Method and device for judging active state of storage link of host machine based on characteristic relaxation regression

Abstract

The invention discloses a host machine storage link activity state discrimination method and device based on feature relaxation regression, and the method comprises the steps: inputting an original training sample, and constructing a training data table; performing digital feature processing on non-digital features in the original sample; marking the sample; obtaining a similar sample set; constructing a target function; solving an objective function, and iteratively calculating a regression matrix, a comprehensive information matrix of a similar sample set and a regression intercept; using the trained discrimination model to predict a sample label of a to-be-tested sample, and completing host machine storage link activity state discrimination; according to the method, from original high-dimensional, redundant and discrete host machine storage link original data, comprehensive information in a similar sample set is extracted, the easy-to-distinguish direction of a heterogeneous sample set is searched, a relaxation coefficient and a loss function are introduced, the relaxation degree of a sample regression sample label is adjusted, rigid regression and over-fitting pain points of a discrimination model are solved, and the reliability of the host machine storage link is improved. And judging the active state of the storage link of the host machine.

Description

Technical Field

[0001] The present invention belongs to the field of financial technology, and relates to a method and device for distinguishing the activity status of a host storage link based on feature relaxation regression in a cloud computing platform in the financial field. Background Art

[0002] As an important computing node of the cloud management platform, the host machine provides computing, network, storage and other computing resources for the operation of various systems in financial technology. It is particularly important to ensure that the host machine provides stable and reliable computing services, timely discover operational risks, especially host storage link risks, and ensure that the host storage services are not interrupted and the business is not degraded.

[0003] When the host storage link is inactive, various business system services running on the host may be affected and may even be unable to provide financial services or conduct business, which may even lead to customer complaints and financial losses. How to timely monitor and determine the activity status of the host storage link has become a problem that needs to be solved.

[0004] The traditional method of inspecting the active status of the host storage link is to log in to the cloud management platform where the host is located and check the status of each storage link of each host one by one. This manual identification method has the following problems: 1. This method is completely manual and depends on the ability and technical level of the inspectors. 2. Inspection is time-consuming, laborious, and inefficient, and cannot meet the timeliness requirements of host storage link operation management and monitoring; 3. Inspection operations are manual, with the risk of misclicks, misoperations, and missed clicks and missed analyses; In addition, for monitoring the active status of the host storage link, many financial institutions will configure monitoring alarms in the cloud management platform. This method has the following shortcomings: 1. The alarm module needs to be embedded in the cloud management platform, relying on the cloud management platform's own alarm configuration and monitoring agility and accuracy. 2. In the event that the cloud management platform itself fails to operate or the alarm configuration is lost, it will not be possible to monitor the active status of the host storage link; 3. It is impossible to conduct offline analysis and inspection of the active status of the storage link; 4. Alarm configuration and confirmation operations still require logging into the cloud management platform, which still carries the risk of accidental clicks and misoperations.

[0005] In recent years, in the analysis models or algorithms for the original information of the host storage link, due to the high dimension of the original data and the coupling, redundancy, and linear relationship between the data dimensions, there are pain points such as high direct judgment calculation overhead, insignificant analysis effect, overfitting of the judgment model, and weak generalization ability.

[0006] Therefore, a method that does not require logging into the cloud management platform and does not rely on the cloud management platform to run can process the original high-dimensional information, eliminate coupling, linear, and redundant dimensional information, and prevent model overfitting, which is of great significance to improving the judgment effect of the active status of the host storage link. Summary of the invention

[0007] The purpose of the present invention is to address the pain points that in the VMware cloud computing platform in the traditional financial field, technical personnel are required to log in to the cloud management platform to perform manual inspections, the inspection results depend on the ability and technical level of the inspectors, the inspection efficiency is low, and the timeliness requirements of the host storage link operation management and monitoring cannot be met. At the same time, the inspection operation is manual, which has the defects of misclick, misoperation risk and missed click, missed analysis problem; or it needs to rely on agile and high-precision alarm configuration, needs to be coupled and embedded in the cloud management platform, and relies on the cloud management platform to run, which cannot meet the offline analysis and inspection needs. There is a pain point that the host storage link activity status monitoring cannot be realized when the cloud management platform itself fails to run or the alarm configuration is lost or fails, and it is still necessary to manually log in to the cloud management platform, which still has many pain points and defects such as misclick and misoperation. A method and device for distinguishing the activity status of the host storage link is proposed, which can be operated without manual operation, does not rely on the operation of the cloud management platform, can process the original high-dimensional information, eliminate coupling, linear, and redundant dimensional information, and can autonomously iterate and regress to learn the comprehensive information of the same sample set and the easily distinguishable direction of the heterogeneous sample set, and introduces a loss function. By adjusting the loss function relaxation coefficient and the loss function adjustment parameter, the relaxation degree of the sample feature regression sample label is adjusted to solve the rigid regression and overfitting pain points of the discriminant model, and realize the sample feature relaxation regression sample label to prevent the model from overfitting.

[0008] According to a first aspect of the present invention, a method for determining the active state of a host storage link based on feature relaxation regression is provided, the method comprising the following steps: (1) Obtain the host name and host storage link status information through the cloud management platform operation data, obtain the original sample, and perform digital feature processing on the non-digital features; (2) Using the host storage link activity status in the original sample as a label, a sample set is obtained based on the label; (3) Reduce the dimension of the feature data of the original sample and set the objective function to perform linear regression on the sample label. The objective function includes: Feature regression, the loss function is set based on the degree of allowable deviation of the regression sample label as the relaxation coefficient; The regression intercept indicates the regression prediction value of the sample feature regression sample label when all sample feature values ​​are 0; Obtain the comprehensive information matrix of the sample set, which integrates all sample information in the sample set and has the smallest difference with the comprehensive information of the sample set; Select the easily distinguishable direction in the heterogeneous sample set, make the samples of different classes distinguishable after the sample dimension reduction based on the regression matrix, and maximize the difference of the comprehensive information matrix between heterogeneous samples; (4) The objective function is solved iteratively to obtain the discriminant model. The sample labels are predicted based on the discriminant model to obtain the host storage link activity status discrimination result.

[0009] Furthermore, the digital feature processing of non-numeric features in the original sample data is specifically as follows: a digital dictionary of non-numeric feature dimensions is established to achieve a one-to-one mapping relationship between each dimensional feature and a number, and the digital processing of the non-numeric feature dimensional features of the original input sample is completed by digitally replacing the original features.

[0010] Furthermore, according to the actual situation of the host storage link activity status, the samples are manually marked, and the mark is used as the sample label. Based on the sample label that all host storage link statuses are active and the sample label that there are inactive links in the host storage link status, corresponding sample sets are obtained respectively.

[0011] Furthermore, the loss function of feature regression allows all sample features to have a certain offset when regressing sample labels. In the loss function of feature regression, the loss function relaxation coefficient is designed based on the range of allowable offsets of sample feature regression sample labels. The larger the value of the loss function relaxation coefficient, the larger the range of allowable offsets of sample feature regression sample labels. The loss function adjustment parameter is designed based on the degree of sample constraint on sample features that exceed the allowable offset range. The larger the value of the loss function adjustment parameter, the greater the degree of constraint on sample feature regression sample labels that exceed the offset range, requiring all samples to be as close to the deviation range as possible.

[0012] Furthermore, the relaxation coefficient in the loss function and the loss function adjustment parameters are adjusted to adjust the degree of relaxation of sample regression sample labels, so as to achieve relaxation regression of sample features to sample labels.

[0013] Furthermore, the comprehensive information of the sample set is obtained by multiplying the sample weight coefficient in the sample set by the sample and summing them up. When the sample eigenvalue is missing or some eigenvalues ​​are N / A, the sample weight coefficient of the sample in the sample set is 0. When the sample eigenvalue is complete and not missing, the sample weight coefficient of the sample in the sample set is the inverse of the number of complete samples.

[0014] Furthermore, the regression matrix, regression intercept and comprehensive information matrix of the sample set are initialized, and the objective function is solved by alternating iteration for the regression matrix, regression intercept and comprehensive information matrix in the objective function. When the number of iterations reaches the preset maximum number of iterations or the difference between the regression intercept, regression matrix and comprehensive information matrix of the sample set of two adjacent iterations is less than the preset difference threshold, the iteration is stopped.

[0015] According to a second aspect of the present invention, there is provided a host storage link activity status determination device based on feature relaxation regression, comprising a memory and one or more processors, wherein the memory stores executable code, and wherein when the processor executes the executable code, the host storage link activity status determination method based on feature relaxation regression is implemented.

[0016] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a program stored thereon, wherein when the program is executed by a processor, the method for determining the activity status of a host storage link based on a feature relaxation regression is implemented.

[0017] According to a fourth aspect of the present invention, there is provided a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method for determining the activity state of a host storage link based on characteristic relaxation regression is implemented.

[0018] The advantages and beneficial effects of the present invention are: The method of the present invention aims at the pain points and risks of missing inspections and misoperations in the traditional host storage link activity status discrimination method, such as the inefficiency of manual inspections, reliance on technical personnel, and the need to log in to the cloud management platform, or the alarm configuration module is coupled and embedded in the cloud management platform, relying on the cloud management platform to run, the alarm effect depends on the accuracy and agility of the alarm configuration, and the pain points of still needing to manually log in to the cloud management platform, etc., and proposes a host storage link activity status discrimination method that can be operated without manual operation, does not rely on the operation of the cloud management platform, can process the original high-dimensional information, and eliminate coupling, linear, and redundant dimensional information. The method performs digital feature processing on the non-digital features of the original storage link data of the host, and realizes the calculation of the original feature formula. By synthesizing all sample information in each type of sample set, and minimizing the difference with the comprehensive information of the sample set, the comprehensive information matrix of each type of sample set is obtained through cyclic alternating iteration. By maximizing the difference in the comprehensive information matrix between heterogeneous samples, selecting the easily distinguishable direction in the heterogeneous sample set, different types of samples can have better discrimination after the sample dimension is reduced. By introducing the loss function and adjusting the relaxation coefficient and loss function adjustment parameters in the loss function, the relaxation degree of sample regression sample label is adjusted, the sample feature relaxation regression sample label is realized, the rigid regression and overfitting problems of the model are solved, and the activity status of the host storage link is determined. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 The flowchart of a method for determining the active state of a host storage link by feature relaxation regression is shown as an exemplary embodiment.

[0021] Figure 2 The sample distribution scatter plot of the input samples in the first and second feature dimensions of the original high-dimensional space is shown as an exemplary embodiment.

[0022] Figure 3 The sample distribution scatter plot of the input samples in the 3rd and 4th feature dimensions of the original high-dimensional space is shown as an exemplary embodiment.

[0023] Figure 4 The sample distribution scatter plot of the input samples in the 5th and 6th feature dimensions of the original high-dimensional space is shown as an exemplary embodiment.

[0024] Figure 5A sample distribution scatter plot of input samples projected into a one-dimensional space is shown as an exemplary embodiment.

[0025] Figure 6 The present invention is a structural diagram of a cloud platform host network link redundancy status determination device shown as an exemplary embodiment. DETAILED DESCRIPTION

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0027] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0028] like Figure 1 As shown, an embodiment of the present invention provides a method for distinguishing the activity status of a host storage link based on feature relaxation regression. The method of the present invention can effectively distinguish the activity status of a host storage link on a VMware cloud platform, but is not limited to the VMware cloud platform, and other cloud platforms can be similarly promoted. The method of the present invention includes the following steps: Step 1: Obtain original samples for training and construct a training data table.

[0029] Step 1.1: Get original samples for training.

[0030] The specific operation is: use the RVTools tool, enter the corresponding cloud management platform IP address, user name and password in the tool, click Collect Data, select Excel as the export format of the data, and you can obtain the cloud management platform operation data (the platform where the RVTools tool is deployed must be able to access the cloud management platform).

[0031] The RVTools tool can export the graphical operation data of the cloud management platform and convert it into Excel format data. Among them, the operation data of the host storage link is in the 15th data table in the Excel book (the table name is vHBA). The table has 13 columns. In addition to the first column for the specific host host name, the 12 columns from the second column to the thirteenth column are the host storage link status information.

[0032] The first column of the table is the host name (vNicHostName) of the host, which indicates which host has storage link information; the second column is the name of the data center where the host is located (vHBADatacenter); the third column is the name of the cluster where the host is located (vHBACluster); the fourth column is the name of the host storage NIC device (vHBADevice); the fifth column is the host storage NIC type (vHBADType); the sixth column is the host storage NIC status (vHBAStatus); the seventh column is the host storage NIC bus (vHBABus); the eighth column is the host storage NIC slot (vHBAPci); the ninth column is the host storage NIC driver name (vHBADriver); the tenth column is the host storage NIC driver mode (vHBAModel); the eleventh column is the host storage NIC unique identifier (vHBAWorldWideName); the twelfth column is the cloud management platform server IP address (vHBAVISDKServer); and the thirteenth column is the cloud management platform instance universal unique identifier (vHBAInstanceUUID).

[0033] In this implementation, the number of storage network cards connected to the host is 2, the storage link is dual-uplink, and the connected storage is centralized storage with optical fiber transmission. The storage link information of a single host is a 2-row 12-column data, and each 1-row 12-column data represents the storage link information of a storage network card; after the host storage link information is converted from row to column, the storage link information of a single host is a 2-column 12-row data. After the sample is straightened into a column, that is, the storage link data of the second column of the host is connected to the storage link data of the first column, and the first row of the storage link data of the second column of the host is arranged next to the last row of the storage link data of the first column. Therefore, in this implementation, the original host storage link information is converted from a 2-row 12-column matrix data to a column vector through row-column conversion and straightening, and the training sample becomes a 24-row column vector.

[0034] The original table is converted into rows and columns, and the two storage link data belonging to the same host are straightened into a column vector according to the order of the link data column numbers, which is used as the training data table. In this implementation, the host is connected to two network cards, and the storage link is dual-uplink. The storage link data of the sample is shown in Table 1.

[0035] Table 1

[0036]

[0037] Among them, "XSZX" is a certain center, "XSZX_SC01" is a certain center production area 01, vmhba64 is VMware storage network card 64, "Fibre Channel" is fiber channel, "online" is online, "0:00:17" indicates a slot position of the host storage network card, "qfle3f" is a network card driver type, "QLogic Inc. FCoE Adapter" is a network card driver mode, "50060B0000C2D603" and "50060B0000C2D602" are the numbers of the unique identifiers of the storage network cards on the sample host.

[0038] This implementation process is divided into a modeling process and a discrimination test process. In this implementation process, the total number of samples is 600. In the modeling process, there are 400 samples in total, of which the host storage link status of the first 200 samples is active, and the host storage link status of the last 200 samples has an inactive storage link. In the discrimination test process, there are 200 samples in total, of which the host storage link status of the first 100 samples is active, and the host storage link status of the last 100 samples has an inactive storage link.

[0039] During the modeling process, the storage link data of the first four samples are shown in Tables 2 and 3.

[0040] Table 2

[0041]

[0042] Table 3

[0043]

[0044] Step 2: Process the non-numeric features in the original sample into digital features. For each dimensional feature in the original sample, establish a digital dictionary of features. This dictionary can realize a one-to-one mapping relationship between each dimensional feature and a number. Through the digital dictionary, in each dimensional feature, the original features are replaced with numbers to complete the digital processing of the original input sample features. Specifically, there are two steps. Among them, data is filled in for the part of the feature dimension in the training data table that is missing a feature value. If a feature dimension has no feature value or displays N / A, the number 0 is used to fill in the data. Therefore, Table 3 can be updated to the following Table 4: Table 4

[0045]

[0046] Step 2.1: Establish a corresponding digital dictionary for the non-numeric feature dimension. The establishment of the digital dictionary is specifically as follows: for the features on a certain non-numeric feature dimension, different numerical values ​​are mapped according to the input order of the features, and the numerical values ​​are the input order of the feature data; The establishment of the digital dictionary is as follows: according to the column number order of the samples in the training data table and the row number order of the sample features in the same column of samples, they are input into the memory in sequence, and the order in which the features of each feature dimension are input into the memory is recorded. For example, if a feature value of the feature dimension appears for the first time, the input order of the feature is recorded as 1, and the feature value is mapped to 1; if it appears for the second time, it is recorded as 2, and the feature value is mapped to 2, and so on, and the feature input order of each feature dimension is counted; In this implementation, the order in which the storage link data in the training data table is input into the memory is as follows: the first row of the storage link data (i.e., the first sample) in the first column of the training data table is first input into the memory, followed by the second row of the storage link data in the first column until all rows of the storage link data in the first column are input into the memory. Then the first row of the storage link data (i.e., the second sample) in the second column is input into the memory, followed by the second row of the storage link data in the second column is input into the memory until all rows of the storage link data in the second column are input into the memory, and so on, and the feature input order of each feature dimension is counted; In this implementation, except for the sixth feature "vHBABus" which is a number, the feature values ​​of the remaining dimensional features are all non-numeric.

[0047] Take the first feature dimension, the name of the data center where the host is located (vHBADatacenter) as an example. Among these 600 samples, there is only one feature value category, and the corresponding feature value is "XSZX". Therefore, the digital dictionary of the first feature dimension, here called 1.dict, is specifically: .

[0048] Take the second feature dimension, the name of the cluster where the host is located (vHBACluster), as an example. Among these 600 samples, there are only 9 feature value categories, and the corresponding feature values ​​are "XSZX_SC01", "XSZX_SC02", "XSZX_SC03", "XSZX_SC04", "XSZX_DMZ01", "XSZX_DMZ02", "XSZX_MGMT01", "XSZX_BG01", and "XSZX_BG02". Among them, the feature value "XSZX_SC01" is first input into the memory, and the input order of this feature value is 1, so the feature value mapping is 1; the feature value "XSZX_SC02" is secondly input into the memory, and the input order of this feature value is 2, so the feature value mapping is 2, and so on, the feature value "XSZX_BG02" is last input into the memory, and the input order of this feature value is 9, so the feature value mapping is 9. Therefore, the digital dictionary of the second feature dimension, here called 2.dict, is: .

[0049] By analogy, the third feature dimension is the storage network card device name (vHBADevice). The digital dictionary of the third feature dimension is called 3.dict here, and is specifically: .

[0050] The fourth feature dimension is the host storage network card type (vHBADType). The digital dictionary of the fourth feature dimension, here called 4.dict, is as follows: .

[0051] The fifth feature dimension is the host storage network card status (vHBAStatus). The digital dictionary of the fifth feature dimension, here called 5.dict, is as follows: .

[0052] The sixth feature dimension is the host storage network card bus (vHBABus). The feature value of this dimension feature is a number itself, and there is no need to establish a digital dictionary.

[0053] The seventh feature dimension is the host storage network card slot (vHBAPci). The host storage link activity status has nothing to do with the storage network card slot. Therefore, the digital dictionary of this feature dimension maps the feature to 1 if there is a feature value, and maps it to 0 if there is no feature value or N / A is displayed. The digital dictionary of the seventh feature dimension is called 7.dict here, and is as follows: .

[0054] The eighth feature dimension is the host storage network card driver name (vHBADriver). The digital dictionary of the eighth feature dimension, referred to as 8.dict, is as follows: .

[0055] The ninth feature dimension is the host storage network card driver mode (vHBAModel). The digital dictionary of the ninth feature dimension, here called 9.dict, is as follows: .

[0056] The 10th feature dimension is the unique identifier of the network card storage network card (vHBAWorldWideName). In this implementation, the host storage link is dual-linked, so the identified storage network card unique identifier should appear in pairs. If only a single storage network card is identified, the storage network card unique identifier only appears alone. If no storage network card is identified, there is no storage network card unique identifier information. The digital dictionary of the 10th feature dimension, here called 10.dict, is specifically: .

[0057] The 11th feature dimension is the IP address of the cloud management platform server (vHBAVISDKServer). The digital dictionary of the 11th feature dimension, referred to here as 11.dict, is as follows: .

[0058] The 12th feature dimension is the universal unique identifier of the cloud management platform instance (vHBAInstanceUUID). The digital dictionary of the 12th feature dimension, here called 12.dict, is as follows: .

[0059] Step 2.2: Based on the digital dictionary established in step 2.1, replace the non-digital features in the sample with digital ones.

[0060] Specifically, according to the digital dictionary of each non-numeric feature established in step 2.1, the non-numeric features in the sample are digitally processed by replacing the features with the numbers they are mapped to. In this way, each feature in the sample is digital and can be processed by calculation.

[0061] Taking the example samples shown in Table 1 and the feature information of the first four samples shown in Table 2 and Table 3 as examples, after feature replacement by non-numeric features, the example samples are shown in Table 5 and the first four samples are shown in Table 6.

[0062] Table 5

[0063]

[0064] Table 6

[0065]

[0066] Step 3: Label the samples. According to the actual status of the host storage link activity in the training samples, the samples are manually labeled, and the label is the sample label.

[0067] The specific sample label is: when the host storage link status is all active, its sample label is recorded as ; When there is an inactive link in the host storage link, its sample label is recorded as .

[0068] In this implementation, the input sample is denoted as , for the input samples, denoted as , whose sample labels are recorded as ,in .

[0069] Step 4: Get a set of samples with the same sample label The sample labels are The sample is a sample set, recorded as set The sample labels are The sample is a sample set, recorded as set .

[0070] This implementation process is divided into a modeling process and a discrimination test process. In this implementation process, the total number of samples is 600. In the modeling process, there are 400 samples in total, of which the host storage link status of the first 200 samples is active, and the host storage link status of the last 200 samples has inactive storage links. In the discrimination test process, there are 200 samples in total, of which the host storage link status of the first 100 samples is active, and the host storage link status of the last 100 samples has inactive storage links. Therefore, in this modeling process, , ; , .

[0071] Step 5: Construct the objective function. The objective function establishment process is as follows: The original input samples have many dimensions, and there are coupling, redundancy, and linear relationships between features. Some features have a low contribution to sample classification. The existence of these features will affect the data analysis effect. Through data dimensionality reduction, these features are eliminated, and linear regression is performed on the sample labels after sample dimensionality reduction.

[0072] Based on the above ideas, the objective function is proposed as follows: (1) in, is the objective function, is the number of samples, is the regression matrix. For the samples, . For the Samples The sample labels, . is the regression intercept. The objective function is minimized, and the sample After feature dimensionality reduction, return to sample labels as much as possible .

[0073] After reducing the dimensionality of the samples in the above objective function, for each sample , Need to be as equal to the sample label as possible , the smaller the deviation, the better. This is a relatively rigid constraint that requires all samples to regress to sample labels as much as possible. This algorithm has the problems of weak generalization ability and model overfitting, that is, it is easy for the model to perform very well on the training data, but perform poorly when testing data outside the implementation process.

[0074] In order to avoid rigid constraints, improve the generalization ability of the algorithm, solve the overfitting problem, and allow the regression sample labels to have a certain deviation after sample dimensionality reduction, the present invention introduces a loss function, and relaxes the objective function by introducing the loss function, so the objective function can be updated as follows: (2) in It is determined by the loss function, which is formula (3): (3) in: , which is called the loss function relaxation coefficient in the present invention. is an adjustment parameter in the loss function, is a constant, which is referred to as the loss function adjustment parameter in the present invention, and .

[0075] In the above formula (3), is the relaxation coefficient of the loss function, reflecting Regression sample labels The degree of permissible deviation, The larger the value, the greater the allowable deviation. Its purpose is to allow all sample regression sample labels to have a certain deviation. Adjust the parameters for the loss function, The larger the value, the Regression sample labels The greater the degree of constraint on samples beyond the deviation range, the more the objective function is minimized, that is, all samples are required to be within the deviation interval as much as possible. If is infinite, all samples are required to be within the deviation interval.

[0076] The overall objective function is proposed as follows: (4) Among them, during this implementation process, ( )exist The sample weight coefficient in the class sample set is determined by formula (5): (5) For the sample set The total number of samples, For the sample set The number of samples in which the sample eigenvalues ​​are missing or some eigenvalues ​​are N / A. , , , therefore, the sample weight coefficient is as follows (6): (6) For each sample Determined by the loss function, as shown in formula (3); is the number of samples, ; is the regression matrix, matrix; For the sample, ; For the sample The sample labels of ; is the regression intercept; For the A collection of class samples; for exist The sample weight coefficient in the class sample set. In this implementation, if the sample feature value is missing or some feature values ​​are N / A, the sample weight coefficient is 0. On the contrary, if the sample feature value is complete and not missing, the sample weight coefficient is In this modeling process, The value of is determined by formula (6); Represents the comprehensive information of the sample set; is the number of sample categories. In this implementation ; , For the , The comprehensive information matrix of the class sample set. To adjust the parameters. is the square of the F-norm of the matrix and the sum of the squares of all elements in the matrix.

[0077] The first term of the objective function is feature regression. The second term is to obtain the comprehensive information matrix of the sample set, which integrates all the sample information in the sample set and has the smallest difference with the comprehensive information of the sample set. The third term is to select the easily distinguishable direction in the heterogeneous sample set, so that after the sample dimensionality is reduced, different types of samples have better distinguishability.

[0078] Let the comprehensive information of the sample set , the objective function is simplified as follows: (7) in, For the matrix The objective function is further decomposed into the trace of , which is equivalent to the optimization as follows: (8) Step 6: Solve the objective function.

[0079] In the objective function, the unknown variables have regression matrices , regression intercept , sample set comprehensive information matrix The objective function is solved by alternating iteration.

[0080] fixed , beg The objective function is minimized. Taking partial derivatives we get: (9) Let formula (9) equal to zero, we can get: (10) in: (11) fixed , , find the regression matrix . To minimize the objective function, Lagrange multipliers are introduced , , for The matrix represents the regression matrix The Lagrangian matrix of . The i in the equation represents the i-th row and the j represents the j-th column. Taking partial derivatives we get: (12) Let formula (12) equal to zero and introduce the Lagrange multiplier condition, The regression matrix can be obtained Iteration formula: (13) fixed , , ,beg . The objective function is decomposed into the following: (14) Minimize the objective function and introduce the Lagrange multiplier , for The matrix represents the comprehensive information matrix The Lagrangian matrix of . The i in the equation represents the i-th row, and the j represents the j-th column. ,right Taking partial derivatives we get: (15) Let formula (15) equal to zero and introduce the Lagrange multiplier condition, The iterative formula of the comprehensive information matrix can be obtained: (16) Similarly, fixed , , ,beg . To minimize the objective function, Lagrange multipliers are introduced , for The matrix represents the comprehensive information matrix The Lagrangian matrix of . The i in the equation represents the i-th row, and the j represents the j-th column. ,right Taking partial derivatives we get: (17) Let formula (17) equal to zero and introduce the Lagrange multiplier condition, Comprehensive information matrix can be obtained The iteration formula is: (18) Step 7: Initialize the regression matrix , regression intercept , sample set comprehensive information matrix , set the adjustment parameters , relaxation coefficient in the loss function And the loss function adjustment parameters .

[0081] Initialize the regression matrix , in this implementation, the regression matrix The initial value is 0.03703704* ,in for And each element is a 1 matrix, that is: (19) Initialize the regression intercept , in this implementation, the regression intercept The initial value of is 0; initialize the comprehensive information matrix of the sample set The initial values ​​are the sample sets in the modeling process. The sum of all samples in divided by the sample set The number of samples and sample set The sum of all samples in divided by the sample set The number of samples.

[0082] In this implementation, the objective function adjustment parameters The initial value of is 1, and the relaxation coefficient in the loss function The initial value of is 1, the loss function adjustment parameter The initial value of is 5.

[0083] Step 8: Alternate iterative calculation of the regression matrix , regression intercept , sample set comprehensive information matrix According to formulas (10), (13), (16), and (18), the iterations are repeated alternately. Each time formulas (10), (13), (16), and (18) are completed, it is a cycle. The output of the previous cycle is the input of the next cycle. The iteration process is repeated. When the number of iterations reaches the preset maximum number of iterations or the regression intercept of two consecutive iterations is , regression matrix , Comprehensive Information Matrix If the difference between is less than the preset difference threshold, the iteration is stopped and the current , , .

[0084] In this implementation, the regression matrix The initial value is 0.03703704* , regression intercept The initial value is 0, the sample set comprehensive information matrix The initial values ​​are the sample sets in the modeling process. The sum of all samples in divided by the sample set The number of samples and sample set The sum of all samples in divided by the sample set The number of samples, adjustment parameters The initial value of is 1, and the relaxation coefficient in the loss function The initial value of is 1, the loss function adjustment parameter The initial value of is 5, and when the number of iterations is 70, the discrimination effect is excellent.

[0085] Step 9: Calculate the predicted label of the input sample in the discriminant model and calculate the accuracy of the discriminant model. Calculate each sample In the regression matrix , regression intercept The regression labels under , compared with the actual sample labels, calculate the accuracy of the algorithm. If the accuracy is high and the discrimination effect is good, the regression matrix is ​​output , regression intercept ; Otherwise, reinitialize the regression matrix , regression intercept , sample set comprehensive information matrix , adjust the relaxation coefficient of the loss function And the loss function adjustment parameters , adjust the tuning parameters , and iterate again according to steps 7 and 8 until the discrimination effect reaches the expected requirements. The label prediction formula is: (20) In this modeling process, when the initial values ​​of the parameters are as in step 8 and the number of iterations is 70, the discrimination effect is excellent and the regression matrix The output value is as follows: Table 7, Regression intercept The output value is 3.27298333.

[0086] Regression Matrix for Matrix, for convenience of display, for display.

[0087] Table 7

[0088]

[0089] In this implementation, the predicted labels for the first 4 samples in the modeling process are calculated as follows:

[0090]

[0091]

[0092]

[0093] Step 10: Get the original test sample, process the non-digital features in the original test sample into digital features to obtain the input test sample, and calculate the input test sample to predict the sample label in the discriminant model. In this discriminant test process, the following steps are specifically included: Step 10.1: Input the original samples in the discrimination test process through step 1. The first four original test samples are shown in Tables 8 and 9.

[0094] Table 8

[0095]

[0096] Table 9

[0097]

[0098] Step 10.2: According to the digital dictionary previously established in step 2, the non-digital features in the input sample of the test process are processed into digital features to obtain the input sample of the discriminant test process. , and its corresponding label is , , , As shown in Table 10 below: Table 10

[0099]

[0100] Step 10.3: Regress the discriminant model matrix obtained in step 9 , regression intercept , in the process of calculating the discriminant test, the sample The predicted label .

[0101] The predicted labels of the first four samples of The calculation is as follows:

[0102]

[0103]

[0104]

[0105] Same as step 10.3, complete the label prediction of all samples during the discriminant test: , .

[0106] As can be seen from step 3, the total number of samples in this implementation is 600. There are 400 samples in the modeling process, of which the host storage link status of the first 200 samples is active, and the host storage link status of the last 200 samples has an inactive storage link. During the discrimination test, there are 200 samples in total, of which the host storage link status of the first 100 samples is active, and the host storage link status of the last 100 samples has an inactive storage link. For samples where the host storage link status is all active, the sample label is recorded as 1; for samples where the host storage link activity status has an inactive storage link, the sample label is recorded as -1. Therefore, during the discrimination test, the sample label is predicted 100% accurately, that is, the actual situation of the host storage link activity status represented by the sample is 100% accurately judged.

[0107] Corresponding to the aforementioned embodiment of a method for determining the activity status of a host storage link based on feature relaxation regression, the present invention further provides an embodiment of a device for determining the activity status of a host storage link based on feature relaxation regression.

[0108] See also Figure 6 An embodiment of the present invention provides a host storage link activity status determination device with feature relaxation regression, comprising a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it is used to implement a host storage link activity status determination method with feature relaxation regression in the above embodiment.

[0109] An embodiment of a host storage link activity state determination device with feature relaxation regression provided by the present invention can be applied to any device with data processing capability, and the device with data processing capability can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capability in which it is located reading the corresponding computer program instructions in the non-volatile memory into the internal memory for execution. From a hardware perspective, if Figure 6 As shown, it is a hardware structure diagram of any device with data processing capability where a host storage link activity state determination device with feature relaxation regression provided by the present invention is located, except Figure 6In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0110] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0111] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0112] An embodiment of the present invention further provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, a method for determining the active state of a host storage link with a feature relaxation regression in the above embodiment is implemented.

[0113] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0114] The present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, the method for determining the active state of a host storage link based on feature relaxation regression is implemented.

[0115] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for determining the active state of a host storage link based on feature relaxation regression, characterized in that: The method comprises the following steps: (1) Obtain the host name and host storage link status information through the cloud management platform operation data, obtain the original sample, and perform digital feature processing on the non-digital features; (2) Using the host storage link activity status in the original sample as a label, a sample set is obtained based on the label; (3) Reduce the dimension of the feature data of the original sample and set the objective function to perform linear regression on the sample label. The objective function includes: Feature regression, the loss function is set based on the degree of allowable deviation of the regression sample label as the relaxation coefficient; The regression intercept indicates the regression prediction value of the sample feature regression sample label when all sample feature values ​​are 0; Obtain the comprehensive information matrix of the sample set, which integrates all sample information in the sample set and has the smallest difference with the comprehensive information of the sample set; Select the easily distinguishable direction in the heterogeneous sample set, make the samples of different classes distinguishable after the sample dimension reduction based on the regression matrix, and maximize the difference of the comprehensive information matrix between heterogeneous samples; (4) The objective function is solved iteratively to obtain the discriminant model. The sample labels are predicted based on the discriminant model to obtain the host storage link activity status discrimination result.

2. According to the method for determining the active state of a host storage link based on feature relaxation regression according to claim 1, it is characterized in that: The digital feature processing of non-numeric features in the original sample data is specifically as follows: a digital dictionary of non-numeric feature dimensions is established to achieve a one-to-one mapping relationship between each dimensional feature and a number, and the digital processing of the features of the non-numeric feature dimensions in the original input sample is completed by digitally replacing the original features.

3. According to the method for determining the active state of a host storage link based on feature relaxation regression of claim 1, it is characterized in that: According to the actual situation of the host storage link activity status, the samples are manually marked as sample labels, and the corresponding sample sets are obtained based on the sample labels of the host storage link status being all active and the sample labels of the host storage link status having inactive links.

4. According to the method for determining the active state of a host storage link based on feature relaxation regression according to claim 1, it is characterized in that: The loss function of feature regression allows all sample features to have a certain offset when regressing sample labels. In the loss function of feature regression, the loss function relaxation coefficient is designed based on the range of allowable offsets of sample feature regression sample labels. The larger the value of the loss function relaxation coefficient, the larger the range of allowable offsets of sample feature regression sample labels. The loss function adjustment parameter is designed based on the degree of sample constraint on sample feature regression sample labels exceeding the allowable offset range. The larger the value of the loss function adjustment parameter, the greater the degree of constraint on sample feature regression sample labels exceeding the offset range, requiring all samples to be as close to the deviation range as possible.

5. According to the method for determining the active state of a host storage link based on feature relaxation regression of claim 4, it is characterized in that: Adjust the relaxation coefficient in the loss function and the loss function adjustment parameters to adjust the degree of relaxation of sample regression sample labels, and achieve sample feature relaxation regression sample labels.

6. According to the method for determining the active state of a host storage link based on feature relaxation regression of claim 1, it is characterized in that: The comprehensive information of the sample set is obtained by multiplying the sample weight coefficient in the sample set by the sample and summing them up. When the sample eigenvalue is missing or some eigenvalues ​​are N / A, the sample weight coefficient of the sample in the sample set is 0. When the sample eigenvalue is complete and not missing, the sample weight coefficient of the sample in the sample set is the inverse of the number of complete samples.

7. According to the method for determining the active state of a host storage link based on feature relaxation regression according to claim 1, it is characterized in that: Initialize the regression matrix, regression intercept and comprehensive information matrix of the sample set, and solve the objective function by alternating iteration. Stop the iteration when the number of iterations reaches the preset maximum number of iterations or the difference between the regression intercept, regression matrix and comprehensive information matrix of the sample set of two adjacent iterations is less than the preset difference threshold.

8. A host storage link activity state determination device based on feature relaxation regression, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, a method for determining the active state of a host storage link by feature relaxation regression according to any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, a method for determining the active state of a host storage link based on feature relaxation regression according to any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, a method for determining the active state of a host storage link based on characteristic relaxation regression as described in any one of claims 1 to 7 is implemented.

Images