An automatic labeling method and system for industrial field time series data

Abstract

The application discloses an automatic labeling method and system for industrial field time series data, and belongs to the technical field of industrial automation. Considering that if there is an error category label in each industrial field time series data, the model error of the overall industrial field time series data set will significantly increase, the labeling of a single industrial field time series data depends on the information fed back by other time series data labeling, and the autocorrelation characteristics of the industrial field time series data are simultaneously considered, MVAR models of each industrial field time series data to be labeled are respectively established, then the conditional probability distribution of each industrial field time series data under a given category is obtained, the likelihood function of all industrial field time series data is obtained by combination, and maximum likelihood estimation is performed to obtain the probability of each industrial field time series belonging to different categories. In the case that there are many sensors in the industrial field, the industrial field time series data can be accurately and automatically labeled.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial automation, and more particularly relates to an automatic labeling method and system for industrial field time series data. BACKGROUND

[0002] Industrial automation refers to the information processing and process control of measurement, manipulation, etc. by machine equipment or production process without direct human intervention, and is one of the important technologies in the field of modern manufacturing in the 21st century. In the fields of textiles, packaging, building materials, papermaking, automobiles, electronics, environmental protection, steel, transportation and water treatment, industrial automation technology involves single equipment, workstations, production lines, and even entire factories in the processing process of these industries, and is the basic supporting technology of modern industry. Enterprises can achieve the purposes of improving production efficiency, ensuring product quality, saving labor costs and ensuring safety by introducing automation technology.

[0003] As a typical application of industrial automation, the automatic labeling technology of industrial field time series data relies on high-tech such as artificial intelligence to inject new vitality into the development of industrial automation. With the large number and variety of sensor devices in industrial scenarios, the collection of "dynamic" data time series (such as wearable devices) has rapidly increased, and how to efficiently and accurately automatically label vector time series data collected by multiple sensors has become increasingly important.

[0004] However, most current automatic labeling methods for time series data often only process scalar time series collected by a single sensor, and need to label each scalar time series collected by each sensor one by one, which is time-consuming, and also does not pay attention to the autocorrelation patterns / features of vector time series data collected by multiple sensors, so the accuracy is low and cannot be applied to the automatic labeling of vector time series data collected by multiple sensors. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the present application provides an automatic labeling method and system for industrial field time series data to solve the technical problem that the prior art cannot accurately and automatically label industrial field time series data when there are many sensors in the industrial field.

[0006] To achieve the above purpose, in a first aspect, the present application provides an automatic labeling method for industrial field time series data, comprising:

[0007] respectively establishing MVAR models for each industrial field time series data to be labeled, and then obtaining the conditional probability distribution thereof under a given category;

[0008] The conditional probability distribution of each industrial field time series data under a given category is combined to obtain a likelihood function of all industrial field time series data, and then maximum likelihood estimation is performed to obtain the probability of each industrial field time series belonging to different categories, and then the category of the industrial field time series data is labeled;

[0009] The nth industrial field time series data is a sequence composed of sensor data vectors at all time points in the nth processing process; the sensor data vector at each time point is a vector composed of the collected data of all sensors on the processing pipeline at the time point.

[0010] Further preferably, the MVAR model established for the nth industrial field time series data is:

[0011]

[0012] wherein, represents the probability of Y n,t occurring under the condition that the historical data before t time point is data n,t ; Y k is a sequence composed of sensor data vectors at the first t time points in the nth industrial field time series data; K is the total number of preset categories; α k is the probability of any industrial field time series data belonging to the kth category under the condition that the historical data before t time point is data ; Φ(·) is a multivariate cumulative distribution function of Gaussian distribution with zero mean and unit covariance matrix; Ω k is an m×m size coefficient matrix corresponding to the kth category; p k，0 is the kth order in the preset order set of the MVAR model; all are parameters of the MVAR model, Θ n，k is an m-dimensional coefficient vector, all are m×m size coefficient matrices; m is the total number of sensors on the processing pipeline.

[0013] Further preferably, the conditional probability distribution of the nth industrial field time series data under a given category is:

[0014]

[0015] wherein, Z k = 1 indicates that the nth industrial field time series data belongs to the kth category label; p is the maximum order in the preset order set of the MVAR model;

[0016] Further preferably, the logarithmic expression of the likelihood function of all industrial field time series data is log L:

[0017]

[0018] f k n，k，t k n，t n，k n，t-1 n，t-p

[0019] wherein N is the total number of industrial field time series data; T is the total number of time points of the process.

[0020] Further preferably, the method of maximum likelihood estimation comprises the following steps:

[0021] initializing τ n，k , and Ω k ; wherein τ n，k is the probability of the nth industrial field time series data belonging to the kth category; n = 1,..., N;

[0022] In the E step: traversing k = 1,..., K, if then updating τ n，k to 1, otherwise updating τ n，k to 0; wherein

[0023] In the M step:

[0024] updating α k to

[0025] obtaining a set composed of all n when τ n，k = 1 as the current set I k ;

[0026] updating to

[0027] updating Ω k to wherein |I k | is the number of elements in the set I k .

[0028] Alternating the E step and the M step for iteration until α k , and Ω k all converge, at this time τ n，k ​​​​​​​This is the probability that the nth industrial site time series data belongs to the kth category.

[0029] More preferably, the total number of categories K and the preset order set {p1, ..., p} of the MVAR model are... k The settings for} include:

[0030] K from N K Arbitrarily select from the ordered set of n integers, p1, ..., p2, ... k All from N p By arbitrarily selecting from an ordered set of n integers, we obtain multiple sets of K, p1, ..., p k The value of N; K and N p All are integers greater than or equal to 2;

[0031] Calculate each group K, p1, ..., p using the following formula. k BIC value under the given value;

[0032]

[0033] in,

[0034]

[0035]

[0036] η mix =N

[0037] Let K and p1, ..., p k The set of all possible values ​​for these two types of parameters; For set The total number of elements in

[0038] Find the set of K,p1,…,p values ​​that have the lowest BIC values. k To set the total number of categories K and the preset order set {p1,…,p} of the MVAR model. k}

[0039] In a second aspect, the present invention provides an automatic labeling system for industrial field time series data, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the automatic labeling method for industrial field time series data provided in the first aspect of the present invention.

[0040] In a third aspect, the present application also provides a computer readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls a device in which the storage medium is located to perform the automatic labeling method of industrial field time series data according to the first aspect of the present application.

[0041] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0042] 1. The present application provides an automatic labeling method for industrial field time series data, which takes into account that incorrect class labeling of each industrial field time series data will significantly increase the model error of the overall industrial field time series data set, and the labeling of a single industrial field time series data depends on the information fed back by other time series data labeling, while also taking into account the autocorrelation characteristics of industrial field time series data, an MVAR model is established for each industrial field time series data to be labeled, and then the conditional probability distribution of each industrial field time series data under a given class is obtained, and the likelihood function of all industrial field time series data is obtained by simultaneous solution, and maximum likelihood estimation is performed to obtain the probability of each industrial field time series belonging to different classes, which can accurately and automatically label industrial field time series data in the case of more sensors in the industrial field.

[0043] 2. Further, in the automatic labeling method for industrial field time series data provided by the present application, the EM algorithm is improved by adding small noise when performing maximum likelihood estimation to solve the calculation bottleneck and numerical underflow problem, effectively reducing the calculation amount of the label solving process, reducing the time used for classification, greatly improving the calculation efficiency, and being able to quickly process the automatic labeling problem of a large number of industrial time series, and good results can be achieved when the length of the industrial time series is large or the number of types is large.

[0044] 3. Further, in the automatic labeling method for industrial field time series data provided by the present application, it is considered that in the original BIC standard, a model with a larger K obtains a higher probability than a model with a smaller K, which violates the simplicity criterion in many scenarios, leading to a decrease in the performance of the model under the selected parameters, and thus is not applicable, therefore, the BIC standard is improved, and an extended BIC standard is used, so that the model can better fit the industrial field time series data under the selected parameters, so as to select the total number of classes K and the preset order set {p1, …, p k} of the MVAR model, ensuring the consistency of the model selection problem.

[0045] 4. Further, the automatic labeling method for industrial field time series data provided by the present application can be quickly deployed to the industrial field and is applicable to the labeling of various industrial time series data sets. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The automatic labeling method flow chart of industrial field time series data provided for embodiment 1 of the present application;

[0047] Figure 2 The performance curve of the improved EM algorithm and the MVAR method using the traditional EM algorithm under different lengths of industrial field time series data provided for embodiment 1 of the present application;

[0048] Figure 3 The performance curve of the improved EM algorithm and the MVAR method using the traditional EM algorithm under different numbers of sensors provided for embodiment 1 of the present application. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0050] Embodiment 1

[0051] An automatic labeling method of industrial field time series data, as shown in Figure 1 , comprises:

[0052] S1, respectively, for each industrial field time series data to be labeled, an MVAR model (mixed autoregressive vector model) is established, and then the conditional probability distribution thereof under a given category is obtained; wherein the nth industrial field time series data is a sequence composed of sensor data vectors at all time points in the nth processing process; the sensor data vector at each time point is a vector composed of the collected data of all sensors on the processing pipeline at that time point;

[0053] The present application improves the MVAR model to be applied to industrial time series data clustering. The MVAR model is composed of K Gaussian VAR components of m-dimensional vector Y n,t , which is used to model a single time series showing multimodal in marginal distribution or conditional distribution, and is not used for time series clustering. Specifically, in the present embodiment, the MVAR model established for the nth industrial field time series data is:

[0054]

[0055] wherein, represents the conditional probability distribution of Y under the condition that the historical data before t time is datan,t The probability of occurrence; Y n,t The sequence is formed by the sensor data vectors at the first t moments in the nth industrial field time series data; K is the total number of preset categories; α k The historical data before time t is the data. Given the condition, the probability that any industrial field time series data belongs to the k-th category; to avoid unidentification due to component label interchange, let α1≥α2≥…≥α k ≥0, and Φ(·) is the multivariate cumulative distribution function of a Gaussian distribution with zero mean and unit covariance matrix; Ω k p is the m×m coefficient matrix corresponding to the k-th category; k The k-th order in the preset order set of the MVAR model; These are all parameters of the MVAR model, Θ k,0 It is an m-dimensional coefficient vector. Both are coefficient matrices of size m×m; m is the total number of sensors on the production line. In this model, K and p k The parameter α is a parameter set by the user. k , and Ω k To be requested.

[0056] Furthermore, for the nth industrial site time series data, assume its category label Z n =(Z n,1 Z n,2 ...Z n,k Z is a one-hot encoding, i.e. n,1 Z n,2 ...Z n,k Only one value in the set can be 1, and the rest can be 0. A value of 1 indicates that the industrial site time series data belongs to that category. Specifically, the conditional probability distribution of the nth industrial site time series data under a given category is:

[0057]

[0058] Among them, Z n,k =1 indicates that the nth industrial site time series data belongs to the kth category label; p is the maximum order in the preset order set of the MVAR model;

[0059] S2, the conditional probability distribution of each industrial field time series data under a given category is combined to obtain the likelihood function of all industrial field time series data, and then maximum likelihood estimation is performed to obtain the probability of each industrial field time series belonging to different categories, and then the category of the industrial field time series data is labeled;

[0060] Specifically, the logarithmic expression log L of the likelihood function of all industrial field time series data is:

[0061]

[0062] f k (e n,k,t ,Ω k )=p(Y n,t |Z n,k =1,Y n,t-1 ,…,Y n,t-p )

[0063] Where N is the total number of industrial field time series data; T is the total number of time points of the process.

[0064] If each industrial field time series has an error category label, the model error of the entire industrial field time series data set will be significantly increased (equivalently, the likelihood function value will be reduced), and the labeling of a single time series depends on the information feedback from the labeling of other sequence data; Therefore, for the conditional probability of each industrial field time series data under a given category label, the likelihood function of the entire industrial field time series data set under a given model set is constructed, and the category label of each industrial field time series data is continuously optimized in the iteration process of maximum likelihood estimation, and the probability value of each industrial field time series data belonging to each category is batched to select the best label. In an optional implementation, the category label of the industrial field time series can be labeled as the category corresponding to the maximum probability; in another optional implementation, considering that the improved EM algorithm is used, the probability value only exists between 0 and 1, and the category label of the industrial field time series can also be labeled as the category corresponding to the probability of 1.

[0065] It should be noted that in an optional implementation, the EM algorithm can be used for maximum likelihood estimation; in another optional implementation, the EM algorithm is improved for maximum likelihood estimation, and by adding small noise, the calculation bottleneck and numerical underflow problem are solved, and the time for classification is reduced; Specifically, the improved EM algorithm is used to perform maximum likelihood estimation on the logarithmic expression log L of the likelihood function, including the following steps:

[0066] 1) Initialize τ n,k , and Ωk ; where τ n,k is the probability that the nth industrial site time series belongs to the kth class; n = 1,...,N;

[0067] In particular, in an alternative embodiment, for each n = 1,...,N, one k is randomly selected, where τ n,k = 1 and all other As for the class VAR parameters The key is to choose K pairs of parameters that are essentially different. Another approach is to perform a standard VAR model estimation for each industrial site time series, resulting in N sets of Then, among the N results, K values are selected for initialization.

[0068] 2) In the E-step: iterate over k = 1,...,K, if then update τ n,k to 1, otherwise update τ n,k to 0; where

[0069] In particular, in the E-step, the original EM algorithm is improved by adding a small noise limit, in the following way: when solving the probability τ n,k that the nth industrial site time series belongs to the kth class, a small noise limit is added, in particular: for k = 1,...,K, let When all γ k go to zero at the same speed, we want to get the limit. To do this in a continuous way, express the ratio as where When γ goes to 0 and the ratio becomes constant in the E-step of the EM algorithm, we get the new iteration formula:

[0070]

[0071] This means that when k is equal to the value of k that makes ψ n,k take the minimum value, then update τ n,k to 1, otherwise update τ n,k to 0 (where k = 1,2,...,K).

[0072] 3) In the M-step:

[0073] Update α k to Other parameters can be further simplified because of the calculation of τ n,k ;

[0074] Get τ n,kthe set consisting of all n when α k ;

[0075] update Ω to

[0076] update Ω k to where |I k | is the number of elements in the set I k ;

[0077] 4) iterate the E step and M step alternately until α k , and Ω k all converge, at this time τ n,k is the probability that the nth industrial field time series data belongs to the kth category.

[0078] By improving the EM algorithm, part of the calculation in the traditional EM algorithm is removed, so as to avoid the underflow problem caused by calculating the very large negative value of the exponential when the dimension m or the time length T is particularly large. And using the improved EM algorithm, only N times of calculation is needed for each iteration, while N×K times of calculation is needed for each iteration in the traditional EM algorithm, which greatly reduces the calculation cost.

[0079] Further, before the automatic labeling method starts iteration, the total number of categories K and the preset order set {p1,…,p k} of the MVAR model are selected: the total number of categories K and the preset order set {p1,…,p k} of the MVAR model can be determined based on empirical values, or can be selected based on the BIC standard. However, in the original BIC standard, the model with a larger K obtains a higher probability than the model with a smaller K, which violates the simplicity criterion in many scenarios, resulting in reduced performance of the model selected under the parameters, and thus is not applicable. Therefore, preferably, in an optional implementation, the total number of categories K and the preset order set {p1,…,p k} of the MVAR model can be selected by using an improved extended BIC standard for model parameter selection to ensure the consistency of the model selection problem, which specifically includes:

[0080] K is randomly selected from an ordered set of N K integers, p1,…,p k are randomly selected from an ordered set of N p integers, and a plurality of groups of values of K, p1,…,p k are obtained; N K and N p are integers greater than or equal to 2.

[0081] The following formula is used to calculate each group K, p1,..., p k The BIC value under the value;

[0082]

[0083] Wherein,

[0084]

[0085]

[0086] η mix =N

[0087] S j K and p1,..., p k The set of all possible values of these two types of parameters; The total number of elements in the set It should be noted that p k The same value can be selected, but due to the nature of clustering, only their combination is important, therefore, the size of the model set The number of K combinations from a group of N p elements when the elements are allowed to repeat;

[0088] Get a group K, p1,..., p k The BIC value is the smallest, to set the total number of categories K and the preset order set of MVAR model {p1,..., p k}. Specifically, the selected order set and the number of categories are substituted into the above expression to obtain the BIC value, and in general, the model with lower BIC value is more recommended.

[0089] In order to further illustrate the performance of the industrial field time series data automatic labeling method provided by the present application, the automatic labeling effects of other various existing advanced methods and the method (MVAR limit model) are compared, in order to better verify the method proposed by the present application, RI and NMI are used as evaluation indexes, and the expressions are as follows:

[0090]

[0091]

[0092] Where TP represents true positive, TN represents true negative, N is the total number of time series, |G i | and |A j | is the number of time series of clusters G i and A j N ij =|G​i ∩A j | represents the time series belonging to the set G i and A j The intersection of the two values of RI and NMI close to 1 indicate high automatic labeling performance.

[0093] The automatic labeling performance of the method of the present application and the k-DBA, k-Shape, k-SC, u-shapelet and USSL on the "WAFER" data set is compared and summarized in Table 1. The "WAFER" data comes from a real manufacturing process from a semiconductor microelectronic manufacturing application involving hundreds of processing steps and online sensors for quality control. As can be seen from Table 1, the accuracy of the present application is higher in the RI and NMI performance indicators compared to other models and methods, verifying the effectiveness of the present application.

[0094] Table 1

[0095] k-DBA K-Shape k-GAK k-SC U-SHAPELET USSL MVAR limit model RI 0.4925 0.4925 0.5008 0.4925 0.4925 0.8246 0.8227 NMI 0 0.001 0.0054 0.001 0.001 0.0492 0.1432

[0096] Figure 2 and Figure 3 The improved EM algorithm in the present application, i.e. the improved MVAR limit method with small noise limit (denoted as MVAR Limit) and the MVAR method using the traditional EM algorithm (denoted as MVAR) are compared in scalability, wherein, Figure 2 is the performance curve of the improved EM algorithm in the present application and the MVAR method using the traditional EM algorithm under different lengths of industrial field time series data, Figure 3 is the performance curve of the improved EM algorithm in the present application and the MVAR method using the traditional EM algorithm under different numbers of sensors, and the lost data points or the points marked with red circles indicate the failure cases in the current problem setting; as can be seen from the figure, compared to the traditional EM algorithm, the running time of the improved EM algorithm is shorter and the failure cases are fewer in the case of particularly large time vector dimension caused by too long time series length or too many sensors.

[0097] In summary, the application develops a new scalable clustering algorithm for self-correlation vector time series. In particular, the fundamental feature of (vector) time series, which is self-correlation, is explicitly exploited for the first time. A new EM-type clustering algorithm is developed based on the MVAR model. Further, a new EM algorithm is designed by employing a smaller noise restriction to solve the computational bottleneck and numerical underflow problem, and to reduce the time used for classification. Further, a novel and unusual extension of the BIC criterion is designed to select the number of clusters and the MVAR order for the automatic parameter selection before the iteration of the automatic labeling method. The application uses the MVAR model for time series clustering, realizes the automatic classification and labeling of industrial time series data, designs a computational optimization algorithm, and thus greatly improves the computational performance, meeting the requirements of rapidity and accuracy of time series data labeling in industrial sites.

[0098] Embodiment 2

[0099] An automatic labeling system for industrial site time series data, comprising a memory and a processor, the memory storing a computer program, and the processor executing the computer program to execute the automatic labeling method for industrial site time series data provided in Embodiment 1 of the application.

[0100] The related technical solutions are the same as those in Embodiment 1, and will not be repeated here.

[0101] Embodiment 3

[0102] A computer readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls a device in which the storage medium is located to execute the automatic labeling method for industrial site time series data provided in Embodiment 1 of the application.

[0103] The related technical solutions are the same as those in Embodiment 1, and will not be repeated here.

[0104] Those skilled in the art will readily understand that the above description is only of the preferred embodiments of the application and is not intended to limit the application, and any modifications, equivalent replacements and improvements made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. An automatic annotation method for industrial field time series data, characterized in that, include: MVAR models are established for each industrial site time series data to be labeled, and then the conditional probability distribution of the data under a given category is obtained. By combining the conditional probability distributions of time series data from various industrial sites under a given category, the likelihood functions of all industrial site time series data are obtained. Then, maximum likelihood estimation is performed to obtain the probability that each industrial site time series belongs to a different category, and the category of the industrial site time series data is labeled accordingly. Among them, the nth industrial site time series data is a sequence composed of sensor data vectors at all times during the nth processing process; the sensor data vector at each time moment is a vector composed of the data collected by all sensors on the processing line at that time moment. The MVAR model established for the nth industrial site time series data is as follows: in, The data refers to historical data prior to time t. Under the conditions The probability of occurrence; The sequence is formed by the sensor data vectors at the first t moments in the nth industrial field time series data; K is the total number of preset categories; The historical data before time t is the data. Given the condition, the probability that any industrial field time series data belongs to the k-th category; Φ(·) is the multivariate cumulative distribution function of a Gaussian distribution with zero mean and unit covariance matrix; This is the coefficient matrix of size m×m corresponding to the k-th category; The k-th order in the preset order set of the MVAR model; These are all parameters of the MVAR model. It is an m-dimensional coefficient vector. All are coefficient matrices of size m×m; m is the total number of sensors on the processing line.

2. The automatic annotation method for industrial field time series data according to claim 1, characterized in that, The conditional probability distribution of the nth industrial site time series data under a given category is: in, This indicates that the nth industrial site time series data belongs to the kth category label; Preset the maximum order in the order set for the MVAR model; ; ; .

3. The automatic annotation method for industrial field time series data according to claim 2, characterized in that, Logarithmic expression of the likelihood function for all industrial field time series data for: = Where N is the total number of time series data from the industrial site; This represents the total number of time steps in the processing.

4. The automatic annotation method for industrial field time series data according to claim 3, characterized in that, The method for performing maximum likelihood estimation includes the following steps: initialization , and ;in, Let be the probability that the nth industrial site time series data belongs to the kth category; n = 1, ..., N; In step E: iterate through k=1,…,K, if Then Update to 1, otherwise, Updated to 0; where, , ; In step M: Will Updated to ; Get When The set of all n at time n is used as the current set. ; Will Updated to ; Will Updated to ;in, For set The number of elements in the middle; Iterate alternating between E-steps and M-steps until... , and All converge, at this point This is the probability that the nth industrial site time series data belongs to the kth category.

5. The automatic annotation method for industrial field time series data according to claim 3 or 4, characterized in that, The total number of categories K and the preset order set of the MVAR model { The settings for} include: K from N K Arbitrarily select from the ordered set of integers, All from N p Multiple sets can be obtained by randomly selecting from an ordered set of integers. The value of N; K and N p All are integers greater than or equal to 2; Calculate each group using the following formula. BIC value under the given value; in, = For K and The set of all possible values ​​for these two types of parameters; For set The total number of elements in the card (S) j )= ; Get the group with the smallest BIC value To set the total number of categories K and the preset order set of the MVAR model { } 6. An automatic annotation system for industrial field time series data, characterized in that, include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the automatic labeling method for industrial field time series data as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is run by a processor, it controls the device where the storage medium is located to perform the automatic labeling method for industrial field time series data as described in any one of claims 1-5.

Images