WO2022142120A1 - 基于人工智能的数据检测方法、装置、服务器及存储介质 - Google Patents

基于人工智能的数据检测方法、装置、服务器及存储介质 Download PDF

Info

Publication number
WO2022142120A1
WO2022142120A1 PCT/CN2021/097410 CN2021097410W WO2022142120A1 WO 2022142120 A1 WO2022142120 A1 WO 2022142120A1 CN 2021097410 W CN2021097410 W CN 2021097410W WO 2022142120 A1 WO2022142120 A1 WO 2022142120A1
Authority
WO
WIPO (PCT)
Prior art keywords
time series
series data
vector
data
time
Prior art date
Application number
PCT/CN2021/097410
Other languages
English (en)
French (fr)
Inventor
陈桢博
郑立颖
徐亮
Original Assignee
平安科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2022142120A1 publication Critical patent/WO2022142120A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/903Querying
    • G06F16/90335Query processing
    • G06F16/90348Query processing by searching ordered data, e.g. alpha-numerically ordered data

Definitions

  • the present application relates to the technical field of intelligent decision-making in artificial intelligence, and in particular, to a data detection method, device, server and storage medium based on artificial intelligence.
  • the server needs to monitor the operating data in real time, so as to detect the abnormal state of the application and the server hardware in time and issue an alarm.
  • the anomaly detection model can be used to monitor the indicators of multiple objects in the server system, such as applications and hardware, so as to determine whether the applications and hardware are in an abnormal state.
  • the inventor realized that the existing anomaly detection model is only obtained by using historical time series data as training samples for training, and data deviation is likely to occur when predicting current time series data, resulting in a decrease in the accuracy of anomaly prediction, and the increase of abnormal data.
  • the detection effect is not good.
  • the main purpose of the present application is to provide a data detection method, device, server and storage medium based on artificial intelligence, aiming at improving the detection accuracy of abnormal data.
  • the present application provides a data detection method based on artificial intelligence, which is applied to a server, where the server stores a time series data prediction model, and the time series data prediction model includes a pre-trained first Transformer model and a second Transformer. model, the method includes:
  • the historical sequence vector and the current sequence vector are associated to obtain an associated sequence vector
  • the present application also provides a data detection device based on artificial intelligence, which is applied to a server, where the server stores a time series data prediction model, and the time series data prediction model includes a pre-trained first Transformer model and a second Transformer model, the data detection device includes:
  • the acquisition module is used to acquire the current time series data of the target detection indicator
  • a generating module is used to generate a current time sequence vector according to the current time sequence data, and determine the historical time sequence vector of the historical time sequence data of the target detection index;
  • an association module for associating the historical sequence vector with the current sequence vector through the first Transformer model to obtain an associated sequence vector
  • a prediction module configured to input the associated time series vector into the second Transformer model to obtain the prediction data of the next time series of the current time series data
  • the acquisition module is further configured to acquire target time series data of the next time series of the current time series data
  • a detection module configured to perform anomaly detection on the target time series data through the predicted data and the target time series data.
  • the present application further provides a server, the server includes a processor, a memory, and a computer program stored on the memory and executable by the processor, the server stores a time series data prediction model,
  • the time series data prediction model includes a pre-trained first Transformer model and a second Transformer model, and when the computer program is executed by the processor, the following steps are implemented:
  • the historical sequence vector and the current sequence vector are associated to obtain an associated sequence vector
  • the present application further provides a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, wherein when the computer program is executed by the processor, the following steps are implemented:
  • the historical sequence vector and the current sequence vector are associated to obtain the associated sequence vector
  • the present application provides a data detection method, device, server and storage medium based on artificial intelligence.
  • the present application obtains the current time series data of the target detection index, generates the current time series vector according to the current time series data, and determines the historical time series data of the target detection index.
  • the Transformer model is used to process the current time series vector and the historical time series vector, and output prediction data with higher accuracy, which greatly improves the detection accuracy of abnormal data.
  • FIG. 1 is a schematic flowchart of steps of a data detection method based on artificial intelligence provided by an embodiment of the present application
  • Fig. 2 is a schematic flow chart of sub-steps of the data detection method in Fig. 1;
  • FIG. 3 is a schematic diagram of a scenario for implementing the output associated time series data provided by this embodiment
  • FIG. 4 is a schematic block diagram of a data detection device based on artificial intelligence provided by an embodiment of the present application
  • Fig. 5 is a schematic block diagram of sub-modules of the data detection device in Fig. 4;
  • FIG. 6 is a schematic block diagram of the structure of a server according to an embodiment of the present application.
  • Embodiments of the present application provide an artificial intelligence-based data detection method, device, server, and storage medium.
  • the data detection method can be applied to a server, and the server can be a single server or a server cluster composed of multiple servers.
  • FIG. 1 is a schematic flowchart of steps of an artificial intelligence-based data detection method provided by an embodiment of the present application.
  • the data detection method includes steps S101 to S105.
  • Step S101 acquiring current time series data of the target detection index.
  • the abnormal fluctuation of the server performance index may reflect the abnormal state of the application, so the abnormality can be detected through the time series data of the server performance index.
  • the target detection indicator in this embodiment may include, for example, at least one of the following indicators: CPU usage, IO usage, memory usage, bandwidth usage, network ingress traffic, network egress traffic, access volume, access time-consuming, Downloads, new users, and active users.
  • the time series data refers to time series data, for example, a data column recorded by the same indicator in time sequence, and the time series data may be the number of periods or the number of points in time.
  • the current time series data includes the time series data of the target detection indicators in the current time series, and the current time series can be flexibly set by the user. For example, the current time series data is the time series data of multiple memory usage rates recorded within one hour.
  • the current time series data in this embodiment may be time series data of a single target detection indicator, such as time series data of CPU usage, or time series data of bandwidth usage;
  • the current time series data may also be time series data of multiple target detection indicators, for example, time series data may include memory usage and network egress traffic.
  • the current time series data of the target detection index is obtained, and the historical time series data of the target detection index is obtained.
  • the historical time series data is the time series data of target detection indicators within a preset time period, and the historical time series data in this embodiment may also be time series data about a single target detection indicator or multiple target detection indicators, and can be based on actual settings.
  • the historical time series data is long time series data
  • the current time series data is short time series data.
  • the time length corresponding to the historical time series data is greater than or equal to the first preset time length
  • the time length corresponding to the current time series data is less than or equal to the second preset time length
  • the time length corresponding to the historical time series data is greater than the time corresponding to the current time series data. length, the first preset time length and the second preset time length can be set according to the actual situation.
  • the time length corresponding to the historical time series data is 1 week; if the current time series data is the time series of the target detection indicators recorded in the current 1 hour sequence data, the time length corresponding to the current time series data is 1 hour.
  • target detection indicators are often associated with the distribution of historical data.
  • short time series data for anomaly detection may reduce the accuracy of anomaly detection, but using long time series data for anomaly detection Detection will bring a large amount of calculation. Therefore, in this embodiment, the historical time series data of the long time series data and the current time series data of the short time series data are used to predict the next time series, the abnormality detection accuracy is high, and the historical time series data of the long time series data can be updated at a lower frequency , thereby saving computation.
  • Step S102 generating a current time series vector according to the current time series data, and determining a historical time series vector of the historical time series data of the target detection index.
  • the current time series data is converted into the current time series vector, and the historical time series vector of the historical time series data of the target detection index is determined.
  • the historical time series vector can be generated in real time, or can be directly obtained from the historical data.
  • the historical time series data of the detection indicators are converted into the process information and result information of the historical time series vectors, so the historical time series vectors can be reused instead of being generated in real time, thereby saving the amount of calculation and improving the efficiency of anomaly detection.
  • the time length corresponding to the historical sequence vector is greater than or equal to the first preset time length, and the time length corresponding to the current sequence vector is less than or equal to the second preset time length. It should be noted that the time length corresponding to the historical time series vector generated by using the historical time series data of the long time series data is greater than or equal to the first preset time length, and the time length corresponding to the current time series vector generated by using the current time series data of the short time series data.
  • the time length is less than or equal to the second preset time length, the accuracy of anomaly detection can be improved, and the historical time series data of long time series data can be updated at a lower frequency and need not be changed for a long period of time, thereby saving the amount of calculation.
  • the time length corresponding to the generated historical time series vector is 1 week; if the current time series data is the target detection indicators recorded in the current 1 hour. For time series data, the time length corresponding to the generated current time series vector is 1 hour.
  • determining the historical sequence vector of the target detection index includes: sub-steps S1021 to S1023 .
  • Sub-step S1021 Acquire historical time series data of the target detection index, and determine first time information corresponding to the historical time series data.
  • the historical time series data may include multiple detection data of target detection indicators, and the first time information includes a first time series, a first time interval, a first time interval, and/or a first time point of each detection data.
  • the first time series is the time length corresponding to the historical time series data.
  • the first time series is 1 week, that is, the historical time series data includes multiple detection data of the target detection indicators within the range of one week; the first time interval is the corresponding time series of the historical time series data.
  • the interval range of the time length if the first time interval is (14th, 21st), that is, the historical time series data includes multiple detection data of the target detection indicators within the 14th to 21st days; the first time interval is the historical time series data.
  • the first time information includes m time point information of historical time series data, and k and m are positive integers greater than or equal to 1; wherein, each time point information corresponds to one detection data , m time point information indicates that there are m detection data in the historical time series data, each target detection index corresponds to one time series data, and k target detection indicators indicate that there are k*m detection data in the historical time series data.
  • Sub-step S1022 Generate a first time series vector according to the historical time series data and the first time information, and generate a first correction vector according to the historical time series data, the first time information and the preset function.
  • the preset function includes a sine function, a cosine function, or a combined function of a sine function and a cosine function, which is not specifically limited in this embodiment.
  • the first time series vector is a matrix vector of k*m
  • the first correction vector is a matrix vector of e*m.
  • the first time information includes m time point information of historical time series data, and k and m are positive integers greater than or equal to 1; the first time information is generated according to the historical time series data and the first time information.
  • a time series vector comprising: arranging historical time series data of k target detection indicators according to m time point information to generate a k*m matrix vector, and using the matrix vector as the first time series vector.
  • the historical time series data may include multiple detection data of target detection indicators, each time point information corresponds to one detection data, and m time point information indicates that there are m detection data in the historical time series data, and each target detection indicator Corresponding to a time series data, k target detection indicators indicate that there are k*m detection data in the historical time series data.
  • the matrix vector of k*m is the first time sequence vector, and k represents the width or length of the matrix vector.
  • generating the first correction vector according to the historical time series data, the first time information and the preset function includes: obtaining a preset function, and substituting the historical time series data and the first time information into the preset function to obtain a historical Correction information of the time series data; generate a first correction vector according to the historical time series data and the correction information of the historical time series data.
  • the preset function is a sine function, a cosine function, or a combined function of a sine function and a cosine function including time characteristic parameters, for example, the preset function is A cos(2 ⁇ xt), where A is a preset coefficient and t is a time characteristic parameter , x is an unknown parameter set, which can be used to characterize the detection data in the historical time series data.
  • the historical time series data and the first time information are substituted into the preset function, the correction information of the detection data at different time points is calculated, and the correction information of the detection data at different time points is arranged to obtain the first correction. vector, where the first correction vector is used to correct the first time series vector, so that the subsequently calculated prediction data is more accurate, and the accuracy of abnormality detection in this embodiment of the present application is also higher.
  • the number of correction information calculated by the preset function is e
  • the first time information includes m time point information of the historical time series data; according to the m time point information, the e correction information are arranged to generate The matrix vector of e*m, and the matrix vector is used as the first correction vector.
  • e represents the length of the matrix vector
  • m represents the width of the matrix vector
  • e represents the width of the matrix vector, which is not specifically limited in this embodiment.
  • Sub-step S1023 splicing the first time sequence vector and the first correction vector to obtain a historical time sequence vector.
  • the splicing method includes row splicing or column splicing, which can be flexibly used and transposed according to the actual situation.
  • the first time series vector is a matrix vector of k*m
  • the first correction vector is a matrix vector of e*m
  • the matrix vector of (k+e)*m is obtained by splicing
  • the matrix vector of (k+e)*m is obtained by splicing vector as the historical time series vector.
  • the first time series vector is a matrix vector of m*k
  • the first correction vector is a matrix vector of m*e
  • the matrix vector of m*(k+e) is obtained by splicing, and the m*(k+e) The matrix vector of , as the historical time series vector.
  • generating the current time sequence vector according to the current time sequence data includes: acquiring second time information corresponding to the current time sequence data; generating a second time sequence vector according to the current time sequence data and the second time information, The second time information and the preset function are used to generate the second correction vector; the second time sequence vector and the second correction vector are spliced to obtain the current time sequence vector.
  • the current time series data may include a plurality of detection data of target detection indicators
  • the second time information includes a second time series, a second time interval, a second time interval and/or a second time point of each detection data.
  • the second time sequence vector is a matrix vector of k*n
  • the second correction vector is a matrix vector of e*n
  • the matrix vector of (k+e)*n is obtained by splicing
  • the (k+e)*n matrix vector is obtained by splicing.
  • the matrix vector of as the historical time series vector.
  • a historical time series vector of historical time series data of the target detection indicator is obtained from a memory. It should be noted that the embodiment of the present application uses historical time series data of long time series data to generate historical time series vectors, the historical time series vectors generated historically can be stored in a memory, and the historical time series vectors can be stored in a long period of time (preset time period). Multiplexing, it can be called directly through the memory when it is used again, without frequent operations, which saves the amount of calculation.
  • Step S103 associate the historical sequence vector with the current sequence vector through the first Transformer model to obtain the associated sequence vector.
  • the server stores a time series data prediction model, and the time series data prediction model includes a pre-trained first Transformer model and a second Transformer model.
  • the first Transformer model includes a Transformer operation layer and a feedforward neural network layer FNN; through the Transformer operation layer, the historical time series vector and the current time series vector are operated to obtain associated time series data; through the feedforward neural network layer, The associated time series data is mapped to obtain the associated time series vector.
  • the Transformer operation layer is used to perform the Transformer operation on the historical time series vector and the current time series vector
  • the feedforward neural network layer FNN is used to perform a fully connected layer mapping on the operation results output by the Transformer operation layer, and output the associated time series data to achieve historical time series.
  • the correlation between the target detection indicators in the vector and the current time series vector will cover the current time series data and historical time series data after the two are associated, so as to predict the next time series more accurately and improve the accuracy of anomaly detection.
  • the associated time sequence (x3) of the historical time sequence vector history_window and the current time sequence vector current_window is calculated through the Transformer operation layer and the feedforward neural network layer FNN.
  • the historical time series vector history_window and the current time series vector current_window are input to the first Transformer model, and the first Transformer model includes a Transformer operation layer and a feedforward neural network layer FNN.
  • the number of first Transformer timing models may be n, where n is a positive integer greater than or equal to 2. If the number of the first Transformer timing model can be n, the input historical timing vector and the current timing vector will be calculated through each first Transformer timing model in turn according to the set level, until the nth first Transformer timing model completes the calculation And output to get the associated time series data.
  • Step S104 input the associated time series vector into the second Transformer model to obtain the prediction data of the next time series of the current time series data.
  • the second Transformer model has the same structure as the aforementioned first Transformer model, and the two have commonalities, including the Transformer operation layer and the feedforward neural network layer FNN, and the difference lies in the different model parameters.
  • the associated time series vector is processed by the second Transformer model, and the prediction data of the next time series of the current time series data is output.
  • the current time series data is associated with a similar time period in the historical time series data, so as to obtain the prediction data of the next time series of the current time series data.
  • the associated time series data x3 is input into the second Transformer time series model, that is, the associated time series data x3 is subjected to a Transformer operation again, and the fully connected layer mapping Dense is performed through the feedforward neural network layer FNN, and finally the output is completed to obtain the current time series. Forecast data for the next time series of data
  • Step S105 acquiring target time series data of the next time series of the current time series data, and performing anomaly detection on the target time series data through the prediction data and the target time series data.
  • the target time series data is the time series data of the next time series of the current time series data, wherein the current time series data corresponds to the second time series, and the second time series is the time length corresponding to the current time series data , for example, the second time series is 1 hour, that is, the current time series data includes multiple detection data of target detection indicators within 1 hour.
  • the time length corresponding to the target time series data is the same as the time length corresponding to the current time series data, that is, the target time series data corresponds to the second time series, for example, including multiple detection data of target detection indicators within 1 hour.
  • the target time series data is the time series data of the next time series of the current time series data.
  • the time interval corresponding to the target time series data is obtained by adding the second time series to the second time interval corresponding to the current time series data. For example, when the second time series is 1 hour, The second time interval corresponding to the current time series data is (12 o'clock, 13 o'clock), and the time interval corresponding to the target time series data is (13 o'clock, 14 o'clock).
  • the predicted data and the target time series data are compared to obtain a relative error between the predicted data and the target time series data; an abnormality detection is performed on the target time series data according to the relative error to obtain an abnormality detection result.
  • the abnormality detection result includes abnormal data and normal data.
  • the abnormality detection result is a data abnormality; when the relative error between the predicted data and the target time series data is less than the preset error threshold, the abnormality detection result The data is normal.
  • the present application uses the associated time series vector of the current time series data and the historical time series data to predict the target time series data of the next time series of the current time series data, and the predicted data can adapt to the change trend of the target detection index itself over time, which improves the accuracy of anomaly detection. Rate.
  • an alarm when it is determined that the target time series data is abnormal data, an alarm may be performed.
  • the server can play an abnormal data alarm sound, or light up an abnormal data warning light, or send the corresponding alarm information to the relevant responsible personnel.
  • the notification methods of the alarm information include but are not limited to WeChat, SMS, email, enterprise-level communication platform, etc.
  • the artificial intelligence-based data detection method obtains the current time series data of the target detection index, generates the current time series vector according to the current time series data, and determines the historical time series vector of the historical time series data of the target detection index, and then passes the first time series data.
  • the Transformer model associates the historical time series vector with the current time series vector to obtain the associated time series vector, and then inputs the associated time series vector into the second Transformer model to obtain the prediction data of the next time series of the current time series data, and finally obtains the next time series of the current time series data.
  • target time series data of a time series and perform anomaly detection on the target time series data through the prediction data and the target time series data.
  • the Transformer model is used to process the current time series vector and the historical time series vector, and output prediction data with higher accuracy, which greatly improves the detection accuracy of abnormal data.
  • the data detection method further includes: acquiring a plurality of sample data, the sample data includes first time series data, second time series data and marked time series data, and the collection time of the first time series data is later than the second time series data.
  • Through the first preset Transformer model associate the first time sequence vector and the second time sequence vector to obtain a target time sequence vector;
  • Input the target time series vector into the second preset Transformer model to obtain the predicted time series data; update the model parameters of the first preset Transformer model and the second preset Transformer model according to the predicted time series data and the marked time series data; continue to
  • the updated first preset Transformer model and the second preset Transformer model are iteratively trained until the first preset Transformer model and the second preset Transformer model converge, and a time series data prediction model is obtained.
  • the time-series data prediction model is obtained by training the associated time-series vectors of the first time-series data and the second time-series data, which can effectively improve the effect of model training, and the trained time-series data prediction model can be used to detect abnormal data.
  • the collection time of the first time series data is later than that of the second time series data.
  • the first time series data is, for example, current time series data
  • the second time series data is, for example, historical time series data.
  • the current time series data includes time series data of target detection indicators in the current time series.
  • the current time series data is time series data of multiple memory usage rates recorded within one hour.
  • the historical time series data includes the time series data of the target detection indicators in the historical time series, for example, the historical time series data is the time series data of the target detection indicators recorded in the past one week.
  • the second time series data is long time series data
  • the first time series data is short time series data.
  • the time length corresponding to the second time series data is greater than or equal to the first preset time length
  • the time length corresponding to the first time series data is less than or equal to the second preset time length
  • the time length corresponding to the second time series data is greater than the first time series.
  • the time length corresponding to the data, the first preset time length and the second preset time length can be set according to the actual situation.
  • the second time series data of the long time series data and the first time series data of the short time series data are used to predict the next time series, the abnormality detection accuracy is high, and the second time series data of the long time series data can be Updates are done less frequently, saving computation.
  • the first time series data is converted into a first time series vector
  • the second time series data is converted into a second time series vector.
  • the second time series vector can be generated from the second time series data and stored in the historical data of the memory, and can be directly obtained from the historical data in subsequent use, so the second time series vector can be reused without repeated generation, thus saving calculation and improve the efficiency of model training.
  • the time length corresponding to the second time sequence vector is greater than or equal to the first preset time length, and the time length corresponding to the first time sequence vector is less than or equal to the second preset time length. It should be noted that the time length corresponding to the second time series vector generated by the second time series data of the long time series data is greater than or equal to the first preset time length, while the first time series vector generated by using the first time series data of the short time series data The corresponding time length is less than or equal to the second preset time length, which can improve the accuracy of data anomaly detection assisted by the trained model, and the second time series data of long time series data can be updated at a lower frequency, and in a longer period of time. No changes are required over time, thus saving computation.
  • generating the second time series vector according to the second time series data includes: acquiring the second time series data of the target detection index, and determining the first time information corresponding to the second time series data; according to the second time series data and the first time series data; The time information generates a first time sequence vector, and a first correction vector is generated according to the second time sequence data, the first time information and the preset function; the first time sequence vector and the first correction vector are spliced to obtain a second time sequence vector.
  • generating the first time series vector according to the first time series data includes: acquiring second time information corresponding to the first time series data; generating a second time series vector according to the first time series data and the second time information, and A second time series vector is generated from the time series data, the second time information and the preset function; the second time series vector and the second correction vector are spliced to obtain the first time series vector.
  • the first time series data may include a plurality of detection data of target detection indicators
  • the second time information includes a second time series, a second time interval, a second time interval and/or a second time point of each detection data.
  • the first preset Transformer sequence model may be one or more, and the first preset Transformer model includes a Transformer operation layer and a feedforward neural network layer FNN; through the Transformer operation layer, the second sequence vector and the first sequence vector are processed operation to obtain the associated time series data; through the feedforward neural network layer, the associated time series data is mapped to obtain the associated time series vector.
  • the Transformer operation layer is used to perform Transformer operation on the second time series vector and the first time series vector
  • the feedforward neural network layer FNN is used to perform a fully connected layer mapping on the operation results output by the Transformer operation layer, and output the associated time series data to achieve
  • the correlation between the second time series vector and the target detection indicators in the first time series vector, after the two are correlated, will cover the first time series data and the second time series data, so as to predict the next time series more accurately and improve the accuracy of abnormal detection.
  • the structure of the second preset Transformer model is the same as that of the first preset Transformer model described above, and the two have commonalities, including the Transformer operation layer and the feedforward neural network layer FNN, the difference lies in the different model parameters.
  • the associated time series vector is processed by the second preset Transformer model, and the predicted time series data of the next time series is output.
  • the first time series data is associated with a similar time period in the second time series data, thereby obtaining the predicted time series data of the next time series of the first time series data.
  • the model parameters of the first preset Transformer model and the second preset Transformer model are updated, and the updated first preset Transformer model and the second preset Transformer model are continued to be updated. Iterative training is performed until the first preset Transformer model and the second preset Transformer model converge, and a time series data prediction model is obtained.
  • the time series data prediction model is obtained by training the associated time series vector of the first time series data and the second time series data, the model training effect is better, and the detection accuracy of abnormal data is higher.
  • the first loss function of the first Transformer preset model and the second model parameter of the second preset Transformer model are calculated through the predicted time series data and the labeled time series data; the first loss function is adjusted based on the first loss function.
  • the trained first preset Transformer model and the second preset Transformer model are in a convergent state, and a time series data prediction model is obtained.
  • the first preset Transformer model and the second preset Transformer model are fused to obtain the target Transformer model; through the predicted time series data and the marked time series data, the loss function of the target Transformer model is calculated, and based on the loss The function determines the model parameters of the target Transformer model, iteratively trains the target Transformer model according to the model parameters, and determines whether the trained target Transformer model is in a convergent state; if it is determined that the trained target Transformer model is in a convergent state, the time series data prediction model is obtained.
  • the loss function is, for example, an MSE (mean-square error, mean square error) loss function, and the Adam optimization algorithm is used to iteratively train the target Transformer model.
  • the trained abnormal data detection model can predict the next window of time series data (target time series data of the next time series of the first time series data) based on the input second time series data and the first time series data.
  • the adaptive gradient algorithm AdaGrad
  • RMSProp root mean square propagation
  • FIG. 4 is a schematic block diagram of an artificial intelligence-based data detection apparatus provided by an embodiment of the present application.
  • the artificial intelligence-based data detection device 300 includes: an acquisition module 301 , a generation module 302 , an association module 303 , a prediction module 304 and a detection module 305 .
  • the data detection apparatus 300 can be applied to a server, where the server stores a time series data prediction model, and the time series data prediction model includes a pre-trained first Transformer model and a second Transformer model, wherein:
  • an acquisition module 301 configured to acquire current time series data of the target detection index
  • a generating module 302 configured to generate a current time sequence vector according to the current time sequence data, and determine the historical time sequence vector of the historical time sequence data of the target detection index;
  • the association module 303 is used to associate the historical sequence vector with the current sequence vector through the first Transformer model to obtain an associated sequence vector;
  • a prediction module 304 configured to input the associated time series vector into the second Transformer model to obtain the prediction data of the next time series of the current time series data
  • the obtaining module 301 is further configured to obtain the target sequence data of the next sequence of the current sequence data
  • the detection module 305 is configured to perform anomaly detection on the target time series data by using the predicted data and the target time series data.
  • the time length corresponding to the historical sequence vector is greater than or equal to the first preset time length, and the time length corresponding to the current sequence vector is less than or equal to the second preset time length.
  • the generating module 302 includes:
  • the acquisition sub-module 3021 is used to acquire the historical time series data of the target detection index, and determine the first time information corresponding to the historical time series data;
  • Generating submodule 3022 for generating a first time sequence vector according to the historical time sequence data and the first time information, and generating a first correction vector according to the historical time sequence data, the first time information and a preset function;
  • the splicing sub-module 3023 is used for splicing the first time sequence vector and the first correction vector to obtain a historical time sequence vector.
  • the target detection indicators are k
  • the first time information includes m time point information of historical time series data
  • k and m are positive integers greater than or equal to 1
  • the generating module 302 is further configured to:
  • the historical time series data of the k target detection indicators are arranged to generate a matrix vector of k*m, and the matrix vector is used as the first time series vector.
  • the generation module 302 is also used to:
  • a first correction vector is generated according to the historical time series data and the correction information of the historical time series data.
  • the first Transformer model includes a Transformer operation layer and a feedforward neural network layer; the association module 303 is also used for:
  • the historical time sequence vector and the current time sequence vector are operated to obtain associated time sequence data
  • the associated time series data is mapped to obtain an associated time series vector.
  • the data detection apparatus 300 is further used for:
  • the sample data includes first time series data, second time series data and marked time series data, the collection time of the first time series data is later than the second time series data;
  • the first time sequence vector and the second time sequence vector are associated to obtain a target time sequence vector
  • the apparatuses provided by the above embodiments may be implemented in the form of a computer program, and the computer program may be executed on the server as shown in FIG. 6 .
  • FIG. 6 is a schematic block diagram of the structure of a server according to an embodiment of the present application.
  • the server stores a time series data prediction model, and the time series data prediction model includes a pre-trained first Transformer model and a second Transformer model.
  • the server includes a processor, a memory and a network interface connected through a system bus, wherein the memory may include a storage medium and an internal memory, and the storage medium may be non-volatile or volatile.
  • the storage medium may store an operating system and a computer program.
  • the computer program includes program instructions, which, when executed, can cause the processor to execute any data detection method based on artificial intelligence.
  • the processor is used to provide computing and control capabilities to support the operation of the entire server.
  • the internal memory provides an environment for running the computer program in the storage medium.
  • the processor can execute any artificial intelligence-based data detection method.
  • the network interface is used for network communication, such as sending assigned tasks.
  • the network interface is used for network communication, such as sending assigned tasks.
  • FIG. 6 is only a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the server to which the solution of the present application is applied. More or fewer components are shown in the figures, either in combination or with different arrangements of components.
  • the processor may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general-purpose processors, digital signal processors (Digital Signal Processors, DSP), application specific integrated circuits (Application Specific Integrated circuits) Circuit, ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor can be a microprocessor or the processor can also be any conventional processor or the like.
  • the processor is configured to run a computer program stored in the memory to implement the following steps:
  • the historical sequence vector and the current sequence vector are associated to obtain an associated sequence vector
  • the time length corresponding to the historical sequence vector is greater than or equal to the first preset time length, and the time length corresponding to the current sequence vector is less than or equal to the second preset time length.
  • the processor when implementing the determining the historical time series vector of the historical time series data of the target detection index, is configured to:
  • the first time sequence vector and the first correction vector are spliced to obtain a historical time sequence vector.
  • the target detection indicators are k
  • the first time information includes m time point information of historical time series data, and k and m are positive integers greater than or equal to 1;
  • the processor When implementing the generation of the first time sequence vector according to the historical time sequence data and the first time information, the processor is configured to implement:
  • the historical time series data of the k target detection indicators are arranged to generate a matrix vector of k*m, and the matrix vector is used as the first time series vector.
  • the processor when the processor generates the first correction vector according to the historical time series data, the first time information and the preset function, the processor is configured to:
  • a first correction vector is generated according to the historical time series data and the correction information of the historical time series data.
  • the first Transformer model includes a Transformer operation layer and a feedforward neural network layer; the processor is performing the process of passing the first Transformer model and performing the process between the historical sequence vector and the current sequence vector. Association, when the associated timing vector is obtained, it is used to realize:
  • the historical time sequence vector and the current time sequence vector are operated to obtain associated time sequence data
  • the associated time series data is mapped to obtain an associated time series vector.
  • the processor is further configured to:
  • the sample data includes first time series data, second time series data and marked time series data, the collection time of the first time series data is later than the second time series data;
  • the first time sequence vector and the second time sequence vector are associated to obtain a target time sequence vector
  • Embodiments of the present application further provide a computer-readable storage medium, where a computer program is stored on the computer-readable storage medium, wherein when the computer program is executed by a processor, the following steps are implemented:
  • the historical sequence vector and the current sequence vector are associated to obtain the associated sequence vector
  • the computer-readable storage medium may be an internal storage unit of the server described in the foregoing embodiments, such as a hard disk or a memory of the server.
  • the computer-readable storage medium may also be an external storage device of the server, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card equipped on the server , Flash Card (Flash Card) and so on.

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Supply And Distribution Of Alternating Current (AREA)

Abstract

一种基于人工智能的数据检测方法、装置、服务器及存储介质,该方法包括:获取目标检测指标的当前时序数据(S101);根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量(S102);通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量(S103);将关联时序向量输入至所述第二Transformer模型,得到当前时序数据的下一时序的预测数据(S104);获取当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测(S105)。该方法能够提高异常数据的检测准确度。

Description

基于人工智能的数据检测方法、装置、服务器及存储介质
本申请要求于2020年12月31日提交中国专利局、申请号为202011639634.3、发明名称为“基于人工智能的数据检测方法、装置、服务器及存储介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能中的智能决策的技术领域,尤其涉及一种基于人工智能的数据检测方法、装置、服务器及存储介质。
背景技术
随着计算机科学技术的不断发展,互联网中越来越多的应用运行于服务器上,会产生大量的运行数据。运行数据的异常波动有可能反应了应用或者服务器硬件的异常状态,因此服务器需要对运行数据进行实时监测,从而及时发现应用和服务器硬件的异常状态并进行告警。目前,可以通过异常检测模型负责监控服务器系统中关于应用和硬件等多个对象的指标,从而确定应用和硬件是否存在异常状态。然而,发明人意识到现有的异常检测模型仅是利用历史时序数据作为训练样本进行训练得到的,在对当前时序数据进行预测时容易发生数据偏离,导致异常预测的精准度下降,异常数据的检测效果不好。
发明内容
本申请的主要目的在于提供一种基于人工智能的数据检测方法、装置、服务器及存储介质,旨在提高异常数据的检测准确度。
第一方面,本申请提供一种基于人工智能的数据检测方法,应用于服务器,所述服务器存储有时序数据预测模型,所述时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型,所述方法包括:
获取目标检测指标的当前时序数据;
根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
将所述关联时序向量输入至所述第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数 据,对所述目标时序数据进行异常检测。
第二方面,本申请还提供一种基于人工智能的数据检测装置,应用于服务器,所述服务器存储有时序数据预测模型,所述时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型,所述数据检测装置包括:
获取模块,用于获取目标检测指标的当前时序数据;
生成模块,用于根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
关联模块,用于通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
预测模块,用于将所述关联时序向量输入至所述第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
所述获取模块,还用于获取所述当前时序数据的下一时序的目标时序数据;
检测模块,用于通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
第三方面,本申请还提供一种服务器,所述服务器包括处理器、存储器、以及存储在所述存储器上并可被所述处理器执行的计算机程序,所述服务器存储有时序数据预测模型,所述时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型,其中所述计算机程序被所述处理器执行时,实现以下步骤:
获取目标检测指标的当前时序数据;
根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
将所述关联时序向量输入至所述第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
第四方面,本申请还提供一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,其中所述计算机程序被处理器执行时,实现以下步骤:
获取目标检测指标的当前时序数据;
根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
通过预先训练好的第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
将所述关联时序向量输入至预先训练好的第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
本申请提供一种基于人工智能的数据检测方法、装置、服务器及存储介质,本申请通过获取目标检测指标的当前时序数据,根据当前时序数据生成当前时序向量,并确定目标检测指标的历史时序数据的历史时序向量,再通过第一Transformer模型,将历史时序向量和当前时序向量进行关联,得到关联时序向量,然后将关联时序向量输入至第二Transformer模型,得到当前时序数据的下一时序的预测数据,最后获取当前时序数据的下一时序的目标时序数据,并通过预测数据和目标时序数据,对所述目标时序数据进行异常检测。通过Transformer模型对当前时序向量和历史时序向量进行处理,输出准确度更高的预测数据,极大提升了异常数据的检测准确度。
附图说明
为了更清楚地说明本申请实施例技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1为本申请实施例提供的一种基于人工智能的数据检测方法的步骤流程示意图;
图2为图1中的数据检测方法的子步骤流程示意图;
图3为实施本实施例提供的输出关联时序数据的一场景示意图;
图4为本申请实施例提供的一种基于人工智能的数据检测装置的示意性框图;
图5为图4中的数据检测装置的子模块的示意性框图;
图6为本申请实施例提供的一种服务器的结构示意性框图。
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本申请保护的范围。
附图中所示的流程图仅是示例说明,不是必须包括所有的内容和操作/步骤,也不是必须按所描述的顺序执行。例如,有的操作/步骤还可以分解、组合或部分合并,因此实际执行的顺序有可能根据实际情况改变。另外,虽然在装置示意图中进行了功能模块的划分,但是在某些情况下,可以以不同于装置示意图中的模块划分。
本申请实施例提供一种基于人工智能的数据检测方法、装置、服务器及存储介质。其中,该数据检测方法可应用于服务器中,该服务器可以为单台的服务器,也可以为由多台服务器组成的服务器集群。
下面结合附图,对本申请的一些实施方式作详细说明。在不冲突的情况下,下述的实施例及实施例中的特征可以相互组合。
请参照图1,图1为本申请实施例提供的一种基于人工智能的数据检测方法的步骤流程示意图。
如图1所示,该数据检测方法包括步骤S101至步骤S105。
步骤S101、获取目标检测指标的当前时序数据。
当应用程序运行于服务器上时,服务器性能指标的异常波动可能反应了应用的异常状态,因此可以通过服务器性能指标的时序数据来检测异常。本实施例中的目标检测指标例如可以包括以下指标中的至少一种:CPU使用率、IO使用率、内存使用率、带宽使用率、网络入口流量、网络出口流量、访问量、访问耗时、下载量、新增用户数和活跃用户。其中,时序数据是指时间序列数据,例如为同一指标按时间顺序记录的数据列,时序数据可以是时期数,也可以时点数。当前时序数据包括当前时序内的目标检测指标的时间序列数据,当前时序可由用户灵活设置,例如,当前时序数据为1小时内记录的多个内存使用率的时间序列数据。
需要说明的是,本实施例中的当前时序数据可以是单个目标检测指标的时序数据,例如可以是CPU使用率的时间序列数据,或者可以是带宽使用率的时间序列数据;本实施例中的当前时序数据还可以是多个目标检测指标的时序数据,例如可以包括内存使用率和网络出口流量的时间序列数据。
在一实施例中,获取目标检测指标的当前时序数据,并获取目标检测指标的历史时序数据。同理,历史时序数据为预设时间段内的目标检测指标的时序数据,本实施例中的历史时序数据也可以是关于单个目标检测指标或者多个目标检测指标的时间序列数据,并可以根据实际情况进行设置。
进一步地,历史时序数据为长时序数据,当前时序数据为短时序数据。其中,历史时序数据对应的时间长度大于或等于第一预设时间长度,当前时序数据对应的时间长度小于或等于第二预设时间长度,历史时序数据对应的时间长度大于当前时序数据对应的时间长度,第一预设时间长度和第二预设时间长度可根据实际情况进行设置。例如,若历史时序数据为过去1周内记录的目标检测指标的时间序列数据,则该历史时序数据对应的时间长度为1周;若当前时序数据为当前1小时内记录的目标检测指标的时间序列数据,则该当前时序数据对应的时间长度为1小时。
需要说明的是,通过异常检测模型对时序数据进行异常检测时,目标检测指标往往与历史数据的分布相关联,采用短时序数据进行异常检测可能降低异常检测的精度,但是采 用长时序数据进行异常检测会带来较大的计算量。因此,本实施例采用长时序数据的历史时序数据和短时序数据的当前时序数据来以预测下一时序,异常检测的精度较高,且长时序数据的历史时序数据能够以较低频率进行更新,从而节省计算量。
步骤S102、根据当前时序数据生成当前时序向量,并确定目标检测指标的历史时序数据的历史时序向量。
其中,将当前时序数据转化为当前时序向量,以及确定目标检测指标的历史时序数据的历史时序向量,历史时序向量可以是实时生成的,也可以从历史数据中直接获取,该历史数据记录有目标检测指标的历史时序数据转化为历史时序向量的过程信息和结果信息,因此历史时序向量可以复用而不必实时生成,从而节省计算量和提高异常检测效率。
在一实施例中,历史时序向量对应的时间长度大于或等于第一预设时间长度,当前时序向量对应的时间长度小于或等于第二预设时间长度。需要说明的是,采用长时序数据的历史时序数据生成的历史时序向量对应的时间长度大于或等于第一预设时间长度,而采用短时序数据的当前时序数据生成的当前时序向量对应的时间长度小于或等于第二预设时间长度,能够提高异常检测的准确度,且长时序数据的历史时序数据能够以较低频率进行更新,在较长一段时间内无需改变,从而节省计算量。
例如,若历史时序数据为过去1周内记录的目标检测指标的时间序列数据,则生成的历史时序向量对应的时间长度为1周;若当前时序数据为当前1小时内记录的目标检测指标的时间序列数据,则生成的当前时序向量对应的时间长度为1小时。
在一实施例中,如图2所示,确定目标检测指标的历史时序向量,包括:子步骤S1021至子步骤S1023。
子步骤S1021、获取目标检测指标的历史时序数据,并确定历史时序数据对应的第一时间信息。
其中,历史时序数据可以包括目标检测指标的多个检测数据,第一时间信息包括第一时序、第一时间区间、第一时间间隔和/或每个检测数据的第一时间点。第一时序为历史时序数据对应的时间长度,如第一时序为1周,即该历史时序数据包括1周范围内的目标检测指标的多个检测数据;第一时间区间为历史时序数据对应的时间长度的区间范围,如第一时间区间为(14日,21日),即该历史时序数据包括14日至21日内的目标检测指标的多个检测数据;第一时间间隔为历史时序数据中的每两个检测数据之间的采集时间间隔;第一时间点为历史时序数据中的每个检测数据对应的时间点,第一时序与第一时间间隔之间的比值为历史时序数据中的检测数据的数量。
在一实施例中,目标检测指标为k个,第一时间信息包括历史时序数据的m个时间点信息,k、m为大于等于1的正整数;其中,每个时间点信息对应一个检测数据,m个时间点信息表示历史时序数据中存在m个检测数据,每个目标检测指标对应一个时序数据,k个目标检测指标表示历史时序数据共存在k*m个检测数据。
子步骤S1022、根据历史时序数据和第一时间信息生成第一时序向量,并根据历史时序数据、第一时间信息和预设函数生成第一修正向量。
其中,预设函数包括正弦函数、余弦函数、或者正弦函数与余弦函数的组合函数,本实施例不做具体限定。例如,第一时序向量为k*m的矩阵向量,第一修正向量为e*m的矩阵向量。
在一实施例中,目标检测指标为k个,第一时间信息包括历史时序数据的m个时间点信息,k、m为大于等于1的正整数;根据历史时序数据和第一时间信息生成第一时序向量,包括:按照m个时间点信息,对k个目标检测指标的历史时序数据进行排列,以生成k*m的矩阵向量,并将矩阵向量作为第一时序向量。需要说明的是,历史时序数据可以包括目标检测指标的多个检测数据,每个时间点信息对应一个检测数据,m个时间点信息表示历史时序数据中存在m个检测数据,每个目标检测指标对应一个时序数据,k个目标检测指标表示历史时序数据共存在k*m个检测数据。按照m个时间点信息确定每个目标检测指标中的多个检测数据的排列顺序,并根据每个目标检测指标中的多个检测数据的排列顺序,对k个目标检测指标的历史时序数据进行排列,得到k*m个检测数据的矩阵向量,该k*m的矩阵向量即为第一时序向量,k代表矩阵向量的宽度或者长度。通过历史时序数据和第一时间信息,能够准确地生成历史时序数据的第一时序向量。
在一实施例中,根据历史时序数据、第一时间信息和预设函数生成第一修正向量,包括:获取预设函数,并将历史时序数据和第一时间信息代入至预设函数,得到历史时序数据的修正信息;根据历史时序数据和历史时序数据的修正信息,生成第一修正向量。其中,预设函数为包括时间特征参数的正弦函数、余弦函数、或者正弦函数与余弦函数的组合函数,例如预设函数为A cos(2πxt),其中A为预设系数,t为时间特征参数,x为设置的未知参数,可用于表征历史时序数据中的检测数据。需要说明的是,将历史时序数据和第一时间信息代入至预设函数,计算出不同时间点的检测数据的修正信息,并将不同时间点的检测数据的修正信息进行排列,得到第一修正向量,该第一修正向量用于修正第一时序向量,使得后续计算出的预测数据更加准确,也使得本申请实施例对异常检测的准确度更高。
示例性的,通过预设函数计算出的修正信息的数量为e,第一时间信息包括历史时序数据的m个时间点信息;按照m个时间点信息,对e个修正信息进行排列,以生成e*m的矩阵向量,并将该矩阵向量作为第一修正向量。其中,e代表矩阵向量的长度,m代表矩阵向量的宽度;或者,m代表矩阵向量的长度,e代表矩阵向量的宽度,本实施例不做具体限定。
子步骤S1023、拼接第一时序向量和第一修正向量,得到历史时序向量。
其中,拼接方法包括行拼接或者列拼接,可根据实际情况进行灵活运用和转置。例如,第一时序向量为k*m的矩阵向量,第一修正向量为e*m的矩阵向量,拼接得到(k+e)*m 的矩阵向量,将该(k+e)*m的矩阵向量作为历史时序向量。或者,例如,第一时序向量为m*k的矩阵向量,第一修正向量为m*e的矩阵向量,拼接得到m*(k+e)的矩阵向量,将该m*(k+e)的矩阵向量作为历史时序向量。
在一实施例中,根据当前时序数据生成当前时序向量,包括:获取与当前时序数据对应的第二时间信息;根据当前时序数据和第二时间信息生成第二时序向量,根据当前时序数据、第二时间信息和预设函数生成第二修正向量;拼接第二时序向量和第二修正向量,得到当前时序向量。其中,当前时序数据可以包括目标检测指标的多个检测数据,第二时间信息包括第二时序、第二时间区间、第二时间间隔和/或每个检测数据的第二时间点。具体地,可参照前述生成历史时序向量的实施例,在此不再赘述。
示例性的,第二时序向量为k*n的矩阵向量,第二修正向量为e*n的矩阵向量,拼接得到(k+e)*n的矩阵向量,将该(k+e)*n的矩阵向量作为历史时序向量。
在一实施例中,从存储器中获取目标检测指标的历史时序数据的历史时序向量。需要说明的是,本申请实施例采用长时序数据的历史时序数据生成历史时序向量,历史生成的历史时序向量能够存储于存储器中,历史时序向量能够在很长一段时间(预设时间段)内复用,再次使用时能够直接通过存储器调用,无需频繁运算,节省了计算量。
步骤S103、通过第一Transformer模型,将历史时序向量和当前时序向量进行关联,得到关联时序向量。
其中,服务器存储有时序数据预测模型,该时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型。
在一实施例中,第一Transformer模型包括Transformer运算层和前馈神经网络层FNN;通过Transformer运算层,对历史时序向量和当前时序向量进行运算,得到关联时序数据;通过前馈神经网络层,对关联时序数据进行映射,得到关联时序向量。其中,Transformer运算层用于对历史时序向量和当前时序向量进行Transformer运算,前馈神经网络层FNN用于对Transformer运算层输出的运算结果进行全连接层映射,输出关联时序数据,以实现历史时序向量和当前时序向量中的各项目标检测指标的关联,两者关联后将覆盖当前时序数据与历史时序数据,以进行更准确地下一时序的预测,提高异常检测的准确度。
示例性的,通过Transformer运算层和前馈神经网络层FNN计算出历史时序向量history_window和当前时序向量current_window的关联时序序列(x3)。具体地,如图3所示,将历史时序向量history_window和当前时序向量current_window输入至第一Transformer模型,第一Transformer模型包括Transformer运算层和前馈神经网络层FNN。Transformer运算层对当前时序向量current_window进行Transformer运算得到x1,x1=Tranformer×S(history_window),×S表示经过了S层Transformer运算,并对当前时序向量S(current_window)进行Transformer运算得到x2,x2=Tranformer×S(current_window),;前馈神经网络层FNN对x1和x2进行全连接层映射,输出关联时序数据x3,x3 =Dense a out(attention out*value out),其中,前馈神经网络层FNN包括查询向量query out、键向量key out和值向量value out,attention out=softmax(query out*key out),value out=Dense v out(x1),key out=Dense k out(x1),query out=Dense q out(x2),Dense表示全连接层,softmax是激活函数。
在一实施例中,第一Transformer时序模型可以为n个,n为大于等于2的正整数。若第一Transformer时序模型可以为n个,则输入的历史时序向量和当前时序向量将按照设置的层级,依次经由每个第一Transformer时序模型进行计算,直至第n个第一Transformer时序模型完成计算并输出,得到关联时序数据。
步骤S104、将关联时序向量输入至第二Transformer模型,得到当前时序数据的下一时序的预测数据。
其中,第二Transformer模型与前述的第一Transformer模型的结构一致,两者存在共性,也包括Transformer运算层和前馈神经网络层FNN,其区别在于模型参数不同。通过第二Transformer模型对关联时序向量进行处理,输出当前时序数据的下一时序的预测数据。通过第二Transformer模型和关联时序向量,将当前时序数据与历史时序数据中的相似时段关联,从而得到当前时序数据的下一时序的预测数据。
示例性的,将关联时序数据x3输入至第二Transformer时序模型,即将关联时序数据x3再经过一次Transformer运算,并通过前馈神经网络层FNN进行全连接层映射Dense,最终完成输出,得到当前时序数据的下一时序的预测数据
Figure PCTCN2021097410-appb-000001
步骤S105、获取当前时序数据的下一时序的目标时序数据,通过预测数据和目标时序数据,对目标时序数据进行异常检测。
在得到当前时序数据的下一时序的预测数据之后,获取当前时序数据的下一时序的目标时序数据,其中,当前时序数据与第二时序相对应,第二时序为当前时序数据对应的时间长度,例如第二时序为1小时,即该当前时序数据包括1小时内的目标检测指标的多个检测数据。目标时序数据对应的时间长度与当前时序数据对应的时间长度相同,即目标时序数据与第二时序相对应,例如包括1小时内的目标检测指标的多个检测数据。目标时序数据是当前时序数据的下一时序的时序数据,因此,目标时序数据对应的时间区间是根据当前时序数据对应的第二时间区间加第二时序得到的,例如当二时序为1小时,当前时序数据对应的第二时间区间是(12时,13时),则目标时序数据对应的时间区间为(13时,14时)。
在一实施例中,将预测数据和目标时序数据进行对比,得到预测数据和目标时序数据之间的相对误差;根据该相对误差对目标时序数据进行异常检测,得到异常检测结果。其中,异常检测结果包括数据异常和数据正常。当预测数据和目标时序数据之间的相对误差大于或等于预设误差阈值,则异常检测结果为数据异常,当预测数据和目标时序数据之间的相对误差小于预设误差阈值,则异常检测结果为数据正常。本申请采用了当前时序数据 和历史时序数据的关联时序向量对当前时序数据的下一时序的目标时序数据进行预测,预测数据能够适应目标检测指标自身随时间的变化趋势,提高了异常检测的准确率。
在一实施例中,当确定目标时序数据为异常数据时,可以进行告警。例如服务器可以播放异常数据警报音,或者,点亮异常数据警示灯,或者将相应的告警信息发送至相关负责人员等。该告警信息的通知方式包括但不限于微信、短信、邮件、企业级通信平台等。
上述实施例提供的基于人工智能的数据检测方法,通过获取目标检测指标的当前时序数据,根据当前时序数据生成当前时序向量,并确定目标检测指标的历史时序数据的历史时序向量,再通过第一Transformer模型,将历史时序向量和当前时序向量进行关联,得到关联时序向量,然后将关联时序向量输入至第二Transformer模型,得到当前时序数据的下一时序的预测数据,最后获取当前时序数据的下一时序的目标时序数据,并通过预测数据和目标时序数据,对目标时序数据进行异常检测。通过Transformer模型对当前时序向量和历史时序向量进行处理,输出准确度更高的预测数据,极大提升了异常数据的检测准确度。
在一实施例中,该数据检测方法还包括:获取多个样本数据,样本数据包括第一时序数据、第二时序数据和标注的时序数据,第一时序数据的采集时间晚于第二时序数据;根据第一时序数据生成第一时序向量,并根据第二时序数据生成第二时序向量;通过第一预设Transformer模型,对第一时序向量和第二时序向量进行关联,得到目标时序向量;将目标时序向量输入至第二预设Transformer模型,得到预测的时序数据;根据预测的时序数据和标注的时序数据,更新第一预设Transformer模型和第二预设Transformer模型的模型参数;继续对更新后的第一预设Transformer模型和第二预设Transformer模型进行迭代训练,直至第一预设Transformer模型和第二预设Transformer模型收敛,得到时序数据预测模型。通过第一时序数据和第二时序数据的关联时序向量训练得到时序数据预测模型,可以有效的提高模型训练的效果,训练好的时序数据预测模型可用于对异常数据的检测。
其中,第一时序数据的采集时间晚于第二时序数据,第一时序数据例如为当前时序数据,第二时序数据例如为历史时序数据。当前时序数据包括当前时序内的目标检测指标的时间序列数据,例如,当前时序数据为1小时内记录的多个内存使用率的时间序列数据。历史时序数据包括历史时序内的目标检测指标的时间序列数据,例如,历史时序数据为过去1周内记录的目标检测指标的时间序列数据。
在一实施例中,第二时序数据为长时序数据,第一时序数据为短时序数据。其中,第二时序数据对应的时间长度大于或等于第一预设时间长度,第一时序数据对应的时间长度小于或等于第二预设时间长度,第二时序数据对应的时间长度大于第一时序数据对应的时间长度,第一预设时间长度和第二预设时间长度可根据实际情况进行设置。需要说明的是,本实施例采用长时序数据的第二时序数据和短时序数据的第一时序数据来以预测下一时序,异常检测的精度较高,且长时序数据的第二时序数据能够以较低频率进行更新,从而 节省计算量。
其中,将第一时序数据转化为第一时序向量,以及将第二时序数据转化为第二时序向量。第二时序向量可以由第二时序数据生成的,并存储于存储器的历史数据中,在后续使用时可以从历史数据中直接获取,因此第二时序向量可以复用而不必重复生成,从而节省计算量和提高模型训练效率。
在一实施例中,第二时序向量对应的时间长度大于或等于第一预设时间长度,第一时序向量对应的时间长度小于或等于第二预设时间长度。需要说明的是,采用长时序数据的第二时序数据生成的第二时序向量对应的时间长度大于或等于第一预设时间长度,而采用短时序数据的第一时序数据生成的第一时序向量对应的时间长度小于或等于第二预设时间长度,能够提高训练好的模型辅助进行数据异常检测的准确度,且长时序数据的第二时序数据能够以较低频率进行更新,在较长一段时间内无需改变,从而节省计算量。
在一实施例中,根据第二时序数据生成第二时序向量,包括:获取目标检测指标的第二时序数据,并确定第二时序数据对应的第一时间信息;根据第二时序数据和第一时间信息生成第一时序向量,并根据第二时序数据、第一时间信息和预设函数生成第一修正向量;拼接第一时序向量和第一修正向量,得到第二时序向量。具体地,可参阅前述数据检测方法的实施例中的对应过程,本实施例在此不再赘述。
在一实施例中,根据第一时序数据生成第一时序向量,包括:获取与第一时序数据对应的第二时间信息;根据第一时序数据和第二时间信息生成第二时序向量,根据第一时序数据、第二时间信息和预设函数生成第二修正向量;拼接第二时序向量和第二修正向量,得到第一时序向量。其中,第一时序数据可以包括目标检测指标的多个检测数据,第二时间信息包括第二时序、第二时间区间、第二时间间隔和/或每个检测数据的第二时间点。具体地,可参照前述数据检测方法中的生成历史时序向量的实施例,在此不再赘述。
其中,第一预设Transformer时序模型可以为一个或多个,第一预设Transformer模型包括Transformer运算层和前馈神经网络层FNN;通过Transformer运算层,对第二时序向量和第一时序向量进行运算,得到关联时序数据;通过前馈神经网络层,对关联时序数据进行映射,得到关联时序向量。其中,Transformer运算层用于对第二时序向量和第一时序向量进行Transformer运算,前馈神经网络层FNN用于对Transformer运算层输出的运算结果进行全连接层映射,输出关联时序数据,以实现第二时序向量和第一时序向量中的各项目标检测指标的关联,两者关联后将覆盖第一时序数据与第二时序数据,以进行更准确地下一时序的预测,提高异常检测的准确度。
其中,第二预设Transformer模型与前述的第一预设Transformer模型的结构一致,两者存在共性,也包括Transformer运算层和前馈神经网络层FNN,其区别在于模型参数不同。通过第二预设Transformer模型对关联时序向量进行处理,输出下一时序的预测的时序数据。通过第二Transformer模型和关联时序向量,将第一时序数据与第二时序数据中 的相似时段关联,从而得到第一时序数据的下一时序的预测的时序数据。
其中,根据预测的时序数据和标注的时序数据,更新第一预设Transformer模型和第二预设Transformer模型的模型参数,继续对更新后的第一预设Transformer模型和第二预设Transformer模型进行迭代训练,直至第一预设Transformer模型和第二预设Transformer模型收敛,得到时序数据预测模型。通过第一时序数据和第二时序数据的关联时序向量训练得到时序数据预测模型,模型训练效果更好,对异常数据的检测精度更高。
在一实施例中,通过预测的时序数据和标注的时序数据,计算第一Transformer预设模型的第一损失函数和第二预设Transformer模型的第二模型参数;基于第一损失函数调整第一预设Transformer模型的模型参数,并基于第二损失函数调整第二预设Transformer模型的模型参数;对经过调整模型参数的第一预设Transformer模型和第二预设Transformer模型进行迭代训练;当确定训练的第一预设Transformer模型和第二预设Transformer模型处于收敛状态,得到时序数据预测模型。
在一实施例中,将第一预设Transformer模型和第二预设Transformer模型进行融合,得到目标Transformer模型;通过预测的时序数据和标注的时序数据,计算目标Transformer模型的损失函数,并基于损失函数确定目标Transformer模型的模型参数,根据模型参数对目标Transformer模型进行迭代训练,并确定训练的目标Transformer模型是否处于收敛状态;若确定训练的目标Transformer模型处于收敛状态,得到时序数据预测模型。其中,损失函数例如为MSE(mean-square error,均方误差)损失函数,并使用Adam优化算法对目标Transformer模型进行迭代训练。需要说明的是,训练完成的异常数据检测模型,便能够基于输入的第二时序数据与第一时序数据,进行下一窗口时序数据(第一时序数据的下一时序的目标时序数据)的预测。可以理解的是,也可以采用适应性梯度算法(AdaGrad)、均方根传播(RMSProp)对目标Transformer模型进行迭代训练,本申请不做具体限定。
请参照图4,图4为本申请实施例提供的一种基于人工智能的数据检测装置的示意性框图。
如图4所示,该基于人工智能的数据检测装置300包括:获取模块301、生成模块302、关联模块303、预测模块304和检测模块305。该数据检测装置300可应用于服务器,所述服务器存储有时序数据预测模型,所述时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型,其中:
获取模块301,用于获取目标检测指标的当前时序数据;
生成模块302,用于根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
关联模块303,用于通过第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
预测模块304,用于将所述关联时序向量输入至第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
所述获取模块301,还用于获取所述当前时序数据的下一时序的目标时序数据;
检测模块305,用于通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
在一个实施例中,所述历史时序向量对应的时间长度大于或等于第一预设时间长度,所述当前时序向量对应的时间长度小于或等于第二预设时间长度。
在一个实施例中,如图5所示,所述生成模块302包括:
获取子模块3021,用于获取所述目标检测指标的历史时序数据,并确定所述历史时序数据对应的第一时间信息;
生成子模块3022,用于根据所述历史时序数据和第一时间信息生成第一时序向量,并根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量;
拼接子模块3023,用于拼接所述第一时序向量和第一修正向量,得到历史时序向量。
在一个实施例中,所述目标检测指标为k个,第一时间信息包括历史时序数据的m个时间点信息,k、m为大于等于1的正整数;生成模块302还用于:
按照所述m个时间点信息,对k个所述目标检测指标的历史时序数据进行排列,以生成k*m的矩阵向量,并将所述矩阵向量作为第一时序向量。
在一个实施例中,生成模块302还用于:
获取预设函数,并将所述历史时序数据和第一时间信息代入至所述预设函数,得到所述历史时序数据的修正信息;
根据所述历史时序数据和所述历史时序数据的修正信息,生成第一修正向量。
在一个实施例中,所述第一Transformer模型包括Transformer运算层和前馈神经网络层;关联模块303还用于:
通过所述Transformer运算层,对所述历史时序向量和当前时序向量进行运算,得到关联时序数据;
通过所述前馈神经网络层,对所述关联时序数据进行映射,得到关联时序向量。
在一个实施例中,该数据检测装置300还用于:
获取多个样本数据,所述样本数据包括第一时序数据、第二时序数据和标注的时序数据,所述第一时序数据的采集时间晚于所述第二时序数据;
根据所述第一时序数据生成第一时序向量,并根据所述第二时序数据生成第二时序向量;
通过第一预设Transformer模型,对所述第一时序向量和所述第二时序向量进行关联,得到目标时序向量;
将所述目标时序向量输入至第二预设Transformer模型,得到预测的时序数据;
根据所述预测的时序数据和所述标注的时序数据,更新所述第一预设Transformer模型和第二预设Transformer模型的模型参数;
继续对更新后的所述第一预设Transformer模型和第二预设Transformer模型进行迭代训练,直至所述第一预设Transformer模型和第二预设Transformer模型收敛,得到所述时序数据预测模型。
需要说明的是,所属领域的技术人员可以清楚地了解到,为了描述的方便和简洁,上述描述的装置和各模块及单元的具体工作过程,可以参考前述基于人工智能的数据检测方法实施例中的对应过程,在此不再赘述。
上述实施例提供的装置可以实现为一种计算机程序的形式,该计算机程序可以在如图6所示的服务器上运行。
请参阅图6,图6为本申请实施例提供的一种服务器的结构示意性框图。该服务器存储有时序数据预测模型,所述时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型。
如图6所示,该服务器包括通过系统总线连接的处理器、存储器和网络接口,其中,存储器可以包括存储介质和内存储器,该存储介质可以是非易失性的也可以是易失性。
存储介质可存储操作系统和计算机程序。该计算机程序包括程序指令,该程序指令被执行时,可使得处理器执行任意一种基于人工智能的数据检测方法。
处理器用于提供计算和控制能力,支撑整个服务器的运行。
内存储器为存储介质中的计算机程序的运行提供环境,该计算机程序被处理器执行时,可使得处理器执行任意一种基于人工智能的数据检测方法。
该网络接口用于进行网络通信,如发送分配的任务等。本领域技术人员可以理解,图6中示出的结构,仅仅是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的服务器的限定,具体的服务器可以包括比图中所示更多或更少的部件,或者组合某些部件,或者具有不同的部件布置。
应当理解的是,处理器可以是中央处理单元(Central Processing Unit,CPU),该处理器还可以是其他通用处理器、数字信号处理器(Digital Signal Processor,DSP)、专用集成电路(Application Specific Integrated Circuit,ASIC)、现场可编程门阵列(Field-Programmable Gate Array,FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。其中,通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
其中,在一个实施例中,所述处理器用于运行存储在存储器中的计算机程序,以实现如下步骤:
获取目标检测指标的当前时序数据;
根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据 的历史时序向量;
通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
将所述关联时序向量输入至所述第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
在一个实施例中,所述历史时序向量对应的时间长度大于或等于第一预设时间长度,所述当前时序向量对应的时间长度小于或等于第二预设时间长度。
在一个实施例中,所述处理器在实现所述确定所述目标检测指标的历史时序数据的历史时序向量时,用于实现:
获取所述目标检测指标的历史时序数据,并确定所述历史时序数据对应的第一时间信息;
根据所述历史时序数据和第一时间信息生成第一时序向量,并根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量;
拼接所述第一时序向量和第一修正向量,得到历史时序向量。
在一个实施例中,所述目标检测指标为k个,第一时间信息包括历史时序数据的m个时间点信息,k、m为大于等于1的正整数;
所述处理器在实现所述根据所述历史时序数据和第一时间信息生成第一时序向量时,用于实现:
按照所述m个时间点信息,对k个所述目标检测指标的历史时序数据进行排列,以生成k*m的矩阵向量,并将所述矩阵向量作为第一时序向量。
在一个实施例中,所述处理器在实现所述根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量时,用于实现:
获取预设函数,并将所述历史时序数据和第一时间信息代入至所述预设函数,得到所述历史时序数据的修正信息;
根据所述历史时序数据和所述历史时序数据的修正信息,生成第一修正向量。
在一个实施例中,所述第一Transformer模型包括Transformer运算层和前馈神经网络层;所述处理器在实现所述通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量时,用于实现:
通过所述Transformer运算层,对所述历史时序向量和当前时序向量进行运算,得到关联时序数据;
通过所述前馈神经网络层,对所述关联时序数据进行映射,得到关联时序向量。
在一个实施例中,所述处理器还用于实现:
获取多个样本数据,所述样本数据包括第一时序数据、第二时序数据和标注的时序数据,所述第一时序数据的采集时间晚于所述第二时序数据;
根据所述第一时序数据生成第一时序向量,并根据所述第二时序数据生成第二时序向量;
通过第一预设Transformer模型,对所述第一时序向量和所述第二时序向量进行关联,得到目标时序向量;
将所述目标时序向量输入至第二预设Transformer模型,得到预测的时序数据;
根据所述预测的时序数据和所述标注的时序数据,更新所述第一预设Transformer模型和第二预设Transformer模型的模型参数;
继续对更新后的所述第一预设Transformer模型和第二预设Transformer模型进行迭代训练,直至所述第一预设Transformer模型和第二预设Transformer模型收敛,得到所述时序数据预测模型。
需要说明的是,所属领域的技术人员可以清楚地了解到,为了描述的方便和简洁,上述描述服务器的具体工作过程,可以参考前述基于人工智能的数据检测方法实施例中的对应过程,在此不再赘述。
本申请实施例还提供一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机程序,其中所述计算机程序被处理器执行时,实现以下步骤:
获取目标检测指标的当前时序数据;
根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
通过预先训练好的第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
将所述关联时序向量输入至预先训练好的第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
需要说明的是,所属领域的技术人员可以清楚地了解到,为了描述的方便和简洁,上述描述的计算机可读存储介质的具体工作过程,可以参考前述数据检测方法实施例中的对应过程,在此不再赘述。
其中,所述计算机可读存储介质可以是前述实施例所述的服务器的内部存储单元,例如所述服务器的硬盘或内存。所述计算机可读存储介质也可以是所述服务器的外部存储设备,例如所述服务器上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等。
应当理解,在此本申请说明书中所使用的术语仅仅是出于描述特定实施例的目的而并 不意在限制本申请。如在本申请说明书和所附权利要求书中所使用的那样,除非上下文清楚地指明其它情况,否则单数形式的“一”、“一个”及“该”意在包括复数形式。
还应当理解,在本申请说明书和所附权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合,并且包括这些组合。需要说明的是,在本文中,术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、方法、物品或者系统不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、方法、物品或者系统所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除在包括该要素的过程、方法、物品或者系统中还存在另外的相同要素。
上述本申请实施例序号仅仅为了描述,不代表实施例的优劣。以上所述,仅为本申请的具体实施方式,但本申请的保护范围并不局限于此,任何熟悉本技术领域的技术人员在本申请揭露的技术范围内,可轻易想到各种等效的修改或替换,这些修改或替换都应涵盖在本申请的保护范围之内。因此,本申请的保护范围应以权利要求的保护范围为准。

Claims (20)

  1. 一种基于人工智能的数据检测方法,其中,应用于服务器,所述服务器存储有时序数据预测模型,所述时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型,所述方法包括:
    获取目标检测指标的当前时序数据;
    根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
    通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
    将所述关联时序向量输入至所述第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
    获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
  2. 如权利要求1所述的数据检测方法,其中,所述历史时序向量对应的时间长度大于或等于第一预设时间长度,所述当前时序向量对应的时间长度小于或等于第二预设时间长度。
  3. 如权利要求1所述的数据检测方法,其中,所述确定所述目标检测指标的历史时序数据的历史时序向量,包括:
    获取所述目标检测指标的历史时序数据,并确定所述历史时序数据对应的第一时间信息;
    根据所述历史时序数据和第一时间信息生成第一时序向量,并根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量;
    拼接所述第一时序向量和第一修正向量,得到历史时序向量。
  4. 如权利要求3所述的数据检测方法,其中,所述目标检测指标为k个,第一时间信息包括历史时序数据的m个时间点信息,k、m为大于等于1的正整数;所述根据所述历史时序数据和第一时间信息生成第一时序向量,包括:
    按照所述m个时间点信息,对k个所述目标检测指标的历史时序数据进行排列,以生成k*m的矩阵向量,并将所述矩阵向量作为第一时序向量。
  5. 如权利要求3所述的数据检测方法,其中,所述根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量,包括:
    获取预设函数,并将所述历史时序数据和第一时间信息代入至所述预设函数,得到所述历史时序数据的修正信息;
    根据所述历史时序数据和所述历史时序数据的修正信息,生成第一修正向量。
  6. 如权利要求1-5中任一项所述的数据检测方法,其中,所述第一Transformer模型包括Transformer运算层和前馈神经网络层;所述通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量,包括:
    通过所述Transformer运算层,对所述历史时序向量和当前时序向量进行运算,得到关联时序数据;
    通过所述前馈神经网络层,对所述关联时序数据进行映射,得到关联时序向量。
  7. 如权利要求1-5中任一项所述的数据检测方法,其中,所述方法还包括:
    获取多个样本数据,所述样本数据包括第一时序数据、第二时序数据和标注的时序数据,所述第一时序数据的采集时间晚于所述第二时序数据;
    根据所述第一时序数据生成第一时序向量,并根据所述第二时序数据生成第二时序向量;
    通过第一预设Transformer模型,对所述第一时序向量和所述第二时序向量进行关联,得到目标时序向量;
    将所述目标时序向量输入至第二预设Transformer模型,得到预测的时序数据;
    根据所述预测的时序数据和所述标注的时序数据,更新所述第一预设Transformer模型和第二预设Transformer模型的模型参数;
    继续对更新后的所述第一预设Transformer模型和第二预设Transformer模型进行迭代训练,直至所述第一预设Transformer模型和第二预设Transformer模型收敛,得到所述时序数据预测模型。
  8. 一种基于人工智能的数据检测装置,其中,应用于服务器,所述服务器存储有时序数据预测模型,所述时序数据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型,所述数据检测装置包括:
    获取模块,用于获取目标检测指标的当前时序数据;
    生成模块,用于根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
    关联模块,用于通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
    预测模块,用于将所述关联时序向量输入至所述第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
    所述获取模块,还用于获取所述当前时序数据的下一时序的目标时序数据;
    检测模块,用于通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
  9. 一种服务器,其中,所述服务器包括处理器、存储器、以及存储在所述存储器上并可被所述处理器执行的计算机程序,所述服务器存储有时序数据预测模型,所述时序数 据预测模型包括预先训练好的第一Transformer模型和第二Transformer模型,其中所述计算机程序被所述处理器执行时,实现以下步骤:
    获取目标检测指标的当前时序数据;
    根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
    通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
    将所述关联时序向量输入至所述第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
    获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
  10. 如权利要求9所述的服务器,其中,所述历史时序向量对应的时间长度大于或等于第一预设时间长度,所述当前时序向量对应的时间长度小于或等于第二预设时间长度。
  11. 如权利要求9所述的服务器,其中,所述处理器在实现所述确定所述目标检测指标的历史时序数据的历史时序向量时,用于实现:
    获取所述目标检测指标的历史时序数据,并确定所述历史时序数据对应的第一时间信息;
    根据所述历史时序数据和第一时间信息生成第一时序向量,并根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量;
    拼接所述第一时序向量和第一修正向量,得到历史时序向量。
  12. 如权利要求11所述的服务器,其中,所述目标检测指标为k个,第一时间信息包括历史时序数据的m个时间点信息,k、m为大于等于1的正整数;所述处理器在实现所述根据所述历史时序数据和第一时间信息生成第一时序向量时,用于实现:
    按照所述m个时间点信息,对k个所述目标检测指标的历史时序数据进行排列,以生成k*m的矩阵向量,并将所述矩阵向量作为第一时序向量。
  13. 如权利要求11所述的服务器,其中,所述处理器在实现所述根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量时,用于实现:
    获取预设函数,并将所述历史时序数据和第一时间信息代入至所述预设函数,得到所述历史时序数据的修正信息;
    根据所述历史时序数据和所述历史时序数据的修正信息,生成第一修正向量。
  14. 如权利要求9-13中任一项所述的服务器,其中,所述第一Transformer模型包括Transformer运算层和前馈神经网络层;所述处理器在实现所述通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量时, 用于实现:
    通过所述Transformer运算层,对所述历史时序向量和当前时序向量进行运算,得到关联时序数据;
    通过所述前馈神经网络层,对所述关联时序数据进行映射,得到关联时序向量。
  15. 一种计算机可读存储介质,其中,所述计算机可读存储介质上存储有计算机程序,其中所述计算机程序被处理器执行时,实现以下步骤:
    获取目标检测指标的当前时序数据;
    根据所述当前时序数据生成当前时序向量,并确定所述目标检测指标的历史时序数据的历史时序向量;
    通过预先训练好的第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量;
    将所述关联时序向量输入至预先训练好的第二Transformer模型,得到所述当前时序数据的下一时序的预测数据;
    获取所述当前时序数据的下一时序的目标时序数据,通过所述预测数据和目标时序数据,对所述目标时序数据进行异常检测。
  16. 如权利要求15所述的计算机可读存储介质,其中,所述历史时序向量对应的时间长度大于或等于第一预设时间长度,所述当前时序向量对应的时间长度小于或等于第二预设时间长度。
  17. 如权利要求15所述的计算机可读存储介质,其中,所述处理器在实现所述确定所述目标检测指标的历史时序数据的历史时序向量时,用于实现:
    获取所述目标检测指标的历史时序数据,并确定所述历史时序数据对应的第一时间信息;
    根据所述历史时序数据和第一时间信息生成第一时序向量,并根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量;
    拼接所述第一时序向量和第一修正向量,得到历史时序向量。
  18. 如权利要求17所述的计算机可读存储介质,其中,所述目标检测指标为k个,第一时间信息包括历史时序数据的m个时间点信息,k、m为大于等于1的正整数;所述处理器在实现所述根据所述历史时序数据和第一时间信息生成第一时序向量时,用于实现:
    按照所述m个时间点信息,对k个所述目标检测指标的历史时序数据进行排列,以生成k*m的矩阵向量,并将所述矩阵向量作为第一时序向量。
  19. 如权利要求17所述的计算机可读存储介质,其中,所述处理器在实现所述根据所述历史时序数据、第一时间信息和预设函数生成第一修正向量时,用于实现:
    获取预设函数,并将所述历史时序数据和第一时间信息代入至所述预设函数,得到所述历史时序数据的修正信息;
    根据所述历史时序数据和所述历史时序数据的修正信息,生成第一修正向量。
  20. 如权利要求15-19中任一项所述的计算机可读存储介质,其中,所述第一Transformer模型包括Transformer运算层和前馈神经网络层;所述处理器在实现所述通过所述第一Transformer模型,将所述历史时序向量和当前时序向量进行关联,得到关联时序向量时,用于实现:
    通过所述Transformer运算层,对所述历史时序向量和当前时序向量进行运算,得到关联时序数据;
    通过所述前馈神经网络层,对所述关联时序数据进行映射,得到关联时序向量。
PCT/CN2021/097410 2020-12-31 2021-05-31 基于人工智能的数据检测方法、装置、服务器及存储介质 WO2022142120A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011639634.3 2020-12-31
CN202011639634.3A CN112732983B (zh) 2020-12-31 2020-12-31 基于人工智能的数据检测方法、装置、服务器及存储介质

Publications (1)

Publication Number Publication Date
WO2022142120A1 true WO2022142120A1 (zh) 2022-07-07

Family

ID=75608975

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/097410 WO2022142120A1 (zh) 2020-12-31 2021-05-31 基于人工智能的数据检测方法、装置、服务器及存储介质

Country Status (2)

Country Link
CN (1) CN112732983B (zh)
WO (1) WO2022142120A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115475900A (zh) * 2022-09-19 2022-12-16 江兴(淮安)汽车部件有限公司 一种传动轴自动化精密锻造的方法和系统
CN115514679A (zh) * 2022-11-11 2022-12-23 浙江万胜智能科技股份有限公司 一种基于通信模块的异常来源监测方法及系统
CN115545362A (zh) * 2022-12-05 2022-12-30 南方电网数字电网研究院有限公司 一种ai与tsd结合的新能源中期功率组合预测方法
CN115840875A (zh) * 2022-11-10 2023-03-24 北京擎天信安科技有限公司 一种基于Anomaly Transformer的毫米波雷达异常信号检测方法及系统
CN116684878A (zh) * 2023-07-10 2023-09-01 北京中科网芯科技有限公司 一种5g信息传输数据安全监测系统
CN116820896A (zh) * 2023-08-30 2023-09-29 南京南瑞信息通信科技有限公司 一种基于物理信号的非侵入式工控终端异常检测方法
CN117216701A (zh) * 2023-09-13 2023-12-12 广州桐富科技发展有限公司 智慧桥梁监测预警方法及系统

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112732983B (zh) * 2020-12-31 2023-09-12 平安科技(深圳)有限公司 基于人工智能的数据检测方法、装置、服务器及存储介质
CN113591958B (zh) * 2021-07-21 2024-03-19 杭州海康威视数字技术股份有限公司 融合物联网数据和信息网数据的方法、装置及设备
CN114510968B (zh) * 2022-01-21 2022-09-16 石家庄铁道大学 一种基于Transformer的故障诊断方法
CN116149896B (zh) * 2023-03-27 2023-07-21 阿里巴巴(中国)有限公司 时序数据异常检测方法、存储介质和电子设备
CN117473514B (zh) * 2023-12-28 2024-03-15 华东交通大学 一种工控系统的智能运维方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180006900A1 (en) * 2016-06-29 2018-01-04 Microsoft Technology Licensing, Llc Predictive anomaly detection in communication systems
CN109815980A (zh) * 2018-12-18 2019-05-28 北京三快在线科技有限公司 用户类型的预测方法、装置、电子设备及可读存储介质
CN110059858A (zh) * 2019-03-15 2019-07-26 深圳壹账通智能科技有限公司 服务器资源预测方法、装置、计算机设备及存储介质
CN110377447A (zh) * 2019-07-17 2019-10-25 腾讯科技(深圳)有限公司 一种异常数据检测方法、装置及服务器
CN111130940A (zh) * 2019-12-26 2020-05-08 众安信息技术服务有限公司 异常数据检测方法、装置及服务器
CN112732983A (zh) * 2020-12-31 2021-04-30 平安科技(深圳)有限公司 基于人工智能的数据检测方法、装置、服务器及存储介质

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104020754B (zh) * 2014-05-16 2017-01-25 国家电网公司 一种变电站一次主设备状态监测信息接入调控系统的方法
CN109542740B (zh) * 2017-09-22 2022-05-27 阿里巴巴集团控股有限公司 异常检测方法及装置
WO2019064599A1 (ja) * 2017-09-29 2019-04-04 日本電気株式会社 異常検知装置、異常検知方法、及びコンピュータ読み取り可能な記録媒体
CN111242793B (zh) * 2020-01-16 2024-02-06 上海金仕达卫宁软件科技有限公司 医保数据异常的检测方法和装置
CN111178645B (zh) * 2020-04-14 2020-12-08 深圳市朝阳辉电气设备有限公司 电力设备异常检测方法、装置、控制设备及存储介质
CN111784062A (zh) * 2020-07-07 2020-10-16 上海携旅信息技术有限公司 点击概率预测方法、系统、电子设备和介质
CN112037001A (zh) * 2020-09-03 2020-12-04 云账户技术(天津)有限公司 打款风险预测模型训练方法、打款风险预测方法及其装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180006900A1 (en) * 2016-06-29 2018-01-04 Microsoft Technology Licensing, Llc Predictive anomaly detection in communication systems
CN109815980A (zh) * 2018-12-18 2019-05-28 北京三快在线科技有限公司 用户类型的预测方法、装置、电子设备及可读存储介质
CN110059858A (zh) * 2019-03-15 2019-07-26 深圳壹账通智能科技有限公司 服务器资源预测方法、装置、计算机设备及存储介质
CN110377447A (zh) * 2019-07-17 2019-10-25 腾讯科技(深圳)有限公司 一种异常数据检测方法、装置及服务器
CN111130940A (zh) * 2019-12-26 2020-05-08 众安信息技术服务有限公司 异常数据检测方法、装置及服务器
CN112732983A (zh) * 2020-12-31 2021-04-30 平安科技(深圳)有限公司 基于人工智能的数据检测方法、装置、服务器及存储介质

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115475900A (zh) * 2022-09-19 2022-12-16 江兴(淮安)汽车部件有限公司 一种传动轴自动化精密锻造的方法和系统
CN115475900B (zh) * 2022-09-19 2023-10-27 江兴(淮安)汽车部件有限公司 一种传动轴自动化精密锻造的方法和系统
CN115840875A (zh) * 2022-11-10 2023-03-24 北京擎天信安科技有限公司 一种基于Anomaly Transformer的毫米波雷达异常信号检测方法及系统
CN115514679A (zh) * 2022-11-11 2022-12-23 浙江万胜智能科技股份有限公司 一种基于通信模块的异常来源监测方法及系统
CN115545362A (zh) * 2022-12-05 2022-12-30 南方电网数字电网研究院有限公司 一种ai与tsd结合的新能源中期功率组合预测方法
CN116684878A (zh) * 2023-07-10 2023-09-01 北京中科网芯科技有限公司 一种5g信息传输数据安全监测系统
CN116684878B (zh) * 2023-07-10 2024-01-30 北京中科网芯科技有限公司 一种5g信息传输数据安全监测系统
CN116820896A (zh) * 2023-08-30 2023-09-29 南京南瑞信息通信科技有限公司 一种基于物理信号的非侵入式工控终端异常检测方法
CN116820896B (zh) * 2023-08-30 2023-11-24 南京南瑞信息通信科技有限公司 一种基于物理信号的非侵入式工控终端异常检测方法
CN117216701A (zh) * 2023-09-13 2023-12-12 广州桐富科技发展有限公司 智慧桥梁监测预警方法及系统
CN117216701B (zh) * 2023-09-13 2024-03-15 华夏安信物联网技术有限公司 智慧桥梁监测预警方法及系统

Also Published As

Publication number Publication date
CN112732983A (zh) 2021-04-30
CN112732983B (zh) 2023-09-12

Similar Documents

Publication Publication Date Title
WO2022142120A1 (zh) 基于人工智能的数据检测方法、装置、服务器及存储介质
US11715001B2 (en) Water quality prediction
US20190354836A1 (en) Dynamic discovery of dependencies among time series data using neural networks
US10379146B2 (en) Detecting non-technical losses in electrical networks based on multi-layered statistical techniques from smart meter data
US20200394462A1 (en) Predicting software performace based on different system configurations
WO2022227388A1 (zh) 日志异常检测模型训练方法、装置及设备
CN114285728B (zh) 一种预测模型训练方法、流量预测方法、装置及存储介质
WO2021047270A1 (zh) 一种网络流量预测方法及通信设备和存储介质
KR20210055560A (ko) 인공지능 서비스에 대한 성능 관리 장치 및 방법
US20160124713A1 (en) Fast, energy-efficient exponential computations in simd architectures
US20210232472A1 (en) Low-latency systems to trigger remedial actions in data centers based on telemetry data
JP2017037645A (ja) スマートアラートのためのシステム及び方法
JP6086875B2 (ja) 発電量予測装置および発電量予測方法
US11205111B2 (en) End of period metric projection with intra-period alerts
US9412267B2 (en) Auto-calibration for road traffic prediction
US8793106B2 (en) Continuous prediction of expected chip performance throughout the production lifecycle
WO2023051974A1 (en) Calibrated decoders for implementations of quantum codes
CN112131274B (zh) 时间序列异常点的检测方法、装置、设备及可读存储介质
WO2020173270A1 (zh) 用于分析数据的方法、设备和计算机存储介质
WO2022218068A1 (zh) 素材投放方法、装置、设备和介质
CN116668351A (zh) 服务质量预测方法、装置、计算机设备及存储介质
US20230038977A1 (en) Apparatus and method for predicting anomalous events in a system
CN115952916A (zh) 基于人工智能的风电功率预测误差修正方法、装置及设备
US20150212974A1 (en) Fast and automated arima model initialization
US8965825B2 (en) Mode determination for multivariate time series data

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21912881

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21912881

Country of ref document: EP

Kind code of ref document: A1