Methods and apparatuses for detecting and localizing faults using machine learning models
A method of pre-processing data for use in training one or more machine learning, ML, models for use in detecting anomalies occurring during execution of one or more procedures at a plurality of network nodes in a network, includes: obtaining procedure level time series data relating to the execution of a first procedure at the plurality of network nodes; and deriving, from the procedure level time series data relating to the execution of the first procedure at the plurality of network nodes, one or more first feature time series of one or more respective first feature values, wherein the one or more first feature time series are for use in training a first ML model associated with the first procedure.
1 . A computer implemented method of pre-processing data implemented using a processing circuitry for use in training one or more machine learning, ML, models for use in detecting anomalies occurring during execution of one or more procedures at a plurality of network nodes in a network, the method comprising:
obtaining procedure level time series data relating to the execution of a first procedure at the plurality of network nodes;
deriving, from the procedure level time series data relating to the execution of the first procedure at the plurality of network nodes, one or more first feature time series of one or more respective first feature values,
wherein the one or more first feature time series are for use in training a first ML model associated with the first procedure; and the one or more first feature time series comprises at least one time series of a count value increase, and wherein the step of deriving one or more first feature time series comprises:
for each event in a graph of hierarchical events, wherein the graph of hierarchical events represents the first procedure:
determining, from the procedure level time series data relating to the execution of the first procedure at the plurality of network nodes, a time series of a count value increase for the event, wherein the count value increase comprises a difference in consecutive counts of occurrences of the event.
2 . The method as claimed in claim 1 further comprising:
obtaining procedure level time series data relating to the execution of a second procedure at the plurality of network nodes; and
deriving, from the procedure level time series data relating to the execution of a second procedure at the plurality of network nodes, one or more second feature time series of one or more respective second feature values, wherein the one or more second feature time series are for use in training a second ML model associated with the second procedure.
3 . The method as claimed in claim 1 wherein the one or more first feature times series comprises at least one time series of an edge value, and wherein the step of deriving the one or more first feature time series further comprises:
for each edge in the graph of hierarchical events, determining a time series of an edge value from a first time series of a count value for a parent event associated with the edge;
and a second time series of a count value for a child event associated with the edge.
4 . The method as claimed in claim 1 wherein the step of deriving the one or more first feature series is based on procedure level time series data relating to the execution of the first procedure at one or more network nodes in a first network node group.
5 . The method as claimed in claim 4 further comprising grouping the plurality of network nodes into a plurality of groups comprising the first network node group based on one or more characteristics associated with the network nodes.
6 . The method as claimed in claim 5 wherein the one or more characteristics comprises one or more of: cell parameters associated with each network node, traffic levels at each network node, technology used at each network node and software version.
7 . The method as claimed in claim 1 wherein the procedure level time series data comprises internal log data logged at each of the plurality of network nodes.
8 . The method as claimed in claim 1 wherein the one or more procedures comprise procedures standardized according to a 3GPP standard.
9 . The method as claimed in claim 1 further comprising pre-processing the procedure level time series data to remove duplicated events.
10 . The method as claimed in claim 1 , further comprising performing the training of the one or more ML models including by
utilising the one or more first feature time series as training data for a first ML model.
11 . The method as claimed in claim 10 further comprising:
obtaining procedure level time series data relating to the execution of a second procedure at the plurality of network nodes;
deriving, from the procedure level time series data relating to the execution of a second procedure at the plurality of network nodes, one or more second feature time series of one or more respective second feature values, wherein the one or more second feature time series are for use in training a second ML model associated with the second procedure; and
utilising the one or more second feature time series as training data for a second ML model.
12 . The method as claimed in claim 1 wherein the first ML model comprises a Long Short-Term Memory (LSTM) autoencoder model.
13 . The method as claimed in claim 12 wherein the first ML model is trained using Encoder-Decoder LSTM architecture.
14 . The method as claimed in claim 1 , further comprising using the one or more ML models for detecting anomalies occurring during execution of one or more procedures at the plurality of network nodes in the network, including by
utilising the first feature time series as input values for a first ML model.
15 . An apparatus for use in training one or more machine learning, ML, models for use in detecting anomalies occurring during execution of one or more procedures at a plurality of network nodes in a network, the apparatus comprising processing circuitry configured to cause the apparatus to:
obtain procedure level time series data relating to the execution of a first procedure at the plurality of network nodes;
derive, from the procedure level time series data relating to the execution of the first procedure at the plurality of network nodes, one or more first feature time series of one or more respective first feature values, wherein the one or more first feature time series are for use in training a first ML model associated with the first procedure; and the one or more first feature time series comprises at least one time series of a count value increase, and wherein the step of deriving one or more first feature time series comprises:
for each event in a graph of hierarchical events, wherein the graph of hierarchical events represents the first procedure:
determining, from the procedure level time series data relating to the execution of the first procedure at the plurality of network nodes, a time series of a count value increase for the event, wherein the count value increase comprises a difference in consecutive counts of occurrences of the event.
16 . The apparatus as claimed in claim 15 wherein the processing circuitry is further operative to:
obtain procedure level time series data relating to the execution of a second procedure at the plurality of network nodes; and
derive, from the procedure level time series data relating to the execution of a second procedure at the plurality of network nodes, one or more second feature time series of one or more respective second feature values, wherein the one or more second feature time series are for use in training a second ML model associated with the second procedure.
17 . The apparatus as claimed in claim 15 , wherein the processing circuitry is further operative for training the one or more machine learning, ML, models for use in detecting anomalies occurring during the execution of one or more procedures at the plurality of network nodes in the network, and is further operative to:
obtain one or more first feature times series from a pre-processing apparatus as claimed in claim 15 ; and
utilise the one or more first feature time series as training data for a first ML model.