Failure prediction

1. A failure prediction system, comprising:

a model-based signature generator generating feature vectors from individual attributes of multi-variate time series data based on sequence importance and attribute importance;

a knowledge database storing feature vectors corresponding to a set of different failure types; and

a set of detectors, each detecting any of the feature vectors generated by the model-based signature generator that are similar to any of the feature vectors corresponding to a respective one of the different failure types stored in the knowledge database based on a similarity threshold and outputting the respective one of the different failure types and a likely time period when the respective one of the different failure types will occur,

wherein the model-based signature generator generates the feature vectors by:

extracting, by a feature extractor, feature values from the individual attributes of the multi-variate time series data and concatenating the feature values into the feature vectors;

assigning, by a feature selector, weights to each of the feature vectors;

converting, by a feature converter, multiple weighted ones of the feature vectors into a new feature vector; and

jointly training the feature extractor, the feature selector, and the feature converter.

2. The failure prediction system of claim 1 , wherein the model-based signature generator generates the feature vectors as a matrix.

3. The failure prediction system of claim 1 , wherein the feature extractor comprises a set of Recurrent Neural Networks.

4. The failure prediction system of claim 1 , wherein the feature vectors corresponding to the set of different failure types form a historical failure record of a target system sourcing the multi-variate time series data.

5. The failure prediction system of claim 1 , wherein the feature vectors corresponding to the set of different failure types are used to form a prediction model.

6. The failure prediction system of claim 5 , wherein the prediction model is trained by minimizing a triplet loss involving a first distance between an anchor and a positive sample, a second distance between the anchor and a negative sample, and a margin between the first and second distances.

7. The failure prediction system of claim 1 , wherein the multi-variate time series data is divided into a set of different regions, the set of different regions comprising normal regions and early symptom regions.

8. A computer-implemented method for failure prediction, comprising:

generating, by a model-based signature generator, feature vectors from multi-variate time series data based on sequence importance and attribute importance;

storing, by a knowledge database, feature vectors corresponding to a set of different failure types; and

detecting, by each of detectors in a set, any of the feature vectors generated by the model-based signature generator that are similar to any of the feature vectors corresponding to the set of different failure types stored in the knowledge database based on a similarity threshold and outputting the respective one of the different failure types and a likely time period when the respective one of the different failure types will occur,

wherein said generating step comprises:

extracting, by a feature extractor, feature values from the individual attributes of the multi-variate time series data and concatenating the feature values into the feature vectors;

assigning, by a feature selector, weights to each of the feature vectors; and

converting, by a feature converter, multiple weighted ones of the feature vectors into a new feature vector,

wherein the method further comprises jointly training the feature extractor, the feature selector, and the feature converter.

9. The computer-implemented method of claim 8 , wherein the feature vectors are generated by the model-based signature generator as a matrix.

10. The computer-implemented method of claim 8 , wherein the feature extractor comprises a set of Recurrent Neural Networks.

11. The computer-implemented method of claim 8 , wherein the feature vectors corresponding to the set of different failure types form a historical failure record of a target system sourcing the multi-variate time series data.

12. The computer-implemented method of claim 8 , further comprising forming a prediction model using the feature vectors corresponding to the set of different failure types.

13. The computer-implemented method of claim 12 , wherein the prediction model is trained by minimizing a triplet loss involving a first distance between an anchor and a positive sample, a second distance between the anchor and a negative sample, and a margin between the first and second distances.

14. A computer program product for failure prediction, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:

generating, by a model-based signature generator of the computer, feature vectors from multi-variate time series data based on sequence importance and attribute importance;

storing, by a knowledge database of the computer, feature vectors corresponding to a set of different failure types; and

detecting, by each of detectors of the computer, any of the feature vectors generated by the model-based signature generator that are similar to any of the feature vectors corresponding to the set of different failure types stored in the knowledge database based on a similarity threshold and outputting the respective one of the different failure types and a likely time period when the respective one of the different failure types will occur,

wherein said generating step comprises:

extracting, by a feature extractor, feature values from the individual attributes of the multi-variate time series data and concatenating the feature values into the feature vectors;

assigning, by a feature selector, weights to each of the feature vectors; and

converting, by a feature converter, multiple weighted ones of the feature vectors into a new feature vector,

wherein the method further comprises jointly training the feature extractor, the feature selector, and the feature converter.

15. The computer program product of claim 14 , wherein the feature vectors are generated by the model-based signature generator as a matrix.