Condition monitoring based device training safety analysis method and system

Abstract

The application relates to the technical field of equipment security analysis, and particularly discloses a device training security analysis method and system based on state monitoring, which comprises the following steps: dividing monitoring data into feature data groups and label data groups; extracting a preset number of data in the feature data groups, performing data expansion based on the extracted data, and obtaining virtual feature data groups; constructing sample sets of each extracted number of virtual feature data groups to label data groups, sending the sample sets to a model provider, and receiving a neural network model of each extracted number fed back by the model provider; calculating error rates and security degrees of the neural network model of each extracted number; receiving error rate requirements and security degree requirements input by a user, matching the neural network model, and taking the neural network model as an analysis model; and the application provides a data extraction and re-expansion scheme, a user can independently select how much real data is provided, and the information security of the equipment is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of equipment safety analysis technology, specifically a method and system for equipment training safety analysis based on condition monitoring. Background Technology

[0002] In the era of smart manufacturing and the Industrial Internet of Things (IIoT), equipment training (such as industrial applications of intelligent robots, automated machinery, or AI models) relies on a large amount of operational data. This data typically comes from equipment condition monitoring systems, such as sensors, control units, and SCADA (Supervisory Control and Data Acquisition) systems. After acquiring the condition data, the current output of the equipment can be obtained, and the relationship between the condition data and the output can be trained, for example, a neural network model. This model is used to analyze certain preset condition data and predict the equipment's output under that condition data.

[0003] However, the existing model training process is generally outsourced to third parties, who specialize in providing model training services. If the model requester provides all the operational data of the equipment, the third party can easily learn the specific details of the equipment, posing a certain risk of information leakage. Summary of the Invention

[0004] The purpose of this invention is to provide a device training security analysis method and system based on status monitoring to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for security analysis of device training based on state monitoring, the method comprising:

[0007] Establish a connection channel with the monitor in the equipment to obtain monitoring data containing the monitor number and monitoring time in real time;

[0008] The monitoring data is divided into feature data groups and tag data groups based on the monitor number;

[0009] A preset number of data points are extracted from the feature data group, and the extracted data is expanded to obtain a virtual feature data group; the monitoring time of the virtual feature data group adopts the monitoring time of the corresponding feature data group.

[0010] Construct a sample set from each number of virtual feature data groups to the label data groups, send it to the model provider, and receive the neural network model from the model provider for each number of extractions.

[0011] Calculate the error rate and security of the neural network model for each sample size;

[0012] It receives user input regarding error rate and security requirements, matches them with a neural network model, and uses this model as the analysis model.

[0013] As a further aspect of the present invention: the step of establishing a connection channel with the monitor in the device and acquiring monitoring data containing the monitor number and monitoring time in real time includes:

[0014] The system queries the registered intelligent modules in the device; the intelligent modules include sensors and controllers.

[0015] Acquire the operating data of each intelligent module; for any intelligent module, compare its operating data with the operating data of other intelligent modules within a preset distance range and calculate the correlation.

[0016] When the absolute value of the correlation reaches a preset threshold, only one of the intelligent modules of the two compared parties is retained;

[0017] After comparing and retaining all intelligent modules, they are used as monitors, and the monitors are numbered.

[0018] Establish a connection channel with the monitor in the equipment to obtain monitoring data containing the monitor number and monitoring time in real time.

[0019] As a further aspect of the present invention: the step of dividing the monitoring data into feature data groups and tag data groups according to the monitor number includes:

[0020] The system receives the target metrics input by the user based on a preset metric receiving port.

[0021] Query the monitor ID of the monitor corresponding to the target indicator and use it as the target ID. Use the monitor ID of the monitor that does not correspond to the target indicator as the non-target ID.

[0022] At any given time, collect the monitoring data corresponding to all target numbers and use them as a tag data group;

[0023] Collect all monitoring data corresponding to non-target numbers and use them as feature data groups;

[0024] The data arrangement order in both the label data group and the feature data group is based on the numbering order.

[0025] As a further aspect of the present invention: the step of extracting a preset number of data from the feature data group and expanding the data based on the extracted data to obtain a virtual feature data group includes:

[0026] The number of samples to be drawn is determined sequentially within a preset range; the number of samples to be drawn is an integer.

[0027] Select a feature data set, and randomly select a certain number of data points from the feature data set as known data;

[0028] Preserve the data locations of known data and create a transit data group; the dimensions of the transit data group are the same as the dimensions of the feature data group.

[0029] For unknown data in the transit data group, query the correlation between the unknown data and each known data; the correlation is the correlation of the intelligent module corresponding to the data;

[0030] Find the extreme values ​​in the historical data of known data, and predict unknown data based on the extreme values ​​and relevance;

[0031] Once all unknown data has been predicted, a virtual feature data set is obtained.

[0032] As a further aspect of the present invention: the process of querying the extreme values ​​in the historical data of known data, and predicting unknown data based on the extreme values ​​and relevance includes:

[0033] ;in, For predicted values ​​of unknown data, The total number of known data. For the first Known data, For the first The maximum value among the historical data of a known set of data. The maximum value in the historical data of the unknown data; Indicates the first The correlation between known data and currently predicted unknown data.

[0034] As a further aspect of the present invention: the step of calculating the error rate and security of the neural network model for each number of samples includes:

[0035] Obtain the error rate of the neural network model for each sample size;

[0036] The safety factor is calculated based on the inverse ratio of the number of samples taken.

[0037] The calculation process for the level of security is as follows:

[0038] In the formula, For safety reasons, The preset correction factor. Indicates the number of samples to be drawn.

[0039] The present invention also provides a device training security analysis system based on status monitoring, the system comprising:

[0040] The monitoring data acquisition module is used to establish a connection channel with the monitor in the device and acquire monitoring data containing the monitor number and monitoring time in real time.

[0041] The data segmentation module is used to segment monitoring data into feature data groups and tag data groups based on the monitor number;

[0042] The data expansion module is used to extract a preset number of data from the feature data group, and expand the data based on the extracted data to obtain a virtual feature data group; wherein, the monitoring time of the virtual feature data group adopts the monitoring time of the corresponding feature data group.

[0043] The sample construction application module is used to construct a sample set from each extracted number of virtual feature data groups to a label data group, send it to the model provider, and receive the neural network model from the model provider for each extracted number of samples.

[0044] The parameter calculation module is used to calculate the error rate and security of the neural network model for each number of samples.

[0045] The model matching module receives user input regarding error rate and security requirements, and matches them with neural network models as the analysis model.

[0046] As a further aspect of the present invention: the monitoring data acquisition module includes:

[0047] A data query unit is used to query the registered intelligent modules in the device; the intelligent modules include sensors and controllers.

[0048] The correlation calculation unit is used to acquire the operating data of each intelligent module, and for any intelligent module, compare the operating data of the intelligent module with the operating data of other intelligent modules within a preset distance range to calculate the correlation.

[0049] The module elimination unit is used to retain only one of the intelligent modules of the two compared parties when the absolute value of the relevance reaches a preset threshold.

[0050] The numbering unit is used to compare and retain all intelligent modules, and then use them as monitors to assign numbers to the monitors.

[0051] The acquisition execution unit is used to establish a connection channel with the monitor in the device and acquire monitoring data containing the monitor number and monitoring time in real time.

[0052] As a further aspect of the present invention: the data segmentation module includes:

[0053] The indicator receiving unit is used to receive the target indicator input by the user based on the preset indicator receiving port;

[0054] The number query unit is used to query the monitor number of the monitor corresponding to the target indicator, which is used as the target number, and the monitor number of the monitor corresponding to the non-target indicator is used as the non-target number.

[0055] The first statistical unit is used to collect monitoring data corresponding to all target numbers at any given time, and to form a tag data group.

[0056] The second statistical unit is used to count the monitoring data corresponding to all non-target numbers, as a feature data group.

[0057] The data arrangement order in both the label data group and the feature data group is based on the numbering order.

[0058] As a further aspect of the present invention: the data expansion module includes:

[0059] A quantity determination unit is used to sequentially determine the extraction quantity within a preset quantity range; the extraction quantity is an integer.

[0060] The data selection unit is used to select a feature data group and randomly select a certain number of data points from the feature data group as known data.

[0061] The data retention unit is used to retain the data location of known data and create a transit data group; the dimension of the transit data group is the same as the dimension of the feature data group.

[0062] The relevance query unit is used to query the relevance between unknown data and each known data in the transit data group; the relevance is based on the relevance of the intelligent module corresponding to the data.

[0063] The prediction execution unit is used to query the maximum and minimum values ​​in the historical data of known data, and predict unknown data based on the maximum and minimum values ​​and relevance.

[0064] The output unit is used to obtain a virtual feature data set after all unknown data has been predicted.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a data extraction and expansion scheme, which fuzzifies the samples provided by the model requester, so that the third party cannot know whether the data is completely real data. The third party's work is only to train the model based on the samples. After the model is trained, the model requester can obtain the error rate and security of each extraction quantity based on the real data stored by itself, and then select how much real data to provide, which greatly improves the information security of the device. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0067] Figure 1 A flowchart for training a security analysis method for condition-based monitoring devices.

[0068] Figure 2 A block diagram illustrating the structural composition of a security analysis system for equipment training based on condition monitoring. Detailed Implementation

[0069] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0070] Figure 1 This is a flowchart of a device training security analysis method based on state monitoring. In this embodiment of the invention, a device training security analysis method based on state monitoring includes:

[0071] Step S100: Establish a connection channel with the monitor in the device to obtain monitoring data containing the monitor number and monitoring time in real time;

[0072] Modules in the equipment that have data acquisition and data transmission functions are called monitors, such as sensors connected to the data transmission bus, or controllers of various components. During the operation of the equipment, these monitors acquire data and upload it to the equipment's management backend. When acquiring data, the monitor needs to insert the monitor number and monitoring time into the data. The monitor number is the monitor's identification, a pre-determined unique label; the monitoring time is the data acquisition time.

[0073] Step S200: Divide the monitoring data into feature data groups and tag data groups according to the monitor number;

[0074] Based on the monitor number, the monitoring data obtained from each monitor can be divided into feature data group and label data group. During model training, features generally refer to independent variables, and labels generally refer to dependent variables. The above segmentation process is essentially a segmentation of monitors, determining which monitors are regarded as independent variables and which are regarded as dependent variables. The function of the trained model will also be updated accordingly, indicating which monitors make data predictions for which other monitors.

[0075] Generally, when the equipment is production equipment, the monitors used as tags are mostly monitors at the equipment's exit point, used to characterize the final production status of the equipment. These monitors mainly acquire product information, not equipment information. For example, a monitor used as a feature can be a module in the equipment used to acquire the operating parameters of various components, while a monitor used as a tag is generally a module used to acquire product parameters, such as weight, volume, and surface roughness. In terms of data volume, the number of elements in a feature data group can be large, while the number of elements in a tag data group is generally small, usually only one.

[0076] Step S300: Extract a preset number of data from the feature data group, and expand the data based on the extracted data to obtain a virtual feature data group; wherein, the monitoring time of the virtual feature data group adopts the monitoring time of the corresponding feature data group.

[0077] The feature data group consists of data acquired by all monitors acting as tags. Its information volume is too large, and if this data is known to a third party (the model provider), the third party can fully understand the device's status, making its security relatively low. To address this, this application provides a data virtualization scheme. A preset amount of data is extracted from the feature data group; that is, a portion of the data is selected and then expanded to obtain a data group with the same dimensions as the feature data group, called a virtual feature data group. The virtual feature data group also contains time stamps, which represent the monitoring time of the feature data group.

[0078] An example of the above process is as follows: Suppose there are 10 monitors, the resulting feature data set will be 10-dimensional. When the number is 3, it means that 3 data points are extracted from the feature data set and their positions are retained. The data at the other 7 positions are unknown. Based on the 3 extracted data points, the unknown data is simulated. After the simulation is completed, the resulting data set is the virtual feature data set.

[0079] Step S400: Construct a sample set of virtual feature data groups to label data groups for each number of samples, send it to the model provider, and receive the neural network model for each number of samples from the model provider.

[0080] For any number of samples, the data extraction process is random (taking the above example as an example, the process of extracting data from 3 positions in a 10-dimensional data set is random). However, regardless of which data are extracted, data expansion will be performed to obtain a virtual feature data set. Therefore, this application directly uses the number of samples as an index to count the virtual feature data set for each number of samples. At the same time, the label data set at the same time is obtained. The label data is the prediction target and the dependent variable, and all data must be retained. After the number of samples is determined, a sample set from the virtual feature data set to the label data set is constructed, and it is divided into a training set and a test set according to a preset ratio to train the neural network model.

[0081] It should be noted that the premise for the virtual feature data group and the label data group to constitute a sample is that their times need to correspond. That is, the virtual feature data group is the independent variable of the label data group at the same moment. In addition, regarding the number of samples, each moment corresponds to one sample. A device has multiple moments, and the device itself can also have multiple moments. Therefore, the number of samples can be very large.

[0082] Step S500: Calculate the error rate and security of the neural network model for each number of samples;

[0083] During the training of a neural network model, the error rate is used to represent the difference between the real data and the output data. After the model trainer provides the model, the model requester can judge the error rate of the model based on their own real data. The larger the number of samples, the more real the data provided by the model requester, and the larger the error rate. At the same time, this application also needs to calculate the security level. The more samples are extracted, the more the operating status of the device is disclosed, and the lower the security level.

[0084] Step S600: Receive the error rate and security requirements input by the user, and match them with a neural network model as the analysis model;

[0085] Users are personnel who need to perform analysis and prediction. The system receives the error rate and safety requirements input by the user, compares them with the error rate and safety of the neural network model for each sampling quantity, and obtains the sampling quantity that meets the requirements. The neural network model corresponding to the sampling quantity is then read as the analysis model. At this time, the user inputs some operating parameters of the equipment (generally simulated operating parameters in the prediction stage), and the system can obtain the label data under those operating parameters. While meeting the safety requirements, the system provides analysis functions.

[0086] It is worth mentioning that the user in this application refers to the model requester, who is the party providing the data, while the third party is the model provider, who is the party training the model based on the data.

[0087] Regarding step S100, the step of establishing a connection channel with the monitor in the device and acquiring monitoring data containing the monitor number and monitoring time in real time includes:

[0088] The system queries the registered intelligent modules in the device; the intelligent modules include sensors and controllers.

[0089] Acquire the operating data of each intelligent module; for any intelligent module, compare its operating data with the operating data of other intelligent modules within a preset distance range and calculate the correlation.

[0090] When the absolute value of the correlation reaches a preset threshold, only one of the intelligent modules of the two compared parties is retained;

[0091] After comparing and retaining all intelligent modules, they are used as monitors, and the monitors are numbered.

[0092] Establish a connection channel with the monitor in the equipment to obtain monitoring data containing the monitor number and monitoring time in real time.

[0093] In one embodiment of the technical solution of this invention, a monitor selection scheme is provided. A device may contain many intelligent modules that can serve as monitors. The system queries the registered intelligent modules in the device, obtains the operating data of each intelligent module, and compares the operating data of any given intelligent module with the operating data of other intelligent modules within a preset distance range to calculate the correlation. When two intelligent modules are sufficiently correlated, one intelligent module is removed, leaving only one. This process is continuously repeated to optimize the number of monitors, reduce the number of selected monitors, and alleviate the pressure of sample acquisition and analysis. Specifically, this optimizes the dimension of the feature data group. After selecting a monitor, it is assigned a number.

[0094] It is worth mentioning that there are two possibilities for two smart modules to be sufficiently correlated: one is positive correlation, and the other is negative correlation. This application considers both of these cases as sufficiently correlated. Therefore, the method for judging the degree of correlation in the above content is to judge whether the absolute value of the correlation reaches a preset threshold.

[0095] Regarding step S200, the step of dividing the monitoring data into feature data groups and tag data groups according to the monitor number includes:

[0096] The system receives the target metrics input by the user based on a preset metric receiving port.

[0097] Query the monitor ID of the monitor corresponding to the target indicator and use it as the target ID. Use the monitor ID of the monitor that does not correspond to the target indicator as the non-target ID.

[0098] At any given time, collect the monitoring data corresponding to all target numbers and use them as a tag data group;

[0099] Collect all monitoring data corresponding to non-target numbers and use them as feature data groups;

[0100] The data arrangement order in both the label data group and the feature data group is based on the numbering order.

[0101] In one example of the technical solution of this invention, the segmentation process of the monitor is described. The user inputs an indicator into the preset indicator receiving port as the target indicator, queries the monitor number of the monitor corresponding to the target indicator as the target number, and takes the monitor number of the monitor corresponding to the non-target indicator as the non-target number. Thus, the monitor is segmented. After the segmentation is completed, the monitoring data corresponding to the target number at the same time is the tag data group, and the monitoring data corresponding to all non-target numbers is the feature data group.

[0102] In the above content, whether it is the tag data group or the feature data group, the correspondence between the element positions and the monitor is preset and fixed.

[0103] Regarding step S300, the step of extracting a preset number of data from the feature data group and expanding the data based on the extracted data to obtain a virtual feature data group includes:

[0104] The number of samples to be drawn is determined sequentially within a preset range; the number of samples to be drawn is an integer.

[0105] Select a feature data set, and randomly select a certain number of data points from the feature data set as known data;

[0106] Preserve the data locations of known data and create a transit data group; the dimensions of the transit data group are the same as the dimensions of the feature data group.

[0107] For unknown data in the transit data group, query the correlation between the unknown data and each known data; the correlation is the correlation of the intelligent module corresponding to the data;

[0108] Find the extreme values ​​in the historical data of known data, and predict unknown data based on the extreme values ​​and relevance;

[0109] Once all unknown data has been predicted, a virtual feature data set is obtained.

[0110] The key point of this application is to first extract data from the feature data set and then expand the data, the process of which is as follows:

[0111] Within a preset range, the number of samples is determined sequentially. For example, if the range is 1 to 10, the number of samples is determined sequentially as 1, 2, 3, ..., 10. After determining the number of samples, for any feature data group, a number of data points are randomly selected from the feature data group. For example, if the number of samples is 3, 3 data points are selected from any feature data group. The selected data points are taken as known data, and the unselected data points are called unknown data.

[0112] For unknown data in the transit data group, query the correlation between the unknown data and each known data. When executing step S200, the correlation between the monitors of each data location has been calculated, so it can be read directly. Query the extreme values ​​in the historical data of the known data, and predict the unknown data based on the extreme values ​​and correlation.

[0113] The process of querying the extreme values ​​in the historical data of known data and predicting unknown data based on the extreme values ​​and relevance includes:

[0114] ;in, For predicted values ​​of unknown data, The total number of known data. For the first Known data, For the first The maximum value among the historical data of a known set of data. The maximum value in the historical data of the unknown data; Indicates the first The correlation between known data and currently predicted unknown data.

[0115] In one example of the technical solution of this invention, the prediction process for unknown data is described. It uses the mean of the predicted data for all known data to the unknown data. The prediction process for each known data to the unknown data is as follows: calculate the ratio of the known data to the maximum value in the historical data. If the correlation is 1, then multiply the ratio by the maximum value in the historical data of the unknown data to obtain the predicted value. However, the correlation is not necessarily 1. Therefore, the correlation parameter is introduced in the multiplication process. The correlation uses the Pearson correlation coefficient.

[0116] Regarding step S500, the step of calculating the error rate and security of the neural network model for each number of samples includes:

[0117] Obtain the error rate of the neural network model for each sample size;

[0118] The safety factor is calculated based on the inverse ratio of the number of samples taken.

[0119] In one example of the technical solution of this invention, for each number of neural network models extracted, after the model trainer provides the model, the model requester can determine the error rate of the model based on its own real data; on this basis, the safety degree is calculated based on the number of extractions, and the safety degree is inversely proportional to the number of extractions; after processing, the error rate and safety degree of each number of neural network models extracted are obtained.

[0120] The calculation process for the level of security is as follows:

[0121] In the formula, For safety reasons, The preset correction factor. Indicates the number of samples to be drawn.

[0122] The calculation process of the safety factor is not complicated. The safety factor is inversely proportional to the number of samples. The reason for using an exponential function and a composite function with a negative coefficient and positive proportion is that as the number of samples increases, the decrease in safety factor becomes smaller. The principle is that the larger the number of samples, the closer the derivative is to zero.

[0123] Figure 2 The present invention provides a structural block diagram of a device training security analysis system based on state monitoring. The system 10 includes:

[0124] The monitoring data acquisition module 11 is used to establish a connection channel with the monitor in the device and acquire monitoring data containing the monitor number and monitoring time in real time.

[0125] The data segmentation module 12 is used to segment the monitoring data into feature data groups and tag data groups according to the monitor number;

[0126] Data expansion module 13 is used to extract a preset number of data from the feature data group, and expand the data based on the extracted data to obtain a virtual feature data group; wherein, the monitoring time of the virtual feature data group adopts the monitoring time of the corresponding feature data group.

[0127] The sample construction application module 14 is used to construct a sample set from each number of virtual feature data groups to a label data group, send it to the model provider, and receive the neural network model for each number of samples from the model provider.

[0128] The parameter calculation module 15 is used to calculate the error rate and security of the neural network model for each number of samples.

[0129] The model matching module 16 is used to receive the error rate requirements and security requirements input by the user, and match them with a neural network model as the analysis model.

[0130] Furthermore, the monitoring data acquisition module 11 includes:

[0131] A data query unit is used to query the registered intelligent modules in the device; the intelligent modules include sensors and controllers.

[0132] The correlation calculation unit is used to acquire the operating data of each intelligent module, and for any intelligent module, compare the operating data of the intelligent module with the operating data of other intelligent modules within a preset distance range to calculate the correlation.

[0133] The module elimination unit is used to retain only one of the intelligent modules of the two compared parties when the absolute value of the relevance reaches a preset threshold.

[0134] The numbering unit is used to compare and retain all intelligent modules, and then use them as monitors to assign numbers to the monitors.

[0135] The acquisition execution unit is used to establish a connection channel with the monitor in the device and acquire monitoring data containing the monitor number and monitoring time in real time.

[0136] Specifically, the data segmentation module 12 includes:

[0137] The indicator receiving unit is used to receive the target indicator input by the user based on the preset indicator receiving port;

[0138] The number query unit is used to query the monitor number of the monitor corresponding to the target indicator, which is used as the target number, and the monitor number of the monitor corresponding to the non-target indicator is used as the non-target number.

[0139] The first statistical unit is used to collect monitoring data corresponding to all target numbers at any given time, and to form a tag data group.

[0140] The second statistical unit is used to count the monitoring data corresponding to all non-target numbers, as a feature data group.

[0141] The data arrangement order in both the label data group and the feature data group is based on the numbering order.

[0142] Furthermore, the data expansion module 13 includes:

[0143] A quantity determination unit is used to sequentially determine the extraction quantity within a preset quantity range; the extraction quantity is an integer.

[0144] The data selection unit is used to select a feature data group and randomly select a certain number of data points from the feature data group as known data.

[0145] The data retention unit is used to retain the data location of known data and create a transit data group; the dimension of the transit data group is the same as the dimension of the feature data group.

[0146] The relevance query unit is used to query the relevance between unknown data and each known data in the transit data group; the relevance is based on the relevance of the intelligent module corresponding to the data.

[0147] The prediction execution unit is used to query the maximum and minimum values ​​in the historical data of known data, and predict unknown data based on the maximum and minimum values ​​and relevance.

[0148] The output unit is used to obtain a virtual feature data set after all unknown data has been predicted.

[0149] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention’s specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for analyzing device training security based on state monitoring, characterized in that, The method includes: Establish a connection channel with the monitor in the equipment to obtain monitoring data containing the monitor number and monitoring time in real time; The monitoring data is divided into feature data groups and tag data groups based on the monitor number; A preset number of data points are extracted from the feature data group, and the extracted data is expanded to obtain a virtual feature data group; the monitoring time of the virtual feature data group adopts the monitoring time of the corresponding feature data group. Construct a sample set from each number of virtual feature data groups to the label data groups, send it to the model provider, and receive the neural network model from the model provider for each number of extractions. Calculate the error rate and security of the neural network model for each sample size; It receives user input regarding error rate and security requirements, matches them with a neural network model, and uses this model as the analysis model.

2. The device training security analysis method based on state monitoring according to claim 1, characterized in that, The steps of establishing a connection channel with the monitor in the device and acquiring monitoring data containing the monitor number and monitoring time in real time include: The system queries the registered intelligent modules in the device; the intelligent modules include sensors and controllers. Acquire the operating data of each intelligent module; for any intelligent module, compare its operating data with the operating data of other intelligent modules within a preset distance range and calculate the correlation. When the absolute value of the correlation reaches a preset threshold, only one of the intelligent modules of the two compared parties is retained; After comparing and retaining all intelligent modules, they are used as monitors, and the monitors are numbered. Establish a connection channel with the monitor in the equipment to obtain monitoring data containing the monitor number and monitoring time in real time.

3. The device training security analysis method based on state monitoring according to claim 1, characterized in that, The step of dividing the monitoring data into feature data groups and tag data groups according to the monitor number includes: The system receives the target metrics input by the user based on a preset metric receiving port. Query the monitor ID of the monitor corresponding to the target indicator and use it as the target ID. Use the monitor ID of the monitor that does not correspond to the target indicator as the non-target ID. At any given time, collect the monitoring data corresponding to all target numbers and use them as a tag data group; Collect all monitoring data corresponding to non-target numbers and use them as feature data groups; The data arrangement order in both the label data group and the feature data group is based on the numbering order.

4. The device training security analysis method based on state monitoring according to claim 2, characterized in that, The steps of extracting a preset number of data from the feature data group and expanding the data based on the extracted data to obtain a virtual feature data group include: The number of samples to be drawn is determined sequentially within a preset range; the number of samples to be drawn is an integer. Select a feature data set, and randomly select a certain number of data points from the feature data set as known data; Preserve the data locations of known data and create a transit data group; the dimensions of the transit data group are the same as the dimensions of the feature data group. For unknown data in the transit data group, query the correlation between the unknown data and each known data; the correlation is based on the correlation of the intelligent module corresponding to the data. Find the extreme values ​​in the historical data of known data, and predict unknown data based on the extreme values ​​and relevance; Once all unknown data has been predicted, a virtual feature data set is obtained.

5. The device training security analysis method based on state monitoring according to claim 4, characterized in that, The process of querying the maximum and minimum values ​​in the historical data of known data, and predicting unknown data based on the maximum and minimum values ​​and relevance includes: ;in, For predicted values ​​of unknown data, The total number of known data. For the first Known data, For the first The maximum value among the historical data of a known set of data. The maximum value in the historical data of the unknown data; Indicates the first The correlation between known data and currently predicted unknown data.

6. The device training security analysis method based on state monitoring according to claim 1, characterized in that, The steps for calculating the error rate and security of the neural network model for each sampling number include: Obtain the error rate of the neural network model for each sample size; The safety factor is calculated based on the inverse ratio of the number of samples taken. The calculation process for the level of security is as follows: In the formula, For safety reasons, The preset correction factor. Indicates the number of samples to be drawn.

7. A device training security analysis system based on status monitoring, characterized in that, The system includes: The monitoring data acquisition module is used to establish a connection channel with the monitor in the device and acquire monitoring data containing the monitor number and monitoring time in real time. The data segmentation module is used to segment monitoring data into feature data groups and tag data groups based on the monitor number; The data expansion module is used to extract a preset number of data from the feature data group, and expand the data based on the extracted data to obtain a virtual feature data group; wherein, the monitoring time of the virtual feature data group adopts the monitoring time of the corresponding feature data group. The sample construction application module is used to construct a sample set from each extracted number of virtual feature data groups to a label data group, send it to the model provider, and receive the neural network model from the model provider for each extracted number of samples. The parameter calculation module is used to calculate the error rate and security of the neural network model for each number of samples. The model matching module receives user input regarding error rate and security requirements, and matches them with neural network models as the analysis model.

8. The device training security analysis system based on state monitoring according to claim 7, characterized in that, The monitoring data acquisition module includes: A data query unit is used to query the registered intelligent modules in the device; the intelligent modules include sensors and controllers. The correlation calculation unit is used to acquire the operating data of each intelligent module, and for any intelligent module, compare the operating data of the intelligent module with the operating data of other intelligent modules within a preset distance range to calculate the correlation. The module elimination unit is used to retain only one of the intelligent modules of the two compared parties when the absolute value of the relevance reaches a preset threshold. The numbering unit is used to compare and retain all intelligent modules, and then use them as monitors to assign numbers to the monitors. The acquisition execution unit is used to establish a connection channel with the monitor in the device and acquire monitoring data containing the monitor number and monitoring time in real time.

9. The device training security analysis system based on state monitoring according to claim 7, characterized in that, The data segmentation module includes: The indicator receiving unit is used to receive the target indicator input by the user based on the preset indicator receiving port; The number query unit is used to query the monitor number of the monitor corresponding to the target indicator, which is used as the target number, and the monitor number of the monitor corresponding to the non-target indicator is used as the non-target number. The first statistical unit is used to collect monitoring data corresponding to all target numbers at any given time, and to form a tag data group. The second statistical unit is used to count the monitoring data corresponding to all non-target numbers, as a feature data group. The data arrangement order in both the label data group and the feature data group is based on the numbering order.

10. The device training security analysis system based on state monitoring according to claim 8, characterized in that, The data expansion module includes: A quantity determination unit is used to sequentially determine the extraction quantity within a preset quantity range; the extraction quantity is an integer. The data selection unit is used to select a feature data group and randomly select a certain number of data points from the feature data group as known data. The data retention unit is used to retain the data location of known data and create a transit data group; the dimension of the transit data group is the same as the dimension of the feature data group. The relevance query unit is used to query the relevance between unknown data and each known data in the transit data group; the relevance is based on the relevance of the intelligent module corresponding to the data. The prediction execution unit is used to query the maximum and minimum values ​​in the historical data of known data, and predict unknown data based on the maximum and minimum values ​​and relevance. The output unit is used to obtain a virtual feature data set after all unknown data has been predicted.

Images