Machine room multi-point environment temperature monitoring air-conditioning energy-saving system based on artificial intelligence

Abstract

The invention provides a machine room multi-point environment temperature monitoring air conditioner energy-saving system based on artificial intelligence, and the system comprises a space temperature field construction module which is used for constructing a space temperature field in combination with the detection temperature, the distance factor and the airflow weight in a grid machine room; the temperature time trend construction module is used for predicting the temperature based on the sliding window and capturing the time trend in combination with the real temperature; the comprehensive matching degree calculation module is used for clustering the space temperature field and the time trend and calculating the comprehensive matching degree in combination with the mode interval weight; and the temperature regulation and control instruction generation module is used for inputting the features into a double-flow attention convolutional network when the comprehensive matching degree abnormities reach a preset number, and generating a temperature regulation and control instruction according to the fault classification probability after spatial flow and time flow feature extraction and channel-space attention fusion. Therefore, on the premise that safety is guaranteed, refined management and energy consumption optimization of machine room temperature control are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental temperature control, and in particular to an artificial intelligence-based multi-point monitoring environmental temperature air conditioner energy-saving system for a computer room. BACKGROUND

[0002] With the development of cloud computing and big data, the power of a single cabinet in a modern data center continues to rise, and the heat density of a computer room significantly increases. Traditional temperature control methods usually rely on single-point temperature measurement and single-air-conditioner control. This method has obvious limitations when facing complex temperature fields. For example, the temperature near the air supply outlet may be low, while the temperature at the top of the cabinet or where the blind plate is missing may be high. Single-point readings cannot reflect the overall temperature distribution and easily hide local hot spots, leading to server load abnormalities, frequency reduction, or even downtime, affecting the stable operation of the computer room. In addition, the dynamic changes in IT equipment load, the redundant switching of air conditioning systems, and the changes in air flow inside the computer room all contribute to the rapid evolution of the temperature field over time and space, exacerbating the limitations of single-point measurement.

[0003] To address the limitations of single-point detection, some computer rooms deploy multi-point temperature monitoring systems to collect temperature data at 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 characteristics arise, limiting the analysis and processing efficiency of the computer room temperature field.

[0004] At the same time, PUE (Power Usage Effectiveness) optimization requires that for every 1 ℃ increase in air conditioner setpoint, 2-4% of refrigeration energy consumption can be saved. However, blindly increasing the set temperature may threaten the safety and stable operation of equipment, further highlighting the need for efficient analysis and utilization of temperature data. Therefore, existing temperature control methods cannot achieve fine-grained management and energy consumption optimization while ensuring safety. SUMMARY

[0005] The present application aims to at least solve one of the above technical defects, particularly the technical defect that existing temperature control methods cannot achieve fine-grained management and energy consumption optimization while ensuring safety.

[0006] The present application provides an artificial intelligence-based multi-point monitoring environmental temperature air conditioner energy-saving system for a computer room, which includes:

[0007] A spatial temperature field construction module is used to grid the computer room. For each type of temperature parameter, the spatial temperature field of each grid is constructed based on its corresponding detected temperature, distance factor, and air flow weight in each grid. The distance factor is determined by the distance between the grid and the refrigeration air outlet, and the air flow weight is determined by the wind speed and direction of the grid.

[0008] a temperature time trend construction module, configured to, for each type of temperature parameter, generate a predicted temperature by a sliding window according to its continuous real temperature in each grid, and capture a temperature time trend of each grid based on each predicted temperature and the corresponding real temperature;

[0009] a comprehensive matching degree calculation module, configured to, for each type of temperature parameter, cluster its spatial temperature field and temperature time trend, generate a plurality of cluster centers, and calculate a comprehensive matching degree in combination with a mode interval to which each cluster center belongs and a weight thereof;

[0010] a temperature regulation instruction generation module, configured to, when the number of comprehensive matching degrees that do not meet the normal mode reaches a preset number, input each spatial temperature field and each temperature time trend into a double-flow attention convolutional network, and generate a temperature regulation instruction of the machine room according to classification probabilities of each type of fault, the double-flow attention convolutional network being configured to extract spatial flow features of the spatial temperature field and time flow features of the temperature time trend, perform weighted fusion by using channel-space double attention, and classify.

[0011] In an embodiment, the spatial temperature field construction module comprises:

[0012] a distance factor calculation unit, configured to, for each type of temperature parameter, obtain a distance from a refrigeration air outlet to each grid, and perform inverse distance weighting calculation on each distance to obtain a distance factor of the type of temperature parameter in each grid;

[0013] an air flow weight calculation unit, configured to, for each type of temperature parameter, obtain a wind speed and a wind direction in each grid, and a connecting line direction of a sensor corresponding to the type of temperature parameter to each grid, and calculate an air flow weight of each grid;

[0014] a spatial temperature field construction unit, configured to, for each type of temperature parameter, perform weighted interpolation calculation based on the distance factor, the air flow weight and the detection temperature of each grid to obtain a spatial temperature field of the type of temperature parameter in each grid.

[0015] In an embodiment, a calculation formula of the spatial temperature field of each grid is:

[0016]

[0017]

[0018]

[0019] wherein, denotes a spatial temperature field of the i-th grid, denotes a distance factor in the i-th grid, denotes an air flow weight of the k-th sensor in the i-th grid, Tik, l represents the detection temperature measured by the kth sensor in the ith grid, Dm, i represents the distance from the refrigeration air outlet m to the ith grid, and p is a positive constant, Dm, j represents the distance from the refrigeration air outlet m to the jth grid, V(x, y, z) represents the air speed at the spatial position (x, y, z) in the grid, is the direction of the line connecting the kth sensor to the grid, represents the wind direction in the grid.

[0020] In an embodiment, the temperature time trend construction module comprises:

[0021] a predicted temperature generation unit configured to, for each type of temperature parameter, acquire continuous real temperatures in each grid, and perform average processing on the real temperatures at continuous time points by using a fixed-length sliding window to obtain a predicted temperature corresponding to a next time point at the end of the sliding window;

[0022] a predicted error value calculation unit configured to determine the real temperature at the corresponding time point of each predicted temperature, and calculate a predicted error value between each predicted temperature and the corresponding real temperature;

[0023] an average error value calculation unit configured to perform average processing on each predicted error value of each type of temperature parameter in each grid to obtain an average error value;

[0024] a temperature time trend construction unit configured to, for each type of temperature parameter, calculate an error dispersion degree according to the average error value and each predicted error value of each grid to obtain a temperature time trend of each grid.

[0025] In an embodiment, the calculation formula of the error dispersion degree is:

[0026]

[0027]

[0028]

[0029]

[0030] wherein, represents the error dispersion degree, represents the average error value, Tik, l+n represents the predicted error value of the kth sensor in the ith grid at the l+n time point, represents the number of predicted error values, Tik, l+n represents the predicted temperature of the kth sensor in the ith grid at the l+n time point, represents the real temperature of the kth sensor in the ith grid at the l+n time point, represents the fixed length of the sliding window.

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

[0032] a fuzzy membership calculation unit configured to calculate the fuzzy membership between each grid's spatial temperature field and temperature time trend and each cluster center;

[0033] a total fuzzy membership calculation unit configured to determine the mode interval to which each cluster center belongs, and accumulate the fuzzy membership corresponding to each cluster center in each mode interval to obtain the total fuzzy membership of each mode interval;

[0034] a comprehensive matching degree calculation unit configured to obtain the comprehensive matching degree by weighted summation and normalization according to the total fuzzy membership and weight of each mode interval.

[0035] In one embodiment, the dual-flow attention convolutional network comprises:

[0036] an input and feature extraction layer configured to extract spatial flow features according to each spatial temperature field, and extract time flow features according to each temperature time trend, and align the time flow features and the spatial flow features in the spatial dimension;

[0037] a dual attention enhancement and feature fusion layer configured to determine spatial channel attention, time channel attention and spatial attention map according to the spatial flow features and the time flow features, and generate a fused feature map;

[0038] an output classification layer configured to output classification probabilities of each category according to the fused feature map.

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

[0040] a spatial flow feature extraction unit configured to convert each spatial temperature field into a spatial flow tensor, and then extract spatial flow features from the spatial flow tensor by using a two-dimensional convolution method;

[0041] a time flow feature extraction unit configured to convert each temperature time trend into a time flow tensor, and then extract time flow features from the time flow tensor by using a one-dimensional convolution method.

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

[0043] a channel weight determination unit configured to perform global average pooling and two-layer bottleneck fully connected on the spatial flow features and the time flow features respectively to obtain the spatial channel attention and the time channel attention;

[0044] The spatial attention map determination unit is configured to splice the spatial stream feature and the time stream feature to obtain a spatial attention map.

[0045] The attention weighting unit is configured to weight the spatial stream feature by using the spatial channel attention first and then using the spatial attention map, to obtain a weighted spatial feature, and weight the time stream feature by using the time channel attention first and then using the spatial attention map, to obtain a weighted time feature.

[0046] The weighted fusion unit is configured to fuse the weighted spatial feature and the weighted time feature pixel by pixel according to spatial positions to obtain a fused feature map.

[0047] In an embodiment, the output classification layer comprises:

[0048] The average pooling result determination unit is configured to perform average pooling on the fused feature map to obtain an average pooling result.

[0049] The classification probability output unit is configured to output classification probabilities of various categories according to the average pooling result.

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

[0051] The artificial intelligence-based computer room multi-point monitoring environment temperature air conditioner energy-saving system provided by the present application can not only accurately reflect the overall temperature distribution of the computer room by using multi-point monitoring data, but also capture the characteristics of dynamic changes of temperature over time, thereby avoiding the problem of traditional single-point detection covering up local hot spots. Through clustering analysis and comprehensive matching degree calculation, the system effectively eliminates data redundancy and local abnormal interference caused by multi-point monitoring, and improves the utilization rate of temperature characteristics. At the same time, when the system is in a comprehensive matching degree anomaly, the system can intelligently identify abnormal temperature control states of the computer room and generate corresponding control instructions by fusing and analyzing spatial stream and time stream features through a double-flow attention convolutional network, thereby realizing fine and dynamic temperature management. Therefore, the present application not only ensures the safety and stability of the computer room operation, but also scientifically improves the air conditioner set point on the premise of ensuring the reliability of the equipment, thereby reducing the refrigeration energy consumption, meeting the requirements of PUE optimization, and finally realizing the unity of computer room energy saving and consumption reduction and stable operation. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0053] Figure 1A structural schematic diagram of an energy-saving system of an air conditioner for multi-point monitoring of ambient temperature in a computer room based on artificial intelligence is provided for an embodiment of the present application.

[0054] Figure 2 A "space-time-feature" three-level fusion framework for temperature monitoring in a computer room is provided for an embodiment of the present application.

[0055] Figure 3 A sliding window prediction value example diagram is provided for an embodiment of the present application.

[0056] Figure 4 An example diagram of a space-time index in a two-dimensional quadrant is provided for an embodiment of the present application.

[0057] Figure 5 An internal structure schematic diagram of a computer device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0059] As shown in Figure 1 The present application provides an energy-saving system of an air conditioner for multi-point monitoring of ambient temperature in a computer room based on artificial intelligence, which comprises:

[0060] The space temperature field construction module 101 is used to grid the computer room, and for each type of temperature parameter, construct the space temperature field of each grid according to its corresponding detected temperature, distance factor and airflow weight in 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 direction of the grid.

[0061] It should be noted that the computer room refers to a space environment carrying servers and cooling equipment, which is divided into multiple grid units, and each grid can be regarded as an independent temperature detection area. The computer room is arranged with eight types of temperature sensors, including supply air, return air, cold and hot channels, cabinet air inlet, cabinet air outlet, equipment interior and computer room environment sensors, which are used to comprehensively reflect the temperature distribution state at different positions and different levels in the computer room. Due to the shielding of cabinets, cable trays and air conditioner outlets inside the computer room, the sensors are arranged with uneven density characteristics. The sensors close to the air outlet often measure a lower temperature than the average level of the channel, while the sensors in the dead angle or hot loop are higher, further exacerbating the unevenness of space temperature measurement.

[0062] The temperature parameter refers to a temperature value collected by a temperature sensor in real time, and is used to represent the temperature state of the grid. The temperature parameter type corresponds to the temperature sensor type one by one, that is, the temperature parameter includes the air temperature, the return air temperature, the cold channel temperature, the hot channel temperature, the cabinet inlet air temperature, the cabinet outlet air temperature, the equipment internal temperature and the computer room environment temperature. The distance factor refers to the distance parameter between each grid relative to the cooling air outlet, which is used to measure the direct effect of the cooling air on the grid. The airflow weight is a correction coefficient obtained according to the wind speed and wind direction information at the grid, which is used to reflect the influence of air flow on the temperature distribution of the grid.

[0063] Specifically, first, the entire space of the computer room is divided into grids, and the originally continuous space is divided into several regular cubic units. For example, the space can be divided into three-dimensional grid units with a side length of 1 meter, so that the computer room is evenly divided into regular cubic regions. The advantage of this processing is that each grid unit can be clearly corresponded to the actually collected temperature data, facilitating the mapping and calculation of temperature in the spatial dimension.

[0064] After completing the grid division, the temperature sensor collection data in each grid is obtained, and further combined with the spatial distance between the grid and the cooling air outlet, the corresponding distance factor is calculated. The factor can quantify the effect of the cooling air attenuation with distance in the propagation process, so that the grid closer to the air outlet has a higher weight in the temperature field calculation, and the influence of the grid far away from the air outlet gradually weakens. At the same time, the wind speed and wind direction detection device arranged in the computer room is called to collect the local airflow parameters of each grid in real time, and the corresponding airflow weight is determined based on these data. The airflow weight can reflect the speed and direction characteristics of the local air flow, thereby effectively depicting the adjusting effect of the airflow in the temperature distribution formation process.

[0065] On this basis, the actual temperature value, the calculated distance factor and the airflow weight are comprehensively weighted and calculated. This calculation method avoids the smoothing distortion caused by relying on remote data in the sparse sensor area, and also prevents the local anomaly in the dense sensor area from being covered by a large number of normal values. As a result, the hot short circuit area or local hot spot can be accurately amplified and presented, rather than being covered by the overall average effect. The finally generated spatial temperature field can not only truly represent the static distribution of temperature in space in the computer room, but also reflect the dynamic characteristics under the interaction of airflow and cooling equipment, providing reliable data support for the fine modeling of the computer room thermal environment.

[0066] It can be understood that by adopting the comprehensive method of detecting temperature, distance factor and air flow weight after griding the machine room, the space temperature field can be constructed, which can make the space temperature field more consistent with the actual situation of the diffusion and attenuation of the cooling air in the machine room. Among them, the detected temperature can directly reflect the real state collected by the sensor, combined with the distance factor, it can reasonably reflect the difference in cooling effect caused by the change of distance from the cooling air outlet in different grids, and the introduction of air flow weight further compensates the influence of wind speed and direction on local temperature distribution. In this way, the obtained space temperature field is more continuous and accurate as a whole, which can effectively reduce the distortion problem caused by uneven distribution of sensors or air flow stratification difference, thereby avoiding the smoothing distortion in sparse areas, preventing the local anomalies in dense areas from being covered up, and weakening the error caused by longitudinal temperature gradient in different height dimensions. Finally, the constructed space temperature field can improve the authenticity and fineness of the temperature distribution modeling of the machine room.

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

[0068] The sliding window refers to a continuous data segment intercepted on the time series, and its length and step can be set according to actual application requirements, which is used to balance the time locality and computational efficiency. The predicted temperature refers to the estimated value of the future time based on the existing real temperature data in the sliding window, which is generated by a sequence modeling method. The temperature time trend refers to the dynamic evolution characteristics of the temperature of each grid with time, which is used to represent the fluctuation and development law of the temperature in the time dimension.

[0069] Specifically, first, the temperature data of each grid position is continuously collected from the temperature sensors arranged in the machine room, and is arranged in order according to the collection time, forming a time series reflecting the temperature change characteristics of the grid. In this way, the dynamic change of temperature can be recorded completely, providing a reliable data basis for trend analysis and prediction.

[0070] Subsequently, for each type of temperature parameter, a fixed-length sliding window is used to segment and intercept on the time series of each grid. The setting of the sliding window can ensure that each time segment not only retains the local continuous historical temperature information, but also provides appropriate context environment, so that the prediction model can accurately capture the short-term temperature change trend. At the end of the sliding window, a time series prediction method is used to generate a corresponding predicted temperature value, which represents the temperature estimate at a future time point, thereby realizing the quantification of future trends based on historical data.

[0071] After generating the predicted temperature values, they are compared with the actual measured temperatures at the same time points, and the differences between the two are calculated to form a prediction error sequence. By analyzing the error over time, the temperature evolution characteristics of each grid at consecutive time points can be extracted, which not only quantifies the deviation between prediction and actual, but also reveals potential trends and abnormal signals. Finally, the temperature time trends of each grid are uniformly stored and characterized, and a complete thermal environment change pattern of the machine room in the time dimension is constructed, providing accurate dynamic input for anomaly detection and refined temperature control.

[0072] It can be understood that by introducing a sliding window based on the continuous real temperature, the time correlation of the temperature sequence can be preserved, avoiding modeling distortion caused by instantaneous fluctuations. Further combining the predicted temperature with the real temperature, the potential evolution direction and change speed of the temperature can be revealed, so that the potential thermal abnormal trend inside the machine room can be more accurately captured. This method can improve the accuracy and robustness of time trend modeling, effectively avoiding false judgments caused by short-term fluctuations of the sensor.

[0073] The comprehensive matching degree calculation module 103 is configured to, for each type of temperature parameter, cluster its spatial temperature field and temperature time trend, generate a plurality of cluster centers, and calculate a comprehensive matching degree in combination with the mode interval to which each cluster center belongs and the weight thereof.

[0074] The cluster center refers to a representative point extracted from a large amount of temperature feature data by a clustering algorithm, which can summarize the common characteristics of different temperature modes. The mode interval refers to the temperature mode range, each interval representing a typical temperature distribution or change. The weight is a measure of the importance or reliability of different mode intervals, which is used to reflect the contribution of each mode to the overall matching degree in the comprehensive calculation. The comprehensive matching degree is an index that measures the degree of conformity of the overall temperature state of the machine room to the expected mode, which is calculated by combining the clustering results of the spatial temperature field and the temperature time trend.

[0075] Specifically, first, the plurality of spatial temperature fields and temperature time trend data corresponding to each type of temperature parameter are input as clustering objects into the K-means clustering algorithm for processing. Through this algorithm, the originally discrete and complex distributed spatial and time feature data can be divided into several groups, and the center of each group is the cluster center, which represents the common characteristics and typical performance of the group in spatial distribution and time evolution. In this way, complex data is structured into several representative modes, providing a basis for subsequent analysis that is easy to quantify and compare.

[0076] Subsequently, the obtained respective cluster centers are mapped into preset mode intervals, each of which corresponds to a typical temperature mode and can accommodate one or more cluster centers falling into. In order to reflect the relative importance of different modes in the comprehensive calculation, different weights can be assigned to the mode intervals to ensure that the influence of different modes can be reasonably distinguished and reflected in the calculation of the comprehensive matching degree, so that the matching degree result more accurately represents the consistency between the computer room temperature state and the typical mode.

[0077] Then, based on the cluster centers and the weights of the mode intervals, the spatial temperature field and the temperature time trend are fused and analyzed. For example, according to the weight of each mode interval, the cluster centers falling into the interval are weighted and calculated to obtain a comprehensive matching degree value. The comprehensive matching degree value can quantify the overall consistency between the current computer room temperature state and the given typical mode, providing a scientific and quantifiable basis for judging whether the computer room is running in a reasonable thermal environment, and can be used for subsequent temperature control or abnormal warning decision.

[0078] It can be understood that the reason for clustering the spatial temperature field and the temperature time trend corresponding to each type of temperature parameter is to divide a large number of scattered temperature feature data into several groups, so as to extract the cluster centers that can represent the common characteristics of each group. Then, the overall matching degree between the current computer room temperature state and the typical temperature mode is quantified by combining the mode interval to which each cluster center belongs and its weight for comprehensive calculation. Through clustering and weighted calculation, the complexity and noise influence of the original temperature data can be effectively reduced, and the typical temperature mode features can be highlighted, while ensuring reasonable allocation of importance between different mode intervals, and the comprehensive matching degree obtained can accurately reflect the rationality of the overall thermal environment of the computer room.

[0079] The temperature regulation instruction generation module 104 is configured to input the respective spatial temperature fields and the respective temperature time trends into a double-flow attention convolutional network when the number of comprehensive matching degrees that do not satisfy the normal mode reaches a preset number, and generate a temperature regulation instruction for the computer room according to the classification probabilities of the respective types of faults output by the double-flow attention convolutional network. The double-flow attention convolutional network is used to extract spatial flow features of the spatial temperature field and time flow features of the temperature time trend, and then perform weighted fusion and classification using channel-spatial double attention.

[0080] The double-flow attention convolutional network is a deep learning model, the spatial flow of which is used to extract local structural features of the spatial temperature field, the time flow of which is used to extract time evolution features of the temperature time trend, and the channel-spatial double attention mechanism is used to enhance the response of key channels and important spatial positions to realize weighted fusion and final classification of features. The temperature regulation instruction is a regulation operation signal for the computer room air conditioner or refrigeration system generated according to the model classification result, which is used to restore or optimize the thermal environment of the computer room.

[0081] Specifically, 8 types of temperature parameters arranged in the machine room are continuously collected in real time, including supply air, return air, cold and hot channels, cabinet inlet and outlet air, equipment internal temperature, and machine room environment temperature, etc. Based on the previously constructed spatial temperature field and temperature time trend, the comprehensive matching degree of each type of temperature parameter with the corresponding normal mode is calculated respectively, so as to quantify the consistency of the current temperature state with the normal mode. When the comprehensive matching degree of 3 or more than 3 types of temperature parameters is lower than the preset threshold, triggering the alarm mechanism, it is determined that there is a potential abnormal thermal environment in the machine room, prompting the need to take control measures.

[0082] Subsequently, the two key indicators corresponding to the 8 types of temperature parameters, i.e. the mean of the spatial temperature field and the volatility of the temperature time trend, are provided as input features to the 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 stream branch, extracts the dynamic features of the temperature change over time through the time stream branch, and uses the channel-spatial attention mechanism to weight and fuse the feature importance of each channel and the influence of the key spatial position, realizing the comprehensive analysis of spatial and temporal information and strengthening the key abnormal features. Finally, the network outputs a classification probability Y, where the categories include but are not limited to normal, cold channel hot spot, hot channel overheating, air conditioner supply air abnormality, cabinet inlet air abnormality, and overall refrigeration deficiency, reflecting the specific abnormal type to which the current temperature state of the machine room belongs. Based on the classification result, the intelligent temperature control technology and air conditioner intelligent or group control scheme can be combined to generate corresponding temperature control instructions, realizing the precise management of the cold channel, hot channel and overall refrigeration of the machine room, thereby optimizing the energy consumption distribution, improving the operation efficiency, and ensuring the safety of the equipment and the stability of the thermal environment in the machine room.

[0083] It can be understood that the input of each spatial temperature field and temperature time trend into the dual-stream attention convolutional network is to fully extract the spatial features of the machine room temperature distribution and its dynamic features over time. By processing the spatial stream and the time stream respectively, the key changes of the temperature state at different grid positions and different time nodes can be captured. Subsequently, the channel-spatial dual attention mechanism is used to weight and fuse each channel and the key spatial position, which can highlight the abnormal features and suppress irrelevant information, thereby realizing the comprehensive analysis and strengthening of spatial and temporal information. Finally, the temperature control instructions are generated according to the classification probability of each type of fault output by the network, which can convert the detected abnormal state into specific control measures in a timely manner, realizing the precise adjustment of the machine room temperature. Thus, the accuracy and response speed of temperature anomaly recognition can be effectively improved, making the control instructions more scientific and reliable, thereby ensuring the safe operation of the machine room equipment and optimizing the overall energy consumption management.

[0084] In order to facilitate the understanding of the scheme of the present application, specific examples are provided below for illustration.

[0085] AsFigure 2 As shown in the figure, a "space-time-feature" three-level fusion framework for monitoring the temperature of the machine room is shown. First, eight temperature parameters including supply air temperature, return air temperature, cold aisle temperature, hot aisle temperature, cabinet inlet air temperature, cabinet outlet air temperature, equipment internal temperature, and machine room environment temperature are obtained through multi-point position temperature collection. Then, the space and time data of each temperature parameter are fused separately. The space data fusion is realized by dividing the machine room space into grids and combining the space distribution and airflow weight to construct the space temperature field. The time data fusion is realized by constructing the fluctuation of the temperature parameter through the sliding window and the error dispersion degree measurement. After the space and time data fusion, the space temperature field and the time trend of the eight temperature parameters are fused to realize the channel-space attention weighted fusion through the double-flow attention convolutional network to improve the feature utilization rate. Finally, if the space comprehensive matching degree meets the requirements, the temperature control instruction is output to support the second-level decision and energy consumption optimization of the intelligent temperature control of the machine room.

[0086] In the above embodiment, by constructing the space temperature field and the temperature time trend in the machine room, the overall temperature distribution of the machine room can be accurately reflected by using the multi-point monitoring data, and the characteristics of the dynamic change of the temperature with time can be captured, thereby avoiding the problem that the local hot spot is hidden by the traditional single-point detection. The system effectively eliminates the data redundancy and local abnormal interference caused by multi-point monitoring through clustering analysis and comprehensive matching degree calculation, and improves the utilization rate of temperature features. At the same time, when the comprehensive matching degree is abnormal, the system can intelligently identify the abnormal temperature control state of the machine room and generate the corresponding control instruction by means of the fusion analysis of the space flow and time flow features through the double-flow attention convolutional network, so as to realize the fine and dynamic temperature management. Therefore, the safety and stability of the operation of the machine room are ensured, and the air conditioning set point is scientifically improved on the premise of ensuring the reliability of the equipment, so as to reduce the refrigeration energy consumption and meet the requirements of PUE optimization, and finally realize the unity of the energy saving and consumption reduction and the stable operation of the machine room.

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

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

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

[0090] The space temperature field construction unit is configured to, for each type of temperature parameter, perform a weighted interpolation calculation based on a distance factor, an air flow weight, and a detected temperature of each grid to obtain a space temperature field of the type of temperature parameter at each grid.

[0091] Specifically, based on the grid layout of the machine room, the spatial distances from the cooling air outlet to each grid position are first calculated one by one to ensure that the positional relationship of different grids in the geometric space can be accurately described. After the distance calculation is completed, inverse distance weighting operation is performed on the distance data, which enables grids closer to the air outlet to be given a higher weight in the cold quantity diffusion model, and grids farther away to be given a lower weight, thereby obtaining a distance factor that reflects the real diffusion law of cold air in space, thereby ensuring that the physical characteristics of the cold air during transmission in the machine room can be reasonably mapped.

[0092] On this basis, the air speed sensor and the air direction sensor arranged in the machine room are further called to collect air flow speed and flow direction data covering each grid position in real time and take them as dynamic input parameters. At the same time, the transmission path of the air flow is matched and calculated in combination with the connection direction between the temperature sensor and the target grid, thereby quantifying the effect of the air flow in different directions on heat transfer. Through this analysis, an air flow weight accurately representing the influence of the air flow on the temperature distribution can be generated, so that the temperature change characteristics of different grids under the action of the air flow can be clearly presented.

[0093] Finally, the distance factor, the air flow weight, and the real-time collected temperature value are subjected to a weighted interpolation calculation. In this process, the spatial geometric relationship and the air flow dynamics characteristics are fused with the measured temperature data, which not only ensures that the model faithfully reflects the physical characteristics of cold air diffusion, but also effectively utilizes the measured results of the temperature sensor. Through fusion calculation, a space temperature field of each type of temperature parameter at each grid in the machine room is finally constructed, thereby realizing fine modeling and comprehensive presentation of the thermal environment distribution in the machine room.

[0094] In this embodiment, the distance from the cooling air outlet to each grid needs to be obtained and inverse distance weighted calculation is performed in order to highlight the dominant role of the near distance on the temperature distribution in the modeling process, avoid the excessive influence of the far distance point on the result, and obtain a more reasonable distance factor. Further, the wind speed and wind direction of each grid are obtained, and the airflow weight is calculated in combination with the connection direction of the temperature sensor to the grid, which can truly reflect the influence degree of the airflow transmission path on the temperature diffusion, so that the modeling result can reflect the aerodynamic characteristics. Finally, the weighted interpolation calculation is performed based on the distance factor, the airflow weight and the actual detected temperature, which can organically integrate the spatial geometric relationship, the airflow distribution law and the measured data, so as to construct a more accurate spatial temperature field. Thus, the deviation caused by single factor modeling can be avoided, the accuracy and reliability of the spatial temperature distribution modeling can be effectively improved, and a more scientific basis for the monitoring and regulation of the machine room thermal environment can be provided.

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

[0096]

[0097]

[0098]

[0099] wherein, represents the spatial temperature field of the i-th grid, represents the distance factor in the i-th grid, represents the airflow weight of the k-th sensor in the i-th grid, represents the detected temperature measured by the k-th sensor in the i-th grid, represents the distance from the cooling air outlet m to the i-th grid, and p is a positive constant, represents the distance from the cooling air outlet m to the j-th grid, represents the wind speed at the spatial position (x, y, z) in the grid, is the connection direction of the k-th sensor to the grid, represents the wind direction in the grid.

[0100] It should be noted that the formula calculates the temperature of each grid as a weighted average of all sensor temperature values falling into the grid, and the weight is determined by two parts, one is the distance factor which decays with geometric distance and is normalized, and the other is the airflow weight which reflects the consistency of local airflow strength and direction, and the angle between the sensor connection direction and the grid wind direction is reflected by the wind speed normalization and the cosine term. Therefore, the formula first limits the influence of remote sensors on the target point by the distance decay mechanism, avoiding the "bias" of interpolation results in sparse areas near the target point by high-density remote sensors; then the contribution of each sensor is adjusted in direction and strength by the airflow related term, which suppresses the undue influence of sensors in the opposite direction on the grid temperature and amplifies the signals of sensors with consistent airflow direction and large wind speed.

[0101] The numerator is the sum of the weighted temperatures, and the denominator is the sum of the corresponding weights, so as to ensure that the output is a normalized local weighted average, which not only preserves the true gradient of the spatial temperature field, but also amplifies the signal of the thermal short circuit or local hot spot without being diluted by the surrounding large number of normal values. Thus, a grid temperature estimation which is robust to uneven sensor distribution, airflow stratification at different heights, and local anomalies can be obtained.

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

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

[0104] A predicted error value calculation unit is configured to determine the real temperature at the corresponding time point of each predicted temperature, and calculate a predicted error value between each predicted temperature and the corresponding real temperature.

[0105] An average error value calculation unit is configured to average each predicted error value of each type of temperature parameter in each grid to obtain an average error value.

[0106] A temperature time trend construction unit is configured to, for each type of temperature parameter, calculate the error dispersion degree according to the average error value and each predicted error value of each grid to obtain the temperature time trend of each grid.

[0107] The predicted error value refers to the difference between the predicted temperature and the corresponding real temperature. The average error value refers to the statistical result obtained by averaging a plurality of predicted error values within a certain time range.

[0108] Specifically, within each grid, the real temperature data acquired by the deployed sensors are first continuously collected, which form a sequence evolving over time in the time dimension. In order to weaken the influence of abnormal fluctuations at a single time point, the sequence is input into a fixed-length sliding window. The sliding window moves step by step on the time axis, covering a continuous segment of temperature data each time, and by averaging all temperatures within the window, short-term fluctuations can be effectively smoothed. Thus, at the end of the sliding window, the predicted temperature corresponding to the next time point is calculated, realizing the prediction function based on short-term historical data. Subsequently, the predicted temperature is compared with the real temperature actually collected at that time point, and the difference between the two is calculated, which is the prediction error value. By continuously calculating the prediction error value 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] Further, the multiple prediction error values generated for each type of temperature parameter within each grid are statistically processed, specifically by averaging all prediction error values to obtain the average error value of the grid. In this way, the influence of a single extreme error point on the overall result can be weakened, making the error evaluation more stable and representative. After obtaining the average error value, the actual distribution of each prediction error value within the grid is combined to calculate its dispersion. Dispersion reflects the volatility of prediction error over time. If the error distribution is more concentrated, it means that the temperature change is more stable. If the dispersion is high, it means that the deviation between temperature prediction and reality fluctuates greatly over time. By analyzing the dispersion, the corresponding temperature time trend can be generated, thereby revealing the change pattern and fluctuation rule of the temperature in different regions of the machine room in the time dimension.

[0110] In this embodiment, by performing sliding window averaging on continuous temperature data and generating predicted temperatures, short-term fluctuations can be smoothed and a quantitative estimate of future temperature changes can be provided, thereby enabling early capture of temperature trends and reducing the interference of incidental noise on judgment. Subsequently, the prediction error is calculated and the average error is further calculated, which helps to stabilize error evaluation, reduce the influence of a single abnormal measurement point on overall trend analysis, and make the temperature change evaluation of each grid more representative and reliable. By constructing a temperature time trend based on the dispersion degree of the prediction error, the fluctuation rule of the temperature in the time dimension can be reflected, the recognition ability of local abnormalities and sudden temperature rise can be enhanced, and physically meaningful time series features can be provided for temperature control decisions, thereby overall improving the accuracy, robustness and prediction guidance of the machine room temperature monitoring.

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

[0112]

[0113]

[0114]

[0115]

[0116] wherein, represents the error dispersion degree, represents the average error value, represents the prediction error value of the kth sensor in the ith grid at the l+n time point, represents the number of prediction error values, represents the predicted temperature of the kth sensor in the ith grid at the l+n time point, represents the true temperature of the kth sensor in the ith grid at the l+n time point, represents the fixed length of the sliding window.

[0117] In this embodiment, the formula is used to generate the predicted temperature by dividing the continuous time series into fixed-length sliding windows, averaging the temperature data in each window, reflecting the short-term historical trend and suppressing the instantaneous fluctuations. Then, the predicted temperature is compared with the true temperature to obtain the prediction error at each time point, and the average value of the prediction error of multiple sliding windows is calculated to obtain the overall deviation level. Finally, the dispersion degree of the error relative to the average error is calculated to quantify the fluctuation of each grid in the time dimension. This method can effectively reveal the stability and fluctuation characteristics of the temperature parameter in continuous time, highlight abnormal temperature rise or sudden changes, smooth daily minor fluctuations, ensure the reliability and robustness of time trend analysis, provide a quantitative basis with clear physical meaning and statistical stability for constructing the time evolution model of each grid temperature, and help subsequent temperature control optimization and abnormal warning.

[0118] In one example, the continuous temperature time series data is divided into fixed-length sliding windows, each window containing n consecutive time point temperature data, and the data at the lth time point can be represented as The n time point temperature data in each sliding window is averaged to obtain the predicted temperature at the next time point at the end of the window. In order to evaluate the smoothing effect of the sliding window, the predicted temperature at the l+n time point is compared with the true temperature to calculate the prediction error. As Figure 3 shown, for the prediction values of s sliding windows, the overall deviation of each grid in the time dimension can be quantified by averaging the prediction error. On this basis, the dispersion degree of the error relative to the average error is calculated to measure the fluctuation of the temperature of each grid in the time dimension.

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

[0120] a fuzzy membership degree calculation unit configured to calculate fuzzy membership degrees between the spatial temperature field and the temperature time trend of each grid and each cluster center;

[0121] a total fuzzy membership degree calculation unit configured to determine a mode interval to which each cluster center belongs, and to accumulate the fuzzy membership degrees corresponding to each cluster center in each mode interval to obtain a total fuzzy membership degree of each mode interval;

[0122] a comprehensive matching degree calculation unit configured to obtain a comprehensive matching degree by weighted summation and normalization according to the total fuzzy membership degrees of each mode interval and weights.

[0123] The fuzzy membership degree refers to the similarity between the spatial temperature field and the temperature time trend of each grid and each cluster center, and is used to quantify the attribution possibility of single grid data under different cluster categories. The total fuzzy membership degree is a comprehensive value obtained by accumulating the fuzzy membership degrees corresponding to all cluster centers in the same mode interval, and is used to reflect the overall coverage and representativeness of the mode interval to the grid temperature state.

[0124] Specifically, the fuzzy membership degree calculation can be realized by comparing the spatial temperature features and time sequence features of each grid with all cluster centers one by one. The comparison method can adopt a continuous membership function, such as a Gaussian membership function or an exponential function. By calculating the difference between the grid features and the cluster center features, the result is mapped to a continuous membership value, thereby reflecting the relative attribution possibility of each grid data under different cluster centers. This operation can ensure flexible mapping of grid data under multiple modes, so that the local temperature features can not only reflect the most matched cluster center, but also reflect the partial correlation under other cluster centers.

[0125] Subsequently, by determining the mode interval to which each cluster center belongs, the total fuzzy membership degree of the mode interval is obtained by accumulating the fuzzy membership degrees corresponding to all cluster centers falling into the same mode interval. In actual operation, the mode intervals can be divided according to the feature vectors or mode attributes of the cluster centers, and then the membership degrees in the intervals are accumulated. This step can integrate the local information of a single cluster center into a whole index at the mode interval level, weaken the influence of abnormal values or extreme values, and ensure that the overall representativeness of the mode interval to the grid temperature state is more accurate.

[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] By this, the value of the comprehensive matching degree is limited between 0 and 1, and the closer the value is to 1, the higher the fit degree of the current data with the normal mode; on the contrary, when it is lower than a certain threshold (for example, 0.8), it means that the current state may be abnormal, and the corresponding alarm mechanism can be triggered, so as to realize the matching quantification of the temperature parameter in the space and time dimensions and the abnormal monitoring.

[0136] It can be understood that, through the continuous calculation of the fuzzy membership degree, the flexible attribution of the grid data between multiple clustering centers is realized, the information loss is minimized, and the local details are preserved; secondly, the accumulation of the total fuzzy membership degree can integrate the clustering center information, reduce the interference of single-point abnormality on the mode judgment, and improve the representativeness of the mode interval; finally, the comprehensive matching degree obtained by weighted summation and normalization can quantify the consistency of the entire machine room temperature state and the typical mode, and provide a scientific, robust and physically meaningful decision basis for temperature control optimization, abnormal identification and operation monitoring.

[0137] In one embodiment, the dual-flow attention convolutional network comprises:

[0138] An input and feature extraction layer is configured to extract spatial flow features from each spatial temperature field and extract time flow features from each temperature time trend, and align the time flow features and the spatial flow features in the spatial dimension;

[0139] A dual-attention enhanced and feature fusion layer is configured to determine spatial channel attention, time channel attention and spatial attention map according to the spatial flow features and the time flow features, and generate a fused feature map;

[0140] An output classification layer is configured to output classification probabilities of each category according to the fused feature map.

[0141] In the embodiment, the input and feature extraction layer is configured to extract spatial flow features from the spatial temperature field of the machine room, extract time flow features from each temperature time trend, and align the extracted time flow features and the spatial flow features in the spatial dimension, so as to ensure that the time features of each spatial unit can accurately correspond to the spatial position features corresponding thereto. The spatial flow features represent the local mode and gradient information of temperature change in the three-dimensional grid space, and the time flow features represent the evolution law and dynamic trend of temperature in the time dimension. The alignment operation guarantees the semantic consistency of the features in different dimensions. In implementation, the spatial flow features can be obtained by convolution operation or other spatial feature extraction operators, the time flow features can be obtained by processing the temperature time series through a recurrent neural network or a time convolution network, and the spatial dimension alignment can be completed through index mapping or tensor rearrangement operation.

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

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

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

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

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

[0147] In this embodiment, the spatial flow feature extraction unit is configured to convert each spatial temperature field in the data center into a spatial flow tensor to represent the distribution and change information of the temperature in the three-dimensional grid, and the spatial flow features represent the local spatial temperature patterns and gradient information extracted by the two-dimensional convolution operation, which can reflect the temperature differences between different grids and their spatial correlation. In the implementation, each spatial temperature field can be arranged according to the grid index to form a three-dimensional tensor, and then a two-dimensional convolution kernel is used to perform sliding convolution operation on the tensor to extract a feature map with a spatial local receptive field, thereby realizing efficient coding of the temperature distribution pattern.

[0148] The time flow feature extraction unit is configured to convert each temperature time trend into a time flow tensor to represent the temperature evolution rule of each grid in the time dimension, and the time flow feature is extracted from the time flow tensor by a one-dimensional convolution operation to reflect the fluctuation and trend of the temperature over time. In the implementation process, the continuous temperature time series of each grid can be arranged as a one-dimensional tensor, and the local feature response is calculated by sliding the one-dimensional convolution kernel on the time axis, so as to obtain a time feature vector that can capture short-term changes and long-term trends.

[0149] The operations of extracting the spatial flow and the time flow respectively are performed by convolution operations, which can realize efficient feature coding while preserving local information and continuity features. Such operations can not only reveal the local patterns and gradient differences of the spatial temperature distribution, but also quantify the dynamic changes of the temperature time evolution, providing a reliable basis for subsequent feature fusion and anomaly detection. At the same time, the weight sharing mechanism of convolution calculation improves the calculation efficiency and robustness of feature extraction, so that the spatial and temporal features have both accuracy and stability in the modeling process, providing scientific and quantifiable basis for data center temperature control optimization and state recognition.

[0150] In one example, the temperature field data is represented by a tensor. Specifically, 8 temperature parameters are discretized according to a 1 meter x 1 meter grid to form a grid map with a spatial resolution of H x W, thereby obtaining a spatial flow input:

[0151]

[0152] wherein, is equal to 8, represents the mean distribution of the temperature field at each spatial position, and the spatial temperature field mean is obtained by averaging the data along the z-axis direction in the three-dimensional space. The time fluctuation is represented by a tensor as follows:

[0153]

[0154] Each channel corresponds to the data of S sliding windows of one temperature parameter, which is used to reflect the dynamic fluctuation characteristics of the temperature. The spatial flow coding is realized by a two-dimensional convolution operation, and the coding result is:

[0155]

[0156] The local structural features of the spatial temperature field can be extracted. The time flow coding is realized by a one-dimensional convolution operation, which captures the time evolution pattern of each temperature parameter, and the coding result is:

[0157]

[0158] To realize the cross-modal feature fusion, the time flow feature is remapped to the spatial dimension, and is obtained by repeated expansion along the time axis and reshaping:

[0159]

[0160] Thus, it is aligned with in the spatial dimension, providing a basis for the subsequent fusion of spatial and temporal features.

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

[0162] A channel weight determination unit is configured to perform global average pooling and two-layer bottleneck fully connected on the spatial flow feature and the time flow feature respectively to obtain spatial channel attention and time channel attention.

[0163] A spatial attention map determination unit is configured to concatenate the spatial flow feature and the time flow feature to obtain a spatial attention map.

[0164] An attention weighting unit is configured to weight the spatial flow feature using the spatial channel attention first and then using the spatial attention map to obtain a weighted spatial feature, and weight the time flow feature using the time channel attention first and then using the spatial attention map to obtain a weighted time feature.

[0165] A weighted fusion unit is configured to fuse the weighted spatial feature and the weighted time feature pixel by pixel according to the spatial position to obtain a fusion feature map.

[0166] In the specific implementation process, the channel weight determination unit obtains the global statistical information of each channel by performing global average pooling on the spatial flow feature and the time flow feature respectively, and then generates spatial channel attention and time channel attention through two-layer bottleneck fully connected operation, and then quantifies the importance of each channel.

[0167] The spatial attention map determination unit concatenates the spatial flow feature and the time flow feature in the channel dimension, and generates a spatial attention map through convolution or normalization operation, so as to reflect the feature contribution degree of different positions in the spatial dimension.

[0168] The attention weighting unit first weights the spatial flow feature using the spatial channel attention, and then weights it using the spatial attention map to obtain the weighted spatial feature; at the same time, the time flow feature is weighted using the time channel attention first, and then using the same spatial attention map to obtain the weighted time feature.

[0169] Finally, the weighted fusion unit fuses the weighted spatial feature and the weighted time feature pixel by pixel according to the spatial position to form a fusion feature map, which provides a comprehensive feature representation for the subsequent classification or prediction task.

[0170] In one example, to enhance the important channel responses of spatial stream features and temporal stream features, first, a channel attention module sharing weights is applied to the two streams. Global average pooling is performed on each stream to obtain a global statistic vector of the spatial stream:

[0171]

[0172] A global statistic vector of the temporal stream:

[0173]

[0174] Then, through two layers of bottleneck fully connected operations, a compression ratio r (usually taken as 16) is adopted, and a ReLU activation and a Sigmoid function σ are performed to generate a channel weight vector of the spatial stream and the temporal stream:

[0175]

[0176]

[0177] wherein, .

[0178] Output and are used to adaptively weight the respective channels to highlight key channel information. Then, the features of the two streams are concatenated in the channel dimension:

[0179]

[0180] And a 1x1 convolution and a Sigmoid activation function are used to generate a single-channel spatial attention map:

[0181]

[0182] The spatial attention map reflects the importance of each spatial position, thereby providing a basis for subsequent weighting. Then, a "channel first, spatial second" order attention weighting strategy is adopted to weight the spatial stream features and the temporal stream features respectively:

[0183]

[0184]

[0185] wherein, ⊙ represents element-wise multiplication (Hadamard product), which realizes joint weighting of channels and spaces. Finally, the weighted spatial stream features and the weighted temporal stream features are fused pixel by pixel:

[0186]

[0187] The fusion feature map is formed, so as to realize fine-grained complementary fusion of double-flow features, retain spatial local structure information, and take into account time evolution mode, so that the feature expression is more discriminative and robust.

[0188] Thus, the expression capability of important information of spatial flow features and time flow features can be effectively improved, while redundant and irrelevant features are inhibited, and the selectivity and expression of features in the spatial and time dimensions are enhanced. The channel attention can highlight key channel information, and the spatial attention map can highlight important spatial positions, so that more discriminative feature representation is formed after weighted fusion, the recognition capability and robustness of the model to temperature field and time trend mode are improved, and a reliable feature basis is provided for subsequent classification and anomaly detection.

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

[0190] An average pooling result determination unit is configured to perform average pooling on the fusion feature map to obtain an average pooling result.

[0191] A classification probability output unit is configured to output classification probabilities of each class according to the average pooling result.

[0192] In this embodiment, the average pooling result determination unit is configured to perform global average pooling on the fusion feature map, which is a multi-channel and high-dimensional feature representation obtained by weighted fusion of the spatial flow feature and the time flow feature. The fusion feature map contains important information of each spatial position and time evolution mode, and the average pooling operation can compress the feature dimension by performing global average on the feature map in the spatial dimension, to generate an average pooling result vector that can comprehensively reflect the distribution of the entire input data.

[0193] The classification probability output unit maps the average pooling result to the prediction probability distribution of each class through a fully connected layer, and can normalize the output to between 0 and 1 through a Softmax activation function to obtain the classification probability of each class. This operation can convert the high-dimensional fusion feature vector into classification probability that is easy to interpret and apply, thereby providing a basis for subsequent decision-making.

[0194] By performing average pooling first and then outputting classification probability in sequence, the influence of feature dimension on calculation amount can be effectively reduced, while the global discriminative information of the fusion feature is maintained, so that the prediction of each class is more robust and robust. In addition, this method can highlight the average response of global features and inhibit the interference of local noise on classification, thereby improving the accuracy and reliability of the classification result and ensuring that the discriminative ability of the model to the temperature change mode in the spatial and time dimensions is fully utilized.

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

[0196]

[0197] Subsequently, 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] wherein, is the weight matrix of the classifier, which is initially determined by random assignment, and its value is adjusted according to the input data during the training process until the model converges or reaches a predetermined number of iterations, thereby achieving accurate prediction of each class.

[0200] In one embodiment, the present application also provides a computer device having computer readable instructions stored therein, which, under the control of one or more processors, enable the computer device to implement the various functional modules of the artificial intelligence-based computer room multi-point monitoring environment temperature air conditioning energy-saving system as described in any of the above embodiments.

[0201] As shown schematically in Figure 5 , Figure 5 is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 200 can be provided as a server. Referring 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 further including one or more processors for invoking instructions in the memory 201 to implement various functional modules of the system.

[0202] In addition, the computer device 200 can also include a power supply component 203 for managing the power supply of the device, 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 ServerTM, Mac OSXTM, UnixTM, LinuxTM or Free BSDTM, etc., to provide support for the collaborative operation of various modules of the system.

[0203] Those skilled in the art can understand that, Figure 5The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components, to realize the complete operation of the functional modules of the system.

[0204] Finally, it should be noted that in this document, the 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. Also, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of additional identical elements in the process, method, article or device that includes the element. In this document, "a", "an", "said", "the" and "it" can also include plural forms, unless the context clearly indicates otherwise. A plurality means 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 the specification are described in a progressive manner, each embodiment focusing on the differences from other embodiments, and the various embodiments can be combined as needed, and the same and similar parts refer to each other.

[0206] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An artificial intelligence-based machine room multi-point monitoring ambient temperature air conditioning energy-saving system, characterized in that, The system comprises: a spatial temperature field construction module, after the grid is constructed, for each type of temperature parameter, according to its corresponding detection temperature, distance factor and air flow weight in each grid, the spatial temperature field of each grid is constructed, the distance factor is determined by the distance between the grid and the cooling air outlet, and the air flow weight is determined by the wind speed and wind direction of the grid; a temperature time trend construction module, for each type of temperature parameter, according to its continuous real temperature in each grid, a predicted temperature is generated by a sliding window, and based on each predicted temperature and the corresponding real temperature, the temperature time trend of each grid is captured; a comprehensive matching degree calculation module, for each type of temperature parameter, the spatial temperature field and the temperature time trend are clustered to generate a plurality of cluster centers, and the comprehensive matching degree is calculated in combination with the mode interval to which each cluster center belongs and the weight thereof; a temperature regulation instruction generation module, when the number of comprehensive matching degrees that do not meet the normal mode reaches a preset number, each spatial temperature field and each temperature time trend are input into a double-flow attention convolutional network, and according to the classification probability of each type of fault output by the double-flow attention convolutional network, a temperature regulation instruction of the machine room is generated, and the double-flow attention convolutional network is used to extract the spatial flow feature of the spatial temperature field and the time flow feature of the temperature time trend, and then weighted fusion and classification are performed by using channel-space double attention.

2. The artificial intelligence-based machine room multi-point monitoring ambient temperature air conditioning energy-saving system according to claim 1, characterized in that, The spatial temperature field construction module comprises: a distance factor calculation unit, for each type of temperature parameter, the distance from the cooling air outlet to each grid is obtained, and inverse distance weighted calculation is performed on each distance to obtain the distance factor of the temperature parameter in each grid; an air flow weight calculation unit, for each type of temperature parameter, the wind speed and wind direction in each grid and the connection direction of the sensor corresponding to the temperature parameter to each grid are obtained, and the air flow weight of each grid is calculated; a spatial temperature field construction unit, for each type of temperature parameter, based on the distance factor, air flow weight and detection temperature of each grid, weighted interpolation calculation is performed to obtain the spatial temperature field of the temperature parameter in each grid. 3.The AI-based multi-point monitoring environment temperature air conditioning energy-saving system for a computer room according to claim 2, characterized in that, The calculation formula of the spatial temperature field of each grid is: wherein, represents the spatial temperature field of the i-th grid, represents the distance factor in the i-th grid, represents the air flow weight of the k-th sensor within the i-th grid, represents the detected temperature measured by the k-th sensor within the i-th grid, represents the distance from the cooling air outlet m to the i-th grid, p is a positive constant, represents the distance from the cooling air outlet m to the j-th grid, represents the air speed at the spatial position (x, y, z) within the grid, is the line direction of the k-th sensor to the grid, represents the air direction within the grid.

4. The artificial intelligence-based machine room multi-point monitoring ambient temperature air conditioning energy-saving system according to claim 1, characterized in that, The temperature time trend construction module comprises: a predicted temperature generation unit, for each type of temperature parameter, the continuous real temperature in each grid is obtained, and a fixed-length sliding window is used to average the real temperature at the continuous time points to obtain the predicted temperature corresponding to the next time point at the end of the sliding window; a predicted error value calculation unit, for determining the real temperature at the corresponding time point of each predicted temperature, and calculating the predicted error value between each predicted temperature and the corresponding real temperature; an average error value calculation unit, for each type of temperature parameter, each predicted error value in each grid is averaged to obtain an average error value; The temperature time trend construction unit is configured to calculate a degree of error dispersion of each grid according to the average error value and the respective prediction error value of each grid, and obtain a temperature time trend of each grid for each type of temperature parameter.

5. The artificial intelligence-based machine room multi-point monitoring ambient temperature air conditioning energy-saving system according to claim 4, characterized in that, The calculation formula of the degree of error dispersion is as follows: wherein, denotes the degree of dispersion of the errors, denotes the average error value, denotes the prediction error value of the kth sensor in the ith grid at the l+n time point, denotes the number of prediction error values, denotes the predicted temperature of the kth sensor in the ith grid at the l+n time point, denotes the real temperature of the kth sensor in the ith grid at the l+n time point, denotes the fixed length of the sliding window. 6.The artificial intelligence based multi-point monitoring environment temperature air conditioning energy saving system for a computer room of claim 1, wherein, The comprehensive matching degree calculation module comprises: The fuzzy membership calculation unit is configured to calculate fuzzy membership between the spatial temperature field and the temperature time trend of each grid and each cluster center. The total fuzzy membership calculation unit is configured to determine a mode interval to which each cluster center belongs, and accumulate the fuzzy membership corresponding to each cluster center in each mode interval to obtain a total fuzzy membership of each mode interval. The comprehensive matching degree calculation unit is configured to obtain the comprehensive matching degree by weighted summation and normalization according to the total fuzzy membership and the weight of each mode interval.

7. The artificial intelligence-based machine room multi-point monitoring ambient temperature air conditioning energy-saving system according to claim 1, characterized in that, The double-flow attention convolutional network comprises: The input and feature extraction layer is configured to extract spatial flow features according to each spatial temperature field, and extract time flow features according to each temperature time trend, and align the time flow features and the spatial flow features in the spatial dimension. The double-attention enhancement and feature fusion layer is configured to determine spatial channel attention, time channel attention and a spatial attention map according to the spatial flow features and the time flow features, and generate a fusion feature map. The output classification layer is configured to output classification probabilities of each category according to the fusion feature map.

8. The artificial intelligence-based machine room multi-point monitoring ambient temperature air conditioning energy-saving system according to claim 7, characterized in that, The input and feature extraction layer comprises: The spatial flow feature extraction unit is configured to convert each spatial temperature field into a spatial flow tensor, and extract the spatial flow features from the spatial flow tensor by using a two-dimensional convolution method. The time flow feature extraction unit is configured to convert each temperature time trend into a time flow tensor, and extract the time flow features from the time flow tensor by using a one-dimensional convolution method. 9.The artificial intelligence based multi-point monitoring environment temperature air conditioning energy saving system for a computer room of claim 7, wherein, The double-attention enhancement and feature fusion layer comprises: The channel weight determination unit is configured to perform global average pooling and two-layer bottleneck full connection on the spatial flow features and the time flow features respectively, to obtain the spatial channel attention and the time channel attention. The spatial attention map determination unit is configured to splice the spatial flow features and the time flow features to obtain the spatial attention map. The attention weighting unit is configured to weight the spatial flow features by using the spatial channel attention first and then by using the spatial attention map, to obtain weighted spatial features, and weight the time flow features by using the time channel attention first and then by using the spatial attention map, to obtain weighted time features. The weighted fusion unit is configured to fuse the weighted spatial features and the weighted time features according to spatial positions pixel by pixel to obtain a fusion feature map.

10. The artificial intelligence-based machine room multi-point monitoring ambient temperature air conditioning energy-saving system according to claim 7, characterized in that, The output classification layer comprises: The average pooling result determination unit is configured to perform average pooling on the fusion feature map to obtain an average pooling result. The classification probability output unit is configured to output classification probabilities of each category according to the average pooling result.

Images