System and method for explanation of condition predictions in complex systems

A monitoring and alert apparatus for a machine system comprising sensors generating a plurality of signals over time, comprising:

• a computer system comprising a plurality of inputs to receive time series signals from a plurality of sensors;
• the computer system comprising a processor and a memory adapted with instructions forming:
• logic to receive a labeled point comprising, for a range of time, a first time series from each of the plurality of signals, each first time series being associated with at least one feature vector, each feature vector having components corresponding to one or more signal features, the labeled point having a first label describing a first condition of a plurality of conditions of the machine system, each of the plurality of signals having a corresponding signal projection subspace;
• logic to projecting each of a plurality of model condition points into each signal projection subspace of the plurality of signal projection subspaces,
• each model condition point comprising, for a range of time specific to the model condition point, a respective time series from each of the plurality of signals, each respective time series being associated with at least one feature vector, each feature vector having components corresponding to one or more signal features, the model condition point having a respective label describing one of the plurality of conditions of the machine system, the projecting comprising evaluating the at least one feature vector associated with the respective time series from each of the plurality of signals in the corresponding signal projection subspace;
• logic to project the labeled point into each signal projection subspace, comprising evaluating the at least one feature vector associated with the first time series from each of the plurality of signals in the corresponding signal projection subspace,
• wherein in each signal projection space corresponding to a signal of the plurality of signals, the at least one feature vector associated with only the signal from the labeled point and each model condition point are evaluated;
• logic to constrain an analysis neighborhood for the labeled point to a set number of model condition points closest to the labeled point as projected in each signal projection subspace corresponding to a signal of the plurality of signals,
• wherein the respective at least one feature vectors associated with the signal from the set number of model condition points are closest to the at least one feature vector associated with the signal from the labeled point;
• logic to calculate, for each signal projection subspace, a first percent of model condition points in the respective analysis neighborhood having a same label as the labeled point in the analysis neighborhood out of the set number of model condition points in the respective analysis neighborhood;
• logic to calculate, for each signal projection subspace, a second percent of model condition points having a same label as the labeled point out of the plurality of model condition points;
• logic to calculate a contribution of each of the plurality of signals to the first condition from the first percent and the second percent calculated for the corresponding signal projection subspace, to form a sorted list of signals and contributions;
• and
• logic to display signals that are higher in the sorted list of signals as the most likely contributors for the machine system's condition corresponding to the labeled point.

A non-transitory machine-readable storage medium storing instructions that when executed by a processor, cause the processor to execute a control method for a machine system comprising sensors generating a plurality of signals over time, the method comprising:

• receiving a labeled point comprising, for a range of time, a first time series from each of the plurality of signals, each first time series being associated with at least one feature vector, each feature vector having components corresponding to one or more signal features, the labeled point having a first label describing a first condition of a plurality of conditions of the machine system, each of the plurality of signals having a corresponding signal projection subspace;
• projecting each of a plurality of model condition points into each signal projection subspace of the plurality of signal projection subspaces,
• each model condition point comprising, for a range of time specific to the model condition point, a respective time series from each of the plurality of signals, each respective time series being associated with at least one feature vector, each feature vector having components corresponding to one or more signal features, the model condition point having a respective label describing one of the plurality of conditions of the machine system, the projecting comprising evaluating the at least one feature vector associated with the respective time series from each of the plurality of signals in the corresponding signal projection subspace;
• projecting the labeled point into each signal projection subspace, comprising evaluating the at least one feature vector associated with the first time series from each of the plurality of signals in the corresponding signal projection subspace,
• wherein in each signal projection space corresponding to a signal of the plurality of signals, the at least one feature vector associated with only the signal from the labeled point and each model condition point are evaluated;
• constraining an analysis neighborhood for the labeled point to a set number of model condition points closest to the labeled point as projected in each signal projection subspace corresponding to a signal of the plurality of signals,
• wherein the respective at least one feature vectors associated with the signal from the set number of model condition points are closest to the at least one feature vector associated with the signal from the labeled point;
• for each signal projection subspace, calculating a first percent of model condition points in the respective analysis neighborhood having a same label as the labeled point in the respective analysis neighborhood out of the set number of model condition points in the analysis neighborhood;
• for each signal projection subspace, calculating a second percent of model condition points in the model having a same label as the labeled point out of the plurality of model condition points;
• calculating a contribution of each of the plurality of signals to the first condition from the first percent and the second percent calculated for the corresponding signal projection subspace, to form a sorted list of signals and contributions;
• and
• adapt the machine system's behavior based on the sorted list.