SYSTEM FOR PREDICTING A 5G USER LEVEL USING TAX LEVEL CHARACTERISTICS AND GRANGER CAUSALITY FOR CHARACTER SELECTION

Abstract

A system for predicting user-level behavior using control-level features and Granger causality includes a host device, a cloud computing server, and controllers. The controllers execute control logic that includes a predictive application (PA) which receives sensor data from the sensors, sends control-level and user-level data to an infrastructure and the cloud computing servers, and accesses prior knowledge data stored in the cloud computing servers' memory. Additional control logic performs a Granger causality check on the user-level and control-level data and uses a predictive model to generate a prediction from the fused user-level and control-level data.A prediction checker is applied to the prediction from the prediction model, and the PA enables the host device to make reliable predictions that allow the host device to adapt to the states of dynamic wireless communication networks, even when a best possible model is not fully trained.

Description

INTRODUCTION

[0001] The present disclosure relates to cellular communication systems and, in particular, to transient changes in the connection quality of cellular networks.

[0002] Cellular networks are increasingly used by a variety of devices, including vehicles. Vehicles often use cellular networks to communicate with other vehicles, infrastructure, and the like. To optimize performance, manage network traffic, and improve user experiences, machine learning techniques are often used for feature selection. However, traditional machine learning techniques often fail to capture relationships between control-plane information and user-plane metrics in dynamic environments.

[0003] Accordingly, while current systems and methods for optimizing performance, managing network traffic, and improving user experience through user-plane metric predictions achieve their intended purpose, there is a need for new and improved systems and methods for predicting user-plane metrics using control plane features that are highly customizable, perform well in dynamic environments, improve mobile communication reliability and user experience, and can be easily applied to new and existing platforms without increasing system complexity. SUMMARY

[0004] In several aspects of the present disclosure, a system for predicting a user plane using control plane features and Granger causality includes a host device having one or more sensors. The one or more sensors detect telematics information, telecommunications information, host device telemetry information, and host device position information. The system further includes one or more cloud computing servers and one or more controllers. Each of the one or more controllers includes a processor, memory, and one or more input / output (I / O) ports. The input / output ports are in communication with the one or more sensors and the one or more cloud computing servers. The memory stores programmatic control logic. The processor executes the programmatic control logic.The programmatic control logic includes a prediction application (PA). The PA includes at least a first, a second, a third, a fourth, a fifth, and a sixth control logic section. The first control logic receives sensor data from the one or more sensors. The second control logic sends control plane data and user plane data to an infrastructure and to the cloud computing servers via the input / output ports. The third control logic accesses prior knowledge data stored in the memory of the cloud computing server. The fourth control logic performs a Granger causality check on the user plane data and the control plane data. The fifth control logic uses a prediction model to generate a prediction from fused user plane data and control plane data. The sixth control logic applies a prediction checker to the prediction from the prediction model.

[0005] The PA enables the host device to make reliable predictions that allow the host device to adapt to conditions of dynamic wireless communication networks even when a best-fit model is not fully trained.

[0006] In another aspect of the present disclosure, the host device further defines a vehicle that communicates with the cloud computing servers via a wireless communication network. The first control logic further includes obtaining telematics information, telecommunications information, vehicle telemetry information, and vehicle position information from the one or more sensors.

[0007] In yet another aspect of the present disclosure, the second control logic further includes control logic for sending control plane data radio resource control (RRC) messages to an infrastructure including one or more cellular towers and to one or more cloud computing servers. The second control logic also receives control plane data RRC messages from the one or more cellular towers and from the one or more cloud computing servers and sends user plane data to the one or more cellular towers and to the one or more cloud computing servers. The second control logic further receives user plane data from the one or more cellular towers and from the one or more cloud computing servers.

[0008] In yet another aspect of the present disclosure, the third control logic further includes control logic for accessing prior knowledge data, including data obtained from sensors of the host device and additional devices, data obtained from Global Positioning System (GPS) satellites, and data obtained from infrastructure including one or more cellular towers. The third control logic further compares data control plane data RRC messages from the host device with the prior knowledge data to generate a partial input to the prediction checker.

[0009] In yet another aspect of the present disclosure, the fourth control logic further includes control logic for applying a Granger causality check to the control plane data RRC messages and to the user plane data. The Granger causality check outputs a P-value predicting when a change in the control plane data RRC messages caused a change in the user plane data. The Granger causality check is applied using at least three different P-value-based selections, including: a threshold P-value selection, a fixed-count P-value selection, and a conditional selection. In the threshold P-value selection, all control plane data RRC features that have a P-value less than a predetermined threshold are used.In fixed-count P-value selection, control-plane data RRC features are ranked from smallest to largest, and a predetermined number of features with the smallest P-values ​​are selected. In conditional selection, control-plane data RRC features from threshold P-value selection and from fixed-count P-value selection are compared. If it is determined that fixed-count P-value selection produced a P-value smaller than the P-value associated with threshold P-value selection, fixed-count P-value selection indicates Granger causality between the control-plane data RRC features and the user-plane data.If it is determined that the threshold P-value selection produced a P-value smaller than the P-value associated with the fixed-number P-value selection, the threshold P-value selection indicates Granger causality between the control-plane data RRC features and the user-plane data. If it is determined that the threshold P-value selection and the fixed-number P-value selection values ​​are equal to each other, then the threshold P-value selection and the fixed-number P-value selection equally indicate Granger causality between the control-plane data RRC features and the user-plane data.

[0010] In yet another aspect of the present disclosure, the fifth control logic further includes control logic for using the output of the Granger causality check to reduce the amount of data fed into the prediction model from a first amount of data to a second amount of data that is smaller than the first and prevents overfeeding of the prediction model. The fifth control logic also uses a long short-term memory (LSTM) model, a recurrent neural network (RNN), and / or an autoregressive integrated moving average (ARIMA) model to generate a multivariate prediction from user-level data and control-level data output from the Granger causality check.

[0011] In yet another aspect of the present disclosure, the fifth control logic further includes control logic for online training the PA by initializing the PA as a null feature model as a reference, sending control plane messages with Granger causality test result information and a prediction error root mean square (RMSE) of a current on-vehicle null feature model to the cloud computing server, evaluating a on-vehicle null feature model prediction RMSE training speed and predictions, and selectively providing a new model to the host device.

[0012] In yet another aspect of the present disclosure, the sixth control logic further includes control logic for receiving data indicating that a trigger event has occurred and for initializing the prediction checker with prior knowledge and with a user-level time series u(n) and a control-level time series C(n) as inputs. The sixth control logic further generates a weighted fused statistical prediction according to: p f = w1 · p s + w2 · p l , where w1 is a first weight applied to a statistical prediction p s generated by a statistical prediction model as applied to the prior knowledge and the user-level time series, w2 is a second weight different from the first weight applied to an LSTM prediction p l is applied, and p fis a fused prediction applied to the fused user plane and control plane data.

[0013] In yet another aspect of the present disclosure, the control logic of the PA further includes: control logic that applies non-uniform upsampling in a data collection and time series formulation for both the user plane and control plane data.

[0014] In yet another aspect of the present disclosure, the non-uniform upsampling further includes control logic that reduces bandwidth consumption and processing complexity from a first level to a second level less than the first level by obtaining user plane and control plane data only upon the occurrence of a trigger event and at one or more additional times after a predefined time δt after the trigger event.If a value of the control plane time series at time δt after the trigger event is greater than or equal to a value of the control plane time series at a subsequent trigger event, the non-uniform upsampling causes the PA to increment the control plane time series to collect data due to the occurrence of a subsequent trigger event, and then, if a value of the control plane time series at time δt after the trigger event is less than the value of the control plane time series at the subsequent trigger event, causes the PA to upsample the control plane time series.

[0015] In yet another aspect of the present disclosure, a method for predicting a user plane using control plane features and Granger causality includes: detecting telematics information, telecommunications information, host device telemetry information, and host device position information with one or more sensors mounted on a host device, using one or more cloud computing servers, and using one or more controllers. Each of the one or more controllers includes a processor, memory, and one or more input / output (I / O) ports. The input / output ports are in communication with the one or more sensors and the one or more cloud computing servers. The memory stores programmatic control logic. The processor executes the programmatic control logic.The programmatic control logic includes a prediction application (PA) having control logic that receives sensor data from the one or more sensors via the input / output ports and sends control plane data and user plane data to an infrastructure and to the cloud computing servers. The PA control logic further accesses prior knowledge data stored in the memory of the cloud computing servers, performs a Granger causality check on the user plane data and the control plane data, and uses a prediction model to generate a prediction from the fused user plane data and control plane data.The control logic of the PA also applies a prediction checker to the prediction from the prediction model and, via the PA, enables the host device to make reliable predictions that allow the host device to adapt to conditions of dynamic wireless communication networks even if a best-possible model is not fully trained.

[0016] In yet another aspect of the present disclosure, the host device further defines a vehicle that communicates with the cloud computing servers via a wireless communication network, and the control logic of the PA further includes control logic for obtaining telematics information, telecommunications information, vehicle telemetry information, and vehicle position information from the one or more sensors.

[0017] In yet another aspect of the present disclosure, the method further comprises sending control plane data radio resource control (RRC) messages to an infrastructure including one or more cellular towers and to one or more cloud computing servers, and receiving control plane data RRC messages from the one or more cellular towers and from the one or more cloud computing servers. The method further comprises sending user plane data to the one or more cellular towers and to the one or more cloud computing servers, and receiving user plane data from the one or more cellular towers and from the one or more cloud computing servers.

[0018] In yet another aspect of the present disclosure, the method further comprises accessing prior knowledge data including: data obtained from sensors of the host device and additional devices, data obtained from global positioning system (GPS) satellites, and data obtained from infrastructure including one or more cellular towers, and comparing data control plane data RRC messages from the host device with the prior knowledge data to generate a partial input to the prediction checker.

[0019] In yet another aspect of the present disclosure, the method further comprises applying a Granger causality test to the control plane data RRC messages and to the user plane data. The Granger causality test outputs a P-value predicting when a change in the control plane data RRC messages caused a change in the user plane data. The Granger causality test is applied using at least three different P-value-based selections, including: a threshold P-value selection, a fixed-count P-value selection, and a conditional selection. In the threshold P-value selection, all control plane data RRC features having a P-value less than a predetermined threshold are used.In fixed-count P-value selection, control-plane data RRC features are ranked from smallest to largest, and a predetermined number of features with the smallest P-values ​​are selected. In conditional selection, control-plane data RRC features from threshold P-value selection and from fixed-count P-value selection are compared. If it is determined that fixed-count P-value selection produced a P-value smaller than the P-value associated with threshold P-value selection, fixed-count P-value selection indicates Granger causality between the control-plane data RRC features and the user-plane data.If it is determined that the threshold P-value selection produced a P-value smaller than the P-value associated with the fixed-number P-value selection, the threshold P-value selection indicates Granger causality between the control-plane data RRC features and the user-plane data. If it is determined that the threshold P-value selection and the fixed-number P-value selection values ​​are equal to each other, then the threshold P-value selection and the fixed-number P-value selection equally indicate Granger causality between the control-plane data RRC features and the user-plane data.

[0020] In yet another aspect of the present disclosure, the method further comprises using an output of the Granger causality test to reduce an amount of data fed into the prediction model from a first amount of data to a second amount of data that is smaller than the first and prevents overfeeding of the prediction model, and using at least one of a long short-term memory (LSTM) model, a recurrent neural network (RNN), and an autoregressive integrated moving average (ARIMA) model to generate a multivariate prediction from user-level data and control-level data output from the Granger causality test.

[0021] In yet another aspect of the present disclosure, the method further comprises online training of the PA by initializing the PA as a null feature model as a reference and sending control plane messages with Granger causality test result information and a root mean square (RMSE) prediction error of a current on-vehicle null feature model to the cloud computing server. The online training of the PA further includes evaluating a vehicle-side null feature model prediction RMSE training speed and predictions, and selectively providing a new model to the host device.

[0022] In yet another aspect of the present disclosure, the method further comprises receiving data indicating that a trigger event has occurred, initializing the prediction checker with prior knowledge and with a user-level time series u(n) and a control-level time series C(n) as inputs, and generating a weighted fused statistical prediction according to: p f = w1 · p s + w2 · p l , where w1 is a first weight applied to a statistical prediction p s generated by a statistical prediction model as applied to the prior knowledge and the user-level time series, w2 is a second weight different from the first weight applied to an LSTM prediction p l is applied, and p f is a fused prediction applied to the fused user plane and control plane data.

[0023] In yet another aspect of the present disclosure, the method further comprises applying non-uniform upsampling in a data collection and time series formulation for both the user plane and control plane data, and reducing bandwidth consumption and processing complexity from a first level to a second level less than the first level by obtaining user plane and control plane data only upon occurrence of a trigger event and at one or more additional times after a predefined time δt after the trigger event.If a value of the control plane time series at time δt after the trigger event is greater than or equal to a value of the control plane time series at a subsequent trigger event, the method causes the PA to increment the control plane time series to collect data due to the occurrence of a subsequent trigger event, and then if a value of the control plane time series at time δt after the trigger event is less than the value of the control plane time series at the subsequent trigger event, the method causes the PA to upsample the control plane time series.

[0024] In yet another aspect of the present disclosure, a method for predicting a user plane using control plane features and Granger causality includes detecting telematics information, telecommunications information, host device telemetry information, and host device position information with one or more sensors mounted on a host vehicle and communicating with one or more cloud computing servers via a wireless communication network. The method further comprises using one or more controllers, each of the one or more controllers having a processor, a memory, and one or more input / output (I / O) ports, the input / output ports in communication with the one or more sensors and the one or more cloud computing servers. The memory stores programmatic control logic.The processor executes the programmatic control logic. The programmatic control logic includes a prediction application (PA). The PA includes control logic for obtaining telematics information, telecommunications information, vehicle telemetry information, and vehicle position information from the one or more sensors. The PA further includes control logic for transmitting, via the input / output ports, control plane data radio resource control (RRC) messages to an infrastructure including one or more cellular towers and to one or more cloud computing servers. The PA further includes control logic for receiving control plane data RRC messages from the one or more cellular towers and from the one or more cloud computing servers and transmitting user plane data to the one or more cellular towers and to the one or more cloud computing servers.The PA further includes control logic for receiving user plane data from the one or more cellular towers and from the one or more cloud computing servers. The PA further includes control logic for accessing prior knowledge data stored in memory of the cloud computing servers, wherein the prior knowledge data includes: data obtained from sensors of the host device and additional devices, data obtained from Global Positioning System (GPS) satellites, and data obtained from infrastructure including one or more cellular towers. The PA further includes control logic for comparing data control plane data RRC messages from the host device with the prior knowledge data to generate a partial input to the prediction checker.The PA further includes control logic for applying a Granger causality check to the control plane data RRC messages and to the user plane data. The Granger causality check outputs a P-value predicting when a change in the control plane data RRC messages caused a change in the user plane data. The Granger causality check is applied using at least three different P-value-based selections, including: a threshold P-value selection, a fixed-count P-value selection, and a conditional selection. In the threshold P-value selection, all control plane data RRC features with a P-value less than a predetermined threshold are used.In fixed-count P-value selection, control-plane data RRC features are ranked from smallest to largest, and a predetermined number of features with the smallest P-values ​​are selected. In conditional selection, control-plane data RRC features from threshold P-value selection and from fixed-count P-value selection are compared. If it is determined that fixed-count P-value selection produced a P-value smaller than the P-value associated with threshold P-value selection, fixed-count P-value selection indicates Granger causality between the control-plane data RRC features and the user-plane data.If it is determined that the threshold P-value selection produced a P-value smaller than the P-value associated with the fixed-number P-value selection, the threshold P-value selection indicates Granger causality between the control-plane data RRC features and the user-plane data. If it is determined that the threshold P-value selection and the fixed-number P-value selection values ​​are equal to each other, then the threshold P-value selection and the fixed-number P-value selection equally indicate Granger causality between the control-plane data RRC features and the user-plane data.The PA further includes control logic for using the output of the Granger causality check to reduce the amount of data fed into the prediction model from a first amount of data to a second amount of data that is smaller than the first and prevents overfeeding of the prediction model. The PA further includes control logic for using a long short-term memory (LSTM) model, a recurrent neural network (RNN), and / or an autoregressive integrated moving average (ARIMA) model to generate a multivariate prediction from user-level data and control-level data output by the Granger causality check.The PA further includes control logic for using a prediction model to generate a prediction from fused user-plane data and control-plane data, including online training of the PA by initializing the PA as a null-feature model as a reference and sending control-plane messages containing Granger causality test result information and a root-mean-square (RMSE) prediction error of a current on-vehicle null-feature model to the cloud computing server. The PA further includes control logic for evaluating the on-vehicle null-feature model's RMSE training speed and predictions, and selectively providing a new model to the host device.The PA further includes control logic for applying a prediction checker to the prediction from the prediction model, including: receiving data indicating that a trigger event has occurred, initializing the prediction checker with prior knowledge and with a user-level time series u(n) and a control-level time series C(n) as inputs, and generating a weighted fused statistical prediction according to: p. f = w1 · p s + w2 · p l , where w1 is a first weight applied to a statistical prediction p s generated by a statistical prediction model as applied to the prior knowledge and the user-level time series, w2 is a second weight different from the first weight applied to an LSTM prediction p l is applied, and p fa fused prediction applied to the fused user plane and control plane data. The PA further includes control logic for applying non-uniform upsampling in the data collection and time series formulation for both the user plane and control plane data, and reducing bandwidth consumption and processing complexity from a first level to a second level less than the first level by obtaining user plane and control plane data only upon the occurrence of a trigger event and at one or more additional times after a predefined time δt after the trigger event.If a value of the control plane time series at time δt after the triggering event is greater than or equal to a value of the control plane time series at a subsequent triggering event, the method causes the PA to increment the control plane time series to collect data due to the occurrence of a subsequent triggering event, and then, if a value of the control plane time series at time δt after the triggering event is less than the value of the control plane time series at the subsequent triggering event, the method causes the PA to upsample the control plane time series. The PA further includes control logic for enabling the host device to make reliable predictions that allow the host device to adapt to conditions of dynamic wireless communication networks even when a best-fit model is not fully trained.

[0025] Further areas of applicability will become apparent from the description provided herein. It should be understood that the description and specific examples are for purposes of illustration only and are not intended to limit the scope of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings described herein are for illustrative purposes only and are not intended to limit the scope of the present disclosure in any way; they show: Fig. 1 is a schematic diagram of a system for predicting a 5G user plane using control plane features and Granger causality for feature selection according to one aspect of the present disclosure; Fig. 2 is a flowchart showing a logical flow of a Granger causality testing section of the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; Fig. 3A is a flowchart illustrating a logical flow of a prediction model of the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; Fig. 3B is a graph illustrating exemplary control plane time series data of the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; Fig. 4 is a flowchart illustrating a logical flow of a prediction model section of the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; Fig. 5 is a schematic diagram showing an online training structure of the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; Fig. 6 is a flowchart illustrating a logical flow of a prediction checker of the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; Fig. 7A is a timeline diagram illustrating data collection and time series formulation in the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; Fig. 7B is a flowchart illustrating a logical flow of an upsampling process of the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure; and Fig. 8 is a schematic diagram illustrating the deployment of predictions and cloud-based model training in a localized area according to the system for predicting a 5G user plane using control plane features and Granger causality for feature selection of Fig. 1 according to one aspect of the present disclosure. DETAILED DESCRIPTION

[0027] The following description is merely exemplary in nature and is not intended to limit the present disclosure, application, or uses.

[0028] With reference to Fig. 1, a system 10 for predicting a user plane 12 using control plane features 14 and Granger causality for feature selection is shown in diagrammatic form. The system 10 includes one or more devices, such as vehicles 16, each having one or more on-board controllers 18. It should be appreciated that the devices may include any of a variety of wirelessly connected host devices, such as cellular phones, portable computers such as tablet computers, laptops, or the like, without departing from the scope or spirit of the present disclosure. However, for simplicity, clarity, and specificity, the following description refers to non-limiting examples where the devices or host devices are vehicles 16.

[0029] The vehicle-mounted controllers 18 are non-generalized electronic control devices that include a preprogrammed digital computer or processor 20, utilize a non-transitory computer-readable medium or memory 22 to store data such as control logic, software applications, instructions, computer code, data, lookup tables, etc., and include a transmit / receive device or input / output (I / O) ports 24. A computer-readable medium includes any type of medium accessible by a computer, such as read-only memory (ROM), random access memory (RAM), a hard disk drive, a compact disc (CD), a digital video disc (DVD), or any other type of storage.A "non-transitory" computer-readable memory 22 excludes wired, wireless, optical, or other communication connections that carry transient electrical or other signals. A non-transitory computer-readable memory 22 includes media where data can be permanently stored and media where data can be stored and later overwritten, such as a rewritable optical disk or an erasable storage device. Computer code includes any type of program code, including source code, object code, and executable code. The processor 20 is configured to execute the code or instructions. In vehicles 12, the controller 18 may be a dedicated Wi-Fi controller or an engine control module, a transmission control module, a body control module, an infotainment control module, etc.The input / output ports 24 are configured to communicate wirelessly using Wi-Fi protocols under IEEE 802.11x, cellular protocols such as Global System for Mobile Communications (GSM), Code Division Multiple Access (CDMA), Wireless Local Loop (WLL), General Packet Radio Service (GPRS), 1G, 2G, 3G, 4G Long Term Evolution (LTE), 5G, or the like.

[0030] The memory 22 may store one or more applications 26. An application 26 is a software program configured to perform a particular function or set of functions. The application 26 may include one or more computer programs, software components, sets of instructions, procedures, functions, objects, classes, instances, related data, or a portion thereof adapted for implementation in suitable computer-readable program code. The applications 26 may be stored in the memory 22 of the on-board controllers 18 in the vehicles 16 or in additional or separate memory, such as in a memory 22 of a cloud computing device, such as a cloud computing server 28.Examples of applications 26 include audio or video streaming services, games, browsers, social media, and an application for predicting a 5G user plane 12 using control plane 14 features and Granger causality for feature selection, or the like. For simplicity and to improve clarity, the application for predicting a 5G user plane 12 using control plane 14 features and Granger causality for feature selection will be referred to as propositional application (PA) 30.

[0031] In several aspects, the vehicles 16 generate, send, and receive data while operating. The data may include telematics information, telecommunications information, vehicle telemetry information, position information, data collected by sensors 31 or systems onboard the vehicle 16, or the like. The user plane 12 carries and communicates the data generated by the vehicles 16, such as data packets processed by protocols such as the Transmission Control Protocol (TCP), the User Graph Protocol (UDP), the Internet Protocol (IP), and the like. In contrast, in the control plane 14, a bearer protocol such as a Radio Resource Control (RRC) protocol processes signaling messages sent between one or more vehicles 16 or between end-user devices such as the one or more vehicles 16 and a base station or infrastructure such asCell phone towers 32 and the like.

[0032] Wireless communications, particularly those using GSM, CDMA, WLL, GPRS, 1G, 2G, 3G, 4G, and 5G, are subject to a variety of signaling issues that can affect the quality and reliability of the data transmitted and received. Accordingly, optimizing user layer 12 metrics such as bandwidth and throughput is important for managing communications network traffic, optimizing communications network performance, and improving the user experience.

[0033] In several aspects, the vehicles 16 send control plane data in the form of RRC messages 34 to infrastructure such as the cellular towers 32 and / or the cloud computing servers 28. Similarly, the vehicles 16 receive control plane data RRC messages 34 from cellular towers 32, cloud computing servers 28, and the like. As RRC messages 34 of the control plane 14 are received by cloud computing servers 28, cellular towers 32, etc., the RRC messages are both added to a database of prior knowledge 36 and compared to prior knowledge 36 stored in the memory 22 of the cloud computing servers 28 and / or the cellular towers 32 and / or in the on-board memory 22 of the vehicle 16. In several aspects, the prior knowledge 36 is a compilation of data received from sensors 31 of a host device such as the cellular towers 32 and / or the cloud computing servers 28.a host vehicle 16, from sensors 31 within additional devices or vehicles 16', and from sensors 31 located on or acting as part of infrastructure such as weather satellites, GPS satellites 37, or the like. However, it should be recognized that the data from additional vehicles 16' may be filtered or otherwise restricted in the prior knowledge database 36 in operation based on predefined and / or changeable metrics. Data from additional vehicles 16' may be obtained from these additional vehicles 16' at a variety of different locations and over a variety of different time periods.It will be appreciated that in order to make accurate and relevant predictions, the system 10 is best operated using only data from additional vehicles 16' generated by such additional vehicles 16' that are in physical proximity and / or temporal proximity to the host vehicle 16. Accordingly, data from GPS satellites, weather satellites, or the like may be relevant to the host vehicle 16 for limited periods of time or for limited physical distances. Accordingly, the data from GPS and weather satellites are filtered accordingly.

[0034] The exact definition of "near" may vary depending on the situation, but should be understood to mean that the host vehicle 16 and additional vehicles 16' are within a predefined physical and / or temporal distance that allows data from the additional vehicles 16' to remain relevant to the host vehicle 16. For example, on a curve on a highway, while traveling at seventy miles per hour and in changeable weather conditions, data generated by additional vehicles 16' will have a half-life that is relatively short compared to data from additional vehicles 16' on a straight roadway traveling at fifteen miles per hour in sunny, warm, dry weather.

[0035] Vehicles 16 also send user plane data 38 to infrastructure such as cellular towers 32 and / or cloud computing servers 28. Similarly, vehicles 16 receive messages containing user plane data 38 from cellular towers 32, cloud computing servers 28, and the like. In several aspects, user plane data 38 and RRC messages 34 are subjected to a Granger causality check 40. Granger causality check is used in assisting feature selection 42 to generate user plane predictions 12 for 5G communications. Granger causality check 40 acts as a predictor, where if Y is "Granger-caused" by X, then past values ​​of X are assumed to contain information that helps predict the value of Y. The Granger causality test 40 outputs a P-value that has a different size or value for different control level 14 data.In some aspects, if the P-value 44 is less than a predefined significance level, it is inferred that X has a predictive influence on Y, thereby indicating Granger causality. While the predefined significance level may vary from application to application, in some examples, the significance level is less than 0.05.

[0036] With reference to Fig. 2, Fig. 3A and Fig. 3B and with continued reference to Fig. 1, a method 200 for using the Granger causality test 40 to predict features is shown in more detail in flowchart form. In one example, the Granger causality test 40 is defined as: u(n)=a0+a1 u(n−1)+…+am u(n−m)+b1 c(n−1)+…+bm c(n−m)+e1(n); u(n)=c0+c1 u(n−1)+…+cm u(n−m)+e2(n);

[0037] Then, a feature test is performed, testing the null hypothesis that b1 = b2 = ...bm = 0. If the null hypothesis is rejected, control-level features c(n) Granger-cause one or more user-level features u(n), where c(n)s are considered as the features used to improve the prediction accuracy of u(n).

[0038] In some examples, at least three different feature selection methods are used based on the p-values ​​of the Granger causality test 40. Specifically, a threshold Φ selection method 202, a fixed number selection method 204, and a conditional selection method 206 are used.

[0039] The method 202 using threshold Φ selects all c(n)s with a P-value 44 that is smaller than the threshold Φ 208 according to: C_1={c(n):p(c(n))<ϕ},|C_1|=N1, where N1 is the feature number 1 in block 210. It is recognized that the threshold Φ is a predefined value that may vary from application to application and situation to situation, but represents a value chosen to ensure, with a given degree of accuracy and confidence, that, based on their P-values, the c(n)s are accurate predictors.

[0040] In contrast, the method 204 calculates all p values ​​from the control plane 14 data using a specified number and ranks 212 the p values ​​from smallest to largest. The method 204 then selects a predetermined number of features having the smallest p values ​​using a specified number. In one example, the method 204 selects the three (3) features having the smallest p values ​​using a specified number, however, the exact set of features selected may vary without departing from the scope or intent of the present disclosure. The selection process may be represented as selecting an N2 number of c(n)s with the lowest p values ​​according to: C_2={c(n):classify(p(c(n)))<=N2},|C_2|=N2, where N2 is the feature number 2 in block 214.

[0041] Finally, the conditional procedure 206 compares the values ​​of N1 and N2 in block 216 to determine whether Granger causation exists. More specifically: If N2 <N1,C_3=C_2 If N2>N1,C_3=C_1

[0042] Accordingly, C_3 is theoretically a Granger cause of u(n) under the limits of the available feature numbers. That is, if N1 < N2, then method 204 using a fixed number has produced the most accurate result, whereas if N2 < N1, then method 202 using threshold Φ has produced the most accurate result. Similarly, if N2 = N1, both the selection method 202 using threshold Φ and the selection method 204 using a fixed number have the same probability of producing accurate results.

[0043] Now, with particular reference to Fig. 3A and Fig. 3B and with continued reference to Fig. 1 and Fig. 2 Twenty (20) different time series data points of the control plane 14 are collected and represented as the function c(t). Granger causality tests 40 determine the probabilities of the c(n)s being Granger-caused in a user-plane time series u(n). As previously stated, the lower a P-value 44, the higher the probability that u(n) is Granger-caused by c(n). Granger causality can therefore be represented as the function P(GC|p) ∝ f(p) in block 46, where f is a decreasing function.

[0044] With reference to Fig. 4 and with continued reference to Fig. 1- Fig. 3, the system 10 uses a predictive model 300 to predict short-term data rates 301. The predictive model 300 starts in block 302 using user plane 12 data (represented as u(i : i + 10)) and control plane 14 data (represented as C(i : i + 10)) as inputs. In several aspects, the Granger causality check 40 is used to verify the input data prior to use in the predictive model 300 to reduce the potential for overfeeding the predictive model 300. Overfeeding the predictive model 300 may cause the predictive model 300 to produce an output that is effectively meaningless because the predictive model 300 itself does not determine which input data is actually useful for the vehicle 16 or the wireless communication network functionality.That is, the Granger causality check 40 reduces the amount of data fed into the predictive model 300 from a first amount of data to a second amount of data that is smaller than the first amount of data to avoid overfeeding the predictive model 300. In addition to preventing overfeeding—a factor that can weaken the look-ahead performance of the system 10—the Granger causality check 40 provides the ability to select "relevant" input variables, thereby improving the look-ahead accuracy of the model compared to models without such inputs.

[0045] The data from the user and control planes 12, 14 are normalized 304 and processed by a prediction algorithm 306. The prediction algorithm 306 shown is a long short-term memory (LSTM), however, the LSTM may be replaced by other feature-based time series prediction models such as a recurrent neural network (RNN), an autoregressive integrated moving average (ARIMA), or the like, without departing from the scope or intent of the present disclosure. The prediction model 300 then performs inverse normalization 308 and generates an output 310 from the user plane 14 (represented as u'(i + 11)), which is then compared to a target output 312 from the user plane 14 (represented as u(i + 11)) to generate a root mean square error (RMSE) 314.In several aspects, the RMSE 314 defines a metric used to measure a prediction error between the prediction output 310 of the user layer 14 and the target output 312, which is the ground truth.

[0046] With reference to Fig. 5 and with continued reference to Fig. 1- Fig. 4 schematically shows an online training structure 400 for the PA 30. The PA 30 is initialized as a zero-feature model (0-feature model) 402, which operates on the controller 18 of the host vehicle 16. The zero-feature model 402 serves as a reference against which further features are subsequently compared. The host vehicle 16 sends current messages from the control plane 14 with result information from the Granger causality test 40 and the predictive RMSE 314 of the current vehicle-side zero-feature model 402 to the cloud computing server 28. Multiple models are trained in parallel in the cloud computing server 28. Depending on comparisons of the predictive RMSE 314 and evaluations of the training speed, the cloud computing server 28 decides whether and when to provide a new model to the vehicle 16, thereby replacing the current model 404 with a new model.Training based on a cloud computing server 28 enables the host vehicle 16 to make reliable predictions even when the best possible model is not fully trained. It is recognized that the amount of features in a given model correlates with relative speeds of training and deployment processes. That is, a single-feature model 406 can be trained at a greater speed than a two-feature model (2-feature model) 408, and so on. In further exemplary applications of the online training structure 400, results of the single-feature model 406 are compared with the two-feature model 408 up to an N-feature model 410. If the two-feature model 408 achieves an empirically better result than the single-feature model 406, then the two-feature model 408 is deployed to the host vehicle 16.Similarly, if the N-feature model 410 achieves a better result than an N-1 feature model (not specifically shown), the N-feature model 410 is provided to the host vehicle 16. It is recognized that training an N-feature model 410 is computationally more complex and therefore slower than the single-feature model 406, the dual-feature model 408, etc. However, the N-feature model 410 is more accurate and reliable than the N-1 feature model because the N-feature model 410 is optimized from Granger causality tests 40 of each of the sub-N feature models 410. Accordingly, the N-feature model 410 is effectively an optimized fine-tuned feature model for the host vehicle 16.

[0047] With reference to Fig. 6 and with continued reference to Fig. 1- Fig. 5, a prediction checker 500 is shown in flowchart form. The prediction checker 500 leverages insights gained from field experimentation to detect and mitigate the impact of prediction outliers on the predictions of the PA 30 as time progresses. In several aspects, the prediction checker 500 is initialized upon receipt of data indicating the occurrence of a trigger event (n) and generates a prediction (m). Trigger events (n) may be any of a variety of different types of changes in situations of the user and / or control plane 12, 14. In one non-limiting example, changes in the control plane 14 may directly change or impact the state of the user plane 12 when a handover event occurs between multiple cellular towers 32.That is, during a handover between cellular towers 32, a bandwidth of communications between the host vehicle 16 and the cellular towers 32 may decrease from a first level to a second level that is smaller than the first. Accordingly, line-of-sight communication between the host vehicle 16 and cellular towers 32, or the like, may be hampered by the presence of large buildings, tunnels, mountains, etc., causing a change in the activity of the control plane 14. That is, an RSRP (Reference Signal Reception Power), an RSRQ (Reference Signal Reception Quality), and / or an RSSI (Received Signal Strength Indicator) may be negatively affected by a blocked line-of-sight, a handover of cellular towers 32, and the like.In another example, when a speed of the host vehicle 16 is low, multiple or frequent handovers may occur between cellular towers 32 within a small physical distance, causing bandwidth degradation and changes in the experience of the user plane 12. In yet another example, when a bandwidth-intensive application, such as a video streaming service application accessed via a human-machine interface (HMI) of a host vehicle 16 or via user devices in electronic communication with the controller 18, terminates, the amount of bandwidth released for use by other systems in the control plane 14 may change dramatically, resulting in changes in the functionality of the user plane 12. Accordingly, the prediction checker 500 checks and adjusts the prediction results of the LSTM prediction algorithm 306.

[0048] Accordingly, the prediction checker 500 uses prior knowledge 36 as well as the user-level time series u(n) and the control-level time series C(n) as inputs. In particular, the user-level time series u(n) and prior knowledge 36 are used as inputs to a statistical prediction model 502. The statistical prediction model 502 may be any of a variety of different statistical prediction model types, including, but not limited to, null hypothesis statistical testing (NHST) models, multivariate Kalman filters, multivariate Bayesian analyses, or the like. The statistical prediction model 502 generates a statistical prediction 504 or a statistical prediction result p s. Accordingly, the control plane time series data C(n) and the user plane time series u(n) are used as inputs to the LSTM prediction algorithm 306. The LSTM prediction algorithm 306 generates a statistical LSTM prediction 506 or an LSTM prediction result p l . The statistical prediction results p s are used to further estimate errors of the LSTM prediction results p l to reduce and make a final prediction p f 508, which is defined by the following: pf=w1⋅ps+w2⋅pl, after which prediction weights (w1 and w2) 510 are adjusted to minimize the following error root mean square (MSE) loss function: L(w1,w2)=E(T−(w1⋅ps+w2⋅pl))2.

[0049] The weights w1 and w2 are then updated using a gradient descent algorithm according to the following: w1=w1−a⋅dLdw1,w2=w2−a⋅dLdw2.

[0050] Then the gradients themselves are calculated according to the following: dLdw1=−2E(T−(w1⋅ps+w2⋅pl)⋅ps) dLdw2=−2E(T−(w1⋅ps+w2⋅pl)⋅pl), where T is the ground truth and a is a learning rate. The validation process is performed at specific trigger events (n), where prior knowledge 36 has a significant influence on the change in data rate. It is recognized that the weights w1, w2, w(j-1), w(j) may vary from application to application and model to model, but it should be understood that they assist the prediction checker 500 and the PA 30 by selecting which of the single-feature, dual-feature, or up to N-feature models 406, 408, 410 is the appropriate model to address a given trigger event (n).

[0051] With reference to Fig. 7A and Fig. Figure 7B uses PA 30 to further improve prediction accuracy by using non-uniform or non-uniform upsampling in a data collection and time series formulation. Specifically, control plane 14 data is collected from RRC messages at the client or host vehicle 16. Key control plane 14 information (C) is extracted from the RRC messages and monitored according to the following: c(n)={1,C detected0,C not detected, while the user level time series can be represented as: u(n)=Thput(n¯:n)n−n¯, where Thput(n:n) is the data throughput or bandwidth of the system. In some aspects, when c(n) = 1, C is detected, the C detected may be a handover between cellular towers 32 or the like. While uniform sampling can be used to some extent, it has several disadvantages in the present context. Specifically, with uniform sampling, it is difficult to align sample points to the exact occurrence of control plane 14 events, thereby introducing additional measurement uncertainty and increasing the potential for error. Likewise, with uniform sampling, a high sampling frequency results in a large volume of data, high bandwidth consumption, and increased computational complexity in the predictive model. Non-uniform or non-uniform sampling offers certain advantages, namely the ability to sample based on the occurrence of control plane 14 events.However, because non-uniform or non-uniform sampling is triggered solely by the occurrence of a specific control plane 14 event, user plane 12 changes between the occurrence of multiple control plane 14 trigger events may be lost. That is, U(n) represents an average data rate between two consecutive control plane 14 events, but fails to capture changes that occur in between. Accordingly, PA 30 applies upsampling at locations of a significant user plane 12 change.

[0052] In a data collection and time series formulation, the PA 30 double-samples the user-plane time series u(n) using an additional sampling point at a time δt after each of the control plane 14 events. δt is obtained by estimating the time at which the most significant fluctuation in throughput occurs following a particular control plane 14 event. Incorporating the δt -based sample data into the sampling procedure improves the prediction accuracy. Now, with particular reference to Fig. 7B, an upsampling method 600 is shown in more detail. In block 602, the PA 30 receives data of a control plane time series c(n). In block 604, the PA 30 appends the upsampled control plane time series c_upsampled to the control plane time series c(n). In block 606, the PA 30 compares the value of the control plane time series c(n) at time δt after a trigger event (n) with the value of the control plane time series c(n) at a subsequent trigger event (n+1). If the value of the control plane time series c(n) at time δt after the trigger event (n) is greater than or equal to the value of the control plane time series c(n) at the subsequent trigger event (n+1), the control plane time series c(n) is incremented to the collected data after the occurrence of a subsequent trigger event n++.However, if the value of the control plane time series c(n) at time δt after the trigger event (n) is smaller than the value of the control plane time series c(n) at the subsequent trigger event (n + 1), an upsampled control plane time series c_upsampled is obtained in block 608 and then applied to the control plane time series c(n).

[0053] With reference to Fig. 8 and with continued reference to Fig. 1- Fig.7B, models generated by the PA 30 may be shared between a host vehicle 16 and additional vehicles 16' having similar characteristics. For example, additional vehicles 16' traveling along the same or substantially similar paths at a particular physical location and during the same time of day may have the same prediction model provided to each of the similarly located additional vehicles 16', and vice versa. That is, one or more of a plurality of vehicles 16 may generate a user plane 12 and / or control plane 14 prediction through the PA 30, which is then uploaded to the cloud computing server 28. The cloud computing server 28 may then provide the user plane 12 and / or control plane 14 prediction to the additional vehicles 16' in the same vicinity.For each vehicle 16 that uploads data to the cloud computing server 28, additional data is used to further train the model in the cloud. That is, as each vehicle 16 reports to the cloud computing server 28, the upload data rate, GPS data, speed, and the like are uploaded and used as additional training information to further refine and increase the accuracy of predictions generated by the PA 30.

[0054] A system and method for predicting a user plane 12 using control plane 14 features and Granger causality for feature selection of the present disclosure offers several advantages. These include the ability to accurately capture complex causal relationships between control plane 14 information and user plane 12 metrics in a highly dynamic 5G environment where changes to the control plane 14 can directly impact the user plane 12. Granger causality provides a robust and reasoned means for selecting features for predictive models in a practical, reliable, and computationally low-impact manner while optimizing performance, improving cellular communication reliability and the user experience, working well in dynamic environments, and being easily adaptable to new and existing platforms without increasing system complexity.

[0055] The description of the present disclosure is merely exemplary in nature, and variations that do not depart from the gist of the present disclosure are intended to be within the scope of the present disclosure. Such variations should not be considered a departure from the spirit and scope of the present disclosure.

Claims

[1] System for predicting a user plane using control plane features and Granger causality, comprising: a host device having one or more sensors, the one or more sensors detecting telematics information, telecommunications information, host device telemetry information, and host device position information; one or more cloud computing servers; one or more controllers, each of the one or more controllers comprising a processor, a memory, and one or more input / output ports (I / O ports), the input / output ports being in communication with the one or more sensors and the one or more cloud computing servers; the memory stores programmatic control logic; the processor executes the programmatic control logic; the programmatic control logic includes a predictive application (PA), and the PA comprises: first control logic for obtaining sensor data from the one or more sensors; a second control logic for sending control plane data and user plane data to an infrastructure and to the cloud computing servers via the input / output ports; a third control logic for accessing prior knowledge data stored in the memory of the cloud computing servers; a fourth control logic for performing a Granger causality test on the user-level data and the control-level data; a fifth control logic for using a prediction model to generate a prediction from fused user-plane data and control-plane data; and a sixth control logic for applying a prediction checker to the prediction of the prediction model; wherein the PA enables the host device to make reliable predictions that allow the host device to adapt to conditions of dynamic wireless communication networks even when a best-fit model is not fully trained. [2] The system of claim 1, wherein the host device further comprises a vehicle that communicates with the cloud computing servers via a wireless communication network; and the first control logic further comprises: Obtaining telematics information, telecommunications information, vehicle telemetry information, and vehicle position information from the one or more sensors. [3] The system of claim 1, wherein the second control logic further comprises: Sending control plane data radio resource control (RRC) messages to an infrastructure including one or more cellular towers and to one or more cloud computing servers; Receiving control plane data RRC messages from the one or more cellular towers and from the one or more cloud computing servers; Sending user plane data to the one or more cell towers and to the one or more cloud computing servers and Receiving user plane data from the one or more cell towers and from the one or more cloud computing servers. [4] The system of claim 1, wherein the third control logic further comprises: Accessing prior knowledge data, including: data obtained from sensors of the host device and additional devices, data obtained from Global Positioning System (GPS) satellites, and data obtained from infrastructure including one or more cell towers; and Comparing data control plane data RRC messages from the host device with the prior knowledge data to generate a partial input to the prediction checker. [5] The system of claim 3, wherein the fourth control logic further comprises: Applying a Granger causality test to the control plane data RRC messages and to the user plane data, wherein the Granger causality test outputs a P-value predicting when a change in the control plane data RRC messages caused a change in the user plane data; where the Granger causality test is applied using at least three different selections based on the P-value, which include: a P-value selection using a threshold, a P-value selection using a fixed number and a conditional selection, where in the threshold P-value selection, all control plane data RRC features are used that have a P-value smaller than a predefined threshold; in the P-value selection, a specified number of control-level data RRC features are ranked from smallest to largest, and a specified number of features having the smallest P-values are selected; and in the conditional selection, control plane data RRC features from the P-value selection using a threshold and from the P-value selection using a fixed number are compared; where then, if it is determined that the P-value selection by a specified number produced a P-value that is less than the P-value associated with the P-value selection by threshold, the P-value selection by a specified number indicates Granger causality between the control plane data RRC features and the user plane data; if it is determined that the threshold P-value selection produced a P-value that is less than the P-value associated with the threshold P-value selection, the threshold P-value selection indicates Granger causality between the control plane data RRC features and the user plane data, and then, if the threshold P-value selection and the fixed-number P-value selection values are determined to be equal to each other, the threshold P-value selection and the fixed-number P-value selection equally indicate Granger causality between the control-plane data RRC features and the user-plane data. [6] The system of claim 5, wherein the fifth control logic further comprises: Using an output of the Granger causality test to reduce an amount of data fed into the predictive model from a first amount of data to a second amount of data that is smaller than the first and prevents overfeeding of the predictive model; Using a long short-term memory (LSTM) model and / or a recurrent neural network (RNN) and / or an autoregressive integrated moving average (ARIMA) model to generate a multivariate prediction from user-level data and control-level data output by the Granger causality test. [7] The system of claim 6, further comprising: a control logic for online training of the PA through the following: Initializing the PA as a zero feature model as a reference; Sending control plane messages containing Granger causality check result information and a prediction error root mean square (RMSE) of a current on-vehicle zero feature model to the cloud computing server; Evaluating a vehicle-side zero-feature model prediction RMSE training speed and predictions and optionally providing a new model to the host device. [8] The system of claim 6, wherein the sixth control logic further comprises: Receiving data indicating that a trigger event has occurred; Initializing the prediction checker with prior knowledge and with a user-level time series u(n) and a control-level time series C(n) as inputs and Generate a weighted fused statistical prediction according to: p f = w1 · p s + w2 · p l , where w1 is a first weight applied to a statistical prediction p sgenerated by a statistical prediction model as applied to the prior knowledge and the user-level time series, w2 is a second weight different from the first weight applied to an LSTM prediction p l is applied, and p f is a fused prediction applied to the fused user-plane and control-plane data. [9] The system of claim 1, wherein the control logic of the PA further includes: a control logic that applies non-uniform upsampling in a data collection and time series formulation for both the user-plane and control-plane data. [10] The system of claim 9, wherein the non-uniform upsampling further comprises: control logic that reduces bandwidth consumption and processing complexity from a first level to a second level that is smaller than the first level by obtaining user plane and control plane data only upon the occurrence of a trigger event and at one or more additional times after a predefined time δt after the trigger event; and If a value of the control plane time series at time δt after the trigger event is greater than or equal to a value of the control plane time series at a subsequent trigger event, then cause the PA to increment the control plane time series to collect data due to the occurrence of a subsequent trigger event, and then cause the PA to upsample the control plane time series if a value of the control plane time series at time δt after the trigger event is less than the value of the control plane time series at the subsequent trigger event.

Images