JP6390388B2 - Data approximation apparatus and data approximation method - Google Patents

Description

  The present invention relates to an apparatus for approximating time-series data.

  2. Description of the Related Art In recent years, various sensors are mounted on automobiles, and various information relating to vehicle travel such as speed, longitudinal acceleration, steering angle, yaw rate, and lateral acceleration can be acquired. In addition, it is expected that this information will be collected as big data and used for research in various fields.

  When considering collecting and analyzing data acquired from a vehicle, compression of the data amount becomes an issue. This is because it is not realistic in terms of capacity to hold all data generated from a running vehicle as it is without being processed.

  As an invention related to this problem, for example, there is a function approximation device described in Patent Document 1. The apparatus has a feature of performing clustering by dividing input data into unit times and obtaining an approximate expression for each created cluster. The input data (sensor data acquired from the vehicle when applied to a vehicle) is classified into similar data and converted into an approximate expression, so that the capacity is greatly compressed compared to the case of holding raw data. be able to.

JP-A-7-065168

  On the other hand, since the approximate expression obtained by the method described in Patent Document 1 approximates a change in value, there is a problem that an error with respect to the original data increases with time. In order to reduce the error, the unit for dividing the input data may be reduced. However, if the unit is made too small, the feature for each cluster is lost and the data compression effect is diminished.

  The present invention has been made in consideration of the above-described problems, and provides a data approximation device that approximates time-series data, and that generates information that accurately represents the characteristics of target time-series data. The purpose is to do.

In order to solve the above-described problem, a data approximation device according to the present invention includes:
A data approximation device that outputs a coefficient of a linear dynamic system representing time series data and a reference value at one or more time points as data approximating a cluster including time series data composed of a plurality of values. Using data acquisition means for acquiring one or more time series data included in a cluster, first calculation means for calculating a coefficient of a linear dynamic system approximating the time series data, and using the linear dynamic system Second calculating means for determining one or more reference values composed of time and value, used when approximating the time series data, the linear dynamic system, and the one or more reference values And output means for outputting as a parameter expressing the target time-series data.

The data approximation apparatus according to the present invention is an apparatus that outputs information for approximating time series data included in a cluster.
The cluster to be approximated may be any cluster as long as it includes one or more data generated in time series (time series data). For example, it may be a set of time series data composed of values periodically output by sensors mounted on the vehicle.
When the target cluster includes a plurality of time series data, any time series data may be approximated, or time series data representing the plurality of time series data (for example, for each relative time in the window) An average value or a set of median values) may be approximated. Time series data representing the plurality of time series data can also be handled as time series data included in the cluster.

The first calculation means is a means for calculating a linear dynamic system for approximating a target cluster, that is, approximating any time series data included in the cluster. A linear dynamic system is a system in which a state changes with the passage of time, and a state at another time can be expressed using a state at a certain time.
The second calculation means is means for determining a reference value for approximating time series data included in a cluster using a linear dynamic system. The linear dynamic system calculated by the first calculation means represents the amount of change from an adjacent time step by a mathematical expression. Therefore, in order to approximate the time series data, it is necessary to give a reference value. By applying the reference value determined by the second calculation means to the linear dynamic system calculated by the first calculation means, time series data can be approximated.

  The data approximation apparatus according to the present invention outputs one or more reference values corresponding to the time series data in association with the linear dynamic system corresponding to the time series data. Thereby, compared with the case where the original time series data is approximated using a single approximate expression, the accuracy of approximation can be further improved.

  Further, the second calculation means applies a target linear dynamic system with one or more temporarily determined reference values as starting points, and generates approximate data that approximates the time series data; and the approximate data And a means for obtaining an approximation error for the time series data and determining the tentatively determined reference value as an output reference value based on the approximation error.

  When approximating time-series data using a linear dynamic system, the approximation error varies greatly depending on how the reference value is taken. Therefore, approximate data is generated by temporarily setting a reference value, and a reference value to be output is determined based on the obtained approximation error. Thereby, a more accurate approximation is possible.

  The time series data includes n time steps, and the second calculation unit temporarily selects m time / value combinations from the n time steps of the time series data. Then, a reference value to be output is determined by searching for a combination that minimizes the approximation error, and m is a minimum value that causes the approximation error to fall below a predetermined threshold value.

The accuracy of approximation of the reference value varies greatly depending on how it is determined. Therefore, after generating candidates for combinations that m reference values take, a combination that minimizes the error when actually approximating is searched. As a result, the approximation accuracy can be maximized.
Further, the number of m is the smallest number in which the error when actually approximating is below a predetermined threshold. Thereby, it is possible to balance the approximation accuracy and the data amount.

Further, the approximation error is the sum, average, square of absolute values of the difference when the difference between the value of the time series data and the value of the approximation data is acquired at one or more time points. It may be one of the sums.

  The approximate error is an error (absolute value of difference) between an approximate value at a certain time (that is, an approximate value obtained as a result of applying a linear dynamic system to a reference value) and a value of the time-series data at the time. It can be obtained by sum, average, sum of squares, and the like.

  The plurality of values constituting the time series data may each be a multi-dimensional vector. For example, when the time-series data is sensor data acquired from a vehicle, it can be a multi-dimensional vector including speed, longitudinal acceleration, steering angle, yaw rate, lateral acceleration, and the like.

  The data approximation device according to the present invention further includes a clustering unit that clusters a plurality of time-series data acquired by the data acquisition unit, and the first and second calculation units are classified into the same cluster. Processing may be performed on a plurality of time-series data, and the linear dynamic system and the reference value may be associated with corresponding clusters.

  As described above, the time series data included in the cluster is preferably time series data similar to each other classified by clustering.

  The data approximation apparatus according to the present invention may further include a combining unit that combines similar clusters among a plurality of clusters generated by clustering.

  By joining similar clusters together, the amount of data generated can be further compressed. For example, when clusters are generated by a hierarchical method, lower-layer clusters can be combined as similar clusters. Further, when clusters are generated by the division optimization method, they can be combined after determining the similarity based on the distance between the clusters.

  When the combining unit combines clusters, the first calculating unit recalculates the linear dynamic system using time series data belonging to the cluster after combining, and the output unit recalculates. The linear dynamic system may be output in association with the reference value corresponding to each cluster before combination.

  When clusters are combined, a linear dynamic system associated with each cluster can be shared. However, when such a method is used, the approximation accuracy is degraded. Therefore, in the present invention, when the clusters are combined, only the linear dynamic system is shared, and the reference value belonging to each cluster is held as it is associated with each cluster. By doing in this way, the fall of the approximation precision by combining a cluster can be suppressed to the minimum.

  In addition, the first calculation means may calculate the coefficient of the linear dynamic system so that the sum of squares of errors between the value of the target time series data and the approximate value represented by the linear dynamic system is minimized. May be determined.

  The linear dynamic system can be suitably obtained by the least square method. That is, it is only necessary to obtain a parameter that minimizes the sum of squares of errors between the value of the time series data to be approximated and the approximated value.

The time series data may be at least one of a vehicle speed, longitudinal acceleration, steering angle, yaw rate, and lateral acceleration.

  The data approximation apparatus according to the present invention can be suitably applied to an apparatus that approximates sensor data generated from a vehicle.

  In addition, this invention can be specified as a data approximation apparatus containing at least one part of the said means. It can also be specified as a data approximation method performed by the data approximation device. Further, it can be specified as a vehicle behavior classification system including the data approximation device and a vehicle that transmits sensor information to the data approximation device. The above processes and means can be freely combined and implemented as long as no technical contradiction occurs.

  ADVANTAGE OF THE INVENTION According to this invention, it is a data approximation apparatus which approximates time series data, Comprising: The data approximation apparatus which produces | generates the information which represented the characteristic of the target time series data accurately can be provided.

1 is a system configuration diagram of a vehicle behavior classification system according to a first embodiment. It is a figure explaining the time series data which a sensor information acquisition part acquires. It is a figure explaining the production | generation of a vehicle behavior symbol. It is a figure explaining the time series data contained in a cluster. It is a figure explaining an approximation error. It is a flowchart figure of the 2nd process in 1st embodiment. It is a system configuration figure of the classification device concerning a second embodiment. It is a flowchart figure of the 3rd process in 2nd embodiment. It is an example which joins a cluster in 2nd embodiment.

(First embodiment)
<System configuration>
Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings.
The vehicle behavior classification system according to the first embodiment is a system including an in-vehicle device 10 mounted on a vehicle and a classification device 20, and classifies the behavior of the vehicle based on information transmitted from the in-vehicle device. , A system to accumulate.

FIG. 1 is a system configuration diagram of the in-vehicle device 10 and the classification device 20 according to the present embodiment.
First, the in-vehicle device 10 will be described. The in-vehicle device 10 is a device that transmits information about the behavior of the vehicle on which the in-vehicle device is mounted to the classification device 20. The in-vehicle device 10 includes a sensor information acquisition unit 11 and a communication unit 12.

  The sensor information acquisition unit 11 is a means for acquiring values (hereinafter referred to as sensor values) from a plurality of sensors mounted on the vehicle. The sensor mounted on the vehicle is a sensor that acquires a value representing the behavior of the vehicle, and includes, for example, a speed sensor, an acceleration sensor, a yaw rate sensor, and a steering angle sensor, but is not limited thereto. The sensor value is transmitted to the classification device 20 as time-series data via the communication unit 12 described later at every predetermined time. Hereinafter, data transmitted from the sensor information acquisition unit 11 is referred to as sensor information, and in particular, data including a plurality of sensor values generated in time series is referred to as time series data.

  The communication unit 12 is a means for transmitting the sensor information acquired by the sensor information acquisition unit 11 to the classification device 20. A protocol and a communication method to be used are not particularly limited as long as information can be transmitted and received by wireless communication.

Next, the classification device 20 will be described. The classification device 20 is a device that receives and accumulates data transmitted from the in-vehicle device 10. Further, when storing data, the data is compressed by the method disclosed in this specification.
The classification device 20 includes a communication unit 21, a cluster generation unit 22, an approximate expression generation unit 23, and a storage unit 24.

  The communication unit 21 is means for receiving sensor information transmitted from the in-vehicle device 10. The protocol and communication method used are the same as those of the communication unit 12.

  The cluster generation unit 22 is means for generating a symbol representing the behavior of the vehicle for each unit time using time series data acquired from the vehicle. A symbol representing the behavior of the vehicle is referred to as a vehicle behavior symbol. In the present embodiment, the vehicle behavior symbol is generated by performing clustering by dividing the time-series data acquired from the vehicle and accumulated for each unit time. The clustering result is stored in the storage unit 24 described later. A specific example of clustering will be described later.

  The approximate expression generation unit 23 is a means for generating an approximate expression corresponding to the cluster generated by the cluster generation unit 22. Specifically, for each cluster, a linear dynamic system that approximates related time series data is generated. The generated linear dynamic system is associated with the cluster and stored in the storage unit 24 described later. Details of the linear dynamic system and a method for generating the linear dynamic system will be described later.

  The storage unit 24 is a nonvolatile storage medium that stores acquired time-series data, vehicle behavior symbols (clusters), linear dynamic systems, and the like. The storage unit 24 is preferably a storage medium that can read and write at high speed and has a large capacity. For example, a flash memory can be suitably used.

  Control of each means demonstrated above is implement | achieved when processing apparatuses (not shown), such as CPU, run a control program. Further, the function may be realized by a field-programmable gate array (FPGA), an application specific integrated circuit (ASIC), or the like, or may be realized by a combination thereof.

<Outline of processing>
The processing performed by the classification device 20 according to the present embodiment mainly includes processing for collecting and storing time series data from the in-vehicle device (first processing), clustering the accumulated time series data, and generating a cluster Can be divided into two processes (second process) that approximate by a mathematical expression. An overview of each process will be described.

  First, the first process will be described with reference to FIG. 2 which is a diagram for explaining data transmitted from the in-vehicle device 10.

<< Sensor information transmission process >>
The sensor information acquisition unit 11 acquires sensor values at a predetermined sampling rate (for example, 10 Hz) from a plurality of sensors (not shown) included in the vehicle. The sensor value may be acquired at a sampling rate higher than the target sampling rate and then smoothed by a filter. For example, after sampling at 100 Hz, it may be downsampled to 10 Hz by a Gaussian filter or the like.
In this example, three sensors of steering angle, speed, and longitudinal acceleration are used. That is, since ten sensor values are obtained for each of the three sensors, 30 sensor values per second are transmitted to the classification device 20 as time-series data (reference numeral 201). The transmission cycle is not necessarily fixed, and the transmission unit is not necessarily fixed.
When receiving the time series data from the vehicle via the communication unit 21, the classification device 20 sequentially stores the data in the storage unit 24. In this example, since there are three sensors, time-series data is stored as a three-dimensional vector.

<< Clustering process >>
Next, the second process performed by the classification device 20 will be described. The second processing is divided into processing for clustering time series data and processing for generating a linear dynamic system for each cluster. First, clustering processing performed by the cluster generation unit 22 will be described with reference to FIG.

The vehicle behavior symbol is a symbol representing the behavior of the vehicle in a unit time (for example, 5 seconds). The vehicle behavior symbol can be obtained by dividing time-series data generated by the vehicle (a three-dimensional vector having information for unit time) into unit times and clustering the divided data. As a result, a train of vehicle behavior symbols as indicated by reference numeral 301 is obtained. Reference numeral 301 represents a vehicle behavior symbol for 80 seconds (5 seconds × 16). For the clustering, for example, an arbitrary method such as K-means or spectral clustering can be used. Further, classification may be performed using other methods as long as the classification result can be obtained by using time-series data as input. A combination of classification and clustering may be used. For example, the remainder processed by the support vector machine (SVM) may be processed by K-means.

Although FIG. 3 shows time-series data acquired from one vehicle, clustering may be performed using time-series data acquired from a plurality of vehicles. In FIG. 3, the time series data is divided every 5 seconds, but the unit time may be other than 5 seconds. In addition, the divided data may partially overlap.
Information about the generated cluster is temporarily stored in the storage unit 24.

<< Generation of linear dynamic system >>
Next, processing for generating a linear dynamic system corresponding to the generated cluster will be described. FIG. 4 is a diagram illustrating a plurality of time-series data belonging to a certain cluster generated by the cluster generation unit 22. In this example, since the unit time is 5 seconds and the sensor value sampling rate is 10 Hz, each time-series data includes 50 sensor values.
Here, it is assumed that three time-series data represented by reference numerals 401 to 403 are classified into the cluster. Note that the time-series data is a vector having a plurality of dimensions, but FIG. 4 shows only one dimension for the sake of simplicity.

  Here, a linear dynamic system for approximating these time series data is generated. Note that when a plurality of time-series data are assigned to one cluster, for example, the plurality of time-series data may be connected in order of time to be approximated time-series data. Also, new time series data representing the plurality of time series data may be generated and used as approximation target time series data. For example, new time series data may be generated by obtaining an average value or median value of the three time series data 401 to 403 for each time. The time series data obtained in this way can also be handled as time series data belonging to the target cluster.

A linear dynamic system is represented by Formula (1), for example. In Expression (1), x _t is a system state at time t (that is, a multidimensional vector representing a set of sensor values), and x _t−1 is a system state at time t−1. A and b are parameters representing state changes with the passage of time. Further, v _t is a term representing an error caused by noise. The parameters A and b are expressed as a matrix, and operations on multidimensional vectors can be performed.

As a method for obtaining a linear dynamic system that approximates certain time series data, there is a least square method.
Specifically, the parameters A and b of the linear dynamic system are determined so that the sum of squares of the difference between the value of the target time-series data and the approximate value is minimized.

<< Parameter determination method >>
A specific method for obtaining the parameters A and b will be described.
First, based on the target time-series data, an input vector and an output vector are generated every time step (sensor value acquisition interval, 0.1 seconds in this embodiment). The input vector is a vector that represents a sensor value at time t−1, and the output vector is a vector that represents a sensor value at time t. Expression (2) is an input vector, and Expression (3) is an output vector. Note that n is the number of dimensions. When there are three types of sensors, n = 3.

  The input vector and the output vector are generated in the range of 2 ≦ t ≦ T. T is the number of time steps (that is, the number of sensor values) included in the target cluster.

Next, an average is obtained from the input vector and the output vector. The average value can be obtained by Equation (4) and Equation (5).

Then, each vector is normalized by subtracting the average from the input vector and the output vector. The normalized result is expressed by Expression (6) and Expression (7).

Next, an input matrix represented by Expression (8) and an output matrix represented by Expression (9) are generated.

Finally, parameter A is obtained by equation (10), and parameter b is obtained by equation (11). X ⁻¹ represents an inverse matrix of X.

The linear dynamic system generated by the above description represents the amount of change in sensor value from a certain point in time. That is, in order to approximate the sensor value, at least one reference sensor value must be given. Therefore, in the present embodiment, the approximate expression generation unit 23 selects a reference sensor value and stores it in association with the cluster. In the following description, the reference sensor value selected by the approximate expression generation unit 23 is referred to as a reference value.
A specific method for selecting the reference value will be described with reference to a flowchart described later.

The effect of associating a reference value with a cluster (ie, a linear dynamic system) will be described with reference to FIG.
Reference numeral 501 is time series data to be approximated, and reference numeral 502 is time series data approximated by a linear dynamic system (hereinafter, approximate data). The black circles in the figure are reference sensor values, that is, reference values.
Although the actual number of reference values is larger than that shown in the figure, the description here assumes that the reference values are one and two for ease of explanation.

FIG. 5A shows an example where there is one reference value. In the case of this example, as can be seen from the figure, the error accumulates with the passage of time, so that the approximation accuracy falls in the second half of the time slot.
FIG. 5B shows an example in which there are two reference values. In this example, the approximate value can be corrected by the second reference value. As described above, when there are a plurality of reference values, the values can be appropriately corrected. Therefore, even when the same linear dynamic system is used, the approximation accuracy can be improved.

  As described above, the classification device according to the present embodiment firstly approximates time-series data belonging to a cluster by a linear dynamic system, and secondly associates a reference value with the cluster, and represents the characteristics of the cluster. Remember as. By doing so, the features of the cluster can be expressed with a small amount of data and high accuracy.

<Process flowchart>
Next, a method of determining the number of reference values and the arrangement position will be described with reference to FIG. 6 which is a flowchart of the second process. Note that the first process is a process of periodically receiving time-series data from the in-vehicle device and accumulating it in the storage unit 24, and thus the description thereof is omitted.
The process shown in FIG. 6 is executed at an arbitrary timing in a state where time-series data is accumulated in the storage unit 24. Note that steps S11 to S12 are processes performed by the cluster generation unit 22, and steps after step S13 are processes performed by the approximate expression generation unit 23.

First, in step S11, all time-series data stored in the storage unit 24 is acquired.
Next, in step S12, the acquired time series data is divided into unit times, converted into feature quantity vectors, and then clustered.

The processes in steps S13 to S19 are processes for associating a linear dynamic system with a reference value for each of the generated clusters.
In step S13, it is determined whether or not processing has been completed for all clusters. If not, the processing is shifted to step S14.

In step S14, after determining a cluster to be processed, a linear dynamic system corresponding to the cluster is obtained by the method described above.
In step S15, initial values are set for variables (two types, N and ε). N is a variable representing the number of reference values associated with the cluster to be processed, and the initial value is 1. Ε is a variable representing an error, and the initial value is a sufficiently large natural number.

  In step S16, it is determined whether or not the error ε is smaller than a predetermined threshold value E. If the error is larger than the threshold value E, the process proceeds to step S17.

Steps S <b> 17 to S <b> 19 are steps for searching for an optimal arrangement of reference values when the number of reference values is N. Specifically, when the target cluster has n time steps, N is extracted from n combinations of time and sensor value, and a linear dynamic system is used as a temporary reference value. Apply to generate approximate data, and calculate the error from the time series data to be approximated. This, while changing the position of the reference value all patterns (i.e., _n C _N pieces) is performed for, searching for a combination of errors is minimized.

In step S17, it is determined whether or not all patterns have been examined for the combination of N reference values. If not, the process proceeds to step S18.
In step S18, a combination of unexamined reference values is temporarily selected, and the linear dynamic system obtained in step S14 is applied to the temporarily selected reference value to obtain approximate data. Then, an average value ε ₁ of errors between the approximate data and the time series data is obtained. Ε ₁ obtained here is, for example, the average value of the deviation amount between the solid line and the dotted line in FIG. Here, if epsilon ₁ is smaller than epsilon, then updates the epsilon in step S19, shifts the process to step S17.
In addition, when calculating | requiring an error, you may normalize a value. Further, instead of the average of errors, a total of errors or a sum of squares may be used.

  If this process is repeated for the number of combinations that can be taken by the N reference values, the combination of the reference values that minimizes the average error when N reference values are used can be specified.

  When all the combinations that can be taken by the N reference values are examined (step S17—Yes), N is incremented, and the process proceeds to step S16. Here, when the error ε is smaller than the predetermined threshold E, the storage unit 24 associates the linear dynamic system and the N reference values temporarily selected in step S18 with the cluster to be processed in step S20. Remember me. At this time, the original time series data is deleted.

  The process described above is repeatedly executed for the number of clusters. Finally, when it is determined in step S13 that the processing for all clusters has been completed, the processing ends.

  As described above, the classification device according to the present embodiment obtains a linear dynamic system for approximating the original time-series data, searches for a combination of reference values that satisfy the approximation error, A reference value is associated and stored. Thereby, compared with the case where raw time-series data is stored as it is, the data amount can be greatly compressed. Also, the approximation accuracy can be improved as compared with the case of storing a single approximate expression.

  In the present embodiment, the generated linear dynamic system and the reference value are stored in the storage unit. However, the linear dynamic system and the reference value may be output to the outside. Alternatively, the value may be approximated based on the accumulated information and the result may be output to the outside.

(Second embodiment)
The second embodiment is an embodiment that further improves the data compression rate by combining similar clusters among the generated clusters.
FIG. 7 is a system configuration diagram of the classification device 30 according to the second embodiment. The classification device 30 according to the second embodiment is different from the first embodiment in that it includes a cluster coupling unit 35. Since other means are the same as those in the first embodiment, description thereof will be omitted. The in-vehicle device 10 that communicates with the classification device 30 is the same as that in the first embodiment, and thus the description thereof is omitted.

In the second embodiment, after completing the second process, the cluster combining unit 35 executes a process (third process) for further data compression by combining the generated clusters.
FIG. 8 is a flowchart of the third process performed by the cluster combining unit 35. This process is executed after the process shown in FIG. 6 is completed.

First, in step S31, the most similar clusters are temporarily combined among the plurality of classified clusters, and the linear dynamic system is processed by the same processing as in step S14 using the time-series data included in each cluster. Ask again. For example, new time series data representing each time series data included in the cluster before the combination is generated and used as the approximation target time series data, and a linear dynamic system is obtained.
Note that the most similar cluster may be determined based on the hierarchy, for example, when a hierarchical method is used for clustering, or determined based on the distance between the clusters, when a division optimization method is used. May be. Moreover, you may determine using another method. Alternatively, a threshold may be set for the degree of similarity, and processing may be performed only for clusters that are somewhat similar.

Next, in step S32, using the reference value associated with each cluster to be combined and the linear dynamic system recalculated in step S31, an error is obtained for each cluster by the same process as in step S18. The average value ε _{2 of the} error is calculated. If the error calculated here is sufficiently small, it can be seen that the approximation error does not increase even if the linear dynamic system is shared.

Next, it is determined at step S33, whether the average error epsilon ₂ is above the threshold E. Here, if not exceed the epsilon ₂ threshold E, processing to transition to step S34, formally bind the target cluster. Specifically, the linear dynamic system obtained in step S31 is stored in the storage unit 24 in association with the combined cluster. At this time, the cluster before joining is included in the cluster after joining as a sub-cluster. The reference values associated with the clusters before joining are held as they are while being associated with the clusters (subclusters) before joining.

  FIG. 9 shows an example in which clusters are combined. FIG. 9A shows a state before joining, and FIG. 9B shows a state after joining. In this example, cluster C including cluster A and cluster B is generated, and clusters A and B are sub-clusters included in cluster C.

In step S33, if the average error epsilon ₂ was above the threshold E, the subjects of the cluster, coupling is determined to be impossible, undo cluster is tentatively bonded. That is, the information stored in the storage unit 24 is maintained.
The above processing is repeatedly executed until the processing is completed for all clusters (step S35).

  As described above, in the second embodiment, the amount of data is further compressed by combining similar clusters and sharing a linear dynamic system within a range where the error satisfies an allowable value. On the other hand, the reference value related to each cluster is not shared but is held as it is. Thereby, further compression of the data amount can be realized without greatly deteriorating the approximation accuracy.

(Modification)
The above embodiment is merely an example, and the present invention can be implemented with appropriate modifications within a range not departing from the gist thereof.

  For example, in the description of the embodiment, as the linear dynamic system, the state at a certain time step is represented by using the state at the previous time step (forward approximation). It may be used. For example, the state at a certain time step may be represented using the state at the next time step (rearward approximation). Further, the state at a certain time step may be expressed by using the state at each adjacent time step both before and after.

  In the description of the embodiment, an example in which sensor information is periodically transmitted from the in-vehicle device 10 by wireless communication has been described. However, the sensor information may be transmitted by any method. For example, it may be transmitted every time a trip is completed, or may be transmitted according to a predetermined schedule. The sensor information does not necessarily have to be transmitted wirelessly, but may be exchanged offline.

  In the description of the embodiment, the speed, the longitudinal acceleration, the steering angle, the yaw rate, and the lateral acceleration are exemplified as information that can be acquired by the sensor. Good. In addition, the processing target data may not be acquired by the sensor as long as it is time-series data, and may be acquired from other than the vehicle.

  In the description of the embodiment, an example in which sensor information acquired from a vehicle is clustered has been described. However, clustering is not necessarily performed, and time-series data is not necessarily sensor information. Even if one or more arbitrary time-series data is processed as it is, the object of the invention can be achieved.

DESCRIPTION OF SYMBOLS 10 In-vehicle apparatus 11 Sensor information acquisition part 12 Communication part 20 Classification apparatus 21 Communication part 22 Cluster generation part 23 Approximate expression generation part 24 Storage part

Claims (13)

A data approximation device that outputs a coefficient of a linear dynamic system representing time series data and a reference value at one or more time points as data approximating a cluster including time series data composed of a plurality of values. ,
Data acquisition means for acquiring one or more time-series data included in the cluster;
First calculating means for calculating a coefficient of a linear dynamic system approximating the time series data;
Second calculating means for determining one or more reference values consisting of time and value, used when approximating the time series data using the linear dynamic system;
An output means for associating the linear dynamic system with the one or more reference values and outputting as a parameter expressing the target time-series data;
A data approximation device. The second calculating means includes
Means for applying the target linear dynamic system starting from one or more temporarily determined reference values and generating approximate data approximating the time series data;
Means for obtaining an approximation error of the approximate data with respect to the time-series data, and determining the provisionally determined reference value as a reference value to be output based on the approximation error;
The data approximation apparatus according to claim 1. The time series data consists of n time steps,
The second calculation means temporarily selects m combinations of time and value from the n time steps of the time series data, and searches for a combination that minimizes the approximation error. A reference value to be output is determined, and m is a minimum value at which the approximation error falls below a predetermined threshold;
The data approximation apparatus according to claim 2. The approximate error is the sum, average, sum of squares of the absolute values of the differences when the difference between the value of the time series data and the value of the approximate data is acquired at one or more time points. One of them,
The data approximation apparatus according to claim 3. The plurality of values constituting the time series data are each a multidimensional vector.
The data approximation apparatus according to claim 1. Clustering means for clustering a plurality of time series data acquired by the data acquisition means,
The first and second calculation means perform processing on a plurality of time-series data classified into the same cluster,
The linear dynamic system and the reference value are each associated with a corresponding cluster;
The data approximation apparatus according to any one of claims 1 to 5. Of the plurality of clusters generated by clustering, the apparatus further includes a combining unit that combines similar clusters.
The data approximation apparatus according to claim 6. When the combining means combines clusters,
The first calculation means recalculates the linear dynamic system using time series data belonging to the combined cluster,
The output means associates and outputs the recalculated linear dynamic system and reference values corresponding to the respective clusters before combining.
The data approximation apparatus according to claim 7. The first calculating means determines a coefficient of the linear dynamic system so that a sum of squares of errors between a value of the target time series data and an approximate value represented by the linear dynamic system is minimized. To
The data approximation apparatus according to claim 1. The time series data is at least one of vehicle speed, longitudinal acceleration, steering angle, yaw rate, lateral acceleration,
The data approximation apparatus according to claim 1. A data approximation device according to any one of claims 1 to 10,
A vehicle having a plurality of sensors and transmitting the acquired sensor information to the data approximation device as time-series data;
A data collection system consisting of Data performed by a data approximation device that outputs a coefficient of a linear dynamic system representing the time series data and a reference value at one or more time points as data approximating a cluster including time series data composed of a plurality of values. An approximation method,
A data acquisition step of acquiring one or more time-series data included in the cluster;
A first calculation step of calculating a coefficient of a linear dynamic system approximating the time series data;
A second calculation step for determining one or more reference values consisting of time and value, used when approximating the time series data using the linear dynamic system;
An output step of associating the linear dynamic system with the one or more reference values and outputting as a parameter expressing time-series data of interest;
A data approximation method, including A data approximation device that outputs a coefficient of a linear dynamic system representing time series data and a reference value at one or more time points as data approximating a cluster including time series data composed of a plurality of values. ,
Data acquisition means for acquiring one or more time-series data included in the cluster;
First calculating means for calculating a coefficient of a linear dynamic system approximating the time series data;
Second calculating means for determining one or more reference values consisting of time and value, used when approximating the time series data using the linear dynamic system;
An output means for associating the linear dynamic system with the one or more reference values and outputting as a parameter expressing the target time-series data;
Have
The second calculating means includes
Applying the target linear dynamic system starting from one or more reference values determined as a starting point, generating approximate data approximating the time series data, and then calculating an approximation error of the approximate data with respect to the time series data Means to obtain,
While increasing the number m of the reference values, m time / value combinations that minimize the approximation error are searched from n time steps of the time series data, and the approximation error is And a means for terminating the search process and outputting the m reference values when a predetermined threshold value is exceeded.
Data approximation device.

Images