WO2020105812A1

WO2020105812A1 - Prediction system and method on basis of parameter improvement through learning

Info

Publication number: WO2020105812A1
Application number: PCT/KR2019/003775
Authority: WO
Inventors: 김도현
Original assignee: 제주대학교 산학협력단
Priority date: 2018-11-22
Filing date: 2019-04-01
Publication date: 2020-05-28
Also published as: KR20200063338A; KR102225370B1

Abstract

The present invention relates to a prediction system on the basis of parameter improvement through learning, the system comprising: a prediction algorithm module for receiving prediction parameters and generating future data; and a learning module for receiving and learning previous data and current data so as to improve the prediction parameters and inputting the prediction parameters to the prediction algorithm module, thereby realizing an improved prediction algorithm. The prediction algorithm module comprises a Kalman filter for receiving and processing a sensor detection value and a sensor error value of a specific environment, and outputting a specific environment prediction value, and the learning module receives a sensor detection value of at least one specific environment and the specific environment prediction value, forms a learned sensor error value learned by a difference value therebetween, and inputs the learned sensor error value to the Kalman filter. By such a configuration, the Kalman filter can predict temperature more accurately by means of the learned sensor error value, and accordingly, an indoor environment can be more efficiently controlled by using the predicted temperature predicted by the Kalman filter.

Description

Prediction system and method based on parameter improvement through learning

The present invention relates to a prediction system, and more specifically, in order to improve the performance of a prediction system through learning, improves prediction parameters by learning prediction algorithm-based prediction parameters using previous and current data, and uses the data Predicting, relates to a prediction system and method based on parameter improvement through learning.

Conventionally, an optimization technique based on prediction of indoor environment parameters has various prediction algorithms to predict indoor environment parameters from a sensor, and a Kalman filter is typically used.

These prediction algorithms contain key parameters and are commonly used in a variety of applications to accurately estimate system conditions under noisy conditions.

On the other hand, data read from the sensor usually has noise that not only affects accuracy, but also causes data spikes. For example, the Kalman filter can intelligently guess the actual state of the system using the previous state (that is, it is not necessary to keep all historical data). This speculation is performed by the Kalman's gain (often referred to as 'K'), which defines whether to give more preference to sensor readings or system self-prediction to accurately estimate the actual state of the system.

By the way, in the indoor environment control system, the sensor value of the indoor environment is received and the sensor value of the indoor environment is learned to control the driver (actuator) of the indoor environment according to the learned sensor value, and the Kalman filter is used in the indoor environment control system. Learning skills are not developed.

Therefore, the present invention was devised to solve the above-mentioned problems, and provides improved data prediction performance results by inputting previous and current data into a learning algorithm to derive improved prediction parameters and inputting them into the prediction algorithm. . Based on the parameter improvement through learning to realize the improved Kalman filter algorithm by deriving the improved prediction parameters by inputting the previous data and the current data into the learning algorithm, and using, for example, the improved Kalman filter, a representative prediction algorithm. It is to provide a prediction system and method.

The configuration according to the first aspect of the present invention for achieving the above object relates to a prediction system based on parameter improvement through learning, a prediction algorithm module that receives prediction parameters and generates future data, and previous data and current It is characterized by realizing an improved prediction algorithm, including a learning module that receives data and learns to improve prediction parameters to input the prediction algorithm module.

The prediction algorithm module includes a Kalman filter that receives and processes sensor detection values and sensor error values of a specific environment to output a specific environment prediction value, and the learning module includes at least one sensor detection value of the specific environment and the specific It is characterized by receiving an environmental prediction value, learning from the difference value, forming a learned sensor error value, and inputting it into the Kalman filter.

Here, the learning module preferably includes a neural network and a sensor error calculation unit that calculates the sensor error value using a learned sensor error value output from the neural network and a preset error factor.

A sensing value of a temperature sensor, a sensing value of a humidity sensor, and a predicted temperature value output from the Kalman filter are input to the input layer of the neural network, and the Kalman filter calculates a predicted temperature using a state variation matrix and a control matrix. Prediction temperature calculation unit, sensor value reading unit in a specific environment, sensor error value reading unit learned in the learning module, covariance value calculation and updating unit for calculating and updating covariance values, and outputting the learned sensor error value and the covariance A Kalman gain calculation unit that receives and calculates a value to output a Kalman gain, a sensor value of the sensor value reading unit, a predicted temperature from the predicted temperature calculation unit, and an actual temperature for estimating a temperature based on the Kalman gain to output a predicted temperature It is preferable to include an estimation unit.

Here, the equation for calculating the predicted temperature by the predicted temperature calculator

And the expression for calculating the covariance and updating unit P _t

And the equation for calculating the updated covariance (P _predicted ) is

Is, the equation for calculating the Kalman gain

Is, the equation for estimating the actual temperature by the actual temperature estimator,

It is preferred.

Where A: is the state transition matrix, A ^T is the transpose of the state transition matrix, B is the control matrix, T _t _{-1 is the} previously calculated temperature, u _{t is the} control vector, T _t : the current sensor Temperature, I: as the identification metric, used to facilitate matrix multiplication, H: the observation metric, z _t : the sensor input (readout), and K: Kalman gain.

The configuration according to the second aspect of the present invention for achieving the above object is in a prediction method based on parameter improvement through learning, a first step in which a prediction algorithm module receives prediction parameters and generates future data, learning And a second step in which the module receives and learns previous data and current data to improve prediction parameters and inputs them to the prediction algorithm module.

Here, the prediction algorithm module includes a Kalman filter that receives and processes sensor detection values and sensor error values of a specific environment to output a specific environment prediction value, and in the second step, the learning module includes at least one specific It is preferable to receive the sensor detection value of the environment and the specific environment prediction value, and learn by the difference value to form a learned sensor error value and input it to the Kalman filter.

According to the prediction system and method based on parameter improvement through learning composed of the above configuration, when the sensor error value is learned by the learning module and input to the Kalman filter, which is an example of the prediction algorithm module, the Kalman filter receiving the learned sensor error value The temperature is accurately predicted, and accordingly, the indoor environment can be more efficiently controlled using the predicted temperature predicted by the Kalman filter of the prediction system.

1 is a conceptual diagram of a prediction system based on parameter improvement through learning according to the present invention;

2 is a configuration diagram of a prediction system based on parameter improvement through learning with a learning-based Kalman filter according to the present invention,

3 is a detailed block diagram of the Kalman filter and the learning module of FIG. 2,

4 is a graph of a result of predicting temperature of a conventional Kalman filter according to various sensor error values;

5 is a graph of temperature prediction results of a Kalman filter learned according to various error factor values,

6 is a sensor error learning result interface screen in a learning-based Kalman filter according to the present invention,

7 is a learning-based Kalman filter prediction result interface screen according to the present invention,

8 is a prediction result interface screen of a learning-based Kalman filter having different sensor error values according to the present invention.

Advantages and features of the present invention, and a method of achieving the same will be described through embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. However, the present embodiments are provided to explain in detail that the technical spirit of the present invention can be easily carried out to a person having ordinary knowledge in the technical field to which the present invention pertains.

In the drawings, the embodiments of the present invention are not limited to the specific form shown and are exaggerated for clarity. In addition, parts indicated by the same reference numerals throughout the specification represent the same components. In this specification, the expression "and / or" is used to mean including at least one of the components listed before and after. In addition, the singular form also includes the plural form unless otherwise specified in the phrase. Also, components, steps, operations and elements referred to as “comprising” or “comprising” as used herein mean the presence or addition of one or more other components, steps, operations, elements and devices.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings, preferred embodiments of the present invention. In the description of the present invention, when it is determined that detailed descriptions of related known functions or configurations may unnecessarily obscure the subject matter of the present invention, detailed descriptions thereof will be omitted.

The prediction system based on parameter improvement through learning according to the present invention is a learning module, an indoor environment that receives control signals of various sensors, controllers, and controllers including temperature and humidity of an indoor environment that provides a sensing signal to the learning module. It includes various actuators. Various system learning algorithms such as an artificial neural network (ANN) can be used for the learning module. More specifically, the learning-based environment control system of the present invention includes a Kalman filter that receives and processes sensor detection values and sensor error values of a specific environment to output a specific environment prediction value, and sensor detection values of at least one specific environment and the A learning module that receives a specific environment prediction value, learns from the difference value to form a learned sensor error value, and inputs it into the Kalman filter, and a predetermined actuator in the indoor environment based on the specific environment prediction value output from the Kalman filter. And a controlling controller.

Sensors can include various sensors to collect data about other environmental parameters, such as temperature, humidity, and the like, from the environment.

In the present invention, it is assumed that a variable error of a temperature sensor value (read value) due to a change in humidity among various sensors. The learning-based environment control system according to the present invention includes a learning module based on a neural network to estimate the amount of error in the current sensor reading and update the sensor error 'R' of the Kalman filter accordingly. The learning module learns the sensor error 'R' input to the Kalman filter. This sensor error 'R' means the difference between the actual temperature and the sensed (measured) temperature.

In the present invention, two main modules are included: a Kalman filter algorithm module and a learning algorithm module. Data is collected from sensors in the environment and input to the learning algorithm module and the Kalman filter algorithm module.

In the Kalman filter, the prediction parameter is most affected by the gain (K). When obtaining the gain (K) of the Kalman filter, it is an important parameter, so if you obtain it correctly, the prediction is made accurately.

The present invention is to derive an improved prediction parameter by inputting previous and current data into a learning algorithm, and inputting it into a prediction algorithm, so that the prediction algorithm provides improved data prediction performance results. More specifically, the advanced Kalman filter algorithm is derived by inputting previous data and current data into the learning algorithm to derive improved prediction parameters (eg, Kalman filter parameters), and using the improved prediction parameters of the representative Kalman filter, which is a representative prediction algorithm. It is to provide a prediction system based on parameter improvement through learning to realize.

1 is a conceptual diagram of a prediction system based on parameter improvement through learning according to the present invention, and FIG. 2 is a configuration diagram of a prediction system based on parameter improvement through learning using a Kalman filter according to the present invention, and FIG. 2 is a detailed block diagram of the Kalman filter and learning module.

As illustrated in FIG. 1, the prediction system 1 based on parameter improvement through learning receives prediction values of various sensors 30 of the environment 5 and outputs prediction values (Kalman filter algorithm) 10 ), Learning algorithm 20 that receives the previous data and the current data (environmental sensor error value) and outputs the learned prediction parameters (Kalman filter parameters).

The controller 4 receives the predicted values output from the prediction algorithm (Kalman filter algorithm) 10 and controls the driver 52 of the environment 5 including various sensors 30.

2, the current sensing value of the temperature sensor is input to the Kalman filter algorithm 10, and a predicted value (future value) is output. Then, the current detection value and the predicted value of the temperature sensor 31 and the detected value of the humidity sensor are input to the learning algorithm (neural network) 20.

As shown in FIG. 3, the Kalman filter includes a sensor value reading unit 12 in a specific environment, a sensor error reading unit 11 reading a sensor error value learned from the learning module 20, and calculating and updating a covariance value. (13) A Kalman gain calculation unit (15) that receives the calculated sensor error values and covariance values and calculates them to output Kalman gain, and a predicted temperature calculation unit that calculates the predicted temperature using the state variation matrix and the control matrix ( 17), an actual temperature estimation unit 18, a sensor error (R) calculation unit 204.

According to the present invention, a neural network is used as a learning module to predict an error rate. Based on the predicted error rate, equation (1) is used to set the R value (error prediction value), which is the input value of the Kalman filter algorithm.

R = Err _pre / C --- Equation 1

Here, C is an error factor and an error value before Err _pre .

State transition metrics and control metrics are well-known techniques mainly used in Kalman filter technology, so detailed descriptions thereof will be omitted.

The actual temperature estimating unit 18 receives the sensor value read by the sensor value reading unit 12, the predicted temperature from the predicted temperature calculating unit 17, and the Kalman gain from the Kalman gain calculation unit 15, and receives the temperature. Estimate and output the estimated temperature T _t .

Meanwhile, it is assumed that the estimated temperature is T _t at the time 't'. The estimated temperature T _t can be used to predict the temperature of t + 1 hour next time. The T _t ₊₁ stepwise calculation using the Kalman filter algorithm is as follows. First, the predicted temperature is calculated from the previously estimated values as follows.

--- Equation 2

Here, A is the state transition matrix, and A ^T is the transpose of the state transition matrix. B is the control matrix, T _t-1 is the previously calculated temperature and u _t is the control vector.

Next, the predicted covariance factor can be updated by the following equation.

--- Equation 3

P _t means process covariance for time t, and P _predicted means updated covariance using the previous covariance and the estimated error being processed.

Here, P _t- ₁ is the previously calculated covariance, and Q is the error estimated in the process.

Using P _predicted , Kalman gain 'K' is obtained by the following equation.

--- Equation 4

Here, H (H ^T is the transpose of H) is the observation matrix, and R is the estimated error in the measurement.

It is assumed that the current sensor predicted temperature is T _t and the sensor input value (reading value) is z _t . Then, the estimated temperature (T _t) of the current time interval of the Kalman filter is

--- Equation 5

Becomes.

Finally, the covariance coefficient P _t for the next iteration is updated as follows. Covariance is a variance that shows the distribution of two or more variances in relation to each other.

--- Equation 6

I is the identification matrix, which is used to facilitate matrix multiplication.

H is used as the observation matrix. K · H is the product of Kalman gain and observation metrics.

The Kalman filter 10 obtains sensor readings from the temperature sensor and removes noise to predict the actual temperature.

The Kalman gain 'K' value can be adjusted using the process covariance matrix P and the prediction error of the sensor reading 'R' as indicated in the equation.

The learning module 20 performs finding the prediction error in the sensor reading 'R'.

As shown in FIG. 3, the neural network 200 in the learning module 20 includes an input layer 201, a hidden layer 202, and an output layer 203. In the input layer 201 of the neural network 200, sensing values of the temperature sensor, sensing values of the humidity sensor, and estimated temperature T _t of the Kalman filter are input.

That is, the estimated temperature T _{t, which} is the output of the actual temperature estimation unit 18, is fed back and enters the input value of the neural network 200. Accordingly, the input values of the neural network 200 are three values of the sensed (measured) temperature and humidity and the estimated temperature T _t .

In the output layer of the neural network 200, an error value err is output.

The R calculation unit of the learning module 20 calculates by dividing the error value err by the error factor C and outputs the learned sensor error. The learned sensor error is read and input by the sensor error reading unit 11 of the Kalman filter.

The results of evaluating the performance of the learning-based Kalman filter by comparing the conventional Kalman filter algorithm prediction results with the learning effects on the learning-based Kalman filter algorithm results according to the present invention will be described with reference to FIGS. 4 to 8.

For performance evaluation, we compared the prediction results of the existing Kalman filter algorithm with the proposed prediction model learning results, and observed the improvement of the predictability results of Kalman filter algorithm results.

In the case of the conventional Kalman filter, results are collected using various sensor error R values. The optimal value of the sensor error R is not fixed and depends on the available data set. Since it is very difficult to manually select the optimum value for the sensor error R in the Kalman filter, an experiment is performed according to the change in the sensor error R value. Then, it is observed that the prediction accuracy of the Kalman filter changes as the sensor error R changes.

4 shows a graph of a result of predicting temperature of a conventional Kalman filter according to various sensor error values. According to various sensor error values R = 5, R = 10, R = 15, and R = 20, the conventional Kalman filter has a different temperature prediction value.

On the other hand, according to the present invention, after training in a learning module composed of an artificial neural network (ANN), the trained module is used to appropriately adjust the variable R to improve the performance of the Kalman filter algorithm.

To obtain the prediction error R, we select the appropriate value of the error factor C as the proportionality constant given in the formula R = err / C. Therefore, the experiment is performed with various values of the error factor C.

5 shows prediction results of a Kalman filter algorithm including a learning module as a variable value of the error factor C. That is, Figure 5 shows a graph of the temperature prediction result of the Kalman filter to which the learning module is applied.

Table 1 below shows a summary of Kalman filter prediction results with or without learning modules for different values of sensor error R and error factor C.

[Table 1]

In the conventional Kalman filter, it can be seen that when the sensor error R value is changed, the prediction accuracy (RMSE) shown in Table 1 is changed. The smaller the value of the prediction accuracy, the higher the performance of the Kalman filter.

Hereinafter, the prediction accuracy (RMSE) will be described in detail as follows.

The mean square error (MSE) is calculated by dividing the sum of the absolute difference between the actual value and the predicted value by the number of data items, as shown in Equation 8 below, between the actual value and the predicted value in the Mean Absolute Deviation (MAD) equation (Expression 7). Is the sum of the squares of the sum of the absolute differences.

--- Equation 7

--- Equation 8

The Root Mean Square Error (RMSE) is obtained by taking the square root of MSE as shown in Equation 9 below.

--- Equation 9

Here, n is the total number of items in the test data set, Ti is the actual temperature (° C) for the i-th instance of the data set, and (Ti) ^ is the corresponding expected temperature. These statistics provide a single number to quantify prediction accuracy. If the predicted result is closer to the actual temperature, the value corresponding to this statistical value will be small.

Table 2 shows a statistical summary of Kalman filter performance results with or without learning modules.

[Table 2]

The results are summarized for various values of the sensor error R used in the case of experiments performed on a Kalman filter without a learning module.

Similarly, in the statistical summary of Kalman filter prediction results for the ANN learning module, various selection values such as error factor C are presented.

Experiments for evaluating the performance of the Kalman filter algorithm are performed by learning a predictive model with various error factor C values. For comparative analysis, the results of Kalman filters (without learning modules) with various sensor error R values were collected. The results are compared with three statistical measures (ie, Mean Absolute Deviation (MAD), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE).

Comparative analysis of Table 2 shows that the Kalman filter with the proposed learning to predict the results of the error factor C = 0.01 surpasses all other settings for all statistical measurements.

Referring to Table 2, the optimal result for the Kalman filter without a learning module is a sensor error R = 15, which has a prediction accuracy of 2.06 in terms of RMSE. Similarly, the optimal result for the Kalman filter using the learning module was found to be error factor C = 0.01.

Compared with the best and worst case results of the Kalman filter without a learning module, the degree of improvement in prediction accuracy by learning in the prediction model is 6.79% and 15.04%, respectively, based on the RMSE metric.

6 is a sensor error learning result interface screen in a learning-based Kalman filter according to the present invention. That is, as a screen showing the result of the Kalman filter input with the learned sensor error value, the learning result of the neural network in FIG.

On the left side of the interface screen of FIG. 6, the input field 61 into which the sensor error value R of the Kalman filter is input, the input field 62 of the error factor for learning the Kalman filter, and the Kalman filter prediction execution button 63 , A learning start button 64 and a learning stop button of the learning module are provided.

6, the sensor error R of the Kalman filter was set to 5, and the error factor C value was set to 0.01.

The original error graph 66 and the prediction error graph 67 according to the learning result are displayed on the right side of the interface screen of FIG. 6.

When comparing the actual temperature graph 66 of FIG. 6 with the prediction error graph 67 according to the learning result, it can be seen that the predicted error and the actual error by the neural network 200 are well aligned. This shows that the learning module 20 has been fully trained on a given data set. After training with the learning module 20, the performance of the Kalman filter algorithm 10 is improved using this trained learning module.

As can be seen from the description of FIG. 6 and Table 1, the best results can be confirmed in the Kalman filter including the learning module with error factor C = 0.01.

7 is an application interface screen for evaluation of a learning-based Kalman filter algorithm using the learning module according to the present invention.

On the left side of the interface screen of FIG. 7, an input field 71 into which a sensor error value R of a Kalman filter is input, an input field 72 of an error factor C for learning a Kalman filter, and a Kalman filter prediction execution button 73 ), A learning start button 74 and a learning stop button of the learning module are provided.

On the right side of the interface screen of FIG. 7, an actual temperature graph 75, a sensing data graph 76, a Kalman filter prediction result graph 77, and a Kalman filter prediction module 78 having a learning module are displayed.

That is, the interface screen of FIG. 7 shows original temperature data, temperature sensor readings, Kalman filter result data, and learned Kalman filter result data. Sensor readings are expressed as Root Mean Square Error (RMSE).

It is performed using different sensor error R values and corresponding results are collected. The prediction temperature (RMSE) of the predicted temperature using the Kalman filter with sensor error R = 5 is 2.25, the RMSE of the sensor reading is 4.74, and the RMSE of the predicted temperature by the learned Kalman filter is 1.92. That is, the RMSE with the Kalman filter is much better than the RMSE of the sensor reading (error 52.32% reduction). In addition, the RMSE of the predicted temperature of the Kalman filter (learned Kalman filter) having a learning module is superior to the RMSE of the predicted temperature using only the Kalman filter.

It can be seen that the learning module improves the prediction accuracy of the Kalman filter algorithm.

8 shows prediction results of the Kalman filter algorithm using the learning module 20. In the interface screen of FIG. 8, the sensor error R of the Kalman filter is set to 20, and the error factor C on the left is set to 0.01.

It is the result of the Kalman filter 10 in which the error factor C = 0.01 is included in the learning module 20. When the error factor C = 0.01 is used, RMSE (Root Mean Square Error) is 1.82.

As can be seen from the above-described embodiment, the Kalman filter including the learning module shows better performance than the existing Kalman filter algorithm in terms of the root mean square error metric.

That is, when the Kalman filter receives the learned sensor error value, the effect is as follows. The prediction in the Kalman filter 10 is most affected by the gain K. This sensor error value means the difference between the actual temperature and the sensed (measured) temperature. It is an important parameter when calculating the gain (K) of the Kalman filter. If this sensor error value is obtained accurately, prediction is also made accurately.

With this configuration, the Kalman filter can accurately predict the sensor value (for example, temperature). Accordingly, the controller of the prediction system can control the indoor environment more efficiently using the sensor value predicted by the Kalman filter.

In the present invention, in a prediction system and method based on parameter improvement through learning, when a sensor error value is learned by a learning module and input to a Kalman filter that is an example of a prediction algorithm module, the Kalman filter receiving the learned sensor error value is configured to detect the temperature. By accurately predicting, it can be used to more efficiently control the indoor environment using the predicted temperature of the prediction system.

Claims

In the prediction system based on parameter improvement through learning,

A prediction algorithm module that receives prediction parameters and generates future data,

A prediction system based on parameter improvement through learning to realize an improved prediction algorithm, including a learning module that receives and learns previous data and current data to improve prediction parameters and inputs them into the prediction algorithm module.
According to claim 1,

The prediction algorithm module includes a Kalman filter that receives sensor processing values and sensor error values of a specific environment and processes them to output a specific environment prediction value,

The learning module receives at least one sensor detection value of the specific environment and the predicted value of the specific environment, learns by the difference value, forms a learned sensor error value, and inputs it to the Kalman filter. Prediction system based on parameter improvement through.
According to claim 2,

The learning module includes a neural network and a sensor error calculation unit that calculates the sensor error value using a learned sensor error value output from the neural network and a preset error factor, and prediction based on parameter improvement through learning. system.
According to claim 3,

A sensing value of a temperature sensor, a sensing value of a humidity sensor, and an estimated temperature value output from the Kalman filter are input to the input layer of the neural network,

The Kalman filter,

Prediction temperature calculation unit that calculates the predicted temperature using the state transition metrics and control metrics, the sensor value reading unit of a specific environment, the sensor error value reading unit learned in the learning module, calculates the covariance value and updates the updated covariance Covariance value calculation and update unit for calculating a value, Kalman gain calculation unit for receiving and calculating the learned sensor error value and the covariance value to output a Kalman gain, sensor value for the sensor value reading unit and the predicted temperature calculation unit A prediction system based on parameter improvement through learning, including an actual temperature estimator for estimating a temperature from the Kalman gain and predicting temperature from and outputting the estimated temperature.
According to claim 4,

The equation for calculating the predicted temperature by the predicted temperature calculator,

ego,

The equation for calculating the covariance and updating unit P t is
And the equation for calculating the updated covariance (P predicted ) is
And

The equation for calculating the Kalman's gain is calculated by the Kalman's gain calculation unit,

ego,

The equation for estimating the actual temperature by the actual temperature estimator,

ego,

here,

A: State transition matrix

A T : Transpose of state transition matrix

B: Control matrix

T t -1 : previously calculated temperature

u t : control vector

T t : Current sensor temperature

I: Identification matrix, used to facilitate matrix multiplication

H: Observation metrics

z t : Sensor input value (read value)

K: Kalman Gain

Q: estimation error
In the prediction method based on parameter improvement through learning,

A first step of the prediction algorithm module receiving prediction parameters and generating future data;

A second method of improving a prediction parameter by learning a learning module by receiving previous data and current data and inputting the predictive algorithm module to the prediction algorithm module.
The method of claim 6,

The prediction algorithm module includes a Kalman filter that receives sensor processing values and sensor error values of a specific environment and processes them to output a specific environment prediction value,

In the second step, the learning module receives at least one sensor detection value of the specific environment and the specific environment prediction value, learns by the difference value, forms a learned sensor error value, and inputs it to the Kalman filter. A prediction method based on parameter improvement through learning.