Data discrimination device and data discrimination method

Abstract

The purpose of the present invention is to improve the comprehensiveness of teacher data. This data discrimination device stores a plurality of reference data items and additional data items to create a reference data group. The data determination device comprises a distance calculation unit, a grouping unit, a frequency calculation unit, and a selection unit. The distance calculation unit calculates a reference point of the reference data group and the distance from the reference point to the plurality of reference data items and the additional data items. The grouping unit groups the plurality of reference data items and the additional data items into a plurality of groups on the basis of the distance calculated by the distance calculation unit. The frequency calculation unit calculates the frequency of the plurality of reference data items and the additional data items for each of the plurality of groups. The selection unit selects a plurality of reference data items and additional data items to be included in the reference data group such that the frequency approaches a predetermined target level.

Description

Data discrimination device and data discrimination method

[0001] The present invention relates to a data discrimination device and a data discrimination method.

[0002] MAP (Model Predictive Control) has traditionally been used for powertrain control. However, in recent years, there has been a demand for higher precision and efficiency in powertrain control, and powertrain control using neural network models has been considered. A neural network model is a mathematical model that mimics the mechanisms of the neural circuits in the human brain.

[0003] Control using a neural network model requires learning using training data. In a neural network model, increasing the amount of training data increases the amount of unnecessary training data that does not affect accuracy. As a result, there is a risk that the calculation accuracy cannot be improved by learning using training data.

[0004] Patent Literature 1 discloses an apparatus for efficiently selecting training data to be used for reconstructing a learning model. The apparatus disclosed in Patent Literature 1 assumes that the training data is a group of normally distributed data, and selects data that is far from the existing training data as additional training data.

[0005] Japanese Patent Application Laid-Open No. 2022-157180

[0006] However, the device disclosed in Patent Document 1 has a problem in that even if there is a sufficient amount of training data, if the distance between the training data and the existing training data is large, the training data is selected as additional training data, resulting in excessive addition of training data. For example, the training data obtained from automobiles contains a mixture of data that shows transient changes and data that shows steady changes. As a result, training data with similar values ​​is likely to be added.

[0007] In view of the above problems, an object of the present invention is to provide a data discrimination device and a data discrimination method that can increase the comprehensiveness of training data.

[0008] To solve the above problems and achieve the present object, one aspect of the present invention is a data discrimination device that accumulates multiple pieces of reference data and additional data to create a reference data group. The data discrimination device includes a distance calculation unit, a grouping unit, a frequency calculation unit, and a selection unit. The distance calculation unit calculates a reference point of the reference data group and the distance from the reference point to the multiple pieces of reference data and additional data. The grouping unit divides the multiple pieces of reference data and additional data into multiple groups based on the distances calculated by the distance calculation unit. The frequency calculation unit calculates the frequencies of the multiple pieces of reference data and additional data for each of the multiple groups. The selection unit selects the multiple pieces of reference data and additional data to be included in the reference data group so that the frequencies approach a predetermined target level.

[0009] A data discrimination method according to one aspect of the present invention is a method for creating a reference data group by accumulating multiple pieces of reference data and additional data. In the data discrimination method, a distance calculation unit first calculates a reference point of the reference data group and the distances from the reference point to the multiple pieces of reference data and additional data. Next, a grouping unit divides the multiple pieces of reference data and additional data into multiple groups based on the distances calculated by the distance calculation unit. Next, a frequency calculation unit calculates the frequencies of the multiple pieces of reference data and additional data for each of the multiple groups. Then, a selection unit selects the multiple pieces of reference data and additional data to be included in the reference data group so that the frequencies approach a predetermined target level.

[0010] According to one aspect of the present invention, it is possible to improve the comprehensiveness of training data. Note that problems, configurations, and effects other than those described above will become apparent from the following description of the embodiment.

[0011] 1 is a block diagram showing an example of the hardware configuration of a data discrimination device according to the first embodiment; FIG. 2 is a conceptual diagram showing the weights and biases of each neuron constituting a neural network model; FIG. 3 is a diagram showing an example of a neural network model according to the first embodiment; FIG. 4 is a diagram showing an overview of Mahalanobis distance according to the first embodiment; FIG. 5 is a diagram showing an increase in the number of data in a conventional reference data group and a coefficient of determination; FIG. 6 is a block diagram showing the functional configuration of a control unit (CPU) according to the first embodiment; FIG. 7 is a flowchart showing data discrimination processing according to the first embodiment; FIG. 8 is a diagram showing a first selection example performed by a data selection unit according to the first embodiment; FIG. 9 is a diagram showing a second selection example performed by a data selection unit according to the first embodiment; FIG. 10 is a diagram showing an example of calculation of a target level according to the first embodiment; FIG. 11 is a flowchart showing data replacement processing according to the first embodiment; FIG. 11 is a diagram showing a data discrimination device according to a second embodiment;

[0012] First Embodiment A data discrimination device and a data discrimination method according to a first embodiment will be described below. Note that common parts in each drawing are given the same reference numerals.

[0013] [Data discrimination device] First, a description will be given of the configuration of a data discrimination device according to the first embodiment. Fig. 1 is a block diagram showing an example of the hardware configuration of a data discrimination device according to an embodiment of the present invention.

[0014] 1 is, for example, an ECU (Electronic Control Unit), and includes an input circuit 101, an input / output port 102, a RAM 103, a ROM 104, a CPU 105, an output circuit 106, and a storage unit 107.

[0015] Signals output by various sensors are input to the input circuit 101. The signals input to the input circuit 101 are sent to the RAM 103 via the input / output port 102. The RAM 103 stores the signals sent via the input / output port 102. The ROM 104 is used as an example of a computer-readable non-transitory storage medium that stores a control program executed by the CPU 105.

[0016] The CPU 105 executes a control program (not shown) stored in the ROM 104 to process the output signal stored in the RAM 103 in accordance with the control program. The CPU 105 performs online learning in accordance with the control program. The CPU 105 temporarily stores the calculation results of the online learning in the ROM 104. The calculation results stored in the ROM 104 are output to the output circuit 106 via the input / output port 102.

[0017] The storage unit 107 is electrically connected to the RAM 103, the ROM 104, and the CPU 105 via the input / output port 102. The storage unit 107 can be a storage device such as a hard disk drive or a solid state drive, a storage area inside a computer, or a storage area on a network.

[0018] The storage unit 107 stores a reference data group, which will be described later. The reference data group is input to an input layer of a neural network model, which will be described later. In this embodiment, multiple types of reference data groups are stored in the storage unit according to the types of calculation results to be output to the output layer of the neural network model. For example, calculation results in powertrain control can include the results of catalyst deterioration diagnosis, motor torque, motor coolant temperature, etc.

[0019] [Neural Network Model] Next, the neural network model according to this embodiment will be described with reference to Fig. 2 and Fig. 3. Fig. 2 is a conceptual diagram showing the weights and biases of each neuron constituting the neural network model. Fig. 3 is a conceptual diagram showing the neural network model.

[0020] A neural network model is a mathematical model that mimics the structure of the neural circuits in the human brain. As shown in Figure 11, a weight w and a bias b are set for each neuron that makes up the neural network model.

[0021] In a neural network model, inputs a1 to an are input to n neurons, and are multiplied by weights w1 to wn set for each neuron. Then, in the next layer of neurons, the inputs a1 to an multiplied by the weights w1 to wn are added (combined), and a bias b is applied to the addition result to obtain the output z. The next layer of neurons outputs a, which is expressed by the function f(z).

[0022] Furthermore, a function called an activation function is defined for each neuron. The activation function may be a logistic function (sigmoid function), a ramp function (ReLU (Rectified Linear Unit) function), or the like, as appropriate. FIG. 11 shows an example in which the neuron is activated as the input x is greater than 0, and the output y (= f(x)) approaches 1, whereas the neuron is deactivated as the input x is smaller than 0, and the output y approaches 0. For example, when the input x is "5", the output y of the activation function is "1", and the neuron outputs "1" to the neuron in the next layer.

[0023] As shown in Figure 3, the neural network model has one layer formed by multiple neurons. The neural network model has an input layer to which training data is input, an output layer to which calculation results are output, and an intermediate layer located between the input layer and the output layer.

[0024] The hidden layer is constructed by stacking multiple layers, each consisting of multiple neurons. A neural network model can approximate more complex input-output relationships by increasing the number of neurons and hidden layers. However, there is a trade-off between approximation accuracy and model size, so a compromise that satisfies both requirements is selected.

[0025] In the neural network model of this embodiment, for example, the frequency of the exhaust gas sensor signal is set in the input layer, and the catalyst deterioration diagnosis is set in the output layer, and the weight and bias of each neuron are learned by machine learning (supervised). This makes it possible to approximate the input-output relationship. For example, a known algorithm such as the backpropagation method can be applied as the machine learning algorithm.

[0026] [Mahalanobis distance] Next, the Mahalanobis distance according to this embodiment will be described with reference to Fig. 4. Fig. 4 is a diagram showing an overview of the Mahalanobis distance.

[0027] The Mahalanobis distance is a distance measure used in statistics and multivariate analysis. As shown in Figure 4, the Mahalanobis distance is characterized by its ability to easily determine whether a value is normal or abnormal, even if the distance from the center point is similar, by combining multidimensional variances.

[0028] When the input data vector is X, the mean vector is μ, and the variance-covariance matrix is ​​Σ, the Mahalanobis distance D Ｍ is calculated by the following formula (1).

[0029]

[0030] The input data vector X is calculated by the following equation (2).

[0031]

[0032] The mean value vector μ is calculated by the following equation (3).

[0033]

[0034] The variance-covariance matrix Σ is calculated by the following equation (4): where E is the expected value.

[0035]

[0036] [Increase in the number of data and coefficient of determination in the conventional method] Next, the relationship between the number of training data and the coefficient of determination in the conventional method will be described with reference to Fig. 5. Fig. 5 is a diagram showing the number of training data and the coefficient of determination in the conventional method.

[0037] The graph on the left side of Fig. 5 shows the frequency of each Mahalanobis distance in data group A consisting of multiple training data. The horizontal axis of the graph on the left side of Fig. 5 represents the Mahalanobis distance, and the vertical axis represents the frequency of the training data. When data group A is input to the input layer, the coefficient of determination, which indicates the estimation accuracy of the calculation result output from the output layer, is 0.833.

[0038] The graph on the right side of Figure 5 shows the frequency of each Mahalanobis distance in data group B, which consists of multiple pieces of training data. The horizontal axis of the graph on the right side of Figure 5 is the Mahalanobis distance, and the vertical axis is the frequency of the training data. Data group B is data group A to which multiple pieces of training data with Mahalanobis distances of 0.5 to 1 have been added.

[0039] When data group B is input to the input layer, the coefficient of determination, which indicates the estimation accuracy of the calculation result output from the output layer, is 0.841. Therefore, when data group B is input to the input layer, the coefficient of determination is improved by only 0.08 compared to when data group A before the addition of multiple pieces of training data with Mahalanobis distances of 0.5 to 1 is input to the input layer.

[0040] In this way, when high-frequency training data is added to a group of already accumulated data, it is clear that the addition does not contribute to improving the estimation accuracy, and the storage unit ends up using up its storage capacity to store data that does not contribute to improving the estimation accuracy.

[0041] [Functional Configuration of Control Unit] Next, the functional configuration of the CPU 105 (control unit) of this embodiment will be described with reference to Fig. 6. Fig. 6 is a block diagram showing the functional configuration of the CPU 105 (control unit).

[0042] Hereinafter, data selected as training data will be referred to as reference data. Data newly input to the CPU 105 as candidates for training data will be referred to as additional data. The additional data is input to the CPU 105 from the input circuit 101 via the input / output port 102. Note that the additional data may also be input to the CPU 105 from the RAM 103.

[0043] The CPU 105 accumulates a plurality of reference data to create a reference data group. The reference data is input to the input layer of the neural network model. The CPU 105 selects reference data to be included in the reference data group from the plurality of reference data and additional data already accumulated. The plurality of reference data and additional data are provided with explanatory data indicating the time of acquisition. Therefore, the CPU 105 can identify the time of acquisition of each piece of data.

[0044] 6, the CPU 105 has a distance calculation unit 501, a grouping unit 502, a frequency calculation unit 503, and a data selection unit 504. The frequency calculation unit 503 corresponds to the frequency calculation unit according to the present invention.

[0045] The distance calculation unit 501 calculates the Mahalanobis distance from the center of gravity of the reference data group to the plurality of reference data and additional data, and sends the calculated Mahalanobis distance to the grouping unit 502.

[0046] The grouping unit 502 divides the plurality of reference data and additional data into a plurality of groups according to the Mahalanobis distance. In this embodiment, the Mahalanobis distance is divided into a predetermined range to set a plurality of groups. The grouping unit 502 sends the grouping results to the frequency calculation unit 503.

[0047] The frequency calculation unit 503 calculates the frequency of the reference data in each group, and sends the calculation result to the data selection unit 504. The frequency of the reference data corresponds to the frequency according to the present invention.

[0048] The data selection unit 504 selects multiple reference data and additional data to be included in the reference data group so that the frequency of the reference data in each group is uniform. For example, if the additional data is assigned to a low-frequency group, the data selection unit 504 includes the additional data as reference data in the reference data group.

[0049] On the other hand, when the additional data is sorted into a high frequency group, the data selection unit 504 interchanges the reference data of the high frequency group with the additional data. Also, when the additional data is sorted into a high frequency group, the data selection unit 504 may not include the additional data in the reference data group.

[0050] [Data Determination Process] Next, the data determination process executed by the CPU 105 of this embodiment will be described with reference to Fig. 7. Fig. 7 is a flowchart showing the data determination process according to the first embodiment.

[0051] First, CPU 105 acquires a plurality of additional data (S1). Next, data selection unit 504 of CPU 105 determines whether the total data volume of the reference data group has reached a predetermined storage capacity of storage unit 107 (S2). If it is determined in step S2 that the total data volume of the reference data group has not reached the predetermined storage capacity of storage unit 107 (NO in S2), CPU 105 proceeds to step S5.

[0052] In step S2, when it is determined that the total data amount of the reference data group has reached the predetermined storage capacity of the storage unit 107 (if S2 is determined as YES), the data selection unit 504 of the CPU 105 determines whether or not any of the multiple reference data included in the reference data group was acquired before a predetermined threshold (S3). The predetermined threshold is determined in advance as the time to update.

[0053] In step S3, when it is determined that the acquisition time of the plurality of reference data included in the reference data group is not earlier than the predetermined threshold (NO in S3), the CPU 105 proceeds to step S5.

[0054] In step S3, when it is determined that the acquisition time of the plurality of reference data included in the reference data group is before a predetermined threshold (if S3 is judged YES), the data selection unit 504 of the CPU 105 replaces the reference data before the predetermined threshold with the additional data (S4). For example, if the predetermined threshold is six months, and there is data older than six months, the data is replaced with the additional data.

[0055] If the determinations in steps S2 and S3 are NO, or after the processing of step S4, the distance calculation unit 501 of the CPU 105 calculates the Mahalanobis distances of the plurality of reference data and the plurality of additional data from the center of gravity of the reference data group (S5).

[0056] Next, the grouping unit 502 of the CPU 105 divides the Mahalanobis distance into predetermined ranges to set a plurality of groups, and groups the plurality of reference data and the plurality of additional data according to the Mahalanobis distance calculated in step S5 (S6).

[0057] Next, the frequency calculation unit 503 of the CPU 105 calculates the frequency distribution of the reference data in each group (S7). Subsequently, the data selection unit 504 of the CPU 105 selects data to be included in the reference data group from the plurality of reference data and the plurality of additional data, based on the frequency distribution of the reference data (S8). In step S8, the data selection unit 504 selects the plurality of reference data and the plurality of additional data so that the frequencies in each group are uniform. After processing step S8, the CPU 105 terminates the data discrimination process.

[0058] [First Selection Example] Next, a first selection example performed by the data selection unit 504 will be described with reference to Fig. 8. Fig. 8 is a diagram showing a first selection example performed by the data selection unit 504 when selecting data to be included in the reference data group.

[0059] The graph on the left side of Fig. 8 shows the frequency distribution of multiple reference data and multiple additional data in each group before selection according to this embodiment. The horizontal axis of the graph on the left side of Fig. 8 represents the Mahalanobis distance, and the vertical axis represents the frequency of the reference data and the additional data.

[0060] Hereinafter, the Mahalanobis distance D Ｍ is 0 to 0.5 (0≦D Ｍ <0.5) is the first group, and the Mahalanobis distance D Ｍ is 0.5 to 1.0 (0.5≦D Ｍ <1.0) is the second group. Ｍ is 1.0 to 1.5 (1.0≦D Ｍ <1.5) is the third group, and the Mahalanobis distance D Ｍ is 1.5 to 2.0 (1.5≦D Ｍ <2.0) is the fourth group. Ｍ is 2.0 to 2.5 (2.0≦D Ｍ <2.5) is the fifth group, and the Mahalanobis distance D Ｍ is 2.5 to 3.0 (1.5≦D Ｍ <2.0) is the fourth group.

[0061] As shown in the graph on the left side of Figure 8, in data group A, which is made up of multiple reference data and multiple additional data before selection, the frequency of the reference data and additional data in the third group is the highest. In this embodiment, the target level of frequency is set to 0.4. The target level is set appropriately depending on the storage capacity of the storage unit 107 that stores the reference data group and the number of groups.

[0062] The frequencies of the reference data and additional data of the third group exceed the target level. On the other hand, the frequencies of the reference data and additional data of the other groups do not reach the target level. When such data group A is input to the input layer of the neural network model, the coefficient of determination, which indicates the estimation accuracy of the calculation results output from the output layer, is approximately 0.42.

[0063] The graph on the right side of Fig. 8 shows the frequency distribution of the plurality of reference data and the plurality of additional data in each group after selection according to this embodiment. The horizontal axis of the graph on the right side of Fig. 8 represents the Mahalanobis distance, and the vertical axis represents the frequency of the reference data and the additional data.

[0064] As shown in the graph on the right side of Figure 8, data group B, which is composed of multiple reference data and multiple additional data, is obtained by deleting some of the multiple reference data and multiple additional data in the third group of data group A (hatched in the graph on the left side of Figure 8).As a result, the frequency of the reference data and additional data in the third group has fallen to the target level of 0.4.The frequencies of the reference data and additional data in the second, fourth, and fifth groups have increased.On the other hand, the frequencies of the reference data and additional data in the first group have decreased.

[0065] When this data group B is input to the input layer of the neural network model, the coefficient of determination of the calculation results output from the output layer is approximately 0.42. This shows that deleting some of the reference data and additional data from the third group does not affect the coefficient of determination. Therefore, even if some of the reference data and additional data are deleted from the group with frequencies exceeding the target level, it is possible to ensure the comprehensiveness of the training data while preventing unnecessary use of the storage capacity of the storage unit 107.

[0066] [Second Selection Example] Next, a second selection example performed by the data selection unit 504 will be described with reference to Fig. 9. Fig. 9 is a diagram showing a second selection example performed by the data selection unit 504 when selecting data to be included in the reference data group.

[0067] The graph on the left side of Fig. 9 shows the frequency distribution of the plurality of reference data and the plurality of additional data in each group before selection according to this embodiment. The horizontal axis of the graph on the left side of Fig. 9 represents the Mahalanobis distance, and the vertical axis represents the frequency of the reference data and the additional data.

[0068] As shown in the graph on the left side of Figure 9, in data group C, which is made up of multiple reference data and multiple additional data before selection, the frequency of the reference data and additional data in the third group is the highest. In this embodiment, the target level of frequency is set to 0.3. The target level is set appropriately depending on the storage capacity of the storage unit 107 that stores the reference data group and the number of groups.

[0069] The frequencies of the reference data and additional data in the second and third groups exceed the target level. The additional data (hatched b) is also grouped into the fourth group. As a result, the frequencies of the reference data and additional data in the fourth group reach the target level. The frequencies of the reference data and additional data in the other groups do not reach the target level. When this data group C is input to the input layer of the neural network model, the coefficient of determination of the calculation results output from the output layer is approximately 0.42.

[0070] The graph on the right side of Fig. 9 shows the frequency distribution of the plurality of reference data and the plurality of additional data in each group after selection according to this embodiment. The horizontal axis of the graph on the right side of Fig. 9 represents the Mahalanobis distance, and the vertical axis represents the frequency of the reference data and the additional data.

[0071] As shown in the graph on the right side of Fig. 9, data group D, which is made up of the plurality of reference data and the plurality of additional data after selection, is obtained by deleting some of the data (hatched a in the graph on the left side of Fig. 9) from the plurality of reference data and the plurality of additional data in the second and third groups in data group C. As a result, the frequency of the reference data and the additional data in the second and third groups has fallen to the target level of 0.3.

[0072] Furthermore, the fourth group of data set D includes additional data (hatched area b in the graph on the left side of FIG. 9) as reference data. As a result, the frequency of the reference data and additional data in the fourth group reaches the target level of 0.3. The frequencies of the reference data and additional data in the first, fifth, and sixth groups are also increasing.

[0073] When such a data group D is input to the input layer of a neural network model, the coefficient of determination of the calculation results output from the output layer is approximately 0.78. In this way, by selecting multiple reference data and multiple additional data to create data group D so as to approach the target level (0.3 in this example), the comprehensiveness of the training data can be improved. As a result, the coefficient of determination of the calculation results using the neural network model can be improved.

[0074] [Target Level] Next, an example of calculating the target level will be described with reference to Fig. 10. Fig. 10 is a diagram showing an example of calculating the target level.

[0075] As described above, the target level is calculated according to the storage capacity (rated value) of the storage unit 107 and the number of groups. The number of groups is an integer equal to or greater than two. The reference data group shown in FIG. 10 is divided into four groups with Mahalanobis distances of 1 to 4. The frequency of a Mahalanobis distance of 1 is 0.25, the frequency of Mahalanobis distances of 2 and 3 is 0.3, and the frequency of a Mahalanobis distance of 4 is 0.15.

[0076] The target level can be set to, for example, the median of the frequencies. The median of the frequencies of the four groups shown in FIG. 10 is 0.25. Therefore, the target level is set to 0.25, which is the median of the frequencies. As a result, the data selection unit 504 (see FIG. 6) reduces some of the multiple reference data or multiple additional data in the groups with Mahalanobis distances of 2 and 3 so that the frequencies of the groups with Mahalanobis distances of 2 and 3 approach 0.25. Furthermore, the data selection unit 504 includes the additional data assigned to the groups with Mahalanobis distances of 1 and 4 as reference data in the reference data group.

[0077] The target levels may be distributed equally among the multiple groups, or may be distributed by applying a weighted value to each group according to its importance, thereby making it possible to appropriately distribute the limited storage capacity of the storage unit 107 among the multiple groups.

[0078] [Data Selection Processing] Next, the data selection processing performed by the data selection unit 504 in step S8 of Fig. 7 will be described with reference to Fig. 11. Fig. 11 is a flowchart showing the data selection processing.

[0079] First, the data selection unit 504 compares the frequency of one of all groups with the target level (S11), and then determines whether the frequency of the one group is higher than the target level (S12).

[0080] In step S12, when it is determined that the frequency of one group is higher than the target level (if S12 is determined as YES), the data selection unit 504 deletes some of the multiple data items assigned to one group (S13). In step S13, the data selection unit 504 determines the number of data items to delete so that the frequency of one group approaches the target level. After processing in step S13, the data selection unit 504 proceeds to processing in step S16.

[0081] In step S12, when it is determined that the frequency of one group is equal to or lower than the target level (if S12 is determined as NO), the data selection unit 504 adds the additional data assigned to one group as reference data (S14). After processing in step S14 or after processing in step S13, the data selection unit 504 determines whether there is any group whose frequency has not been compared with the target level (S15).

[0082] In step S15, when it is determined that there is a group for which the target level and frequency have not been compared (if S15 is a YES determination), the data selection unit 504 returns to the processing of step S11. On the other hand, when it is determined that there is no group for which the target level and frequency have not been compared (if S15 is a NO determination), the data selection unit 504 ends the data selection processing.

[0083] By selecting the reference data and additional data in this way, it is possible to prevent the total amount of data in the reference data group from becoming excessive. Also, by including newly observed additional data as reference data in a group where the frequency (frequency) of the reference data is lower (too low) than the target level, it is possible to increase the comprehensiveness of the reference data (teaching data).

[0084] Second Embodiment A data discrimination device according to a second embodiment will now be described. The data discrimination device according to the present invention is not limited to being applied to online learning performed on an ECU, but may also be applied to offline learning. Fig. 12 is a diagram showing the data discrimination device according to the second embodiment.

[0085] 12 , the data discrimination device according to the second embodiment is a personal computer 2. The personal computer 2 receives additional data (teaching data) from a vehicle 3 via a wireless connection. After receiving the additional data, the personal computer 2 performs offline learning using a neural network model or the like. Then, the personal computer 2 transmits a learning value, which is the output of the neural network model or the like, to the vehicle 3.

[0086] The control unit of the personal computer 2 divides the multiple reference data (teaching data) in the reference data group into groups A to E according to the Mahalanobis distance. The control unit of the personal computer 2 does not add corresponding additional data to groups A to C, whose frequencies have reached the target level. On the other hand, the control unit adds corresponding additional data to group D, whose frequencies have not reached the target level. This allows the control unit to increase the comprehensiveness of the teaching data at an early stage. It also makes it possible to prevent the total amount of data in the reference data group from becoming excessive.

[0087] The present invention is not limited to the embodiments described above and shown in the drawings, and various modifications can be made without departing from the spirit of the invention as defined in the claims.

[0088] Furthermore, the above-described embodiments have been described in detail to clearly explain the present invention, and are not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with another configuration.

[0089] REFERENCE SIGNS LIST 1... data discrimination device, 2... personal computer (data discrimination device), 101... input circuit, 102... input / output port, 103... RAM, 104... ROM, 105... CPU, 106... output circuit, 107... storage unit, 501... distance calculation unit, 502... grouping unit, 503... frequency calculation unit, 504... data selection unit

Claims

1. A data discrimination device that accumulates a plurality of reference data and additional data to create a reference data group, comprising: a distance calculation unit that calculates a reference point of the reference data group and the distance from the reference point to the plurality of reference data and the additional data; a grouping unit that divides the plurality of reference data and the additional data into a plurality of groups based on the distance calculated by the distance calculation unit; a frequency calculation unit that calculates the frequency of the plurality of reference data and the additional data for each of the plurality of groups; and a selection unit that selects the plurality of reference data and the additional data to be included in the reference data group so that the frequency approaches a predetermined target level.

2. The data discrimination device according to claim 1, wherein the target level is set according to the storage capacity of a storage unit that stores the reference data group and the number of the plurality of groups.

3. The data discrimination device according to claim 2, wherein the target level is set for each of the plurality of groups by applying a weighting value.

4. The data discrimination device according to claim 1, wherein the selection section discards at least one reference data from the plurality of reference data in a group in which the frequency is greater than the target level.

5. The data discrimination device according to claim 1, wherein the selection unit adds the additional data that has been grouped into a group in which the frequency is smaller than the target level as reference data.

6. The data discrimination device according to claim 1, wherein the plurality of reference data and the additional data are provided with explanatory data according to the chronological order in which they were observed, and when the total amount of data in the reference data group reaches a predetermined storage capacity of the storage unit, the selection unit replaces, based on the explanatory data, reference data included in the reference data group and having a relatively old observation time with additional data having a relatively new observation time that is assigned to the same group.

7. The data discrimination device according to claim 6, wherein the grouping section further divides the plurality of reference data and the additional data in the plurality of groups into groups based on the explanatory data.

8. The data discrimination device according to claim 1, wherein the reference point is the center of gravity of the reference data group.

9. The data discrimination device according to claim 8, wherein the distance from the reference point is Mahalanobis distance, Euclidean distance, or Manhattan distance.

10. A data discrimination method for accumulating a plurality of reference data and additional data to create a reference data group, wherein a distance calculation unit calculates a reference point of the reference data group and the distance from the reference point to the plurality of reference data and the additional data, a grouping unit divides the plurality of reference data and the additional data into a plurality of groups based on the distance calculated by the distance calculation unit, a frequency calculation unit calculates the frequencies of the plurality of reference data and the additional data for each of the plurality of groups, and a selection unit selects the plurality of reference data and the additional data to be included in the reference data group so that the frequency approaches a predetermined target level.

Images