Machine room data intelligent sensing method for performing knowledge distillation by using star large model, storage medium and equipment

Abstract

The invention discloses a machine room data intelligent sensing method for performing knowledge distillation by using a star large model, a storage medium and equipment. The method comprises the following steps: collecting historical machine room data in various abnormal scenes from log files of a machine room; constructing a random forest model for computer room data key feature extraction; taking the large star model as a teacher model, taking the random forest model as a student model, respectively inputting historical machine room data into the teacher model and the student model, guiding the student model to train according to knowledge distillation loss between the teacher model and the student model until a knowledge distillation loss function converges, and completing training of the random forest model; collecting machine room data in real time, and inputting the machine room data into the trained random forest model to extract key features; and inputting the extracted key features into a fanotify component for machine room data anomaly identification. According to the invention, a lightweight random forest model is used to realize machine room data intelligent perception, thereby realizing abnormity identification.

Description

Technical Field

[0001] The present invention relates to the technical field of computer room data anomaly identification, and in particular to a computer room data intelligent perception method, storage medium and device for performing knowledge distillation using a star big model. Background Art

[0002] Data centers host core enterprise businesses, such as financial transactions, online services, and cloud computing. Any anomalies within the data center, such as server downtime or network latency, can lead to business failures. Real-time monitoring, intelligent analysis, and automated response to data center data can proactively detect anomalies, significantly reducing failure risks, improving service quality, and ensuring data security.

[0003] On the one hand, the dimensions of computer room data are complex, and truly abnormal data points may be masked by the noise effects of multiple unrelated dimensions, making it difficult for existing computer room data perception technologies to capture anomalies from high-dimensional data. On the other hand, in order to improve the perception accuracy of computer room data, most current models using complex structures, such as deep neural networks and integrated machine learning models, are used for computer room data perception. This requires the use of high-performance GPUs or TPUs for training and inference, significantly increasing hardware investment. At the same time, complex structure models require longer training and inference time, which cannot meet the requirements of real-time monitoring of computer rooms. Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention provides a method, storage medium and device for intelligent perception of computer room data using the Star Big Model for knowledge distillation. Knowledge distillation is used to optimize the random forest model for computer room data perception, greatly reducing the inference delay. At the same time, the optimized random forest model can extract key features from high-dimensional computer room data for computer room anomaly identification, greatly improving the accuracy of anomaly identification.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution: a method for intelligent perception of computer room data using a large star model for knowledge distillation, characterized by comprising the following steps: Step S1: Collect historical data of the computer room under various abnormal scenarios from the log files of the computer room; Step S2: Construct a random forest model for extracting key features of computer room data; Step S3: Use the Xingchen Big Model as the teacher model and the Random Forest model as the student model. Input the historical computer room data into the teacher model and the student model respectively. Guide the student model training according to the knowledge distillation loss between the teacher model and the student model until the knowledge distillation loss function converges, completing the training of the Random Forest model. Step S4: Collect computer room data in real time and input it into the trained random forest model to extract key features; Step S5: Input the extracted key features into the fanotify component for anomaly identification of computer room data.

[0006] Furthermore, step S3 includes the following sub-steps: Step S3.1: Input historical computer room data into the Xingchen big model to extract key features; Step S3.2: Use the historical data center data as the input of the random forest model, extract the key features from the Xingchen model as the soft labels of the random forest model, use the manually labeled key features in the historical data as the true labels of the random forest model, train the random forest model, and calculate the knowledge distillation loss function; Step S3.3: Adjust the hyperparameters of the random forest model, determine the pruning path of each subtree in the random forest model according to the pruning benefit, and update the random forest model; Step S3.4: Repeat steps S3.2-S3.3 with the updated random forest model until the knowledge distillation loss function converges, completing the training of the random forest model.

[0007] Furthermore, the knowledge distillation loss function The calculation process is:

[0008] in, Represents the error loss between the key features extracted by the random forest model and the true label, , Indicates the number of samples involved in training, i express N The index of J Indicates the number of key features extracted from each sample, j express J The index of Indicates the i The first sample extracted j Key features, Indicates the i In the sample j True labels, express The weight coefficient of Represents the error loss between the key features extracted by the random forest model and the soft labels, , Indicates the i In the sample j A soft tag, express The weight coefficient of Represents the error loss of the intermediate features between the random forest model and the Xingchen model, , Indicates the i The intermediate features of samples in the random forest model, Indicates the i The samples have the middle features in the star model. express The weight coefficient of .

[0009] Furthermore, the process of determining the pruning path of each subtree in the random forest model according to the pruning cost is as follows: i. For each node in the random forest model, calculate the pruning benefit of pruning the corresponding subtree; ii. Find the internal node with the minimum pruning benefit from pruning the corresponding subtree from all internal nodes in the random forest model, subtract the corresponding subtree, and get the new tree; iii. Repeat steps i-ii until only the root node remains in all trees.

[0010] Furthermore, the pruning benefit of pruning the corresponding subtree is The calculation process is:

[0011] in, Represents a node t The subtree with root Represents a subtree The number of leaf nodes on Indicates t The node is the error loss between the key features extracted from the leaf node and the true label, Represents a node t The root subtree The error loss between the extracted key features and the true labels, Indicates t The node is the error loss between the key features extracted from the leaf node and the soft label, Represents a node t The root subtree The error loss between the extracted key features and the soft labels.

[0012] Furthermore, the specific process of step S4 is: inputting the real-time collected computer room data into the trained random forest model, calculating the score of the computer room data in each dimension, and extracting the computer room data with the top m scores as key features.

[0013] Furthermore, the scoring of each dimension of the computer room data The calculation process is:

[0014] in, represents the number of subtrees in the trained random forest model, h express The index of Indicates that the data of a certain dimension computer room is in h The scores in the subtrees, , Indicates the h The number of nodes in the subtree, Indicates the s Nodes The amount of reduction in importance.

[0015] Furthermore, the specific process of step S5 is: configuring the rules for identifying anomalies in the computer room data in the fanotify component, matching the extracted key features with the rules for identifying anomalies in the computer room data, and if the matching degree exceeds the set threshold, determining that there is an anomaly in the computer room data; otherwise, the computer room data is operating normally.

[0016] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the method for intelligent perception of computer room data using the Star Big Model for knowledge distillation.

[0017] Furthermore, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the method for intelligent perception of computer room data using the Star Big Model for knowledge distillation is implemented.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The method for intelligent perception of computer room data using the Xingchen Big Model for knowledge distillation in the present invention uses the Xingchen Big Model as a teacher model. The Xingchen Big Model has cross-domain knowledge and can quickly and accurately capture the key features in the computer room data. Through knowledge distillation, the random forest model can learn the generalized knowledge of the Xingchen Big Model in different computer room scenarios. This enables the optimized random forest model to better maintain performance stability and reduce the risk of overfitting when facing a new computer room environment. In addition, the random forest model after knowledge distillation has a reduced number of parameters and reduced computational complexity, and can run quickly under limited computing resources, meeting the needs of real-time perception of computer room data. (2) The present invention calculates the pruning benefits of pruning subtrees in the random forest model based on the knowledge distillation loss function, which can more accurately evaluate the contribution of each subtree to the performance of the random forest model and only prune those parts that have a small impact on the knowledge distillation loss, thereby achieving more accurate pruning while ensuring the performance of the random forest model; The present invention utilizes an optimized random forest model to extract key features from high-dimensional computer room data for computer room anomaly identification, greatly improving anomaly identification accuracy and reducing inference latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the method for intelligent perception of computer room data using the Xingchen large model for knowledge distillation in the present invention; Figure 2 Flowchart of knowledge distillation in the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention will be further explained below with reference to the accompanying drawings.

[0021] like Figure 1 This is a flow chart of the intelligent perception method for computer room data using the Star Big Model for knowledge distillation in the present invention. The anomaly identification method includes the following steps: Step S1: Collect historical computer room data under various abnormal scenarios from the log files of the computer room, wherein the abnormal scenarios include: hardware equipment failure caused by server downtime, network equipment abnormality, storage system problem or power distribution failure, environmental abnormality caused by temperature and humidity out of control, computer room water leakage or air quality abnormality, operational errors caused by configuration errors, physical damage or abuse of authority, external threats caused by power outages, network attacks or natural disasters, and software abnormalities caused by operating system crashes, application server failures or log system abnormalities; therefore, for various abnormal scenarios, the computer room data that need to be collected include: hardware equipment data, environmental monitoring data, network performance data, security log data and business continuity data, and the abnormalities of the computer room can be identified through intelligent perception of these computer room data.

[0022] Step S2: Construct a random forest model for extracting key features of computer room data. Since computer room data often contains a large number of features and complex patterns, the original random forest model cannot fully learn due to structural limitations, resulting in large errors in the extraction of key features of computer rooms.

[0023] Step S3: Therefore, in the present invention, the Xingchen model is used as the teacher model, the random forest model is used as the student model, and the historical computer room data is input into the teacher model and the student model respectively. The student model training is guided by the knowledge distillation loss between the teacher model and the student model until the knowledge distillation loss function converges and the training of the random forest model is completed. The Xingchen model has cross-domain knowledge and can quickly and accurately capture the key features in the computer room data. Through knowledge distillation, the random forest model can learn the generalized knowledge of the Xingchen model in different computer room scenarios. This enables the optimized random forest model to better maintain performance stability and reduce the risk of overfitting when facing a new computer room environment. In addition, the random forest model after knowledge distillation can run quickly under limited computing resources due to the reduction in the number of parameters and the reduction in computational complexity, meeting the needs of real-time perception of computer room data. Figure 2 , including the following sub-steps: Step S3.1: Input historical computer room data into the Xingchen big model to extract key features; Step S3.2: Use the historical data center data as the input of the random forest model, extract the key features from the Xingchen model as the soft labels of the random forest model, use the manually labeled key features in the historical data as the true labels of the random forest model, train the random forest model, and calculate the knowledge distillation loss function; The knowledge distillation function in the present invention contains the error loss between the key features extracted by the random forest model and the true label , the error loss between the key features extracted by the random forest model and the soft labels And the error loss of the intermediate features between the random forest model and the Xingchen model ,in, Used to constrain the consistency of the output of the random forest model with the true label to ensure that the random forest model learns the basic pattern; The random forest model learns the knowledge imparted by the Xingchen large model, helping it capture more detailed decision logic. This allows the output distribution of the random forest model to gradually approach that of the Xingchen large model, improving the sensitivity of identifying anomalies in the data center. It is used to align the intermediate layer features of the random forest model and the Starry Sky model, forcing the random forest model to learn the high-level abstract representation of the Starry Sky model. The knowledge distillation loss function in this invention is The calculation process is:

[0024] in, , Indicates the number of samples involved in training, i express N The index of J Indicates the number of key features extracted from each sample,j express J The index of Indicates the i The first sample extracted j Key features, Indicates the i In the sample j True labels, express The weight coefficient of , Indicates the i In the sample j A soft tag, express The weight coefficient of , Indicates the i The intermediate features of samples in the random forest model, Indicates the i The samples have the middle features in the star model. express The weight coefficient of .

[0025] Step S3.3: Adjust the hyperparameters of the random forest model and determine the pruning path for each subtree in the random forest model based on the pruning benefit, thus updating the random forest model. During the pruning process, using the knowledge distillation loss as a guide ensures that the pruned random forest model retains the key knowledge learned from the Xingchen large model. This is crucial for data center data perception tasks, as this key knowledge may involve key features for identifying anomalies in the data center, thereby improving pruning efficiency.

[0026] The process of determining the pruning path of each subtree in the random forest model according to the pruning cost in the present invention is as follows: i. For each node in the random forest model, calculate the pruning benefit of pruning the corresponding subtree :

[0027] in, Represents a node t The subtree with root Represents a subtree The number of leaf nodes on Indicates t The node is the error loss between the key features extracted from the leaf node and the true label, Represents a node t The root subtree The error loss between the extracted key features and the true labels, Indicates t The node is the error loss between the key features extracted from the leaf node and the soft label, Represents a node t The root subtree The error loss between the extracted key features and the soft labels.

[0028] ii. From all internal nodes in the random forest model, find the internal node with the smallest pruning benefit when pruning its corresponding subtree. Subtract the corresponding subtree from the new tree to obtain a new tree. The pruning benefit can more accurately evaluate the contribution of each subtree to the performance of the random forest model. Only those parts that have the least impact on the knowledge distillation loss can be pruned, thus achieving more accurate pruning while maintaining the performance of the random forest model. iii. Repeat steps i-ii until only the root node remains in all trees.

[0029] Step S3.4: Repeat steps S3.2-S3.3 with the updated random forest model until the knowledge distillation loss function converges, completing the training of the random forest model.

[0030] Step S4: Collect computer room data in real time and input it into the trained random forest model to extract key features. This can extract key features from high-dimensional computer room data for computer room anomaly identification, greatly improving anomaly identification accuracy and reducing inference latency. Specifically, input the real-time collected computer room data into the trained random forest model, calculate the score of the computer room data in each dimension, and extract the computer room data with the top m scores as key features.

[0031] Scoring of data in each dimension of the computer room in this invention The calculation process is:

[0032] in, represents the number of subtrees in the trained random forest model, h express The index of Indicates that the data of a certain dimension computer room is in h The scores in the subtrees, , Indicates the h The number of nodes in the subtree, Indicates the s Nodes The amount of reduction in importance.

[0033] Step S5: The extracted key features are input into the fanotify component for use in identifying anomalies in the computer room data. The fanotify component runs directly in kernel space, eliminating the need for frequent context switching and thus preventing performance impacts on various systems in the computer room. Furthermore, the fanotify component's file system monitoring mechanism enables real-time capture of file access, modification, and deletion operations. Specifically, rules for identifying anomalies in the computer room data are configured in the fanotify component. The extracted key features are then matched against these rules. If the match exceeds a set threshold, the computer room data is deemed to be anomaly; otherwise, the computer room data is considered to be operating normally.

[0034] The intelligent perception method of computer room data in the present invention uses the Star Big Model for knowledge distillation to construct a lightweight random forest model through knowledge distillation, which is used to extract key features from high-dimensional computer room data, greatly improving the accuracy of key feature extraction and reducing latency. It also uses the fanotify component to identify abnormal computer room data from key features, which can reduce overhead and false alarm rate, provide reliable support for abnormal warning of computer room data, and thus take relevant measures in advance to reduce the risk of computer room abnormalities.

[0035] In one technical solution of the present invention, a computer-readable storage medium is also provided, which stores a computer program, and the computer program enables a computer to execute the method for intelligent perception and anomaly identification of computer room data using the Star Big Model for knowledge distillation.

[0036] In one technical solution of the present invention, an electronic device is also provided, including: a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the method for intelligent perception and anomaly recognition of computer room data using the Star Big Model for knowledge distillation is implemented.

[0037] In the embodiments disclosed herein, computer storage media may be tangible media that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. Computer storage media may include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media may include electrical connections based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), optical fibers, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0038] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0039] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent perception of computer room data using the Xingchen large model for knowledge distillation, characterized in that: The steps include: Step S1: Collect historical data of the computer room under various abnormal scenarios from the log files of the computer room; Step S2: Construct a random forest model for extracting key features of computer room data; Step S3: Use the Xingchen Big Model as the teacher model and the Random Forest model as the student model. Input the historical computer room data into the teacher model and the student model respectively. Guide the student model training according to the knowledge distillation loss between the teacher model and the student model until the knowledge distillation loss function converges, completing the training of the Random Forest model. Step S4: Collect computer room data in real time and input it into the trained random forest model to extract key features; Step S5: Input the extracted key features into the fanotify component for anomaly identification of computer room data.

2. The method for intelligent perception of computer room data using the Xingchen big model for knowledge distillation according to claim 1 is characterized in that: Step S3 includes the following sub-steps: Step S3.1: Input historical computer room data into the Xingchen big model to extract key features; Step S3.2: Use the historical data center data as the input of the random forest model, extract the key features from the Xingchen model as the soft labels of the random forest model, use the manually labeled key features in the historical data as the true labels of the random forest model, train the random forest model, and calculate the knowledge distillation loss function; Step S3.3: Adjust the hyperparameters of the random forest model, determine the pruning path of each subtree in the random forest model according to the pruning benefit, and update the random forest model; Step S3.4: Repeat steps S3.2-S3.3 with the updated random forest model until the knowledge distillation loss function converges, completing the training of the random forest model.

3. The method for intelligent perception of computer room data using the Xingchen large model for knowledge distillation according to claim 1 is characterized in that: The knowledge distillation loss function The calculation process is: in, Represents the error loss between the key features extracted by the random forest model and the true label, , Indicates the number of samples involved in training, i express N The index of J Indicates the number of key features extracted from each sample, j express J The index of Indicates the i The first sample extracted j Key features, Indicates the i In the sample j True labels, express The weight coefficient of Represents the error loss between the key features extracted by the random forest model and the soft labels, , Indicates the i In the sample j A soft tag, express The weight coefficient of Represents the error loss of the intermediate features between the random forest model and the Xingchen model, , Indicates the i The intermediate features of samples in the random forest model, Indicates the i The samples have the middle features in the star model. express The weight coefficient of .

4. The method for intelligent perception of computer room data using the Xingchen large model for knowledge distillation according to claim 3 is characterized in that: The process of determining the pruning path of each subtree in the random forest model according to the pruning cost is as follows: i. For each node in the random forest model, calculate the pruning benefit of pruning the corresponding subtree; ii. Find the internal node with the minimum pruning benefit from pruning the corresponding subtree from all internal nodes in the random forest model, subtract the corresponding subtree, and get the new tree; iii. Repeat steps i-ii until only the root node remains in all trees.

5. The method for intelligent perception of computer room data using the Xingchen large model for knowledge distillation according to claim 4 is characterized in that: The pruning benefit of pruning the corresponding subtree The calculation process is: in, Represents a node t The subtree with root Represents a subtree The number of leaf nodes on Indicates t The node is the error loss between the key features extracted from the leaf node and the true label, Represents a node t The root subtree The error loss between the extracted key features and the true labels, Indicates t The node is the error loss between the key features extracted from the leaf node and the soft label, Represents a node t The root subtree The error loss between the extracted key features and the soft labels.

6. The method for intelligent perception of computer room data using the Xingchen large model for knowledge distillation according to claim 5 is characterized in that: The specific process of step S4 is: inputting the real-time collected computer room data into the trained random forest model, calculating the score of the computer room data in each dimension, and extracting the computer room data with the top m scores as key features.

7. The method for intelligent perception of computer room data using the Xingchen large model for knowledge distillation according to claim 6 is characterized in that: Scoring of computer room data in each dimension The calculation process is: in, represents the number of subtrees in the trained random forest model, h express The index of Indicates that the data of a certain dimension computer room is in h The scores in the subtrees, , Indicates the h The number of nodes in the subtree, Indicates the s Nodes The amount of reduction in importance.

8. The method for intelligent perception of computer room data using the Xingchen big model for knowledge distillation according to claim 1 is characterized in that: The specific process of step S5 is as follows: configuring the rules for identifying abnormalities in the computer room data in the fanotify component, matching the extracted key features with the rules for identifying abnormalities in the computer room data, and determining that the computer room data is abnormal if the matching degree exceeds the set threshold; Otherwise, the data in the computer room operates normally.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the method for intelligent perception of computer room data using the Star Big Model for knowledge distillation as described in any one of claims 1-8.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for intelligent perception of computer room data using a large star model for knowledge distillation as described in any one of claims 1 to 8 is implemented.

Images