Driving data classification device

Abstract

To provide a travel data classification device capable of automatically classifying travel data without a manual operation.SOLUTION: A travel data classification device 10 comprises a storage 12 for storing a learned model to estimate a track name corresponding to a point where a vehicle is traveling when travel data is collected by using input data being time-series data of an explanatory variable which includes a plurality of types of movement average values of output torque of a prime mover of the vehicle. The learned model is learned through supervised learning using training data which includes: input data: and correct answer data indicating the track names corresponding to the points where the vehicle was traveling when the travel data was collected. The travel data classification device 10 includes a processor 11 for executing, using the learned model stored in the storage 12: estimation processing to output estimated values corresponding to the track names from the input data; and classification processing to perform classification by applying the travel data to the track names based on the estimated values.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] This invention relates to a driving data classification device that classifies driving data collected while a vehicle is in motion.

Background Art

[0002] Data may be collected by sensors mounted on a vehicle while the vehicle is in motion. Patent Document 1 discloses a test vehicle traveling on a test course. Patent Document 1 also discloses a management system that displays images representing each test vehicle in a reduced view of a test course on which a plurality of test vehicles are traveling.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] When analyzing driving data collected while a vehicle is in motion, the collected driving data may be classified according to the type of road. Classifying by visually checking the content of the collected driving data takes time and manpower.

Means for Solving the Problems

[0005] Hereinafter, means for solving the above problems and their operational effects will be described. A driving data classification device for solving the above problems classifies driving data including the output torque of the prime mover of the vehicle collected by the sensor while a vehicle equipped with the sensor is in motion, The test course was pre-classified into sections. by applying it to a plurality of road names. Multiple track names include both straight sections and uphill sections.The driving data classification device comprises a storage device and a processing circuit. The storage device stores a trained model that estimates the name of the road corresponding to the point where the vehicle was traveling when the driving data was being collected, using input data which is time-series data of explanatory variables including multiple types of moving average values ​​of the output torque of the vehicle's prime mover. The trained model is supervised learning using training data which includes the input data and ground truth data indicating the name of the road corresponding to the point where the vehicle was traveling when the driving data was being collected. The processing circuit uses the trained model stored in the storage device to perform an estimation process that outputs an estimated value corresponding to the road name from the input data and a classification process that classifies the driving data by applying it to the road name based on the estimated value. [Effects of the Invention]

[0006] The above-mentioned driving data classification device can automatically classify driving data without human intervention. [Brief explanation of the drawing]

[0007] [Figure 1] Figure 1 is a schematic diagram showing the configuration of a driving data classification device. [Figure 2] Figure 2 is a graph illustrating one example of how training data is acquired. [Figure 3] Figure 3 is an explanatory diagram illustrating the concept of a long-term memory neural network. [Figure 4] Figure 4 is a flowchart showing the learning process flow. [Figure 5] Figure 5 is a flowchart showing the processing flow for estimation and classification. [Figure 6] Figure 6 is a graph showing the estimated values. [Modes for carrying out the invention]

[0008] Below, one embodiment of the driving data classification device will be described with reference to Figures 1 to 6. <Configuration of the driving data classification device 10> Figure 1 shows the configuration of the driving data classification device 10. The driving data classification device 10 is a computer. The driving data classification device 10 comprises a processor 11, which is a processing circuit, a storage device 12, which is a storage device, and a memory 13, which is an auxiliary storage device. The driving data classification device 10 also comprises an input device 14, a display device 15, an input / output interface 16, and a communication device 17. The input device 14 is, for example, a keyboard. The input device 14 may also be a touch panel. Each of these components—the processor 11, storage device 12, memory 13, input device 14, display device 15, input / output interface 16, and communication device 17—is connected via a bus 18.

[0009] The storage 12 stores a program. The processor 11 executes the program stored in the storage 12 and performs various processes. The driving data classification device 10 classifies the driving data. The driving data is data collected by sensors while a vehicle equipped with sensors is in motion. Specifically, the driving data includes time-series data of output torque detected by a torque sensor mounted on the vehicle. The driving data classification device 10 classifies this driving data by matching it to one of several pre-set driving route names. In short, the driving data classification device 10 estimates which driving route the driving data was acquired on. Then, it classifies the driving data by matching it to a driving route name according to the estimation result.

[0010] <Output Torque> The torque sensor is connected to the vehicle's drive shaft. The drive shaft is the output shaft of the vehicle's prime mover. For example, the prime mover is an engine. The prime mover may also be an electric motor. Furthermore, the prime mover may be a hybrid system combining an engine and an electric motor.

[0011] The drive shaft is connected to the drive wheels. The torque sensor detects the torque of the drive shaft. This torque, which is output from the prime mover, is referred to here as output torque. When a vehicle is in motion, even if the prime mover output is constant, the output torque changes depending on the road conditions. For example, in vehicle development, driving data including such output torque is sometimes collected and time-series data is analyzed.

[0012] <Classification of driving data> When analyzing driving data, the collected data may be classified by the type of road. For example, if driving data is collected on a test course that includes multiple sections such as sharp curves, straightaways, and uphill roads, the data may be classified by section. Similarly, if driving data is collected over long distances including mountain roads, highways, and suburban roads, the data may be classified by road type, such as "mountain roads," "unpaved roads," "highways," and "general roads." The name of the road may also include the name of the region where the data was collected. Driving data may also be classified by the region where it was collected. For example, "Jibuzaka, Nagano Prefecture," "Hakone, Shizuoka Prefecture," "Mt. Rokko, Hyogo Prefecture," and "Mt. Ibuki, Shiga Prefecture."

[0013] Traditionally, the classification of such driving data has been done manually by visually checking the data displayed on a display device in the form of a graph or other format. Figure 2 is a graph of time-series data of output torque. For example, when classifying manually, the worker determines from the shape of the graph and the magnitude of the values ​​that the driving data up to time t_1 is data from driving on a straight road. The worker then classifies the driving data up to time t_1 into classification number "1", which indicates the road name "straight". The worker then determines from the shape of the graph and the magnitude of the values ​​that the driving data after time t_1 is data from driving on an uphill road. The worker then classifies the driving data after time t_1 into classification number "2", which indicates the road name "uphill road".

[0014] Thus, visually checking and classifying the content of driving data takes time and effort. In particular, when a large amount of unclassified driving data is accumulated, the labor required for classification also becomes enormous.

[0015] <Trained model> In the storage 12 of the driving data classification device 10, a trained model for estimating the road name corresponding to the location where the vehicle was traveling when collecting the driving data based on the driving data is stored. The driving data classification device 10 automatically classifies unclassified driving data using this trained machine learning model. A large amount of unclassified driving data to be classified is accumulated in the storage 12 of the driving data classification device 10. Such unclassified driving data can be moved to the storage 12 of the driving data classification device 10 via a recording medium. Also, driving data may be transmitted from the vehicle to the driving data classification device 10 via the communication device 17.

[0016] The trained model is trained with supervised learning using training data including driving data and correct answer data indicating the road name corresponding to the location where the vehicle was traveling when collecting the driving data. The training data can be created by a person confirming and assigning correct answer data to the driving data as in the conventional case, as exemplified by referring to FIG. 2.

[0017] The driving data classification device 10 uses a long short-term memory neural network that can be handled while retaining information on the transition along the time axis of time-series data as a trained model. The long short-term memory neural network is a so-called LSTM (Long Short-Term Memory) neural network. The LSTM neural network is a type of recurrent neural network.

[0018] The output torque collected during running undergoes changes according to the condition of the road surface on which the vehicle is running. The time-series data of the output torque is information useful for estimating the name of the road on which the vehicle is running. The running data classification device 10 uses data including multiple types of moving average values of the output torque as explanatory variables.

[0019] In FIG. 2, the moving average value of the output torque is indicated by a dashed line. The multiple types of moving average values are moving average values with different numbers of data for calculating the average value. For example, the output torque is calculated every 0.01 seconds. One of the multiple types of moving average values is the 100-second moving average value of the output torque. One of the multiple types of moving average values is the 20-second moving average value of the output torque. One of the multiple types of moving average values is the 5-second moving average value of the output torque. For example, the explanatory variables include several such types of moving average values. The running data may include data other than the output torque as explanatory variables. For example, as explanatory variables other than the output torque, information on the steering angle, accelerator opening, vehicle speed, etc. may be included. All of these may be included as explanatory variables, or some of these may be included as explanatory variables. Information other than these may be included as explanatory variables.

[0020] The running data classification device 10 inputs the time-series data for the default time of the explanatory variables as input data X into the learned model. For example, when the explanatory variables consist of the steering angle, accelerator opening, vehicle speed, 100-second moving average value, 20-second moving average value, and 5-second moving average value of the output torque, the explanatory variables include six types of data.

[0021] For example, when classifying one hour's worth of driving data, the driving data classification device 10 extracts explanatory variables from the driving data within a predetermined time window WdS, as shown in Figure 2. The driving data classification device 10 then uses the time-series data of the extracted explanatory variables as input data X to be input into the trained model. For example, if the predetermined time is 10 seconds and the explanatory variables are collected every hundredth of a second, the input data X for 10 seconds is a set of 1000 explanatory variables collected continuously in time series. Specifically, the input data X is a set from the first collected explanatory variable X(1) to the last collected explanatory variable X(1000) within the 10 seconds.

[0022] Each explanatory variable contains multiple types of information. For example, in the case of explanatory variables consisting of six types of data as described above, each collected data is a six-dimensional vector consisting of six values. In this case, the input layer of the neural network that makes up the trained model has six nodes.

[0023] Figure 3 schematically illustrates the structure of a recurrent neural network. In Figure 3, the vertical arrows indicate the direction of forward propagation in the neural network where explanatory variables are input. In Figure 3, "n" indicates the time-series order of the explanatory variables in the input data X.

[0024] The neural network shown on the far right of Figure 3, which receives the collected data X(n) as input, takes the 6-dimensional vector of collected data X(n) as input and outputs the target variable y as output. The target variable is the classification number of the road name. In this case, the collected data X(n) is the neural network that receives collected data X(1000) as input. This neural network is a fully connected neural network that propagates forward to the output layer that outputs the target variable.

[0025] As shown in Figure 3, the hidden layer of the neural network that receives the last collected data X(1000) in a 10-second period reflects the output of the hidden layer of the neural network that receives the collected data X(999) in the previous timing.

[0026] The hidden layer of the neural network that receives the collected data X(999) as input reflects the output of the hidden layer of the neural network that received the collected data X(998) at the previous time point. In this way, each neural network that receives each piece of collected data as input reflects the output of the hidden layer of the neural network that received the collected data at the previous time point.

[0027] An LSTM neural network is a recurrent neural network that incorporates a mechanism called an LSTM block in each hidden layer of such a recurrent neural network, allowing for the adjustment of the propagation of time-series information. An LSTM block consists of a cell that keeps errors internal and prevents the vanishing of gradients, an input gate that controls the input to the cell, an output gate that controls the output from the cell, and a forget gate that prevents excessive error retention in the cell.

[0028] The trained model stored in storage 12 is pre-trained using supervised training data that includes correct classification data for road names. The large amount of data needed to create the training data is collected by repeatedly conducting driving experiments with an experimental vehicle. When the data is collected, multiple types of moving average values ​​of the output torque are also calculated.

[0029] For example, when classifying road names based on one hour of driving data, the driving data classification device 10 extracts time-series data of explanatory variables within a predetermined window WdS range by moving the window WdS at a constant stride from the one hour of driving data. Then, it inputs the extracted time-series data of explanatory variables into a trained model to calculate an estimated value of the road name classification number.

[0030] The window width (WdS), i.e., the length of the default time, and the stride length should be set according to the type of road name to be classified, so that appropriate classification can be performed. <Learning Process> Next, the training process for obtaining a trained model will be described. Model training is performed by supervised learning using training data that includes correct answer data. The training process is performed by a computer equipped with a processing circuit and a memory device, similar to the driving data classification device 10. The computer's memory device stores a sufficient amount of training data to complete the training and obtain a trained model. Alternatively, the training data may be stored in the storage 12 of the driving data classification device 10 and the training process may be performed by the driving data classification device 10 itself.

[0031] As shown in Figure 3, when the learning process begins, the computer first reads the training data stored in memory during step S100. Only a portion of the training data stored in memory is read at this stage.

[0032] In step S110, the computer calculates an estimated value for the classification number indicating the road name. Specifically, the computer inputs the loaded data into a model that is trained using supervised learning to calculate the estimated value. When the training process starts, the model parameters are in their initial state. After calculating the estimated value from each piece of loaded data, the computer proceeds to step S120.

[0033] In step S120, the computer performs learning. Specifically, the computer calculates an error function representing the difference between the target variable contained in the loaded data and the calculated estimate. The computer then performs learning by adjusting the model parameters to minimize the error function. After updating the model parameters, the computer proceeds to step S130.

[0034] In step S130, the computer determines whether or not the learning process is complete. The conditions for determining whether the learning process is complete can be set appropriately based on the number of learning runs, the magnitude of the error function, etc. For example, the learning process may be determined to be complete when the error function becomes sufficiently small. Alternatively, if the number of runs required to obtain a model with sufficient estimation accuracy is known, the learning process may be determined to be complete when that number of runs is reached.

[0035] If the computer determines in step S130 that learning is not complete (step S130: NO), it returns to step S100. The computer then repeats the process from step S100 to step S130. In other words, the computer continues learning by gradually reading data and updating the model parameters until it determines that learning is complete. If the computer determines in step S130 that learning is complete (step S130: YES), it proceeds to step S140.

[0036] In step S140, the computer records the parameters of the trained model in a storage device. Then, the computer terminates this series of training processes. Thus, the trained model data is obtained through the training process. The storage device 12 of the driving data classification device 10 stores the trained model data that has been trained through the training process.

[0037] <Estimation and Classification Processing> Next, we will explain the estimation and classification processes when classifying unclassified driving data using the driving data classification device 10.

[0038] As described above, the storage 12 of the driving data classification device 10 stores unclassified driving data. The driving data classification device 10 inputs this unclassified driving data as input data X into the trained model and classifies the driving data by assigning it to a driving route name.

[0039] The unclassified driving data stored in storage 12 differs from the training data in that it is not provided with correct answer data. The driving data classification device 10 classifies the driving route names of the driving data by performing a series of processes shown in Figure 5. This series of processes is performed by the processor 11.

[0040] As shown in Figure 5, the processor 11 first performs a data reading process in step S200. This reads the data stored in the storage 12. Similar to the learning process described above, the data read here is a portion of the driving data stored in the storage 12. Once the driving data has been read, the processor 11 proceeds to step S210.

[0041] In step S210, the processor 11 inputs the loaded data as input data X into the trained model and calculates an estimated value indicating the road name. After calculating the estimated value using the input data X, the computer proceeds to step S220. In step S220, the processor 11 associates the input data X used for the calculation with the calculated estimated value and records it in the storage 12.

[0042] Next, in step S230, the processor 11 determines whether it has processed all the data stored in the storage 12. That is, it determines whether it has completed the calculation of estimated values ​​for all the driving data stored in the storage 12. If it determines in step S230 that not all the driving data has been processed (step S230: NO), the processor 11 returns to step S200. Then, the processor 11 reads some of the remaining driving data for which estimated values ​​have not been calculated and executes the processing from step S210 onward. In this way, the processor 11 repeats the processing from step S200 to step S230 until it has calculated estimated values ​​for all the unclassified driving data stored in the storage 12.

[0043] Then, if the processor determines that all data has been processed in step S230 (step S230: YES), the processor 11 proceeds to step S240.

[0044] In this driving data classification device 10, the series of processes from step S200 to step S230 described above corresponds to an estimation process that uses a trained model stored in the storage 12 to output an estimated value corresponding to the road name from the driving data.

[0045] In step S240, the processor 11 performs post-processing on the estimated data recorded in the storage 12. The rectangular symbols in Figure 6 represent the estimated data before post-processing. The estimated data linked to the driving data contains information indicating the name of the road the vehicle was traveling on at the time the driving data was collected. However, due to the influence of instantaneous fluctuations in output torque and other factors, the estimated values ​​fluctuate significantly in places, as shown in Figure 6. If the road name is determined and the driving data is classified based on estimated values ​​that include such fluctuations, the driving data may become fragmented, and it may not be possible to perform the analysis correctly.

[0046] Therefore, this driving data classification device 10 applies post-processing to the estimated data to suppress fluctuations in the estimated data. Specifically, as shown in Figure 6, the processor 11 extracts the estimated values ​​contained in the window WdL and calculates the mode. Then, it adopts the calculated mode as the representative value for the range of that window WdL and updates the extracted estimated values ​​to values ​​equal to the mode. This post-processing is applied to all estimated values ​​while moving at a constant stride. By applying this post-processing, the estimated data becomes data with suppressed fluctuations, as shown by the solid line in Figure 6.

[0047] Once the post-processing in step S240 is complete, the processor 11 proceeds to step S250. In step S250, the processor determines the road name based on the estimated value associated with the driving data. Specifically, the processor 11 assigns information about the road name corresponding to the estimated value associated with the driving data to the driving data. As a result, the driving data classification device 10 classifies the driving data according to one of several pre-prepared road names. Once the classification of all driving data is complete, the processor 11 proceeds to step S260.

[0048] In this driving data classification device 10, the processes in steps S250 and S260 described above correspond to a classification process that assigns driving data to road names based on estimated values ​​and classifies them accordingly.

[0049] In step S260, the processor 11 records the classified driving data, which has been assigned road name information and whose classification is complete, into the storage 12. Then, the processor 11 terminates this series of processes.

[0050] <Operation of this embodiment> The output torque collected during driving changes according to the road surface conditions on which the vehicle is traveling. In other words, the time-series data of output torque is useful information for estimating the name of the road on which the vehicle is traveling. The input data X is time-series data of explanatory variables, including multiple types of moving average values ​​of output torque. The driving data classification device 10 estimates the name of the road from the input data X using a trained model learned through supervised learning.

[0051] <Effects of this embodiment> (1) The driving data classification device 10 can automatically classify driving data without human intervention using a trained model. The driving data classification device 10 includes multiple types of moving average values ​​of output torque as explanatory variables. By increasing the number of explanatory variables, overfitting is suppressed. Therefore, the driving data classification device 10 can automatically classify driving data with high accuracy without human intervention.

[0052] (2) In the driving data classification device 10, the trained model is a long-short-term memory neural network. The trained model stored in the storage 12 is a long-short-term memory neural network. A long-short-term memory neural network can handle time-series data while retaining information about its progression along the time axis. According to the driving data classification device 10, classification can be performed by reflecting the time-series progression of explanatory variables included in the input data X for a predetermined time period. The driving data classification device 10 can perform classification with higher accuracy.

[0053] (3) In the above embodiment, it was shown as an example that steering angle information is included as an explanatory variable. When the road is curved, the steering angle changes when the vehicle approaches the curve. Therefore, time-series data of the steering angle is useful information for estimating the name of the road the vehicle is traveling on. If the driving data classification device 10 uses driving data that includes time-series data of the steering angle, the driving data classification device 10 will be able to estimate the name of the road based on more information. Therefore, the driving data classification device 10 can perform classification with higher accuracy.

[0054] (4) In the above embodiment, it was shown as an example that accelerator opening information is included as an explanatory variable. When the accelerator opening is changed, the output torque changes. If the driving data classification device 10 uses driving data that includes accelerator opening and output torque, the driving data classification device 10 can perform estimation processing that also reflects the effect of changes in accelerator opening. Therefore, the driving data classification device 10 can perform classification with higher accuracy.

[0055] (5) In the above embodiment, we have illustrated the inclusion of vehicle speed information as an explanatory variable. Even if the output torque is the same, the vehicle speed changes depending on the road surface conditions. Therefore, vehicle speed is also information that is useful for estimating the name of the road. If the driving data classification device 10 uses driving data that includes time-series data of vehicle speed, the driving data classification device 10 will be able to estimate the name of the road based on more information. Therefore, the driving data classification device 10 can perform classification with higher accuracy.

[0056] (6) An example was given of including the name of the driving area in the name of the driving route. In this case, the name of the driving route can be classified based on the driving data. That is, even vehicles that are not equipped with GPS can be automatically classified by applying the driving data to the name of the driving area.

[0057] <Example of changes> This embodiment can be implemented with the following modifications. This embodiment and the following modifications can be combined with each other to the extent that they do not contradict each other technically.

[0058] The specific post-processing method is not limited to the method exemplified in the above embodiment. For example, post-processing using the maximum value as the representative value or post-processing using the minimum value as the representative value may be employed. Alternatively, post-processing using the median or mean as the representative value may be employed. When using the median or mean as the representative value, it is sufficient to apply it to the road name that is closest to the estimated value after post-processing.

[0059] The types of moving averages to be included in the explanatory variables are not limited to those exemplified in the embodiments described above. For example, the types of moving averages are not limited to three. There may be two types of moving averages. There may be four or more types of moving averages.

[0060] In the above embodiment, the driving data classification device 10 includes a processor 11, which is a processing circuit, and a storage device 12, which is a memory device, to perform software processing. However, this is merely an example. For example, the driving data classification device 10 may include a dedicated hardware circuit (e.g., an ASIC) that processes at least a part of the software processing performed in the above embodiment. That is, the driving data classification device 10 may have any of the following configurations (A) to (C): (A) The driving data classification device 10 includes an execution device that performs all processing according to a program, and a memory device that stores the program. That is, the driving data classification device 10 includes a software execution device. (B) The driving data classification device 10 includes an execution device that performs part of the processing according to a program, and a memory device. Furthermore, the driving data classification device 10 includes a dedicated hardware circuit that performs the remaining processing. (C) The driving data classification device 10 includes a dedicated hardware circuit that performs all processing. Here, there may be multiple software execution devices and / or dedicated hardware circuits. In other words, the above process may be performed by a processing circuitry comprising at least one of one or more software execution devices and one or more dedicated hardware circuits. The storage device for storing the program, i.e., the computer-readable medium, includes any available medium accessible by a general-purpose or dedicated computer. [Explanation of Symbols]

[0061] 10... Driving data classification device, 11... Processor, 12... Storage, 13... Memory, 14... Input device, 15... Display device, 16... Input / Output interface, 17... Communication device, 18... Bus

Claims

1. This is a driving data classification device that, while driving a vehicle equipped with sensors, classifies driving data, including the output torque of the vehicle's prime mover, by applying it to multiple driving route names, which are pre-classified into sections on a test course. The aforementioned names of multiple roads include straight sections and uphill sections. A trained model that estimates the name of the road corresponding to a point where the vehicle was traveling when the driving data was being collected, using input data which is time-series data of explanatory variables including multiple types of moving average values ​​of the output torque of the vehicle's engine, wherein the trained model is stored in a storage device that has been supervised and trained using training data which includes the input data and ground truth data indicating the name of the road corresponding to a point where the vehicle was traveling when the driving data was being collected. The system includes a processing circuit that performs an estimation process to output an estimated value corresponding to the road name from the input data using the trained model stored in the memory device, and a classification process to classify the driving data by applying it to the road name based on the estimated value. Driving data classification device.

2. The aforementioned trained model is a long- and short-term memory neural network. The driving data classification device according to claim 1.

3. The plurality of road names further include sharp curves, The aforementioned explanatory variables include information about the steering angle. The driving data classification device according to claim 1.

Images