JP2020061042A - Data discriminating apparatus and data discriminating method - Google Patents

Abstract

Kind Code: A1 It is possible to discriminate input data at high speed even when the scale of learning data is large. A data discrimination device (10) discriminates input data for a trained neural network (20). The data discriminating device comprises an autoencoder (11) and an evaluation means (12) for evaluating an error between input data and output data of the autoencoder when the input data is input to the autoencoder. , and determines whether or not the input data is within the learning range of the neural network based on the evaluation result of the evaluation means. [Selection drawing] Fig. 1

Description

  The present invention relates to a data discriminating apparatus and a data discriminating method for discriminating input data for a neural network.

  Research is being conducted in many fields to develop products and optimize control by applying artificial intelligence using neural networks. For example, a neural network trained to classify time series data acquired from a sensor mounted on a vehicle under a predetermined criterion is useful for designing a vehicle or optimizing a vehicle control method. However, as a general property of neural networks, it shows a high correct answer rate for input data with high similarity to the data used for learning, but correct answer for data outside the learning range with low similarity. It is known that the rate drops significantly.

  In Patent Document 1, whether the input data is within the learning range or not depends on how much the signal strength of the output layer of the learned neural network with respect to the input data has a margin with respect to the determination threshold value for binarizing the output. A discriminating device for discriminating whether or not is described. According to this document, when the output value has a sufficient margin with respect to the threshold value, it is determined that the data is within the learning range and the comparison process with the learning data is omitted. The amount of processing is reduced as compared with the method of comprehensively searching for a match with.

JP 2001-75937 A

  However, when the signal strength of the output layer is close to the threshold value, the discriminating apparatus described in the above-mentioned document discriminates whether or not the input data is within the learning range by exhaustive search of the learning data. Therefore, as the scale of the learning data increases, the number of comparisons increases, the processing speed decreases, and the memory area for storing the learning data runs short.

  Therefore, an object of the present invention is to provide a data discriminating apparatus and a data discriminating method capable of discriminating input data at high speed even when the scale of learning data is large.

One aspect of the present invention is a data discriminating apparatus for discriminating input data for a learned neural network,
An auto-encoder,
The input data to the neural network, and an evaluation means for evaluating the error between the output data of the self-encoder when the input data is input to the self-encoder,
Based on the evaluation result of the evaluation means, it is determined whether the input data to the neural network is within the learning range of the neural network,
It is a data discriminating device.

Another aspect of the present invention is a data discriminating method in which a computer discriminates input data to a learned neural network,
A learning step of learning the self-encoder using data common to the data set used for learning the neural network;
An evaluation step of evaluating an error between input data to the neural network and output data of the self-encoder when the input data is input to the self-encoder.
A determination step of determining whether or not the input data to the neural network is within the learning range of the neural network based on the evaluation result of the evaluation step.
This is a data discrimination method.

  According to the present invention, even when the scale of learning data is large, input data can be discriminated at high speed.

FIG. 3 is a block diagram showing a configuration of a data discriminating apparatus according to an embodiment of the present disclosure. FIG. 3 is a schematic diagram showing a network structure of the self-encoder according to the first embodiment. FIG. 4 is a diagram showing an example of time-series data in the first embodiment. 5 is a graph showing the results of verifying the distribution of errors for each model in Example 1.

  Hereinafter, a data discriminating apparatus and a data discriminating method according to the present disclosure will be described with reference to the drawings.

  FIG. 1 is a functional block diagram showing the configuration of an automatic classification device 1 including a data discrimination device 10 of this embodiment. The automatic classification device 1 includes a classifier 20 for classifying input data according to a predetermined standard, and a data discriminating device 10 for discriminating input data.

  The classifier 20 is a neural network trained using a learning data set (hereinafter referred to as learning data), and outputs a binary or multilevel signal representing the result of classifying the input data as classification information. To do. As the classifier 20, any neural network including a multilayer perceptron and a recursive neural network can be used. Further, either supervised learning or unsupervised learning may be used as a learning framework, and a known method such as an error backpropagation method can be adopted as a learning algorithm. In Example 1 described later, a convolutional neural network is used.

In the data discriminating apparatus 10, whether the input data input to the classifier 20 is data belonging to the range of learning data used for learning of the classifier 20 (hereinafter, referred to as in-learning data) or unknown data ( Hereinafter, it is determined as non-learning data) and the determination result is output as data type information. Here, “within the range of the learning data” represents a range in which the learned classifier 20 can be expected to output an appropriate classification result, out of arbitrary input data that can be input to the classifier 20. , Refers to the following, for example.
-When there are multiple models in the device, when the classifier is trained with the data acquired from the sensor attached to a specific model as the learning data, the input obtained from the sensor attached to the same model data.
Input data acquired in the same operation mode as the operation mode when the classifier is trained with the data acquired in the specific operation mode as the learning data when the device has a plurality of operation modes implemented.

  The data discriminating apparatus 10 includes a self-encoder 11, an evaluation unit 12, and a determination unit 13. The self-encoder 11 is a neural network trained so that the output data reproduces the input data. The self-encoder 11 can be a known neural network such as a multilayer neural network having four or more layers in addition to a three-layer perceptron having the same number of nodes in the input / output layers, but the input data is learned by the classifier 20. In order to more accurately determine whether or not the data is internal data, it is preferable to have a common structure with the classifier 20. In Example 1 described later, a convolutional neural network is used for the self-encoder 11 in accordance with the classifier 20.

  The evaluation unit 12, which is the evaluation means of this embodiment, compares the input data with the output data of the self-encoder 11 when the input data is input to the self-encoder 11, and compares the input data with the output data. Calculate the error. The determination unit 13, which is the determination unit according to the present exemplary embodiment, determines whether the input data is within the learning range based on the error calculated by the evaluation unit 12. The output of the determination unit 13 is at least in a format capable of discriminating whether or not the input data is within the learning range. For example, a binary signal corresponding to the in-learning data and the unknown data or the in-learning data is used. It is possible to use a multivalued signal in which a certain accuracy is divided into three or more steps.

  Here, the self-encoder 11 is trained by using data common to the learning data used for learning of the classifier 20. That is, the data set used for learning the self-encoder 11 and the data set used for learning the classifier 20 substantially match.

  In other words, the weighting factor of the self-encoder 11 obtained as a result of the machine learning is obtained by comparing the data belonging to the data set used for learning of the classifier 20 and the relevant data with the error output from the evaluation unit 12 as an objective function. It is approximately equal to one of the local solutions of the optimization problem that minimizes the error with the output data of the self-encoder 11 when input to the self-encoder 11. However, as long as the weighting factor is determined by learning, it does not always match the exact local solution, so the variation that is allowed in the convergence judgment of a general learning algorithm becomes "a value approximately equal to the local solution". Shall be included.

  By using a common data set with the classifier 20 for learning of the self-encoder 11, the self-encoder 11 inputs typical characteristics of input data within the learning range of the classifier 20 as a result of learning. Acquire the ability to express efficiently with less dimensions than data. That is, when the data within the learning range is input to the learned self-encoder 11, the output data of the self-encoder 11 shows high similarity to the input data. On the other hand, when the data outside the learning range, which has low similarity to the learning data, is input, the middle layer of the self-encoder 11 cannot well represent the input data, and the self-encoder 11 has a The similarity between the output data and the input data is reduced.

  Therefore, if the learned self-encoder 11 is given, the distance (error) between the input data and the output data of the self-encoder 11 corresponding to it is calculated by using an appropriate distance function. , It is possible to evaluate the similarity between the input data and the learning data of the classifier 20.

The arrow in FIG. 1 indicates a data processing procedure in the data discriminating apparatus 10 of the present embodiment. Further, it is assumed that the self-encoder 11 has been learned before the input data determination process is performed. That is, the data discriminating method in the present embodiment includes at least the following steps.
(1) A learning step of learning the self-encoder 11 using a data set used for learning the classifier 20.
(2) An evaluation step of evaluating an error between the input data to the classifier 20 and the output data of the self-encoder 11 when the input data is input to the self-encoder 11.
(3) A determination step of determining whether the input data to the classifier 20 is within the learning range of the classifier 20 based on the evaluation result of the evaluation step.

  In the above description, the learned classifier 20 is not limited to one in which the weighting coefficient learned in the learning phase is fixed, but includes one in which the weighting coefficient is updated using the input data input in the operation phase. Let's assume. In that case, the learning data of the classifier 20 refers to all data used for learning of the classifier 20 up to the present time. Further, when updating the weighting factor of the classifier 20 in the operation phase, it is preferable to update the weighting factor of the self-encoder 11 accordingly.

  The automatic classification device 1 as hardware includes at least one processing device and a storage device including a temporary storage area and a non-transitory storage area. The processing device exhibits the control function described below by reading and executing the program stored in the storage device. The storage device serves as a storage place for programs and data and also as a work space when the processing device executes the programs.

  Each functional block illustrated in FIG. 1 may be implemented as independent hardware such as an ASIC on the control circuit, or may be implemented as software as a functional unit of a program executed by the processing device. The hardware that realizes the function of the automatic classification device 1 may be distributed to a plurality of devices that are physically separated and connected via a network. Further, the program can be provided by supplying a program that realizes the function of the automatic classification device 1 to a system or a device via a network or a storage medium, and a processing device installed in the system or the device reads and executes the program. .

  Hereinafter, as an application example of the above-described embodiment, a method of identifying the similarity of an automatic transmission using a self-encoder will be described. In the present embodiment, in an automatic transmission including a lockup clutch that locks up a fluid coupling, a self-encoder is assumed when a classifier learned to classify time series data of lockup control is assumed. Consider using to determine whether the input data to this classifier is within the learning range.

  FIG. 2 is a schematic diagram showing the network structure of the self-encoder 11 in this embodiment. In this embodiment, an auto-encoder including an 8-layer convolutional neural network is used. That is, the self-encoder 11 includes convolutional layers C1 and C2 that perform a convolution operation, pooling layers P1 and P2 that performs a pooling operation, deconvolution layers D1 and D2 that performs a deconvolution operation, and an ampling layer that performs an ampling operation. It is composed of U1 and U2.

  The time-series data acquired from the sensor attached to the vehicle is input to the self-encoder 11. The measurement data of each sensor is given as a real vector having a dimension corresponding to the length of the measurement time. Therefore, the size (dimension) of input data is 1 × (measurement time) × (number of channels), where the number of sensors is the number of channels.

  In the convolutional layers C1 and C2, each output is connected to some inputs, and the connection relation is fixed. The number of neurons connected to the output is called the kernel size. It is preferable to use an odd number such as 3 or 5 for the kernel size. The number of weighting factors adjusted by learning is the same as the kernel size. The stride is a set value indicating how many input sides are shifted and connected when connecting to the next output. The stride of the convolutional layers C1 and C2 is, for example, 1. Stride is usually 1.

  In the pooling layers P1 and P2, similarly to the convolutional layers C1 and C2, each output is connected to the kernel size input, and the connection relation is fixed. Unlike the convolutional layers C1 and C2, the pooling layers P1 and P2 do not use weighting factors. Instead, the pooling layers P1 and P2 output the maximum signal value of the connected inputs. This is called max pooling. The kernel size and stride of the pooling layers P1 and P2 are, for example, 2 and 2, respectively.

  The deconvolutional layers D1 and D2 are obtained by inverting the connection directions of the convolutional layers C1 and C2 with respect to input / output, and a kernel-sized output is connected to each input. In addition, the amplifiering layers U1 and U2 are obtained by reversing the connection directions of the pooling layers P1 and P2 with respect to the input / output, and the kernel-sized output is connected to each input. From the above network structure, the size of the output signal of the deconvolution layer D1 becomes equal to the size of the input data.

Further, the error function L1 which is the evaluation unit of the present embodiment receives the output signal of the deconvolution layer D2 of the self-encoder 11 (output data of the self-encoder 11) and the input data, and calculates these errors. calculate. An example of the error function L1 is Euclidean Loss below. However, N in the formula is the size of the input data, y _n is the value of the input data, and y _n with the hat symbol is the output value of the deconvolution layer D1. Euclidean loss is a combination of input data, output data, and four arithmetic operations for the number of pieces of input data.

(Verification of discrimination performance)
Hereinafter, using the learning data used for learning of the classifier 20 relating to lockup control of the automatic transmission as a material, the similarity between the input data to the classifier 20 and the learning data of the classifier 20 is calculated using the self-encoder 11. The method and results of the evaluated verification experiment will be described. The learning data of the classifier is time series data when lockup control for locking up the lockup clutch of the automatic transmission is performed. Further, the learning data used for learning of the self-encoder 11 is the time-series data when lockup control is performed as in the classifier.

  The lockup clutch is a mechanism that directly connects the input side and the output side of a fluid coupling (torque converter) interposed between the engine output shaft and the input shaft of the automatic transmission. The lockup clutch is engaged when the vehicle speed exceeds a certain level when the vehicle starts, and is disengaged temporarily when the shift speed of the automatic transmission is changed. If the slip amount of the lock-up clutch is appropriately controlled, the vehicle can be comfortably accelerated by the driver, while if the control is not appropriate, the driver may be shocked.

  FIG. 3 shows an example of time series data used for verification (only four channels are shown). Here, as an example of time series data, time series measurement data composed of 14 signals is used. Of these, four signals are used from the measurement data of the engine such as engine torque and engine speed. Further, ten signals are used from the measurement data of the automatic transmission, such as the input rotation speed to the transmission mechanism of the automatic transmission (AT input speed in FIG. 3) and the control command value to the automatic transmission. The parameters represented by the 14 signals include, specifically, the engine rotation speed (engine speed) and angular acceleration, engine torque, accelerator opening, torque converter slip speed and angular acceleration, torque converter target slip rotation, It includes the input rotational speed and angular acceleration of the transmission mechanism of the automatic transmission, the output rotational speed and angular acceleration of the transmission mechanism, the oil temperature of the automatic transmission fluid, and the hydraulic command value for the hydraulic control device of the automatic transmission. The input data to the classifier 20 of this embodiment is time-series data of these 14 parameters, and the data discriminating apparatus 10 discriminates the same 14-parameter time-series data using the self-encoder 11. .

As the time-series measurement data set, the time-series measurement data randomly collected from the physical vehicle running test in the development phase is used. Physical vehicle driving tests are conducted on straight roads on the test course. Each of the 14 signals included in each time-series measurement data is data of 1 × 400 pixels. That is, the size of the input data in this verification is 1 × 400 × 14. As a time series measurement data set, 5,536 time series data were collected. These time series data are all obtained from vehicles equipped with the same type of automatic transmission. This data set is divided into 4,429 data for learning and 1,107 data for verification. In addition, the time series data is normalized by Expression 1. However, x is the original value, μ is the average, σ is the standard deviation, and x ′ is the normalized value.
x ′ = (x−μ) / σ (Equation 1)

  In the learning data, "OK" means that the driver feels comfortable when the lockup control is performed and "NG" means that the driver does not feel comfortable. Whether or not the time-series data as shown in FIG. 3 is appropriate from the viewpoint of the shift feeling by learning the classifier by using the result of pre-classification by a skilled engineer according to a predetermined criterion as a teacher signal. A classifier for classifying is obtained. As a concrete configuration of such a classifier, the following papers (Kawakami, T., Ide, T. et al., SAE Technical Paper, 2018-01-0399, 2018, which were previously submitted by the group of the inventors, were presented. ). This classifier is a convolutional neural network having 6 convolutional layers and 1 fully connected layer, and the size of input data is 1 × 400 × 14.

  Here, the classifier using the neural network generally requires a large amount of training data in order to achieve high performance. Data padding techniques are commonly used in the field of deep learning. Therefore, as the learning data for the classifier and the self-encoder, data that is artificially padded by random time shift is used for the data included in the time series measurement data set described above. The total number of inflated learning data is 48,719.

  The structure of the self-encoder used for verification is as described with reference to FIG. The number of layers may be further increased, but the verification result this time shows the result obtained by using eight layers.

  Euclidean loss is used for the error function L1. This error function L1 is a function for evaluating the error between the input data and the output signal of the deconvolution layer D1 in the learning process of the self-encoder, and is used for learning the input data and the self-encoder after learning. It is also a function to evaluate the similarity with the obtained data. That is, in this embodiment, the smaller the error E calculated by the error function L1, the higher the similarity between the input data to the learned classifier and the learning data used for learning of the classifier. .

  In the learning of the self-encoder, the gradient of the error function is calculated by the back propagation method using the above-mentioned learning data. Also, the weighting coefficient is optimized so as to minimize the error E by using Adam (Adaptive Moment Estimation).

(inspection result)
In FIG. 4, time-series data obtained from a vehicle equipped with one of the three types of automatic transmissions for the learned self-encoder is input to the self-encoder 11, and the error is calculated by the error function L1. The result of evaluation is shown. The horizontal axis shows the error and the vertical axis shows the number of cases. The model (A) is the same model as that used when acquiring the learning data. The model (B) has the same mechanical configuration as the model (A), but has a different control method. The model (C) is different from the model (A) in mechanical configuration and control method. That is, in this example, the data obtained from the model (A) is the in-learning data, and the data obtained from the model (C) is the non-learning data. The handling of the model (B) will be described later. The difference in mechanical structure indicates, for example, whether the lockup clutch is a multi-plate clutch or a single-plate clutch.

  The number of data for each model was 226, and the error obtained for each data was evaluated. Many of the errors obtained from the data of the model (A) are small, and the representative values are distributed in the section of 0 or 4 (the range of the error is 0 or more and less than 8). The error obtained from the data of the model (B) has many small values, though not as large as that of the model (A). On the other hand, the error obtained from the data of the model (C) was a large value, and the representative value was distributed in the section of 8 or more (the range of the error of 8 or more).

  Therefore, when the error is smaller than the predetermined determination threshold, it is determined that the data is obtained from the same model as the learning data, and when the error is larger than the determination threshold, it is the data obtained from a different model from the learning data. It is possible to discriminate whether or not the input data is the in-learning data by using the threshold value function for determining as the determining unit 13 (FIG. 1). According to the result of FIG. 4, it is possible to accurately determine at least the model (A) and the model (C) by setting the error determination threshold value to 8.

  In the verification result of the present embodiment, the model (B) can be treated as in-learning data based on the same mechanical configuration, or can be treated as non-learning data by focusing on the difference in control method. Is. In either case, although the identification accuracy is lower than that when only data of either model (A) or model (C) is input, input data that is likely to be within the learning range and outside the learning range It is possible to identify the input data that is highly likely to be

  As described above, by using the self-encoder 11 in the data discriminating apparatus 10, it is possible to evaluate the degree of similarity between the learning data used for learning of the classifier 20 and the input data for the learned classifier 20. I was able to. Thereby, when the learned classifier 20 outputs the result of classifying the input data, whether the input data is within the learning range of the classifier 20 (that is, whether the classification result of the classifier 20 is reliable or not). It has become possible to identify (whether or not).

  Since neural networks are black boxes (that is, it is difficult to give intuitive meanings to the weighting factors obtained by learning), reliability is a problem when using neural networks for important control. . One of the factors that reduce the reliability is a mismatch between the range covered by the data used for learning and the range of data that may be input in the operation phase. The data discriminating apparatus and the data discriminating method using the self-encoder described in this embodiment provide one method for solving such a problem by determining the reliability of the output of the neural network.

(Summary of this embodiment)
In this embodiment,
A data discriminating device (10) for discriminating input data to a learned neural network (20),
An auto-encoder (11),
Evaluation means (12) for evaluating an error between input data to the neural network (20) and output data of the self-encoder (11) when the input data is input to the self-encoder (11). And have,
Based on the evaluation result of the evaluation means (12), it is determined whether the input data to the neural network (20) is within the learning range of the neural network (20).
It is a data discriminating device.

  According to this configuration, the error between the input data to the neural network and the output data of the corresponding auto-encoder is obtained, so that the input data is not compared with a large number of learning data. It is determined whether or not the learning range is within the learning range. Therefore, the discrimination process can be performed at high speed even when the scale of the learning data becomes large. Further, since it is not necessary to hold the learning data for the purpose of determining at least whether it is within the learning range, the shortage of the memory area is unlikely to occur.

Furthermore, according to the data discriminating apparatus (10) of the present embodiment,
The weighting factor of the self-encoder (11) uses the error evaluated by the evaluation means (12) as an objective function, and the data belonging to the data set used for learning of the neural network (20) and the data. It is approximately equal to one of the local solutions for minimizing the error from the output data of the self-encoder (11) when input to the self-encoder (11).

  With this configuration, it is possible to provide a data discriminating apparatus capable of discriminating with high accuracy whether or not the input data is within the learning range of the neural network.

Furthermore, according to the data discriminating apparatus (10) of the present embodiment,
The evaluation means (12) uses four rules based on the input data to the neural network (20) and the output of the self-encoder (11) when the input data is input to the self-encoder (11). The error is calculated by a combination of calculations.

  With this configuration, the amount of calculation for calculating the error is small, and the input data determination process can be performed at high speed.

Furthermore, according to the data discriminating apparatus (10) of the present embodiment,
The input data to the neural network (20) is time series data that can be acquired by using a sensor attached to the device,
It is determined whether the input data to the neural network (20) corresponds to a particular type of device.

  According to this configuration, it is possible to determine whether or not the time series data input to the neural network is derived from a specific type of device used for learning the neural network, and the output of the neural network is output. Can provide information about the reliability of the.

In addition, another aspect of this embodiment is
A data discriminating method in which a computer discriminates input data to a learned neural network (20),
A learning step of learning the self-encoder (11) using data common to the data set used for learning the neural network (20);
An evaluation step of evaluating an error between input data to the neural network (20) and output data of the self-encoder (11) when the input data is input to the self-encoder (11);
A determination step of determining whether or not the input data to the neural network (20) is within the learning range of the neural network (20) based on the evaluation result of the evaluation step.
This is a data discrimination method.

  According to this method, the error between the input data to the neural network and the output data of the corresponding auto-encoder is calculated, so that the input data can be processed without comparing the input data with many learning data. It is determined whether or not the learning range is within the learning range. Therefore, the discrimination process can be performed at high speed even when the scale of the learning data becomes large. Further, since it is not necessary to hold the learning data for the purpose of determining at least whether it is within the learning range, the shortage of the memory area is unlikely to occur.

Furthermore, according to the data discrimination method of the present embodiment,
In the evaluation step, a combination of four arithmetic operations is performed from the input data to the neural network (20) and the output of the self-encoder (11) when the input data is input to the self-encoder (11). Calculate the error by

  According to this method, the amount of calculation for calculating the error is small, and the input data determination process can be performed at high speed.

Furthermore, according to the data discrimination method of the present embodiment,
The input data to the dataset and the neural network (20) is time series data that can be obtained using a sensor attached to the device,
In the determining step, it is determined whether the input data to the neural network (20) corresponds to the particular type of device used to acquire the dataset.

  According to this method, it is possible to determine whether or not the time-series data input to the neural network is derived from a specific type of device used for learning the neural network, and the output of the neural network is output. Can provide information about the reliability of the.

Furthermore, according to the data discrimination method of the present embodiment,
The input data to the neural network (20) and the data used for learning of the self-encoder (11) in the learning step are both the rotational speed and angular acceleration of the engine mounted on the vehicle, and the data mounted on the vehicle. The input rotational speed and the angular acceleration input to the transmission mechanism of the automatic transmission, and the output rotational speed and the angular acceleration output by the transmission mechanism are included.

  According to this method, the above-mentioned parameters exist in common in the vehicle equipped with the engine and the automatic transmission, and therefore the development man-hours are reduced by discriminating the data of the apparatus in which a plurality of models exist when developing the vehicle. It becomes possible. For example, in order to optimize the automatic transmission control method and the setting variables on the control program in order to adapt the automatic transmission to the vehicle, the result of adapting the automatic transmission of a specific model to the vehicle is It is possible to judge whether or not it can be applied even in conformity with the model of (3) by the judgment result using the self-encoder. In this case, it is possible to reduce the number of man-hours for data collection when adapting the automatic transmissions of a plurality of models to the vehicle, as compared with the case of individually collecting the data of each model.

  The technique according to the present disclosure is not limited to time-series data obtained from a sensor attached to a vehicle, and can be used for the purpose of discriminating input data for an arbitrary neural network.

10 ... Data discriminating device / 11 ... Self-encoding device / 12 ... Evaluation means (evaluation unit) / 20 ... Neural network (classifier)

Claims (8)

A data discriminating device for discriminating input data for a learned neural network,
An auto-encoder,
The input data to the neural network, and an evaluation means for evaluating the error between the output data of the self-encoder when the input data is input to the self-encoder,
Based on the evaluation result of the evaluation means, it is determined whether the input data to the neural network is within the learning range of the neural network,
Data discriminator. When the weight coefficient of the self-encoder is input to the self-encoder with data belonging to a data set used for learning of the neural network with the error evaluated by the evaluator as an objective function. Which is approximately equal to one of the local solutions for minimizing the error with the output data of the self-encoder,
The data discriminating apparatus according to claim 1. The evaluation means calculates an error by a combination of four arithmetic operations from the input data to the neural network and the output of the self-encoder when the input data is input to the self-encoder.
The data discrimination device according to claim 1. Input data to the neural network is time-series data that can be obtained using a sensor attached to the device,
Determining whether the input data to the neural network corresponds to a particular type of device,
The data discriminating apparatus according to any one of claims 1 to 3. A data discrimination method in which a computer discriminates input data to a learned neural network,
A learning step of learning the self-encoder using data common to the data set used for learning the neural network;
An evaluation step of evaluating an error between input data to the neural network and output data of the self-encoder when the input data is input to the self-encoder.
A determination step of determining whether or not the input data to the neural network is within the learning range of the neural network based on the evaluation result of the evaluation step.
Data discrimination method. In the evaluation step, an error is calculated by a combination of four arithmetic operations from the input data to the neural network and the output of the self-encoder when the input data is input to the self-encoder,
The data discriminating method according to claim 5. Input data to the dataset and the neural network is time series data that can be obtained using a sensor attached to the device,
In the determining step, determining whether the input data to the neural network corresponds to a particular type of device used to obtain the dataset.
The data discriminating method according to claim 5 or 6. The input data to the neural network and the data used for learning of the self-encoder in the learning step are the rotational speed and the angular acceleration of the engine mounted on the vehicle, and the automatic transmission mounted on the vehicle. The input rotation speed and the angular acceleration input to the speed change mechanism, and the output rotation speed and the angular acceleration output by the speed change mechanism are included.
The data discrimination method according to any one of claims 5 to 7.

Images