Log-based predictive maintenance using multiple-instance learning
A method of building a model for predicting failure of a machine, including parsing ( 41 ) daily machine event logs of one or more machines to extract data for a plurality of features, parsing ( 42 ) service notifications for the one or more machine to extract failure information data, creating ( 43 ) bags from the daily machine event log data and failure information data for multiple instance learning by grouping daily event log data into the bags based on a predetermined predictive interval, labeling each bag with a with a known failure as positive, and bags without known failures as negative, where a bag is a set of feature vectors and an associated label, where each feature vector is an n-tuple of features, transforming ( 44 ) the multiple instance learning bags into a standard classification task form, selecting ( 45 ) a subset of features from the plurality of features, and training ( 46 ) a failure prediction model using the selected subset of features.
1. A computer-implemented method of building a model for predicting failure of a machine, the method comprising the steps of:
parsing daily machine event logs of one or more machines to extract data for a plurality of features, the features associated with predicting machine failure;
parsing service notifications for the one or more machine to extract failure information data;
creating bags from the daily machine event log data and failure information data for multiple instance learning by grouping daily event log data into the bags based on a predetermined predictive interval, assigning a positive label to each bag with a known failure, and assigning a negative label to each bag having no known failures, wherein each bag comprises a respective set of feature vectors and an associated label, and wherein each feature vector is an n-tuple of features;
transforming each feature vector in a bag with a negative label into a corresponding negative example, and for each bag with a positive label, creating a respective meta-positive example by computing a respective mean of the feature vectors in each bag with a positive label;
selecting a subset of features from the plurality of features, wherein the subset of features defines a function for a failure prediction model to predict a bag label, and the selecting comprises:
creating multiple candidate subsets of features, wherein each candidate subset of features comprises features from all bags assigned the positive label and a random subsampling of features from bags assigned the negative label,
learning a sparse linear classifier on each candidate subset of features, wherein learning the sparse linear classifier on each candidate subset of features comprises learning a respective set of weights for each candidate subset of features; and
averaging the respective set of weights learned for each candidate subset of features and selecting features with the highest absolute weights as the subset of features;
training the failure prediction model using the selected subset of features; and
using the failure prediction model in deployment to monitor new instances of daily machine event logs wherein a prediction failure alert is triggered in response to receiving a failure prediction score that exceeds a predefined threshold.
2. The method of claim 1 , wherein features include event codes, event code variations, and numerical values associated with the event codes and event code variations.
3. The method of claim 1 , wherein training the failure prediction model using the selected subset of features comprises learning the failure prediction model using the selected subset of features and all event log instances for the selected subset of features.
4. The method of claim 3 , wherein learning the failure prediction model comprises finding a set of weights w that minimizes
λ
2
w
1
2
+
∑
j
max
{
1
-
y
j
w
T
x
j
,
0
}
,
wherein λ>0 is a user-specified regularization parameter, y j ∈{+1, −1} is a label for bag j, and x j is a vector of the selected subset of features for bag j.
5. The method of claim 1 , further comprising:
evaluating a label prediction of a current bag by the trained failure prediction model using a maximal prediction score of all instances in the current bag.
6. A computer-implemented method of building a model for predicting failure of a machine, the method comprising the steps of:
parsing daily machine event logs of one or more machines to extract data for a plurality of features;
parsing service notifications for the one or more machine to extract failure information data;
creating bags from the daily machine event log data and failure information data for multiple instance learning, wherein each bag comprises a respective set of feature vectors and an associated label, wherein the associated label is a positive label or a negative label, and wherein each feature vector is an n-tuple of features;
transforming the multiple instance learning bags into a standard classification task form by transforming each feature vector in a bag with a negative label into a corresponding negative example, and for each bag with a positive label, creating a respective meta-positive example by computing a respective mean of the feature vectors in each bag with a positive label;
selecting a subset of features from the plurality of features, wherein the subset of features defines a function for a failure prediction model to predict a bag label, and the selecting comprises:
creating multiple candidate subsets of features, wherein each candidate subset of features comprises features from all bags assigned the positive label and a random subsampling of features from bags assigned the negative label,
learning a sparse linear classifier on each candidate subset of features, wherein learning the sparse linear classifier on each candidate subset of features comprises learning a respective set of weights for each candidate subset of features, and
averaging the respective set of weights learned for each candidate subset of features and selecting features with the highest absolute weights as the subset of features; and
training the failure prediction model using the selected subset of features; and
using the failure prediction model in deployment to monitor new instances of daily machine event logs wherein a prediction failure alert is triggered in response to receiving a failure prediction score that exceeds a predefined threshold.
7. The method of claim 6 , wherein creating bags from the daily machine event log data and failure information data for multiple instance learning comprises grouping daily event log data into the bags based on a predetermined predictive interval, labeling each bag with a known failure as positive, and labeling bags without known failures as negative.
8. The method of claim 6 , wherein training a failure prediction model using the selected subset of features comprises learning the failure prediction model using the selected subset of features and all event log instances for the selected subset of features by finding a set of weights w that minimizes
λ
2
w
1
2
+
∑
j
max
{
1
-
y
j
w
T
x
j
,
0
}
,
wherein λ>0 is a user-specified regularization parameter, y j ∈{+1, −1} is a label for bag j, and x j is a vector of the selected subset of features for bag j.
9. A non-transitory program storage device readable by a computer, tangibly embodying a program of instructions executed by the computer to perform the method steps for building a model for predicting failure of a machine, the method comprising the steps of:
parsing daily machine event logs of one or more machines to extract data for a plurality of features, the features associated with predicting machine failure;
parsing service notifications for the one or more machine to extract failure information data;
creating bags from the daily machine event log data and failure information data for multiple instance learning by grouping daily event log data into the bags based on a predetermined predictive interval, assigning a positive label to each bag with a known failure, and assigning a negative label to each bag having no known failures, wherein each bag comprises a respective set of feature vectors and an associated label, and wherein each feature vector is an n-tuple of features;
transforming each feature vector in a bag with a negative label into a corresponding negative example, and for each bag with a positive label, creating a respective meta-positive example by computing a respective mean of the feature vectors in each bag with a positive label;
selecting a subset of features from the plurality of features, wherein the subset of features defines a function for a failure prediction model to predict a bag label, and the selecting comprises,
creating multiple candidate subsets of features, wherein each candidate subset of features comprises features from all bags assigned the positive label and a random subsampling of features from bags assigned the negative label,
learning a sparse linear classifier on each candidate subset of features, wherein learning the sparse linear classifier on each candidate subset of features comprises learning a respective set of weights for each candidate subset of features, and
averaging the respective set of weights learned for each candidate subset of features and selecting features with the highest absolute weights as the subset of features;
training the failure prediction model using the selected subset of features; and
using the failure prediction model in deployment to monitor new instances of daily machine event logs wherein a prediction failure alert is triggered in response to receiving a failure prediction score that exceeds a predefined threshold.
10. The computer readable program storage device of claim 9 , wherein features include event codes, event code variations, and numerical values associated with the event codes and event code variations.
11. The computer readable program storage device of claim 9 , wherein training flail the failure prediction model using the selected subset of features comprises learning the failure prediction model using the selected subset of features and all event log instances for the selected subset of features.
12. The computer readable program storage device of claim 11 , wherein learning the failure prediction model comprises finding a set of weights w that minimizes
λ
2
w
1
2
+
∑
j
max
{
1
-
y
j
w
T
x
j
,
0
}
,
wherein λ>0 is a user-specified regularization parameter, y j ∈{+1, −1} is a label for bag j, and x j is a vector of the selected subset of features for bag j.
13. The computer readable program storage device of claim 9 , the method further comprising evaluating a label prediction of a current bag by the trained failure prediction model using a maximal prediction score of all instances in flail the current bag.