JPH04271486A - High-speed sorter using neural network - Google Patents

Abstract

PURPOSE:To obtain a high-speed sorter to discriminate many kinds of signals with the level and sort fast the signals at the time of the monitoring and control of a plant, etc. CONSTITUTION:By receiving plural input signals, adding these to an operational element 11, setting those signal levels so that the more those signal levels can be close to a designated value A, the larger the signal levels can be, and when these signals are sorted and the number of pieces to be extracted is M, adding a bias value B corresponding to the number M of the pieces by a multiplication element 12 and inputting the bias value to the neural network, the parallel processing is performed, and the signals of the number of pieces close to the value designated by the signal level and designated can be sorted approximately instantaneously.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-speed sorter using a neural network that discriminates a large number of similar signals according to their levels and utilizes the information processing for monitoring and controlling plants.

0002

[Prior technology] Gas turbine inlet temperature control, boiler evaporator tube temperature control, fuel cell cell stack temperature control, fuel cell reformer temperature control, heating furnace temperature control, chemical reaction tank (pipe) temperature control, room temperature control, water pipe network When controlling a distribution system such as pressure control, you can take in many analog input signals of the same type and monitor them using signals with high, low, or intermediate signal levels, or use them as a control variable. In some cases, the controlled object may be controlled by operating the operating end of the device.

[0003] These control systems require signal processing to find signals with large, small, or intermediate signal levels (eg, current values, voltage values) among a large number of analog input signals. This signal processing is usually performed by repeatedly comparing the signal levels of two input signals. Also, when checking whether the signal level of the analog input signal exceeds the upper limit value or falls below the lower limit value,
It is necessary to compare the signal level of the input signal with an upper limit value or a lower limit value for each input signal.

0004

[Problem to be Solved by the Invention] In the above method, iterative processing is required to obtain a signal with a high signal level, a low signal, an intermediate signal, or a signal close to a specified value. ,it takes time. Therefore, if such signal processing is employed, an extra delay is added to the feedback control loop, making it impossible to improve controllability.

[0005] Particularly when high-speed control is required, delays caused by this signal processing cannot be tolerated. In this case, it is also conceivable to select one of the many input signals and use it as the control amount. However, if the sensor that outputs the selected input signal becomes abnormal, or if the selected input signal fluctuates differently from other input signals, the entire distribution system will be controlled based on the selected input signal. On the contrary, there is a risk of disturbing the distribution system.

The present invention has been made in view of the above circumstances, and utilizes the parallel processing properties of a mutually inhibiting neural network to eliminate or minimize the repetitive operation of the signal processing. The purpose is to provide a high-speed sorter using a neural network that can sort signals at high speed.

[0007]

[Means and operations for solving the problems] In order to achieve the above object, the present invention provides a plurality of input signals Xi (i=1, 2
, 3, ...), and their signal levels are set to the specified value A.
An input preprocessing element that converts the signal so that the signal level becomes larger as it approaches , and when the number of signals to be sorted and extracted is M, a bias value B corresponding to this number M is added and sent to the neural net. It was constructed from a neural net using mutual inhibition coupling to determine the number of signals specified by B that have a high signal level among a plurality of input signals.

[0008] With the above configuration, among a plurality of input signals, the number of signals designated by B whose signal level is close to the designated value A can be instantly determined. Note that if the predicted maximum (small) value is used as the value of A, M signals having high (low) signal levels can be determined. Furthermore, if the average value or intermediate value of the input signal level obtained separately is taken as the value of A, M signals close to the average value or intermediate value can be determined. In addition, M=
If it is set to 1, the one input signal closest to the value A is determined.

[0009]

[Embodiment] An embodiment will be described below with reference to the drawings. Figure 2
shows an example of a neural net implemented using analog electronic circuits. In the figure, one amplifier (amplifier) AP1 to APN of the circuit corresponds to the cell body of one neuron in the living body. A feedback line of an inverted voltage vj (hereinafter, a lowercase letter is used as an inverted sign) of an output voltage Vj of the j-th amplifier is connected to an input line of the i-th amplifier via a resistor (its reciprocal conductance is Tij). This connection corresponds to inhibitory synaptic connections in living organisms.

[0010] Then, the current ΣTij flowing through the resistor
・vj (=-ΣTij・Vj) and the current Ii flowing in from the outside are added, and the capacitor Ci and the resistor ri
Grounded via. As a result, the i-th amplifier A
An input voltage Ui is applied to Pi. As the output function of the amplifier AP, a sigmoid function whose input/output static characteristic g(Ui) is shown in FIG. 3 is used, and the amplifier APi outputs a voltage Vi corresponding to the input voltage Ui. Note that the static characteristic g(Ui) is not limited to a sigmoid function, but may be another function with similar characteristics.

The characteristics of the neural network shown in FIG. 2 can be expressed by equations (1) to (4). dUi Ui Ci ・
=- −ΣTij・vj +Ii
......(1)
dt Ri j
1 1 N = + ΣTij

......(2) Ri ri
j=1 Vi (t)=-vi (t)=g[Ui (t)
] ………(3)
1 g(x)=[1+tanh(x/xo)]
......(
4) 2 The neural net in Figure 2 works well when the synaptic connections are symmetrical (
Tij=Tji, Tii=0), (5) The energy function (Lyapunov function) shown in equation (5) operates in the direction of decreasing. It is known that the minimum point of the energy value E in equation (5) corresponds to the attractor (steady state) of the circuit.

1
1
Vi E=-Σ(-Tij)Vi
Vj −ΣIi Vi +Σ ◆ g−1
(v) dV 2 ij
i
i Ri o 1

1 Vi = ΣTi
jVi Vj −ΣIi Vi +Σ ◆
g-1(v)dV2ij
i i
Rio

......(5)
However, the unconverted symbol in the above formula is an integral symbol. Note that the third term on the right side of equation (5) is a small value under normal usage conditions, and therefore will be ignored in the following explanation. Here, this embodiment will be explained as follows. M-1<Ii<M (where M is the number of signals to be sorted) In this case, considering the first and second terms on the right side of equation (5), only M artificial neurons are “1” and the others are If the minimum energy value of “0” is EM, then EM = M(M-1)/2-(M selected in descending order)
sum of Ii) ...(6) EM+1 -EM
=M- (size is Mth + 1st I)
...(7) EM −EM−1 =
M-1- (I whose size is Mth)
...(8) becomes.

Since M-1<Ii<M here, (
7) The formula is positive and remains positive even if M becomes larger, and (8
) is negative and remains negative even if M becomes smaller. Therefore,……>EM+2 >EM+1 >EM <EM
−1 < EM-2 <……, that is, when the input Ii to the neural net is M-1 < Ii < M, only the artificial neuron that receives M inputs selected in descending order is in the state “1”. The state where the rest is "0" becomes a stable state with minimum energy, and at this time, by sensing the artificial neuron that is in the state "1", we can send M number of signals to this neural network with a high signal level. Can judge signals. Note that this neural net operates in parallel,
If the time constant Ci .Ri of the circuit is made sufficiently small, the circuit has the ability to perform the above-mentioned signal sorting instantaneously.

Next, the input signal to this device is Xi (i=
1, 2, 3, ...), the specified value is A, and the input signal Xi
The following signal conversion is also performed for
-A)2 +A2 ......(9
) That is, due to the nature of the quadratic function, the closer the signal level Xi is to A, the larger the value of Yi becomes.

Further, M is the number of signals to be sorted and B is the bias value to be added to the signal Yi, (1) 0<Yi
<1, Ii = Yi +B, (B = M-
1) (2) If 0<Yi <1 is not satisfied, I
If the coefficients α and B are determined so that M-1<Ii<M as i = α・Yi +B, then when the number M of signals to be sorted is specified, the input Ii to the neural net will be M-1<Ii Ii <M is satisfied, and the signal level Xi is A
The closer Ii corresponds to the signal, the higher the signal level becomes. Therefore, by using this Ii as the input signal to the neural network described above, M signals whose signal level is close to the value A among the inputs to the present device can be sorted instantly.

FIG. 1 is a block diagram of an embodiment of a high-speed sorter. This device uses the neural network of mutually inhibiting connections shown in Figure 2.
Equipped with net 2. Neural net 2 is N (N is 2
(an integer greater than or equal to) input terminals and output terminals. A sense element (circuit, block) 3 is connected to the output end (amplifier output) of the neural net 2. The sense element 3 uses the neural net 2 to determine which amplifier output is "1" and outputs the signal number.

The input preprocessing element 1 is composed of an arithmetic element 11, a multiplication element 12, and a bias addition element 13. Input signal (input current) from temperature sensor, pressure sensor, etc.
I to XN are supplied to the calculation element 11 and the multiplication element 12. Calculation element 11 uses specified value A to perform calculation 2A-
Xi and transmits the result to the multiplication element 12.

The multiplier element 12 multiplies the input signal Xi by the signal 2A-Xi from the arithmetic element 11 to obtain the signal Yi (=
(2A-Xi)Xi) as bias addition element 13
Send to. The bias addition element 13 adds a bias B to the signal Yi and transmits it to the neural net 2 as a signal Ii =Yi +B. If it is necessary to standardize the signal Yi, the bias addition element 13 performs the calculation α·Yi +B and transmits it as the signal Ii (α: coefficient).

The neural net 2 operates upon receiving the signals Ii, and corresponds to a steady state of minimum energy according to the signal state, that is, an input signal close to the specified value A specified by the bias value B. The amplifier is at "1" and the others are at "0" instantaneously. This amplifier state is “1”
” is determined by the sense element 3 and output.

Although the above embodiment has been explained assuming a current signal as the input signal, the present invention is not limited to this.
The present invention can be similarly applied to cases where the input signal is a voltage signal or the like. Furthermore, in the above embodiments, the neural net is constructed of analog circuits, but it may also be constructed of digital arithmetic circuits. In this case, the function of the neural network is implemented using the differential equation of equation (1) converted into a difference equation. In this case, if the input signal is an analog signal, it is converted into a digital signal by an A/D converter and processed.

According to this embodiment, the lifetime of equipment in the distribution system, such as the gas turbine inlet temperature, fuel cell reformer temperature, and high-pressure polyethylene reaction tube temperature, must not exceed the limit temperature, but efficiency and When precise temperature control is required to keep the temperature as high as possible from a quality standpoint, by setting the specified value to a high temperature, you can instantly sort high-temperature outputs from a large number of temperature sensor outputs and By determining the average value of the signal and controlling it as a control amount, the distribution system can be controlled well and with high reliability.

[0022] Also, set the specified value to an intermediate value, 1 to 2.
By specifying the number of excluded signals, you can instantly sort the signals excluding signals with a high possibility of input abnormality, and use these to set the highly reliable average value as the control variable.
Good and highly reliable control of the distribution system is possible.

[0023] In addition, when it is necessary to ensure a minimum water pressure in any area, such as in a water distribution network, setting the specified value to a low value near the limit value can prevent water pipes from breaking, etc. By sorting several pressure values in the vicinity of the limit value excluding abnormal values and operating the water supply equipment using their average value as the control amount, economical and favorable pressure control of the water distribution pipe network can be achieved. Furthermore, by using the attribute values of parts as input signals, it is possible to instantly sort a specified number of parts that have near desired attribute values from among a large number of parts whose attribute values vary.

[0024]

[Effects of the Invention] As explained above, according to the present invention, a plurality of input signals are converted so that the closer they are to a specified value, the higher the signal level is, and a bias value is added based on the number of signals to be sorted and extracted. Next, we added a neural network that uses mutually inhibiting connections to a configuration that performs parallel processing, so that out of a large number of input signals, the signal level is close to the specified value, and the specified number of signals can be processed almost instantaneously. It can be sorted into

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram of an embodiment of a high-speed sorter according to the present invention.

FIG. 2 is a diagram illustrating a configuration of a mutually inhibiting neural network.

[Figure 3] Input/output characteristic diagram of neural net amplifier.

[Explanation of symbols]

1 Input preprocessing element 2 Neural net 3 Sense element 11 Arithmetic element 12 Multiplication element 13 Bias addition element

Claims (1)

[Claims] Claim 1: Generated by receiving a plurality of input signals, converting the signals so that the closer the signal level is to a specified value, the greater the signal level, and then adding a bias value based on the number of signals to be sorted and extracted. an input preprocessing element that transmits a signal, and a neural network using mutual inhibition coupling that determines which of the signals input through the input preprocessing element has a signal level greater than a specified value and a specified number of signals. A high-speed sorter using a neural net, which is characterized by being equipped with a net.