JPH03218558A - Time series data inputting system to neuro-computer - Google Patents

Abstract

PURPOSE:To execute the recognition of time series data generally, highly precisely and efficiently by extracting the time series data required for the recognition of the time series data based on a window function generated beforehand, and inputting this time series data to the input layer of a neuro-computer. CONSTITUTION:An arithmetic unit 4 extracts the time series data required for the recognition of the inputted time series data based on the window func tion generated beforehand, and inputs these extracted time series data required for the recognition to the input layer constituting the neuro-computer 5. Then, after executing the learning by inputting a teacher signal to an output layer in addition, the inference of the event of coming time, etc., is executed. Thus, by extracting the data required for the recognition of the time series data detected by a sensor, etc., based on the window function, and inputting it to the neuro-computer, the recognition of the time series data can be executed generally, highly precisely and efficiently.

Description

[Detailed Description of the Invention] [Summary] Regarding a time-series data input method for inputting time-series data into a neurocomputer, time-series data detected by a sensor, etc. is extracted using a window function and inputted into a neurocomputer, and the time-series data is input into a neurocomputer. With the aim of making recognition general, highly accurate, and efficient, we extract the time series data necessary for recognition based on a window function generated in advance for time series data, and Configure it so that time series data is input to the human-powered layer of the neurocomputer. [Industrial Application Field] The present invention relates to a time-series data input method for inputting time-series data into a neurocomputer. It is hoped that neurocomputers will be able to handle not only spatial pattern data such as pattern tJk, but also time series data that is constantly input from sensors, thereby expanding the range of applications. [The conventional technology and the invention try to solve the problem!] 613 Conventionally, neurocomputers have dealt with spatial patterns. However, when neurocombiators are introduced, they need to process temporally changing patterns in order to be used more widely. The present invention extracts time series data detected by sensors etc. using a window function and inputs it into a 22P computer, thereby making it possible to recognize time series data in a general, highly accurate and efficient manner. [Means for solving the problem] The means for solving the problem will be explained with reference to Figure 1. In Figure 1, time-series data changes over time. The computer unit 4 extracts the time series data necessary for recognition based on the window function generated in advance for the time series data, which is data detected by sensors etc. The neurocomputer 5 extracts the time series data necessary for recognition. It is a neurocomputer (e.g., a three-layer neurocomputer) that performs learning by inputting time-series data into an input layer and inputting a teacher signal into an output layer, and then performs inference, etc. As shown in Figure 1, Enka Uniso} 4 extracts the time series data necessary for recognition based on the function generated in advance for the input time series data, and the input layer forming the neurocomputer 5 extracts the time series data necessary for recognition. The extracted time series data necessary for recognition are input to the input layer.Then, a teacher signal is also input to the output layer to perform learning.
After that, we make inferences about the next time's events.

Therefore, by extracting the data necessary for recognition using a window function from time-series data detected by sensors, etc., and inputting it to a neurocomputer, it is possible to recognize time-series data generally, with high precision, and efficiently. becomes possible.

【Example〕

Next, the configuration and operation of one embodiment of the present invention will be explained in detail using FIGS. 1 to 3. In Figure 1,
Time-series data is data that changes over time, and is, for example, temperature data that changes over time as measured by a temperature sensor for each part of a blast furnace. Sampling Uninote 1 is a Uninote that synchronizes analog time series data with a sampling signal at predetermined intervals, extracts (samples) the value at that time, and converts it into digital time series data. The time-series data storage unit 2 stores the time-series data sampled by the sampling unit 1. = tz number q formation unit, t3 is a unit that generates a window function for extracting time series data necessary for the neurocomputer 5 to recognize (for example, Fig. 2 (mouth) generates a window function Uninote).

The calculation unit 4 calculates the window function generated by the window function generation unit 3 on digital time series data, and extracts the time series data necessary for recognition.

At this time, the time series data after the process shown in FIG. 2 (c) may be generated by extracting only a predetermined portion from the time series data using a window function used in the process shown in FIG. Alternatively, time series data may be generated by weighting and extracting a predetermined portion. The neurocomputer 5 inputs the time series data extracted by the arithmetic unit 4 to the input layer, inputs the teacher signal to the output layer, learns the current event by the so-called bank propagation method, and learns the event at the next time. It is a computer that has learning functions such as making predictions. Next, the operation of the configuration shown in FIG. 1 will be explained in detail using FIGS. 2 and 3. In FIG. 2, the horizontal axis represents time t, and the vertical axis represents, for example, the temperature X of a certain part of the blast furnace.

(11 Figure 2 (A) Figure L: Input the original analog time series data to the sampling unit 1, sample it, convert it to digital time series data, and save it in the time series data storage unit 2. 》 For example, the window function in Figure 2 (W) generated by the window function generation UniSoto 3 is calculated on the digital time series data taken out from the time series data storage UniSoto 2, and the window function as shown in Figure 2 (W) is calculated. 》Extract the time series data in the thick line part and send it to each node of the input layer of the neurocomputer 5 as λ,.,,λ,1,...λ
、． ．． Input (give) as . +31'! As shown in FIG. 3, each node of the input layer of the neurocomputer 5 has λ,. ,,λ(11 '''λ,,,,) is given in {2} above, and the current situation (for example, normally corresponding data) is given as a teacher signal to the output layer, and the buffer is calculated based on the error at that time. , and repeats adjustments such as weighting between each layer using the provagation method. After performing this series of learning, inference is made about the situation from the current time to the next time.

As described above, by inputting the time-series data necessary for recognition extracted by applying a window function to the time-series data into the Neuro Combiator, the time-series data can be improved. It becomes possible to efficiently obtain a high recognition rate in ml. FIG. 2 shows an explanatory diagram of the operation of the present invention. Figure 2 (a) shows the original time series data. The horizontal axis represents time t, the vertical axis represents temperature X, and 0 represents the current time. This time series data shows temperature changes in a certain part of the blast furnace. Figure 2 (portion) shows the window function used for processing. this is,
The window function from time 1 λ to current time 0 is 1. The window function may have a value other than 1 to provide weighting. FIG. 2(c) shows time series data after processing.

This is done from the original time series data in Figure 2 (A) to Figure 2 (A).
(Example) The time series data extracted by the window function used for processing is shown, and the thick line part is the extracted time series data, which is input to the neurocomputer 5 as input NL. (1) The window function is determined in advance by the administrator empirically based on time-series patterns in which abnormalities were discovered in the past. (I) It is different from -λ of the window function in Figure 2 (A), but for example, regarding the temperature X in a certain part of the blast furnace in Figure 2 (A), it is Time t1,
There is a time t2 when the minimum value is reached, and the nearest maximum value t is determined from this time t. And one τ=Q−t,
Find the value of −τ of the window function as . This interval from -τ to O becomes the time series data necessary to confirm whether a certain part of the blast furnace is normal or abnormal. FIG. 3 shows an example of the configuration of a 221 computer according to the present invention. It consists of 3N layers: input layer, intermediate layer, and output layer, and each layer has each node indicated by O. Here, each node of the input layer has time series data λl.corresponding to the thick line portion in FIG. 2(C). ,,λ,,,,λ. ,...
・λ. ．． , as shown in the figure, and use the pack propagation method to calculate the weights between the input layer and the hidden layer, and between the hidden layer and the output layer, based on the error signal that is the difference between the output of the output layer and the teacher signal. Adjust W, etc., and repeat the learning process. After completing the study, 1! Based on poem M, inferences are made at the next time (inferences about whether the blast furnace situation is normal or abnormal, etc.). [Effects of the Invention] As explained above, according to the present invention, data necessary for recognition is extracted from time-series data based on a window function and inputted to a neurocomputer, so Computers can generally recognize time-series data with high precision and efficiency.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention, FIG. 2 is an explanatory diagram of the operation of the present invention, and FIG. 3 is an example of the configuration of a neurocomputer according to the present invention. In the figure, 1 represents a sampling unit, 2 represents a time series data storage unit, 3 represents a core leap number generation unit, 4 represents a calculation unit, and 5 represents a neurocomputer.

Claims (1)

[Claims] In a time-series data input method for inputting time-series data into a neurocomputer, time-series data necessary for recognition is extracted based on a window function generated in advance for the time-series data. 1. A method for inputting time-series data to a neurocomputer, characterized in that the time-series data is input to an input layer of the neurocomputer.