Time series data augmentation and fault prediction model training

Abstract

Embodiments of the present disclosure provide a time series data augmentation method and a fault prediction model training method, a device, a storage medium, and a program product. A data sliding-based data augmentation mode is introduced in time series data having stochastic volatility, and the augmentation mode changes relative data positions in the time series data, without changing data content in the time series data, a series length, and original stochastic characteristics of the data, so that augmented time series data is kept consistent with original time series data in an overall pattern, new data noise would not introduced, key feature information would not be lost, and failing to identify the time series data caused by a great change in the overall pattern would not occur, thereby successfully augmenting the time series data.

Description

[0001] Time series data enhancement and fault prediction model training technology field

[0002]

[0001] The present disclosure relates to the field of cloud computing technology, and more particularly to time series data enhancement and fault prediction model training.

[0003]

[0002] With the development of cloud computing technology, large-scale cloud service providers in the industry have millions or even more servers, and the number of internal components of these servers has even reached tens of millions. These components cause a large number of failures every year, such as hard disk failures, central processing unit (CPU) failures, and memory failures. Among them, hard disk failures can lead to data loss and performance degradation, while CPU and memory failures can directly cause server downtime. These failures can cause huge losses to cloud service providers and users.

[0004] In cloud computing scenarios, data-driven fault prediction technology plays an important role. It learns fault patterns from sample data and uses this information to predict faults in advance, safeguarding the stability of cloud service products. However, in cloud computing scenarios, time series data generated over time suffers from an uneven distribution of normal and faulty sample data, which directly affects the performance of fault prediction models.

[0005]

[0004] Various aspects of the present disclosure provide a time series data enhancement and fault prediction model training method, device, storage medium, and program product for balancing the distribution of normal sample data and fault sample data of time series data and improving the performance of the fault prediction model.

[0006]

[0005] An embodiment of the present disclosure provides a time series data enhancement method, comprising: determining original time series data and a corresponding enhancement number N, wherein the original time series data includes L observation data, the L observation data have random volatility, L and N are positive integers, and L±2, N>1; sliding a portion of the observation data in the original time series data to obtain target time series data including the L observation data; if the number of existing target time series data is less than the enhancement number N, sliding a portion of the observation data in the previous target time series data to obtain next target time series data; wherein the positional relationship of the L observation data in the original time series data and each target time series data is different.

[0007]

[0006] The embodiment of the present disclosure also provides a time series data enhancement method, comprising: obtaining multiple basic time series data in a fault prediction scenario; marking the multiple basic time series data to obtain fault time series data and normal time series data; using the fault time series data as original time series data, and using the above method to perform data enhancement on the fault time series data to obtain enhanced fault time series data; training an initial fault prediction model based on the normal time series data and the enhanced fault time series data to obtain a target fault prediction model.

[0008]

[0007] An embodiment of the present disclosure further provides an electronic device, comprising: a memory and a processor; the memory stores a computer program, and the processor is coupled to the memory and configured to execute the computer program to implement the steps in the above method.

[0009]

[0008] The embodiment of the present disclosure further provides a computer-readable storage medium storing a computer program / instruction. When the computer program / instruction is executed by a processor, the processor is enabled to implement the steps in the above method.

[0010]

[0009] In the disclosed embodiment, a data enhancement method based on data sliding is introduced into time series data with random volatility. This enhancement method changes the relative data position in the time series data without changing the data content, sequence length, and the random characteristics originally presented by the data. This allows the enhanced time series data to remain consistent with the original time series data in terms of overall form. This neither introduces new data noise nor loses key feature information, nor does it become unrecognizable due to significant changes in the overall form. This successfully enhances the time series data, thereby resolving the problem of uneven distribution of normal sample data and fault sample data faced by time series data.

[0011]

[0010] The drawings described herein are intended to provide a further understanding of the present disclosure and constitute a part of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are intended to explain the present disclosure and do not constitute an improper limitation of the present disclosure. In the drawings:

[0012] FIG1 is a schematic diagram of a time series data enhancement method according to an exemplary embodiment of the present disclosure;

[0013] FIG. 2a is a schematic diagram illustrating the states of various marker bits in time series data provided by an exemplary embodiment of the present disclosure;

[0014]

[0013] FIG2b is a schematic diagram of a data sliding process provided by an exemplary embodiment of the present disclosure;

[0015]

[0014] FIG2c is a schematic diagram of another data sliding process provided by another exemplary embodiment of the present disclosure;

[0016] FIG2 d is a schematic diagram of another data sliding process provided by another exemplary embodiment of the present disclosure;

[0017] FIG2e is a schematic diagram of another data sliding process provided by another exemplary embodiment of the present disclosure;

[0018] FIG. 2f is a schematic diagram of a data sliding process provided by another exemplary embodiment of the present disclosure;

[0019] FIG2g is a simple schematic diagram of a fault time series data enhancement method provided by an exemplary embodiment of the present disclosure;

[0020]

[0019] FIG2h is a schematic diagram showing a comparison of fault time series data before and after fault time series data enhancement according to an exemplary embodiment of the present disclosure;

[0021]

[0020] FIG2i is a flow chart of a method for enhancing fault time series data during model training provided by another exemplary embodiment of the present disclosure;

[0022]

[0021] FIG3 is a flow chart of a fault prediction model training method provided by another exemplary embodiment of the present disclosure;

[0023] FIG4 is a schematic diagram of the structure of a time series data enhancement device provided by another exemplary embodiment of the present disclosure;

[0024]

[0023] FIG5 is a schematic structural diagram of a fault prediction model training device provided by another exemplary embodiment of the present disclosure;

[0025] FIG6 is a schematic diagram of the structure of an electronic device provided by another exemplary embodiment of the present disclosure.

[0026] To make the objectives, technical solutions, and advantages of the present disclosure more clearly apparent, the technical solutions of the present disclosure will be described clearly and completely below in conjunction with specific embodiments of the present disclosure and the corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present disclosure, and are not all of the embodiments. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present disclosure without inventive effort are within the scope of protection of the present disclosure.

[0027]

[0026] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse. In addition, the various models involved in this disclosure (including but not limited to language models or large models) are in compliance with relevant laws and standards.

[0027] In cloud computing scenarios, data-driven fault prediction technology plays an important role. Fault prediction technology refers to the technology that evaluates and predicts the future fault trends or possibilities of a server under normal operating conditions. Through fault prediction technology, timely measures can be taken before a fault occurs to avoid server failure and ensure its safe, stable and long-term operation. The main methods of fault prediction technology include at least methods based on time series analysis, methods based on regression analysis, methods based on classification, methods based on machine learning models, etc. Among them, the method based on machine learning models is a hot topic of current research. Its basic principle is to analyze and process a large amount of sample data such as historical operation data and status monitoring of internal components of cloud servers, learn the fault rules in the sample data, and use these fault rules to train fault prediction models to predict future fault trends in advance.

[0028]

[0028] In cloud computing scenarios, various states or indicators of devices such as cloud servers can be observed, and the state data or indicator data observed at each observation point is used as observation data. This observation data is continuously generated over time. In the disclosed embodiments, observation data within a set unit time can be referred to as time series data. The set unit time can be 1 minute, 3 minutes, 10 minutes, 1 hour, etc., depending on application requirements. In the disclosed embodiments, this time series data can be used as training samples to train a fault prediction model for cloud computing scenarios. In other words, the sample data used to learn and train the fault prediction model is time series data. However, training the model often requires a large amount of data to achieve good results. Although server failures are inevitable, their probability of occurrence is generally low. This results in a small amount of sample data related to failures, while a large amount of sample data indicating normal operation is available. Of course, in some special scenarios, there may be more abnormal samples and fewer normal samples. That is to say, in the cloud computing scenario, the time series data generated over time has the problem of unbalanced distribution of normal sample data and fault sample data, which will directly affect the performance of the fault prediction model.

[0029]

[0029] In view of the technical problem that in a cloud computing scenario, time series data generated over time has an unbalanced distribution of normal sample data and fault sample data, in an embodiment of the present disclosure, for time series data with random volatility, a data enhancement method based on data sliding is introduced in the time series data with random volatility. This enhancement method changes the relative data position in the time series data without changing the data content, sequence length and the random characteristics originally presented by the data in the time series data, so that the enhanced time series data remains consistent with the original time series in terms of overall form, neither introducing new data noise nor losing key feature information, nor being unable to be recognized due to a large change in the overall form. The time series data is successfully enhanced, and the problem of unbalanced distribution of normal sample data and fault sample data faced by the time series data can be solved.

[0030] FIG1 is a flow chart of a time series data enhancement method according to an embodiment of the present disclosure. As shown in FIG1 , the method includes the following steps.

[0031]

[0031] 101. Determine original time series data and its corresponding enhancement number N, where the original time series data includes L observation data, the L observation data have random fluctuations, L and N are positive integers, and L±2, N>L

[0032]

[0032] 102. Slide part of the observation data in the original time series data to obtain target time series data containing L observation data.

[0033]

[0033] 103. If the amount of existing target time series data is less than the number of enhancements N, some observation data in the previous target time series data are slid to obtain the next target time series data; wherein the positional relationships of the L observation data in the original time series data and each target time series data are different.

[0034]

[0034] In the embodiments of the present disclosure, time series data refers to a sequence of data that is arranged sequentially over time within a preset time period, representing a series of states, processes, activities, or phenomena that occur over time. In the embodiments of the present disclosure, time series data includes multiple observations observed at multiple observation time points. These observations may vary depending on the observed object. Generally, the observations in time series data exhibit certain variation characteristics. Some time series data exhibits a certain variation trend, i.e., it varies regularly, such as gradually increasing or decreasing. Other time series data exhibits random fluctuations, i.e., it exhibits irregular and sudden variations. For example, the data in time series data may generally exhibit a regular and smooth variation trend, but at a certain time point or time period, the data may suddenly rise or fall; or the data in time series data may generally exhibit a regular and sudden variation trend, but at a certain time point or time period, the data may not show any significant fluctuations; or the data in time series data may generally exhibit a regular variation trend, but at a certain time point or time period, the observation data may be missing. These variations are all irregular and sudden variations.

[0035]

[0035] The embodiments of the present disclosure focus on time series data with random fluctuations, such as time series data related to cloud server performance. Time series data related to cloud server performance generally encompasses any time-varying data generated by the cloud server operating system, application services, database activity, and other system components. For example, time series data can include system performance indicators such as CPU utilization, memory utilization (including physical memory, swap space, etc.), disk I / O operation data (read / write rate, queue length), network traffic (inbound and outbound bandwidth usage), process or thread counts, and system load averages. For example, time series data of CPU utilization collected over a period of one minute, one hour, or one day at a collection frequency of one second. For another example, time series data can include the number and duration of SQL query executions in database activities, changes in the number of database connections, tablespace usage growth, data lock contention data, database transaction volume, and so on. For example, this data can be time series data of database transaction volume collected over a period of 1 minute, 30 minutes, or an hour at a 10-second frequency. Another example is hardware monitoring data such as server temperature, fan speed, and power supply status. For example, this data can be time series data of fan speed collected over a period of 5 minutes or an hour at a 1-second frequency.

[0036] In this embodiment, the time series data generated over time is divided into normal time series data and fault time series data. Normal time series data refers to sequence data that conforms to the changing characteristics of the corresponding indicators of the sample data and conforms to the characteristics of the time series data. Fault time series data refers to time series data that does not conform to the changing characteristics of the corresponding indicators of the sample data and / or does not conform to the characteristics of the time series data. In other words, if all data contained in a time series data is normal data, then the time series data is considered normal time series data. Conversely, if a time series data contains at least some abnormal data, then the time series data is considered fault time series data.

[0037]

[0037] In the time series data generated by the operation of a cloud server, there may be a smaller amount of fault time series data related to faults, while there may be a larger amount of normal time series data related to normal operation. Alternatively, there may be a larger amount of fault time series data related to faults, while there may be a smaller amount of normal time series data related to normal operation, resulting in an uneven distribution of normal time series data and fault time series data. To address this, data enhancement can be performed on the smaller amount of time series data (e.g., normal time series data or fault time series data) to ensure a balanced distribution of normal time series data and fault time series data. In this embodiment, data enhancement refers to increasing the amount of fault time series data at the granularity of "time series data" based on the original time series data, with the goal of maintaining or substantially maintaining the data content, sequence length, and the original random characteristics of the data. For example, if the number of fault time series data that require data enhancement is N, then N fault time series data can be generated based on the existing fault time series data, with the same data content, sequence length, and original random characteristics as the original fault time series data.

[0038] In this embodiment, before performing data enhancement, the original time series data for data enhancement and the corresponding enhancement quantity are first determined. The original time series data includes a certain number of observation data. For ease of description and distinction, the number of observation data included in the original time series is represented as L, and the number of enhancements corresponding to the original time series data is represented as N. L and N are positive integers, and L±2, N>1. Generally, L±N, L observation data have random fluctuations.

[0039]

[0039] The embodiments of the present disclosure do not limit the specific implementation method for determining the original time series data. For example, time series data with frequent irregular data fluctuations can be selected from multiple existing time series data as the original time series data. Alternatively, time series data with the largest data fluctuation trend can be selected from multiple existing time series data as the original time series data. Alternatively, time series data with frequent irregular data fluctuations or the largest data fluctuation trend can be selected from time series data generated in real time during the data enhancement process as the original time series data, and so on. However, the above examples are merely illustrative and do not constitute limitations on the present disclosure.

[0040]

[0040] The embodiments of the present disclosure do not limit the specific implementation method for determining the number of enhancements N corresponding to the original time series data. Alternatively, the number of enhancements N can be determined based on the purpose of performing data enhancement on the original sequence data. For example, if the purpose is to train a target prediction model, the number of enhancements N for the original sequence data can be determined based on the amount of historical time series data used to train the target prediction model. For another example, if the purpose is to correct the prediction accuracy of the target prediction model, the prediction accuracy of the target prediction model to be improved can be estimated, and based on the prediction accuracy to be improved, the amount of time series data required to improve the prediction accuracy to be improved can be determined, and this amount of data can be used as the number of enhancements N, and so on. However, the above examples are merely illustrative and do not limit the present disclosure. More preferred implementations can be found in the relevant descriptions of the following embodiments and will not be elaborated upon here.

[0041]

[0041] Further, after determining the original time series data and its corresponding number of enhancements N, data enhancement is performed on the original time series data to obtain N target time series data. When performing data enhancement on the original time series data, part of the observed data can be used as the target time series data, part of the observed data can be copied and spliced ​​to obtain the target time series data, new data can be added to the original time series data to obtain the target time series data, and so on. However, the above-mentioned method of data enhancement on the original time series data may introduce new noise data and may also lose key feature information. In other words, it may change the data content, sequence length, and original random characteristics of the original time series data, making the enhanced target time series data inconsistent with the original time series data in terms of overall form, thereby affecting the accuracy of fault identification of the fault prediction model trained based on the target time series data.

[0042] To ensure the accuracy of the fault prediction model trained based on time series data, the data content, sequence length, and randomness of the original time series data should be minimized. Based on this, a preferred implementation method for data enhancement of the original time series in the disclosed embodiments is to slide a portion of the observation data in the original time series data to obtain target time series data containing L observation data. The target time series data obtained by sliding a portion of the observation data in the original time series data only changes the relative data positions in the time series data compared to the original time series data, but does not change the data content, sequence length, and randomness of the original time series data. This ensures that the enhanced time series data remains consistent with the original time series in overall form, without introducing new data noise, losing key feature information, or being unrecognizable due to significant changes in overall form. This successfully enhances the time series data, thereby resolving the imbalanced distribution of normal and faulty sample data faced by time series data.

[0043] In this embodiment, the position and quantity of the partial observation data used in the sliding operation are not limited. Optionally, the quantity of the partial observation data used in the sliding operation can be the number of continuous data with abnormal change trends within a randomly selected range in the original time series data; accordingly, the position of the partial observation data used in the sliding operation is the position within the randomly selected range; or, the amount of the partial observation data used in the sliding operation can be the number of data corresponding to the maximum data fluctuation trend range in the original time series data; accordingly, the position of the partial observation data used in the sliding operation is the position corresponding to the maximum data fluctuation trend range, and so on. The above examples are merely illustrative and do not limit the present disclosure.

[0044] In the above embodiment, the number of enhancements N for the original time series data has been determined. To ensure that the number of enhancements reaches N and to avoid wasting computing resources due to the number of enhancements exceeding N, in this embodiment, after each new target time series data is obtained, the number of existing target time series data is counted, and the relationship between the number of existing target time series data and the number of enhancements N is determined. If the amount of existing target time series data is less than the number of enhancements N, a portion of the observation data in the previous target time series data is slid to obtain the next target time series data. If the amount of existing target time series data is equal to the number of enhancements N, the enhancement operation is terminated. The specific implementation of sliding a portion of the observation data in the previous target time series data to obtain the next target time series data can be found in the relevant description of the following embodiment and is not elaborated here.

[0045]

[0045] In various embodiments of the present disclosure, a data enhancement method based on data sliding is introduced into time series data with random volatility. This enhancement method changes the relative data position in the time series data without changing the data content, sequence length, and the random characteristics originally presented by the data in the time series data. This ensures that the enhanced time series data remains consistent with the original time series data in terms of overall form. This neither introduces new data noise nor loses key feature information, nor does it become unrecognizable due to significant changes in the overall form. This successfully enhances the time series data, thereby solving the problem of uneven distribution of normal sample data and fault sample data faced by time series data.

[0046] In the disclosed embodiments, an observation data segment with a set step size can be determined from the original time series. The data contained in the observation data segment is the aforementioned partial observation data. The step size can be understood as the time interval between the time points corresponding to the observation data at the two endpoints of the observation data segment determined in the time series data. It should be noted that during N sliding operations of partial observation data based on a piece of original time series data, the step size of the partial observation data used in each sliding operation can be fixed (referred to as a synchronization length), or the step size used in each sliding operation can be dynamically variable (referred to as a variable step size). In other words, the N sliding operations can be divided into a sliding operation with a synchronization length and a sliding operation with a variable step size. For ease of description, the synchronization length is referred to as the first step size and is denoted by K1. The variable step size determined during each sliding operation is referred to as the second step size and is denoted by K2. K1 is a positive integer, K is 1, K2 is a positive integer, K2>1, and K2NKL is a positive integer.

[0047]

[0047] The following describes the technical solution of the present disclosure in detail based on two sliding modes: a sliding mode with synchronous length and a sliding mode with variable step length.

[0048] In this embodiment, sliding a portion of the observation data in the original time series data to obtain target time series data containing L observation data includes: determining a first step length K1; sliding K1 observation data starting from a first position in the original time series data to a region starting from a second position based on the first step length K1 to obtain the target time series data, wherein the first position is different from the second position. The disclosed embodiment does not limit the locations of the first position and the second position in the original time series data. Optionally, as shown in FIG2a, the first position is the first position in the original time series data, and the second position is the last position in the original time series data. Alternatively, the first position is the last position in the original time series data, and the second position is the first position in the original time series data. Alternatively, the first position is an intermediate position in the original time series data, and the second position is another position in the original time series data. The other position can be the first position or the last position, and the intermediate position refers to another position in the time series data other than the first position and the last position. This embodiment focuses on the case where the first position is the last position in the original time series data, and the second position is the first position in the original time series data. This is because only a small portion of the observed data changes in form; a significant change in the overall form will not affect the performance of the fault prediction model. The above examples are merely illustrative and do not limit the present disclosure.

[0049]

[0049] In some embodiments, determining the first step length K1 includes: determining the first step length K1 according to the number of enhancements N; or determining the first step length K1 according to the number of observations L; or determining the first step length K1 according to the number of enhancements N and the number of observations L; or randomly determining the first step length KL.

[0050] This embodiment does not limit the implementation method of determining the first step length K1 based on the enhancement number N. Alternatively, the enhancement number N can be used as the first step length KL, or a step length coefficient X1 (0<X1<1) of the first step length K1 can be introduced, and the product of the enhancement number N and the step length coefficient X1 of the first step length K1 can be rounded to an integer, and the integer value can be used as the first step length KL. Alternatively, an offset 81 (81 is an integer) can be introduced, and the absolute value of the sum of the enhancement data N and the offset 81 can be used as the first step length K1, and so on. The above examples are merely illustrative and do not limit the present disclosure.

[0051]

[0051] In this embodiment, the embodiment does not limit the implementation method of determining the first step length K1 according to the number of observation data L. Optionally, an offset 32 ​​(32 is an integer) can be introduced, and the absolute value of the difference between the numerical value of the number of observation data L and the numerical value of the offset 32 ​​can be used as the first step length K1; or, a step length coefficient X2 (0 < X2 < 1) of the first step length K1 is introduced, and the product of the number of observation data L and the step length coefficient X2 of the first step length K1 is rounded, and the integer value is used as the first step length K1; or, a preset weight Y1 (Y1 > 0) is introduced, and the product of the number of observation data L and the preset weight Y1 is rounded, and the integer value is used as the first step length K1, etc. The above examples are only illustrative and do not limit the present disclosure.

[0052]

[0052] This embodiment does not limit the implementation method for determining the first step length K1 based on the number of enhancements N and the number of observations L. Alternatively, the number of enhancements N can be used as the numerator, the number of observations included in the time series data can be used as the denominator, the quotient can be calculated and rounded, and the integer value can be used as the first step length K1, in which case L>N. Alternatively, a step length coefficient X3 (0<X3<1) for the first step length K1 can be introduced, the number of enhancements N can be used as the numerator, the number of observations included in the time series data can be used as the denominator, the quotient can be calculated and rounded, and the product of the integer value and the first step length coefficient X3 can be used as the first step length K1. Alternatively, an offset value (where 3 is an integer) can be introduced, the number of enhancements N can be used as the numerator, the number of observations included in the time series data can be used as the denominator, the quotient can be calculated and rounded, and the absolute value of the difference between the integer value and the offset value can be used as the first step length K1, and so on. The above examples are merely illustrative and do not limit the present disclosure.

[0053]

[0053] Accordingly, in this embodiment, if the amount of existing target time series data is less than the enhancement number N, sliding a portion of the observation data in the previous target time series data to obtain the next target time series data includes: sliding K1 observation data starting from the first position in the previous target time series data to the region starting from the second position according to the first step length K1, to obtain the next target time series data. This means that whether sliding a portion of the observation data in the original sequence data to generate the target time series data or sliding a portion of the observation data in a certain target time series data to generate the next target time series data, the step length used is K1, that is, sliding K1 observation data.

[0054]

[0054] It should be noted that when sliding a portion of the observation data in the previous target time series data to obtain the next target time series data, the sliding modes of the partial observation data include: same-position sliding and different-position sliding. Same-position sliding means that the first position in each sliding process corresponds to the same position point in the time series as the first position in the previous sliding process, and the second position in each sliding process corresponds to the same position point in the time series as the second position in the previous sliding process. Different-position sliding means that the first position and / or second position in each sliding process corresponds to a different position point in the time series as the first position and / or second position in the previous sliding process. In the different-position sliding mode, the positional distribution of the first position and the second position is as follows: in each sliding process, the first position is different from the first position in the previous sliding process of the partial observation data, and the second position may be different from the second position in the previous sliding process. Alternatively, in each sliding process, the first position is the same as the first position in the previous sliding process, and the second position may be different from the second position in the previous sliding process. Alternatively, in each sliding process, the first position is different from the first position in the previous sliding process of the partial observation data, and the second position is the same as the second position in the previous sliding process.

[0055]

[0055] For the case where, in each sliding process of partial observation data in the previous target time series data, the first position is the same as the first position in the sliding process of the previous partial observation data, and the second position can be the same as the second position in the sliding process of the previous partial observation data, an example is taken in which the first position in the previous sliding process is the first position in the previous target time series data, and the second position is the last position in the previous target time series data. As shown in FIG2b, first, based on the first target time series data obtained in the previous sliding, according to the first step length K1, K1 observation data 1 starting from the first position in the first target time series data are slid to the last position of the first target time series data to obtain the second target time series data; the above steps are repeated in sequence until the number of existing target time series data is equal to the enhancement number N. Alternatively, an example is given in which the first position in the previous sliding process is the last position in the previous target time series data, and the second position is the first position in the previous target time series data. As shown in FIG2c , based on the first target time series data obtained in the previous sliding process, K1 observation data 1 starting from the last position in the first target time series data are slid to the first position of the first target time series data according to the first step length K1 to obtain the second target time series data.

[0056]

[0056] In each sliding process of partial observation data in the previous target time series data, the first position is different from the first position in the previous sliding process of partial observation data, and the second position can be different from the second position in the previous sliding process of partial observation data. The following example illustrates the situation where the first position in the previous sliding process is the first position in the previous target time series data, and the second position is the last position in the previous target time series data. As shown in FIG2d, first, the original time series data or the previous target time series data is circularly slid once according to the first step length K1 to obtain the first target time series data. Based on the first target time series data obtained from the previous sliding process, K1 observation data 1 starting from the first middle position in the first target time series data and moving to the right are slid to the second middle position in the first target time series data according to the first step length K1, and K1 observation data 2 starting from the second middle position and moving to the right are slid to the first middle position to obtain the second target time series data. The first intermediate position and the second intermediate position are different. The first intermediate position and the second intermediate position are respectively the first position and the second position in the current sliding process. The third intermediate position is used in conjunction with the first intermediate position to define K1 pieces of observation data 1. It should be noted that because the data is constantly sliding, the portion of observation data 1 corresponding to the same data number is different during each sliding process. This rule also applies to the following embodiments and will not be described in detail in the following embodiments.

[0057]

[0057] For each sliding operation of partial observation data in the previous target time series data, the first position is different from the first position in the previous sliding operation of partial observation data, and the second position is the same as the second position in the previous sliding operation of partial observation data. The following example illustrates the situation where the first position in the previous sliding operation of partial observation data is the first position in the previous target time series data, and the second position is the last position in the previous target time series data. As shown in FIG2e, first, according to the first step length K1, K1 observation data 1 starting from the first position in the original time series data or the previous target time series data are slid to the last position to obtain the first target time series data. Based on the first target time series data obtained in the previous sliding operation, according to the first step length K1, K1 observation data 1 starting from the third position in the first target time series data are slid to the last position in the first target time series data, and observation data 2 to the right of observation data 1 are slid to the left to the third middle position to obtain the second target time series data. The third middle position and the last position are the first position and the second position in the current sliding operation, respectively.

[0058]

[0058] For each sliding operation of partial observation data in the previous target time series data, the first position is the same as the first position in the previous sliding operation of partial observation data, and the second position is different from the second position in the previous sliding operation of partial observation data. An example is provided in which the first middle position in the previous sliding operation of partial observation data is the first position of the previous target time series data, and the second position is the last position in the previous target time series data. As shown in FIG2f , first, according to the first step length K1, K1 observation data 1 starting from the first position in the original time series data or the previous target time series data are slid to the last position to obtain the first target time series data. Based on the first target time series data obtained from the previous sliding operation, according to the first step length K1, K1 observation data 1 starting from the first position in the first target time series data are slid to the fourth middle position in the first target time series data. The remaining observation data 2 starting from the fourth middle position to the right are slid to the last position. The observation data 2 with a length of K1 beyond the last position are slid to the first position to obtain the second target time series data. The first position and the fourth middle position are respectively the first position and the second position in this sliding process.

[0059]

[0059] Furthermore, in the synchronous long sliding mode, after obtaining the next target time series data, it is determined whether the next target time series data is the same as the original time series data. If the next target time series data is the same as the original time series data, the first step length K1 is adjusted. Then, based on the adjusted first step length K1, K1 observations starting from the first position in the previous target time series data are slid into the area starting from the second position to obtain the next target time series data. The above steps are repeated until the number of existing target time series data equals the enhancement number N. Case 2: Sliding Mode with Variable Step Size

[0060] In this embodiment, sliding a portion of the observation data in the original time series data to obtain target time series data containing L observation data includes: determining a first step length K1; and sliding K1 observation data starting from a first position in the original time series data to a region starting from a second position based on the first step length K1 to obtain the target time series data, wherein the first position is different from the second position. This embodiment does not limit the locations of the first position and the second position in the original time series data; for details, please refer to the relevant description of the above embodiment and will not be repeated here. This embodiment also does not limit the implementation method of determining the first step length K1; for details, please refer to the relevant description of the above embodiment and will not be repeated here.

[0061]

[0061] Accordingly, in this embodiment, if the number of existing target time series data is less than the enhancement number N, a portion of the observation data in the previous target time series data is slid to obtain the next target time series data, including: determining a second step length K2; according to the second step length K2, sliding K2 observation data starting from the third position in the previous target time series data to the area starting from the fourth position to obtain the next target time series data, wherein the third position is different from the fourth position, the third position and the first position may be the same or different, and the fourth position and the second position may be the same or different.

[0062] In some embodiments, determining the second step length K2 includes: determining the second step length K2 based on the number of enhancements N; or determining the second step length K2 based on the number of observations L; or determining the second step length K2 based on the number of enhancements N and the number of observations L; or randomly determining the second step length K2; or determining the current second step length K2 based on the first step length K1 or the previous second step length K2. It should be noted that, in each sliding process of a portion of the observations in the previous target time series data, the second step length may be different each time, or, in multiple sliding processes, the second step length may be the same in some sliding processes and different in others.

[0063]

[0063] This embodiment does not limit the implementation method of determining the second step length K2 based on the enhancement number N. Optionally, a weighting coefficient Y2 (Y2>0) that changes randomly over time can be introduced, and the product of the enhancement number N and the weighting coefficient Y2 can be rounded up, and the integer value can be used as the second step length K2. Alternatively, a variable step length coefficient X4 of the first step length K1 can be introduced.

[0064] (0 < X4 < 1, and the magnitude varies randomly over time), the product of the enhancement number N and the step coefficient X4 of the second step size K2 is rounded to an integer, and the integer value is used as the second step size K2. Alternatively, an offset bl is introduced (which is an integer and its magnitude varies randomly), and the absolute value of the sum of the enhancement data N and the offset is used as the second step size K2, and so on. The above examples are merely illustrative and do not limit the present disclosure.

[0065]

[0064] This embodiment does not limit the implementation method of determining the second step size K2 based on the number of observation data L. Alternatively, an offset 84 (84 is an integer whose magnitude varies randomly over time) may be introduced, and the absolute value of the difference between the number of observation data L and the offset 84 may be used as the second step size K2. Alternatively, a variable step size coefficient X5 (0 < X5 < 1 and varies randomly over time) may be introduced for the second step size K2, and the product of the number of observation data L and the variable step size coefficient X5 of the second step size K2 may be rounded to an integer, and the resulting integer value may be used as the second step size K1. Alternatively, a preset weight Y3 (Y3 > 0 and varies randomly over time) may be introduced, and the product of the number of observation data L and the preset weight Y3 may be rounded to an integer, and the resulting integer value may be used as the first step size K1, and so on. The above examples are merely illustrative and do not limit the present disclosure.

[0066] This embodiment does not limit the implementation method for determining the second step size K1 based on the number of enhancements N and the number of observations L. Alternatively, a variable step size coefficient X6 (0 < X6 < 1 and varying randomly over time) can be introduced for the second step size K2. First, the number of enhancements N is used as the numerator, the L observations included in the time series data are used as the denominator, and a quotient is calculated. The product of the quotient and the variable step size coefficient X6 for the second step size K2 is then rounded, and the integer value is used as the second step size K2. In this case, L > N. Alternatively, an offset 35 (which is an integer and varies randomly over time) can be introduced. The number of enhancements N can be used as the numerator, the L observations included in the time series data are used as the denominator, the quotient is calculated and rounded, and the absolute value of the difference between the integer value and the offset 35 is used as the second step size K2. Alternatively, a variable step size coefficient X7 (0 < X7 < 1 and randomly changing over time) of the second step size K2 and a weighting coefficient Y4 (0 < X2 < 1, 0 < Y4 < 1, and both X2 and Y4 randomly changing over time) of the second step size K2 may be introduced. First, the number of enhancements N is used as the numerator, and the L observations included in the time series data are used as the denominator. A quotient is calculated, and then the product of the quotient, the variable step size coefficient Y4, and the weighting coefficient of the second step size K2 is integerized, and the integer value is used as the second step size K2, and so on. The above examples are merely illustrative and do not limit the present disclosure.

[0067]

[0066] This embodiment does not limit the implementation method of determining the second step length K2 of the next sliding according to the first step length K1 or the previous second step length K2. Optionally, an offset 36 (36 is an integer and its size varies randomly) is introduced, and the absolute value of the difference between the first step K1 and the offset 36 is used as the second step K2 of the next slide. Alternatively, a variable step coefficient X8 (0 < X8 < 1 and its size varies randomly) is introduced for the second step K2, and the product of the previous second step K2 and the step coefficient X8 of the second step K2 is rounded up, and the resulting integer value is used as the second step K2 of the next slide. Alternatively, a weighting coefficient Y5 (0 < Y5 < 1, Y5 varies randomly) of the first step K1 and a weighting coefficient Y6 (0 < Y5 < 1, Y5 varies randomly) of the previous second step K2 are introduced, and the sum of the product of the first step K1 and its weighting coefficient Y5 and the product of the previous second step K2 and its weighting coefficient Y5 is rounded up, and the resulting integer value is used as the second step K2 of the next slide, and so on. The above examples are merely illustrative and do not limit the present disclosure.

[0068]

[0067] It should be noted that when sliding a portion of the observation data in the previous target time series data to obtain the next target time series data, the sliding modes of the portion of the observation data include: sliding at the same position with a variable step length and sliding at different positions with a variable step length. Sliding at the same position with a variable step length means that the third position in each sliding process corresponds to the same position point in the time series as the third position in the previous sliding process, and the fourth position in each sliding process corresponds to the same position point in the time series as the first position in the previous sliding process, and the step length of each sliding is different. Sliding at different positions with a variable step length means that at least one of the third position and the fourth position in each sliding process corresponds to a different position point in the time series than the corresponding position in the previous sliding process. In the sliding mode of sliding at the same position with a variable step length, the positional distribution of the third position and the fourth position is as follows: the third position is the same position as the third position in the previous sliding process, and the fourth position is the same position as the fourth position in the previous sliding process. In the sliding mode with different positions and variable step lengths, the distribution of the third position and the fourth position is as follows: in each sliding process, the third position is different from the third position in the previous sliding process, and the fourth position is different from the fourth position in the previous sliding process; alternatively, in each sliding process, the third position is the same as the third position in the previous sliding process, and the fourth position is different from the fourth position in the previous sliding process; alternatively, in each sliding process, the third position is different from the third position in the previous sliding process, and the fourth position is the same as the fourth position in the previous sliding process, and so on. For exemplary descriptions of the sliding mode with the same position and variable step length and the sliding mode with different positions and variable step lengths, please refer to the description of the synchronous step length sliding mode above and will not be described in detail here.

[0069]

[0068] Furthermore, in the variable step size sliding mode, after obtaining the next target time series data, it is determined whether the next target time series data is the same as the original time series data and the previous target time series data; if the next target time series data is the same as the original time series data or the previous target time series data, the second step size K2 is adjusted, and the next target time series data is regenerated based on the previous target time series data according to the adjusted step size.

[0070] Whether using the aforementioned sliding method with a synchronous length or a sliding method with a variable step length, the same implementation method can be used for determining the original time series data and determining the number of enhancements N corresponding to the original time series data. In some preferred embodiments, determining the number of enhancements N corresponding to the original time series data includes: determining the number of enhancements N based on the fault prediction scenario to which the original time series data belongs; wherein different fault prediction scenarios have different requirements for fault prediction accuracy; and / or determining the number of enhancements N based on observation indicators corresponding to observation data in the original time series data; wherein different observation indicators have different importance in the fault prediction scenario; and / or determining the number of enhancements N based on the ratio of the original time series data to another type of time series data in the fault prediction scenario, with different ratios corresponding to different numbers of enhancements N. This embodiment does not limit the classification of fault prediction scenarios. For example, fault prediction scenarios can be classified based on the prediction object. Fault predictions can include memory fault prediction, CPU fault prediction, network fault prediction, application fault prediction, virtual machine fault prediction, container fault prediction, and so on. The prediction accuracy of fault prediction in these different prediction scenarios varies, so the corresponding enhancement number N will also vary. This embodiment does not limit the classification of observation indicators. For example, the observation data can be observation indicators required in the fault prediction scenario. For example, in the CPU fault prediction scenario, the observation indicators can be CPU utilization, power consumption, temperature, etc.; in the memory fault prediction scenario, the observation indicators can be memory utilization, remaining memory, memory fragmentation, etc. Different observation indicators have different importance in the fault prediction scenario. The more important the observation indicator, the more samples may be required, and the enhancement number N can be larger. Conversely, the less important the observation indicator, the fewer samples may be required, and the enhancement number N can be smaller.

[0071]

[0070] Further, the enhancement number N may be determined according to the fault prediction scenario to which the original time series data belongs; or, the enhancement number N may be determined according to the observation index corresponding to the observation data in the original time series data; or, the enhancement number N may be determined according to the ratio of the number of the original time series data and the time series data of another type in the fault prediction scenario; or, the enhancement number N may be determined according to the fault prediction scenario to which the original time series data belongs and the observation index corresponding to the observation data in the original time series data; or, the enhancement number N may be determined according to the fault prediction scenario to which the original time series data belongs and the ratio of the number of the original time series data and the time series data of another type in the fault prediction scenario; or, the enhancement number N may be determined according to the observation index corresponding to the observation data in the original time series data and the ratio of the number of the original time series data and the time series data of another type in the fault prediction scenario; or, the enhancement number N may be determined according to the fault prediction scenario to which the original time series data belongs, the observation index corresponding to the observation data in the original time series data, and the ratio of the number of the original time series data and the time series data of another type in the fault prediction scenario.

[0072]

[0071] In some preferred embodiments, determining the original time series data includes: receiving multiple basic time series data reported by electronic devices, applications, and / or virtualized instances in a fault prediction scenario, the basic time series data including various observation indicators observed at the L1 observation time point in the fault prediction scenario. This embodiment focuses on performance indicators of the electronic devices, applications, and / or virtualized instances; then, marking the multiple basic time series data, such as manually or automatically, to obtain fault time series data and normal time series data; and using a relatively small number of the fault time series data and the normal time series data as the original time series data. The method for determining whether the basic time series data is faulty varies in different application scenarios, and this is not limited in the present embodiment.

[0073]

[0072] Further, after obtaining N target time series data containing L observation data, the original time series data and each target time series data can be used as training samples to perform model training on the initial fault prediction model to obtain a target fault prediction model; the target fault prediction model is used to predict the basic time series data generated in the fault scenario to obtain a prediction result, and the accuracy information of the fault prediction is statistically analyzed; and the enhancement number N is updated according to the accuracy information of the target fault prediction model when performing fault prediction based on the basic time series data in the fault prediction scenario.

[0074]

[0073] Optionally, calculating accuracy information during fault prediction includes: determining accuracy evaluation criteria for each operating indicator of a target fault prediction model; obtaining a prediction result of the target prediction model for the target operating indicator; and determining accuracy information of the target prediction model for the target operating indicator based on the accuracy evaluation criteria for the target operating indicator and the prediction result of the target prediction model for the target operating indicator. The accuracy information may be either meeting the standard or failing to meet the standard, and may also be rated as excellent, good, poor, etc., as long as the accuracy information can indicate the accuracy of the prediction result.

[0075] Optionally, updating the enhancement number N includes: determining, based on the accuracy information of the target prediction model for the target operating indicator, whether the target prediction model's prediction result for the target operating indicator meets the target standard; and if not, determining a new enhancement number N based on the difference between the accuracy of the prediction result and the user's expected accuracy, and updating the original enhancement number N. The new enhancement number N is greater than the original enhancement number N.

[0076]

[0075] Further, after updating the enhancement number N, the relationship between the number of existing target time series data and the updated enhancement number N is continuously determined. If the number of existing target time series data is less than the updated enhancement number N, part of the observation data in the previous target time series data is continuously slid to obtain the next target time series data. When the number of existing target time series data is equal to the updated enhancement number N, the current target fault prediction model is updated according to the normal time series data and the enhanced fault time series data to obtain an updated target fault prediction model. The updated target fault prediction model is used to predict the technical time series data generated in the fault scenario to obtain a new prediction result. The accuracy information of the updated fault prediction is statistically analyzed to further determine whether the enhancement number N needs to be updated.

[0077]

[0076] Figure 2g is a simple schematic diagram of fault time series data enhancement; Figure 2h is the original time series data, the horizontal axis is the observation time point, and the vertical axis is the data value of the observed original time series data. The original time series data takes the fault time series data as an example. As shown in Figure 2h, series 1 represents the data interval where the fault occurs in the fault time series data; by sliding the fault time series data, new target time series data can be obtained, namely, the time series data corresponding to series 2, series 3, series 4 and series 5; among which, the time series data corresponding to series 2 is obtained by sliding on the basis of the time series data corresponding to series 1 (i.e., the original time series data); the time series data corresponding to series 3 is obtained by sliding on the basis of the time series data corresponding to series 2 (i.e., the original time series data); the time series data corresponding to series 4 is obtained by sliding on the basis of the time series data corresponding to series 3 (i.e., the original time series data); and the time series data corresponding to series 5 is obtained by sliding on the basis of the time series data corresponding to series 4 (i.e., the original time series data). In FIG2h, considering that the data length and data content of each target time series data are the same, multiple target time series data are superimposed and displayed, and the different locations of the fault data areas in each target time series data are highlighted.

[0078] In the embodiment shown in FIG. 2i, fault time series data is used as an example of data enhancement to describe the application of the time series data enhancement method provided by the present disclosure in fault prediction model training. As shown in FIG. 2i, the method includes the following steps.

[0079]

[0078] Step 21: Acquire multiple basic time series data in the fault prediction scenario, each basic time series data includes L observation data.

[0080]

[0079] Step 22: Mark multiple basic time series data to obtain fault time series data and normal time series data.

[0081]

[0080] Step 23: Take the fault time series data as the original time series data, and determine the original time series data and its corresponding enhancement number N.

[0082]

[0081] Step 24: Based on the original time series data, some observation data in the original time series data are slid to obtain target time series data containing L observation data.

[0083]

[0082] Step 25: Determine the relationship between the number of existing target time series data and the number of enhancements N. If the number of existing target time series data is less than the number of enhancements N, proceed to step 26. If the number of existing target time series data is equal to the number of enhancements N, proceed to step 27. FIG2g is a simple diagram of the enhancement process from 1 to N time series data.

[0084]

[0083] Step 26: Slide part of the observation data in the previous target time series data to obtain the next target time series data, and return to step 25. As shown in FIG211, the selected part of the observation data is series 1 before and after the peak. After sliding the part of the observation data four times, the fault time series data containing five peaks, series 1, series 2, series 3, series 4, and series 5, are obtained.

[0085]

[0084] Step 27: Train the initial fault prediction model based on the normal time series data and the enhanced fault time series data to obtain the target fault prediction model.

[0086]

[0085] Step 28: Use the target fault prediction model to predict the basic time series data generated in the fault scenario to obtain a prediction result, and calculate the accuracy information of the target fault prediction model when performing fault prediction based on the basic time series data in the fault prediction scenario.

[0087]

[0086] Step 29: Update the enhancement number N according to the accuracy information of the target fault prediction model when performing fault prediction based on the basic time series data in the fault prediction scenario, and return to step 25.

[0088]

[0087] Specifically, when the accuracy information during fault prediction is low, for example, lower than a set accuracy threshold, the number of enhancements N can be increased to continue enhancing the fault time series data and further balance the distribution between normal sample data and fault sample data, so as to improve the model accuracy obtained by model training during the re-model training process.

[0089]

[0088] In the embodiment of the present disclosure, during the model training process, a data enhancement method based on data sliding is introduced for time series data with random volatility. This enhancement method changes the relative data position in the time series data without changing the data content, sequence length and the random characteristics originally presented by the data in the time series data, so that the enhanced time series data remains consistent with the original time series data in terms of overall form. New data noise will not be introduced, key feature information will not be lost, and the data will not be unrecognizable due to a large change in the overall form. The enhancement of the time series data is successfully achieved, and the problem of uneven distribution of normal sample data and fault sample data faced by the time series data can be solved, and the accuracy of the model trained based on these sample data is ultimately improved.

[0090]

[0089] Furthermore, the model training results are fed back to the data enhancement link. When the accuracy information during fault prediction is low, for example, below a set accuracy threshold, the enhancement number N can be increased to continue enhancing the fault time series data and further balance the distribution between normal sample data and fault sample data, so that the model accuracy obtained by model training can be improved during the re-model training process.

[0091]

[0090] FIG3 is a flow chart of a fault prediction model training method provided by an embodiment of the present disclosure. As shown in FIG3, the method includes the following steps.

[0092]

[0091] 301. Acquire multiple basic time series data in a fault prediction scenario.

[0093]

[0092] 302. Mark multiple basic time series data to obtain fault time series data and normal time series data.

[0094]

[0093] 303. Taking the fault time series data as the original time series data, the method in the embodiment related to FIG1 is used to perform data enhancement on the fault time series data to obtain enhanced fault time series data.

[0095]

[0094] 304. Train the initial fault prediction model based on the normal time series data and the enhanced fault time series data to obtain a target fault prediction model.

[0096]

[0095] The method in the embodiment related to FIG1 is used to perform data enhancement on the fault time series data to obtain enhanced fault time series data. For details, please refer to the description in the embodiment related to FIG1 and will not be repeated here.

[0097]

[0096] In addition, the detailed implementation and beneficial effects of each step in the method of this embodiment have been described in detail in the aforementioned embodiments and will not be elaborated here.

[0098] FIG4 is a schematic diagram of the structure of a time series data enhancement device provided in an embodiment of the present disclosure. As shown in FIG4 , the device includes the following modules.

[0099]

[0098] A determination module 41 is used to determine the original time series data and its corresponding enhancement number N. The original time series data includes L observation data, the L observation data have random fluctuations, L and N are positive integers, and L±2, N>L

[0100]

[0099] The sliding module 42 is used to slide part of the observation data in the original time series data to obtain target time series data containing L observation data.

[0101]

[0100] The sliding module 42 is further configured to slide part of the observation data in the previous target time series data to obtain the next target time series data if the amount of the existing target time series data is less than the enhancement amount N.

[0102]

[0101] Among them, the positional relationships of the L observation data in the original time series data and each target time series data are different.

[0103]

[0102] In this embodiment, when the sliding module 42 is used to slide part of the observation data in the original time series data to obtain the target time series data including L observation data, it is specifically used to: determine a first step length K1, K1 is a positive integer, K1>1; according to the first step length K1, slide K1 observation data starting from a first position in the original time series data to a region starting from a second position to obtain the target time series data; wherein the second position is different from the first position.

[0104]

[0103] In this embodiment, when the sliding module 42 is used to slide part of the observation data in the previous target time series data to obtain the next target time series data, it is specifically used to: according to the first step length K1, slide K1 observation data starting from the first position in the previous target time series data to the area starting from the second position to obtain the next target time series data; or determine the second step length K2, K2 is a positive integer, K2>1 WK1; according to the second step length K2, slide K2 observation data starting from the third position in the previous target time series data to the area starting from the fourth position to obtain the next target time series data; the third position is different from the fourth position.

[0105]

[0104] Optionally, the first position is the first position in the original time series data, and the second position is the last position in the original time series data; wherein, the third position is the same as or different from the first position, and the fourth position is the same as or different from the second position.

[0106]

[0105] Further optionally, after the next target time series data is obtained by the sliding module 42, it further includes: a judgment module 43 and an adjustment module 44; wherein the judgment module 43 is used to judge whether the next target time series data is the same as the original time series data or the previous target time series data; if the next target time series data is the same as the original time series data or the previous target time series data, the adjustment module 44 adjusts the first step length K1 or the second step length K2, and regenerates the next target time series data based on the previous target time series data according to the adjusted step length.

[0107]

[0106] Optionally, when the sliding module 42 is used to determine the first step length K1, it is specifically used to: determine the first step length K1 according to the number of enhancements N; or determine the first step length K1 according to the number of observation data L; or determine the first step length K1 according to the number of enhancements N and the number of observation data L; or randomly determine the first step length KL

[0108]

[0107] In some embodiments, when the determination module 41 is used to determine the enhancement number N corresponding to the original time series data, it is specifically used to: determine the enhancement number N according to the fault prediction scenario to which the original time series data belongs; wherein different fault prediction scenarios have different requirements for fault prediction accuracy; and / or determine the enhancement number N according to the observation indicators corresponding to the observation data in the original time series data; wherein different observation indicators have different importance in the fault prediction scenario; and / or determine the enhancement number N according to the ratio of the original time series data and another type of time series data in the fault prediction scenario, and different ratios correspond to different enhancement numbers N.

[0109]

[0108] In some embodiments, when the determination module 41 is used to determine the original time series data, it is specifically used to: receive multiple basic time series data reported by electronic devices, applications and / or virtualization instances in the fault prediction scenario; mark the multiple basic time series data to obtain fault time series data and normal time series data; and use a relatively small number of time series data among the fault time series data and the normal time series data as the original time series data.

[0110]

[0109] Further optionally, it also includes: a training module and a prediction update module; wherein the training module is used to use the original time series data and each target time series data as training samples to perform model training on the initial fault prediction model to obtain a target fault prediction model; the prediction update module is used to update the enhancement number N according to the accuracy information of the target fault prediction model when performing fault prediction based on the basic time series data in the fault prediction scenario.

[0111]

[0110] The detailed implementation and beneficial effects of the steps executed in each module of this embodiment have been described in detail in the previous embodiments and will not be elaborated on here.

[0111] FIG5 is a schematic diagram of the structure of a fault prediction model training device provided in an embodiment of the present disclosure. As shown in FIG5, the device includes the following modules.

[0112]

[0112] The acquisition module 51 is used to obtain multiple basic time series data in the fault prediction scenario.

[0113]

[0113] The marking module 52 is used to mark multiple basic time series data to obtain fault time series data and normal time series data.

[0114]

[0114] The enhancement module 53 is used to use the fault time series data as the original time series data and perform data enhancement on the fault time series data using the time series data enhancement method in any of the above embodiments to obtain enhanced fault time series data.

[0115]

[0115] The training module 54 is used to train the initial fault prediction model according to the normal time series data and the enhanced fault time series data to obtain the target fault prediction model.

[0116]

[0116] The detailed implementation methods and beneficial effects of the steps executed in each module in this embodiment have been described in detail in the aforementioned embodiments and will not be elaborated on here.

[0117] FIG6 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. As shown in FIG6 , the electronic device includes a memory 60a and a processor 60b. The memory 60a stores a computer program, and the processor 60b is coupled to the memory 60a to execute the computer program to perform the following steps.

[0118]

[0118] Determine original time series data and its corresponding enhancement number N, where the original time series data includes L observation data, the L observation data have random fluctuations, L and N are positive integers, and L±2, N>1; slide some observation data in the original time series data to obtain target time series data including the L observation data; if the number of existing target time series data is less than the enhancement number N, slide some observation data in the previous target time series data to obtain the next target time series data; wherein the positional relationship of the L observation data in the original time series data and each target time series data is different.

[0119]

[0119] In this embodiment, when the processor 60b is used to slide part of the observation data in the original time series data to obtain target time series data including L observation data, it is specifically used to: determine a first step length K1, K1 is a positive integer, K1>1; according to the first step length K1, slide K1 observation data starting from a first position in the original time series data to an area starting from a second position to obtain the target time series data; wherein the second position is different from the first position.

[0120]

[0120] In this embodiment, when the processor 60b is used to slide part of the observation data in the previous target time series data to obtain the next target time series data, it is specifically used to: according to the first step length K1, slide K1 observation data starting from the first position in the previous target time series data to the area starting from the second position to obtain the next target time series data; or determine the second step length K2, K2 is a positive integer, K2>1 WK1; according to the second step length K2, slide K2 observation data starting from the third position in the previous target time series data to the area starting from the fourth position to obtain the next target time series data; the third position is different from the fourth position.

[0121]

[0121] Optionally, the first position is the first position in the original time series data, and the second position is the last position in the original time series data; wherein, the third position is the same as or different from the first position, and the fourth position is the same as or different from the second position.

[0122]

[0122] Further optionally, after obtaining the next target time series data, the processor 60b is further used to: determine whether the next target time series data is the same as the original time series data or the previous target time series data; if the next target time series data is the same as the original time series data or the previous target time series data, adjust the first step length KI or the second step length K2, and regenerate the next target time series data based on the previous target time series data according to the adjusted step length.

[0123]

[0123] Optionally, when the processor 60b is used to determine the first step length K1, it is specifically used to: determine the first step length K1 according to the number of enhancements N; or determine the first step length K1 according to the number of observation data L; or determine the first step length K1 according to the number of enhancements N and the number of observation data L; or randomly determine the first step length KL

[0124]

[0124] In some embodiments, when the processor 60b is used to determine the enhancement number N corresponding to the original time series data, it is specifically used to: determine the enhancement number N according to the fault prediction scenario to which the original time series data belongs; wherein different fault prediction scenarios have different requirements for fault prediction accuracy; and / or determine the enhancement number N according to the observation indicators corresponding to the observation data in the original time series data; wherein different observation indicators have different importance in the fault prediction scenario; and / or determine the enhancement number N according to the ratio of the original time series data and another type of time series data in the fault prediction scenario, and different ratios correspond to different enhancement numbers N.

[0125]

[0125] In some embodiments, when the processor 60b is used to determine the original time series data, it is specifically used to: receive multiple basic time series data reported by electronic devices, applications and / or virtualization instances in the fault prediction scenario; mark the multiple basic time series data to obtain fault time series data and normal time series data; and use a relatively small number of time series data among the fault time series data and the normal time series data as the original time series data.

[0126]

[0126] Further optionally, the processor 60b is further configured to: perform model training on the initial fault prediction model using the original time series data and each target time series data as training samples to obtain a target fault prediction model; and update the enhancement quantity N according to accuracy information of the target fault prediction model when performing fault prediction based on the basic time series data in the fault prediction scenario.

[0127]

[0127] The detailed implementation and beneficial effects of each step in the method of this embodiment have been described in detail in the above embodiments and will not be elaborated here.

[0128]

[0128] Accordingly, the electronic device having the same composition structure as FIG6 includes: a memory 60a and a processor 60b; the memory 60a stores a computer program, and the processor 60b is coupled to the memory 60a for executing the computer program, and is also used to perform the following steps.

[0129]

[0129] Acquire multiple basic time series data in a fault prediction scenario; label the multiple basic time series data to obtain fault time series data and normal time series data; use the fault time series data as original time series data, and enhance the fault time series data using the method in the embodiment related to FIG1 to obtain enhanced fault time series data; train an initial fault prediction model based on the normal time series data and the enhanced fault time series data to obtain a target fault prediction model.

[0130]

[0130] Furthermore, as shown in FIG6 , the electronic device further includes: a communication component 60c, a display 60d, a power supply component 60e, an audio component 60f, and other components. FIG6 only schematically illustrates some components, which does not mean that the electronic device only includes the components shown in FIG6 . The electronic device of this embodiment can be implemented as an electronic device such as a desktop computer, a laptop computer, a smartphone, or an IoT device, or as a server-side device such as a conventional server, a cloud server, or a server array. If the electronic device of this embodiment is implemented as a terminal device such as a desktop computer, a laptop computer, or a smartphone, it may include the components within the dashed box in FIG6 ; if the electronic device of this embodiment is implemented as a server-side device such as a conventional server, a cloud server, or a server array, it may not include the components within the dashed box in FIG6 .

[0131]

[0131] The detailed implementation methods and beneficial effects of the steps executed in each module in this embodiment have been described in detail in the aforementioned embodiments and will not be elaborated on here.

[0132] The embodiments of the present disclosure further provide a computer-readable storage medium storing a computer program / instruction. When the computer program / instruction is executed by a processor, the processor is enabled to implement the steps in the above method embodiments.

[0132]

[0133] The present disclosure also provides a computer program product, which includes a computer program / instruction. When the computer program / instruction is executed by a processor, the processor is enabled to implement the steps in the above method embodiments.

[0133]

[0134] The above-mentioned memory can be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0134]

[0135] The communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as WiFi, 2G, 3G, and 4G / LTE.

[0135] 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component further includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0136]

[0136] The display includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, slides, and gestures on the touch panel. The touch sensor may not only sense the boundaries of a touch or slide action, but also detect the duration and pressure associated with the touch or slide operation.

[0137]

[0137] The power supply assembly provides power to various components of the device in which the power supply assembly is located. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device in which the power supply assembly is located.

[0138]

[0138] The audio component may be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), and when the device where the audio component is located is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal may be further stored in a memory or sent via a communication component. In some embodiments, the audio component further includes a speaker for outputting the audio signal.

[0139]

[0139] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to magnetic disk storage, compact disc read-only memory (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0140] The present disclosure is described with reference to the flow chart and / or block diagram of the method, equipment (system) and computer program product according to the embodiment of the present disclosure.It should be understood that each flow process and / or box in the flow chart and / or block diagram and the combination of the flow process and / or box in the flow chart and / or block diagram can be realized by computer program instructions.These computer program instructions can be provided to the processor of general-purpose computer, special-purpose computer, embedded processor or other programmable data processing device to produce a machine, so that the instruction executed by the processor of computer or other programmable data processing device produces the device for realizing the function specified in flow chart one flow process or multiple flow processes and / or one box or multiple boxes of block diagram.

[0141]

[0141] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0142]

[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operating steps are performed on the computer or other programmable device to produce computer-implemented processing, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0143]

[0143] In a typical configuration, a computing device includes one or more processors (Central Processing Unit, CPU), input / output interfaces, network interfaces and memory.

[0144]

[0144] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory.

[0145] (flash RAM). Memory is an example of a computer-readable medium.

[0146]

[0145] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, disk storage or other magnetic storage devices, or any other non-transmitting media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0147]

[0146] It should also be noted that the terms "comprise," "include," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a..." does not preclude the presence of other identical elements in the process, method, commodity, or apparatus comprising the element.

[0148]

[0147] The above are merely embodiments of the present disclosure and are not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present disclosure should be included within the scope of the claims of the present disclosure.

Claims

Claims 1. A time series data enhancement method, comprising: Determining original time series data and its corresponding enhancement number N, wherein the original time series data includes L observation data, the L observation data have random fluctuations, L and N are positive integers, and L±2, N>1; sliding a portion of the observation data in the original time series data to obtain target time series data including the L observation data; If the number of existing target time series data is less than the enhancement number N, some observation data in the previous target time series data are slid to obtain the next target time series data; The positional relationships of the L observation data in the original time series data and each target time series data are different.

2. The method according to claim 1, wherein: Sliding a portion of the observation data in the original time series data to obtain target time series data including the L observation data includes: determining a first step length K1, where K1 is a positive integer and K1>1; and sliding K1 observation data starting from a first position in the original time series data to a region starting from a second position based on the first step length K1 to obtain the target time series data; wherein the second position is different from the first position.

3. The method according to claim 2, wherein: Sliding part of the observation data in the previous target time series data to obtain the next target time series data includes: sliding K1 observation data starting from the first position in the previous target time series data to the region starting from the second position according to the first step length K1, to obtain the next target time series data; or determining a second step length K2, K2 is a positive integer, K2>1 / K1; sliding K2 observation data starting from the third position in the previous target time series data to the region starting from the fourth position according to the second step length K2, to obtain the next target time series data; the third position is different from the fourth position.

4. The method according to claim 3, wherein: The first position is the first position in the original time series data, and the second position is the last position in the original time series data; wherein the third position is the same as or different from the first position, and the fourth position is the same as or different from the second position.

5. The method according to claim 3, wherein: After obtaining the next target time series data, the method further includes: determining whether the next target time series data is the same as the original time series data or the previous target time series data; if the next target time series data is the same as the original time series data or the previous target time series data, adjusting the first step length K1 or the second step length K2, and regenerating the next target time series data based on the previous target time series data according to the adjusted step length.

6. The method according to claim 2, wherein: Determining the first step length K1 includes: determining the first step length K1 according to the enhancement number N; or, determining the first step length K1 according to the number of observation data L; or, determining the first step length K1 according to the enhancement number N and the number of observation data L; or, randomly determining the first step length KL 7. The method according to claim 1, wherein: Determine the number of enhancements N corresponding to the original time series data, including: The enhancement number N is determined based on the fault prediction scenario to which the original time series data belongs; different fault prediction scenarios have different requirements for fault prediction accuracy; and / or the enhancement number N is determined based on observation indicators corresponding to observation data in the original time series data; different observation indicators have different importance in the fault prediction scenario; and / or the enhancement number N is determined based on the ratio of the original time series data to another type of time series data in the fault prediction scenario, with different ratios corresponding to different enhancement numbers N.

8. The method according to any one of claims 1 to 7, wherein: Determining original time series data includes: receiving multiple basic time series data reported by electronic devices, applications, and / or virtualized instances in a fault prediction scenario; marking the multiple basic time series data to obtain fault time series data and normal time series data; and using a relatively small number of time series data from the fault time series data and the normal time series data as the original time series data.

9. The method according to claim 8, further comprising: The original time series data and each target time series data are used as training samples to perform model training on the initial fault prediction model to obtain a target fault prediction model; and the enhancement number N is updated according to accuracy information of the target fault prediction model when performing fault prediction based on the basic time series data in the fault prediction scenario.

10. A fault prediction model training method, comprising: Obtain multiple basic time series data in fault prediction scenarios; Marking the multiple basic time series data to obtain fault time series data and normal time series data; Taking the fault time series data as original time series data, and performing data enhancement on the fault time series data using the method according to any one of claims 1 to 9 to obtain enhanced fault time series data; An initial fault prediction model is trained according to the normal time series data and the enhanced fault time series data to obtain a target fault prediction model.

11. An electronic device, comprising: A memory and a processor; wherein the memory stores a computer program, and the processor is coupled to the memory for executing the computer program to implement the steps in the method according to any one of claims 1 to 9 and claim 10.

12. A computer-readable storage medium storing a computer program / instruction, wherein: When the computer program / instructions are executed by a processor, the processor is enabled to implement the steps of the method according to any one of claims 1 to 9 and claim 10.

13. A computer program product, wherein: The computer program product comprises a computer program / instruction, and when the computer program / instruction is executed by a processor, the processor is enabled to implement the steps of the method according to any one of claims 1 to 9 and claim 10.

Images