Technology in trucking is rapidly transforming every day. Each year, the number of hardware devices in trucks and the software systems that track them continue to grow. As the amount of technology on the road and inside the back office expands, new solutions are being put in place to bridge the gap between raw data and impactful insights. The latest revolution in fleet safety is Artificial Intelligence (AI) and Machine Learning (ML). So what does this mean for fleets?
Artificial Intelligence (AI) Terms
AI is an umbrella term for a computer software’s ability to perceive its environment, attempt to mimic human behavior and learn to accomplish a given goal. The idea of machines capable of thinking and learning for themselves has been a popular notion amongst computer scientists since the 1950s, but the way that it has impacted the world has been more subtle.
Because AI is such a broad term, many companies like to oversell their solution’s AI capability, when in reality their technology is little more than a particularly complicated algorithm. Sometimes, methods of AI are often no more than traditional systems that use simple equations, if-then rules, and decision trees.
At its best, however, AI can be a highly technical and powerful tool for analyzing and gaining insights from vast amounts of data. The ways in which AI is being applied are abundant, but ML, the foundation of advanced AI, is profoundly impacting the way that fleet professionals perform their day-to-day operations.
In the last decade, enormous progress in AI has evolved into a more narrowly defined subfield called Machine Learning (ML). ML is a type of analysis that machines with access to vast amounts of data conduct which enables them to learn and identify optimal outcomes, without human help. While the standard approach to AI focuses on hard-coded algorithms or rule-based systems to mimic human behavior, ML provides a way to move beyond these methods by finding patterns that are difficult for humans to identify and then using those patterns to predict events that haven’t yet occurred.
The real magic behind this technology is the ‘training’ process, where the ML models improve their predictions over time. In the case of transportation safety, ML takes data on driver attributes, behaviors, and safety events and runs hundreds of thousands of simulations, tweaking the potential impact of each data type in every simulation to find the relative importance for each data point or feature. In the end, the finely tuned and highly complex model can accurately identify at-risk drivers, drivers likely to leave a fleet, and much more. Trying to obtain this level of analysis and accuracy without the use of advanced computing is impossible, which speaks to the real power of ML.
Developing a robust ML algorithm is no simple task for data scientists without a reference point from a previous model. Meanwhile, safety managers handle many forms of data and require a variety of reports to do their job effectively, which in turn means that numerous different models and processes would be needed to reach an ideal outcome.
Transfer learning is a method used by data scientists to utilize previous sophisticated models as a starting point for new models. Doing so jump-starts the development process on a new task or problem, which allows ML algorithms to grow exponentially.
In trucking, fleets can use data from various systems integrated over time to see what has happened in the past, identify patterns, and predict how and when the next preventable accident will occur. As you continue to add more data, systems, and features to your models, the potential for more accurate predictions increases. However, this will also increase the complexity of the problem and can quickly outpace the abilities of traditional ML. These cases require more robust and sophisticated techniques like deep learning.
One modern deep learning architecture, recurrent neural networks (RNNs), allows deep learning models to directly evaluate a driver’s history of accidents and other events without first requiring data scientists to choose a single way of summarizing that data, such as natural language processing or feature extraction. Another advantage of deep learning is the ability to learn mathematical representations of customer-specific terminology automatically, which are called embeddings.
Predict Accidents with Machine Learning
ML algorithms predict at-risk drivers by using integrated safety data stored within a fleet’s AI-powered Driver Management Platform. While AI, ML, and other predictive analytics may seem confusing, integration and automation can simplify a safety manager’s workload.
Identifying Metrics that Cause Accidents
Two historical data sets are required to solve a prediction problem with ML:
- Input data that will be available in the future (past accidents, incidents, and driver behavioral data)
- Output data centered around desired predictions (potential accidents or at-risk drivers)
Using these leading indicators, ML can absorb massive amounts of data to build models around patterns of behaviors and predict potential outcomes. Systems with ML can automatically learn patterns without being explicitly programmed and can improve their predictions with experience.
The biggest issue with gathering meaningful insights from data across multiple disparate systems is that the different formats of information are incomprehensible to the model. Integrating safety data into one place, like the Idelic Safety Suite®, provides clean and consistent data for natural language processing to occur, which makes it possible for ML to measure behaviors and predict accidents.
By utilizing natural language processing, systems can better enable safety teams to take advantage of ML no matter which format or data points are most important. Whether a fleet has very few data points or is experiencing a data overload across disparate systems, ML models can be tailored to the specific needs of each fleet.
Speaking specifically to ML, metrics that fleets collect that prove the most useful to data scientists include:
Algorithms Based on Past Data
The task of predicting at-risk drivers begins with pre-accident data. The model forecasts which drivers it believes are most susceptible to an accident and compares the predictions to actual outcomes which depict accuracy. Then, the model adjusts the inputs and recomputes. If the accuracy improved, the model knows it’s on the right track. If it deviates, the model can begin to understand why. The formula is slightly tweaked again and then runs for additional cycles. The model goes through this process hundreds of thousands of times, ‘learning’ as it makes adjustments until it identifies the most significant events causing accidents.
Safety managers can adjust these predictions based on extenuating circumstances. For example, if a driver has an incident that is deemed to be significant but it’s determined that the driver was not at fault, safety managers can reclassify that safety event, and the model will adjust accordingly. If the driver could have responded differently but still made their best effort to rectify the situation, changing the classification of the feature will reflect the weighing of that data point in the model.
Additionally, the model is continuously rerunning and optimizing as a fleet gathers new data. Driver data is then input into this model in real-time and flagged before an accident occurs. Addressing issues and administering training proactively allows fleets to get at-risk drivers back on the road safely through a regularly updated and available Driver Watch List.
Driver Watch List
Once integrated data runs through an ML model, the Driver Management Platform determines which drivers are at risk for an accident and assigns them a score. For example, Safety Suite’s AI-generated Driver Watch List is simple, intuitive, and includes details on a driver’s safety events, risk level, and breakdown of their score. These processes occur behind the scenes so that safety teams can leave the technical jargon aside and easily make sense of the data.
ML processes enable safety managers not only to pinpoint who their at-risk drivers are but also which safety events are causing a sub-par Risk Score. With proprietary ML techniques, events include the percentage that they contribute to a score. This allows safety teams to quickly understand which behaviors are the riskiest for an individual driver and helps them assign the best course of action for remedial training.
Benefits of Using Systems with Machine Learning
The safest fleets are realizing the potential of ML and are quickly seeing the benefits. Before evaluating solutions, safety teams must first identify the underlying issues and challenges that are being faced.
Problems with Non-ML Solutions for Safety, Operations, and Risk Management
With so much data across a variety of technology systems, fleet professionals often rely on time-consuming methods of organizing information to gain insight into their culture and practices. These methods include:
CSA scores were introduced in 2010 and are a victim of their time: Brought into being when the industry had recognized the value of data and performance scoring, but before the advent of AI and ML technologies. As a result, CSA scores can give fleets some idea of their performance in the seven core BASICs but are a static measure that must undergo an intense review process for changes to be made to the system. Though CSA scores serve as a useful benchmark for the industry as a whole, fleets that rely on them to make performance decisions about their practices are likely to fall behind fleets that use the latest and greatest technology.
Relying on disparate, siloed data to track driver performance raises several issues. Collating data from unconnected sources greatly increases the reporting process, while also increasing the likelihood of human error in manual reporting. These tasks get much harder when managing every driver operating within a fleet.
Each year, new third-party systems enter the trucking technology market, exponentially increasing the amount of data generated by each fleet. The issue is no longer having enough driver data available—the problem now is that there is too much data for safety, operations, and risk managers to digest.
Fleet professionals are only human, and people have a tendency to search for information that reinforces their preexisting beliefs. This mentality is known as confirmation bias and can lead individuals to find a false correlation between incidents which can lead to poor overall conclusions. Without realizing it, fleet operators may pick out data that they think is relevant, while in actuality, there are far more significant and unforeseen issues hiding in other data points. Finding meaningful insight is nearly impossible without using a comprehensive Driver Management Platform with ML capability.
A Day in the Life of a Fleet Professional with ML
Fleets taking advantage of AI and ML aren’t just preventing accidents, they are also significantly reducing driver turnover and insurance costs while saving valuable time and resources in the process. When combined with a centralized data management software, ML analytics tools can gather information faster, make better predictions, and streamline fleet safety operations. With an ML-based platform, every department can unlock key metrics:
Safety – Proactive Accident Prevention
A practical and compelling example of ML in trucking is predicting at-risk drivers and potential accidents. After consolidating data from disparate systems into a single platform, ML models can analyze millions of points of driver data and predict which drivers are most at risk of a crash. This gives safety managers the opportunity to proactively intervene with training for the drivers who need it most. The most robust ML models can even tell managers which skills their drivers most need training on, saving time, money, and lives for all involved.
Operations – Reduce Turnover Through Retention by Prevention
With high turnover and a driver shortage plaguing the trucking industry, it’s increasingly difficult for operations managers to retain experienced and reliable drivers. With systems using ML, fleets are able to retain drivers with more ease than ever before. Proactive intervention for at-risk drivers can help fleets prevent an accident that would typically result in termination. Even better, ML models that predict which drivers are at risk of leaving their fleet voluntarily are under production now, with promising applications for the future.
Risk Management – Reduce & Forecast Liability
With an ML system capable of predicting a driver’s accident risk, risk management officers at a fleet are capable of making more accurate forecasts of expected liability. The ability to say, with confidence, that an identified percent of drivers will experience a crash in the next quarter is an extremely valuable report. This is also information that insurance carriers find incredibly useful and fleets can use this information to gain a fair rate at renewals.
Additionally, an increasingly cited form of negligence in crash-related trials is “failure to equip fleet with the latest technology.” By outfitting their operation with the best available ML technology, fleets can insulate themselves from this type of negligence.
Compliance – Get Ahead of CSA Scores
When a fleet’s BASICs percentiles fall into non-compliance, they risk direct intervention and even a potential halt in operations. ML’s ability to predict accident risk can help compliance managers focus their attention on BASICs percentiles that can’t be addressed with training, such as maintenance, HOS, or driver fitness. By taking a major BASICs percentile off the plate of a fleet’s compliance team, ML can raise the effectiveness and focus of a fleet’s most important liability reduction department.