KR20220068044A - Apparatus for diagnosing artificial neural network-based inference model and method thereof - Google Patents

Abstract

The present invention relates to an apparatus and method for diagnosing an artificial neural network-based inference model. The present invention is to provide an apparatus and method for diagnosing an artificial neural network-based inference model, which can identify sensor data having bad influence on performance of the inference model and improve analysis ability of sensor data having an influence on determination results of the inference model by detecting neurons involved in a wrong decision when an artificial neural network-based inference model makes a wrong decision, and detecting a contribution level of each neuron by performing a cumulative back propagation to the detected neurons. The present invention comprises: an artificial neural network-based inference model; and a control unit which detects neurons involved in a wrong decision when the inference model makes a wrong decision, and detects a contribution level of each neuron by performing a cumulative back propagation to the detected neurons.

Description

Diagnosis apparatus and method of inference model based on artificial neural network

The present invention relates to a technology for detecting sensor data that affects the performance of an artificial neural network-based reasoning model.

In general, an artificial neural network (ANN) is a field of artificial intelligence, and is an algorithm made by a machine to learn by simulating a human neural structure. Recently, it has been applied to image recognition, voice recognition, natural language processing, and the like, and has shown excellent effects. An artificial neural network consists of an input layer that receives an input, a hidden layer that actually learns, and an output layer that returns the result of an operation. A deep neural network (DNN) with multiple hidden layers is also a kind of artificial neural network.

Artificial neural networks allow computers to learn on their own based on data. When trying to solve a problem using an artificial neural network, what you need to prepare is a suitable artificial neural network model and data to be analyzed. An artificial neural network model to solve a problem is trained based on data. Before training the model, it is necessary to first divide the data into two types. That is, the data should be divided into a training dataset and a validation dataset. The training dataset is used to train the model, and the validation dataset is used to verify the performance of the model.

There are several reasons for validating an artificial neural network model. An artificial neural network developer tunes a model by modifying the model's hyper parameters based on the model's validation results. In addition, the model is verified to select which model is suitable among various models. The reason for the need to verify the model is explained in more detail as follows.

The first is to predict accuracy. As a result, the purpose of artificial neural networks is to achieve good performance on out-of-sample data that is not used for training. Therefore, after creating the model, it is essential to check how well the model will perform on out-of-sample data. However, since the model should not be verified using the training dataset, the accuracy of the model should be measured using a validation dataset separate from the training dataset.

The second is to improve the performance of the model by tuning the model. For example, overfitting can be prevented. Overfitting is when the model is overtrained on the training dataset. For example, if the training accuracy is high but the validation accuracy is low, the possibility of overfitting may be suspected. And it can be understood in more detail through training loss and validation loss. If overfitting occurs, it is necessary to prevent overfitting to increase the verification accuracy. Overfitting can be prevented by using methods such as regularization or dropout.

A model (hereinafter, an inference model) on which the learning process and verification process have been completed can be applied to various systems and used for various purposes. For example, an inference model for classifying a road surface (general, snowy road, sandy road, muddy road, etc.) based on various sensor data may be provided in the vehicle to classify the road surface while driving.

Since such an inference model cannot completely classify the road surface according to the driving situation, methods for detecting the contribution by sensor data that have caused this result in the case of incorrectly classifying the road surface have been proposed so far. none.

Matters described in this background section are prepared to improve understanding of the background of the invention, and may include matters that are not already known to those of ordinary skill in the art to which this technology belongs.

In the present invention, when an artificial neural network-based inference model makes an erroneous decision, the neurons involved in the erroneous decision are detected, and the contribution of each neuron is detected by performing cumulative back propagation on the detected neurons. By doing so, it is possible to identify sensor data that adversely affects the performance of the inference model, and at the same time, a diagnostic apparatus for an artificial neural network-based reasoning model capable of improving the analysis ability for sensor data affecting the determination result of the inference model and to provide a method therefor.

The objects of the present invention are not limited to the objects mentioned above, and other objects and advantages of the present invention not mentioned can be understood by the following description, and will be more clearly understood by the examples of the present invention. It will also be readily apparent that the objects and advantages of the present invention may be realized by the means and combinations thereof indicated in the appended claims.

In order to achieve the above object, an apparatus for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention includes: an artificial neural network-based reasoning model; And when the inference model makes an erroneous decision, a controller for detecting the neurons involved in the erroneous decision, and performing cumulative back propagation on the detected neurons to detect the contribution of each neuron. can

In an embodiment of the present invention, the controller may analyze the effect of sensor data input to the corresponding neuron on the erroneous determination based on the contribution of each neuron.

In an embodiment of the present invention, the control unit may detect a failure of the sensor based on the analysis result.

In an embodiment of the present invention, the controller may determine that a sensor corresponding to sensor data input to a neuron having the greatest contribution is a failure.

In an embodiment of the present invention, the inference model may determine the road surface on which the vehicle is driving as any one of general, snowy, sand, and mud based on various sensor data.

In an embodiment of the present invention, the various sensor data may include at least one of a steering angle, a steering angular velocity, a yaw rate, a wheel speed, a BPS (Brake Pedal Sensor) value, a master cylinder pressure, a lateral acceleration, and a longitudinal acceleration.

In order to achieve the above object, a method for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention includes: outputting, by an artificial neural network-based reasoning model, a determination result; detecting, by the controller, neurons involved in the erroneous judgment when the inference model makes an erroneous decision; and performing, by the controller, cumulative back propagation on the detected neurons to detect the contribution of each neuron.

An embodiment of the present invention may further include the step of analyzing, by the controller, an effect of sensor data input to a corresponding neuron on the erroneous determination based on the contribution of each neuron.

In an embodiment of the present invention, the analyzing may include detecting a failure of a sensor based on the analysis result.

An embodiment of the present invention may include determining a sensor corresponding to sensor data input to a neuron having the greatest contribution as a failure.

The apparatus and method for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention as described above, when the artificial neural network-based reasoning model makes an erroneous decision, detects neurons involved in the erroneous decision, and the detection By performing cumulative back propagation on one neuron to detect the contribution of each neuron, it is possible to identify sensor data that adversely affects the performance of the inference model.

In addition, the apparatus and method for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention as described above, when the artificial neural network-based reasoning model makes an erroneous decision, detects neurons involved in the erroneous decision, By performing cumulative back propagation on the detected neurons to detect the contribution of each neuron, it is possible to improve the ability to analyze sensor data affecting the judgment result of the inference model.

1 is a block diagram of a diagnosis apparatus of an artificial neural network-based reasoning model according to an embodiment of the present invention;
2 is an exemplary diagram illustrating a process in which a control unit provided in a diagnosis apparatus according to an embodiment of the present invention tests an artificial neural network-based reasoning model;
3 is an exemplary diagram illustrating a process in which a control unit included in a diagnostic apparatus according to an embodiment of the present invention detects neurons involved in an erroneous judgment of an inference model;
4 is an exemplary diagram illustrating a process in which a control unit included in a diagnosis apparatus according to an embodiment of the present invention detects a contribution degree from neurons that have contributed to an erroneous judgment of an inference model;
5 is a flowchart of a diagnosis method of an artificial neural network-based reasoning model according to an embodiment of the present invention;
6 is a block diagram illustrating a computing system for executing a method for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention.

Hereinafter, some embodiments of the present invention will be described in detail with reference to exemplary drawings. In adding reference numerals to the components of each drawing, it should be noted that the same components are given the same reference numerals as much as possible even though they are indicated on different drawings. In addition, in describing the embodiment of the present invention, if it is determined that a detailed description of a related known configuration or function interferes with the understanding of the embodiment of the present invention, the detailed description thereof will be omitted.

In describing the components of the embodiment of the present invention, terms such as first, second, A, B, (a), (b), etc. may be used. These terms are only for distinguishing the elements from other elements, and the essence, order, or order of the elements are not limited by the terms. In addition, unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

1 is a block diagram of a diagnosis apparatus of an artificial neural network-based reasoning model according to an embodiment of the present invention.

As shown in FIG. 1 , the apparatus for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention includes an inference model 10 , an input unit 20 , an output unit 30 , and a control unit 40 . can do. In this case, according to a method of implementing the apparatus for diagnosing an artificial neural network-based inference model according to an embodiment of the present invention, each component may be combined with each other to be implemented as one, or some components may be omitted.

Looking at each of the above components, first, the inference model 10 is an artificial neural network that has been trained, and may be composed of an input layer, a hidden layer, and an output layer, and in this case, the input layer Each of the hidden layer and the output layer may be composed of a plurality of neurons.

The reasoning model 10 is not limited to a specific reasoning model and may be various inference models. For example, an inference model that classifies a road surface on which a vehicle is traveling into general, snowy, sand, mud, etc. may classify the road surface based on sensor data input from various sensors.

The inference model 10 may be stored in a memory, and the memory may include a flash memory type, a hard disk type, a micro type, and a card type (eg, an SD card ( Secure Digital Card) or XD Card (eXtream Digital Card), RAM (Random Access Memory), SRAM (Static RAM), ROM (Read-Only Memory), PROM (Programmable ROM), EEPROM It may include a storage medium of at least one type of (Electrically Erasable PROM), a magnetic memory (MRAM, Magnetic RAM), a magnetic disk, and an optical disk type memory.

When the artificial neural network-based reasoning model makes an erroneous decision, the memory detects the neurons involved in the erroneous decision, and performs cumulative back propagation on the detected neurons to detect the contribution of each neuron Various logics, algorithms, and programs required in the process can be stored.

The input unit 20 may input sensor data obtained from various sensors into the inference model 10 . In this case, the various sensors may include a steering angle sensor, a steering angular velocity sensor, a yaw rate sensor, a wheel speed sensor, a brake pedal sensor (BPS), a master cylinder pressure sensor, a lateral acceleration sensor, and a longitudinal acceleration sensor.

Here, the steering angle sensor may measure a rotation angle of the steering wheel by the driver's manipulation. The steering angular velocity sensor may measure the rotational speed of the steering wheel by the driver's manipulation. The yaw rate sensor may measure the speed at which the rotation angle changes based on a vertical line passing through the center of the vehicle. The wheel speed sensor is mounted on each wheel (Front Left, Front Right, Rear Left, Rear Right) of the vehicle to measure the rotational speed of each wheel. BPS detects whether or not the brake pedal is depressed by the driver. That is, the BPS may output 1 as sensing information, for example, when the brake pedal is pressed, and 0 when not pressed. The master cylinder pressure sensor is mounted on the master cylinder that converts mechanical force generated when the brake pedal is pressed into hydraulic force, and can measure the pressure generated at this time. The lateral acceleration sensor may measure the acceleration acting in the lateral direction of the vehicle. The longitudinal acceleration sensor may measure acceleration acting in the longitudinal direction (driving direction) of the vehicle.

The output unit 30 may output the contribution level for each sensor data under the control of the control unit 40 .

The control unit 40 may perform overall control so that each of the components can perform their functions normally. The control unit 40 may be implemented in the form of hardware, or may be implemented in the form of software, or may be implemented in the form of a combination of hardware and software. Preferably, the control unit 40 may be implemented as a microprocessor, but is not limited thereto.

In particular, when the artificial neural network-based inference model makes an erroneous decision, the controller 40 detects the neurons involved in the erroneous decision, and performs cumulative back propagation on the detected neurons to perform accumulative back propagation on each neuron. Various controls required in the process of detecting the contribution of each star can be performed.

The controller 40 may analyze the effect of sensor data input to the corresponding neuron on the erroneous determination based on the contribution of each neuron.

The control unit 40 may detect a failure of the sensor based on the analysis result. For example, a sensor corresponding to sensor data input to a neuron having the greatest contribution may be determined as a failure.

Hereinafter, the operation of the controller 40 will be described in detail with reference to FIGS. 2 to 4 .

2 is an exemplary diagram illustrating a process in which a control unit included in a diagnosis apparatus according to an embodiment of the present invention tests an artificial neural network-based reasoning model.

As shown in Fig. 2, the control unit 40 sequentially targets the neurons located in the lowest layer of the inference model 10, sensor data 1, sensor data 2, sensor data 3, sensor data 4, ... , input sensor data n. In this case, the sensor data may include, for example, a steering angle, a steering angular velocity, a yaw rate, a wheel speed, a BPS value, a master cylinder pressure, a lateral acceleration, a longitudinal acceleration, and the like.

The inference model 10 is a model in which the learning process and the verification process have been completed, and it can be seen that a correct judgment (Class b) is made, but an incorrect judgment (Class a) is made. In this case, the control unit 40 may receive information indicating that 'Class b' is a correct decision (correct answer) from the user in advance.

As a result, the control unit 40 may determine that the inference model 10 should be determined as 'Class b' based on the input sensor data, but incorrectly determined as 'Class a'.

)을 검출할 수 있다.Looking at this in more detail, the control unit 40 activates each layer in the artificial neural network composed of N layers based on the following [Equation 1] to [Equation 4] in the process of the inference model 10 making a erroneous judgment. neurons (

) can be detected.

[Equation 1]

가 입력되었을 때 1번 레이어(

)의 출력(활성화 뉴런)은

가 된다. 따라서, i번째 레이어의 출력은 하기의 [수학식 1] 및 [수학식 2]를 통해 검출될 수 있다. 이때, i가 1, 2, …, N-1인 경우에는 하기의 [수학식 2]을 통해 검출될 수 있고, i가 N인 경우에는 하기의 [수학식 3]을 통해 검출될 수 있다.here,

When is entered, layer 1 (

) of the output (activating neuron) is

becomes Accordingly, the output of the i-th layer may be detected through the following [Equation 1] and [Equation 2]. In this case, i is 1, 2, ... , N-1 may be detected through the following [Equation 2], and when i is N, it may be detected through the following [Equation 3].

[Equation 2]

는 i번째 레이어의 활성화 함수(ReLU, Softmax)를 나타낸다. 이때, ReLU는 정류 선형 유닛(Rectified Linear Unit)을 나타낸다.Here, W _i represents the weight of the i-th layer, b _i represents the bias of the i-th layer,

denotes the activation function (ReLU, Softmax) of the i-th layer. In this case, ReLU represents a rectified linear unit.

[Equation 3]

)에서 가장 큰 값을 가지는 원소(a)와 대응되는 클래스(Class a)이며, 이는 하기의 [수학식 4]를 통해 검출될 수 있다.The judgment result of the inference model 10 is the output value of the last layer (

) is a class (Class a) corresponding to element (a) having the largest value, which can be detected through [Equation 4] below.

[Equation 4]

과 동일한 크기를 가지는 벡터로서, 일례로 e_i=a = [1,0,0,0]와 같다.Here, e _i has only a value corresponding to class i

As a vector having the same size as , for example, e _{i = a} = [1,0,0,0].

3 is an exemplary diagram illustrating a process in which a control unit included in a diagnosis apparatus according to an embodiment of the present invention detects neurons involved in an erroneous judgment of an inference model.

) 중에서, 하기의 [수학식 5] 내지 [수학식 8]에 기초하여 오판에 관여한 뉴런들과 오판에 반대한 뉴런들을 검출할 수 있다. 즉, 제어부(40)는 하기의 [수학식 5]에 기초하여 N번째 레이어의 입력(

)에서 오판에 기여한 뉴런(

)을 검출할 수 있고, 하기의 [수학식 6]에 기초하여 i=2,3,…, N-2번째 레이어의 입력(

)에서 오판에 기여한 뉴런(

)을 검출할 수 있다.The control unit 40 controls the neurons (

), based on the following [Equation 5] to [Equation 8], it is possible to detect the neurons involved in the misjudgment and the neurons opposing the misjudgment. That is, the control unit 40 is the input of the Nth layer based on the following [Equation 5] (

) that contributed to the misjudgment in (

) can be detected, and based on the following [Equation 6], i = 2, 3, ... , the input of the N-2th layer (

) that contributed to the misjudgment in (

) can be detected.

[Equation 5]

[Equation 6]

는 N개의 레이어로 구성된 인공신경망 내 각 레이어별로 활성화된 뉴런(

) 중에서 오판에 기여한 뉴런을 나타내고,

는

의 j번째 원소를 나타낸다.here,

is an activated neuron (

) represents the neurons that contributed to the misjudgment,

Is

represents the j-th element of

[Equation 7]

)에서 오판에 기여한 입력(

)을 검출할 수 있다. 결국, 제어부(40)는 하기의 [수학식 7]을 통해 오판에 기여한 뉴런(I(ㆍ))을 검출할 수 있다. 이때, +1은 오판에 대한 기여를 나타내고 -1은 오판에 대한 반대를 나타낸다.The control unit 40 receives the input of the first layer through [Equation 7]

) from the input that contributed to the misjudgment (

) can be detected. As a result, the controller 40 may detect the neuron I(·) contributing to the misjudgment through Equation 7 below. In this case, +1 represents the contribution to the misjudgment and -1 represents the opposite to the misjudgment.

[Equation 8]

In FIG. 3 , 311 , 312 , 313 , 314 , and 315 represent neurons contributing to the misjudgment of the inference model 10 , and 321 , 322 , 323 , and 324 represent neurons opposing the misjudgment of the inference model 10 .

4 is an exemplary diagram illustrating a process in which a control unit included in a diagnosis apparatus according to an embodiment of the present invention detects a contribution degree from neurons that have contributed to an erroneous judgment of an inference model.

)들을 대상으로 누적 역전파(back propagation)를 수행하여 각 뉴런별 기여도를 검출할 수 있다. 일례로, 제어부(40)는 하기의 [수학식 9]를 통해 누적 역전파를 수행하여 추론 모델(10)의 오판에 기여한 각 뉴런별 기여도(

)를 검출할 수 있다.The control unit 40 is a neuron (

), the contribution of each neuron can be detected by performing cumulative back propagation. As an example, the control unit 40 performs cumulative backpropagation through the following [Equation 9], and the contribution degree (

) can be detected.

[Equation 9]

는 활성화된 뉴런(

)의 k번째 원소(뉴런)가 추론 모델(10)의 오판에 기여한 정도(값)를 나타낸다.here,

is the activated neuron (

) represents the degree (value) that the k-th element (neuron) contributed to the misjudgment of the inference model 10 .

)를 획득할 수 있다. 이때,

이 클수록 k번째 입력(센서데이터)이 정답(Class b)보다 오판(Class a)에 더 기여한 것을 의미한다. 도 4에서는 뉴런(311)에 입력된 센서데이터 2가 추론 모델(10)의 오판에 가장 기여도가 큰 것을 알 수 있다.The control unit 40 performs cumulative backpropagation through [Equation 9], and the final contribution (

) can be obtained. At this time,

It means that the kth input (sensor data) contributed more to the misjudgment (Class a) than the correct answer (Class b). In FIG. 4 , it can be seen that the sensor data 2 input to the neuron 311 has the greatest contribution to the misjudgment of the inference model 10 .

5 is a flowchart illustrating a method for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention.

First, the artificial neural network-based reasoning model 10 outputs a determination result ( 501 ).

Thereafter, when the inference model 10 makes an erroneous decision, the controller 40 detects neurons involved in the erroneous decision ( 502 ).

Thereafter, the controller detects the contribution of each neuron by performing cumulative back propagation on the detected neurons ( S503 ).

6 is a block diagram illustrating a computing system for executing a method for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention.

Referring to FIG. 6 , the above-described method for diagnosing an artificial neural network-based reasoning model according to an embodiment of the present invention may be implemented through a computing system. The computing system 1000 includes at least one processor 1100 , a memory 1300 , a user interface input device 1400 , a user interface output device 1500 , a storage 1600 connected through a system bus 1200 , and A network interface 1700 may be included.

The processor 1100 may be a central processing unit (CPU) or a semiconductor device that processes instructions stored in the memory 1300 and/or the storage 1600 . The memory 1300 and the storage 1600 may include various types of volatile or nonvolatile storage media. For example, the memory 1300 may include a read only memory (ROM) 1310 and a random access memory (RAM) 1320 .

Accordingly, the steps of a method or algorithm described in connection with the embodiments disclosed herein may be directly implemented in hardware, a software module executed by the processor 1100 , or a combination of the two. A software module may be a storage medium (i.e., memory 1300 and/or It may also reside in storage 1600 . An exemplary storage medium is coupled to the processor 1100 , the processor 1100 capable of reading information from, and writing information to, the storage medium. Alternatively, the storage medium may be integrated with the processor 1100 . The processor and storage medium may reside within an application specific integrated circuit (ASIC). The ASIC may reside within the user terminal. Alternatively, the processor and storage medium may reside as separate components within the user terminal.

The above description is merely illustrative of the technical spirit of the present invention, and various modifications and variations will be possible without departing from the essential characteristics of the present invention by those skilled in the art to which the present invention pertains.

Accordingly, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but to explain, and the scope of the technical spirit of the present invention is not limited by these embodiments. The protection scope of the present invention should be construed by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

10: Inference model
20: input unit
30: output unit
40: control unit

Claims (12)

Inference model based on artificial neural network; and
When the inference model makes an erroneous decision, a control unit that detects neurons involved in the erroneous decision, and performs cumulative back propagation on the detected neurons to detect the contribution of each neuron
A diagnostic device of an artificial neural network-based inference model comprising a.
The method of claim 1,
The control unit is
An apparatus for diagnosing an artificial neural network-based reasoning model, characterized in that based on the contribution of each neuron, the effect of sensor data input to the corresponding neuron on the erroneous judgment is analyzed.
3. The method of claim 2,
The control unit is
A diagnostic apparatus for an artificial neural network-based reasoning model, characterized in that detecting a sensor failure based on the analysis result.
4. The method of claim 3,
The control unit is
A diagnostic device for an artificial neural network-based inference model, characterized in that the sensor corresponding to the sensor data input to the neuron with the greatest contribution is judged as a failure.
The method of claim 1,
The inference model is
A diagnostic device for an artificial neural network-based inference model, characterized in that it determines the road surface on which the vehicle is driving as any one of general, snowy, sand, and mud based on various sensor data.
6. The method of claim 5,
The various sensor data,
A diagnostic device of an artificial neural network-based inference model including at least one of a steering angle, a steering angular velocity, a yaw rate, a wheel speed, a BPS (Brake Pedal Sensor) value, a master cylinder pressure, a lateral acceleration, and a longitudinal acceleration.
outputting, by an artificial neural network-based reasoning model, a judgment result;
detecting, by the controller, neurons involved in the erroneous judgment when the inference model makes an erroneous decision; and
Detecting, by the control unit, the contribution level of each neuron by performing cumulative back propagation on the detected neurons
A diagnostic method of an artificial neural network-based inference model comprising a.
8. The method of claim 7,
analyzing, by the control unit, an effect of sensor data input to a corresponding neuron on the erroneous judgment based on the contribution of each neuron
A diagnostic method of an artificial neural network-based inference model further comprising a.
9. The method of claim 8,
The analyzing step is
detecting a failure of the sensor based on the analysis result
A diagnostic method of an artificial neural network-based inference model comprising a.
10. The method of claim 9,
The step of detecting a failure of the sensor comprises:
Step of judging the sensor corresponding to the sensor data input to the neuron with the greatest contribution as a failure
A diagnostic method of an artificial neural network-based inference model comprising a.
8. The method of claim 7,
The inference model is
A diagnosis method of an artificial neural network-based inference model, characterized in that the road surface on which the vehicle is driving is determined as any one of general, snowy, sand, and mud based on various sensor data.
12. The method of claim 11,
The various sensor data,
A diagnostic method of an artificial neural network-based inference model including at least one of a steering angle, a steering angular velocity, a yaw rate, a wheel speed, a BPS (Brake Pedal Sensor) value, a master cylinder pressure, a lateral acceleration, and a longitudinal acceleration.

Images