its neighborhood. : ;

[0094] Calculate the initial confidence score :

[0095] ;

[0096] To avoid extreme fluctuations in isolated nodes, adjacency weights between nodes are used. Perform a local smoothing:

[0097] ;

[0098] in, Indicates local consistency residuals; This represents a stabilization constant, used to prevent the denominator from being zero or too small, thus ensuring numerical stability. This represents the typical amplitude of the region, used for scale normalization (in this embodiment, the historical average of the region is used); Represents the initial (unsmoothed) original confidence score of node i; , , and Indicates the weighting coefficient; This represents the health status score of node i; This represents the data reliability score for node i; This represents the confidence coefficient, which controls the mixing ratio of a node's self-score and the average confidence level of its neighborhood. This represents the normalized collaboration weights, i.e., the adjacency weights between nodes. Normalization yields the result;

[0099] S22. Analyze the rate of change, direction of change, and persistence of local disturbances of air parameters using a continuous time window sliding method, and identify whether there are microscale diffusion or aggregation trends within the spatial distribution framework. Specifically:

[0100] Based on adjacency rights between nodes And combine the confidence level to form the observation weight matrix. :

[0101] ; ;

[0102] Define the graph Laplace Solve the spatial smoothness minimization problem (for each component):

[0103] ;

[0104] For each grid point, an exponentially weighted historical term is introduced for time-series smoothing:

[0105] ;

[0106] in, L represents the observation weight matrix, i.e., a diagonal matrix; D represents the graph Laplacian matrix; and D represents the degree matrix, a diagonal matrix whose elements are defined as follows: , which is the sum of the connectivity weights of node i and all its neighbors; W represents the weighted adjacency matrix between nodes, where each element is . ; This represents the spatial smoothing intensity coefficient, which controls the trade-off between the observation fitting term and the graph smoothing term; Represents vectorized observation data (by concatenating the same component from each node into a column vector); This represents the spatial estimation solution vector of the graph Laplace smoothing (corresponding to...) (A smooth version); This represents the time smoothing coefficient (exponential weight), used to estimate the current time step. Compared with historical estimates Combine;

[0107] S23. By constructing a cross-regional factor chain, external factors such as cabinet heat load changes, duct resistance, and fluctuations in the operation of neighboring equipment are coupled with air quality trends for evaluation, quantifying the potential driving forces of anomaly formation and spread, specifically:

[0108] Nodes are categorized into three types based on their credibility and original label:

[0109] Trusted Sets : ;

[0110] Suspicious Collection : ;

[0111] Low Trust Set : ;

[0112] For each sensor Calculate the difference between the estimate and the neighborhood smoothing estimate: ;

[0113] like (Adaptive threshold, depending on the historical fluctuation standard deviation) is then determined to be a real sudden anomaly; otherwise, it is considered a sensor anomaly or noise.

[0114] For each spatial point, the fusion value z is obtained using the weighted Huber loss:

[0115] ;

[0116] ;

[0117] In implementation, iterative reweighted least squares (IRLS) is used for solving the problem, with initial weight values ​​using... ;

[0118] If a certain i is determined to be a true burst, then the fusion result at that point allows for higher sensitivity (local). Decrease, or weight (Temporary boost) to ensure that genuine scene changes are not overly smoothed out;

[0119] in, and This represents the set confidence threshold used for stratification; in this embodiment, it is taken as... , ; This represents a measure of the difference between node observations and fusion estimates. Indicates the adaptive consistency threshold, used for comparison. Is it significant? denoted by Huber loss function, used for robust handling of residuals; z represents fusion estimate, which is the robust fusion value obtained by minimizing Huber loss on observations (weighted) at a certain spatial point or grid point; The switching point (threshold) is when the residual Use quadratic loss; otherwise, it grows linearly.

[0120] S24. Based on the disturbance trend and factor risk quantification results, a fusion probability model is used to output the comprehensive credibility and possible causes of air quality anomalies, providing verifiable input for subsequent early warning strategies. Specifically:

[0121] Define an interpretable indicator that transitions from multidimensional to scalar (regional AQI-like index):

[0122] ;

[0123] First, principal component analysis (PCA) is used to obtain the variance contribution rate of each component, and then its correlation with equipment risk is scored. Fusion, generating weights :

[0124] ;

[0125] Clustering nodes into risk level zones (based on) With uncertainty U), it is used for differentiated threshold management and alarm strategies, and uncertainty is quantified:

[0126] ;

[0127] Output triples for each region / grid point And return the node-level anomaly label (whether it is determined to be a real outbreak);

[0128] in, This represents the normalized value of the m-th component after fusion. The comprehensive air quality index for region p; This represents the dynamic weight of the m-th component in the comprehensive index (summing to 1); M represents the total number of components. This represents the score of the m-th component in terms of its contribution to the principal components or variance (a data-driven importance measure). The equipment risk sensitivity score (engineering prior) represents the m-th component, characterizing the strength of the variable's impact on equipment health or failure risk; and Indicates the adjustment coefficient; This represents the quantified value of the uncertainty in region p; This represents the fusion center value of region p.

[0129] S3. Construct adaptive and dynamically evolving air quality thresholds to achieve intelligent early warning of multi-regional and multi-dimensional air conditions in the data center. Through initial threshold generation, dynamic adjustment of component weights, rolling evolution of thresholds, and feedback closed-loop optimization, the thresholds can take into account real-time response, historical statistics, and abnormal sudden events, ensuring the long-term stability, robustness, and intelligence of the early warning system. The specific implementation process is as follows:

[0130] S31. Based on the historical fusion index sequence and abnormal events, calculate the regional mean and standard deviation, and generate the initial threshold for each region through weighted combination to provide a benchmark for subsequent dynamic adjustments. Simultaneously, consider the influence of multi-dimensional components to achieve targeted initial value setting, specifically:

[0131] Select the fusion index sequence within the historical time period H Corresponding exception event label Calculate the statistical distribution for each region (e.g., the time of alarm occurrence):

[0132] ; ;

[0133] Define initial threshold ;

[0134] For multidimensional components, a weighted combination can be used: ;

[0135] Where H represents the set of historical time periods, used to statistically analyze the historical distribution of the regional integration index; This represents the average historical fusion index of region p; This represents the standard deviation of the historical fusion index of region p; Represents the average value function; Represents the standard deviation function; This represents the safety factor, controlling the degree to which the initial threshold exceeds the historical average. This represents the weight of the multidimensional components, used to fuse the contribution ratio of multidimensional air states to the threshold.

[0136] S32. Utilize regional uncertainty to real-time adjust the contribution weights of each component to the threshold. High uncertainty reduces the current measurement weight, while historical weights ensure stability, achieving adaptive sensitivity of the threshold to different regions and states. Specifically:

[0137] Calculate the uncertainty adjustment factor: ;

[0138] Update component weights: ;

[0139] in, This represents the uncertainty adjustment coefficient; Indicates the uncertainty-sensitive adjustment parameter; Indicates reference uncertainty; This represents the historical average weight, used to compensate for the current weight under high uncertainty.

[0140] S33. Based on the adjusted weights and fusion index, the threshold is updated on a rolling basis, while amplification corrections are made for sudden abnormal events to achieve rapid response of the threshold to real-time changes and long-term stable balance, specifically as follows:

[0141] Rolling update threshold: ;

[0142] If step S23 marks the current region as having a real sudden anomaly If the threshold is temporarily lowered by a certain percentage, it will ensure that high-risk events trigger alarms in a timely manner. The updated strategy is as follows: ;

[0143] in, Indicates that region p at time... Adaptive threshold; This indicates the rolling update rate, controlling the threshold's response speed to new data; This represents the abnormal amplification correction coefficient, which adjusts the threshold ratio when a real abnormality occurs. Indicates that region p at time... The fusion index of the m-th component; This indicates that the m-th component is at time [time]. Adaptive weights;

[0144] S34. Statistical analysis of historical false alarms and missed alarms, adjusting the rolling update rate and safety factor through FP and FN indicators to achieve closed-loop optimization; threshold parameters adaptively evolve over time to improve early warning accuracy and long-term system stability, forming a complete adaptive threshold management system, specifically:

[0145] Select historical window H as the validation interval and count false positives and false negatives:

[0146] ;

[0147] ;

[0148] Feedback adjustment parameters: ;

[0149] Update security level : ;

[0150] in, This indicates the number of false negatives in region p (real anomalies were not triggered by the threshold). This indicates the number of false alarms in region p (threshold triggered but no actual anomaly). and A coefficient representing the feedback adjustment of the rolling update rate; This indicates that the safety factor is adjusted based on feedback. Weight parameters; This indicates an indicator function used to count whether an event has occurred (1 for occurrence, 0 for non-occurrence).

[0151] S4. Through four levels—early warning level determination, trigger condition optimization, alarm strategy generation, and feedback self-learning—multi-level early warning triggering driven by dynamic thresholds is achieved. A stable and sensitive alarm process is constructed by combining uncertainty, time accumulation, and spatial neighborhood information. Simultaneously, thresholds, trigger parameters, and strategy priorities are automatically adjusted based on alarm execution feedback, enabling the system to possess the self-evolutionary capability to continuously adapt to environmental changes. The specific implementation process is as follows:

[0152] S41. By comparing regional indices with dynamic thresholds and combining them with uncertainty correction rules, environmental conditions are categorized into normal, light, moderate, and severe alert levels, enabling accurate classification of air quality deviations at different degrees. This process not only utilizes real-time indices but also incorporates threshold offset ranges and reliability into the judgment, ensuring that the alert levels are both sensitive and stable, providing a clear stratification basis for subsequent triggering and strategy generation. Specifically:

[0153] Define multi-level threshold intervals:

[0154] ;

[0155] An uncertainty correction factor is introduced to correct the rank:

[0156] ; ;

[0157] The corrected continuous scores are obtained and mapped back to discrete levels, rounded up to the nearest integer, and the textual levels are mapped to ordered scores (normal = 0, mild = 1, moderate = 2, severe = 3):

[0158] ;

[0159] in, Indicates that region p at time... The warning level; U max Indicates the uncertainty. The reference upper bound is normalized to [0,1]. In this embodiment, the 95th percentile of historical uncertainty is selected to avoid being increased by extreme anomalies. This represents the uncertainty after normalization; Indicates the corrected strength coefficient; The sign function indicates the direction of deviation of the current indicator relative to the threshold; Indicates the amount of continuous level correction; This means that consecutive scores will be limited to the allowed range of levels (0 to 3); This indicates the corrected continuous grade score; Indicates the lower bound buffer size of the threshold; Indicates the upper bound of the threshold buffer size;

[0160] S42. By setting a time accumulation coefficient and a neighborhood coordination factor, the system assesses the continuity and spatial consistency of early warning triggering, avoiding false alarms caused by short-term fluctuations or localized sporadic events. The final triggering condition is met either by continuous judgment or neighborhood resonance, making early warning triggering more consistent with air diffusion patterns and the characteristics of actual data center scenarios, and providing a reliable triggering signal for strategy execution. Specifically:

[0161] Introducing a time accumulation factor: ;

[0162] And introduce the neighborhood synergy factor: ;

[0163] By combining time accumulation and neighborhood collaboration, set the trigger conditions:

[0164] ;

[0165] in, This represents the time accumulation factor, i.e., the factor accumulated over the past sliding window. Inside, satisfy The proportion; Represents the neighborhood collaboration factor, that is, the neighborhood factors at the current time. The proportion of regions exceeding the threshold; This represents a binary trigger flag; it is 1 if the trigger condition is met, and 0 otherwise. Indicates the cumulative time trigger threshold; Indicates the threshold for triggering neighborhood collaboration;

[0166] S43. Match different levels of execution actions based on the warning level and trigger result, and dynamically adjust the strategy strength according to the importance of the area and the real-time status, so that the alarm is not just a notification, but is transformed into an executable operation and maintenance command, realizing a coherent closed loop from detection to handling, specifically:

[0167] Define alarm policies based on level and trigger identifier:

[0168] ;

[0169] By utilizing historical feedback mechanisms, the strategy response level is fine-tuned, and node-level priorities are introduced. For example, higher weights in critical areas trigger stricter measures, and alarm commands are output accordingly.

[0170] in, This indicates that for region p, at time... Alarm actions or strategies to be executed (e.g., "prompt monitoring", "restrict ventilation / activate air purification", "full emergency response");

[0171] S44. Based on actual alarm effects, false alarms and missed alarms, and air quality recovery, the threshold range, trigger parameters, and alarm priorities are automatically optimized to enable the early warning system to have continuous evolution capabilities. By regularly calculating accuracy and adjusting the update rate and strategy sensitivity, the system can always keep pace with changes in the computer room environment and maintain high reliability and low false alarm rate during long-term operation.

[0172] Example 2, as Figure 2 As shown, the present invention proposes a data center air quality intelligent early warning system based on sensor networks, which is used to execute a data center air quality intelligent early warning method based on sensor networks proposed in Embodiment 1. It includes: a sensor node and self-organizing network module, an edge preprocessing and multi-dimensional fusion module, an adaptive threshold and predictive decision module, and a multi-level early warning and intelligent alarm execution module.

[0173] The sensor node and self-organizing network module is responsible for actual physical sensing and reliable data acquisition. It includes multiple types of sensor nodes (temperature and humidity, particulate matter, gas, wind speed and direction, etc.) and low-power wireless self-organizing gateways. The nodes have local self-testing, health scoring, location identification, adjustable sampling rate and node-end calibration capabilities. It supports inter-node collaborative weight calculation and redundant acquisition strategies, and provides edge time synchronization and secure communication channels. It also includes a field operation and maintenance subsystem for automatic calibration, filter replacement reminders and physical protection mechanisms.

[0174] The edge preprocessing and multidimensional fusion module is deployed on the edge server. It receives the corrected multidimensional state and local anomaly markers, completes credibility assessment, neighborhood credibility propagation, graph-temporal weighted interpolation and robust fusion, outputs a spatially continuous fusion estimate and region-level index with uncertainty quantification, and generates fused data packets for use by the upper layer.

[0175] The adaptive threshold and prediction decision module uses regional indicators, uncertainty and historical event knowledge base to build and evolve dynamic thresholds. It includes component weight adaptor, threshold rolling updater, burst amplification and buffering strategy, and short-term trend prediction engine based on time series model. The module also implements threshold verification feedback and online learning mechanism, and uses false alarm / false negative statistics for parameter self-optimization, thereby outputting adaptive threshold and confidence information for each region.

[0176] The multi-level early warning and intelligent alarm execution module is responsible for transforming threshold decisions into executable operation and maintenance actions. It includes, but is not limited to, an early warning level determiner, time accumulation and neighborhood collaboration triggers, an alarm strategy library (including a set of hierarchical response actions such as alert monitoring, local ventilation adjustment, air purifier start / stop, rack load limiting, etc.), alarm issuance and execution interfaces (for management terminals, mobile push and field controllers), and a closed-loop feedback acquisition unit for recording processing effects and feedback learning.

[0177] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A data center air quality intelligent early warning method based on sensor networks, characterized in that, The specific implementation steps include the following: S1. Deploy multiple types of sensor nodes in key areas of the data center to build a self-organizing collaborative network, collect multi-dimensional air data in real time, and perform node-end correction and local anomaly identification on the collected data. S2. Improve data reliability through dynamic weight fusion between nodes and local anomaly identification, and establish a correlation model between air parameters and data center structure using spatial mapping and time series analysis to identify micro-scale diffusion trends, specifically including: The calibration parameters of each node are mapped to the rack coordinates, air supply and exhaust paths, and cold aisle structure of the data center, constructing a dynamic correspondence framework between multi-dimensional air conditions and spatial layout, specifically: A1. Map the calibration parameters of each node to the rack coordinates, air supply and exhaust paths, and cold aisle structure of the data center to construct a dynamic correspondence framework between multi-dimensional air conditions and spatial layout, specifically including: A2. Extract the corrected multidimensional vector and local anomaly marker from the node data packet, and calculate the node-neighborhood consistency residual; A3. Calculate the initial confidence score and perform local smoothing using the adjacency weights between nodes to obtain the final confidence score; An observation weight matrix is ​​formed based on the adjacency weights and credibility between nodes. Spatial and temporal analysis is then performed by combining graph Laplace smoothing and temporal smoothing. By analyzing the rate and direction of change of air parameters and the persistence of local disturbances using a continuous time window sliding method, microscale diffusion or aggregation trends can be identified within a spatial distribution framework. Specifically: B1. Based on adjacency rights between nodes And combine the confidence level to form the observation weight matrix. : ； ； B2. Defining the Laplace Graph Solve the problem of minimizing spatial smoothness: ； B3. Introduce an exponentially weighted historical term for each grid point and perform time-series smoothing: ； in, L represents the observation weight matrix, i.e., a diagonal matrix; D represents the graph Laplacian matrix; and D represents the degree matrix, a diagonal matrix whose elements are defined as follows: , which is the sum of the connectivity weights of node i and all its neighbors; W represents the weighted adjacency matrix between nodes, where each element is . ; This represents the spatial smoothing intensity coefficient, which controls the trade-off between the observation fitting term and the graph smoothing term; Represents vectorized observation data; This represents the spatial estimation solution vector of the graph Laplace smoothing; This represents the time smoothing coefficient, used to estimate the current time step. Compared with historical estimates Combine; Indicates node i at time The confidence score after neighborhood smoothing; By constructing a cross-regional factor chain, external factors such as cabinet heat load changes, air duct resistance, and fluctuations in the operation of neighboring equipment are coupled with air quality trends for evaluation, quantifying the potential driving forces behind the formation and spread of anomalies. Specifically: C1. Nodes are divided into three categories based on their credibility and original label: Trusted Sets : ; Suspicious Collection : ; Low Trust Set : ; C2, for each sensor Calculate the difference between the estimate and the neighborhood smoothing estimate: ; like If it is true, it is determined to be a real sudden anomaly; otherwise, it is considered a sensor malfunction or noise. C3. For each spatial point, use the weighted Huber loss to calculate the fusion value z: ； ； In implementation, iterative reweighted least squares is used for solving the problem, with initial weight values ​​using... ; C4. If a certain i is determined to be a true burst, then the fusion result at that point is allowed to have higher sensitivity. in, and This indicates the set confidence threshold. This represents a measure of the difference between node observations and fusion estimates. Indicates the adaptive consistency threshold, used for comparison. Is it significant? denoted by Huber loss function, used for robust handling of residuals; z represents fusion estimate, which is the robust fusion value obtained by minimizing Huber loss on observations at a certain spatial point or grid point; The switching point, i.e., the threshold, is when the residual... Use quadratic loss; otherwise, it grows linearly. This indicates an adaptive threshold that depends on the historical standard deviation of fluctuations; This indicates a local abnormality; 1 indicates abnormality, and 0 indicates normality. Based on the perturbation trend and factor risk quantification results, a fusion probability model is used to output the comprehensive credibility and possible causes of air quality anomalies. S3. Construct an adaptive dynamic threshold mechanism, combining historical statistics and real-time feedback to achieve threshold rolling optimization; S4. Generates a graded alarm strategy based on multi-level early warning judgment and triggering conditions, and continuously optimizes the early warning accuracy and response efficiency through closed-loop feedback.

2. The intelligent early warning method for data center air quality based on sensor networks according to claim 1, characterized in that, Step S1 specifically includes: Temperature and humidity sensors, particulate matter sensors, gas concentration sensors, and wind speed and direction sensors are installed in the cold aisle, hot aisle, server rack aisle, air conditioning outlet, and underfloor return air area of ​​the computer room. Construct an air state vector for a single node, which includes ambient temperature, relative humidity, particulate matter concentration, gas concentration, air velocity, and wind direction; Construct a node self-organizing and cooperative network. Each node forms a self-organizing network through low-power wireless communication. The cooperative weight is dynamically calculated based on the distance between nodes and their health status. Local data fusion and preliminary anomaly identification are then performed. The nodes collect air state vectors at an adaptive sampling frequency and attach a unique identifier and timestamp. Key areas are collected at high frequency and redundantly supplemented by neighboring nodes. The collected data undergoes drift correction, filtering, delay compensation, and airflow smoothing at the node end, and also supports preliminary anomaly warning at the node end.

3. The intelligent early warning method for data center air quality based on sensor networks according to claim 2, characterized in that, Constructing a self-organizing and collaborative network of nodes specifically includes: Before each sensor node collaborates with its neighboring nodes in real time, a collaboration weight is calculated. This weight takes into account the distance between nodes, the health status of nodes, and the reliability of data. Each node selects the optimal set of neighbors based on the cooperation weight, and performs weighted fusion of the data of the neighboring nodes to generate a local cooperative air state vector; The network performs dynamic topology maintenance based on node health status and environmental changes, including removing failed nodes, adding new nodes, and adaptively reconstructing the neighbor set and weights.

4. The intelligent early warning method for data center air quality based on sensor networks according to claim 3, characterized in that, Step S3 specifically includes: Based on historical fusion index sequences and outliers, calculate the regional mean and standard deviation, and generate the initial threshold for each region through weighted combination. The contribution weights of each component to the threshold are adjusted in real time using regional uncertainty, and stability is ensured by combining historical weights. The thresholds are updated on a rolling basis based on the adjusted weights and fusion index, while amplification corrections are made for sudden abnormal events. By statistically analyzing historical false alarms and missed alarms, and adjusting the rolling update rate and safety factor based on the false alarm and missed alarm indicators, closed-loop optimization can be achieved.

5. The intelligent early warning method for data center air quality based on sensor networks according to claim 4, characterized in that, Based on historical fusion index sequences and outliers, the regional mean and standard deviation are calculated, and an initial threshold for each region is generated through weighted combination, specifically including: Select the fusion index sequence and corresponding abnormal event labels within the historical time period, and calculate the statistical distribution for each region, including the mean and standard deviation; The initial threshold is defined based on a weighted combination of the mean and standard deviation, with a safety factor taken into account. An initial threshold is generated by weighting the multidimensional components, with the component weights based on historical statistics and risk correlation.

6. The intelligent early warning method for data center air quality based on sensor networks according to claim 5, characterized in that, Step S4 specifically includes: By comparing regional indices with dynamic thresholds and combining uncertainty correction rules, environmental conditions are classified into normal, mild, moderate, and severe warning levels. By setting a time accumulation coefficient and a neighborhood coordination factor, the continuity and spatial consistency of the early warning trigger are judged to avoid false alarms caused by short-term fluctuations or local sporadic events. Different levels of execution actions are matched according to the warning level and the triggering result, and the strategy intensity is dynamically adjusted according to the importance of the area and the real-time status. Based on the actual alarm effect, false alarms and missed alarms, and air quality recovery, the threshold range, trigger parameters, and alarm priority are automatically optimized.

7. The intelligent early warning method for data center air quality based on sensor networks according to claim 6, characterized in that, By setting a time accumulation coefficient and a neighborhood coordination factor, the persistence and spatial consistency of early warning triggering are judged, specifically including: Introduce a time accumulation factor to calculate the proportion of times the threshold condition is met within the past sliding window; Introduce a neighborhood collaboration factor to calculate the proportion of regions exceeding the threshold within the neighborhood at the current time; By combining time accumulation and neighborhood collaboration, trigger conditions are set, and an early warning is triggered when either continuous judgment or neighborhood resonance is satisfied.

8. A data center air quality intelligent early warning system based on sensor networks, used to execute the data center air quality intelligent early warning method based on sensor networks as described in any one of claims 1 to 7, characterized in that, include: The sensor node and self-organizing network module is responsible for actual physical sensing and reliable data acquisition. It includes multiple types of sensor nodes and low-power wireless self-organizing gateways. The nodes have local self-testing, health scoring, positioning identification, adjustable sampling rate and node-end calibration capabilities, and support inter-node collaborative weight calculation and redundant acquisition strategies. The edge preprocessing and multidimensional fusion module is deployed on the edge server. It receives the corrected multidimensional state and local anomaly markers, completes credibility assessment, neighborhood credibility propagation, graph time-series weighted interpolation and robust fusion, and outputs a spatially continuous fusion estimate with uncertainty quantification and a region-level index. The adaptive threshold and prediction decision module uses regional indicators, uncertainty and historical event knowledge base to build and evolve dynamic thresholds. It includes component weight adaptor, threshold rolling updater, burst amplification and buffering strategy, and short-term trend prediction engine based on time series model to realize threshold verification feedback and online learning mechanism. The multi-level early warning and intelligent alarm execution module is responsible for transforming threshold decisions into executable operation and maintenance actions. It includes an early warning level determiner, time accumulation and neighborhood collaboration triggers, alarm policy library, alarm issuance and execution interface, and closed-loop feedback acquisition unit.

Images