US20190263417A1 - Systems and methods for driver scoring with machine learning - Google Patents

Abstract

Systems and methods for using machine learning classifiers to identify anomalous driving behavior in vehicle driver data obtained from vehicle telematics devices are provided. In one example, a vehicle telematics device receives vehicle driver data from sensors, identifies anomalies in the vehicle driver data by using an unsupervised machine learning process, calculates a driver risk score by using the anomalies identified in the vehicle driver data, and transmits the risk score to a remote server system. In another example, a server system receives vehicle driver data from a plurality of vehicle telematics devices, identifies anomalies in the vehicle driver data by using an unsupervised machine learning process, and calculates a driver risk score by using the anomalies identified in the vehicle driver data.

Description

FIELD
• The embodiments relate to generating a risk score for vehicle drivers by using accumulated vehicle data and unsupervised machine learning classifiers.
BACKGROUND
• Telematics is the integrated use of telecommunications and informatics. Telematics units are installed in vehicles to provide a variety of telematics functionality in the vehicle. This functionality includes, but is not limited to, emergency warning systems, navigation functionality, safety warnings, and automated driving assistance. Telematics units are also capable of recording vehicle information/data related to the operation of the vehicle and providing that information for analysis, whether in real-time or not, such as during a time when the vehicle is being serviced. The vehicle information/data (telematics data) generated by a telematics unit can be used in a variety of applications, such as fleet tracking, shipment tracking, insurance calculations, and in vehicle management and service.
SUMMARY
• The embodiments are best summarized by the claims that follow below. Briefly, systems and methods are disclosed for generating driver risk scores by using adaptive models with machine learning classifiers.
BRIEF DESCRIPTION OF THE DRAWINGS
• FIG. 1 is a diagram conceptually illustrating vehicle telematics system that can be used to generate driver risk scores.
• FIG. 2A is a block diagram illustrating a vehicle telematics device that can be used to generate vehicle data for computing driver risk scores.
• FIG. 2B is a block diagram illustrating a remote server system that can be used to accumulate vehicle driver data for a plurality of vehicles and drivers and compute driver risk scores.
• FIG. 3 is a conceptual block diagram of the unsupervised machine learning system, including example information that can be useful in computing driver risk scores.
• FIG. 4 is a conceptual block diagram of an unsupervised machine learning system using a plurality of isolation forests that can be used to identify safe and unsafe driving behavior.
• FIG. 5 is a flowchart of a method for determining driver risk scores, in accordance with one embodiment.
• FIG. 6 is a flowchart of a method for calculating a driver risk score by using machine learning classifiers (e.g., isolation forests and/or other classifiers) and ensemble scoring, in accordance with another embodiment.
• FIG. 7 is a flowchart of a method for engaging vehicle safety systems based on a driver risk score, in accordance with yet another embodiment.
• FIG. 8 is a conceptual block diagram of a supervised machine learning system.
• FIG. 9 is a conceptual block diagram the supervised machine learning system, including example information.
• FIG. 10 illustrates example training data for a supervised machine learning system.
• FIG. 11 is a conceptual block diagram of a system for testing the labeled training features ofFIG. 10 and determining a biased driver risk score.
• FIG. 12 shows an example plot of unlabeled data for different vehicles.
• FIG. 13 is a conceptual block diagram of a generic unsupervised machine learning system.
• FIG. 14 is another conceptual block diagram of the unsupervised machine learning system for determining driver risk scores.
• FIG. 15 is an example score distribution that may be determined.
• FIG. 16 is a chart to compare results of vehicle behavior (e.g., driver behavior) for three factors.
• FIG. 17 is a chart to compare the results of vehicle behavior with an average to determine safe, average, and unsafe drivers.
DETAILED DESCRIPTION
• Turning now to the drawings, systems and methods for risk scoring for a vehicle operator (e.g., driver) by using unsupervised machine learning classifiers in accordance with embodiments that are disclosed. Many vehicles are equipped with a telematics device which can collect a variety of sensor data. This data can be analyzed to produce a driver risk score, based on past information, to predict future driving risks.
• Vehicle telematics systems in accordance with a variety of embodiments can utilize a variety of advance machine learning techniques to identify driver behavior and generate driver risk scores that more accurately assess a particular driver's risk than prior art techniques. Unsupervised learning classifiers (e.g., isolation forests and/or other classifiers) can identify relationships in uncategorized vehicle driver data to produce a driver risk score, in accordance with the embodiments. Unsupervised learning classifiers can detect anomalies in vehicle driver data by isolating anomalies from normal points in the data set. These anomalies in data points can generally indicate safe or unsafe driver behavior. While an individual unsafe data point may not indicate a risky driver, a collection of unsafe data points can start to indicate a risk. In several embodiments, isolation forests, an unsupervised learning machine learning classifier, can be used to detect such anomalies. Isolation forests, which are discussed in detail herein, can make random divisions of a data set while detecting anomalies. Several isolation forests constructed for the same data set will each individually make different random choices and therefore each isolation forest can be slightly different. In some embodiments, the system can combine the output of a group of isolation forests by using an ensemble learning process to generate a driver risk score.
• Systems and methods for driver risk scoring by using unsupervised machine learning classifiers in accordance with the disclosed embodiments are described in more detail herein.
• Vehicle Telematics Systems
• FIG. 1 is a conceptual diagram of avehicle telematics system100, in accordance with an embodiment. Vehicle telematics systems described herein can use a variety of data regarding the operation of a vehicle to generate a operator risk score (e.g., driver risk score) for the operator (e.g., driver) of the vehicle. Thevehicle telematics system100 includes one or more vehicle telematics devices (110,110′, etc.) typically mounted in or on a vehicle (102,102′, etc.).
• For explanatory purposes, the embodiments are described herein by using vehicles operated by drivers. A driver is a vehicle operator of various types of vehicles. A vehicle operator can be a vehicle driver, a ship captain, an airplane pilot, a drone pilot, a motorcycle rider, a bicycle rider, an automated operator, and/or a remote operator, among other operators. A vehicle (102,102′, etc.) can be any car, truck, bus, train, airplane, helicopter, drone, motorcycle, bicycle, watercraft, land craft, and/or aircraft, among other vehicles. A vehicle (102,102′, etc.) can be manned, unmanned, motorized, unmotorized, directly operated, remotely operated, artificial intelligence operated, self-driving, self-flying, and/or self-sailing, among other things. A vehicle can be operated by an operator (e.g., driver) and/or operated at least in part by an automated system (e.g., self-driving system, etc.).
• Many other configurations of thevehicle102 are within the scope of the present technology. As one of many examples,FIG. 1 shows thevehicle102′ including avehicle telematics device110′ having amobile communication device116′. Thevehicle telematics device110′ is coupled to avehicle data bus112′ and an I/O interface114′. Thedevices110′,112′,114′, and116′ function like thedevices110,112,114, and116, but may have different physical configurations.
• Thevehicle telematics device110 can be coupled to a connector and/or a wire harness in communication with avehicle data bus112 of thevehicle102 to obtain power and exchange signals with one or more vehicle devices or sensors. Thevehicle telematics device110 can further be coupled to a wired or wireless input/output (I/O)interface114 and/or amobile communications device116 as appropriate to the requirements of specific applications of the embodiments.
• In a variety of embodiments, thevehicle telematics device110 communicates with theremote server system130 via themobile communications device116 over anetwork120. In a variety of embodiments, thenetwork120 is the Internet. In many embodiments, thenetwork120 is any wired or wireless network, such as a cellular network, between thevehicle telematics device110 and/or theremote server system130. In a number of embodiments, theremote server system130 is implemented by using a single server system. In several embodiments, theremote server system130 is implemented by using multiple server systems.
• In a variety of embodiments, thevehicle telematics device110 is installed in avehicle102 having thevehicle data bus112. In several embodiments, thevehicle telematics device110 is connected to a vehicle diagnostic connector that provides access to thevehicle data bus112. Thevehicle telematics device110 can obtain data from any of a variety of vehicle devices connected to thevehicle data bus112 utilizing any of a variety of techniques as appropriate to the requirements of specific applications of embodiments. Vehicle devices can include, but are not limited to, engine sensors, electronic control unit (ECU) devices, alternator sensors, vibration sensors, voltage sensors, oxygen sensors, Global Positioning System (GPS) receivers, ignition devices, weight sensors, wireless network devices, and/or acceleration determination devices. Systems and methods for connecting to a vehicle data bus that can be utilized in accordance with the embodiments are described in SAE J1978, titled “OBD II Scan Tool,” first published by SAE International of Troy, Mich. on Mar. 1, 1992 and last updated Apr. 30, 2002. Systems and methods for obtaining data from devices connected to a vehicle data bus are described in SAE J1979, titled “E/E Diagnostic Test Modes,” first published by SAE International on Dec. 1, 1991 and last updated Aug. 11, 2014. The disclosures of SAE J1978 and SAE J1979 are hereby incorporated by reference in their entirety. In a number of embodiments, the vehicle telematics device is connected directly, either wired or wirelessly, to one or more sensors within thevehicle102 and/or does not utilize thevehicle data bus112.
• Thevehicle telematics device110 can include any of a variety of sensors and/or devices, including those described herein with respect to the vehicle data bus and any described in more detail herein, to obtain data regarding the status of the vehicle and its environment. Thevehicle telematics device110 can also communicate with any of a variety of sensors and/or devices by using the I/O interface114. The I/O interface114 can be any connection, including wired and wireless connections, as appropriate to the requirements of specific applications of the embodiments.
• In several embodiments, thevehicle telematics device110 is capable of executing scripts to read data and/or perform particular processes. These scripts can be pre-loaded on the device and/or obtained from theremote server system130,vehicle data bus112, and/or the I/O interface114 as appropriate to the requirements of specific applications of the embodiments. Thevehicle telematics device110 can be self-powered and/or connected into the electrical system of thevehicle102 in which thevehicle telematics device110 is installed. In a variety of embodiments, the vehicle telematics device is powered via thevehicle data bus112 and/or the I/O interface114.
• In many embodiments, one of thesensor devices240 of thevehicle telematics device110 is a Global Positioning System (GPS) receiver in order to determine the location, speed, and/or acceleration of thevehicle102. In many embodiments, one of thesensor devices240 of thevehicle telematics device110 is a multidimensional accelerometer to acquire acceleration and/or speed of thevehicle102.
• In a variety of embodiments, thevehicle telematics device110 and/orremote server system130 provides a user interface allowing for visualizing and interacting with the data transmitted and/or received between the systems. In several embodiments, thevehicle telematics device110 and/orremote server system130 provides an interface, such as an application programming interface (API) or web service that provides some or all of the data to third-party systems for further processing. Access to the interface can be open and/or secured by using any of a variety of techniques, such as by using client authorization keys, as appropriate to the requirements of specific applications.
• Although a specific architecture of a vehicle telematics system in accordance with embodiments are discussed herein and illustrated inFIG. 1, a variety of architectures, including sensors and other devices and techniques not specifically described herein, can be utilized in accordance with embodiments. Furthermore, the processes described herein can be performed by using any combination the vehicle telematics device, mobile communications device, and/or the remote server systems as appropriate to the requirements of specific applications of the embodiments.
• Vehicle Telematics Devices and Remote Server Systems
• FIG. 2A is a conceptual block diagram of thevehicle telematics device110. Vehicle telematics devices and remote server systems in accordance with the embodiments can transmit and receive data regarding the performance of a vehicle and/or driver risk scores. Thevehicle telematics device110 includes aprocessor210 in communication withmemory230. Thevehicle telematics device110 can also include one ormore communication interfaces220 capable of sending and receiving data. In a number of embodiments, thecommunication interface220 is in communication with theprocessor210, thememory230, and/or the sensor device(s)240. In several embodiments, thememory230 is any form of storage configured to store a variety of data, including, but not limited to, avehicle telematics application232,sensor data234, andtelematics data236. In many embodiments, thevehicle telematics application232,sensor data234, and/ortelematics data236 are stored by using an external server system and received by thevehicle telematics device110 by using thecommunications interface220.
• Sensor devices240 can include RPM sensors, voltage sensors, GPS receivers, noise sensors, vibration sensors, acceleration sensors, weight sensors, and any other device capable of measuring data regarding a vehicle as appropriate to the requirements of specific applications of the embodiments.Sensor devices240 can be included within thevehicle telematics device110 and/or located external to thevehicle telematics device110. Thevehicle telematics device110 can communicate with external sensor devices by using thecommunications interface220, such as via a vehicle data bus, I/O interface (including serial interfaces), mobile communications device, and/or a network connection as appropriate to the requirements of specific applications of embodiments. In a variety of embodiments, a vehicle telematics device is connected to a diagnostic connector (e.g. an OBD II port) in a vehicle. In some embodiments, information collected fromsensor devices240 and/orsensor data234 can be used as vehicle driver data in a variety of machine learning processes for driver scoring as described in more detail herein. For example, vehicle driver data (e.g., raw data collected by vehicle telematics device) can include, without limitation, vehicle speed, acceleration data, and steering data over time and location for the detection of speeding, hard acceleration, hard deceleration, vehicle swerving, time of day, date, traffic information, weather, continuous hours driven. The system can combine raw data from a vehicle telematics device with server-generated data, such as location, time, date, weather, and/or other data. It should be readily appreciated by one having ordinary skill that these are merely illustrative examples and any such information can be used as appropriate to the requirements of specific applications.
• FIG. 2B is a conceptual block diagram of a remote server system, in accordance with an embodiment. Theremote server system130 includes aprocessor252 in communication withmemory260. Theremote server system130 can also include one ormore communications interfaces254 capable of sending and receiving, such as with a vehicle telematics device. In a number of embodiments, the communication interface is in communication with theprocessor252 and/or thememory260. In several embodiments, thememory260 is any form of storage configured to store a variety of data, including, but not limited to, aserver application262, anoperating system264,vehicle driver data266, andhistorical data268. In many embodiments, theserver application262, anoperating system264,vehicle driver data266, and/orhistorical data268 are stored by using an external server system and received by theremote server system130 by using theremote communications interface254.
• Theprocessor210 andprocessor252 can be directed, by thevehicle telematics application232 and theserver application262 respectively, to perform a variety of driver risk scoring processes. Driver risk scoring processes can include obtaining data vehicle driver data and generating driver risk scores by using a vehicle telematics device and/or a remote server system by using a variety of machine learning systems. Driver risk scoring processes that can be performed in accordance with embodiments are described in more detail herein.
• Although specific architectures for vehicle telematics devices and remote server systems in accordance with embodiments are conceptually illustrated inFIGS. 2A-B, any of a variety of architectures, including those that store data or applications on disk or some other form of storage and are loaded into memory at runtime, can also be utilized. Additionally, any of the data utilized in the system can be cached and transmitted once a network connection (such as a wireless network connection via the communications interface) becomes available. In a variety of embodiments, a memory includes circuitry such as, but not limited to, memory cells constructed by using transistors, that are configured to store instructions. Similarly, a processor can include logic gates formed from transistors (or any other device) that dynamically perform actions based on the instructions stored in the memory. In several embodiments, the instructions are embodied in a configuration of logic gates within the processor to implement and/or perform actions described by the instructions. In this way, the systems and methods described herein can be performed utilizing both general-purpose computing hardware and by single-purpose devices.
• Rule-Based Algorithms
• Traditionally, rule-based algorithms calculate a driver risk score via algorithms that use manually inputted rules. Driver risk scores are manually calculated using rule-based algorithms with the data frequently forced into a bell curve to facilitate the calculation of the driver risk score. Examples of manually inputted rules are the following: “If the driver has more than ten instances of hard braking over seven days, then the driver risk score is reduced; and if the driver has less than three instances of hard braking over seven days, then the driver risk score is increased.”
• Unfortunately, there are several problems with manual, rule-based algorithms. A first problem is manual thresholding, which does not enable an accurate general application to different locales and diverse geography. For example, ten hard braking instances might be considered excessive in Omaha, Nebr., but might be considered low in New York City. So, it might be fair for the manual, rule-based algorithms to ding such a driver in Omaha, while it might by unfair to ding such a driver in New York City. A second problem with manual, rule-based algorithms is unjustified data transformations. For example, forcing the raw data received from vehicle telematics devices into a bell curve, for the sake of easy calculations, might not accurately reflect reality. A third problem with manual, rule-based algorithms is the rules are arbitrary, which leads to suboptimal algorithm performance. For example, a rule that states, for example, “ten instances of hard braking are excessive”, may be a best guess by an expert. Unfortunately, such a best guess is highly unlikely to be better than an estimate that is empirically determined from raw data received from vehicle telematics devices. A fourth problem is that manually inputted rules incorporate human biases. For example, an expert in vehicle driving safety may consider hard left turns to be unsafe, when evidence may show hard left turns are not actually unsafe.
• Supervised Machine Learning Systems
• Traditionally, supervised learning classifiers may be used to identify relationships in labeled vehicle telematics information. A known (or labeled) set of vehicle telematics device information, which can be referred to as a training set, can be used to train the machine learning classifier. Once the machine learning classifier is trained by using the labeled training set, the machine learning classifier can classify unknown sets of vehicle telematics device information. Correctly classified information can be added to the training set to continuously improve the performance of the machine learning system. Similarly, information related to incorrectly classified vehicle telematics information can also be added to the training data set to improve the precision of the machine learning system.
• Supervised machine learning classifiers analyze information collected by vehicle telematics devices to produce a driver risk score. Supervised learning classifiers identify relationships in labeled vehicle telematics information. A known (or labeled) set of vehicle telematics device information, which can be referred to as a training set, is used to train the machine learning classifier. Once the machine learning classifier is trained by using the labeled training set, the machine learning classifier classifies unknown sets of vehicle telematics device information. Correctly classified information is added to the training set to continuously improve the performance of the machine learning system. Similarly, information related to incorrectly classified vehicle telematics information is added to the training data set to improve the precision of the machine learning system.
• FIG. 8 is a conceptual block diagram of a supervisedmachine learning system800. Thesystem800 receivesraw data802 from one or more vehicle telematics devices. From theraw data802, thesystem800 performs feature extraction to generateFeature1 through Feature n. Each feature is a description about theraw data802. Example features are described with reference toFIG. 9. Thesystem800 inputs the features into amodel812, which predicts an outcome probability Pr(outcome).
• FIG. 9 is a conceptual block diagram of the supervisedmachine learning system800, including example feature extraction of information from the available vehicle/driver data. Thesystem800 receivesdriver data802 from one or more vehicle telematics devices. From the driver data902, thesystem800 performs feature extraction to generate a speeding description, a hard acceleration description, a hard deceleration description, a serving description, and so on. Each feature is a description about thedriver data802. At the feature extraction stage, thesystem800 is not yet calculating a driver risk score. The system is, for example, keeping track of the number of times an event (e.g., speeding) occurs (e.g., frequency). In this example, the system inputs the features into amodel812, which predicts an outcome probability Pr(Driver has accident), which is the probability of the driver having an accident.
• FIG. 10 illustrates example training data for a supervised machine learning system. Supervised learning involves procedures training the model, including providing labeled data to a model. When an outcome is known, the supervised system can label data points. In example 1000, Driver A has an outcome of an accident. Accordingly, the supervised system can label the following features for Driver A: speeding 25% of the driving time;hard accelerations 60% of acceleration time; andhard decelerations 40% of deceleration time. The supervised system labels those features as being associated with an accident. In example 1020, Driver B has an outcome of no accident. According, the supervised system applies a label to the following features for Driver B: speeding 15% of the driving time;hard accelerations 40% of acceleration time; andhard decelerations 60% of deceleration time. The supervised system labels those features as being associated with no accident, or the system does not those features. The supervised system can train the models for hundreds or thousands of labeled data points (e.g., labeled features).
• FIG. 11 is a conceptual block diagram of asystem1100 for testing the labeled training features ofFIG. 10. The system generates amodel1102 based on labeled features, which are generated as discussed with reference toFIG. 10. As shown inFIG. 11, the system receives unlabeled data, input that unlabeled data into themodel1102, and generate a probability Pr(X), which is the probability of event X happening. In this example, the supervised system calculates an accident probability Pr(accident) equal to 75%, given the unlabeled data inputted into the trainedmodel1102.
• Unfortunately, supervised machine learning requires an adequate amount of labeled historical data to train models. Meanwhile, a system may not have been labeling features when certain events occur (e.g., drivers get into accidents) over time in a database of historical driver data. Lack of an adequate database of historical labeled features, renders the supervised machine learning system inaccurate and unreliable.
• Unsupervised Machine Learning Systems
• Unsupervised machine learning systems, in accordance with embodiments, are designed to solve problems associated with rule-based algorithms and supervised machine learning systems. Unsupervised machine learning classifiers can analyze information collected by vehicle telematics devices to produce a driver risk score. Unsupervised learning classifiers (e.g., isolation forests and/or other classifiers) can identify relationships in uncategorized vehicle driver data to produce a driver risk score. Unsupervised learning classifiers can detect anomalies in vehicle driver data by isolating anomalies from normal points in the data set. These anomalies in data points can generally indicate safe or unsafe driver behavior. While an individual unsafe data point may not indicate a risky driver, a collection of unsafe data points can start to indicate a risk. In several embodiments, the system can use isolation forests, an unsupervised learning machine learning classifier, to detect such anomalies. Isolation forests, which are described herein, can make random divisions of a data set while detecting anomalies. Several isolation forests constructed for the same data set will each individually make different random choices and therefore each isolation forest can be slightly different. In some embodiments, the system can combine the output of a group of isolation forests by using an ensemble learning process to generate a driver risk score.
• FIG. 12 shows anexample plot1200 of unlabeled data, in accordance with an embodiment. The axes do not mean anything specific. Theplot1200 is provided for explanatory purposes. The unsupervised system receives the unstructured, unlabeled data shown in theplot1200. The unsupervised system calculates some sort of judgment in terms of “same versus different” for each unlabeled data point. In the example ofFIG. 12, the unsupervised system determinesdata points1202,1204, and1206 are “different” and determines all the other data points are “same”.
• FIG. 13 is a conceptual block diagram of an unsupervisedmachine learning system1300, in accordance with an embodiment. Thesystem1300 receives rawvehicle driver data1302 from one or more vehicle telematics devices. The system receives a set of features for data extraction. For example, a human can determine that thesystem1300 will extract data associated with the following features: acceleration, deceleration, and speed. Each feature is a description about thevehicle driver data1302. Example features are described with reference toFIG. 3. From thevehicle driver data1302, thesystem1300 performs feature extraction to generate data associated withFeature1 through Feature n. Thesystem1300 inputs the features into amodel1312, which assigns a label (e.g., L₁, . . . , L_n) to the data. Themodel1312 includes any algorithm that uses a machine learning classifier to perform detection of anomalies (e.g., outliers, “different” data, etc.). Unsupervised learning classifiers for themodel1312 can include (but are not limited to) k-means clustering, mixture models, hierarchical clustering, anomaly detection, artificial neural networks, expectation-maximization algorithms, principal component analysis, independent component analysis, singular value decomposition, isolation forests, and/or a combination of unsupervised learning classifiers.
• FIG. 3 is a conceptual block diagram of an unsupervisedmachine learning system300, including example driving features that may be extracted from the vehicle driver data in order that it may be classified. Machine learning systems in accordance with embodiments can receive vehicle driver data from a vehicle telematics device to identify baseline driver behavior and/or anomalies in the vehicle driver data. Themachine learning system300 includesvehicle driver data302. Thevehicle driver data302 can include a variety of information such as speeding304,hard acceleration306,hard deceleration308, and/or swerving310, which is evidence about a driver's driving behavior or habits. An unsupervisedmachine learning classifier312 can receive information as input and use it to generate alabel314. In some embodiments,label314 can indicate anomalies in vehicle driver data such as labeling the data according to a binary decision (e.g., SAME or DIFFERENT). In various embodiments, SAME labels can indicate SAFE driver actions and/or DIFFERENT labels can indicate UNSAFE driver actions. Accordingly, the unsupervised system can operate under the assumption that an average driver (e.g., SAME driver) is a SAFE driver, while an anomalous driver (e.g., DIFFERENT driver) is an UNSAFE driver. However, it should be noted that any labels (including any number of labels) can be utilized as appropriate to the requirements of specific applications of embodiments.
• An isolation forest is one technique, among others, to detect data anomalies (e.g., “different” data, outliers, etc.). Isolation forests are based on the fact that anomalies are data points that are few and different. Because of these properties, anomalies are susceptible to a mechanism called isolation. Isolation forests introduce the use of isolation as a more effective and efficient means to detect anomalies than the commonly used basic distance and density measures. Moreover, an isolation forest algorithm has a relatively low linear time complexity and a relatively small memory requirement, compared to other machine learning methods. An isolation forest algorithm can isolate observations by randomly selecting a feature and then randomly selecting a split value between the maximum and minimum values of the selected feature. Isolating anomaly observations is easier because only a few conditions are needed to separate those cases from the normal (e.g., “same”) observations. On the other hand, isolating normal observations requires more conditions. Accordingly, an anomaly score can be calculated as the number of conditions required to separate a given observation. The isolation forest algorithm constructs the separation by creating isolation trees (e.g., random decision trees). Then, the system calculates a score as a path length to isolate the observation.
• Theunsupervised system300 randomly builds many decision trees for each feature. Theunsupervised system300 then averages the tree distance between a point and its nearest neighbors. The clustering includes training observations, new regular observations, and new abnormal observations. The training observations represent training set data. The new regular observations represent driver behavior that theunsupervised system300 may consider to be normal (e.g., “same”). The new abnormal observations represent driver behavior that theunsupervised system300 may consider to be an anomaly (e.g., outlier or “different”).
• If using isolation forests, vehicle driver data is clustered according to the driving behaviors indicated in the vehicle driver data to identify baseline driver behavior. By clustering the vehicle driver data to identify large clusters of driver behavior and assign those as the baseline driver behavior, labels can be applied to the clustered vehicle driver data to identify safe drivers as most drivers do not regularly have accidents, so that these clusters identifying baseline driver behavior can identify a variety of characteristics of ordinary driving without relying on external labels being applied to the source data. Anomalies in the vehicle driver data can be identified by locating pieces of vehicle driver data that are statistically deviant from the baseline driver behavior. Additionally, the identification of baseline driver behavior can overcome a variety of difficulties in applying machine learning techniques when the vehicle driver data does not include labels identifying the vehicle driver data.
• The unsupervisedmachine learning classifier312 can include a variety of classifiers including (but not limited to) supervised learning classifiers, unsupervised learning classifiers, and/or a combination of several classifiers. Supervised learning classifiers can include (but are not limited to) artificial neural networks, nearest neighbor algorithms, decision trees, support vector machines, random forests, ensembles of classifiers, and/or a combination of supervised learning classifiers. In a variety of embodiments, supervised learning classifiers can be further adapted to be unsupervised learning classifiers for the identification of driver behavior classification as appropriate to the requirements of specific applications of embodiments. Unsupervised learning classifiers can include (but are not limited to) k-means clustering, mixture models, hierarchical clustering, anomaly detection, artificial neural networks, expectation-maximization algorithms, principal component analysis, independent component analysis, singular value decomposition, isolation forests, and/or a combination of unsupervised learning classifiers.
• Additionally, machine learning classifiers can optionally use feature extraction to combine information in a way that still meaningfully represents the data. It should be readily apparent to one having ordinary skill in the art that many feature extraction techniques are available such as (but not limited to) principal component analysis, independent component analysis, isomap analysis, and/or partial least squares, and that feature extraction itself is optional.
• FIG. 4 is conceptual block diagram of an unsupervisedmachine learning system400, in accordance with an embodiment. Themachine learning system400 uses ensemble learning with isolation forests (e.g., ensemble of isolation forests) to generate driver risk scores. Isolation forests are shown for explanatory purposes. However, themachine learning system400 can use any of a variety of machine learning classifiers to generate driver risk scores.
• The unsupervisedmachine learning system400 includesvehicle driver data402. This driver data can be input one ormore isolation forests404. Isolation forests (or iForests) are an machine learning classifier that can detect anomalies in a data set rather than modelling normal points and are generally made of many isolation trees (or iTrees). In many embodiments, anomalies in vehicle telematics device information can be more susceptible to isolation than normal points. Information can be isolated through isolation trees. In a variety of embodiments, isolation trees are binary trees, although a variety of data structures can be utilized to express isolation trees as appropriate to the requirements of specific applications of embodiments. An attribute can randomly be selected at each node and a random split point can be chosen that can divide the data set into two sub-regions. This process can be repeated until every point in the data set is isolated from the rest of the points. The top of an isolation tree will generally contain points that are isolated quickly.
• An isolation forest classifier can construct many isolation trees, each slightly different due to the random split points. Anomalies in a data set are generally points that have a shorter average path length than normal points in a set of isolation trees within an isolation forest. Detected anomalies within isolation forests can be used to label an isolation forest as SAFE and/or UNSAFE. Systems and methods for generating isolation forests are described in “Isolation Forests” to F. T. Liu, K. M. Ting, and Z.-H. Zhou inProc. ICDM2008, pages 413-422, 2008, the disclosure of which is hereby incorporated by reference in its entirety. In some embodiments, the output of many isolation forests can be aggregated to generate a driver risk score by using ensemble scoring.
• Although specific systems for using machine learning classifiers in accordance with embodiments are shown inFIG. 3 and systems for generating driver risk scores by using isolation forests are described herein with respect toFIG. 4, the system can use any of a variety of machine learning classifiers to generate driver risk scores, in accordance with embodiments.
• The system can parallelize the computations of the machine learning classifiers (e.g., ensemble isolation forests) across as many processing cores that are available. For example, if a server has n processors, then each of the n processors can work to perform computations for one machine learning classifier (e.g., one isolation forest).
• FIG. 14 is another conceptual block diagram of the unsupervisedmachine learning system400, in accordance with an embodiment.FIG. 15 illustrates how the unsupervised system can calculate a driver risk score. Thesystem400 sends each data point to multiple isolation forest. Each isolation forest is different from the rest. Thesystem400 calculates a risk score based on the one or more labels generated by the one or machine learning classifiers, which are isolation forests in the example ofFIG. 14. The driver risk score is further described with reference toFIG. 6.
• FIG. 15 is anexample score distribution1500, in accordance with an embodiment. In this example, the system filters data to consider a population of313 sedans. The filtering is an attempt to compare similar vehicles. For example, it would be inappropriate to compare sedans with dump trucks. The system considers raw data captured by vehicle telematics devices over the course of one month. The system uses the following features: acceleration, deceleration, and speed, which are discussed with reference toFIG. 3. The system uses the ensemble isolation forest method, which is described with reference toFIG. 4. This example uses two hundred and fifty-six (256) isolation forest votes for thescore distribution1500. As shown inFIG. 15, about 23 vehicles had a score between 2 and 10; about 120 vehicles had a score between 11 and 20, and so on.
• FIG. 16 is acomparison1600 of results of vehicle behavior (e.g., driver behavior), in accordance with an embodiment. The system uses the following features: acceleration, deceleration, and speed, which are discussed with reference toFIG. 3. The numbers on the vertical axis represent normalized scores. Regarding the acceleration feature, asedan2 is relatively far away from the average acceleration, while asedan1 is relatively close to the average acceleration. Regarding the deceleration feature,sedan2 is close to the average deceleration, whilesedan1 is also not too far away from the average deceleration. Regarding speed, there is no data forsedan2, whilesedan1 is not that far away from the average speed. Accordingly, forsedan1, the system calculates a driver risk score of 9, which indicates a relatively low level of risk. Forsedan2, the system calculates a driver risk score of 95, which indicates a relatively high level of risk. Note that magnitude and spectral differences. Isolation forests can learn spectral patterns as well as dynamic thresholds.
• FIG. 17 is acomparison1700 of results of vehicle behavior associated with unsafe drivers, in accordance with an embodiment. The system uses the following features: acceleration, deceleration, and speed, which are discussed with reference toFIG. 3. The numbers on the vertical axis represent normalized scores. In this example, the behaviors ofsedan3,sedan4, andsedan5 are compared. Each sedan (e.g., driver) is substantially far away from the average acceleration, the average deceleration, and/or the average speed. Accordingly, the system calculates each sedan to be associated with a driver risk score of 100, which indicates the highest rating for the level of risk.
• Methods of Driver Risk Scoring with Unsupervised Machine Learning
• FIG. 5 is a flowchart of amethod500 for determining driver risk scores, in accordance with an embodiment. Driver risk scoring processes can include gathering vehicle driver data from sensor devices and calculating driver risk scores. The system receives (502) vehicle driver data by (but not limited to) a vehicle telematics device and/or a remote server system. A vehicle telematics device can receive vehicle driver data directly from vehicle telematics device sensors. Additionally, a vehicle telematics device can receive vehicle driver data, such as (but not limited to) historical driver data, from a remote server system. Similarly, a remove server can receive vehicle driver data from a plurality of vehicle telematics devices that is collected by one or more sensors of the vehicle telematics devices.
• The system identifies (504) anomalies in vehicle driver data with an adaptive machine learning model. In many embodiments, the model includes a machine learning classifier, such as (but not limited to) an unsupervised learning machine learning classifier, that can label points in the driver data that are the SAME (part of a cluster) and DIFFERENT (an anomaly). Driver data labeled SAME can indicate SAFE driver actions. Similarly, driver data labeled DIFFERENT can indicate UNSAFE driver actions. A vehicle telematics device can receive the adaptive machine learning model, such as (but not limited to) coefficients of the model, from a remote server system.
• The system calculates (506) a driver risk score by using anomalies identified in vehicle driver data. This calculation can be performed by (but is not limited to) one or more processors within the vehicle telematics device. This calculation can alternatively or additionally be performed by one or more processors of the remote server system.
• A driver may receive feedback from the telematics device in real time about his driving behavior. The driver may receive a warning (e.g. visible warning light, audible message, vibration from vibrator) from the telematics device if his/her driver score is unsafe or a periodic compliment (e.g. visible positive reinforcement light, audible message, vibration from vibrator) if the driver score is safe or improves. The feedback regarding driving behavior may be sent remotely from the remote server system to the telematics device and provided to the user with driver risk score calculations being performed by the remote server system.
• In some embodiments, an individual driver can be tracked across several vehicles to calculate a driver risk score. In such cases, vehicle driver data from multiple vehicles can be used to calculate a driver risk score. Alternatively, a driver risk score can be calculated for an individual vehicle for a specific date range. In some embodiments, a driver risk score for a vehicle from two date ranges can be compared to see if driving has improved or become riskier.
• Although a variety of processes for generating a driver risk score are discussed with respect toFIG. 5, any of a variety of processes capable of identifying risks from anomalies and clusters in vehicle driver data can be utilized as appropriate to the requirements of specific applications in accordance with several embodiments.
• FIG. 6 is a flowchart of amethod600 for calculating a driver risk score by using machine learning classifiers (e.g., isolation forests and/or other classifiers) and ensemble scoring, in accordance with an embodiment. Driver risk scoring processes can include any of a variety of machine learning classifiers to identify anomalies and clusters in vehicle driver data. Anomalies in vehicle driver data are distinguishable from the clusters of vehicle driver data. Isolation forests are just one example of machine learning classifiers that can be used to distinguish clusters of vehicle driver data from anomalies in vehicle driver data. The anomalies are the outliers from the clusters generated with the models.
• The system can receive (601) vehicle driver data. For example, a remote server system can receive vehicle driver data from a plurality of vehicle telematics devices. As another example, a vehicle telematics device, or a computing device coupled to the vehicle telematics device, can receive vehicle driver data from sensors and/or from a remote server system. The remote server system, the vehicle telematics device, and/or another computing device can receive vehicle driver data periodically, or in real time as a driver is driving a vehicle.
• The system can filter (602) the received vehicle driver data so that proper comparisons are made between similar vehicles. For example, the system can filter data so that sedan to sedan comparisons are made with the data, and data from sedans are not compared with data from dump trucks. As another example, the system can filter data so that a sports car is compared to other sports cars, and not to pickup trucks. However, embodiments are not limited to these examples. Any filtering is permissible to compare any vehicle type.
• The system generates (604) one or more unsupervised machine learning models by using the vehicle driver data. The models are used to classify the vehicle driver data to identify clusters and distinguish anomalies with little to no bias. One such unsupervised machine learning model that can be used is isolation forests.
• The system identifies (606) anomalies in the vehicle driver data by using the models. The anomalies are the outliers from the clusters generated with the models.
• The system uses the identified anomalies in the driver data to label (608) the output of each model as SAFE or UNSAFE. A specific vehicle driver data point in a cluster of vehicle driver data can be labeled SAFE and indicate safe driver actions. A specific vehicle driver data point that is an anomaly outside the clusters can be labeled UNSAFE and indicate unsafe driver actions.
• The system can calculate (610) a driver risk score by using the number of models labeled UNSAFE and using an ensemble scoring machine learning classifier. It should be readily apparent to one having ordinary skill that ensemble isolation forests are merely illustrative and any variety of processes to evaluate a collection of labeled machine learning classifiers to generate a driver risk score as appropriate to many embodiments. In various embodiments, a risk score can be calculated by using the following equation:
• $\mathrm{Risk}\mathrm{Score}\left(v_i\right)=\frac{100\times \mathrm{Num\_UNUSAFE}}{n},$
• where v_iis a set of vehicle driver data, n is a total number of models, and Num_UNSAFE is the number of models with an output of UNSAFE. The system can compare driver risk score to a threshold to determine what sort of risk it indicates. As an illustrative example, where driver risk scores can range from 0-100, a driver risk score of 9 can indicate a low level of risk and a driver risk score of 95 can indicate a high level of risk. However, it should be noted that the system can use any risk scoring range, including those where a risk score below a threshold value indicates a high risk and a risk score above a threshold value includes a low risk, as appropriate to the requirements of specific applications of embodiments.
• In many embodiments, driver risk scoring processes can include sending safety control signals to a vehicle for the safe control thereof when a calculated driver risk score is at or above a predetermined threshold. As an illustrative example, these safety control signals can (but are not limited to) impose limitations on vehicle speed and/or minimum driving distances from other vehicles, force a vehicle to change from a user-controlled driving mode to a self-driving mode, and/or alert other drivers to the unsafe driver by for example (but not limited to) honking a vehicle horn and/or flashing vehicle headlights until a driver stops, shuts off, and/or exits the vehicle.
• Accordingly, the system can use a variety of processes to generate drive risk scores by using machine learning classifiers (such as isolation forests) and then to control the operation of a vehicle, in accordance with embodiments.
• FIG. 7 is a flowchart of amethod700 for engaging a vehicle safety system based on a driver risk score, in accordance with an embodiment. The system obtains, receives, or calculates (702) a driver risk score. Calculation of driver risk score is discussed with reference toFIG. 6. The system identifies (704) one or more safety violations based on the driver risk score and/or the raw data received from a vehicle telematics device associated with the vehicle. For example, the system may identify the vehicle is traveling at one hundred miles per hour. The system generates (706) one or more safety control signals based on the identified safety violation. For example, the system can generate a safety control signal to instruct the engine (e.g., via an engine control unit coupled to the processor) to reduce rotations per minute (rpm).
• The system engages (708) one or more safety systems based on the one or more safety control signals. For example, the system slows the rpm of the engine based on a safety control signal, and thereby slows down (e.g., reduces velocity) the vehicle. As another example, the system can include, or communicate with, a display device and/or an audio device mounted in the vehicle and coupled to the processor to alert the driver of the unsafe driving, based on a safety control signal. As another example, the system can be coupled to the braking system of a vehicle (e.g., via an antilock brake control unit coupled to the processor) to apply the brakes and further slow down the vehicle based on the control signal. As another example, the system can be coupled to the transmission system of a vehicle (e.g., via a transmission control unit coupled to the processor) to shift down to a lower gear to slow down the vehicle, based on the control signal. As another example, the system can be coupled to the ignition system of a vehicle (e.g., via an ignition or engine control unit coupled to the processor) to shut the engine off, perhaps after bringing the vehicle to a stop at a safe location, based on the control signal. The vehicle can be temporarily disabled from operation in this case. As another example, the safety system can be coupled to an autonomous driving (e.g., autopilot) system coupled to the processor to engage the autopilot based on the safety control signal to automatically drive the vehicle to a safe location.
• Unsupervised machine learning has several advantages. A first advantage is the “rules” that have been learned reflect the dataset and the dataset only. For example, the “rules” are not arbitrary manually-inputted definitions and do not reflect human biases. A second advantage is online learning. For example, the system can learn from a plurality of vehicle telematics devices in real time or not real time, as appropriate. A third advantage is an administrator can quickly and easily add a new feature (e.g., hard reversing or whatever) or remove an unwanted feature (e.g., acceleration or whatever). The quick adding or removal of features enables rapid geographic segmentation and learning.
• Although the embodiments have been described in certain specific aspects, many additional modifications and variations would be apparent to those skilled in the art. In particular, any of the various processes described herein can be performed in alternative sequences and/or in parallel (on the same or on different computing devices) in order to achieve similar results in a manner that is more appropriate to the requirements of a specific application. It is therefore to be understood that the embodiments that are disclosed can be practiced otherwise than specifically described without departing from the scope and spirit of this disclosure. Thus, the embodiments should be considered in all respects as illustrative and not restrictive. It will be evident to the person skilled in the art to freely combine several or all of the embodiments discussed here as deemed suitable for a specific application. Throughout this disclosure, terms like “advantageous”, “exemplary” or “preferred” indicate elements or dimensions that are particularly suitable (but not essential) to one or more embodiments, and may be modified wherever deemed suitable by the skilled person, except where expressly required. Accordingly, the scope of the embodiments should be determined not by those illustrated and described, but by the appended claims and their equivalents.

Claims (21)

1. A vehicle telematics device, comprising:

a processor;

a communications device coupled to the processor;

one or more sensor devices coupled to the processor; and

a memory coupled to the processor;

wherein the vehicle telematics device:

receives a set of unstructured vehicle driver data from the one or more sensor devices;

identifies anomalies in the set of unstructured vehicle driver data by using an unsupervised machine learning process that identifies relationships in the unstructured vehicle driver data;

calculates a driver risk score by using the anomalies identified in the set of unstructured vehicle driver data and the identified relationships in the uncategorized vehicle driver data; and

transmits the driver risk score to a remote server system by using the communications device;

wherein using the unsupervised machine learning process further comprises generating a plurality of isolation forests that distinguish clusters of the set of uncategorized vehicle driver data from anomalies in the set of uncategorized vehicle driver data.

2. The vehicle telematics device ofclaim 1, wherein the set of unstructured vehicle driver data is selected from a group consisting of vehicle speed, vehicle acceleration, vehicle deceleration, and vehicle swerving.

3. (canceled)

4. The vehicle telematics device ofclaim 1, wherein each isolation forest in the plurality of isolation forests comprises a plurality of isolation trees.

5. The vehicle telematics device ofclaim 1, wherein an isolation forest in the plurality of isolation forests without identified anomalies is labeled SAFE to indicate safe driving and wherein an isolation forest with identified anomalies is labeled UNSAFE to indicate unsafe driving.

6. The vehicle telematics device ofclaim 1, wherein the vehicle telematics device calculates the driver risk score by using an ensemble scoring process.

7. The vehicle telematics device ofclaim 6, wherein calculating the driver risk score by using an ensemble scoring process is evaluated by the vehicle telematics device by using the following expression:

$\mathrm{Risk}\mathrm{Score}\left(v_i\right)=\frac{100\times \mathrm{Num\_UNUSAFE}}{n}$

where v_iis a vehicle with a specific set of driver data, n is a total number of isolation forests in the plurality of isolation forests, and Num_UNSAFE is the number of isolation forests in the plurality of isolation forests labeled UNSAFE.

8. The vehicle telematics device ofclaim 7, wherein the driver risk score above a predetermined value indicates an unsafe driver.

9. The vehicle telematics device ofclaim 1, wherein the vehicle telematics device further calculates the driver risk score for a specific date range.

10. The vehicle telematics device ofclaim 1, wherein the communications device is a wireless device.

11. A method for driver risk scoring, the method comprising:

receiving a set of unstructured vehicle driver data from one or more sensor devices by using a vehicle telematics device, wherein the vehicle telematics device comprises a processor, a memory coupled to the processor, a communications device coupled to the processor, and the one or more sensor devices coupled to the processor;

identifying, using the vehicle telematics device, anomalies in the set of unstructured vehicle driver data by using an unsupervised machine learning process that identifies relationships in the unstructured vehicle driver data;

calculating, using the vehicle telematics device, a driver risk score by using the anomalies identified in the set of unstructured vehicle driver data and the identified relationships in the unstructured vehicle driver data; and

transmitting the driver risk score to a remote server system by using the communications device;

wherein using the unsupervised machine learning process further comprises generating a plurality of isolation forests that distinguish clusters of the set of uncategorized vehicle driver data from anomalies in the set of uncategorized vehicle driver data by using the vehicle telematics device.

12. The method ofclaim 11, wherein the set of vehicle driver data is selected from the group consisting of vehicle speed, vehicle acceleration, vehicle deceleration, and vehicle swerving.

13. (canceled)

14. The method ofclaim 1, wherein each isolation forest in the plurality of isolation forests comprises a plurality of isolation trees.

15. The method ofclaim 1, wherein an isolation forest in the plurality of isolation forests without identified anomalies is labeled SAFE to indicate safe driving and the isolation forest with identified anomalies is labeled UNSAFE to indicate unsafe driving by using the vehicle telematics device.

16. The method ofclaim 1, wherein calculating the driver risk score further comprises calculating the driver risk score by using an ensemble scoring process and the vehicle telematics device.

17. The method ofclaim 16, wherein calculating the driver risks core by using an ensemble scoring process is evaluated by the vehicle telematics device by using the following expression:

$\mathrm{Risk}\mathrm{Score}\left(v_i\right)=\frac{100\times \mathrm{Num\_UNUSAFE}}{n}$

where v_iis a vehicle with a specific set of driver data, n is a total number of isolation forests in the plurality of isolation forests, and Num_UNSAFE is the number of isolation forests in the plurality of isolation forests labeled UNSAFE.

18. The method ofclaim 17, wherein the driver risk score above a predetermined value indicates an unsafe driver.

19. The method ofclaim 11, further comprising calculating the driver risk score for a specific date range by using the vehicle telematics device.

20. The method ofclaim 11, wherein the communications device is a wireless device.

21.-59. (canceled)

Images