All forms of analytics are based on the idea that "history repeats itself." By examining the historical record of business activities, future business requirements can be better met and risks can be averted.
Fleet operators already gather fleet performance information on drivers and equipment, much of which is transmitted back to the company from onboard telematics systems. Those systems include GPS or other automated vehicle location (AVL) technology, electronic on-board recorders (EOBR), as well as hand-held computers and smart phones. The type of onboard vehicle technology used differs from fleet to fleet, depending upon its purpose. For example, home food delivery fleets use hand-held computers and collect information in "batches" at the end of the route day. Commercial, for-hire truckload carriers may use wide-area mobile communications (mobile-com) systems, communicating with drivers and collecting data in "real-time" to meet critical customer service requirements, such as ensuring just-in-time deliveries.
Though different types of fleets will use different technology systems, the issue of synergizing diverse data sets to focus on key problem areas is the same for each system type, as the table below demonstrates. All data from diverse sets can be gathered and studied in relation to key problem areas. In the case of fleet risk management, those areas could be, but are not limited to, fleet performance, accident and loss, and driver turnover.
Regardless of the type of fleet, these collected fleet data can be transformed into knowledge-driven, risk management business intelligence using database mining, pattern recognition and predictive modeling technologies to recognize when history is likely to be repeated and to guard against future losses. Each of these types of analytic techniques produces actionable business intelligence that can be incorporated into business operations to improve business performance. Database mining and pattern recognition both use historical data to define specific business procedures, whereas predictive modeling provides businesses the ability to address more complex and subtle business challenges.
Telematics systems data in database mining and pattern recognition
Database mining is used to correlate related business events. Beverage distributors, for example, are very good at forecasting beverage sales with minimal shelf waste by using historical sales analysis, which includes seasonal preferences, current weather conditions and special circumstances. One such analysis in Texas identified that soft drink sales from a particular convenience store would increase by 10% each time average temperatures rose more than 10 degrees above 70 degrees. In this example, on weekends when the temperature averaged 100 degrees, soft drink sales were, on average, 30% higher than when the temperature averaged 70 degrees.
This analysis, however, couldn't explain sudden surges in demand that occurred in limited regions of the city until the GIS information associated with these stores was correlated with athletic event schedules at a nearby stadium. The stadium events caused an additional 40% increase in adjacent stores' sales. Associating these consumer events through database mining allowed the beverage company to meet its customers' needs with minimal product waste.
Pattern recognition is another form of analysis which relies on identifying a number of data objects that, in certain combinations, correlate to a specific characteristic. Pattern recognition can be used to analyze repeated work experiences and improve trucking fleet performance. For example, critical event recorders use geographic locations and time of day derived from on-board GPS systems to identify driving anomalies like hard braking and rapid lane changes. Using mobile communication systems to collect the history of a fleet driving the same route creates a database record of fleet-critical events. A fleet safety manager can use simple pattern recognition to determine if similar incidents occur in the same locations and if a particular driver incident is the result of bad luck or if the driver needs training. This information can also be used to determine whether changing route schedules will reduce the number of incidents. These improvements can be automated through mobile communication dispatch systems, which change load assignment schedules or include driver alert bulletins in the route assignments.
Pulling it all together: Using predictive modeling to forecast the future
Pattern recognition is useful when businesses operate in fixed patterns, but what if the business changes are very dynamic? For instance, how does a fleet deal with a need such as identifying which drivers are likely to have accidents or which drivers are most likely to quit next month? The availability of extensive GIS data now provides fleet managers the ability to address complex, subtle and dynamic issues through the use of predictive modeling analytics.
Unlike database mining and pattern recognition, which use historical data analysis to develop a static model of business performance, predictive modeling uses historical data to create a new type of business tool that accurately forecasts the likelihood of a future event. It does so by correlating diverse data objects that include historical and real-time data gathered through telematics systems, alongside other intuitive and non-intuitive data, such as a driver's age or commute time to work. The resulting business intelligence can be used to intervene against predicted losses and optimize fleet efficiency with amazing accuracy.
Performing effective analytics of any type requires that critical data needed to make associations in the model reside in a single relational database, instead of in isolated business systems throughout the organization, as is often the case. This is particularly true for predictive modeling, which uses arrays of many data objects - also known as predictive factors or contributing factors - in the completed model. The process for creating a good predictive model involves placing all relevant business data in a relational database and then applying sets of algorithms that find the data relationships that correspond to specific business events, like driver accidents or driver resignations, with high statistical certainty.
The predictive model is then fed ongoing business operating information (as opposed to the historic data used to design the model) and the model becomes a type of "analytic machine" which identifies similar patterns in employee behavior prior to the undesired event. For example, a rise in random driver "critical driving events" - incidents such as hard braking or rapid lane changes - have proven to be an indicator that the driver may soon have an accident. However, to make an accurate prediction, the critical event would have to be taken into consideration alongside 15-to-20 other data objects, such as the driver's prior work experience, work schedule, or fatigue level, to name a few. Once the model indicates an at-risk driver, an alert gives employers the opportunity to intervene (with countermeasures created by examining the predictive or contributive factors), and prevent many of the undesired incidents from occurring.
A simple example of a proven predictive model is a fatigue model. Sleep science has determined that all people have a cyclical "natural" work/sleep pattern. An individual's alertness and ability to work at high levels of performance are negatively impacted when the individual works too long or doesn't get enough rest. Over a span of several days of too much work and not enough sleep, a person's ability to perform effectively is diminished. The availability of electronic driver logs aboard vehicles allows this analysis to be conducted on individual drivers and their fatigue levels can be accurately measured and scored. When a fatigue level becomes too extreme, the driver's probability of having critical events, or even an accident, rises to dangerous levels. A dispatcher with access to this real-time information will realize that a driver needs time off and can delay or divert load assignments from that driver until the driver is better rested. Fleets can then also target at-risk drivers with personalized training programs that teach drivers to better deal with driving situations in which they experience problems or to better manage their work/sleep schedules.
Predictive models can also directly address the challenges of accident prevention and driver retention. Driver safety modeling, for example, involves identifying at-risk drivers and understanding the core factors that are creating the risk situation. Often drivers don't realize they are at risk. Family issues, undesirable work assignments or misunderstandings with dispatchers can all contribute to rising levels of stress, which manifest themselves in reduced driver performance. The ability to use telematics systems, with GIS data, provides companies with insight into how their distributed workers are performing.
By scoring driver performance using measures such as on time deliveries (using customer proof-of-delivery on hand-held computers or mobile-com systems), route compliance (using AVL tracking), starting work promptly, minimizing critical events, and good fuel management (using EOBR systems), drivers can be ranked in terms of their performance, in both absolute and relative terms. Dispatchers, safety directors and operations managers can focus on improving the performance of lower ranked drivers and supporting high ranking drivers whose performance begins to slip. This system of reinforcing positive behaviors and correcting negative behaviors leads to improvement in the overall driver population. The ability to target specific drivers for corrective action allows management to focus resources where they provide the most benefit. If a company has thousands of drivers, knowing which hundred need attention is of enormous benefit.
Predicting what comes next
Analysts are just beginning to uncover the potential benefits of predictive modeling, as more varied and complex problems are explored and more effective countermeasures are created. Today predictive models that address issues of driver fatigue, driver performance and driver retention are either in operation or under development, and the results are nothing short of amazing. In one fleet, accident rates were reduced by over 40% in less than six months. In the near future, models that focus on fuel usage, fleet maintenance and fleet optimization (matching driver/vehicle teams to load challenges) seem equally feasible.
The limiting factor in creating all analytic systems is the quality and quantity of data to build the analytic models. The rapid proliferation of mobile data and telematics systems and services is making available extensive, high quality data that lend themselves to the analysis of ever-more-complex and subtle business issues. The challenge of using these data effectively, without disrupting successful business operations, remains an impediment to adoption. But more and more companies are recognizing that a program of integrated analytics, targeted improvement programs and timely feedback of performance measurement is allowing fleet performance to improve at a rate and to an extent that they would have thought impossible just a few years ago.