Artificial intelligence-based computer room multi-point monitoring environment temperature air conditioning energy-saving system

Abstract

The application provides an artificial intelligence-based computer room multi-point monitoring environment temperature air conditioner energy-saving system, comprising: a space temperature field construction module, used for constructing a space temperature field in a gridded computer room in combination with a detected temperature, a distance factor and an air flow weight; a temperature time trend construction module, used for predicting a temperature based on a sliding window and capturing a time trend in combination with a real temperature; a comprehensive matching degree calculation module, used for clustering the space temperature field and the time trend and calculating a comprehensive matching degree in combination with a mode interval weight; and a temperature regulation instruction generation module, used for inputting the above features into a double-flow attention convolutional network when an abnormal comprehensive matching degree reaches a preset number, performing space flow and time flow feature extraction and channel-space attention fusion, and generating a temperature regulation instruction according to a fault classification probability. Thus, under the premise of safety, fine management of computer room temperature control and energy consumption optimization are realized.

Description

Technical Field

[0001] This application relates to the field of environmental temperature control technology, and in particular to an energy-saving air conditioning system for multi-point monitoring of ambient temperature in a computer room based on artificial intelligence. Background Technology

[0002] With the development of cloud computing and big data, the power consumption of single server racks in modern data centers continues to rise, significantly increasing the heat density of server rooms. Traditional temperature control methods typically rely on single-point temperature measurement and single air conditioner control, which has significant limitations when facing complex temperature fields. For example, the temperature near the air vent may be lower, while the temperature at the top of the rack or where the blind panel is missing may be higher. Single-point readings cannot reflect the overall temperature distribution, easily masking local hotspots, leading to abnormal server load, frequency reduction, or even server crashes, affecting the stable operation of the server room. In addition, factors such as dynamic changes in IT equipment load, redundant switching of air conditioning systems, and changes in airflow within the server room cause the temperature field to evolve rapidly over time and space, exacerbating the limitations of single-point measurements.

[0003] To address the limitations of single-point monitoring, some data centers have deployed multi-point temperature monitoring systems. These systems continuously collect temperature data from different locations to obtain more complete temperature information. However, as the number of collection points increases, problems such as data redundancy, local anomalies, and low utilization of temperature features arise, limiting the efficiency of analyzing and processing the temperature field in the data center.

[0004] Meanwhile, PUE (Power Usage Effectiveness) optimization requires that raising the air conditioning setpoint by 1°C can save 2–4% of cooling energy consumption. However, blindly raising the set temperature may threaten the safety and stable operation of the equipment, further highlighting the need for efficient analysis and utilization of temperature data. Therefore, existing temperature control methods are insufficient to achieve refined management and energy consumption optimization while ensuring safety. Summary of the Invention

[0005] The purpose of this application is to at least address one of the aforementioned technical deficiencies, particularly the technical deficiency that existing temperature control methods cannot achieve refined management and energy consumption optimization while ensuring safety.

[0006] This application provides an artificial intelligence-based multi-point monitoring and energy-saving air conditioning system for computer rooms, the system comprising:

[0007] The spatial temperature field construction module is used to construct the spatial temperature field of each grid after the computer room is gridded, based on the detection temperature, distance factor and airflow weight corresponding to each grid for each type of temperature parameter. The distance factor is determined by the distance between the grid and the cooling air outlet, and the airflow weight is determined by the wind speed and wind direction of the grid.

[0008] The temperature time trend construction module is used to generate predicted temperatures for each type of temperature parameter by using a sliding window based on the continuous real temperatures in each grid, and to capture the temperature time trend of each grid based on each predicted temperature and the corresponding real temperature.

[0009] The comprehensive matching degree calculation module is used to cluster the spatial temperature field and temperature time trend of each type of temperature parameter, generate multiple cluster centers, and calculate the comprehensive matching degree by combining the pattern interval to which each cluster center belongs and its weight.

[0010] The temperature control command generation module is used to input the temperature fields of each space and the temperature time trend of each space into a dual-stream attention convolutional network when the number of comprehensive matching degrees that do not meet the normal mode reaches a preset number. Based on the classification probability of various faults, it generates temperature control commands for the computer room. The dual-stream attention convolutional network is used to extract the spatial flow features of the spatial temperature field and the temporal flow features of the temperature time trend, and then uses channel-space dual attention to perform weighted fusion and classification.

[0011] In one embodiment, the space temperature field construction module includes:

[0012] The distance factor calculation unit is used to obtain the distance from the cooling air outlet to each grid for each type of temperature parameter, and to perform inverse distance weighting calculation on each distance to obtain the distance factor of that type of temperature parameter in each grid.

[0013] The airflow weight calculation unit is used to obtain the wind speed and direction in each grid for each type of temperature parameter, as well as the direction of the connection from the sensor to each grid for that type of temperature parameter, and to calculate the airflow weight of each grid.

[0014] The spatial temperature field construction unit is used to perform weighted interpolation calculations for each type of temperature parameter based on the distance factor, airflow weight, and detection temperature of each grid, so as to obtain the spatial temperature field of that type of temperature parameter in each grid.

[0015] In one embodiment, the formula for calculating the spatial temperature field of each grid is:

[0016]

[0017]

[0018]

[0019] in, This represents the spatial temperature field of the i-th grid. Represents the distance factor in the i-th grid. This represents the airflow weight of the k-th sensor within the i-th grid. This represents the detected temperature measured by the k-th sensor within the i-th grid. Let p represent the distance from the cooling air outlet m to the i-th grid, where p is a positive constant. This represents the distance from the cooling air outlet m to the j-th grid. This represents the wind speed at spatial location (x, y, z) within the grid. It is the direction of the line connecting the k-th sensor to the grid. This indicates the wind direction within the grid.

[0020] In one embodiment, the temperature-time trend building module includes:

[0021] The predicted temperature generation unit is used to obtain continuous real temperatures within each grid for each type of temperature parameter, and to average the real temperatures at continuous time points using a fixed-length sliding window to obtain the predicted temperature corresponding to the next time point at the end of the sliding window.

[0022] The prediction error calculation unit is used to determine the actual temperature at the time point corresponding to each predicted temperature, and to calculate the prediction error value between each predicted temperature and its corresponding actual temperature.

[0023] The average error value calculation unit is used to average the prediction error values ​​of each type of temperature parameter in each grid to obtain the average error value.

[0024] The temperature time trend construction unit is used to calculate the error dispersion for each type of temperature parameter based on the average error value and individual prediction error values ​​of each grid, and obtain the temperature time trend for each grid.

[0025] In one embodiment, the formula for calculating the degree of error dispersion is:

[0026]

[0027]

[0028]

[0029]

[0030] in, Indicates the degree of error dispersion. This represents the average error value. This represents the prediction error value of the k-th sensor within the i-th grid at time point l+n. This indicates the number of prediction error values. This represents the predicted temperature at time point (l+n) of the k-th sensor within the i-th grid. This represents the actual temperature at time (l+n) of the k-th sensor within the i-th grid. This indicates the fixed length of the sliding window.

[0031] In one embodiment, the comprehensive matching degree calculation module includes:

[0032] Fuzzy membership calculation unit is used to calculate the spatial temperature field and temperature time trend of each grid, and the fuzzy membership degree between each cluster center;

[0033] The total fuzzy membership calculation unit is used to determine the pattern interval to which each cluster center belongs, and in each pattern interval, the fuzzy membership corresponding to each cluster center is accumulated to obtain the total fuzzy membership of each pattern interval.

[0034] The comprehensive matching degree calculation unit is used to obtain the comprehensive matching degree by weighted summation and normalization based on the total fuzzy membership degree and weight of each mode interval.

[0035] In one embodiment, the two-stream attention convolutional network includes:

[0036] The input and feature extraction layer is used to extract spatial flow features based on each spatial temperature field, and to extract temporal flow features based on each temperature time trend, and to align the temporal flow features and spatial flow features in the spatial dimension.

[0037] The dual attention enhancement and feature fusion layer is used to determine the spatial channel attention, temporal channel attention, and spatial attention map based on spatial flow features and temporal flow features, and to generate a fused feature map.

[0038] The output classification layer is used to output the classification probability of each category based on the fused feature map.

[0039] In one embodiment, the input and feature extraction layer includes:

[0040] The spatial flow feature extraction unit is used to extract spatial flow features from the spatial flow tensor after converting each spatial temperature field into a spatial flow tensor using a two-dimensional convolution method.

[0041] The time flow feature extraction unit is used to convert the time trends of each temperature into a time flow tensor, and then extract time flow features from the time flow tensor using a one-dimensional convolution method.

[0042] In one embodiment, the dual attention enhancement and feature fusion layer includes:

[0043] The channel weight determination unit is used to perform global average pooling and two-layer bottleneck full connection on spatial flow features and temporal flow features respectively to obtain spatial channel attention and temporal channel attention;

[0044] The spatial attention map determination unit is used to stitch together spatial flow features and temporal flow features to obtain a spatial attention map.

[0045] The attention weighting unit is used to first weight spatial flow features with spatial channel attention, and then with spatial attention map weighting to obtain weighted spatial features; and to first weight temporal flow features with temporal channel attention, and then with spatial attention map weighting to obtain weighted temporal features.

[0046] The weighted fusion unit is used to fuse weighted spatial features and weighted temporal features pixel by pixel according to spatial location to obtain a fused feature map.

[0047] In one embodiment, the output classification layer includes:

[0048] The average pooling result determination unit is used to perform average pooling on the fused feature map to obtain the average pooling result;

[0049] The classification probability output unit is used to output the classification probability of each category based on the average pooling result.

[0050] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0051] This application provides an AI-based multi-point monitoring system for ambient temperature and air conditioning energy saving in computer rooms. By constructing a spatial temperature field and temperature time trend within the computer room, it can not only accurately reflect the overall temperature distribution of the computer room using multi-point monitoring data, but also capture the dynamic characteristics of temperature changes over time, thus avoiding the problem of traditional single-point detection masking local hotspots. The system effectively eliminates data redundancy and local anomaly interference caused by multi-point monitoring through cluster analysis and comprehensive matching degree calculation, improving the utilization rate of temperature features. Simultaneously, when the comprehensive matching degree is abnormal, the system uses a dual-stream attention convolutional network to fuse and analyze spatial and temporal flow features, intelligently identifying abnormal temperature control states in the computer room and generating corresponding control commands, achieving refined and dynamic temperature management. Therefore, this application not only ensures the safety and stability of computer room operation, but also scientifically improves the air conditioning setpoint while ensuring equipment reliability, thereby reducing cooling energy consumption, meeting PUE optimization requirements, and ultimately achieving a balance between energy saving and stable operation in the computer room. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1A schematic diagram of the structure of an AI-based multi-point monitoring ambient temperature air conditioning energy-saving system for computer rooms provided in this application embodiment;

[0054] Figure 2 The "space-time-feature" three-level fusion framework for data center temperature monitoring provided in this application embodiment;

[0055] Figure 3 Example diagram of sliding window prediction values ​​provided in embodiments of this application;

[0056] Figure 4 Example diagram of the spatial-temporal indices provided in this application in a two-dimensional quadrant;

[0057] Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] like Figure 1 As shown, this application provides an artificial intelligence-based multi-point monitoring and energy-saving air conditioning system for computer rooms, the system comprising:

[0060] The spatial temperature field construction module 101 is used to construct the spatial temperature field of each grid after the gridded computer room is set up. For each type of temperature parameter, the module constructs the spatial temperature field of each grid based on the detection temperature, distance factor and airflow weight corresponding to each grid. The distance factor is determined by the distance between the grid and the cooling air outlet, and the airflow weight is determined by the wind speed and wind direction of the grid.

[0061] It should be noted that a server room refers to the spatial environment that houses servers and cooling equipment. It is internally divided into multiple grid units, each of which can be considered an independent temperature monitoring area. The server room is equipped with eight types of temperature sensors, including those for supply air, return air, hot and cold aisles, rack intake air, rack exhaust air, internal equipment, and the overall server room environment. These sensors comprehensively reflect the temperature distribution at different locations and levels within the server room. Due to obstructions from racks, cable trays, and air conditioning vents, the sensor placement within the server room is uneven. Sensors near air vents often measure temperatures lower than the average level of the aisle, while sensors in dead zones or hot loops measure higher temperatures, further exacerbating the unevenness of temperature measurement.

[0062] Temperature parameters refer to the real-time temperature values ​​collected by temperature sensors, used to characterize the ambient temperature state of the grid. Temperature parameter types correspond one-to-one with temperature sensor types; that is, temperature parameters include air temperature, return air temperature, cold aisle temperature, hot aisle temperature, rack inlet air temperature, rack outlet air temperature, equipment internal temperature, and data center ambient temperature. Distance factor refers to the distance between each grid and the cooling air outlet, used to measure the direct impact of cooling air on the grid. Airflow weight is a correction coefficient obtained based on wind speed and direction information at the grid, used to reflect the influence of airflow on the temperature distribution of the grid. Spatial temperature field refers to the overall temperature distribution formed at different grid locations within the data center at the same time, used to reflect the differences and patterns of temperature in the spatial dimension.

[0063] Specifically, the first step is to divide the entire computer room space into a grid, breaking down the originally continuous space into several regular cubic units. For example, the space can be divided into three-dimensional grid units with sides of 1 meter, so that the computer room is uniformly divided into regular cubic regions. The advantage of this approach is that each grid unit can be clearly mapped to the actual collected temperature data, facilitating the mapping and calculation of temperature in the spatial dimension.

[0064] After meshing, temperature sensor data is acquired within each mesh, and this data is further combined with the spatial distance between the mesh and the cooling air outlet to calculate a distance factor. This factor quantifies the attenuation of cold air with distance during propagation, giving higher weights to meshes closer to the outlet in the temperature field calculation, while the influence of meshes farther from the outlet gradually weakens. Simultaneously, wind speed and direction detection devices deployed in the computer room are used to collect local airflow parameters for each mesh in real time, and the corresponding airflow weights are determined based on this data. Airflow weights reflect the speed and direction characteristics of local airflow, effectively characterizing the moderating role of airflow in the temperature distribution formation process.

[0065] Based on this, the detected actual temperature value, the calculated distance factor, and the airflow weight are comprehensively weighted. This calculation method avoids smoothing distortion caused by reliance on remote data in sparse sensor areas, and also prevents local anomalies in dense sensor areas from being masked by a large number of normal values. Therefore, thermal short-circuit areas or local hotspots can be accurately amplified and presented, rather than being masked by the overall averaging effect. The resulting spatial temperature field not only realistically represents the static spatial distribution of temperature within the computer room, but also reflects the dynamic characteristics of the interaction between airflow and cooling equipment, providing reliable data support for refined modeling of the computer room's thermal environment.

[0066] It is understandable that constructing a spatial temperature field by combining temperature detection, distance factor, and airflow weights after gridding the computer room can make the spatial temperature field more consistent with the actual situation of cooling air diffusion and attenuation within the computer room. Specifically, the detected temperature directly reflects the true state collected by the sensors; the distance factor reasonably reflects the differences in cooling effect caused by variations in the distance of different grids from the cooling air outlet; and the introduction of airflow weights further compensates for the influence of wind speed and direction on local temperature distribution. Thus, the resulting spatial temperature field is more continuous and accurate overall, effectively reducing distortion caused by uneven sensor distribution or airflow stratification differences. This avoids smoothing distortion in sparse areas, prevents local anomalies in dense areas from being masked, and reduces errors caused by longitudinal temperature gradients at different height dimensions. Ultimately, the constructed spatial temperature field improves the realism and precision of the computer room temperature distribution model.

[0067] The temperature time trend construction module 102 is used to generate a predicted temperature for each type of temperature parameter by means of a sliding window based on the continuous real temperature in each grid, and to capture the temperature time trend of each grid based on each predicted temperature and the corresponding real temperature.

[0068] In this context, a sliding window refers to a continuous data segment extracted from a time series. Its length and step size can be set according to actual application requirements to balance temporal locality and computational efficiency. Predicted temperature refers to the estimated value for future times generated using sequence modeling methods based on existing real temperature data within the sliding window. The temporal temperature trend refers to the dynamic evolution of temperature over time for each grid cell, characterizing the fluctuations and development patterns of temperature in the time dimension.

[0069] Specifically, temperature data is first continuously collected from temperature sensors deployed in the computer room environment at each grid location. This data is then processed chronologically to form a time series reflecting the temperature change characteristics of that grid. This method allows for a complete record of dynamic temperature changes, providing a reliable data foundation for trend analysis and prediction.

[0070] Subsequently, for each type of temperature parameter, a fixed-length sliding window is used to segment the time series of each grid. The sliding window setting ensures that each time segment retains both local, continuous historical temperature information and provides adequate context, enabling the prediction model to accurately capture short-term temperature trends. At the end of the sliding window, a corresponding predicted temperature value is generated using time series forecasting methods. This value represents the temperature estimate for a future point in time, thus achieving the quantification of future trends based on historical data.

[0071] After generating the predicted temperature value, it is compared with the actual measured temperature at the same time point, and the difference between the two is calculated to form a prediction error sequence. By analyzing the variation of the error over time, the temperature evolution characteristics of each grid at continuous time points can be extracted. This not only quantifies the magnitude of the deviation between prediction and reality but also reveals potential trends and anomaly signals. Finally, the temperature time trends of each grid are uniformly stored and characterized to construct a complete thermal environment change pattern of the computer room in the time dimension, providing accurate dynamic input for anomaly detection and refined temperature control.

[0072] It is understandable that introducing a sliding window based on continuous real temperature data can preserve the temporal correlation of the temperature series and avoid modeling distortion caused by instantaneous fluctuations. Further combining the predicted temperature with the actual temperature reveals the potential evolution direction and rate of change of the temperature, thus more accurately capturing potential thermal anomaly trends within the computer room. This approach improves the accuracy and robustness of temporal trend modeling and effectively avoids misjudgments caused by short-term sensor fluctuations.

[0073] The comprehensive matching degree calculation module 103 is used to cluster the spatial temperature field and temperature time trend of each type of temperature parameter, generate multiple cluster centers, and calculate the comprehensive matching degree by combining the pattern interval to which each cluster center belongs and its weight.

[0074] Cluster centers are representative points extracted from a large amount of temperature feature data using clustering algorithms, summarizing the common characteristics of different temperature patterns. Pattern intervals refer to the range of temperature patterns, with each interval representing a typical temperature distribution or variation. Weights measure the importance or reliability of different pattern intervals, reflecting the contribution of each pattern to the overall matching degree during comprehensive calculation. The overall matching degree is an indicator calculated by weighting the clustering results of the spatial temperature field and temperature time trend, measuring the degree to which the overall temperature state of the computer room conforms to the expected pattern.

[0075] Specifically, the multiple spatial temperature fields and temporal temperature trend data corresponding to each type of temperature parameter are first treated as clustering objects and processed by the K-means clustering algorithm. This algorithm divides the originally discrete and complex spatial and temporal feature data into several clusters, with the center of each cluster serving as the cluster center, representing the common characteristics and typical manifestations of that cluster in terms of spatial distribution and temporal evolution. In this way, the complex data is structured into several representative patterns, providing basic units that are easy to quantify and compare for subsequent analysis.

[0076] Subsequently, the obtained cluster centers are mapped to preset pattern intervals. Each pattern interval corresponds to a typical temperature pattern and can accommodate one or more cluster centers. To reflect the relative importance of different patterns in the comprehensive calculation, each pattern interval can be assigned different weights to ensure that the influence of different patterns can be reasonably distinguished and reflected when calculating the comprehensive matching degree, so that the matching degree result more accurately represents the consistency between the computer room temperature status and the typical pattern.

[0077] Next, based on the weights of cluster centers and pattern intervals, a fusion analysis of the spatial temperature field and temperature temporal trend is performed. For example, according to the weights of each pattern interval, the cluster centers falling into that interval are weighted to obtain a comprehensive matching degree value. This comprehensive matching degree value can quantify the overall consistency between the current data center temperature status and a predetermined typical pattern, providing a scientific and quantifiable basis for judging whether the data center is operating in a reasonable thermal environment, and can be used for subsequent temperature control or anomaly early warning decisions.

[0078] It's understandable that clustering the spatial temperature field and temporal trend corresponding to each type of temperature parameter is done to divide a large amount of scattered temperature feature data into several groups, thereby extracting cluster centers that can represent the common characteristics of each group. Then, by combining the pattern interval to which each cluster center belongs and its weight, a comprehensive calculation is performed to quantify the overall matching degree between the current data center temperature status and typical temperature patterns. Through clustering and weighted calculation, the complexity and noise impact of the original temperature data can be effectively reduced, highlighting the characteristics of typical temperature patterns, while ensuring a reasonable allocation of importance among different pattern intervals. The resulting comprehensive matching degree can accurately reflect the rationality of the overall thermal environment of the data center.

[0079] The temperature control command generation module 104 is used to input the spatial temperature fields and the temperature time trends of each space into the dual-stream attention convolutional network when the number of comprehensive matching degrees that do not meet the normal mode reaches a preset number. Based on the classification probability of various faults, it generates temperature control commands for the computer room. The dual-stream attention convolutional network is used to extract the spatial flow features of the spatial temperature field and the temporal flow features of the temperature time trend, and then use channel-space dual attention to perform weighted fusion and classification.

[0080] The dual-stream attention convolutional network is a deep learning model. Its spatial stream is used to extract local structural features of the spatial temperature field, while its temporal stream is used to extract the temporal evolution features of the temperature trend. The channel-spatial dual attention mechanism is used to enhance the response of key channels and important spatial locations to achieve weighted fusion of features and final classification. The temperature control command is an adjustment operation signal for the air conditioning or refrigeration system of the computer room, generated based on the model classification results, to restore or optimize the thermal environment of the computer room.

[0081] Specifically, eight types of temperature parameters are continuously collected in real time within the computer room. These parameters include supply air, return air, hot and cold aisle air, rack intake and exhaust air, equipment internal temperature, and computer room ambient temperature. Based on the previously established spatial temperature field and temperature time trend, the overall matching degree between each type of temperature parameter and its corresponding normal mode is calculated, thereby quantifying the consistency between the current temperature state and the normal mode. When the overall matching degree of three or more of these eight types of temperature parameters is lower than a preset threshold, triggering an alarm mechanism, it is determined that there is a potentially abnormal thermal environment in the computer room, indicating that control measures need to be taken.

[0082] Subsequently, two key indicators corresponding to these eight temperature parameters—the mean of the spatial temperature field and the volatility of the temperature over time—are provided as input features to a dual-stream attention convolutional network for analysis. The dual-stream attention convolutional network extracts the spatial temperature distribution features of each grid through the spatial flow branch and extracts the dynamic features of temperature changes over time through the temporal flow branch. It then utilizes a channel-spatial attention mechanism to weightedly fuse the feature importance of each channel and the influence of key spatial locations, achieving comprehensive analysis of spatial and temporal information and enhancing key anomaly features. Finally, the network outputs a classification probability Y, where categories include, but are not limited to, normal, cold aisle hotspot, hot aisle overheating, abnormal air conditioning supply, abnormal cabinet air intake, and insufficient overall cooling, reflecting the specific anomaly type of the current temperature state in the computer room. Based on this classification result, intelligent temperature control technology and intelligent or group control solutions for air conditioning can be combined to generate corresponding temperature control commands, achieving precise management of the cold aisle, hot aisle, and overall cooling in the computer room. This optimizes energy consumption distribution, improves operational efficiency, and ensures equipment safety and the stability of the computer room's thermal environment.

[0083] It is understandable that inputting various spatial temperature fields and temporal temperature trends into a dual-stream attention convolutional network is to fully extract the spatial characteristics of the computer room temperature distribution and its dynamic characteristics over time. By processing the spatial and temporal flows separately, key changes in temperature status at different grid locations and time points can be captured. Subsequently, a channel-space dual attention mechanism is used to weightedly fuse each channel and key spatial locations, highlighting abnormal features and suppressing irrelevant information, thereby achieving comprehensive analysis and enhancement of spatial and temporal information. Finally, temperature control commands are generated based on the classification probabilities of various faults output by the network. This allows detected abnormal states to be promptly transformed into specific control measures, achieving precise temperature regulation of the computer room. This effectively improves the accuracy and response speed of temperature anomaly identification, making control commands more scientific and reliable, thus ensuring the safe operation of computer room equipment and optimizing overall energy consumption management.

[0084] To facilitate understanding of the scheme in this application, specific examples are provided below.

[0085] like Figure 2 As shown in the figure, a three-level fusion framework of "space-time-feature" is presented for data center temperature monitoring. First, eight temperature parameters are acquired through multi-point temperature data collection, including supply air temperature, return air temperature, cold aisle temperature, hot aisle temperature, rack inlet air temperature, rack outlet air temperature, equipment internal temperature, and data center ambient temperature. Next, spatial and temporal data fusion is performed separately for each temperature parameter. Spatial data fusion constructs a spatial temperature field by dividing the data center space into a grid and combining spatial distribution and airflow weights, while temporal data fusion constructs the temperature parameter fluctuation using a sliding window and error dispersion measurement. After completing spatial and temporal data fusion, the spatial temperature fields and temporal trends of these eight temperature parameters are fused into feature data. A dual-stream attention convolutional network is used to achieve channel-spatial attention weighted fusion to improve feature utilization. Finally, if the spatial overall matching degree meets the requirements, a temperature control command is output to support second-level decision-making and energy consumption optimization for intelligent temperature control in the data center.

[0086] In the above embodiments, by constructing a spatial temperature field and temperature time trend within the computer room, not only can the overall temperature distribution of the computer room be accurately reflected using multi-point monitoring data, but the characteristics of temperature dynamic changes over time can also be captured, thus avoiding the problem of traditional single-point detection masking local hotspots. This system effectively eliminates data redundancy and local anomaly interference caused by multi-point monitoring through cluster analysis and comprehensive matching degree calculation, improving the utilization rate of temperature features. Simultaneously, when the comprehensive matching degree is abnormal, the system uses a dual-stream attention convolutional network to fuse and analyze spatial and temporal flow features, intelligently identifying abnormal temperature control states in the computer room and generating corresponding control commands, achieving refined and dynamic temperature management. Therefore, this application not only ensures the safety and stability of computer room operation but also scientifically improves the air conditioning setpoint while ensuring equipment reliability, thereby reducing cooling energy consumption, meeting PUE optimization requirements, and ultimately achieving a balance between energy saving and stable operation of the computer room.

[0087] In one embodiment, the space temperature field construction module 101 includes:

[0088] The distance factor calculation unit is used to obtain the distance from the cooling air outlet to each grid for each type of temperature parameter, and to perform inverse distance weighting calculation on each distance to obtain the distance factor of that type of temperature parameter in each grid.

[0089] The airflow weight calculation unit is used to obtain the wind speed and direction in each grid for each type of temperature parameter, as well as the direction of the connection from the sensor to each grid for that type of temperature parameter, and to calculate the airflow weight of each grid.

[0090] The spatial temperature field construction unit is used to perform weighted interpolation calculations for each type of temperature parameter based on the distance factor, airflow weight, and detection temperature of each grid, so as to obtain the spatial temperature field of that type of temperature parameter in each grid.

[0091] Specifically, based on the gridded layout of the computer room, the spatial distance from the cooling air outlet to each grid location is first calculated to ensure that the positional relationship of different grids in geometric space can be accurately depicted. After the distance calculation is completed, an inverse distance weighting operation is performed on these distance data. This method assigns higher weights to grids closer to the air outlet in the cold air diffusion model, while assigning lower weights to grids that are relatively farther away. This results in a distance factor that reflects the actual diffusion law of cold air in space, thus ensuring that the physical characteristics of cold air transmission within the computer room are reasonably mapped.

[0092] Building upon this foundation, wind speed and direction sensors deployed within the computer room are further utilized to collect real-time airflow velocity and direction data covering each grid location, using this data as dynamic input parameters. Simultaneously, by combining the direction of the connection between temperature sensors and the target grid, the airflow transmission path is matched and calculated, thereby quantifying the effect of airflow on heat transfer in different directions. This analysis generates accurate airflow weights that characterize the degree of airflow's influence on temperature distribution, clearly revealing the temperature change characteristics of different grids under the influence of airflow.

[0093] Finally, the obtained distance factor, airflow weight, and real-time acquired temperature values ​​are weighted and interpolated. In this process, spatial geometry and airflow dynamics characteristics are integrated with the measured temperature data, ensuring both a faithful reflection of the physical properties of cold air diffusion in the model and effective utilization of the temperature sensor's measured results. Through this fusion calculation, the spatial temperature field of various temperature parameters in each grid within the computer room is ultimately constructed, thereby achieving a refined modeling and comprehensive representation of the computer room's thermal environment distribution.

[0094] In this embodiment, the reason for obtaining the distance from the cooling air outlet to each grid and performing inverse distance weighting calculation is to highlight the dominant role of near-distance in temperature distribution during the modeling process, avoiding excessive influence of far-distance points on the results, thus obtaining a more reasonable distance factor. Furthermore, obtaining the wind speed and direction of each grid and calculating the airflow weight based on the direction of the line connecting the temperature sensor to the grid can accurately reflect the influence of the airflow transmission path on temperature diffusion, enabling the modeling results to reflect aerodynamic characteristics. Finally, weighted interpolation calculation based on the distance factor, airflow weight, and actual measured temperature can organically integrate spatial geometric relationships, airflow distribution patterns, and measured data, thereby constructing a more accurate spatial temperature field. This avoids the bias caused by single-factor modeling, effectively improving the accuracy and reliability of spatial temperature distribution modeling, and providing a more scientific basis for the monitoring and control of the computer room's thermal environment.

[0095] In one embodiment, the formula for calculating the spatial temperature field of each grid is:

[0096]

[0097]

[0098]

[0099] in, This represents the spatial temperature field of the i-th grid. Represents the distance factor in the i-th grid. This represents the airflow weight of the k-th sensor within the i-th grid. This represents the detected temperature measured by the k-th sensor within the i-th grid. Let p represent the distance from the cooling air outlet m to the i-th grid, where p is a positive constant. This represents the distance from the cooling air outlet m to the j-th grid. This represents the wind speed at spatial location (x, y, z) within the grid. It is the direction of the line connecting the k-th sensor to the grid. This indicates the wind direction within the grid.

[0100] It should be noted that the formula calculates the temperature of each grid as a weighted average of the temperature measurements from all sensors falling within that grid. The weights are determined by two coupled parts: one is a distance factor that decays with geometric distance and is normalized; the other is an airflow weight reflecting the strength and directionality of the local airflow, expressed through wind speed normalization and a cosine term that reflects the angle between the sensor connection direction and the grid wind direction. Therefore, the formula first uses a distance attenuation mechanism to limit the influence of distant sensors on the target point, avoiding the "biasing" of interpolation results in sparse nearby areas by high-density distant sensors. Then, it uses airflow-related terms to adjust the directionality and intensity of each sensor's contribution, suppressing the undue influence of sensors in the upwind direction on the grid temperature and amplifying the signals from sensors with airflow directions consistent with the grid and higher wind speeds.

[0101] The numerator is the sum of weighted temperature measurements, and the denominator is the sum of the corresponding weights, thus ensuring that the output is a normalized local weighted average. This preserves the true gradient of the spatial temperature field while amplifying the signal of thermal short circuits or local hotspots without being diluted by a large number of surrounding normal values. This results in a grid temperature estimate that is robust to uneven sensor distribution, airflow stratification at different altitudes, and local anomalies.

[0102] In one embodiment, the temperature-time trend construction module 102 includes:

[0103] The predicted temperature generation unit is used to obtain continuous real temperatures within each grid for each type of temperature parameter, and to average the real temperatures at continuous time points using a fixed-length sliding window to obtain the predicted temperature corresponding to the next time point at the end of the sliding window.

[0104] The prediction error calculation unit is used to determine the actual temperature at the time point corresponding to each predicted temperature, and to calculate the prediction error value between each predicted temperature and its corresponding actual temperature.

[0105] The average error value calculation unit is used to average the prediction error values ​​of each type of temperature parameter in each grid to obtain the average error value.

[0106] The temperature time trend construction unit is used to calculate the error dispersion for each type of temperature parameter based on the average error value and individual prediction error values ​​of each grid, and obtain the temperature time trend for each grid.

[0107] The prediction error value refers to the difference between the predicted temperature and the corresponding actual temperature. The average error value refers to the statistical result obtained by averaging multiple prediction error values ​​over a certain time range.

[0108] Specifically, within each grid, real temperature data is first continuously collected from deployed sensors, forming a time-varying sequence. To mitigate the impact of anomalous fluctuations at individual time points, this sequence is input into a fixed-length sliding window. The sliding window moves gradually along the time axis, covering a segment of continuous temperature data each time. By averaging all temperatures within the window, short-term fluctuations are effectively smoothed. Thus, at the end of the sliding window, the predicted temperature for the next time point is calculated, enabling prediction based on short-term historical data. Subsequently, this predicted temperature is compared with the actual temperature collected at that time point, and the difference is calculated as the prediction error. By continuously calculating the prediction error at each time point, the deviation between prediction and reality can be evaluated in real time, thereby quantifying the prediction performance at different grid locations.

[0109] Furthermore, the multiple prediction error values ​​generated for each type of temperature parameter within each grid are statistically processed. Specifically, all prediction error values ​​are averaged to obtain the average error value for that grid. This method weakens the impact of individual extreme error points on the overall result, making the error assessment more stable and representative. After obtaining the average error value, the dispersion is calculated based on the actual distribution of each prediction error value within the grid. Dispersion reflects the temporal fluctuation of the prediction error. If the error distribution is relatively concentrated, it indicates that the temperature change is relatively stable; if the dispersion is high, it indicates that the deviation between the temperature prediction and the actual situation fluctuates significantly over time. By analyzing the dispersion, a corresponding temperature time trend can be generated, thereby revealing the time-dimensional change patterns and fluctuation laws of temperature in different areas within the computer room.

[0110] In this embodiment, by averaging continuous temperature data using a sliding window and generating predicted temperatures, short-term fluctuations can be smoothed and quantitative estimates of future temperature changes can be provided. This allows for early detection of temperature trends and reduces the interference of occasional noise in the judgment. Subsequently, calculating the prediction error and further averaging it helps stabilize the error assessment, reducing the impact of individual abnormal measurement points on the overall trend analysis, and making the temperature change assessment of each grid more representative and reliable. Constructing a temperature time trend based on the dispersion of the prediction error reflects the fluctuation pattern of temperature over time, enhancing the ability to identify local anomalies and sudden temperature rises. It also provides physically meaningful time-series characteristics for temperature control decisions, thus improving the overall accuracy, robustness, and predictive guidance of computer room temperature monitoring.

[0111] In one embodiment, the formula for calculating the degree of error dispersion is:

[0112]

[0113]

[0114]

[0115]

[0116] in, Indicates the degree of error dispersion. This represents the average error value. This represents the prediction error value of the k-th sensor within the i-th grid at time point l+n. This indicates the number of prediction error values. This represents the predicted temperature at time point (l+n) of the k-th sensor within the i-th grid. This represents the actual temperature at time (l+n) of the k-th sensor within the i-th grid. This indicates the fixed length of the sliding window.

[0117] In this embodiment, the formula divides the continuous time series into fixed-length sliding windows, averages the temperature data within each window, and generates a predicted temperature that reflects short-term historical trends while suppressing instantaneous fluctuations. Subsequently, the predicted temperature is compared with the actual temperature to obtain the prediction error at each time point. The overall deviation level is then obtained by averaging the prediction errors of multiple sliding windows. Finally, the dispersion of the error relative to the average error is calculated to quantify the fluctuation of each grid in the time dimension. This method effectively reveals the stability and fluctuation characteristics of temperature parameters over continuous time, highlighting abnormal temperature rises or sudden changes while smoothing out minor daily fluctuations. It ensures the reliability and robustness of time trend analysis, providing a physically meaningful and statistically robust quantitative basis for constructing time evolution models of temperature for each grid, which is helpful for subsequent temperature control optimization and anomaly early warning.

[0118] In one example, continuous temperature time series data is divided into fixed-length sliding windows, each containing temperature data for n consecutive time points. The data at time point l can be represented as... Averaging the temperature data from n time points within each sliding window yields the predicted temperature for the next time point after the window ends. To evaluate the smoothing effect of the sliding window, the predicted temperature at time point (l+n) is compared with the actual temperature, and the prediction error is calculated. Figure 3 As shown, for the predicted values ​​of s sliding windows, the overall deviation of each grid in the time dimension can be quantified by the average prediction error. Based on this, the temperature fluctuation of each grid in the time dimension is measured by calculating the dispersion of the error relative to the average error.

[0119] In one embodiment, the comprehensive matching degree calculation module 103 includes:

[0120] Fuzzy membership calculation unit is used to calculate the spatial temperature field and temperature time trend of each grid, and the fuzzy membership degree between each cluster center;

[0121] The total fuzzy membership calculation unit is used to determine the pattern interval to which each cluster center belongs, and in each pattern interval, the fuzzy membership corresponding to each cluster center is accumulated to obtain the total fuzzy membership of each pattern interval.

[0122] The comprehensive matching degree calculation unit is used to obtain the comprehensive matching degree by weighted summation and normalization based on the total fuzzy membership degree and weight of each mode interval.

[0123] Fuzzy membership degree refers to the degree of similarity between the spatial temperature field and temperature temporal trend of each grid and each cluster center. It is used to quantify the probability of a single grid data point belonging to different cluster categories and is an indicator represented by a continuous value. Total fuzzy membership degree represents the comprehensive value of the sum of the fuzzy membership degrees corresponding to all cluster centers within the same pattern interval. It is used to reflect the overall coverage and representativeness of the grid temperature state of that pattern interval.

[0124] Specifically, fuzzy membership calculation can be achieved by comparing the spatial temperature features and time-series features of each grid cell with all cluster centers one by one. The comparison method can employ continuous membership functions, such as Gaussian or exponential membership functions. By calculating the difference between the grid features and the cluster center features, the result is mapped to a continuous membership value, thus reflecting the relative probability of each grid cell's data belonging to different cluster centers. This operation ensures flexible mapping of grid data across multiple modes, allowing local temperature features to reflect both the best-matching cluster center and its partial correlation with other cluster centers.

[0125] Subsequently, by determining the pattern interval to which each cluster center belongs, the total fuzzy membership degree of all cluster centers falling within the same pattern interval is obtained by summing the fuzzy membership degrees of that interval. In practice, cluster centers can first be divided into pattern intervals based on their feature vectors or pattern attributes, and then the membership degrees within each interval can be summed. This step integrates the local information of individual cluster centers into an overall indicator at the pattern interval level, reducing the influence of outliers or extreme values, while ensuring that the pattern interval is more accurately representative of the overall grid temperature state.

[0126] Finally, based on the total fuzzy membership degree and preset weights of each mode interval, the comprehensive matching degree is obtained through weighted summation and normalization. This step allows the comprehensive matching degree to reasonably reflect the contribution of each mode interval to the overall data center temperature status by assigning different weights to the importance of the mode intervals. Through this hierarchical calculation method, both local grid temperature characteristics and overall mode distribution are taken into account, thereby achieving a comprehensive matching assessment between the data center temperature status and typical modes.

[0127] In one example, the spatial and temporal dimensions of a single temperature parameter can be combined for analysis in a two-dimensional quadrant, such as... Figure 4 As shown, taking 10 grids as an example, the spatial temperature field mean and temporal fluctuation of each grid are represented as two indicators. Subsequently, the metrics for all grids were... One-dimensional K-means clustering analysis was used to generate m pattern intervals. to And assign corresponding weights to each interval. to Based on this, cluster centers are determined, and the K-means algorithm is applied again to the 10 grid data within the current window to obtain m cluster centers. to These center points represent the main features of the current dataset.

[0128] Next, the membership degree is calculated for each pattern interval. The sum of the fuzzy membership degrees corresponding to the cluster centers falling within this interval is expressed as: :

[0129]

[0130] Among them, membership degree The following can be calculated based on the distance relationship between data points and cluster centers:

[0131]

[0132] Where m is the number of clusters and z is the fuzziness index, used to control the smoothness of the membership function. Representing data points With cluster center The Euclidean distance between them Representing data points With all cluster centers The Euclidean distance between them.

[0133] Based on the total membership degree of the pattern intervals, calculate the overall matching degree (Match) between all data points and the m pattern intervals:

[0134]

[0135] This method limits the overall matching degree value to between 0 and 1. The closer the value is to 1, the higher the degree of fit between the current data and the normal pattern. Conversely, when it is below a certain threshold (e.g., 0.8), it indicates that the current state may be abnormal, which can trigger the corresponding alarm mechanism, thereby realizing the quantification of the matching degree of temperature parameters in the spatial and temporal dimensions and the monitoring of anomalies.

[0136] It is understandable that by continuously calculating fuzzy membership, flexible assignment of grid data among multiple cluster centers is achieved, minimizing information loss and preserving local details. Secondly, the accumulation of total fuzzy membership can integrate cluster center information, reduce the interference of single-point anomalies on pattern judgment, and improve the representativeness of pattern intervals. Finally, the comprehensive matching degree obtained by weighted summation and normalization can quantify the consistency between the temperature status of the entire computer room and typical patterns, providing a scientific, robust, and physically meaningful decision-making basis for temperature control optimization, anomaly identification, and operation monitoring.

[0137] In one embodiment, the two-stream attention convolutional network includes:

[0138] The input and feature extraction layer is used to extract spatial flow features based on each spatial temperature field, and to extract temporal flow features based on each temperature time trend, and to align the temporal flow features and spatial flow features in the spatial dimension.

[0139] The dual attention enhancement and feature fusion layer is used to determine the spatial channel attention, temporal channel attention, and spatial attention map based on spatial flow features and temporal flow features, and to generate a fused feature map.

[0140] The output classification layer is used to output the classification probability of each category based on the fused feature map.

[0141] In this embodiment, the input and feature extraction layer is used to extract spatial flow features from the spatial temperature field of the computer room, and simultaneously extract temporal flow features from various temperature time trends. The extracted temporal flow features and spatial flow features are aligned spatially to ensure that the temporal features of each spatial unit accurately correspond to its corresponding spatial location features. Spatial flow features represent the local patterns and gradient information of temperature changes in a three-dimensional grid space, while temporal flow features represent the evolutionary patterns and dynamic trends of temperature over time. The alignment operation ensures semantic consistency between features of different dimensions. In practice, spatial flow features can be obtained through convolution operations or other spatial feature extraction operators, while temporal flow features are obtained by processing the temperature time series through a recurrent neural network or a temporal convolutional network. Spatial dimension alignment is then completed through index mapping or tensor rearrangement operations.

[0142] Subsequently, the dual-attention enhancement and feature fusion layer generates spatial channel attention, temporal channel attention, and a spatial attention map based on spatial and temporal flow features, thereby highlighting the features of key regions and key time periods and generating a fused feature map. During implementation, the spatial and temporal channels can be weighted separately to calculate the feature response intensity, and the weights of important channels can be enhanced through an attention mechanism. Simultaneously, a spatial attention map is generated to highlight spatial regions with significant temperature changes. Finally, these attention-weighted features are fused to form a feature map that comprehensively reflects spatial and temporal dynamics.

[0143] Finally, the output classification layer calculates the classification probability of each category based on the fused feature map and outputs the final result. This layer can be implemented using a fully connected network or a convolutional classifier, mapping the fused feature map to the probability distribution of each category to achieve the identification of different temperature control states, abnormal patterns, or temperature evolution types. In implementation, by fusing spatial and temporal features and applying an attention mechanism, the sensitivity to local anomalies and dynamic changes can be enhanced while maintaining an accurate characterization of the overall temperature pattern, thereby achieving higher accuracy and robustness in the classification process. This operation can effectively improve the reliability of temperature control state identification, providing a scientific basis for data center temperature control optimization and anomaly early warning.

[0144] In one embodiment, the input and feature extraction layer includes:

[0145] The spatial flow feature extraction unit is used to extract spatial flow features from the spatial flow tensor after converting each spatial temperature field into a spatial flow tensor using a two-dimensional convolution method.

[0146] The time flow feature extraction unit is used to convert the time trends of each temperature into a time flow tensor, and then extract time flow features from the time flow tensor using a one-dimensional convolution method.

[0147] In this embodiment, the spatial flow feature extraction unit is used to convert the temperature fields of each space in the computer room into spatial flow tensors to represent the distribution and variation information of temperature in a three-dimensional grid. The spatial flow features represent the local spatial temperature patterns and gradient information extracted by two-dimensional convolution operations, which can reflect the temperature differences and spatial correlations between different grids. In practice, each spatial temperature field can be arranged according to the grid index to form a three-dimensional tensor, and then a sliding convolution operation is performed on the tensor through a two-dimensional convolution kernel to extract feature maps with spatial local receptive fields, thereby achieving efficient encoding of temperature distribution patterns.

[0148] The temporal flow feature extraction unit transforms the temporal trends of each temperature into a temporal flow tensor to represent the temperature evolution pattern of each grid in the time dimension. The temporal flow features then extract local temporal dynamic information from the temporal flow tensor through a one-dimensional convolution operation, reflecting the fluctuations and trends of temperature over time. In implementation, the continuous temperature time series of each grid can be arranged into a one-dimensional tensor, and the local feature response can be calculated by sliding a one-dimensional convolution kernel along the time axis, thereby obtaining a temporal feature vector that can capture both short-term changes and long-term trends.

[0149] By performing convolutional operations on the extraction of spatial and temporal flows separately, efficient feature encoding can be achieved while preserving local information and continuous features. This type of operation not only reveals local patterns and gradient differences in spatial temperature distribution but also quantifies the dynamic changes in temperature evolution over time, providing a reliable foundation for subsequent feature fusion and anomaly detection. Furthermore, the weight-sharing mechanism of convolutional computation improves the computational efficiency and robustness of feature extraction, ensuring that spatial and temporal features possess both accuracy and stability during modeling. This provides a scientific and quantifiable basis for optimizing data center temperature control and status identification.

[0150] In one example, the temperature field data is represented using tensors. Specifically, the eight temperature parameters are discretized into a 1m × 1m grid to form a grid map with a spatial resolution of H × W, thus obtaining the spatial flow input.

[0151]

[0152] in, Equals 8, This represents the average temperature field distribution at various spatial locations. The spatial temperature field average is obtained by averaging the data along the z-axis in three-dimensional space. Temporal fluctuations are represented by a tensor as follows:

[0153]

[0154] Each channel corresponds to S sliding windows of data for a single temperature parameter, reflecting the dynamic fluctuation characteristics of the temperature. Spatial stream coding is implemented using two-dimensional convolution operations, and its encoding result is as follows:

[0155]

[0156] It can extract local structural features of the spatial temperature field. The temporal flow coding employs a one-dimensional convolution operation to capture the temporal evolution patterns of each temperature parameter, and the coding result is as follows:

[0157]

[0158] To achieve cross-modal feature fusion, temporal features are remapped to the spatial dimension by repeatedly expanding and reshaping along the time axis.

[0159]

[0160] Thus, it is in spatial dimension with Alignment provides a foundation for subsequent fusion processing of spatial and temporal features.

[0161] In one embodiment, the dual attention enhancement and feature fusion layer includes:

[0162] The channel weight determination unit is used to perform global average pooling and two-layer bottleneck full connection on spatial flow features and temporal flow features respectively to obtain spatial channel attention and temporal channel attention;

[0163] The spatial attention map determination unit is used to stitch together spatial flow features and temporal flow features to obtain a spatial attention map.

[0164] The attention weighting unit is used to first weight spatial flow features with spatial channel attention, and then with spatial attention map weighting to obtain weighted spatial features; and to first weight temporal flow features with temporal channel attention, and then with spatial attention map weighting to obtain weighted temporal features.

[0165] The weighted fusion unit is used to fuse weighted spatial features and weighted temporal features pixel by pixel according to spatial location to obtain a fused feature map.

[0166] In the specific implementation process, the channel weight determination unit obtains global statistical information for each channel by performing global average pooling on spatial flow features and temporal flow features respectively. Then, after two layers of bottleneck full connection operations, spatial channel attention and temporal channel attention are generated, thereby quantifying the importance of each channel.

[0167] The spatial attention map determination unit concatenates spatial flow features and temporal flow features along the channel dimension, and generates a spatial attention map through convolution or normalization operations, thereby reflecting the feature contribution of different locations in the spatial dimension.

[0168] The attention weighting unit first uses spatial channel attention to weight the spatial flow features, and then uses the spatial attention map to weight them to obtain weighted spatial features; at the same time, the temporal flow features are first weighted using temporal channel attention, and then weighted using the same spatial attention map to obtain weighted temporal features.

[0169] Finally, the weighted fusion unit fuses the weighted spatial features and weighted temporal features pixel by pixel according to their spatial location to form a fused feature map, providing a comprehensive feature representation for subsequent classification or prediction tasks.

[0170] In one example, to enhance the important channel responses of both spatial and temporal flow features, a channel attention module with shared weights is first applied to both flows. Global average pooling is then performed on each flow to obtain the global statistical vector of the spatial flow.

[0171]

[0172] Global statistical vector of time flow:

[0173]

[0174] Subsequently, through two layers of bottleneck fully connected operations, using a compression ratio r (usually 16), and passing through ReLU activation and the Sigmoid function σ, channel weight vectors for spatial and temporal flows are generated:

[0175]

[0176]

[0177] in, .

[0178] Output and This is used to adaptively weight each channel to highlight key channel information. Then, the features from the two streams are concatenated along the channel dimension:

[0179]

[0180] A single-channel spatial attention map is generated using 1×1 convolution and the sigmoid activation function.

[0181]

[0182] This spatial attention map reflects the importance of each spatial location, thus providing a basis for subsequent weighting. A sequential attention weighting strategy of "channel first, space later" is then used to weight spatial flow features and temporal flow features separately.

[0183]

[0184]

[0185] Here, ⊙ represents element-wise multiplication (Hadamard product), achieving joint weighting of channel and spatial features. Finally, the weighted spatial and temporal features are fused pixel-by-pixel:

[0186]

[0187] A fused feature map is formed, thereby achieving fine-grained complementary fusion of dual-stream features. This preserves spatial local structural information while taking into account temporal evolution patterns, making the feature representation more discriminative and robust.

[0188] This approach effectively enhances the expressive power of key information from both spatial and temporal flow features, while suppressing redundant and irrelevant features, thus strengthening the selectivity and expressiveness of features in both spatial and temporal dimensions. Channel attention highlights crucial channel information, and spatial attention maps emphasize important spatial locations, resulting in more discriminative feature representations after weighted fusion. This improves the model's ability to recognize and its robustness against temperature field and temporal trend patterns, providing a reliable feature foundation for subsequent classification and anomaly detection.

[0189] In one embodiment, the output classification layer includes:

[0190] The average pooling result determination unit is used to perform average pooling on the fused feature map to obtain the average pooling result;

[0191] The classification probability output unit is used to output the classification probability of each category based on the average pooling result.

[0192] In this embodiment, the average pooling result determination unit is used to perform global average pooling on the fused feature map, which is a multi-channel, high-dimensional feature representation obtained by weighted fusion of the aforementioned spatial flow features and temporal flow features. This fused feature map contains important information about each spatial location and temporal evolution pattern, while the average pooling operation, by globally averaging the feature map in the spatial dimension, can compress the feature dimension and generate an average pooling result vector that comprehensively reflects the distribution of the entire input data.

[0193] The classification probability output unit then maps the average pooling result to the predicted probability distribution of each category through a fully connected layer. The output is then normalized to between 0 and 1 using the Softmax activation function, yielding the classification probability for each category. This operation transforms the high-dimensional fused feature vector into easily interpretable and applicable classification probabilities, thus providing a basis for subsequent decision-making.

[0194] By performing average pooling before outputting classification probabilities, the impact of feature dimensionality on computation can be effectively reduced, while preserving the global discriminative information of the fused features, making the predictions for each category more robust and effective. Furthermore, this method highlights the average response of global features and suppresses the interference of local noise on classification, thereby improving the accuracy and reliability of classification results and ensuring that the model's ability to discriminate temperature change patterns in both spatial and temporal dimensions is fully utilized.

[0195] In one example, global average pooling is performed on the fused feature map to obtain the average pooling result vector:

[0196]

[0197] Then, based on the results of average pooling, the final class classification probability is output through a fully connected layer and a softmax activation function:

[0198]

[0199] in, It is the weight matrix of the classifier, which is initially determined by random assignment. During the training process, its value is continuously adjusted according to the input data until the model converges or reaches the predetermined number of iterations, thereby achieving accurate prediction of each category.

[0200] In one embodiment, this application also provides a computer device storing computer-readable instructions, which, under the control of one or more processors, enable the computer device to implement various functional modules of the AI-based multi-point monitoring ambient temperature air conditioning energy-saving system for computer rooms as described in any of the above embodiments.

[0201] Indicatively, such as Figure 5 As shown, Figure 5 This is a schematic diagram of the internal structure of a computer device 200 provided in an embodiment of this application. The computer device 200 can be provided as a server. (Refer to...) Figure 5 The computer device 200 includes a memory 201 for storing programs and data related to system functions, and a processing component 202, which further includes one or more processors for calling instructions in the memory 201 to implement various functional modules of the system.

[0202] In addition, the computer device 200 may also include a power supply component 203 for managing the device's power supply; a wired or wireless network interface 204 for enabling data communication with external networks or other devices; and an input / output interface 205 for interacting with users or other terminals. The computer device 200 can run an operating system stored in the memory 201, such as Windows Server™, Mac OSX™, Unix™, Linux™, or Free BSD™, providing support for the coordinated operation of various system modules.

[0203] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements to achieve the complete operation of each functional module of the system.

[0204] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the related listed items.

[0205] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0206] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An energy-saving air conditioning system for multi-point monitoring of ambient temperature in a computer room based on artificial intelligence, characterized in that, The system includes: The spatial temperature field construction module is used after the computer room is gridded. For each type of temperature parameter, it obtains the distance from the cooling air outlet to each grid, performs inverse distance weighting calculation on each distance to obtain the distance factor of that type of temperature parameter in each grid, obtains the wind speed and direction in each grid, and the direction of the line connecting the corresponding sensor of that type of temperature parameter to each grid, and calculates the airflow weight of each grid. Based on the distance factor, airflow weight, and detected temperature of each grid, it performs weighted interpolation calculation to obtain the spatial temperature field of that type of temperature parameter in each grid. The calculation formula for the spatial temperature field of each grid is as follows: in, This represents the spatial temperature field of the i-th grid. Represents the distance factor in the i-th grid. This represents the airflow weight of the k-th sensor within the i-th grid. This represents the detected temperature measured by the k-th sensor within the i-th grid. Let p represent the distance from the cooling air outlet m to the i-th grid, where p is a positive constant. This represents the distance from the cooling air outlet m to the j-th grid. This represents the wind speed at spatial location (x, y, z) within the grid. It is the direction of the line connecting the k-th sensor to the grid. This indicates the wind direction within the grid; The temperature time trend construction module is used to generate a predicted temperature for each type of temperature parameter by means of a sliding window based on the continuous real temperature in each grid, and to capture the temperature time trend of each grid based on each predicted temperature and the corresponding real temperature. The comprehensive matching degree calculation module is used to cluster the spatial temperature field and temperature time trend of each type of temperature parameter, generate multiple cluster centers, and calculate the comprehensive matching degree by combining the pattern interval to which each cluster center belongs and its weight. The temperature control command generation module is used to input each of the spatial temperature fields and each of the temperature time trends into a dual-stream attention convolutional network when the number of comprehensive matching degrees that do not meet the normal mode reaches a preset number. Based on the classification probabilities of various faults, it generates the temperature control command for the computer room. The dual-stream attention convolutional network is used to extract the spatial flow features of the spatial temperature field and the temporal flow features of the temperature time trend, and then use channel-space dual attention to perform weighted fusion and classification.

2. The AI-based multi-point monitoring and energy-saving air conditioning system for computer rooms according to claim 1, characterized in that, The temperature-time trend construction module includes: The predicted temperature generation unit is used to obtain continuous real temperatures within each grid for each type of temperature parameter, and to average the real temperatures at continuous time points using a sliding window of fixed length to obtain the predicted temperature corresponding to the next time point after the end of the sliding window. The prediction error calculation unit is used to determine the actual temperature at the time point corresponding to each predicted temperature, and to calculate the prediction error value between each predicted temperature and its corresponding actual temperature. An average error value calculation unit is used to average the prediction error values ​​of each type of temperature parameter in each grid to obtain an average error value. A temperature time trend construction unit is used to calculate the degree of error dispersion for each type of temperature parameter based on the average error value and each prediction error value of each grid, and obtain the temperature time trend of each grid.

3. The AI-based multi-point monitoring and energy-saving air conditioning system for computer rooms according to claim 2, characterized in that, The formula for calculating the degree of error dispersion is as follows: in, This indicates the degree of error dispersion. This represents the average error value. This represents the prediction error value of the k-th sensor within the i-th grid at time point l+n. This indicates the number of prediction error values. This represents the predicted temperature at time point (l+n) of the k-th sensor within the i-th grid. This represents the actual temperature at time point (l+n) of the k-th sensor within the i-th grid. This indicates the fixed length of the sliding window.

4. The AI-based multi-point monitoring and energy-saving air conditioning system for computer rooms according to claim 1, characterized in that, The comprehensive matching degree calculation module includes: A fuzzy membership calculation unit is used to calculate the spatial temperature field and temperature-time trend of each grid, and the fuzzy membership degree between each cluster center; The total fuzzy membership calculation unit is used to determine the pattern interval to which each cluster center belongs, and to accumulate the fuzzy membership degrees corresponding to each cluster center in each pattern interval to obtain the total fuzzy membership degree of each pattern interval. The comprehensive matching degree calculation unit is used to obtain the comprehensive matching degree by weighted summation and normalization based on the total fuzzy membership degree and weight of each of the mode intervals.

5. The AI-based multi-point monitoring and energy-saving air conditioning system for computer rooms according to claim 1, characterized in that, The dual-stream attention convolutional network includes: The input and feature extraction layer is used to extract spatial flow features based on each of the spatial temperature fields, and to extract temporal flow features based on each of the temperature time trends, and to align the temporal flow features and the spatial flow features in the spatial dimension. A dual attention enhancement and feature fusion layer is used to determine spatial channel attention, temporal channel attention, and spatial attention map based on the spatial flow features and the temporal flow features, and to generate a fused feature map; The output classification layer is used to output the classification probability of each category based on the fused feature map.

6. The AI-based multi-point monitoring and energy-saving air conditioning system for computer rooms according to claim 5, characterized in that, The input and feature extraction layer includes: The spatial flow feature extraction unit is used to convert each of the spatial temperature fields into a spatial flow tensor, and then extract the spatial flow features from the spatial flow tensor using a two-dimensional convolution method. The time flow feature extraction unit is used to convert each of the temperature time trends into a time flow tensor, and then extract the time flow features from the time flow tensor using a one-dimensional convolution method.

7. The AI-based multi-point monitoring and energy-saving air conditioning system for computer rooms according to claim 5, characterized in that, The dual attention enhancement and feature fusion layer includes: The channel weight determination unit is used to perform global average pooling and two-layer bottleneck full connection on the spatial flow features and the temporal flow features respectively to obtain the spatial channel attention and the temporal channel attention; A spatial attention map determination unit is used to stitch together the spatial flow features and the temporal flow features to obtain the spatial attention map; The attention weighting unit is used to first weight the spatial flow features with the spatial channel attention, and then with the spatial attention map to obtain weighted spatial features; and to first weight the temporal flow features with the temporal channel attention, and then with the spatial attention map to obtain weighted temporal features. The weighted fusion unit is used to fuse the weighted spatial features and the weighted temporal features pixel by pixel according to their spatial location to obtain a fused feature map.

8. The AI-based multi-point monitoring and energy-saving air conditioning system for computer rooms according to claim 5, characterized in that, The output classification layer includes: The average pooling result determination unit is used to perform average pooling on the fused feature map to obtain the average pooling result; The classification probability output unit is used to output the classification probability of each category based on the average pooling result.

Images