KR20110021402A - Tracking detector using neural-network and detecting method thereof - Google Patents

Abstract

PURPOSE: A device and a method for detecting a tracking are provided to accurately detect the tracking by learning specific information in a tracking current waveform through an artificial neural network. CONSTITUTION: A current waveform detector(10) detects a current waveform in a distribution board and a load end. A specific point detector(20) detects at least one specific point related to tracking from the detected current waveform. A tracking detector(30) determines whether the tracking occurs on a region with the current waveform and displays the determination result to the outside.

Description

Tracking Detector Using Neural-network and Detecting Method

The present invention relates to an apparatus and method for detecting a tracking phenomenon caused by poor insulation between lines of indoor wiring. More particularly, the present invention relates to a characteristic of a current waveform in a case where no tracking occurs and when a tracking occurs. A tracking detection apparatus and method using a neural-network for detecting whether a tracking phenomenon occurs in the indoor wiring by judging the current waveform actually detected at the distribution panel and the load terminal using the result of learning the information. will be.

In general, electrical wires installed indoors of electrical consumers insulate the wire conductors from each other using a wire coating made of an insulating material in order to prevent damage due to leakage currents flowing between the wire conductors or between the wire conductors and the earth. It is configured to.

Most of the wire sheaths are made of synthetic resin, which is an organic insulating material. When the wire sheaths are deformed or damaged due to various reasons, and the insulation property is deteriorated, a tracking phenomenon in which leakage current flows along the wire sheath is caused. Is generated.

In this case, the tracking phenomenon means that as the contaminants such as moisture or dust accumulate on the surface of the wire sheath or the wire is overheated, the insulating material constituting the wire sheath is denatured or deteriorated, resulting in poor insulation between the wire conductors. It means a phenomenon that a leakage current flows through the wire sheath.

As such, when the leakage current flowing through the wire covering is gradually increased, the insulating layer between the wire conductors is carbonized by microdischarge to form a carbonization conductor, and when the leakage current increases further along the carbonization conductor, Along with the breakdown of the insulation between the conductors, there is a serious problem that the wire covering will ignite (tracking fire).

On the other hand, even in the case of the electrical equipment according to the prior art, in order to prevent damage caused by leakage current, safety devices such as a circuit breaker or a fuse are provided in a distribution panel or a load, but leakage currents (several to Since the size of several hundred mA) is very small compared to the rated current of the safety devices (several to several decades A), even if the wire covering material is ignited by the tracking phenomenon, the safety devices do not operate, leading to a fire.

In addition, since the leakage current generated by the tracking phenomenon is very fine as described above and the current characteristics are very random according to the type of the load, even if various current sensors and detectors are used in the prior art. There is a problem that it is not possible to accurately detect whether or not the tracking phenomenon occurs.

The present invention is to solve the above-mentioned problems of the prior art, an object of the present invention by learning the characteristic information of the current waveform appearing when tracking occurs in the neural-network in advance, if any current waveform is inputted It is an object of the present invention to provide a tracking detection apparatus and method for quickly and accurately determining whether a tracking phenomenon occurs in a portion where a current waveform is detected through the artificial neural network.

Another object of the present invention is to determine the number of hidden layer nodes, the learning rate and the momentum coefficient of the artificial neural network by using an optimization algorithm, and select some characteristics having a high influence on the determination of tracking occurrence among the characteristics of the current waveform. It is to provide a tracking detection apparatus and method that can further improve the reasoning ability and the calculation speed of the artificial neural network by using as an input of the neural network.

Tracking apparatus using an artificial neural network according to the present invention for achieving the above object is a current waveform detection unit for detecting a current waveform at the distribution panel or load end, detecting at least one feature point associated with tracking from the detected current waveform A feature point detector, and a tracking detector for determining whether tracking has occurred in a region where the current waveform is detected according to an output value obtained by substituting the detected feature points into the completed artificial neural network, and displaying the determination result to the outside. However, the feature points are data related to the impulse irregular waveform appearing in the current waveform when the tracking occurs, and the learning is input to the feature points respectively detected from a plurality of current waveforms for the case where the tracking is not generated and when the tracking occurs. To identify whether tracking has occurred It is characterized by being made by supervised learning.

In addition, the tracking detection unit during the learning, characterized in that for determining the number of hidden layer nodes, the learning rate and the momentum coefficient of the artificial neural network using a PSO algorithm.

In addition, there are a plurality of feature points, and during the learning, the tracking detector selects some feature points of the plurality of feature points using a PSO algorithm to input the artificial neural network.

In addition, a tracking detection method using an artificial neural network according to the present invention includes a first step of detecting a current waveform at a distribution panel or a load end, and a second step of detecting at least one feature point related to tracking from the current waveform detected at the first step. And a third step of determining whether tracking has occurred in a region where the current waveform is detected according to an output value obtained by substituting the feature point detected in the second step into an artificial neural network that has been learned. Values associated with irregular waveforms in the form of impulses appearing in the current waveform when the tracking occurs, respectively, and the learning is performed by inputting the feature points detected from the plurality of current waveforms for the case where the tracking does not occur and when the tracking occurs. By supervised learning to identify It characterized.

As described above, the tracking detection apparatus and method using the artificial neural network according to the present invention, after learning the characteristic information (that is, irregular impulse-shaped waveform) appearing in the current waveform when the tracking occurs in advance in the artificial neural network, input the actual current waveform Since it is a method of determining whether or not tracking has occurred according to the learning result of the artificial neural network at the time, even if there is no prior information on the actual current waveform, it is possible to accurately and quickly detect whether or not tracking has occurred, thereby preventing damage caused by tracking. There is an advantage that can be prevented.

In addition, the tracking detection apparatus and method according to the present invention optimizes the inference capability and calculation speed of the artificial neural network by optimizing the input data, the number of nodes of the hidden layer, the learning rate and the momentum coefficient, etc. necessary for learning and inference of the artificial neural network by the PSO algorithm. The advantage is that it can be further improved.

Hereinafter, with reference to the accompanying drawings a preferred embodiment of the tracking detection device and method using an artificial neural network according to the present invention will be described in detail.

In general, in the case where tracking occurs and in the case where tracking does not occur, the voltage waveform at the load end is almost unchanged in both cases. However, in the case of the current waveform, when the tracking occurs, an irregular peak in the form of an impulse appears.

Specifically, when tracking is not generated, as shown in FIG. 1A, a current sinusoidal waveform is shown. However, when tracking is generated, as shown in FIGS. 1B to 1D, near the pole of the waveform. Impulse waveforms appear irregular.

In addition, the current waveform of the modified form as described above may appear in various patterns depending on the type of load and the progress of tracking.

The present invention utilizes the difference of the current waveform according to whether or not tracking has occurred, and to accurately and quickly determine whether tracking has occurred using only the current waveform detected at the distribution panel or the load terminal.

2 is a block diagram showing a configuration of a tracking detection apparatus using an artificial neural network according to an embodiment of the present invention.

The tracking detection device of FIG. 2 includes a current waveform detector 10 for detecting a current waveform at a distribution panel or a load end, a feature point detector 20 for detecting a feature point from the detected current waveform, and tracking using the detected feature point. It is configured to include a tracking detection unit 30 for determining whether or not this has occurred.

The current waveform detector 10 receives a current waveform at a distribution panel or a load end, and is configured to receive data on a current waveform by using a current transformer (CT) 11 and a DA converter 12. .

At this time, the sampling period and the number of data may be set differently as needed. However, in the present invention, as an example, 50000 pieces of data are input at a sampling period of 50 kHz / s at 1 second intervals.

Meanwhile, the feature point detector 20 detects feature points necessary to determine whether tracking has occurred from the 50000 pieces of sampling data input from the current waveform detector 10. In the present invention, the feature point detector 20 may include a data operator. 21), the data filter 22, the FFT converter 23, and the wavelet converter 24 are configured to detect 12 feature points of <Feature 1> to <Feature 12>.

Specific contents and meanings of the <feature 1> to <feature 12> are as follows.

< Characteristic 1> RMS _o : This is a value obtained by calculating the effective value of the detected current waveform using the data calculator 21. This is a characteristic point necessary for determining whether tracking has occurred through contrast with RMS_f which will be described later. .

<Feature 2> High _ peak : This is a value obtained by obtaining the maximum value of the detected current waveform by using the data calculator 21, and an impulse waveform that appears when tracking occurs on an upper peak of the detected current waveform. It is a feature point that is necessary to determine whether or not.

<Feature point 3> Low _ peak : This is a value obtained by obtaining the minimum value of the detected current waveform by using the data calculator 21. Whether or not an impulse waveform appears when tracking occurs on a lower peak of the detected current waveform. It is a necessary feature point to determine.

<Characteristic 4> RMS _f : This value is obtained by using the data filter 22 and the data operator 21. The Butterworth low pass filter having a cutoff frequency of 65 Hz is applied to the detected current waveform. It is the value calculated after the effective value.

This is to remove the high frequency current signal generated by the tracking from the input current waveform, which is a feature point necessary to determine whether tracking has occurred compared to the RMS_o.

Whether it natneunji a waveform of an impulse type shown in case of tracking to the input current waveform represented by a value obtained by computing the kurtosis of a current waveform input by the data calculator 21, <feature point 5> kurtosis (Kurtosis) It is a feature point necessary to judge.

<Feature 6> Asymmetry ( Skewness ) : This value is obtained by calculating the asymmetry of the current waveform input by using the data calculator 21, the tracking by using the phenomenon that the current waveform of the sinusoidal waveform is distorted when tracking occurs. It is a feature point necessary to determine whether or not it has occurred.

<Feature 7> Average of spectral magnitude : This value is obtained by using the data filter 22 and the FFT converter 23. After applying a high pass filter having a cutoff frequency of 1 kHz to the detected current waveform, the FFT transform is performed. The mean value of the size is obtained.

The average of the spectral magnitudes is that the impulse-shaped waveform that appears when tracking occurs is a high frequency, and is a feature point necessary to determine whether tracking has occurred by extracting the waveform generated by tracking.

<Feature 8> Standard deviation of spectral magnitude : This value is obtained by using the data filter 22 and the FFT converter 23. After applying a high pass filter having a cutoff frequency of 1 kHz to the detected current waveform, the FFT is applied. The standard deviation of magnitudes is obtained by converting.

The standard deviation value of the spectral magnitude is also obtained by using a high frequency impulse waveform that appears when tracking occurs, and after determining the waveform generated by tracking with the average of the spectral magnitudes to determine whether or not the irregular waveform generated by tracking. It is a necessary feature point.

<Feature 9> High _ peak average : This is a value obtained by using the wavelet transducer 24, a phenomenon in which an impulse waveform is irregularly generated at the upper peak and the lower peak of the current waveform when tracking occurs.

That is, a value obtained by finding upper peaks among 60 upper peaks and lower peaks appearing during a sampling time (1 second) using a wavelet, and then averaging the upper peak values. It is a feature point necessary for determining whether or not a part of an impulse waveform by tracking is irregularly included.

<Feature point 10> High peak _ standard deviation: This is using the phenomenon that the shape of the impulse waveform irregularly generated in the upper peak and the lower peak of the obtained value, the tracking in case of the current waveform using the Wavelet converter 24.

That is, after finding the upper peaks among the 60 upper peaks and the lower peaks that appear during the sampling time (1 second) by using the wavelet, the input current waveform together with the high_peak average described above is obtained by calculating the standard deviation of the upper peak values. It is a feature point necessary for determining whether or not a part of an impulse waveform by tracking is irregularly included.

<Feature 11> Low _ peak average : This value is obtained by using the wavelet converter 24, the phenomenon that the impulse waveform is irregularly generated at the upper peak and the lower peak of the current waveform when the tracking occurs.

That is, after finding the lower peaks among the 60 upper peaks and the lower peaks appearing during the sampling time (1 second) by using the wavelet, the input current waveform together with the low_peak standard deviation described later as a value obtained by averaging the lower peak values. It is a feature point necessary for determining whether or not a part of an impulse waveform by tracking is irregularly included.

<Feature 12> Low _ peak standard deviation : This is a value obtained by using the wavelet transducer 24, and uses the phenomenon that the impulse waveform is irregularly generated at the upper peak and the lower peak of the current waveform when tracking occurs.

That is, after finding the lower peaks among the 60 upper peaks and the lower peaks that appear during the sampling time (1 second) using the wavelet, the input current waveform together with the low_peak average described above is obtained by obtaining the standard deviation of the lower peak values. It is a feature point necessary for determining whether or not a part of an impulse waveform by tracking is irregularly included.

Next, the tracking detector 30 determines whether or not tracking has occurred on the line detecting the current waveform by using 12 feature points input from the feature point detector 20, and the controller 31. The memory 32 which pre-stores an algorithm for determining whether or not tracking has occurred and information on a pre-execution result of the algorithm, and a display unit for displaying it externally by a control signal of the controller 31 when tracking occurs. 33).

In the present invention, the control unit 31 uses an artificial neural network as an inference engine for determining whether tracking has occurred, and uses a PSO (Particle Swarm Optimization) algorithm to optimize the artificial neural network.

In general, the artificial neural network is an intelligent algorithm that mimics the process of information transmission in the human nervous system. As shown in FIG. 3, the artificial neural network is largely composed of an input layer, a hidden layer, and an output layer composed of a plurality of nodes. It is widely used for pattern recognition.

The learning method of the artificial neural network configured as described above is largely classified into supervised learning useful for pattern classification and prediction of input data, and unsupervised learning useful for clustering and separating input data.

In the present invention, as the learning method for determining whether or not tracking occurs, the supervised learning method of learning in the state of knowing the output value of the input is used. As the learning algorithm, the input data is propagated in the direction of the input layer from the input layer to the output value. After obtaining, the error back propagation algorithm was used to reduce the error by propagating the error between the output value and the target value in the direction of the output layer from the output layer to change the connection weight between each node.

Looking at the learning method in the artificial neural network in detail, first, in the first step, the connection weights between the input layer and the hidden layer, and between the hidden layer and the output layer are initialized using a random function.

Next, in the second step, forward propagation is performed on the input data input to the input layer in the direction of the output layer.

At this time, the input layer outputs the input data as it is, and in the hidden layer, the sum of the value obtained by multiplying the output value of the input layer by the connection weight between each node of the input layer and the hidden layer, as shown in [Equation 1]. It is done.

Here, net _j is an input value of neuron j (hidden layer), o _i is an active value of neuron i (input layer), w _ji is a connection weight value between neurons i and j, and θ _j is a threshold value of neuron j.

In the hidden layer, the input value obtained in [Equation 1] is taken as an activation function and outputted. As the active function of the hidden layer, a sigmoid function such as [Equation 2] is used.

Here, o _j is the active value of neuron j (hidden layer).

Similarly, in the output layer, the sum of the output value of the hidden layer (that is, the active value) multiplied by the connection weight value between the hidden layer and each node of the output layer as the input of the output layer as shown in [Equation 1], and the input value 2] Calculate the final output of the neural network by taking the form of the activation function.

Next, in the third step, the error is calculated by comparing the final output value with the target value, and if the error is within an appropriate range, the learning is stopped, otherwise the backpropagation algorithm according to the delta rule generalized from the output layer to the input layer direction. After adjusting the connection weight value, the process is repeated until the error is a suitable level using a new input pattern.

In this case, since the error back propagation algorithm using the generalized delta rule is widely known, a detailed description thereof will be omitted.

Meanwhile, in order to determine whether tracking occurs using data actually input by the controller 31, as described above, a plurality of experiments that know a target value (that is, a value corresponding to whether tracking occurs in the present invention) as described above. After the learning using the data (that is, 12 feature points extracted from the current waveform in the present invention) is performed in advance, the connection weight information determined according to the learning result should be stored in the memory 32.

To this end, in the present invention, learning was performed using 1000 feature point data pairs extracted from 1000 experimental current waveforms for tracking occurrences and no occurrences, and the training data is previously stored in the memory 32. It is desirable to have.

At this time, in order to learn the artificial neural network, the number of nodes, the momentum coefficient and the learning rate in the hidden layer should be determined in advance. Even in the case of learning using the same learning data, the learning performance of the artificial neural network depends on how the values are set. Since the reasoning ability is affected, there is a need to set these values optimally.

In addition, all 12 above-mentioned feature points may be used as input data in determining whether tracking has occurred, but in order to improve inference performance and calculation speed of an artificial neural network, only those feature points affecting the judgment may be selected and used as input data. There is also a need.

Accordingly, in the present invention, the process of the control unit 31 to optimize the selection of the feature point, the number of nodes of the hidden layer, the learning rate and the momentum coefficient, etc. using the above-described PSO algorithm stored in the memory 32 for this purpose. Do this.

The Particle Swarm Optimization (PSO) algorithm is an optimization algorithm that mimics the movement of living organisms such as birds and bees and their principles. At this time, the cluster is called swarm, and each individual object in the cluster is called particle.

In the PSO algorithm, each particle means location information of a solution to be searched for and finds an optimal location while moving this location. The movement of each particle position is determined by calculating the direction and speed of movement of each particle using the position of the best particle in the swarm and the best position information of each particle's trace.

The method of applying and optimizing the PSO algorithm used in the present invention to optimize the selection of feature points, the number of nodes of the hidden layer, the learning rate, and the momentum coefficient for determining whether tracking has occurred is performed in the following five steps. Particles in the PSO algorithm consist of 16 elements, as shown in Table 1.

[Table 1] Particle composition in PSO

[Step 1]: Initialization Step

[Step 1-1] n particles are randomly generated in a search space to form an initial swarm. In the present invention, n = 50 is set, and the search space of each conversion coefficient is set to [-10, 10]. That is, 50 particles were randomly generated in the range of [-10, 10].

[Step 1-2] Randomly generate initial particle velocity within [-v _jmax , v _jmax ]. At this time, v _jmax means the maximum moving speed of the j-th element of the particle and set to 20% of the search range.

[Step 1-3] Evaluate the evaluation index by applying each particle of the initial swarm to the fuzzy controller and set it to pbest. In the present invention, a mean squared error (MSE) defined as Equation 3 is used as an evaluation index.

Where m is the number of data, y _i is the actual output, and y _i ^* is the output of the artificial neural network.

[Step 1-4] Find the optimal value among the evaluation index of the initial particle and select it as gbest.

[Step 2]: Inertia Load and Speed Compensation

[Step 2-1] Next, calculate the inertia load value using [Equation 4] below. Here, t means the number of repetitions, set wmax = 0.9, wmin = 0.4, maxgen means the maximum number of repetitions in the PSO algorithm was set to 150 in the present invention.

[Step 2-2] Next, calculate the speed of each particle using Equation 5 below.

Where v _jk (t) is the velocity for the j th element of the k th particle in the t th iteration, and c ₁ , c ₂ is 2.0 as an arbitrary constant. In addition, r ₁ and r ₂ are values between 0 and 1, and random values are generated every time [Formula 5] is calculated, and x _jk is the value of the j th element of the k th particle.

[Equation 5] is a formula for determining the amount of change in the direction and speed of moving the particles, the value of pbest which is the best position in the movement path of each particle and gbest is the position of the best particle is applied.

[Step 3]: Each particle New locations for

When the [Step 2] is completed, the new position of each particle is determined using Equation 6 below. v _jk (t) is the velocity calculated from [Step 2-2], and this value is added to the current particle. At this time, if each element of x _jk (t + 1) that is newly calculated is out of the search range [-10, 10], it is corrected to the upper or lower limit of the search range.

[Step 4] particle Evaluation and pbest , gbest  renewal

[Step 4-1] The evaluation index is obtained by performing evaluation on each particle whose position is updated in the same manner as in [Step 1-3].

[Step 4-2] The pbest of each particle is updated using the evaluation index obtained in [Step 4-1].

At this time, the evaluation index of each particle is compared with the previous pbest. If the evaluation index of the previous pbest is larger than the evaluation index of the new particle, the new particle is reset to pbest.

[Step 4-3] Compare the best index of pbest particles with the best index of gbest, and reset the gbest if it has less evaluation index.

[Step 5] End Assessment

If the termination condition, that is, t> magent, the program is terminated. Otherwise, the process of [Step 2] to [Step 4] is repeated. When the execution of the PSO algorithm is finished, the finally generated gbest has the best location information.

According to the execution of the PSO algorithm as described above, the control unit 31 can obtain an optimal value showing the best learning result while changing the type of feature points, the number of nodes of the hidden layer, the learning rate and the momentum coefficient.

In the present embodiment, in the case of using two to five feature points for the design of the artificial neural network to determine whether the tracking occurs, by applying the above-described PSO algorithm directly, the results as shown in Table 2 below. Could get

At this time, the selected data number means each number of the aforementioned feature points. For example, when two feature points are used, the selected data numbers 3 and 6 have an asymmetry of Low_peak of <feature point 3> and feature point 6, respectively. Means.

<Table 2> Actual results of PSO algorithm

As can be seen from the result of <Table 2>, when applying the tracking detection method according to the present invention, four of the 12 feature points extracted from the input current waveform (that is, <Feature point 1>, <Feature point 2> , <Feature 3> and <Feature 6>) are input to the artificial neural network, and the number of nodes in the hidden layer is set to 9, the learning rate is 0.0102, and the momentum coefficient is set to 00327. Reasoning ability was evaluated to be the best.

On the other hand, if the type of the feature point to be used as an input, the number of nodes of the hidden layer, the learning rate and the momentum coefficient are optimally determined by the above process, the controller 31 uses each of the optimized values by the error backpropagation algorithm. The weight of the connection between nodes is determined by learning.

In this embodiment, in the learning process of the artificial neural network, the connection weight value was learned by setting the target value to -1 for the data in which the tracking was generated and the target value to 1 for the steady state data.

As described above, once the learning weight of the artificial neural network is completed and stored in the memory 32, when the actual input data is input from the feature point detecting unit 20, the controller 31 controls the artificial to which the learned connection weight is applied. The output value of the neural network algorithm is used to determine whether tracking of the inputted real data has occurred.

At this time, if the output value of the artificial neural network algorithm is greater than 0, the controller 31 determines that the current waveform is a steady state in which no tracking is generated, and if the output value of the artificial neural network algorithm is 0 or less than 0, the tracking is generated. You will be judged.

On the other hand, the control unit 31 transmits a control signal for displaying the determination result to the display unit 33 to enable the monitor to quickly and accurately recognize whether or not the tracking occurs.

Therefore, the tracking detection apparatus and method using the artificial neural network according to the present invention has the advantage that it is possible to prevent the damage caused by the tracking by detecting whether the tracking occurs by determining the current waveform actually input in real time.

In this case, the display unit 33 may be preferably implemented by various types of display devices such that a monitor, such as a display device, an alarm lamp device, or an alarm sound device, may recognize a determination result of the controller 31.

1A to 1D are diagrams showing various types of current waveforms when and when tracking is not generated in indoor wiring;

Figure 2 is a block diagram showing the configuration of a tracking detection apparatus using an artificial neural network according to an embodiment of the present invention, and

3 is a view showing the configuration of the artificial neural network used in the tracking detection device using an artificial neural network according to an embodiment of the present invention.

Claims (6)

A current waveform detector for detecting a current waveform at the distribution panel or the load terminal; A feature point detector for detecting at least one feature point associated with tracking from the detected current waveform; And It includes a tracking detection unit for determining whether tracking has occurred in the current waveform is detected according to the output value obtained by substituting the detected feature point in the artificial neural network has completed the learning, and displays the determination result to the outside, Each of the feature points is data related to an irregular waveform in the form of an impulse that appears in the current waveform when tracking occurs. The learning is performed by supervised learning to identify whether tracking has occurred by inputting the feature points respectively detected from a plurality of current waveforms for the case where tracking has not occurred and when the tracking has occurred. Tracking detection device using artificial neural network. The method of claim 1, The tracking detection unit in the learning, using a PSO algorithm to determine the number of hidden layer nodes, the learning rate and the momentum coefficient of the artificial neural network tracking apparatus using an artificial neural network. The method of claim 2, The feature point is a plurality, The learning tracking unit, the tracking detection unit using the artificial neural network, characterized in that the input of the artificial neural network by selecting a feature point of the plurality of feature points using a PSO algorithm. Detecting a current waveform at a distribution panel or at a load end; A second step of detecting at least one feature point related to tracking from the current waveform detected in the first step; And And a third step of determining whether tracking has occurred in a region where the current waveform is detected according to an output value obtained by substituting the feature point detected in the second step into an artificial neural network having completed learning. Each of the feature points is a value associated with an irregular waveform in the form of an impulse that appears in the current waveform when tracking occurs. The learning is performed by supervised learning to identify whether tracking has occurred by inputting the feature points respectively detected from a plurality of current waveforms for the case where tracking has not occurred and when the tracking has occurred. Tracking detection method using artificial neural network. The method of claim 4, wherein In the learning, tracking detection method using an artificial neural network, characterized in that for determining the number of hidden layer nodes, learning rate and momentum coefficient of the artificial neural network using a PSO algorithm. The method of claim 5, The feature point is a plurality, At the time of learning, a tracking detection method using an artificial neural network, characterized in that the input of the artificial neural network by selecting a feature point of the plurality of feature points using a PSO algorithm.

Images