Wireless communication network voice quality monitoring
In voice communication quality monitoring, both user plane and control plane signaling are gathered during network operation, and correlated. Offline (that is, not in real time), a predictive machine learning model is trained using the signaling data. The model is subsequently used to monitor network operation in real time. The model label is instances of voice quality degradation gleaned from probing the user plane media. The are control plane traffic patterns correlated to the voice quality degradation incidents. After training, when monitoring voice quality in real time on the network, only the control plane signaling is monitored. The machine learning model recognizes learned control plane signaling patterns and infers corresponding user plane voice quality degradation incidents. Settings of the model are controlled to achieve a desired precision/recall tradeoff. Because only control plane signaling is monitored in real time, the approach can be applied across all voice communications in a network (or portion of a network). The machine learning model is re-trained as necessary to reflect changes in the network.
1 . A method of monitoring voice quality in a wireless communication network implemented with separate user plane and control plane architectures, comprising, in real time as voice communications are transmitted through the network:
monitoring control plane signaling for all voice communications over at least a first portion of the wireless communication network;
extracting features from the control plane signaling; and
inferring voice quality degradation incidents in the user plane media by applying the features to a previously trained machine learning model that predicts voice quality degradation incidents based on the control plane signaling features.
2 . The method of claim 1 wherein the machine learning model was trained using a predetermined sensitivity setting that adjusts a precision/recall trade-off between a number of voice quality degradation incidents correctly inferred and a number of false positive inferences of voice quality degradation incidents.
3 . The method of claim 1 further comprising adjusting one or more network parameters in response to the number of inferred voice quality degradation incidents.
4 . The method of claim 1 further comprising, prior to the monitoring, extracting, or inferring steps, and in an offline process that does not occur in real time, training a machine learning model by:
obtaining user plane and control plane signaling for a plurality of voice communications over the first portion of the wireless network;
for each voice communication:
temporally dividing the voice communication into a plurality of windows;
analyzing a Real-time Transport Protocol media flow of the user plane media to determine voice quality degradation incidents; and
tagging each window that contains a voice quality degradation incident; and
training a predictive machine learning model wherein, for each window of each voice communication, a label is a tag indicating whether the window contains a voice quality degradation incident, and features are control plane signaling events time-correlated to the window;
whereby the trained machine learning model predicts user plane voice quality degradation incidents based on control plane signaling event features.
5 . The method of claim 4 wherein temporally dividing the voice communication into a plurality of windows comprises:
temporally dividing the voice communication into a plurality of bins; and
wherein each window comprises a predetermined, odd number of contiguous bins.
6 . The method of claim 5 wherein the windows overlap such that each bin is the center bin of a unique window, and further comprising padding, as necessary, short voice communications and the beginning and end of each voice communication with bins indicating no voice quality degradation incident.
7 . The method of claim 5 wherein tagging each window that contains a voice quality degradation incident comprises tagging only the windows that contain a voice quality degradation incident in the center bin of the window.
8 . The method of claim 5 wherein tagging each window that contains a voice quality degradation incident comprises tagging the windows that contain a voice quality degradation incident in any of a predetermined number of center-most bins of the window.
9 . The method of claim 4 wherein control plane signaling events corresponding to the window comprise packet core control plane signaling events time-correlated to the window.
10 . The method of claim 4 wherein control plane signaling events corresponding to the window comprise cell trace record events time-correlated to the window.
11 . A computational resource communicatively coupled to a wireless communication network implemented with separate user plane and control plane architectures, and operative to monitor voice communications in the wireless communication network, comprising:
communication circuitry; and
processing circuitry operatively coupled to the communication circuitry and configured to, in real time as voice communications are transmitted through the network:
monitor control plane signaling for all voice communications over at least a first portion of the wireless communication network;
extract features from the control plane signaling; and
infer voice quality degradation incidents in the user plane media by applying the features to a previously trained machine learning model that predicts voice quality degradation incidents based on the control plane signaling features.
12 . The resource of claim 11 wherein the machine learning model was trained using a predetermined sensitivity setting that adjusts a precision/recall trade-off between a number of voice quality degradation incidents correctly inferred and a number of false positive inferences of voice quality degradation incidents.
13 . The resource of claim 11 wherein the processing circuitry is further configured to adjust one or more network parameters in response to the number of inferred voice quality degradation incidents.
14 . The resource of claim 11 wherein the processing circuitry is further configured to, prior to monitoring, extracting, or inferring, and in an offline process that does not occur in real time, train a machine learning model by:
obtaining user plane and control plane signaling for a plurality of voice communications over the first portion of the wireless network;
for each voice communication:
temporally dividing the voice communication into a plurality of windows;
analyzing a Real-time Transport Protocol media flow of the user plane media to determine voice quality degradation incidents; and
tagging each window that contains a voice quality degradation incident; and
training a predictive machine learning model wherein, for each window of each voice communication, a label is a tag indicating whether the window contains a voice quality degradation incident, and features are control plane signaling events time-correlated to the window;
whereby the trained machine learning model predicts user plane voice quality degradation incidents based on control plane signaling event features.
15 . The resource of claim 14 wherein the processing circuitry is configured to temporally divide the voice communication into a plurality of windows by:
temporally dividing the voice communication into a plurality of bins; and
wherein each window comprises a predetermined, odd number of contiguous bins.
16 . The resource of claim 15 wherein the windows overlap such that each bin is the center bin of a unique window, and wherein the processing circuitry is further configured to pad, as necessary, short voice communications and the beginning and end of each voice communication with bins indicating no voice quality degradation incident.
17 . The resource of claim 15 wherein the processing circuitry is configured to tag each window that contains a voice quality degradation incident by tagging only the windows that contain a voice quality degradation incident in the center bin of the window.
18 . The resource of claim 15 wherein the processing circuitry is configured to tag each window that contains a voice quality degradation incident by tagging the windows that contain a voice quality degradation incident in any of a predetermined number of center-most bins of the window.
19 . The resource of claim 14 wherein control plane signaling events time-correlated to the window comprise packet core control plane signaling events time-correlated to the window.
20 . The resource of claim 14 wherein control plane signaling events time-correlated to the window comprise cell trace record events time-correlated to the window.