Physics-informed partial least squares regression modeling for failure detection in power electronic devices

1 . A system, comprising:

at least one memory storing machine-executable instructions; and

at least one processor configured to access the at least one memory and execute the machine-executable instructions to:

generate a feature set comprising qualified switching parameters to predict one or more faults of switching devices, wherein generating the feature set comprises:

iteratively ranking a level of effectiveness of each candidate switching parameter within a candidate set of candidate switching parameters; and

iteratively removing one or more candidate switching parameters within the candidate set based on the ranking, until a number of retained candidate switching parameters is consistent with a determined size of the feature set;

based on the feature set, predict a fault of a particular switching device, wherein predicting the fault comprises:

predicting, by a machine learning model, a value of a switching parameter of the particular switching device based on one or more switching parameters corresponding to the feature set; and

based on a difference between an actual value of the switching parameter and the predicted value of the switching parameter, generating a test statistic; and

in response to the test statistic identifying an onset of a fault, generate an anomaly signal based on the generated first-test statistic.

2 . The system of claim 1 , wherein the switching parameter comprises an on-state current I ds for the particular switching device.

3 . The system of claim 2 , wherein the value for the on-state current I ds is predicted using values of at least one of body-diode voltage V bd , voltage between drain and source V ds , and squared voltage between gate and source V 2 gs .

4 . The system of claim 1 , wherein the difference is part of a time-series of residuals, and wherein, when the machine-executable instructions are executed, the test statistic is generated by performing a cumulative sum (CUSUM) test on a mean of the residuals in the time-series.

5 . The system of claim 1 , wherein the switching parameter comprises an on-state current I ds for the particular switching device, the machine learning model comprises a first machine learning model, the difference comprises a first difference, the test statistic comprises a first test statistic, and wherein the at least one processor is configured to access the at least one memory and execute the machine-executable instructions to:

predict, by a second machine learning model, a value of a drain-source resistance R ds of the particular switching device, wherein the prediction for the drain-source resistance R ds is based on a present value of another switching parameter of the particular switching device and a prior value of R ds ;

determine a second difference between the predicted value of R ds and an actual value of drain-source resistance R ds ;

generate a second test statistic based on the second difference; and

selectively generate the anomaly signal based on the first test statistic and the second test statistic.

6 . The system of claim 1 , wherein the at least one processor is configured to access the at least one memory and execute the machine-executable instructions to determine on-state values for the one or more switching parameters.

7 . The system of claim 1 , wherein the at least one processor is configured to access the at least one memory and execute the machine-executable instructions to perform:

evaluating first subset levels of effectiveness corresponding to first subsets of the candidate switching parameters, the first subsets of the candidate switching parameters each having a first subset size less than the determined size of the feature set,

deleting a candidate switching parameter resulting in a lowest decrease in a subset level of effectiveness in each of the first subsets, wherein deleting causes generating of second subsets, each second subset having a second subset size that is one less than the first subset size;

evaluating second subset levels of effectiveness corresponding to the second subsets of the candidate switching parameters; and

generating the feature set based on the second subset levels of effectiveness.

8 . The system of claim 1 , wherein the machine learning model is a partial least squares regression model.

9 . The system of claim 1 , wherein the at least one processor is further configured to execute the machine-executable instructions to: cause a physical change in one or more circuitry components of the switching device to reduce or block electrical current flowing through the one or more circuitry components in response to the test statistic identifying an onset of a fault.

10 . A method, comprising:

generating a feature set comprising qualified switching parameters to predict faults of switching devices, wherein generating the feature set comprises:

iteratively ranking a level of effectiveness of each candidate switching parameter within a candidate set of candidate switching parameters; and

iteratively removing one or more candidate switching parameters within the candidate set based on the ranking, until a number of retained candidate switching parameters is consistent with a determined size of the feature set;

based on the feature set, predicting a fault of a particular switching device, wherein predicting the fault comprises:

predicting, by a machine learning based model, a value of a switching parameter of the particular switching device;

based on a difference between an actual value of the switching parameter and the predicted value of the switching parameter, generating a test statistic; and

in response to the test statistic identifying an onset of a fault, generating an anomaly signal related to the presence of a fault.

11 . The method of claim 10 , wherein the switching parameter comprises an on-state current I ds for the particular switching device.

12 . The method of claim 11 , wherein the value for the on-state current I ds is predicted using values of at least one of body-diode voltage V bd , voltage between drain and source V ds , and squared voltage between gate and source V 2 gs .

13 . The method of claim 10 , wherein the difference is part of a time-series of residuals, and wherein the test statistic is generated by performing a cumulative sum (CUSUM) test on a mean of the residuals in the time-series.

14 . The method of claim 10 , wherein the switching parameter comprises an on-state current I ds for the switching device, the machine learning model comprises a first machine learning model, the difference comprises a first difference, the test statistic comprises a first test statistic, the method further comprising:

predicting, using a second machine learning model, a value of a drain-source resistance R ds of the switching device, wherein the prediction for the drain-source resistance is based on a present value of another switching parameter of the switching device and a prior value of drain-source resistance R ds ;

determining a second difference between the predicted value of R ds and an actual value of drain-source resistance R ds ;

generating a second test statistic based on the second difference; and

determining the fault in the particular switching device based on the first test statistic and the second test statistic.

15 . The method of claim 10 , wherein the switching parameter comprises an on-state current I ds for the switching device, the method further comprising:

predicting, using a second machine learning model, a value of a drain-source resistance R ds of the particular switching device, wherein the prediction for the drain-source resistance is based on a present value of another switching parameter of the particular switching device and a prior value of R ds ;

determining a second difference between the predicted value of R ds and an actual value of R ds ;

generating a second test statistic based on the second difference; and

updating the first machine learning model based on a comparison of the second test statistic to a threshold value.

16 . The method of claim 10 , wherein the predicted value for the another switching parameter of the switching device is determined based on on-state values for the one or more switching parameters.

17 . The method of claim 10 , further comprising:

determining the number of switching parameters in the first set of switching parameters;

evaluating first subset levels of effectiveness corresponding to first subsets of the candidate switching parameters, the first subsets of the candidate switching parameters each having a first subset size less than the determined size of the feature set,

deleting a candidate switching parameter resulting in a lowest decrease in a subset level of effectiveness in each of the first subsets, wherein deleting causes generating of second subsets, each second subset having a second subset size that is one less than the first subset size;

evaluating the effectiveness of subsets of switching parameters from the first set of switching parameters, the subsets of switching parameters having a subset size equal to the number of switching parameters in the first set of switching parameters less (i+1); and

selecting the one or more switching parameters for the switching device to be used by the first machine learning model based on the effectiveness of the subsets of switching parameters from the first set of switching parameters.

18 . The method of claim 10 , wherein the one or more switching parameters comprise body-diode voltage V bd , voltage between drain and source V ds , and squared voltage between gate and source V 2 gs ; and wherein the switching parameter is an on-state current I ds for the particular switching device.

19 . A method for detecting faults in power inverters of electric or hybrid vehicles, comprising:

generating a feature set comprising qualified switching parameters to predict faults of switching devices, wherein generating the feature set comprises:

iteratively ranking a level of effectiveness of each candidate switching parameter within a candidate set of candidate switching parameters; and

iteratively removing one or more candidate switching parameters within the candidate set based on the ranking, until a number of retained candidate switching parameters is consistent with a determined size of the feature set;

based on the feature set, predicting a fault of a particular switching device, wherein predicting the fault comprises:

predicting, by a partial least squares model a value for an on-state current I ds of the particular switching device; wherein the predicted value for the on-state current I ds is based on values of the qualified switching parameters for the particular switching device;

based on a difference between an actual value of the switching parameter and the predicted value of the switching parameter, generating a test statistic; and

in response to the test statistic identifying an onset of a fault, generating an anomaly signal related to the presence of a fault.

20 . The method of claim 19 , wherein the first set of switching parameters comprises body-diode voltage V bd , voltage between drain and source V ds , voltage between gate and source V gs , V gs *V ds , and squared voltage between gate and source V 2 gs comprising:

evaluating mean square errors (MSE) of models based on different subsets of switching parameters from the feature set; and

selecting a feature set of size i, where i is less than the number of features in the feature set, that yields a MSE that is closest in value to the MSE for a subset.