Method, system and program product for data transmission with a reduced data volume

The invention relates to a method for data transmission with a reduced data volume.

It further relates to a corresponding computer device or sys- tern for data transmission with a reduced data volume.

In addition, the invention relates to a program product comprising code for directing a computer device or system to carry out a method for data transmission with a reduced data volume.

Systems which automatically collect and process very large volumes of data are increasingly used and implemented in the fields of, e.g., industry, energy, transportation and health- care. Examples of such data may be measurement and/or system information of an automated factory or in any industrial control and factory automation, or medical records and/or image data in healthcare service. For diagnostic, error-detection and predictive-maintenance purposes or to be able to show compliance with internal policies and external regulations, it is often necessary to record such collected and/or pro ^¬ cessed data over a long time period, so that it can be re ^¬ trieved and evaluated whenever required. For this purpose, the data to be recorded is event-triggered or periodically transmitted to a data storage, which may be a single device or may be spread across multiple storage devices such as in a storage area network. The storage devices can be shared by multiple client computing devices, e.g., via a network such as in a cloud storage implementation.

In the applications mentioned above, the volume of the data to be recorded is very large and may lead to high transmis ^¬ sion and storage costs. One approach to reduce the costs is to perform data compression before the transmission. However, as the data is not only large in size but usually also very complex, the challenge is to design a data compression scheme that is easy to implement. Another approach could be to have a central data collector where the raw data is collected. This device would then perform data manipulation, such as averaging. The challenge here is that the quality of such mani ^¬ pulated data is very bad because it deviates largely from the actual raw data. From Z. Banko et al . : "PCA Driven Similarity for Segmented Univariate Time Series", Hungarian Journal of Industrial Chemistry, 37 (2009) 59-67 it is known to measure the similarity of different time series of data collected in, e.g., chemical process industry. The originally multivariate time series are projected in single dimensions by means of princi ^¬ pal component analysis (PCA) . For each dimension, the univariate time series obtained are aligned by means of dynamic time warping (DTW) , and the similarity of the time series is calculated from the distances between the aligned data points. To reduce the DTW computational effort, each uni ^¬ variate time series is approximated by a sequence of seg ^¬ ments, each one represented by the mean value of its data points (piecewise aggregate approximation (PAA) ) or a linear function (piecewise linear approximation (PLA) ) , and the DTW algorithm is applied to the representations.

It is an object of the invention to provide data transmission to a logging/data storage with a reduced data volume while ensuring high data transfer quality and reducing the costs largely.

According to the invention, this is achieved by a computer- implemented method for data transmission with a reduced data volume, which method comprises:

- receiving a multivariate data stream including a plurality of data points;

- sampling the data stream in consecutive time intervals and generating consecutive data packages which each contain data points sampled in a window of a predetermined number of said consecutive sampling time intervals;

- calculating, by means of principal component analysis, for each data package projections of its data points onto a set of principal components;

- aligning, by means of dynamic time warping, for each newly generated data package and for each principal component the projected data points to respective projected data points of an earlier generated data package and calculat- ing distances between the aligned data points;

- determining the maximum of the calculated distances and comparing it to a threshold value; and

- transmitting the newly generated data package to a data sink and setting the newly generated data package to be the earlier generated data package for the next aligning step if the maximum distance exceeds the threshold value or otherwise inhibiting the transmission and keeping the earlier generated data package for the next aligning step. Rather than data manipulation or data compression, the method according to the invention generates and selects packages of data to be transmitted on the basis of some data change in each current (new) data package as compared to an earlier (old) data package. If a non-negligible or substantial change in the data is detected, the current data package is trans ^¬ mitted for storing in, e.g., a cloud since it reveals new in ^¬ formation. Otherwise, i.e. if no or only a negligible change is detected, the data package will not be transmitted. In this way, the total volume of transmitted data is signifi- cantly reduced without changing the data and thus compromis ^¬ ing the data quality.

The change detection method is statistically based, discover ^¬ ing the data change that is contained in multivariate data. First, the multivariate data is investigated by principal component analysis (PCA) , and the variance in the data is captured by a given number of k principal components, prefer ^¬ ably the k top principal components (i.e. eigenvectors with the largest eigenvalues) . Subsequently, dynamic time warping (DTW) is applied to compare the principal components between new data and old data in order to detect a data change. Dy ^¬ namic time warping is a well-known technique for determining the similarity between two temporal sequences of data by warping and aligning the sequences and then measuring a distance-like quantity between them.

The earlier (old) data package used for the comparison with the current (new) data package is kept and used until a data change is detected. Once a data change has been detected, the current (new) data package replaces the earlier (old) one and is be used for subsequent comparisons.

The maximum distance used for the comparison with the thresh- old value may be the largest one of individual maximum dis ^¬ tances calculated for each principal component or a cumula ^¬ tive value of these individual maximum distances. The indi ^¬ vidual maximum distances, in turn, are determined from the distances (individual or accumulated, weighted or direct) be- tween the aligned data points.

The threshold value can be preferably adapted to historical maximum distances to make it more robust to outliers. For ex ^¬ ample, the threshold value may be calculated from an average of the maximums determined for a predetermined number of the most recently data packages plus a margin percentage of, e.g., 5% .

These and further embodiments will be apparent from the de- tailed description and examples that follow. For this, refer ^¬ ence is made to the accompanying drawings, which show, by way of example, preferred embodiments of the present invention, and in which:

FIG. 1 illustrates an example of an industrial automation system which retrieves and collects data for trans ^¬ mitting to a data logging system; FIG. 2 is a flowchart depicting steps of an exemplary embodiment of a method for data transmission with a reduced data volume; and FIG. 3 is a timing diagram exemplarily illustrating generating and selecting of data packages for transmis ^¬ sion and storage.

FIG. 1 is a simplified block scheme illustrating an industri- al automation system 1. A control system 2 receives measure ^¬ ment signals 3 from sensing means 4 in a system 5 under control, e.g. a technical process or system, and outputs actuat ^¬ ing signals or control commands 6 to operating means 7 (e.g. actors) of the technical process 5. The control system 2 may be decentralized where distributed control components 8 con ^¬ tain and perform automation functions for different process areas whereas central components 9 contain and perform opera ^¬ tor control, monitoring, evaluation and other central coordination functions. One such central function is to automati- cally retrieve and collect very large amounts of data 10 for condition monitoring purposes in the broadest sense, which may cover, inter alia, fault diagnostic, predictive-mainte ^¬ nance, product quality, environmental performance, plant safety, plant performance and so on. The data 10 can be col- lected continuously or periodically so that a data stream is produced. The data 10 is usually multivariate and multidimen ^¬ sional (dependent and/or independent) and can include all technical and non-technical, raw and processed data which is gathered or generated by the automation system 1. The automa- tion system 1 illustrated is only exemplary and representa ^¬ tive of any other system which automatically collects huge volumes of data for recording. The processing and analysis of the collected data 10 for the above-mentioned purposes and/or legal provisions for data retention require recording of the data 10 over a long time period. To this end, and in view of the huge data volume to be recorded, the data 10 is transmit ^¬ ted to an adequate data storage 11 outside the automation system 1, which may be a single device or, as shown, spread across multiple storage devices in a cloud storage network. Furthermore, in order to reduce transmission and storage costs, the volume of the data 10 to be transmitted is reduced by a computing device or system 12 which may be part of the automation system 1, and, in particular, part the of central components 9.

The computing device or system 12 comprises a number of units 13, 14, 15, 16, 17, the functions of which are explained in detail below with reference to FIG. 2.

FIG. 2 is a flowchart exemplarily depicting steps a method for reducing the volume of the data 10 to be transmitted without changing the data 10. In a step 18, a receiving unit 13 receives the collected data 10 in the form of an incoming multivariate data stream of a plurality of data points. Assuming the data stream 10 is mul ^¬ tivariate with n variables, it can be represented as a ma ^¬ trix :

with the first sub-script denoting the time t and the second sub-script denoting the i-th variable. Hence, Vn represents the first variable at time 1. As the data arrive continuous ^¬ ly, the matrix has its time dimension t infinite.

In a step 19, a sampling unit 14 samples the data stream in consecutive time intervals (time slots) and generates consec- utive data packages P which each contain data points sampled in a predetermined number of the consecutive time intervals. Thus, the incoming data stream 10 is partitioned into a se ^¬ quence of time windows w (see FIG 3) of equal size, each con ^¬ sisting of a pre-selected number of sampling time intervals. In a step 20, a calculating unit 15 calculates, by means of principal component analysis (PCA) , for each newly generated data package P projections of its data points onto a set of principal components. The projected data is hereinafter de- noted as Z = z _lr ... , z _k . The orthogonality of the principal components ensures that the variance in the data 10 can be captured by a few components, where using the first k princi ^¬ pal components with the largest eigenvalues will cover about 90% of data variance. An approximate value of k can be around 30. Hence, data reduction can be performed by projecting the data points on a couple of dimensions. In reality, the data 10 contains often missing values, which, however, the PCA method can deal with. In step 21, the projected data Z = z _1; ...,z _k of the data package P generated for the first time is stored as a refer ^¬ ence projection data R = ri, ...,r _k .

In a step 22, the calculating unit 15 compares, for each of the k principal components, the projected data Z of the cur ^¬ rent (newly generated) data package P to the corresponding reference projection data R. To this end, the projected data Z and R are aligned by means of dynamic time warping (DTW) . Each possible mapping of aligned projected data points z± and r _j (1 < i,j < k) represents a warping path in an i x j matrix. With optimal alignment of Z and R, the cost function of the warping path is minimal, which cost function is the square of the distances d( z i,r _j ) = (z±-r _j ) ² between the aligned data points z±, r _j is minimal. A dynamic programming approach is usually employed to evaluate the warping path. In addition, a window can be applied to calculate the distance of the data points following into it. Finally, a cumulative distance is represented as the distance between the two se ^¬ quences (projected data Z and R) . Therefore, similar sequenc- es have smaller distance and vice versa.

For more information on DTW it is referred to Ratanamahatana C. A., Keogh E.: "Everything you know about dynamic time warping is wrong", Proc. 3rd Workshop on Mining Temporal and Sequential Data, in conjunction with 10th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2004), August 22-25, 2004. In a step 23, a judging unit 16 determines whether there is a non-negligible or substantial change in the data of the cur ^¬ rent (new) data package P compared to the earlier (old) ref ^¬ erence data package. To this end, the maximum distance d _max among all the distances calculated for the k dimensions

(principal components) is determined and, in a step 24, com ^¬ pared to a threshold value Th. The maximum distance d _max is regarded as the overall distance between the projected data Z of the current data package and the reference projection data R. The maximum distance d _max will be low or zero if these two sets of data are more or less identical, which means the data has remained unchanged. In this case, the maximum distance d _max will remain below the threshold value Th and, as a conse ^¬ quence, the current data package P will not be transmitted to the data storage 11. This means that the transmitting unit 17 will be disabled from transmitting. Conversely, if the maxi ^¬ mum distance d _max , exceed the threshold value Th, a non- negligible or substantial change in the data is considered to be detected and the judging unit 16 enables in a step 25 the transmitting unit 17 to transmit the current data package P (sampled from the original data stream 10) to the data stor ^¬ age 11.

The threshold value Th can be preferably adapted to histori ^¬ cal maximum distances to make it more robust to outliers. For example, the threshold value may be calculated from an aver ^¬ age of the maximums determined for a predetermined number of the, e.g., 10 most recently data packages, plus a margin per ^¬ centage of, e.g., 5%. FIG. 3 exemplarily shows a sequence of data packages Pi, P ₂ , P _h , P _h+ i, P _h+2 which are generated from the data stream 10 at the times t _lr t ₂ , t _h , t _h+ i, t _h+2 . Each data package P covers a time windows w of the same size, in which the data stream 10 is sampled in a pre-selected number of time intervals. The current (new) data packages are marked by solid lines and the earlier (older) ones by dotted lines. The first data package contains new information and is therefore transmitted to the storage 11 which is indicated by an arrow 26.

A curve designated by reference number 27 depicts data chang ^¬ es of the data 10 sampled over the time t. As explained above, this change is detected for each new data package, e.g. P ₂ , by projecting its data points onto a set of princi- pal components, aligning for each principal component, by means of dynamic time warping, the projected data points to respective projected data points of an earlier generated data package, e.g. Pi, and calculating distances between the aligned projected data points of the two data packages Pi, P2 · The maximum d _max of the calculated distances is a measure of the data change and is compared to the threshold value Th. The new data package PI is only transmitted to the storage 11 if the maximum distance d _max exceeds the threshold value Th, which is not the case for the data package P2. The earlier data package Pi is therefore kept as a reference for the com ^¬ parison with the next new data package P3, which function of a reference is here indicated by hatching of the data package Pi. The data package P _h+ i is the first one where the calculated maximum distance d _max exceeds the threshold value Th. As a re ^¬ sult, the data package P _h+ i is transmitted to the data storage 11 and the data package P _h+ i replaces the earlier data package P _h+ i as the reference used for next comparisons.