Machine learning with segment-aligned multisensor trace data

A method for machine learning with segment-aligned multisensor trace data, the method comprising:

• a machine-learning system receiving, from a set of sensors, a set of time series from a set of sensors that each comprises a chronological sequence of measurements of an attribute, of a set of attributes of a mechanical assembly unit, of an instance, of a set of instances of measured entities;
• the machine-learning system selecting for each received time series a corresponding initial segmentation method of a set of segmentation methods,
• where a first selected initial segmentation method partitions measurements comprised by a corresponding first time series of the received time series into a first sequence of at least two contiguous target segments, and
• where the first selected initial segmentation method determines a number of segments in the first sequence and a number of measurements comprised by each segment of the first sequence;
• the machine-learning system designating one series of the received time series to be a first standard series for a first attribute of the set of attributes, where the first standard series comprises measurements of the first attribute;
• the machine-learning system time-shifting measurements of each target series of the set of received time series, such that each time-shifted measurement is time-aligned with a corresponding measurement of the first standard series,
• where measurement of each target series is a measurement of the first attribute,
• where the time-shifting aligns each target series to the first standard series by performing a method of dynamic time warping (DTW), and
• where the machine-learning system determines that a first target segment of the first target series, of the set of received time series, should be segment-aligned to a first standard segment of the first standard series by determining that a number of measurement points of the first target segment that have been aligned by the DTW procedure to any measurement point of the first standard segment is at least as large as a number of measurement points of any other segment of the first target series that have been aligned by the DTW procedure to any measurement point of the first standard segment;
• the machine-learning system modifying a corpus configured to train a self-learning application by incorporating each segment-aligned series into the corpus, said corpus being a data storage device;
• the machine-learning system training the self-learning application by submitting the modified corpus to the self-learning application, wherein said training the self-learning application comprises:
• (i) using training data to perform the training, said training data comprising each time-shifted measurement that is time-aligned with the corresponding measurement of the first standard series, (ii) teaching the self-learning application to make a prediction of a likely failure of the mechanical assembly unit's functional quality-assurance test, wherein said making the prediction is in response to the self-learning application identifying adverse conditions, as measured by the set of sensors, on the set of attributes of the mechanical assembly unit as the mechanical assembly unit leaves an assembly line of a manufacturing process that manufactured the mechanical assembly unit, and (iii) gaining experience, by the self-learning application, that allows the self-learning application to infer a semantic meaning from behavior of the set of attributes of the mechanical assembly unit as the mechanical assembly unit leaves the assembly line and to use the inferred semantic meaning to predict the likely failure of the mechanical assembly unit's functional quality-assurance test;
• and
• executing the trained self-learning application using the inferred semantic meaning to predict the likely failure of the mechanical assembly unit's functional quality-assurance test,
• wherein the machine-learning system uses a method of Lavielle to select an initial segmentation method for each received time series,
• wherein the selection of each initial segmentation method comprises:
• using the method of Lavielle to assign a contrast-function value to each received time series, and determining an optimal number of segments L into which a particular received time series should be segmented by choosing a smallest value of L for which a second derivative of the particular received time series' contrast function is greater than a threshold S.

A computer program product, comprising a computer-readable hardware storage device having a computer-readable program code stored therein, the program code configured to be executed by a machine-learning system comprising a processor, a memory coupled to the processor, and a computer-readable hardware storage device coupled to the processor, the storage device containing program code configured to be run by the processor via the memory to implement a method for machine learning with segment-aligned multisensor trace data, the method comprising:

• the machine-learning system receiving, from a set of sensors, a set of time series from a set of sensors that each comprises a chronological sequence of measurements of an attribute, of a set of attributes of a mechanical assembly unit, of an instance, of a set of instances of measured entities;
• the machine-learning system selecting for each received time series a corresponding initial segmentation method of a set of segmentation methods,
• where a first selected initial segmentation method partitions measurements comprised by a corresponding first time series of the received time series into a first sequence of at least two contiguous target segments, and
• where the first selected initial segmentation method determines a number of segments in the first sequence and the number of measurements comprised by each segment of the first sequence;
• the machine-learning system designating one series of the received time series to be a first standard series for a first attribute of the set of attributes, where the first standard series comprises measurements of the first attribute;
• the machine-learning system time-shifting measurements of each target series of the set of received time series, such that each time-shifted measurement is time-aligned with a corresponding measurement of the first standard series,
• where measurement of each target series is a measurement of the first attribute,
• where the time-shifting aligns each target series to the first standard series by performing a method of dynamic time warping (DTW), and
• where the machine-learning system determines that a first target segment of the first target series, of the set of received time series, should be segment-aligned to a first standard segment of the first standard series by determining that a number of measurement points of the first target segment that have been aligned by the DTW procedure to any measurement point of the first standard segment is at least as large as a number of measurement points of any other segment of the first target series that have been aligned by the DTW procedure to any measurement point of the first standard segment;
• the machine-learning system modifying a corpus configured to train a self-learning application by incorporating each segment-aligned series into the corpus, said corpus being a data storage device;
• the machine-learning system training the self-learning application by submitting the modified corpus to the self-learning application, wherein said training the self-learning application comprises:
• (i) using training data to perform the training, said training data comprising each time-shifted measurement that is time-aligned with the corresponding measurement of the first standard series, (ii) teaching the self-learning application to make a prediction of a likely failure of the mechanical assembly unit's functional quality-assurance test, wherein said making the prediction is in response to the self-learning application identifying adverse conditions, as measured by the set of sensors, on the set of attributes of the mechanical assembly unit as the mechanical assembly unit leaves an assembly line of a manufacturing process that manufactured the mechanical assembly unit, and (iii) gaining experience, by the self-learning application, that allows the self-learning application to infer a semantic meaning from behavior of the set of attributes of the mechanical assembly unit as the mechanical assembly unit leaves the assembly line and to use the inferred semantic meaning to predict the likely failure of the mechanical assembly unit's functional quality-assurance test;
• and
• executing the trained self-learning application using the inferred semantic meaning to predict the likely failure of the mechanical assembly unit's functional quality-assurance test,
• wherein the machine-learning system uses a method of Lavielle to select an initial segmentation method for each received time series,
• wherein the selection of each initial segmentation method comprises:
• using the method of Lavielle to assign a contrast-function value to each received time series, and determining an optimal number of segments L into which a particular received time series should be segmented by choosing a smallest value of L for which a second derivative of the particular received time series' contrast function is greater than a threshold S.