How do we identify potential troubles and address them before they manifest into real troubles?
Cable has made incredible progress in network and service reliability/availability. From vast improvements of the early days to the introduction of lifeline services like voice, we have reduced customer reported troubles from 40 percent to under 3 percent in an amazingly short time. Network troubles have followed a similar trajectory. But where do we go from here? Operators continue to improve as they deploy new tools and techniques for process improvement. But there’s more work to be done. Heightened competition gives customers many new choices and higher expectations than ever. The industry has to respond in aggressive and innovative ways.
One such response has been to use CableLabs’ Proactive Network Maintenance Reference Implementation. Those that have deployed this pre-equalization, impulse coefficient tool (see SCTE’s tutorial on pre-equalization coefficients at www.scte.org/ standards/netops) have been able to remotely identify RF impedance mismatches to under three feet. Using this tool over time, allows operators to become more and more proactive.
Another tool being introduced in various forms is one that takes advantage of full bandwidth spectrum analysis, enabling the capture of RF performance across the full frequency range.
This software-enabled tool takes advantage of the inherent capabilities of chips inside cable modems, set-top boxes, eMTAs, E-DTAs and gateways. Being able to pull spectral data from anywhere, at any time, and at any resolution is incredibly powerful for identifying potential customer-impacting network issues and resolving them. Correlating this data with other data sets and using it in places like MDUs provides a very powerful problem solver indeed.
As good and valuable as these tools are, they are still reactive. The question is how do we identify potential troubles and address them before they manifest into real problems, much less impact a customer? To start, we can take lessons from other industries and apply advanced techniques available today for predictive modeling. One business that we can learn from is the cement industry where significant issues were seen with an 1800 rpm mixer drum motor and cement screw. Downtime was being experienced from dry ball bearings, loose belts and shaft misalignments, all causing expensive replacement of parts and equally costly preventative maintenance procedures.
The truck owners performed vibration analyses and counted the vibrations per second which they used to set thresholds which when exceeded signaled a problem was beginning to happen. It was a useful predictive indicator, but it did not isolate the source of the vibration.
The source was found by performing a FFT on the data and observing the amplitude and frequency response, which correlated to the specific type of equipment that was beginning to have issues. This was accomplished by deploying accelerometer sensors on the trucks and using a handheld analyzer. Maintenance costs were dropped 60 percent, and nearly 98 percent of breakdowns were eliminated (more on this analysis is on our web site).
Note that this was accomplished in 1979! Today, we have big data analytics and far more advanced signal processing techniques.
With the continued deployment of sensor technology, an almost unlimited amount of previously unrelated and unavailable data is now available to us. By leveraging this plethora of information we will allow operations professionals unprecedented visibility and centralized control of their networks.
There are three distinct data mining applications cable can take advantage of.
Classification is used to identify a risk of failure by finding variables that are strongly related to a variable of interest and helps develop predictive models. Regression looks for similar variable correlation to aid in foreseeing problems in stochastic processes. Clustering is where a variable of interest is undetermined and attempts are made to compare, assign, measure and identify new, informative variable clusters of data.
Using data mining algorithms also identifies new data sets that predict future values of statistical univariate or multivariate series that are based on historical and what seems like unrelated observations. Additionally, spectral analysis of this data allows for the identification of unseen periodicity properties of the system being analyzed.
Finally, applying advanced correlation techniques and algorithms can vastly improve the relationship between seemingly unrelated sets of data which then provide better precision in truly predictive analyses.
These techniques yield incredibly valuable information that can be used to improve the actual measurement methods themselves, thus improving the quality of the data being gathered and subsequently the predictive capabilities altogether. Furthermore, correlation can be performed in multiple dimensions and even be further enhanced by using various correlation filters or estimators.
Employing these techniques, we will be able to solidify our place atop the network operations world by identifying and correcting customer impacting events before they even happen.