Machine learning based device operational status self-checking method and system

Abstract

The present application relates to the technical field of data processing, more particularly, the present application relates to a device running state self-checking method and system based on machine learning, the method comprising: obtaining a current sequence of a monitoring device, using the noise level of the current sequence to adaptively adjust the subsample size of the SCIForest algorithm, and the subsample size is negatively correlated with the noise level; randomly selecting several data sets with the same size as the subsample size from the current sequence, using the SCIForest algorithm to construct a decision tree, and deploying the trained decision tree to the monitoring device to determine the running state of the monitoring device based on the comparison result of the abnormal score determined by the trained decision tree and the preset value. The present application can improve the efficiency of the running state self-checking of the monitoring device, and at the same time improve the accuracy of the self-checking result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing. More particularly, the present application relates to a device running state self-checking method and system based on machine learning. BACKGROUND

[0002] The self-checking function of monitoring the running state of a device refers to the device automatically detecting and diagnosing faults according to its own running parameters. With the continuous development of monitoring devices, real-time running state self-checking can effectively improve device efficiency and reduce the probability of faults. For a monitoring device, anomalies in current data are often a sign that the device is about to malfunction. By monitoring anomalies in current data in real time, anomalies in device operation can be discovered in a timely manner, so that preventive measures can be taken to avoid faults.

[0003] In related technologies, a device power-off monitoring method and system are disclosed in patent application file CN118708936A. The method includes: collecting device performance data corresponding to a regional device in real time, extracting power-related data in the device performance data; capturing the time sequence correlation corresponding to the power-related data, extracting the time domain features corresponding to the power-related data; constructing a data isolation tree corresponding to the power-related data, calculating an abnormal score value corresponding to the power-related data; identifying abnormal power data in the power-related data, and filtering out abnormal devices in the regional device according to the abnormal power data.

[0004] However, the above scheme does not consider the influence of noise when identifying abnormal power data. In actual applications, the performance data of a device is often disturbed by factors such as environmental changes, power fluctuations, and sensor errors, resulting in noise. Noise can disturb the normal fluctuations of data, thereby affecting the accuracy of the anomaly detection model. Due to the strong randomness and volatility of noise, it may be misjudged as normal variation, thereby affecting the construction of the isolation tree, leading to normal data being misjudged as abnormal, or failing to correctly identify the real abnormal data, thereby affecting the judgment of the device state. SUMMARY

[0005] In order to solve the problem of being unable to accurately identify the abnormal running state of a device, the present application provides a device running state self-checking method and system based on machine learning.

[0006] According to a first aspect of the present application, a device running state self-checking method based on machine learning is provided, comprising:

[0007] Obtaining a current sequence of a monitoring device;

[0008] Adaptively adjusting the subsample size of the SCIForest algorithm using the noise level of the current sequence, the subsample size being negatively correlated with the noise level;

[0009] Randomly select several data sets of the same size as the subsample from the current sequence, use the SCIForest algorithm to build a decision tree, and deploy the trained decision tree to the monitoring device to self-check the status of the monitoring device based on the comparison results of the anomaly score determined by the trained decision tree and the preset value;

[0010] The noise level is obtained by calculating the noise performance of any data in the current series. The noise performance is positively correlated with the fitting loss of the data within a preset local range centered on the data, as well as the ratio of the range of the data within the local range to the number of data categories.

[0011] With the goal of maximizing the between-class variance, the data in the current series are divided into two categories. The average noise performance of the current series is weighted by the ratio of the data volume of the category with larger noise performance to the other category to obtain the noise level of the current series.

[0012] The anomaly detection algorithm employed in the decision tree construction of this invention effectively addresses the poor performance of the isolation forest algorithm in detecting cluster anomalies. It can also process massive amounts of data, improving the efficiency of anomaly detection. Furthermore, adaptively adjusting the subsample size based on the noise level of the current sequence enhances the noise immunity of anomaly detection and improves computational accuracy. This ensures the accuracy of anomaly detection results and enhances the accuracy of self-tests of the operating status of monitoring equipment.

[0013] Preferably, the noise performance of any data in the current sequence satisfies the following relationship:

[0014] ;

[0015] Where, The first in the current sequence Noise performance of individual data; is the data sequence number in the current sequence; For the The value range of data within the local range of each data; For the The number of data categories within the local range of each data; For the The local range of the data The fitted value of the data; For the The local range of the data The value of the data; For the The total number of data within the local range of each data; is the absolute value symbol; is the hyperbolic tangent function.

[0016] The application can avoid misjudging normal data changes as noise-induced changes, and ensures the accuracy of noise performance.

[0017] Preferably, the method for obtaining the fitting value comprises:

[0018] The least square method is used to perform curve fitting on the data in the local range, to obtain the fitting value of each data in the local range.

[0019] Preferably, the size of the sub-sample of the SCIForest algorithm is adaptively adjusted according to the noise level of the current sequence, and the following relationship is satisfied:

[0020] ;

[0021] In the formula, is the adjusted size of the sub-sample; is a preset initial size of the sub-sample; is the noise level of the current sequence; is an exponential function with a natural constant as the base; is a rounding-up symbol.

[0022] The size of the sub-sample is adaptively adjusted, when the noise level is high, the influence of noise can be reduced by reducing the size of the sub-sample; when the noise level is high, the noise can be avoided from being ignored by increasing the size of the sub-sample, so that the calculation accuracy can be improved.

[0023] Preferably, the method for obtaining the anomaly score comprises:

[0024] The current current is obtained, and all the trained decision trees are traversed using the current current to obtain the path length of the current current in each trained decision tree.

[0025] The average value of the path length of the current current in the trained decision tree is calculated and normalized to obtain the anomaly score of the current current.

[0026] Preferably, the comparison result between the anomaly score determined based on the trained decision tree and the preset value is used to self-check the state of the monitoring device, comprising:

[0027] The preset value is obtained, and when the anomaly score is greater than or equal to the preset value, it is determined that the running state of the monitoring device is abnormal at the current time;

[0028] When the anomaly score is less than the preset value, it is determined that the running state of the monitoring device is normal at the current time.

[0029] The application can accurately identify the abnormality of the running state of the monitoring device.

[0030] Preferably, the method further comprises:

[0031] a preset update period;

[0032] In one update period, the abnormal score of each sampling time in the second half period of the update period is determined by using the decision tree constructed by the current sequence collected in the first half period of the update period.

[0033] The application introduces a dynamic update period mechanism to avoid the influence of current trend difference on the detection result, can more accurately track the trend change of current fluctuation, avoid misjudgment caused by different trends, and improve the accuracy of the self-check result of the monitoring device running state.

[0034] According to the second aspect of the application, a device running state self-check system based on machine learning is provided, the system comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to realize the steps of the first aspect of the application.

[0035] The application has the following effects:

[0036] 1. The abnormality detection algorithm used in the application can improve the self-check efficiency of the monitoring device running state by constructing multiple decision trees to process and analyze data in parallel.

[0037] 2. The application can improve the noise resistance of the abnormality detection algorithm by adaptively adjusting the sub-sample size in combination with the noise level of the current sequence, and when the noise level is high, a smaller sub-sample size can be set to reduce the interference of noise; when the noise level is low, a larger sub-sample size can be set to avoid noise being ignored and improve the calculation accuracy, so that the abnormal running state of the monitoring device can be accurately identified, and the accuracy of the running state self-check result of the monitoring device can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0038] The above and other objects, features and advantages of the exemplary embodiments of the application will be more apparent from the following detailed description read in conjunction with the accompanying drawings, in which several embodiments of the application are shown by way of example, and wherein like reference numerals refer to like parts throughout the several views, and wherein:

[0039] Figure 1 is a step flow diagram of the device running state self-check method based on machine learning of the embodiment of the application. DETAILED DESCRIPTION

[0040] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0041] The specific implementation of the present application will be described in detail below with reference to the drawings.

[0042] With reference to Figure 1 The device operation state self-checking method based on machine learning comprises steps S1-S3, and specifically as follows:

[0043] S1: Obtain a current sequence of a monitored device.

[0044] Specifically, a current sensor electrically connected to the monitored device can be used to collect the current of the monitored device at a certain frequency, such as 10 Hz, to obtain the current sequence of the monitored device. The present application does not particularly limit the data collection frequency.

[0045] S2: Use the noise level of the current sequence to adaptively adjust the subsample size of the SCIForest algorithm, and the subsample size is negatively correlated with the noise level.

[0046] It should be noted that the SCIForest algorithm is a commonly used isolation-based machine learning anomaly detection algorithm based on the improvement of the IForest algorithm. In view of the poor performance of the original algorithm in detecting cluster anomalies, SCIForest proposes a new segmentation standard function and a random hyperplane construction method. In the SCIForest algorithm, the size of the subsample set is usually fixed. However, since the monitored device is usually located outdoors or in a noisy environment, the current may be disturbed by noise to varying degrees, and the influence of noise changes over time, which is uncertain. A subsample set of fixed size may not be able to fully adapt to such noise fluctuations, resulting in inaccurate anomaly detection results, i.e. noise may be misjudged as abnormal, or normal data may be misjudged as abnormal due to noise. Therefore, the present application improves the traditional SCIForest algorithm, dynamically adjusts the subsample size of the SCIForest algorithm by analyzing the noise level of the current sequence, thereby improving the anti-noise ability of the algorithm, and further improving the accuracy and efficiency of the current anomaly detection of the monitored device.

[0047] Among them, the noise level refers to data that can reflect the influence of noise on the current sequence. For example, when the current sequence is greatly affected by noise, the noise level of the current sequence is high, and vice versa, if the current sequence is less affected by noise, the noise level of the current sequence is low.

[0048] Optionally, when the noise level of the current sequence is large, reducing the subsample size of the SCIForest algorithm can reduce the calculation amount while reducing the influence of noise on the SCIForest algorithm; when the noise level of the current sequence is small, increasing the subsample size of the SCIForest algorithm can improve the calculation accuracy of the SCIForest algorithm.

[0049] Specifically, the determination of the noise level of the current sequence can be achieved by the following steps:

[0050] Step one: calculate the noise performance of any data in the current sequence, the noise performance is positively correlated with the fitting loss of the data in the preset local range centered on the data, and the ratio of the value range of the data in the local range to the number of categories of the data;

[0051] It should be noted that due to the complex working environment of the monitoring device, there are often various types of noise in the current sequence collected by the current sensor, such as random noise caused by internal electronic components of the device, and burst noise caused by external environmental interference. These noises usually cause the current data to fluctuate sharply within a certain range. Therefore, the application calculates the ratio of the value range of the data in the local range to the number of categories of the data, which can evaluate the possibility of the data being noise.

[0052] It should be further noted that the load change of the monitoring device can cause the current to fluctuate greatly, thereby misjudging the normal data change as fluctuation caused by noise. However, since normal data change usually shows gradual change of current in a short time, the change trend is relatively regular, while noise usually shows random fluctuation. Therefore, the application uses this feature to measure the fitting difficulty of the data in the local range by the fitting loss of the data in the local range, so as to distinguish the reason for the change of the data and avoid misjudging the normal data change as fluctuation caused by noise.

[0053] Wherein, the noise performance refers to the possibility of the data being noise. The local range refers to a data segment centered on any data, containing a plurality of data, such as 15 data. In this embodiment, the data contained in the local range is 15, and the amount of data contained in the local range is not particularly limited in this embodiment.

[0054] Optionally, when the length of the data segment on either side of any data in the current sequence is less than 7, the data segment on that side is discarded, and the continuous data segment composed of the data and the 7 data on the other side is taken as the local range of the data.

[0055] Specifically, the noise performance of any data in the current sequence satisfies the following relationship:

[0056] ;

[0057] Where, The first in the current sequence Noise performance of individual data; is the data sequence number in the current sequence; For the The value range of data within the local range of each data; For the The number of categories of data in the local range of data, where the number of categories refers to the number of different values ​​in the local range. For example, when the data in the local range is {1,3,4,2,1,3}, the number of categories of data in the local range is 4, that is, {1,3,4,2}; For the The local range of the data The fitted value of the data; For the The local range of the data The value of the data; For the The total number of data within the local range of each data; is the absolute value symbol; is the hyperbolic tangent function.

[0058] in, , For the The maximum value of the data in the local range of the data, For the The minimum value of the data in the local range of the data.

[0059] Reflects the first The change intensity of the data in the local range of the data is greater. The greater the value, the more drastic the change of the data in the local range. On the contrary, the smaller the value is, the more likely the data in the local range is to be noise. The smaller the possibility that a piece of data is noise, the smaller the noise performance of the corresponding data.

[0060] The first in the current sequence The average fitting loss of the data in the local range of the data, the greater the value, the more difficult the fitting of the data in the local range, that is, the change of the data in the local range is more random, which indicates that the change intensity in the local range is more likely to be caused by noise; on the contrary, the smaller the value, the smaller the fitting difficulty of the data in the local range, that is, the change of the data in the local range has regularity, which indicates that the change intensity in the local range is more likely to be normal current change.

[0061] In another embodiment, the noise performance can also be calculated in other ways, such as not determined by a product calculation formula, but determined by a summation calculation formula.

[0062] In an example embodiment of the present application, the fitting value of the data can be determined by the following steps:

[0063] The least square method is used to perform curve fitting on the data in the local range to obtain the fitting value of each data in the local range.

[0064] Optionally, the ridge regression method can also be used to perform curve fitting on the data in the local range, and of course, a suitable method can also be selected according to the specific situation to perform curve fitting, and the selected fitting method is not particularly limited in the present embodiment. It should be noted that the least square fitting method for data is a prior art, which will not be described in detail herein.

[0065] Step 2: The data in the current sequence is divided into two categories with the goal of maximizing the inter-class variance, and the average noise performance of the current sequence is weighted by the ratio of the data amount of the category with greater noise performance to the other category to obtain the noise level of the current sequence.

[0066] It should be noted that the average noise performance of the current sequence can only reflect the overall noise performance of the current sequence, and cannot reflect the real distribution of the noise in the current sequence. For example, when the data with greater noise performance is less and the data with smaller noise performance is more in the current sequence, the overall noise performance of the current sequence will be low after taking the average, which may be misjudged as the current sequence being less affected by noise, resulting in a lower sub-sample size of the SCIForest algorithm. Therefore, the current sequence is divided into two categories by the inter-class variance maximization method, and the average noise performance of the current sequence is weighted by the ratio of the data amount of the category with greater noise performance to the other category to accurately evaluate the influence of the current sequence by noise.

[0067] Specifically, the noise level of the current sequence satisfies the following relationship:

[0068] ;

[0069] Where, is the noise level of the current sequence; is the average value of the noise performance of the data in the current series; The amount of data in the category with the largest noise in the classification results; It is the amount of data in the category with less noise in the classification results.

[0070] Furthermore, after determining the noise level of the current sequence, the subsample size of the SCIForest algorithm can be adaptively adjusted using the noise level of the current sequence. Specifically, the subsample size of the SCIForest algorithm is adaptively adjusted using the noise level of the current sequence to satisfy the following relationship:

[0071] ;

[0072] Where, is the adjusted subsample size; To preset the initial subsample size, the present invention =100; is the noise level of the current sequence; For the natural constant An exponential function with base ; The symbol for rounding up.

[0073] It should be noted that the initial subsample size set in the present invention is a larger value. The present invention only adjusts the initial subsample size downward. Of course, a suitable initial subsample size can also be set according to specific circumstances. This embodiment does not impose any special restrictions on the setting of the initial subsample size.

[0074] S3: Randomly select several data sets with the same size as the subsample from the current sequence, use the SCIForest algorithm to build a decision tree, and deploy the trained decision tree to the monitoring device to self-check the status of the monitoring device based on the comparison results of the anomaly score determined by the trained decision tree and the preset value.

[0075] Optionally, when randomly selecting several data sets of the same size as the subsample from the current sequence, the data sets may be selected using simple random sampling or stratified sampling. The present invention does not particularly limit the selected sampling method. It should be noted that the process of constructing a decision tree based on the selected data set using the SCIForest algorithm and training the constructed decision tree is a prior art and is not described in detail in this embodiment.

[0076] The integration scale set in the present invention is 50, that is, the number of decision trees constructed is 50, that is, the number of data sets selected is 50.

[0077] In an example embodiment of the present application, the determination of the anomaly score can be achieved by the following steps:

[0078] The current current is obtained, and the current current is used to traverse all the trained decision trees to obtain the path length of the current current in each trained decision tree; the average value of the path length of the current current in the trained decision trees is calculated and normalized to obtain the anomaly score of the current current.

[0079] In the SCIForest algorithm, the process of converting the average value of the path length into an anomaly score by normalization is prior art, which will not be described in detail in this embodiment.

[0080] In an example embodiment of the present application, the self-checking of the monitoring device can be achieved by the following steps:

[0081] The preset value is obtained, and when the anomaly score is greater than or equal to the preset value, it is determined that the running state of the monitoring device is abnormal at the current time; when the anomaly score is less than the preset value, it is determined that the running state of the monitoring device is normal at the current time.

[0082] Optionally, the preset value can be set to 0.56, and if it is determined that the anomaly score of the current current is greater than or equal to 0.56, it is determined that the current current is abnormal, which indicates that the monitoring device is in an abnormal running state at the current time, and an alarm or a corresponding processing mechanism can be triggered to handle the fault in time.

[0083] Optionally, if the anomaly score of the current current is less than 0.56, it is determined that the current current is normal data, which indicates that the monitoring device is in a normal running state at the current time, thereby achieving self-checking of the running state of the monitoring device.

[0084] In an example embodiment of the present application, it further comprises:

[0085] A preset update period; in an update period, the decision trees constructed by the current sequence collected in the first half period of the update period are used to determine the anomaly scores of each sampling time in the second half period of the update period.

[0086] For example, the update period can be set to 2 For any update period, the current data in the first period is collected, the decision trees are constructed by using the improved SCIForest algorithm, and the trained decision trees are deployed to the monitoring device, and the monitoring device in the second the running state in the first time period of the next update cycle to perform anomaly detection. the running state in the first time period of the next update cycle to perform anomaly detection. the running state in the first time period of the next update cycle to perform anomaly detection. = 1 min.

[0087] The application further provides a device running state self-checking system based on machine learning, which comprises a memory and a processor, and the memory stores a computer program, the computer program integrates the function of the device running state self-checking method based on machine learning, and when the computer program is executed, the device running state self-checking method based on machine learning can improve the efficiency of the self-checking of the running state of the monitoring device and improve the accuracy of the self-checking result.

[0088] In the description of the present specification, the meaning of "a plurality of" and "several" is at least two, for example, two, three or more, etc., unless otherwise explicitly specified.

[0089] Although the present specification has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that such embodiments are provided only by way of example. Those skilled in the art will think of many changes, changes and alternatives without departing from the idea and spirit of the present application. It should be understood that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application.

Claims

1. A device operation status self-checking method based on machine learning, characterized in that: include: Obtain the current sequence of the monitoring device; Adaptively adjusting a subsample size of an SCIForest algorithm using a noise level of the current sequence, wherein the subsample size is negatively correlated with the noise level; Randomly selecting a number of data sets of the same size as the subsample from the current sequence, constructing a decision tree using the SCIForest algorithm, and deploying the trained decision tree to the monitoring device to self-check the status of the monitoring device based on a comparison result between anomaly scores determined by the trained decision tree and preset values; The noise level is obtained by calculating the noise performance of any data in the current sequence, wherein the noise performance is positively correlated with the fitting loss of data within a preset local range centered on the data, and the ratio of the range of the data within the local range to the number of data categories; the noise performance of any data in the current sequence is: ； The first in the current sequence Noise performance of individual data; is the data sequence number in the current sequence; For the The value range of data within the local range of each data; For the The number of data categories within the local range of each data; For the The local range of the data The fitted value of the data; For the The local range of the data The value of the data; For the The total number of data within the local range of each data; is the absolute value symbol; is the hyperbolic tangent function; With the goal of maximizing the between-class variance, the data in the current sequence is divided into two categories. The average noise performance of the current sequence is weighted by the ratio of the data volume of the category with greater noise performance to the other category to obtain the noise level of the current sequence.

2. The device operation status self-check method based on machine learning according to claim 1 is characterized in that: The method for obtaining the fitting value includes: The least square method is used to perform curve fitting on the data within the local range to obtain fitting values ​​of the data within the local range.

3. The device operation status self-checking method based on machine learning according to claim 1 is characterized in that: The subsample size of the SCIForest algorithm is adaptively adjusted using the noise level of the current sequence to satisfy the following relationship: ； Where, is the adjusted subsample size; To preset the initial subsample size; is the noise level of the current sequence; The natural constant An exponential function with base ; The symbol for rounding up.

4. The device operation status self-checking method based on machine learning according to claim 1 is characterized in that: The method for obtaining the anomaly score includes: Get the current at the current moment, use the current at the current moment to traverse all the trained decision trees, and get the path length of the current at the current moment in each trained decision tree; The average path length of the current at the current moment in the trained decision tree is calculated and normalized to obtain the abnormality score of the current at the current moment.

5. The device operation status self-checking method based on machine learning according to claim 4 is characterized in that: The self-checking and monitoring device status based on the comparison result of the abnormality score determined based on the trained decision tree and the preset value includes: Obtaining a preset value, and when the abnormality score is greater than or equal to the preset value, determining that the operating state of the monitoring device is abnormal at the current moment; When the abnormality score is less than the preset value, it is determined that the operating status of the monitoring device is normal at the current moment.

6. The device operation status self-checking method based on machine learning according to claim 4 is characterized in that: The method further comprises: Preset update cycle; In an update cycle, the decision tree constructed using the current sequence collected in the first half of the update cycle is used to determine the abnormality score of each sampling moment in the second half of the update cycle.

7. The equipment operation status self-checking system based on machine learning is characterized by: The device operation status self-inspection system based on machine learning includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

Images