Intelligent adjustment method for maintaining positive pressure of large data center machine room based on deep learning

Abstract

The invention discloses a deep learning-based intelligent adjustment method for maintaining positive pressure of a data center machine room. The method comprises the steps of obtaining a training set F; performing data cleaning on the F, supplementing missing values and processing abnormal values, and performing discrete processing on continuous variable values to form a derivative feature set F1; performing feature data extraction on the F1, and extracting a feature data set S and a test data set C; processing the S by adopting an entropy and information gain ratio algorithm to form a decision tree data set TREE; performing post pruning on the TREE, deleting branches which contribute little to the generalization ability of the model, and forming a training set D; constructing a decision tree model by using the data of the D, and adjusting model parameters in the training process; and for the obtained decision result, evaluating the accuracy and the flooding ability of the decision tree model by using the data in the C. According to the invention, indoor positive pressure maintenance of the data center and deep learning and intelligent adaptive adjustment of different operation environments can be realized, and each functional room of the data center is maintained to be in positive pressure operation.

Description

Technical Field

[0001] The present invention relates to the field of computer science and technology, and in particular to an intelligent adjustment method for maintaining positive pressure in a computer room of a large data center based on deep learning. Technical Background

[0002] Dust management is crucial for large data center rooms, as it is directly related to the safety, economy, and compliance of equipment operation. Dust, as a hidden threat, can penetrate the protective layer of electronic equipment and deposit on circuit boards, causing short circuits, overheating, and even equipment failures, threatening the stability and safety of data centers. Maintaining positive pressure in the data center room can not only effectively curb dust accumulation and extend the service life of core equipment such as servers and switches, but also keep the fans or heat sinks of cooling equipment clean, avoid a decrease in cooling efficiency, maintain energy efficiency, and help keep the data center power efficiency ratio (PUE) at a normal level, promoting energy conservation and emission reduction.

[0003] As the scale of a single data center continues to expand, data center buildings are modularized in the park, the area of ​​the computer room is growing in an open manner, the functions of each room are more independent, and the cleaning requirements are also changing. The mainstream industrial fresh air equipment on the market that is suitable for dust filtration and air renewal in a single room can no longer meet the needs of maintaining indoor positive pressure and keeping indoor cleanliness in large-scale clustered data center computer rooms. There is an urgent need for a deep learning, intelligently adjustable, and sustainable and autonomous adjustment method to maintain indoor positive pressure in data center computer rooms for one or several buildings, one or several floors. Summary of the invention

[0004] In response to the problems existing in the prior art, the present invention provides an intelligent adjustment method for maintaining positive pressure in a large data center computer room based on deep learning, which can achieve deep learning and intelligent adaptive adjustment of the indoor positive pressure of the data center and different operating environments, and maintain positive pressure operation in various functional rooms of the data center.

[0005] A method for maintaining positive pressure in a data center room based on deep learning, comprising the following steps:

[0006] Step 1: Data collection and extraction: Obtain training set F: Collect various relevant data inside and outside the room, including indoor pressure P, for each room. in , Outdoor pressure P out 、Inlet and outlet ventilation damper opening θ in and θ ex , the fan speed V of the fresh air system, the time interval T for opening and closing the fresh air equipment, and the various types of collected data are preliminarily screened based on the time interval T for opening and closing the fresh air equipment, and the maximum and minimum values ​​are filtered out to form a training set F;

[0007] Step 2: Data preprocessing: clean the data in the training set F, supplement missing values ​​and process outliers, discretize the continuous variable values, and form a derived feature set F1 based on the influence relationship between the opening and closing time interval T of the fresh air equipment and the indoor and outdoor pressure difference P;

[0008] Step 3: Extract feature data set S and test data set C: Extract feature data from the derived feature set F1, and extract the indoor and outdoor pressure difference P, the time interval between opening and closing of the fresh air equipment T, the fan speed V, and the ventilation inlet and outlet opening θ. in and θ ex、 Generate feature data set S for each time period as main feature data i = {P i , T i , V i ,θ in ,θ ex}, then extract the feature data set S for each time period i , generate key feature data set S = {S1+S2+S3+···+S n}; Then, a standard sample data is extracted by human intervention and mathematical methods to form the relevant data of the test data set C, C = {δP1+δP2+δP3+···+δP n}, where i = 0, 1, 2, ... n;

[0009] Step 4: Form a decision tree data set TREE: Use entropy and information gain ratio algorithms to process the feature data set S, select the best features and splitting points, split the data based on the selected best features and splitting points to create the root node of the tree, and recursively perform the same process on each child node until the stopping condition is reached to form a decision tree data set TREE. The stopping condition is that all samples in the node belong to the same class, or the maximum tree depth is reached, or the number of samples in the node is less than a preset threshold.

[0010] Step 5: Decision tree pruning: Post-prune the decision tree data set TREE, delete the branches that contribute less to the generalization ability of the model from the bottom up, and form the training set D;

[0011] Step 6: Decision tree model construction and training: Use the data of training set D to build a decision tree model, and adjust the model parameters during the training process. The model parameters include the maximum depth of the tree and the minimum number of leaf node samples;

[0012] Step 7: Decision tree model evaluation: Use the data in the test data set C to evaluate the accuracy and generalization ability of the decision tree model. For inappropriate decision tree models, re-discrete the collected data and repeat the first six steps until the decision tree model evaluation results meet the test requirements.

[0013] Furthermore, in the first step, a full range of sensors are deployed to collect various relevant data inside and outside the room.

[0014] Furthermore, the missing values ​​in the second step are specifically supplemented by the missing data of the fan speed V and the damper closing status of the fresh air system where some fans are not started, and the inlet and outlet ventilation damper opening θ in and θ ex Perform zero filling processing to change the indoor pressure P of some points with negative pressure difference to in , Outdoor pressure P out Modify to 0;

[0015] The specific processing of abnormal values ​​is as follows: identifying and eliminating outliers through Z-score or IQR method;

[0016] Discrete processing is specifically: for the regular pressure data P, T, V and θ within a period of time in and θ ex , use discrete mathematical models for discrete processing to improve data generalization capabilities.

[0017] Furthermore, the decision tree model constructed in the sixth step is used to identify abnormal data and misjudged data in the training set D classification, wherein the specific method of identifying abnormal data and misjudged data is as follows: 1) Identify the inlet and outlet ventilation damper opening θ in the training data in and θ ex Whether it is consistent with the actual opening percentage, and eliminate inappropriate data; 2) Identify whether the fan speed V of the fresh air system of some modules in the training data exceeds the actual operating speed range, and eliminate inappropriate data.

[0018] Furthermore, in the seventh step, the accuracy and generalization ability of the decision tree model are evaluated using the data in the test set C, and the confusion matrix, ROC curve and AUC value indicators are used to evaluate the model performance. The standard for the decision tree model evaluation results to meet the test requirements is to achieve an accuracy of 98% and an AUC value of 0.95.

[0019] The present invention utilizes deep learning and remote intelligent control technology to collect operating environment, evaluate operating quality, conduct deep learning of historical data, and conduct autonomous adaptive adjustment of one or more data center computer rooms to achieve continuous maintenance of positive pressure in the data center and continuous cleaning of air dust, providing a deep learning-based intelligent adjustment method for maintaining positive pressure in large data center computer rooms for large data centers. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of a data management method for maintaining positive pressure inside a data center based on a decision tree model of the present invention;

[0021] Figure 2It is a planar logic diagram of the positive pressure maintenance system of the data room between floors of a large data center. Specific embodiments

[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] See also Figure 1 The embodiment of the present invention provides a method for maintaining a positive pressure intelligent adjustment in a data center computer room based on deep learning, comprising the following steps:

[0024] Step 1: Data collection and extraction

[0025] Data collection: Figure 2 As shown, high-precision sensors are installed in each room to collect data in real time;

[0026] Indoor pressure P in (Unit: Pa)

[0027] Outdoor pressure P out (Unit: Pa)

[0028] Ventilation opening θ in and θ ex (Unit: Percentage, 0-100%)

[0029] Fresh air system fan speed V (unit: m / s)

[0030] Power consumption E (unit: kW·h)

[0031] Fresh air equipment on and off time interval T (unit: minutes)

[0032] Preliminary screening: Perform preliminary screening on the collected data with a time interval of T = 5 minutes as the dimension, filter out the maximum and minimum values, and form a training set F; use the sliding window method (for example, 3 consecutive time points) to smooth the data and reduce the impact of noise.

[0033] Specific example: Data center A has 10 computer rooms, each of which collects data every 5 minutes, for a total of 7 days of data. After preliminary screening, a training set F containing about 20,160 records is formed.

[0034] Step 2: Data Preprocessing

[0035] Cleaning data: 1) Supplement missing values, fill in the missing data V of some fans not started and the missing data θ of the damper closed status with 0, and modify the P value of some points with negative pressure difference to 0; 2) Process outliers, identify and eliminate outliers through Z-score or IQR method.

[0036] Discretization processing: Discretize the continuous variable values ​​in the data P, T, V, E and θ, for example, divide the indoor pressure Pin into three levels: low, medium and high.

[0037] Derived feature generation: According to the opening and closing time interval T of the fresh air equipment and the indoor and outdoor pressure difference P = P in -P out The influence relationship forms the derived feature set F1;

[0038] Specific examples:

[0039] During the preprocessing of the data collected by Data Center A, it was found that the power consumption E in some time periods had obvious outliers. These outliers were eliminated using the IQR method and the missing values ​​were supplemented. The preprocessed data set F1 contains about 20,160 records for subsequent analysis.

[0040] Step 3: Extract feature data set S and test data set C

[0041] Feature data extraction: Extract the main feature data: indoor and outdoor pressure difference P, fresh air equipment opening and closing time interval T, fan speed V and ventilation inlet and outlet opening θ in and θ ex , generate the feature data set Si for each time period = {P i ,T i ,V i ,θ in ,θ ex}.

[0042] Generate key feature data set S: Merge the feature data sets Si of each time period to form a key feature data set S = {S1+S2+S3+...+S n}.

[0043] Generate test data set C: Use manual intervention and mathematical methods to calculate the pressure drop rate δp of each room, and arrange them in ascending order according to the δp value, extract a standard sample data, and form a test set: C = {δP1+δP2+δP3+...+δP n}, where i = 0, 1, 2...n.

[0044] The specific algorithm of the method of manual intervention of test data set C is as follows:

[0045] ① Collect sensor data from each room, including fresh air fan speed V, inlet and outlet opening θ in and θ ex , indoor and outdoor pressure difference P.

[0046] ②Record the changes in these data and mark the corresponding pressure drop rate δp.

[0047] A mathematical model based on the above formula is established to predict the pressure drop rate δp of each room. The specific formula is as follows:

[0048] δp=f(V,θ in ,θ ex ,P)

[0049] Here, f is a function, which can be customized according to the actual application scenario. A simple typical model example used in the present invention is:

[0050] δp=C V *V+C in *θ in +C ex *θ ex +C P *P

[0051] in:

[0052] C V is a coefficient related to fan speed;

[0053] C in is the coefficient related to the opening of the fresh air inlet damper;

[0054] C ex is the coefficient related to the opening of the fresh air outlet damper;

[0055] C P is the coefficient related to the indoor and outdoor pressure difference;

[0056] Assume that the following fitting coefficients are obtained by using the linear regression algorithm to perform statistics on the collected data:

[0057] C V =0.5, C in =0.3, C ex =0.4, C P =0.2

[0058] So, for a room, if it is measured that:

[0059] Fresh air fan speed V = 10m / s

[0060] Air door opening C in =0.8, C ex =1

[0061] Indoor and outdoor pressure difference P = 5Pa

[0062] Then the pressure drop rate δp in the room during the time T can be calculated as:

[0063] δp=0.5*10+0.3*0.8+0.4*1+0.2*5=5+0.24+0.4+1=6.64Pa / s

[0064] Using the collected data, the changing pattern of δp under different conditions is analyzed.

[0065] From this, we can calculate the δP data set of each room within a continuous span of several days T involved in the sample standard data, which is used as the test data set C.

[0066] C={δP1+δP2+δP3+···+δP n}, where i = 0, 1, 2, ... n;

[0067] Specific examples:

[0068] Data Center A selected 7 days of data as the training set and randomly selected 1,440 records (i.e., 1 day of data) for calculation to form the test set C for evaluating the performance of the model.

[0069] The feature data set S contains the data for the remaining 6 days, totaling 18,720 records.

[0070] Step 4: Form the decision tree data set TREE

[0071] The entropy value and information gain ratio algorithm is used to process the feature data set S, select the best features and splitting points, and create the root node of the tree based on the selected best features and splitting points. The same process is recursively performed on each child node until the stopping condition is reached to form a decision tree data set TREE;

[0072] (1) Entropy algorithm

[0073] The entropy value is used to measure the purity of the data set S. The higher the entropy, the greater the uncertainty of the data set; conversely, the smaller the uncertainty, the algorithm is:

[0074]

[0075] where p i It indicates the proportion of samples of the i-th category containing feature A to the total samples.

[0076] (2) Information Gain Algorithm

[0077] Information gain IG(A) refers to the amount of information gain obtained after partitioning the data set S using feature A in the set S. It is used to measure the reduction in uncertainty brought about by an attribute segmentation data set. The algorithm is:

[0078]

[0079] in:

[0080] Entropy(S) represents the entropy of data set S.

[0081] Sv represents the data subset whose value is v on attribute A.

[0082] |S| and |Sv| represent the number of samples in the dataset D and subset Dv, respectively.

[0083] Values(A) represents the set of all possible values ​​of attribute A.

[0084] (3) Information Gain Ratio Algorithm

[0085] The information gain ratio algorithm is based on information gain and constrains attributes with many different values ​​by dividing by the intrinsic value of the attribute itself, thereby better reflecting the usefulness of the attribute. The specific algorithm is:

[0086] Define IV(A) as the intrinsic value of the sample, then:

[0087]

[0088] Information gain ratio Gain_Ratio(A), its algorithm formula is:

[0089]

[0090] IV(A) represents the sample intrinsic value of feature A in set S

[0091] IG(A) means calculating the information gain of set S based on feature A.

[0092] Specific examples:

[0093] Based on the 18,720 records in the feature data set S, the entropy value and information gain ratio algorithm were used to select the best features and splitting points, the decision tree root node was constructed, and the data was recursively split until the stopping condition was reached (all samples in the node belonged to the same class, or the maximum tree depth was reached, or the number of samples in the node was less than the preset threshold). The final decision tree data set TREE contained 15 nodes with a maximum depth of 5.

[0094] Step 5: prune the decision tree to form the training set D

[0095] The cost complexity pruning algorithm is adopted to perform post - pruning on the data set D, deleting the branches that contribute less to the model generalization ability from bottom to top to form the training set D;

[0096] The specific steps of the cost complexity pruning algorithm are as follows:

[0097] 1) Initialization: Select a validation set and recursively traverse the entire decision tree starting from the root node;

[0098] 2) Calculate the error: For each internal node t, calculate its error err(Tt) when not pruned and the error err(Tt′) when it is converted into a leaf node;

[0099] 3) Comparison and pruning: For each internal node t, calculate the cost complexity Rα(Tt) and Rα(Tt′) before and after pruning. If Rα(Tt′) < Rα(Tt), then prune this node, that is, convert it into a leaf node;

[0100] 4) Recursion: For the pruned tree, continue to recursively execute the above steps until no further pruning is possible;

[0101] 5) Select the optimal α: By evaluating the performance of the decision tree under different α values on the validation set, select an optimal α value to make the performance of the model optimal on the validation set.

[0102] Specific case:

[0103] Post - pruning is performed on the decision tree, deleting the less - contributing branches, and finally forming the simplified training set D, reducing the risk of overfitting. The simplified training set D contains 14,526 records for subsequent model training.

[0104] Step 6: Decision tree model construction and training

[0105] Use the data in the training set D to construct a decision tree model, and adjust the model parameters (such as the maximum depth of the tree, the minimum number of samples in a leaf node) to optimize the model performance. Use grid search and cross - validation methods to optimize the hyperparameters.

[0106] Specific case:

[0107] Data center A uses the training set D to construct a decision tree model, and optimizes the model parameters through grid search and cross - validation. Finally, the maximum depth of the model is set to 6, and the minimum number of samples in a leaf node is set to 10, achieving good generalization performance. The training set D contains 14,526 records, ensuring that the model has sufficient data support

[0108] Step 7: Decision tree model evaluation

[0109] Model evaluation: Use the data in the test set C to evaluate the accuracy and generalization ability of the decision tree model. Use indicators such as confusion matrix, ROC curve and AUC value to evaluate model performance. For inappropriate decision tree models, re-discretize the data and repeat the first six steps until the model evaluation results meet the requirements.

[0110] Specific examples:

[0111] The accuracy and generalization ability of the decision tree model of data center A were evaluated using the test set C, and it was found that the performance of the model was slightly lower than expected in some time periods. By re-discretizing the data of the specific time period of the computer room and retraining the model, an accuracy of 98% and an AUC value of 0.95 were finally achieved, meeting the needs of the data center.

[0112] Test set C contains 1,440 records, covering data for the entire day, ensuring the comprehensiveness and accuracy of the evaluation.

[0113] Simplified evaluation algorithm:

[0114] Suppose we select data Si from the training set D, Si = {P i ,T i, V i ,θ in ,θ ex} contains five characteristic attributes P d、 θ′ in and θ′ ex As evaluation data, the sample table data is as follows:

[0115] Table 1 Sample data

[0116] <![CDATA[P i (Well)]]> <![CDATA[T i (min)]]> <![CDATA[V i (in)]]> <![CDATA[θ in (%)]]> <![CDATA[θ ex (%)]]> 3 5 1200 40 60 7 5 1330 60 40 5 5 1250 50 50 ... ... ... ... ...

[0117] (1) The data Si in the training set D, Si = {P i ,T i, V i ,θ in ,θ ex The data is input into the decision-making model to guide the operation of the fresh air system:

[0118] 1. T i, V i ,θ in ,θ ex Input to the decision tree model.

[0119] 2. Get the current pressure set P of each area in the room at every time Ti in and outdoor pressure P out ;

[0120] (2) Calculate the new pressure difference

[0121] 1. Calculate the current positive pressure difference set P d ′=P in -P out ;

[0122] (3) Data collection

[0123] 1. Rewrite the records of each time Ti according to the newly collected data i ′,

[0124] S i ′={P i ,T i ,V i ,θ in ,θ ex}

[0125] 2. Collect data from each time period Si to form a new feature set S′,

[0126] S′={S1′+S2′+S3′+...+S n ′}

[0127] (4) Calculate the new test data set C

[0128] 1. Calculate the pressure drop rate δp within each time T according to the re-obtained S′={S1′+S2′+S3′+...+Sn′} i '

[0129] 2. As the test data set C′, C′={δP1′+δP2′+δP3′+···+δP n ′},

[0130] Where i = 0, 1, 2, ... n;

[0131] (5) Compare the test data sets C and C′, and calculate whether the data entries with a positive pressure difference deviation of less than 2Pa have achieved an accuracy of more than 98% and an AUC value of 0.95. If not, the data of a specific time period in the computer room is discretized again and the model is retrained. Finally, an accuracy of 98% and an AUC value of 0.95 are achieved, meeting the operation requirements of the data center.

[0132] This completes the intelligent adjustment and evaluation process of the decision tree model. As time goes by, through continuous intelligent adjustment, the pressure in the room can be kept near the manually set target positive pressure. This algorithm case only considers the impact of internal and external pressures, damper inlet and outlet openings, fan speed, and inlet and outlet time on indoor pressure. In real cases, factors such as indoor and outdoor temperature, humidity, natural wind pressure, and power consumption can be considered comprehensively. The decision tree model is more complex, but the principle is the same; when we use the four variables of internal and external pressure, damper inlet and outlet angles, fan speed, and inlet and outlet time for model training and decision-making, it can effectively adapt to the influence of external factors such as indoor and outdoor temperature, humidity, wind speed, and weather, and intelligently adjust the equipment operation status to adapt to different external conditions.

[0133] The present invention uses a decision tree model to make decisions based on data such as indoor and outdoor temperature, humidity, pressure, fan speed, damper opening, inlet and outlet time of each room in the data center. In a few minutes, it can give appropriate fresh air system adjustment input values ​​according to the operating conditions of different rooms, achieve the purpose of intelligent adjustment of each room to maintain positive pressure continuously, and minimize the cost waste caused by manual intervention and opening the fresh air system all day long. The innovation of the present invention is to use big data and decision tree intelligent adjustment algorithms. By centrally and uniformly controlling the fresh air control units in each room, it can effectively respond to the operation requirements of one or several buildings, one or several floors of computer rooms that need to maintain indoor positive pressure, and provide data center operators with an intelligent adjustment algorithm to maintain a clean operating environment for digital equipment with minimal investment. Compared with traditional fresh air algorithms, this intelligent adjustment method has strong adaptability, flexible expansibility, and high intelligence, and can effectively meet the intelligent adjustment requirements of positive pressure environments in multiple target rooms, saving labor and energy costs.

[0134] The above is only an example embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for maintaining positive pressure in a data center room based on deep learning, characterized in that: The steps include: Step 1: Data collection and extraction: Obtain training set F: Collect various relevant data inside and outside the room, including indoor pressure P, for each room. in , Outdoor pressure P out 、Inlet and outlet ventilation damper opening θ in and θ ex , the fan speed V of the fresh air system, the time interval T for opening and closing the fresh air equipment, and the various types of collected data are preliminarily screened based on the time interval T for opening and closing the fresh air equipment, and the maximum and minimum values ​​are filtered out to form a training set F; Step 2: Data preprocessing: clean the data in the training set F, supplement missing values ​​and process outliers, discretize the continuous variable values, and form a derived feature set F1 based on the influence relationship between the opening and closing time interval T of the fresh air equipment and the indoor and outdoor pressure difference P; Step 3: Extract feature data set S and test data set C: Extract feature data from the derived feature set F1, and extract the indoor and outdoor pressure difference P, the time interval between opening and closing of the fresh air equipment T, the fan speed V, and the ventilation inlet and outlet opening θ. in and θ ex、 Generate feature data set S for each time period as main feature data i = {P i , T i , V i ,θ in ,θ ex }, then extract the feature data set S for each time period i , generate key feature data set S = {S1+S2+S3+···+S n }; Then, a standard sample data is extracted by human intervention and mathematical methods to form the relevant data of the test data set C, C = {δP1+δP2+δP3+···+δP n }, where i = 0, 1, 2, ... n; Step 4: Form a decision tree data set TREE: Use entropy and information gain ratio algorithms to process the feature data set S, select the best features and splitting points, split the data based on the selected best features and splitting points to create the root node of the tree, and recursively perform the same process on each child node until the stopping condition is reached to form a decision tree data set TREE. The stopping condition is that all samples in the node belong to the same class, or the maximum tree depth is reached, or the number of samples in the node is less than a preset threshold. Step 5: Prune the decision tree to form a training set D: Perform post-pruning on the decision tree data set TREE, and delete the branches that contribute less to the generalization ability of the model from the bottom up to form a training set D; Step 6: Decision tree model construction and training: Use the data of training set D to build a decision tree model, and adjust the model parameters during the training process. The model parameters include the maximum depth of the tree and the minimum number of leaf node samples; Step 7: Decision tree model evaluation: Use the data in the test data set C to evaluate the accuracy and generalization ability of the decision tree model. For inappropriate decision tree models, re-discrete the collected data and repeat the first six steps until the decision tree model evaluation results meet the test requirements.

2. The method for maintaining positive pressure in a data center room based on deep learning as claimed in claim 1, characterized in that: In the first step, a full range of sensors are deployed to collect various relevant data inside and outside the room.

3. The method for maintaining positive pressure in a data center room based on deep learning as claimed in claim 1, characterized in that: The second step is to supplement the missing values ​​by: in and θ ex Perform zero filling processing to change the indoor pressure P of some points with negative pressure difference to in , Outdoor pressure P out Modify to 0; The specific processing of abnormal values ​​is as follows: identifying and eliminating outliers through Z-score or IQR method; Discrete processing is specifically: for the regular pressure data P, T, V and θ within a period of time in and θ ex , use discrete mathematical models for discrete processing and improve data generalization capabilities.

4. The method for maintaining positive pressure in a data center room based on deep learning as claimed in claim 1, characterized in that: The decision tree model constructed in the sixth step is used to identify abnormal data and misjudged data in the training set D classification. The specific method of identifying abnormal data and misjudged data is as follows: 1) Identify the opening degree θ of the inlet and outlet ventilation dampers in the training data in and θ ex Whether it is consistent with the actual opening percentage, and eliminate inappropriate data; 2) Identify whether the fan speed V of the fresh air system of some modules in the training data exceeds the actual operating speed range, and eliminate inappropriate data.

5. The method for maintaining positive pressure in a data center room based on deep learning as claimed in claim 1, characterized in that: In the seventh step, the accuracy and generalization ability of the decision tree model are evaluated using the data in the test set C. The confusion matrix, ROC curve and AUC value indicators are used to evaluate the model performance. The standard for the decision tree model evaluation results to meet the test requirements is to achieve an accuracy of 98% and an AUC value of 0.95.

Images